LLVM  14.0.0git
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"
21 #include "llvm/IR/ConstantRange.h"
22 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/IntrinsicInst.h"
25 #include "llvm/IR/PatternMatch.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/KnownBits.h"
29 
30 using namespace llvm;
31 using namespace PatternMatch;
32 
33 #define DEBUG_TYPE "instcombine"
34 
35 // How many times is a select replaced by one of its operands?
36 STATISTIC(NumSel, "Number of select opts");
37 
38 
39 /// Compute Result = In1+In2, returning true if the result overflowed for this
40 /// type.
41 static bool addWithOverflow(APInt &Result, const APInt &In1,
42  const APInt &In2, bool IsSigned = false) {
43  bool Overflow;
44  if (IsSigned)
45  Result = In1.sadd_ov(In2, Overflow);
46  else
47  Result = In1.uadd_ov(In2, Overflow);
48 
49  return Overflow;
50 }
51 
52 /// Compute Result = In1-In2, returning true if the result overflowed for this
53 /// type.
54 static bool subWithOverflow(APInt &Result, const APInt &In1,
55  const APInt &In2, bool IsSigned = false) {
56  bool Overflow;
57  if (IsSigned)
58  Result = In1.ssub_ov(In2, Overflow);
59  else
60  Result = In1.usub_ov(In2, Overflow);
61 
62  return Overflow;
63 }
64 
65 /// Given an icmp instruction, return true if any use of this comparison is a
66 /// branch on sign bit comparison.
67 static bool hasBranchUse(ICmpInst &I) {
68  for (auto *U : I.users())
69  if (isa<BranchInst>(U))
70  return true;
71  return false;
72 }
73 
74 /// Returns true if the exploded icmp can be expressed as a signed comparison
75 /// to zero and updates the predicate accordingly.
76 /// The signedness of the comparison is preserved.
77 /// TODO: Refactor with decomposeBitTestICmp()?
78 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
79  if (!ICmpInst::isSigned(Pred))
80  return false;
81 
82  if (C.isZero())
83  return ICmpInst::isRelational(Pred);
84 
85  if (C.isOne()) {
86  if (Pred == ICmpInst::ICMP_SLT) {
87  Pred = ICmpInst::ICMP_SLE;
88  return true;
89  }
90  } else if (C.isAllOnes()) {
91  if (Pred == ICmpInst::ICMP_SGT) {
92  Pred = ICmpInst::ICMP_SGE;
93  return true;
94  }
95  }
96 
97  return false;
98 }
99 
100 /// This is called when we see this pattern:
101 /// cmp pred (load (gep GV, ...)), cmpcst
102 /// where GV is a global variable with a constant initializer. Try to simplify
103 /// this into some simple computation that does not need the load. For example
104 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
105 ///
106 /// If AndCst is non-null, then the loaded value is masked with that constant
107 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
108 Instruction *
110  GlobalVariable *GV, CmpInst &ICI,
111  ConstantInt *AndCst) {
112  Constant *Init = GV->getInitializer();
113  if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
114  return nullptr;
115 
116  uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
117  // Don't blow up on huge arrays.
118  if (ArrayElementCount > MaxArraySizeForCombine)
119  return nullptr;
120 
121  // There are many forms of this optimization we can handle, for now, just do
122  // the simple index into a single-dimensional array.
123  //
124  // Require: GEP GV, 0, i {{, constant indices}}
125  if (GEP->getNumOperands() < 3 ||
126  !isa<ConstantInt>(GEP->getOperand(1)) ||
127  !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
128  isa<Constant>(GEP->getOperand(2)))
129  return nullptr;
130 
131  // Check that indices after the variable are constants and in-range for the
132  // type they index. Collect the indices. This is typically for arrays of
133  // structs.
134  SmallVector<unsigned, 4> LaterIndices;
135 
136  Type *EltTy = Init->getType()->getArrayElementType();
137  for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
138  ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
139  if (!Idx) return nullptr; // Variable index.
140 
141  uint64_t IdxVal = Idx->getZExtValue();
142  if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
143 
144  if (StructType *STy = dyn_cast<StructType>(EltTy))
145  EltTy = STy->getElementType(IdxVal);
146  else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
147  if (IdxVal >= ATy->getNumElements()) return nullptr;
148  EltTy = ATy->getElementType();
149  } else {
150  return nullptr; // Unknown type.
151  }
152 
153  LaterIndices.push_back(IdxVal);
154  }
155 
156  enum { Overdefined = -3, Undefined = -2 };
157 
158  // Variables for our state machines.
159 
160  // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
161  // "i == 47 | i == 87", where 47 is the first index the condition is true for,
162  // and 87 is the second (and last) index. FirstTrueElement is -2 when
163  // undefined, otherwise set to the first true element. SecondTrueElement is
164  // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
165  int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
166 
167  // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
168  // form "i != 47 & i != 87". Same state transitions as for true elements.
169  int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
170 
171  /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
172  /// define a state machine that triggers for ranges of values that the index
173  /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
174  /// This is -2 when undefined, -3 when overdefined, and otherwise the last
175  /// index in the range (inclusive). We use -2 for undefined here because we
176  /// use relative comparisons and don't want 0-1 to match -1.
177  int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
178 
179  // MagicBitvector - This is a magic bitvector where we set a bit if the
180  // comparison is true for element 'i'. If there are 64 elements or less in
181  // the array, this will fully represent all the comparison results.
182  uint64_t MagicBitvector = 0;
183 
184  // Scan the array and see if one of our patterns matches.
185  Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
186  for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
187  Constant *Elt = Init->getAggregateElement(i);
188  if (!Elt) return nullptr;
189 
190  // If this is indexing an array of structures, get the structure element.
191  if (!LaterIndices.empty())
192  Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
193 
194  // If the element is masked, handle it.
195  if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
196 
197  // Find out if the comparison would be true or false for the i'th element.
199  CompareRHS, DL, &TLI);
200  // If the result is undef for this element, ignore it.
201  if (isa<UndefValue>(C)) {
202  // Extend range state machines to cover this element in case there is an
203  // undef in the middle of the range.
204  if (TrueRangeEnd == (int)i-1)
205  TrueRangeEnd = i;
206  if (FalseRangeEnd == (int)i-1)
207  FalseRangeEnd = i;
208  continue;
209  }
210 
211  // If we can't compute the result for any of the elements, we have to give
212  // up evaluating the entire conditional.
213  if (!isa<ConstantInt>(C)) return nullptr;
214 
215  // Otherwise, we know if the comparison is true or false for this element,
216  // update our state machines.
217  bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
218 
219  // State machine for single/double/range index comparison.
220  if (IsTrueForElt) {
221  // Update the TrueElement state machine.
222  if (FirstTrueElement == Undefined)
223  FirstTrueElement = TrueRangeEnd = i; // First true element.
224  else {
225  // Update double-compare state machine.
226  if (SecondTrueElement == Undefined)
227  SecondTrueElement = i;
228  else
229  SecondTrueElement = Overdefined;
230 
231  // Update range state machine.
232  if (TrueRangeEnd == (int)i-1)
233  TrueRangeEnd = i;
234  else
235  TrueRangeEnd = Overdefined;
236  }
237  } else {
238  // Update the FalseElement state machine.
239  if (FirstFalseElement == Undefined)
240  FirstFalseElement = FalseRangeEnd = i; // First false element.
241  else {
242  // Update double-compare state machine.
243  if (SecondFalseElement == Undefined)
244  SecondFalseElement = i;
245  else
246  SecondFalseElement = Overdefined;
247 
248  // Update range state machine.
249  if (FalseRangeEnd == (int)i-1)
250  FalseRangeEnd = i;
251  else
252  FalseRangeEnd = Overdefined;
253  }
254  }
255 
256  // If this element is in range, update our magic bitvector.
257  if (i < 64 && IsTrueForElt)
258  MagicBitvector |= 1ULL << i;
259 
260  // If all of our states become overdefined, bail out early. Since the
261  // predicate is expensive, only check it every 8 elements. This is only
262  // really useful for really huge arrays.
263  if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
264  SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
265  FalseRangeEnd == Overdefined)
266  return nullptr;
267  }
268 
269  // Now that we've scanned the entire array, emit our new comparison(s). We
270  // order the state machines in complexity of the generated code.
271  Value *Idx = GEP->getOperand(2);
272 
273  // If the index is larger than the pointer size of the target, truncate the
274  // index down like the GEP would do implicitly. We don't have to do this for
275  // an inbounds GEP because the index can't be out of range.
276  if (!GEP->isInBounds()) {
277  Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
278  unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
279  if (Idx->getType()->getPrimitiveSizeInBits().getFixedSize() > PtrSize)
280  Idx = Builder.CreateTrunc(Idx, IntPtrTy);
281  }
282 
283  // If inbounds keyword is not present, Idx * ElementSize can overflow.
284  // Let's assume that ElementSize is 2 and the wanted value is at offset 0.
285  // Then, there are two possible values for Idx to match offset 0:
286  // 0x00..00, 0x80..00.
287  // Emitting 'icmp eq Idx, 0' isn't correct in this case because the
288  // comparison is false if Idx was 0x80..00.
289  // We need to erase the highest countTrailingZeros(ElementSize) bits of Idx.
290  unsigned ElementSize =
291  DL.getTypeAllocSize(Init->getType()->getArrayElementType());
292  auto MaskIdx = [&](Value* Idx){
293  if (!GEP->isInBounds() && countTrailingZeros(ElementSize) != 0) {
294  Value *Mask = ConstantInt::get(Idx->getType(), -1);
295  Mask = Builder.CreateLShr(Mask, countTrailingZeros(ElementSize));
296  Idx = Builder.CreateAnd(Idx, Mask);
297  }
298  return Idx;
299  };
300 
301  // If the comparison is only true for one or two elements, emit direct
302  // comparisons.
303  if (SecondTrueElement != Overdefined) {
304  Idx = MaskIdx(Idx);
305  // None true -> false.
306  if (FirstTrueElement == Undefined)
307  return replaceInstUsesWith(ICI, Builder.getFalse());
308 
309  Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
310 
311  // True for one element -> 'i == 47'.
312  if (SecondTrueElement == Undefined)
313  return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
314 
315  // True for two elements -> 'i == 47 | i == 72'.
316  Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
317  Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
318  Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
319  return BinaryOperator::CreateOr(C1, C2);
320  }
321 
322  // If the comparison is only false for one or two elements, emit direct
323  // comparisons.
324  if (SecondFalseElement != Overdefined) {
325  Idx = MaskIdx(Idx);
326  // None false -> true.
327  if (FirstFalseElement == Undefined)
328  return replaceInstUsesWith(ICI, Builder.getTrue());
329 
330  Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
331 
332  // False for one element -> 'i != 47'.
333  if (SecondFalseElement == Undefined)
334  return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
335 
336  // False for two elements -> 'i != 47 & i != 72'.
337  Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
338  Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
339  Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
340  return BinaryOperator::CreateAnd(C1, C2);
341  }
342 
343  // If the comparison can be replaced with a range comparison for the elements
344  // where it is true, emit the range check.
345  if (TrueRangeEnd != Overdefined) {
346  assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
347  Idx = MaskIdx(Idx);
348 
349  // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
350  if (FirstTrueElement) {
351  Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
352  Idx = Builder.CreateAdd(Idx, Offs);
353  }
354 
355  Value *End = ConstantInt::get(Idx->getType(),
356  TrueRangeEnd-FirstTrueElement+1);
357  return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
358  }
359 
360  // False range check.
361  if (FalseRangeEnd != Overdefined) {
362  assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
363  Idx = MaskIdx(Idx);
364  // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
365  if (FirstFalseElement) {
366  Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
367  Idx = Builder.CreateAdd(Idx, Offs);
368  }
369 
370  Value *End = ConstantInt::get(Idx->getType(),
371  FalseRangeEnd-FirstFalseElement);
372  return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
373  }
374 
375  // If a magic bitvector captures the entire comparison state
376  // of this load, replace it with computation that does:
377  // ((magic_cst >> i) & 1) != 0
378  {
379  Type *Ty = nullptr;
380 
381  // Look for an appropriate type:
382  // - The type of Idx if the magic fits
383  // - The smallest fitting legal type
384  if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
385  Ty = Idx->getType();
386  else
387  Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
388 
389  if (Ty) {
390  Idx = MaskIdx(Idx);
391  Value *V = Builder.CreateIntCast(Idx, Ty, false);
392  V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
393  V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
394  return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
395  }
396  }
397 
398  return nullptr;
399 }
400 
401 /// Return a value that can be used to compare the *offset* implied by a GEP to
402 /// zero. For example, if we have &A[i], we want to return 'i' for
403 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
404 /// are involved. The above expression would also be legal to codegen as
405 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
406 /// This latter form is less amenable to optimization though, and we are allowed
407 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
408 ///
409 /// If we can't emit an optimized form for this expression, this returns null.
410 ///
412  const DataLayout &DL) {
414 
415  // Check to see if this gep only has a single variable index. If so, and if
416  // any constant indices are a multiple of its scale, then we can compute this
417  // in terms of the scale of the variable index. For example, if the GEP
418  // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
419  // because the expression will cross zero at the same point.
420  unsigned i, e = GEP->getNumOperands();
421  int64_t Offset = 0;
422  for (i = 1; i != e; ++i, ++GTI) {
423  if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
424  // Compute the aggregate offset of constant indices.
425  if (CI->isZero()) continue;
426 
427  // Handle a struct index, which adds its field offset to the pointer.
428  if (StructType *STy = GTI.getStructTypeOrNull()) {
429  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
430  } else {
431  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
432  Offset += Size*CI->getSExtValue();
433  }
434  } else {
435  // Found our variable index.
436  break;
437  }
438  }
439 
440  // If there are no variable indices, we must have a constant offset, just
441  // evaluate it the general way.
442  if (i == e) return nullptr;
443 
444  Value *VariableIdx = GEP->getOperand(i);
445  // Determine the scale factor of the variable element. For example, this is
446  // 4 if the variable index is into an array of i32.
447  uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
448 
449  // Verify that there are no other variable indices. If so, emit the hard way.
450  for (++i, ++GTI; i != e; ++i, ++GTI) {
451  ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
452  if (!CI) return nullptr;
453 
454  // Compute the aggregate offset of constant indices.
455  if (CI->isZero()) continue;
456 
457  // Handle a struct index, which adds its field offset to the pointer.
458  if (StructType *STy = GTI.getStructTypeOrNull()) {
459  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
460  } else {
461  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
462  Offset += Size*CI->getSExtValue();
463  }
464  }
465 
466  // Okay, we know we have a single variable index, which must be a
467  // pointer/array/vector index. If there is no offset, life is simple, return
468  // the index.
469  Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
470  unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
471  if (Offset == 0) {
472  // Cast to intptrty in case a truncation occurs. If an extension is needed,
473  // we don't need to bother extending: the extension won't affect where the
474  // computation crosses zero.
475  if (VariableIdx->getType()->getPrimitiveSizeInBits().getFixedSize() >
476  IntPtrWidth) {
477  VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
478  }
479  return VariableIdx;
480  }
481 
482  // Otherwise, there is an index. The computation we will do will be modulo
483  // the pointer size.
484  Offset = SignExtend64(Offset, IntPtrWidth);
485  VariableScale = SignExtend64(VariableScale, IntPtrWidth);
486 
487  // To do this transformation, any constant index must be a multiple of the
488  // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
489  // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
490  // multiple of the variable scale.
491  int64_t NewOffs = Offset / (int64_t)VariableScale;
492  if (Offset != NewOffs*(int64_t)VariableScale)
493  return nullptr;
494 
495  // Okay, we can do this evaluation. Start by converting the index to intptr.
496  if (VariableIdx->getType() != IntPtrTy)
497  VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
498  true /*Signed*/);
499  Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
500  return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
501 }
502 
503 /// Returns true if we can rewrite Start as a GEP with pointer Base
504 /// and some integer offset. The nodes that need to be re-written
505 /// for this transformation will be added to Explored.
506 static bool canRewriteGEPAsOffset(Type *ElemTy, Value *Start, Value *Base,
507  const DataLayout &DL,
508  SetVector<Value *> &Explored) {
509  SmallVector<Value *, 16> WorkList(1, Start);
510  Explored.insert(Base);
511 
512  // The following traversal gives us an order which can be used
513  // when doing the final transformation. Since in the final
514  // transformation we create the PHI replacement instructions first,
515  // we don't have to get them in any particular order.
516  //
517  // However, for other instructions we will have to traverse the
518  // operands of an instruction first, which means that we have to
519  // do a post-order traversal.
520  while (!WorkList.empty()) {
522 
523  while (!WorkList.empty()) {
524  if (Explored.size() >= 100)
525  return false;
526 
527  Value *V = WorkList.back();
528 
529  if (Explored.contains(V)) {
530  WorkList.pop_back();
531  continue;
532  }
533 
534  if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
535  !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
536  // We've found some value that we can't explore which is different from
537  // the base. Therefore we can't do this transformation.
538  return false;
539 
540  if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
541  auto *CI = cast<CastInst>(V);
542  if (!CI->isNoopCast(DL))
543  return false;
544 
545  if (!Explored.contains(CI->getOperand(0)))
546  WorkList.push_back(CI->getOperand(0));
547  }
548 
549  if (auto *GEP = dyn_cast<GEPOperator>(V)) {
550  // We're limiting the GEP to having one index. This will preserve
551  // the original pointer type. We could handle more cases in the
552  // future.
553  if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
554  GEP->getSourceElementType() != ElemTy)
555  return false;
556 
557  if (!Explored.contains(GEP->getOperand(0)))
558  WorkList.push_back(GEP->getOperand(0));
559  }
560 
561  if (WorkList.back() == V) {
562  WorkList.pop_back();
563  // We've finished visiting this node, mark it as such.
564  Explored.insert(V);
565  }
566 
567  if (auto *PN = dyn_cast<PHINode>(V)) {
568  // We cannot transform PHIs on unsplittable basic blocks.
569  if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
570  return false;
571  Explored.insert(PN);
572  PHIs.insert(PN);
573  }
574  }
575 
576  // Explore the PHI nodes further.
577  for (auto *PN : PHIs)
578  for (Value *Op : PN->incoming_values())
579  if (!Explored.contains(Op))
580  WorkList.push_back(Op);
581  }
582 
583  // Make sure that we can do this. Since we can't insert GEPs in a basic
584  // block before a PHI node, we can't easily do this transformation if
585  // we have PHI node users of transformed instructions.
586  for (Value *Val : Explored) {
587  for (Value *Use : Val->uses()) {
588 
589  auto *PHI = dyn_cast<PHINode>(Use);
590  auto *Inst = dyn_cast<Instruction>(Val);
591 
592  if (Inst == Base || Inst == PHI || !Inst || !PHI ||
593  !Explored.contains(PHI))
594  continue;
595 
596  if (PHI->getParent() == Inst->getParent())
597  return false;
598  }
599  }
600  return true;
601 }
602 
603 // Sets the appropriate insert point on Builder where we can add
604 // a replacement Instruction for V (if that is possible).
606  bool Before = true) {
607  if (auto *PHI = dyn_cast<PHINode>(V)) {
608  Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
609  return;
610  }
611  if (auto *I = dyn_cast<Instruction>(V)) {
612  if (!Before)
613  I = &*std::next(I->getIterator());
614  Builder.SetInsertPoint(I);
615  return;
616  }
617  if (auto *A = dyn_cast<Argument>(V)) {
618  // Set the insertion point in the entry block.
619  BasicBlock &Entry = A->getParent()->getEntryBlock();
620  Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
621  return;
622  }
623  // Otherwise, this is a constant and we don't need to set a new
624  // insertion point.
625  assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
626 }
627 
628 /// Returns a re-written value of Start as an indexed GEP using Base as a
629 /// pointer.
630 static Value *rewriteGEPAsOffset(Type *ElemTy, Value *Start, Value *Base,
631  const DataLayout &DL,
632  SetVector<Value *> &Explored) {
633  // Perform all the substitutions. This is a bit tricky because we can
634  // have cycles in our use-def chains.
635  // 1. Create the PHI nodes without any incoming values.
636  // 2. Create all the other values.
637  // 3. Add the edges for the PHI nodes.
638  // 4. Emit GEPs to get the original pointers.
639  // 5. Remove the original instructions.
640  Type *IndexType = IntegerType::get(
641  Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
642 
644  NewInsts[Base] = ConstantInt::getNullValue(IndexType);
645 
646  // Create the new PHI nodes, without adding any incoming values.
647  for (Value *Val : Explored) {
648  if (Val == Base)
649  continue;
650  // Create empty phi nodes. This avoids cyclic dependencies when creating
651  // the remaining instructions.
652  if (auto *PHI = dyn_cast<PHINode>(Val))
653  NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
654  PHI->getName() + ".idx", PHI);
655  }
656  IRBuilder<> Builder(Base->getContext());
657 
658  // Create all the other instructions.
659  for (Value *Val : Explored) {
660 
661  if (NewInsts.find(Val) != NewInsts.end())
662  continue;
663 
664  if (auto *CI = dyn_cast<CastInst>(Val)) {
665  // Don't get rid of the intermediate variable here; the store can grow
666  // the map which will invalidate the reference to the input value.
667  Value *V = NewInsts[CI->getOperand(0)];
668  NewInsts[CI] = V;
669  continue;
670  }
671  if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
672  Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
673  : GEP->getOperand(1);
675  // Indices might need to be sign extended. GEPs will magically do
676  // this, but we need to do it ourselves here.
677  if (Index->getType()->getScalarSizeInBits() !=
678  NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
679  Index = Builder.CreateSExtOrTrunc(
680  Index, NewInsts[GEP->getOperand(0)]->getType(),
681  GEP->getOperand(0)->getName() + ".sext");
682  }
683 
684  auto *Op = NewInsts[GEP->getOperand(0)];
685  if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
686  NewInsts[GEP] = Index;
687  else
688  NewInsts[GEP] = Builder.CreateNSWAdd(
689  Op, Index, GEP->getOperand(0)->getName() + ".add");
690  continue;
691  }
692  if (isa<PHINode>(Val))
693  continue;
694 
695  llvm_unreachable("Unexpected instruction type");
696  }
697 
698  // Add the incoming values to the PHI nodes.
699  for (Value *Val : Explored) {
700  if (Val == Base)
701  continue;
702  // All the instructions have been created, we can now add edges to the
703  // phi nodes.
704  if (auto *PHI = dyn_cast<PHINode>(Val)) {
705  PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
706  for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
707  Value *NewIncoming = PHI->getIncomingValue(I);
708 
709  if (NewInsts.find(NewIncoming) != NewInsts.end())
710  NewIncoming = NewInsts[NewIncoming];
711 
712  NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
713  }
714  }
715  }
716 
717  PointerType *PtrTy =
718  ElemTy->getPointerTo(Start->getType()->getPointerAddressSpace());
719  for (Value *Val : Explored) {
720  if (Val == Base)
721  continue;
722 
723  // Depending on the type, for external users we have to emit
724  // a GEP or a GEP + ptrtoint.
725  setInsertionPoint(Builder, Val, false);
726 
727  // Cast base to the expected type.
728  Value *NewVal = Builder.CreateBitOrPointerCast(
729  Base, PtrTy, Start->getName() + "to.ptr");
730  NewVal = Builder.CreateInBoundsGEP(
731  ElemTy, NewVal, makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
732  NewVal = Builder.CreateBitOrPointerCast(
733  NewVal, Val->getType(), Val->getName() + ".conv");
734  Val->replaceAllUsesWith(NewVal);
735  }
736 
737  return NewInsts[Start];
738 }
739 
740 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
741 /// the input Value as a constant indexed GEP. Returns a pair containing
742 /// the GEPs Pointer and Index.
743 static std::pair<Value *, Value *>
745  Type *IndexType = IntegerType::get(V->getContext(),
746  DL.getIndexTypeSizeInBits(V->getType()));
747 
749  while (true) {
750  if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
751  // We accept only inbouds GEPs here to exclude the possibility of
752  // overflow.
753  if (!GEP->isInBounds())
754  break;
755  if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
756  GEP->getSourceElementType() == ElemTy) {
757  V = GEP->getOperand(0);
758  Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
760  Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
761  continue;
762  }
763  break;
764  }
765  if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
766  if (!CI->isNoopCast(DL))
767  break;
768  V = CI->getOperand(0);
769  continue;
770  }
771  if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
772  if (!CI->isNoopCast(DL))
773  break;
774  V = CI->getOperand(0);
775  continue;
776  }
777  break;
778  }
779  return {V, Index};
780 }
781 
782 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
783 /// We can look through PHIs, GEPs and casts in order to determine a common base
784 /// between GEPLHS and RHS.
787  const DataLayout &DL) {
788  // FIXME: Support vector of pointers.
789  if (GEPLHS->getType()->isVectorTy())
790  return nullptr;
791 
792  if (!GEPLHS->hasAllConstantIndices())
793  return nullptr;
794 
795  Type *ElemTy = GEPLHS->getSourceElementType();
796  Value *PtrBase, *Index;
797  std::tie(PtrBase, Index) = getAsConstantIndexedAddress(ElemTy, GEPLHS, DL);
798 
799  // The set of nodes that will take part in this transformation.
800  SetVector<Value *> Nodes;
801 
802  if (!canRewriteGEPAsOffset(ElemTy, RHS, PtrBase, DL, Nodes))
803  return nullptr;
804 
805  // We know we can re-write this as
806  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
807  // Since we've only looked through inbouds GEPs we know that we
808  // can't have overflow on either side. We can therefore re-write
809  // this as:
810  // OFFSET1 cmp OFFSET2
811  Value *NewRHS = rewriteGEPAsOffset(ElemTy, RHS, PtrBase, DL, Nodes);
812 
813  // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
814  // GEP having PtrBase as the pointer base, and has returned in NewRHS the
815  // offset. Since Index is the offset of LHS to the base pointer, we will now
816  // compare the offsets instead of comparing the pointers.
817  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
818 }
819 
820 /// Fold comparisons between a GEP instruction and something else. At this point
821 /// we know that the GEP is on the LHS of the comparison.
824  Instruction &I) {
825  // Don't transform signed compares of GEPs into index compares. Even if the
826  // GEP is inbounds, the final add of the base pointer can have signed overflow
827  // and would change the result of the icmp.
828  // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
829  // the maximum signed value for the pointer type.
831  return nullptr;
832 
833  // Look through bitcasts and addrspacecasts. We do not however want to remove
834  // 0 GEPs.
835  if (!isa<GetElementPtrInst>(RHS))
837 
838  Value *PtrBase = GEPLHS->getOperand(0);
839  // FIXME: Support vector pointer GEPs.
840  if (PtrBase == RHS && GEPLHS->isInBounds() &&
841  !GEPLHS->getType()->isVectorTy()) {
842  // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
843  // This transformation (ignoring the base and scales) is valid because we
844  // know pointers can't overflow since the gep is inbounds. See if we can
845  // output an optimized form.
846  Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
847 
848  // If not, synthesize the offset the hard way.
849  if (!Offset)
850  Offset = EmitGEPOffset(GEPLHS);
852  Constant::getNullValue(Offset->getType()));
853  }
854 
855  if (GEPLHS->isInBounds() && ICmpInst::isEquality(Cond) &&
856  isa<Constant>(RHS) && cast<Constant>(RHS)->isNullValue() &&
857  !NullPointerIsDefined(I.getFunction(),
859  // For most address spaces, an allocation can't be placed at null, but null
860  // itself is treated as a 0 size allocation in the in bounds rules. Thus,
861  // the only valid inbounds address derived from null, is null itself.
862  // Thus, we have four cases to consider:
863  // 1) Base == nullptr, Offset == 0 -> inbounds, null
864  // 2) Base == nullptr, Offset != 0 -> poison as the result is out of bounds
865  // 3) Base != nullptr, Offset == (-base) -> poison (crossing allocations)
866  // 4) Base != nullptr, Offset != (-base) -> nonnull (and possibly poison)
867  //
868  // (Note if we're indexing a type of size 0, that simply collapses into one
869  // of the buckets above.)
870  //
871  // In general, we're allowed to make values less poison (i.e. remove
872  // sources of full UB), so in this case, we just select between the two
873  // non-poison cases (1 and 4 above).
874  //
875  // For vectors, we apply the same reasoning on a per-lane basis.
876  auto *Base = GEPLHS->getPointerOperand();
877  if (GEPLHS->getType()->isVectorTy() && Base->getType()->isPointerTy()) {
878  auto EC = cast<VectorType>(GEPLHS->getType())->getElementCount();
879  Base = Builder.CreateVectorSplat(EC, Base);
880  }
881  return new ICmpInst(Cond, Base,
883  cast<Constant>(RHS), Base->getType()));
884  } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
885  // If the base pointers are different, but the indices are the same, just
886  // compare the base pointer.
887  if (PtrBase != GEPRHS->getOperand(0)) {
888  bool IndicesTheSame =
889  GEPLHS->getNumOperands() == GEPRHS->getNumOperands() &&
890  GEPLHS->getType() == GEPRHS->getType() &&
891  GEPLHS->getSourceElementType() == GEPRHS->getSourceElementType();
892  if (IndicesTheSame)
893  for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
894  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
895  IndicesTheSame = false;
896  break;
897  }
898 
899  // If all indices are the same, just compare the base pointers.
900  Type *BaseType = GEPLHS->getOperand(0)->getType();
901  if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
902  return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
903 
904  // If we're comparing GEPs with two base pointers that only differ in type
905  // and both GEPs have only constant indices or just one use, then fold
906  // the compare with the adjusted indices.
907  // FIXME: Support vector of pointers.
908  if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
909  (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
910  (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
911  PtrBase->stripPointerCasts() ==
912  GEPRHS->getOperand(0)->stripPointerCasts() &&
913  !GEPLHS->getType()->isVectorTy()) {
914  Value *LOffset = EmitGEPOffset(GEPLHS);
915  Value *ROffset = EmitGEPOffset(GEPRHS);
916 
917  // If we looked through an addrspacecast between different sized address
918  // spaces, the LHS and RHS pointers are different sized
919  // integers. Truncate to the smaller one.
920  Type *LHSIndexTy = LOffset->getType();
921  Type *RHSIndexTy = ROffset->getType();
922  if (LHSIndexTy != RHSIndexTy) {
923  if (LHSIndexTy->getPrimitiveSizeInBits().getFixedSize() <
924  RHSIndexTy->getPrimitiveSizeInBits().getFixedSize()) {
925  ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
926  } else
927  LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
928  }
929 
930  Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
931  LOffset, ROffset);
932  return replaceInstUsesWith(I, Cmp);
933  }
934 
935  // Otherwise, the base pointers are different and the indices are
936  // different. Try convert this to an indexed compare by looking through
937  // PHIs/casts.
938  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
939  }
940 
941  // If one of the GEPs has all zero indices, recurse.
942  // FIXME: Handle vector of pointers.
943  if (!GEPLHS->getType()->isVectorTy() && GEPLHS->hasAllZeroIndices())
944  return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
946 
947  // If the other GEP has all zero indices, recurse.
948  // FIXME: Handle vector of pointers.
949  if (!GEPRHS->getType()->isVectorTy() && GEPRHS->hasAllZeroIndices())
950  return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
951 
952  bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
953  if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
954  // If the GEPs only differ by one index, compare it.
955  unsigned NumDifferences = 0; // Keep track of # differences.
956  unsigned DiffOperand = 0; // The operand that differs.
957  for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
958  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
959  Type *LHSType = GEPLHS->getOperand(i)->getType();
960  Type *RHSType = GEPRHS->getOperand(i)->getType();
961  // FIXME: Better support for vector of pointers.
962  if (LHSType->getPrimitiveSizeInBits() !=
963  RHSType->getPrimitiveSizeInBits() ||
964  (GEPLHS->getType()->isVectorTy() &&
965  (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) {
966  // Irreconcilable differences.
967  NumDifferences = 2;
968  break;
969  }
970 
971  if (NumDifferences++) break;
972  DiffOperand = i;
973  }
974 
975  if (NumDifferences == 0) // SAME GEP?
976  return replaceInstUsesWith(I, // No comparison is needed here.
978 
979  else if (NumDifferences == 1 && GEPsInBounds) {
980  Value *LHSV = GEPLHS->getOperand(DiffOperand);
981  Value *RHSV = GEPRHS->getOperand(DiffOperand);
982  // Make sure we do a signed comparison here.
983  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
984  }
985  }
986 
987  // Only lower this if the icmp is the only user of the GEP or if we expect
988  // the result to fold to a constant!
989  if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
990  (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
991  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
992  Value *L = EmitGEPOffset(GEPLHS);
993  Value *R = EmitGEPOffset(GEPRHS);
994  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
995  }
996  }
997 
998  // Try convert this to an indexed compare by looking through PHIs/casts as a
999  // last resort.
1000  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1001 }
1002 
1004  const AllocaInst *Alloca,
1005  const Value *Other) {
1006  assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1007 
1008  // It would be tempting to fold away comparisons between allocas and any
1009  // pointer not based on that alloca (e.g. an argument). However, even
1010  // though such pointers cannot alias, they can still compare equal.
1011  //
1012  // But LLVM doesn't specify where allocas get their memory, so if the alloca
1013  // doesn't escape we can argue that it's impossible to guess its value, and we
1014  // can therefore act as if any such guesses are wrong.
1015  //
1016  // The code below checks that the alloca doesn't escape, and that it's only
1017  // used in a comparison once (the current instruction). The
1018  // single-comparison-use condition ensures that we're trivially folding all
1019  // comparisons against the alloca consistently, and avoids the risk of
1020  // erroneously folding a comparison of the pointer with itself.
1021 
1022  unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1023 
1025  for (const Use &U : Alloca->uses()) {
1026  if (Worklist.size() >= MaxIter)
1027  return nullptr;
1028  Worklist.push_back(&U);
1029  }
1030 
1031  unsigned NumCmps = 0;
1032  while (!Worklist.empty()) {
1033  assert(Worklist.size() <= MaxIter);
1034  const Use *U = Worklist.pop_back_val();
1035  const Value *V = U->getUser();
1036  --MaxIter;
1037 
1038  if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1039  isa<SelectInst>(V)) {
1040  // Track the uses.
1041  } else if (isa<LoadInst>(V)) {
1042  // Loading from the pointer doesn't escape it.
1043  continue;
1044  } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1045  // Storing *to* the pointer is fine, but storing the pointer escapes it.
1046  if (SI->getValueOperand() == U->get())
1047  return nullptr;
1048  continue;
1049  } else if (isa<ICmpInst>(V)) {
1050  if (NumCmps++)
1051  return nullptr; // Found more than one cmp.
1052  continue;
1053  } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1054  switch (Intrin->getIntrinsicID()) {
1055  // These intrinsics don't escape or compare the pointer. Memset is safe
1056  // because we don't allow ptrtoint. Memcpy and memmove are safe because
1057  // we don't allow stores, so src cannot point to V.
1058  case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1059  case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1060  continue;
1061  default:
1062  return nullptr;
1063  }
1064  } else {
1065  return nullptr;
1066  }
1067  for (const Use &U : V->uses()) {
1068  if (Worklist.size() >= MaxIter)
1069  return nullptr;
1070  Worklist.push_back(&U);
1071  }
1072  }
1073 
1074  Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1075  return replaceInstUsesWith(
1076  ICI,
1078 }
1079 
1080 /// Fold "icmp pred (X+C), X".
1082  ICmpInst::Predicate Pred) {
1083  // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1084  // so the values can never be equal. Similarly for all other "or equals"
1085  // operators.
1086  assert(!!C && "C should not be zero!");
1087 
1088  // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1089  // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1090  // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1091  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1092  Constant *R = ConstantInt::get(X->getType(),
1093  APInt::getMaxValue(C.getBitWidth()) - C);
1094  return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1095  }
1096 
1097  // (X+1) >u X --> X <u (0-1) --> X != 255
1098  // (X+2) >u X --> X <u (0-2) --> X <u 254
1099  // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1100  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1101  return new ICmpInst(ICmpInst::ICMP_ULT, X,
1102  ConstantInt::get(X->getType(), -C));
1103 
1104  APInt SMax = APInt::getSignedMaxValue(C.getBitWidth());
1105 
1106  // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1107  // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1108  // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1109  // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1110  // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1111  // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1112  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1113  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1114  ConstantInt::get(X->getType(), SMax - C));
1115 
1116  // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1117  // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1118  // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1119  // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1120  // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1121  // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1122 
1123  assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1124  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1125  ConstantInt::get(X->getType(), SMax - (C - 1)));
1126 }
1127 
1128 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1129 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1130 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1132  const APInt &AP1,
1133  const APInt &AP2) {
1134  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1135 
1136  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1137  if (I.getPredicate() == I.ICMP_NE)
1138  Pred = CmpInst::getInversePredicate(Pred);
1139  return new ICmpInst(Pred, LHS, RHS);
1140  };
1141 
1142  // Don't bother doing any work for cases which InstSimplify handles.
1143  if (AP2.isZero())
1144  return nullptr;
1145 
1146  bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1147  if (IsAShr) {
1148  if (AP2.isAllOnes())
1149  return nullptr;
1150  if (AP2.isNegative() != AP1.isNegative())
1151  return nullptr;
1152  if (AP2.sgt(AP1))
1153  return nullptr;
1154  }
1155 
1156  if (!AP1)
1157  // 'A' must be large enough to shift out the highest set bit.
1158  return getICmp(I.ICMP_UGT, A,
1159  ConstantInt::get(A->getType(), AP2.logBase2()));
1160 
1161  if (AP1 == AP2)
1162  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1163 
1164  int Shift;
1165  if (IsAShr && AP1.isNegative())
1166  Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1167  else
1168  Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1169 
1170  if (Shift > 0) {
1171  if (IsAShr && AP1 == AP2.ashr(Shift)) {
1172  // There are multiple solutions if we are comparing against -1 and the LHS
1173  // of the ashr is not a power of two.
1174  if (AP1.isAllOnes() && !AP2.isPowerOf2())
1175  return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1176  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1177  } else if (AP1 == AP2.lshr(Shift)) {
1178  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1179  }
1180  }
1181 
1182  // Shifting const2 will never be equal to const1.
1183  // FIXME: This should always be handled by InstSimplify?
1184  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1185  return replaceInstUsesWith(I, TorF);
1186 }
1187 
1188 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1189 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1191  const APInt &AP1,
1192  const APInt &AP2) {
1193  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1194 
1195  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1196  if (I.getPredicate() == I.ICMP_NE)
1197  Pred = CmpInst::getInversePredicate(Pred);
1198  return new ICmpInst(Pred, LHS, RHS);
1199  };
1200 
1201  // Don't bother doing any work for cases which InstSimplify handles.
1202  if (AP2.isZero())
1203  return nullptr;
1204 
1205  unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1206 
1207  if (!AP1 && AP2TrailingZeros != 0)
1208  return getICmp(
1209  I.ICMP_UGE, A,
1210  ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1211 
1212  if (AP1 == AP2)
1213  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1214 
1215  // Get the distance between the lowest bits that are set.
1216  int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1217 
1218  if (Shift > 0 && AP2.shl(Shift) == AP1)
1219  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1220 
1221  // Shifting const2 will never be equal to const1.
1222  // FIXME: This should always be handled by InstSimplify?
1223  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1224  return replaceInstUsesWith(I, TorF);
1225 }
1226 
1227 /// The caller has matched a pattern of the form:
1228 /// I = icmp ugt (add (add A, B), CI2), CI1
1229 /// If this is of the form:
1230 /// sum = a + b
1231 /// if (sum+128 >u 255)
1232 /// Then replace it with llvm.sadd.with.overflow.i8.
1233 ///
1235  ConstantInt *CI2, ConstantInt *CI1,
1236  InstCombinerImpl &IC) {
1237  // The transformation we're trying to do here is to transform this into an
1238  // llvm.sadd.with.overflow. To do this, we have to replace the original add
1239  // with a narrower add, and discard the add-with-constant that is part of the
1240  // range check (if we can't eliminate it, this isn't profitable).
1241 
1242  // In order to eliminate the add-with-constant, the compare can be its only
1243  // use.
1244  Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1245  if (!AddWithCst->hasOneUse())
1246  return nullptr;
1247 
1248  // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1249  if (!CI2->getValue().isPowerOf2())
1250  return nullptr;
1251  unsigned NewWidth = CI2->getValue().countTrailingZeros();
1252  if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1253  return nullptr;
1254 
1255  // The width of the new add formed is 1 more than the bias.
1256  ++NewWidth;
1257 
1258  // Check to see that CI1 is an all-ones value with NewWidth bits.
1259  if (CI1->getBitWidth() == NewWidth ||
1260  CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1261  return nullptr;
1262 
1263  // This is only really a signed overflow check if the inputs have been
1264  // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1265  // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1266  if (IC.ComputeMaxSignificantBits(A, 0, &I) > NewWidth ||
1267  IC.ComputeMaxSignificantBits(B, 0, &I) > NewWidth)
1268  return nullptr;
1269 
1270  // In order to replace the original add with a narrower
1271  // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1272  // and truncates that discard the high bits of the add. Verify that this is
1273  // the case.
1274  Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1275  for (User *U : OrigAdd->users()) {
1276  if (U == AddWithCst)
1277  continue;
1278 
1279  // Only accept truncates for now. We would really like a nice recursive
1280  // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1281  // chain to see which bits of a value are actually demanded. If the
1282  // original add had another add which was then immediately truncated, we
1283  // could still do the transformation.
1284  TruncInst *TI = dyn_cast<TruncInst>(U);
1285  if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1286  return nullptr;
1287  }
1288 
1289  // If the pattern matches, truncate the inputs to the narrower type and
1290  // use the sadd_with_overflow intrinsic to efficiently compute both the
1291  // result and the overflow bit.
1292  Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1294  I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1295 
1297 
1298  // Put the new code above the original add, in case there are any uses of the
1299  // add between the add and the compare.
1300  Builder.SetInsertPoint(OrigAdd);
1301 
1302  Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1303  Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1304  CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1305  Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1306  Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1307 
1308  // The inner add was the result of the narrow add, zero extended to the
1309  // wider type. Replace it with the result computed by the intrinsic.
1310  IC.replaceInstUsesWith(*OrigAdd, ZExt);
1311  IC.eraseInstFromFunction(*OrigAdd);
1312 
1313  // The original icmp gets replaced with the overflow value.
1314  return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1315 }
1316 
1317 /// If we have:
1318 /// icmp eq/ne (urem/srem %x, %y), 0
1319 /// iff %y is a power-of-two, we can replace this with a bit test:
1320 /// icmp eq/ne (and %x, (add %y, -1)), 0
1322  // This fold is only valid for equality predicates.
1323  if (!I.isEquality())
1324  return nullptr;
1325  ICmpInst::Predicate Pred;
1326  Value *X, *Y, *Zero;
1327  if (!match(&I, m_ICmp(Pred, m_OneUse(m_IRem(m_Value(X), m_Value(Y))),
1328  m_CombineAnd(m_Zero(), m_Value(Zero)))))
1329  return nullptr;
1330  if (!isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, 0, &I))
1331  return nullptr;
1332  // This may increase instruction count, we don't enforce that Y is a constant.
1333  Value *Mask = Builder.CreateAdd(Y, Constant::getAllOnesValue(Y->getType()));
1334  Value *Masked = Builder.CreateAnd(X, Mask);
1335  return ICmpInst::Create(Instruction::ICmp, Pred, Masked, Zero);
1336 }
1337 
1338 /// Fold equality-comparison between zero and any (maybe truncated) right-shift
1339 /// by one-less-than-bitwidth into a sign test on the original value.
1341  Instruction *Val;
1342  ICmpInst::Predicate Pred;
1343  if (!I.isEquality() || !match(&I, m_ICmp(Pred, m_Instruction(Val), m_Zero())))
1344  return nullptr;
1345 
1346  Value *X;
1347  Type *XTy;
1348 
1349  Constant *C;
1350  if (match(Val, m_TruncOrSelf(m_Shr(m_Value(X), m_Constant(C))))) {
1351  XTy = X->getType();
1352  unsigned XBitWidth = XTy->getScalarSizeInBits();
1353  if (!match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1354  APInt(XBitWidth, XBitWidth - 1))))
1355  return nullptr;
1356  } else if (isa<BinaryOperator>(Val) &&
1357  (X = reassociateShiftAmtsOfTwoSameDirectionShifts(
1358  cast<BinaryOperator>(Val), SQ.getWithInstruction(Val),
1359  /*AnalyzeForSignBitExtraction=*/true))) {
1360  XTy = X->getType();
1361  } else
1362  return nullptr;
1363 
1364  return ICmpInst::Create(Instruction::ICmp,
1368 }
1369 
1370 // Handle icmp pred X, 0
1372  CmpInst::Predicate Pred = Cmp.getPredicate();
1373  if (!match(Cmp.getOperand(1), m_Zero()))
1374  return nullptr;
1375 
1376  // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1377  if (Pred == ICmpInst::ICMP_SGT) {
1378  Value *A, *B;
1379  SelectPatternResult SPR = matchSelectPattern(Cmp.getOperand(0), A, B);
1380  if (SPR.Flavor == SPF_SMIN) {
1381  if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1382  return new ICmpInst(Pred, B, Cmp.getOperand(1));
1383  if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1384  return new ICmpInst(Pred, A, Cmp.getOperand(1));
1385  }
1386  }
1387 
1388  if (Instruction *New = foldIRemByPowerOfTwoToBitTest(Cmp))
1389  return New;
1390 
1391  // Given:
1392  // icmp eq/ne (urem %x, %y), 0
1393  // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
1394  // icmp eq/ne %x, 0
1395  Value *X, *Y;
1396  if (match(Cmp.getOperand(0), m_URem(m_Value(X), m_Value(Y))) &&
1397  ICmpInst::isEquality(Pred)) {
1398  KnownBits XKnown = computeKnownBits(X, 0, &Cmp);
1399  KnownBits YKnown = computeKnownBits(Y, 0, &Cmp);
1400  if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
1401  return new ICmpInst(Pred, X, Cmp.getOperand(1));
1402  }
1403 
1404  return nullptr;
1405 }
1406 
1407 /// Fold icmp Pred X, C.
1408 /// TODO: This code structure does not make sense. The saturating add fold
1409 /// should be moved to some other helper and extended as noted below (it is also
1410 /// possible that code has been made unnecessary - do we canonicalize IR to
1411 /// overflow/saturating intrinsics or not?).
1413  // Match the following pattern, which is a common idiom when writing
1414  // overflow-safe integer arithmetic functions. The source performs an addition
1415  // in wider type and explicitly checks for overflow using comparisons against
1416  // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1417  //
1418  // TODO: This could probably be generalized to handle other overflow-safe
1419  // operations if we worked out the formulas to compute the appropriate magic
1420  // constants.
1421  //
1422  // sum = a + b
1423  // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1424  CmpInst::Predicate Pred = Cmp.getPredicate();
1425  Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1426  Value *A, *B;
1427  ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1428  if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1429  match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1430  if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1431  return Res;
1432 
1433  // icmp(phi(C1, C2, ...), C) -> phi(icmp(C1, C), icmp(C2, C), ...).
1434  Constant *C = dyn_cast<Constant>(Op1);
1435  if (!C || C->canTrap())
1436  return nullptr;
1437 
1438  if (auto *Phi = dyn_cast<PHINode>(Op0))
1439  if (all_of(Phi->operands(), [](Value *V) { return isa<Constant>(V); })) {
1440  Type *Ty = Cmp.getType();
1441  Builder.SetInsertPoint(Phi);
1442  PHINode *NewPhi =
1443  Builder.CreatePHI(Ty, Phi->getNumOperands());
1444  for (BasicBlock *Predecessor : predecessors(Phi->getParent())) {
1445  auto *Input =
1446  cast<Constant>(Phi->getIncomingValueForBlock(Predecessor));
1447  auto *BoolInput = ConstantExpr::getCompare(Pred, Input, C);
1448  NewPhi->addIncoming(BoolInput, Predecessor);
1449  }
1450  NewPhi->takeName(&Cmp);
1451  return replaceInstUsesWith(Cmp, NewPhi);
1452  }
1453 
1454  return nullptr;
1455 }
1456 
1457 /// Canonicalize icmp instructions based on dominating conditions.
1459  // This is a cheap/incomplete check for dominance - just match a single
1460  // predecessor with a conditional branch.
1461  BasicBlock *CmpBB = Cmp.getParent();
1462  BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1463  if (!DomBB)
1464  return nullptr;
1465 
1466  Value *DomCond;
1467  BasicBlock *TrueBB, *FalseBB;
1468  if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1469  return nullptr;
1470 
1471  assert((TrueBB == CmpBB || FalseBB == CmpBB) &&
1472  "Predecessor block does not point to successor?");
1473 
1474  // The branch should get simplified. Don't bother simplifying this condition.
1475  if (TrueBB == FalseBB)
1476  return nullptr;
1477 
1478  // Try to simplify this compare to T/F based on the dominating condition.
1479  Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1480  if (Imp)
1481  return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1482 
1483  CmpInst::Predicate Pred = Cmp.getPredicate();
1484  Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1485  ICmpInst::Predicate DomPred;
1486  const APInt *C, *DomC;
1487  if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1488  match(Y, m_APInt(C))) {
1489  // We have 2 compares of a variable with constants. Calculate the constant
1490  // ranges of those compares to see if we can transform the 2nd compare:
1491  // DomBB:
1492  // DomCond = icmp DomPred X, DomC
1493  // br DomCond, CmpBB, FalseBB
1494  // CmpBB:
1495  // Cmp = icmp Pred X, C
1497  ConstantRange DominatingCR =
1498  (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1500  CmpInst::getInversePredicate(DomPred), *DomC);
1501  ConstantRange Intersection = DominatingCR.intersectWith(CR);
1502  ConstantRange Difference = DominatingCR.difference(CR);
1503  if (Intersection.isEmptySet())
1504  return replaceInstUsesWith(Cmp, Builder.getFalse());
1505  if (Difference.isEmptySet())
1506  return replaceInstUsesWith(Cmp, Builder.getTrue());
1507 
1508  // Canonicalizing a sign bit comparison that gets used in a branch,
1509  // pessimizes codegen by generating branch on zero instruction instead
1510  // of a test and branch. So we avoid canonicalizing in such situations
1511  // because test and branch instruction has better branch displacement
1512  // than compare and branch instruction.
1513  bool UnusedBit;
1514  bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1515  if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1516  return nullptr;
1517 
1518  // Avoid an infinite loop with min/max canonicalization.
1519  // TODO: This will be unnecessary if we canonicalize to min/max intrinsics.
1520  if (Cmp.hasOneUse() &&
1521  match(Cmp.user_back(), m_MaxOrMin(m_Value(), m_Value())))
1522  return nullptr;
1523 
1524  if (const APInt *EqC = Intersection.getSingleElement())
1525  return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1526  if (const APInt *NeC = Difference.getSingleElement())
1527  return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1528  }
1529 
1530  return nullptr;
1531 }
1532 
1533 /// Fold icmp (trunc X, Y), C.
1535  TruncInst *Trunc,
1536  const APInt &C) {
1537  ICmpInst::Predicate Pred = Cmp.getPredicate();
1538  Value *X = Trunc->getOperand(0);
1539  if (C.isOne() && C.getBitWidth() > 1) {
1540  // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1541  Value *V = nullptr;
1542  if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1543  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1544  ConstantInt::get(V->getType(), 1));
1545  }
1546 
1547  unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1548  SrcBits = X->getType()->getScalarSizeInBits();
1549  if (Cmp.isEquality() && Trunc->hasOneUse()) {
1550  // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1551  // of the high bits truncated out of x are known.
1552  KnownBits Known = computeKnownBits(X, 0, &Cmp);
1553 
1554  // If all the high bits are known, we can do this xform.
1555  if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1556  // Pull in the high bits from known-ones set.
1557  APInt NewRHS = C.zext(SrcBits);
1558  NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1559  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1560  }
1561  }
1562 
1563  // Look through truncated right-shift of the sign-bit for a sign-bit check:
1564  // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] < 0 --> ShOp < 0
1565  // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] > -1 --> ShOp > -1
1566  Value *ShOp;
1567  const APInt *ShAmtC;
1568  bool TrueIfSigned;
1569  if (isSignBitCheck(Pred, C, TrueIfSigned) &&
1570  match(X, m_Shr(m_Value(ShOp), m_APInt(ShAmtC))) &&
1571  DstBits == SrcBits - ShAmtC->getZExtValue()) {
1572  return TrueIfSigned
1573  ? new ICmpInst(ICmpInst::ICMP_SLT, ShOp,
1574  ConstantInt::getNullValue(X->getType()))
1575  : new ICmpInst(ICmpInst::ICMP_SGT, ShOp,
1576  ConstantInt::getAllOnesValue(X->getType()));
1577  }
1578 
1579  return nullptr;
1580 }
1581 
1582 /// Fold icmp (xor X, Y), C.
1584  BinaryOperator *Xor,
1585  const APInt &C) {
1586  Value *X = Xor->getOperand(0);
1587  Value *Y = Xor->getOperand(1);
1588  const APInt *XorC;
1589  if (!match(Y, m_APInt(XorC)))
1590  return nullptr;
1591 
1592  // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1593  // fold the xor.
1594  ICmpInst::Predicate Pred = Cmp.getPredicate();
1595  bool TrueIfSigned = false;
1596  if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1597 
1598  // If the sign bit of the XorCst is not set, there is no change to
1599  // the operation, just stop using the Xor.
1600  if (!XorC->isNegative())
1601  return replaceOperand(Cmp, 0, X);
1602 
1603  // Emit the opposite comparison.
1604  if (TrueIfSigned)
1605  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1606  ConstantInt::getAllOnesValue(X->getType()));
1607  else
1608  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1609  ConstantInt::getNullValue(X->getType()));
1610  }
1611 
1612  if (Xor->hasOneUse()) {
1613  // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1614  if (!Cmp.isEquality() && XorC->isSignMask()) {
1615  Pred = Cmp.getFlippedSignednessPredicate();
1616  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1617  }
1618 
1619  // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1620  if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1621  Pred = Cmp.getFlippedSignednessPredicate();
1622  Pred = Cmp.getSwappedPredicate(Pred);
1623  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1624  }
1625  }
1626 
1627  // Mask constant magic can eliminate an 'xor' with unsigned compares.
1628  if (Pred == ICmpInst::ICMP_UGT) {
1629  // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1630  if (*XorC == ~C && (C + 1).isPowerOf2())
1631  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1632  // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1633  if (*XorC == C && (C + 1).isPowerOf2())
1634  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1635  }
1636  if (Pred == ICmpInst::ICMP_ULT) {
1637  // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1638  if (*XorC == -C && C.isPowerOf2())
1639  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1640  ConstantInt::get(X->getType(), ~C));
1641  // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1642  if (*XorC == C && (-C).isPowerOf2())
1643  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1644  ConstantInt::get(X->getType(), ~C));
1645  }
1646  return nullptr;
1647 }
1648 
1649 /// Fold icmp (and (sh X, Y), C2), C1.
1651  BinaryOperator *And,
1652  const APInt &C1,
1653  const APInt &C2) {
1654  BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1655  if (!Shift || !Shift->isShift())
1656  return nullptr;
1657 
1658  // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1659  // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1660  // code produced by the clang front-end, for bitfield access.
1661  // This seemingly simple opportunity to fold away a shift turns out to be
1662  // rather complicated. See PR17827 for details.
1663  unsigned ShiftOpcode = Shift->getOpcode();
1664  bool IsShl = ShiftOpcode == Instruction::Shl;
1665  const APInt *C3;
1666  if (match(Shift->getOperand(1), m_APInt(C3))) {
1667  APInt NewAndCst, NewCmpCst;
1668  bool AnyCmpCstBitsShiftedOut;
1669  if (ShiftOpcode == Instruction::Shl) {
1670  // For a left shift, we can fold if the comparison is not signed. We can
1671  // also fold a signed comparison if the mask value and comparison value
1672  // are not negative. These constraints may not be obvious, but we can
1673  // prove that they are correct using an SMT solver.
1674  if (Cmp.isSigned() && (C2.isNegative() || C1.isNegative()))
1675  return nullptr;
1676 
1677  NewCmpCst = C1.lshr(*C3);
1678  NewAndCst = C2.lshr(*C3);
1679  AnyCmpCstBitsShiftedOut = NewCmpCst.shl(*C3) != C1;
1680  } else if (ShiftOpcode == Instruction::LShr) {
1681  // For a logical right shift, we can fold if the comparison is not signed.
1682  // We can also fold a signed comparison if the shifted mask value and the
1683  // shifted comparison value are not negative. These constraints may not be
1684  // obvious, but we can prove that they are correct using an SMT solver.
1685  NewCmpCst = C1.shl(*C3);
1686  NewAndCst = C2.shl(*C3);
1687  AnyCmpCstBitsShiftedOut = NewCmpCst.lshr(*C3) != C1;
1688  if (Cmp.isSigned() && (NewAndCst.isNegative() || NewCmpCst.isNegative()))
1689  return nullptr;
1690  } else {
1691  // For an arithmetic shift, check that both constants don't use (in a
1692  // signed sense) the top bits being shifted out.
1693  assert(ShiftOpcode == Instruction::AShr && "Unknown shift opcode");
1694  NewCmpCst = C1.shl(*C3);
1695  NewAndCst = C2.shl(*C3);
1696  AnyCmpCstBitsShiftedOut = NewCmpCst.ashr(*C3) != C1;
1697  if (NewAndCst.ashr(*C3) != C2)
1698  return nullptr;
1699  }
1700 
1701  if (AnyCmpCstBitsShiftedOut) {
1702  // If we shifted bits out, the fold is not going to work out. As a
1703  // special case, check to see if this means that the result is always
1704  // true or false now.
1705  if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1706  return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1707  if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1708  return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1709  } else {
1710  Value *NewAnd = Builder.CreateAnd(
1711  Shift->getOperand(0), ConstantInt::get(And->getType(), NewAndCst));
1712  return new ICmpInst(Cmp.getPredicate(),
1713  NewAnd, ConstantInt::get(And->getType(), NewCmpCst));
1714  }
1715  }
1716 
1717  // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1718  // preferable because it allows the C2 << Y expression to be hoisted out of a
1719  // loop if Y is invariant and X is not.
1720  if (Shift->hasOneUse() && C1.isZero() && Cmp.isEquality() &&
1721  !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1722  // Compute C2 << Y.
1723  Value *NewShift =
1724  IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1725  : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1726 
1727  // Compute X & (C2 << Y).
1728  Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1729  return replaceOperand(Cmp, 0, NewAnd);
1730  }
1731 
1732  return nullptr;
1733 }
1734 
1735 /// Fold icmp (and X, C2), C1.
1737  BinaryOperator *And,
1738  const APInt &C1) {
1739  bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1740 
1741  // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1742  // TODO: We canonicalize to the longer form for scalars because we have
1743  // better analysis/folds for icmp, and codegen may be better with icmp.
1744  if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isZero() &&
1745  match(And->getOperand(1), m_One()))
1746  return new TruncInst(And->getOperand(0), Cmp.getType());
1747 
1748  const APInt *C2;
1749  Value *X;
1750  if (!match(And, m_And(m_Value(X), m_APInt(C2))))
1751  return nullptr;
1752 
1753  // Don't perform the following transforms if the AND has multiple uses
1754  if (!And->hasOneUse())
1755  return nullptr;
1756 
1757  if (Cmp.isEquality() && C1.isZero()) {
1758  // Restrict this fold to single-use 'and' (PR10267).
1759  // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1760  if (C2->isSignMask()) {
1761  Constant *Zero = Constant::getNullValue(X->getType());
1762  auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1763  return new ICmpInst(NewPred, X, Zero);
1764  }
1765 
1766  // Restrict this fold only for single-use 'and' (PR10267).
1767  // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1768  if ((~(*C2) + 1).isPowerOf2()) {
1769  Constant *NegBOC =
1770  ConstantExpr::getNeg(cast<Constant>(And->getOperand(1)));
1771  auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1772  return new ICmpInst(NewPred, X, NegBOC);
1773  }
1774  }
1775 
1776  // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1777  // the input width without changing the value produced, eliminate the cast:
1778  //
1779  // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1780  //
1781  // We can do this transformation if the constants do not have their sign bits
1782  // set or if it is an equality comparison. Extending a relational comparison
1783  // when we're checking the sign bit would not work.
1784  Value *W;
1785  if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1786  (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1787  // TODO: Is this a good transform for vectors? Wider types may reduce
1788  // throughput. Should this transform be limited (even for scalars) by using
1789  // shouldChangeType()?
1790  if (!Cmp.getType()->isVectorTy()) {
1791  Type *WideType = W->getType();
1792  unsigned WideScalarBits = WideType->getScalarSizeInBits();
1793  Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1794  Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1795  Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1796  return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1797  }
1798  }
1799 
1800  if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1801  return I;
1802 
1803  // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1804  // (icmp pred (and A, (or (shl 1, B), 1), 0))
1805  //
1806  // iff pred isn't signed
1807  if (!Cmp.isSigned() && C1.isZero() && And->getOperand(0)->hasOneUse() &&
1808  match(And->getOperand(1), m_One())) {
1809  Constant *One = cast<Constant>(And->getOperand(1));
1810  Value *Or = And->getOperand(0);
1811  Value *A, *B, *LShr;
1812  if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1813  match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1814  unsigned UsesRemoved = 0;
1815  if (And->hasOneUse())
1816  ++UsesRemoved;
1817  if (Or->hasOneUse())
1818  ++UsesRemoved;
1819  if (LShr->hasOneUse())
1820  ++UsesRemoved;
1821 
1822  // Compute A & ((1 << B) | 1)
1823  Value *NewOr = nullptr;
1824  if (auto *C = dyn_cast<Constant>(B)) {
1825  if (UsesRemoved >= 1)
1826  NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1827  } else {
1828  if (UsesRemoved >= 3)
1829  NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1830  /*HasNUW=*/true),
1831  One, Or->getName());
1832  }
1833  if (NewOr) {
1834  Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1835  return replaceOperand(Cmp, 0, NewAnd);
1836  }
1837  }
1838  }
1839 
1840  return nullptr;
1841 }
1842 
1843 /// Fold icmp (and X, Y), C.
1845  BinaryOperator *And,
1846  const APInt &C) {
1847  if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1848  return I;
1849 
1850  const ICmpInst::Predicate Pred = Cmp.getPredicate();
1851  bool TrueIfNeg;
1852  if (isSignBitCheck(Pred, C, TrueIfNeg)) {
1853  // ((X - 1) & ~X) < 0 --> X == 0
1854  // ((X - 1) & ~X) >= 0 --> X != 0
1855  Value *X;
1856  if (match(And->getOperand(0), m_Add(m_Value(X), m_AllOnes())) &&
1857  match(And->getOperand(1), m_Not(m_Specific(X)))) {
1858  auto NewPred = TrueIfNeg ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE;
1859  return new ICmpInst(NewPred, X, ConstantInt::getNullValue(X->getType()));
1860  }
1861  }
1862 
1863  // TODO: These all require that Y is constant too, so refactor with the above.
1864 
1865  // Try to optimize things like "A[i] & 42 == 0" to index computations.
1866  Value *X = And->getOperand(0);
1867  Value *Y = And->getOperand(1);
1868  if (auto *LI = dyn_cast<LoadInst>(X))
1869  if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1870  if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1871  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1872  !LI->isVolatile() && isa<ConstantInt>(Y)) {
1873  ConstantInt *C2 = cast<ConstantInt>(Y);
1874  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1875  return Res;
1876  }
1877 
1878  if (!Cmp.isEquality())
1879  return nullptr;
1880 
1881  // X & -C == -C -> X > u ~C
1882  // X & -C != -C -> X <= u ~C
1883  // iff C is a power of 2
1884  if (Cmp.getOperand(1) == Y && C.isNegatedPowerOf2()) {
1885  auto NewPred =
1887  return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1888  }
1889 
1890  return nullptr;
1891 }
1892 
1893 /// Fold icmp (or X, Y), C.
1895  BinaryOperator *Or,
1896  const APInt &C) {
1897  ICmpInst::Predicate Pred = Cmp.getPredicate();
1898  if (C.isOne()) {
1899  // icmp slt signum(V) 1 --> icmp slt V, 1
1900  Value *V = nullptr;
1901  if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1902  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1903  ConstantInt::get(V->getType(), 1));
1904  }
1905 
1906  Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
1907  const APInt *MaskC;
1908  if (match(OrOp1, m_APInt(MaskC)) && Cmp.isEquality()) {
1909  if (*MaskC == C && (C + 1).isPowerOf2()) {
1910  // X | C == C --> X <=u C
1911  // X | C != C --> X >u C
1912  // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1914  return new ICmpInst(Pred, OrOp0, OrOp1);
1915  }
1916 
1917  // More general: canonicalize 'equality with set bits mask' to
1918  // 'equality with clear bits mask'.
1919  // (X | MaskC) == C --> (X & ~MaskC) == C ^ MaskC
1920  // (X | MaskC) != C --> (X & ~MaskC) != C ^ MaskC
1921  if (Or->hasOneUse()) {
1922  Value *And = Builder.CreateAnd(OrOp0, ~(*MaskC));
1923  Constant *NewC = ConstantInt::get(Or->getType(), C ^ (*MaskC));
1924  return new ICmpInst(Pred, And, NewC);
1925  }
1926  }
1927 
1928  // (X | (X-1)) s< 0 --> X s< 1
1929  // (X | (X-1)) s> -1 --> X s> 0
1930  Value *X;
1931  bool TrueIfSigned;
1932  if (isSignBitCheck(Pred, C, TrueIfSigned) &&
1934  auto NewPred = TrueIfSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGT;
1935  Constant *NewC = ConstantInt::get(X->getType(), TrueIfSigned ? 1 : 0);
1936  return new ICmpInst(NewPred, X, NewC);
1937  }
1938 
1939  if (!Cmp.isEquality() || !C.isZero() || !Or->hasOneUse())
1940  return nullptr;
1941 
1942  Value *P, *Q;
1943  if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1944  // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1945  // -> and (icmp eq P, null), (icmp eq Q, null).
1946  Value *CmpP =
1947  Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1948  Value *CmpQ =
1949  Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1950  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1951  return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1952  }
1953 
1954  // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1955  // a shorter form that has more potential to be folded even further.
1956  Value *X1, *X2, *X3, *X4;
1957  if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1958  match(OrOp1, m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1959  // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1960  // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1961  Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1962  Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1963  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1964  return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1965  }
1966 
1967  return nullptr;
1968 }
1969 
1970 /// Fold icmp (mul X, Y), C.
1972  BinaryOperator *Mul,
1973  const APInt &C) {
1974  const APInt *MulC;
1975  if (!match(Mul->getOperand(1), m_APInt(MulC)))
1976  return nullptr;
1977 
1978  // If this is a test of the sign bit and the multiply is sign-preserving with
1979  // a constant operand, use the multiply LHS operand instead.
1980  ICmpInst::Predicate Pred = Cmp.getPredicate();
1981  if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1982  if (MulC->isNegative())
1983  Pred = ICmpInst::getSwappedPredicate(Pred);
1984  return new ICmpInst(Pred, Mul->getOperand(0),
1986  }
1987 
1988  // If the multiply does not wrap, try to divide the compare constant by the
1989  // multiplication factor.
1990  if (Cmp.isEquality() && !MulC->isZero()) {
1991  // (mul nsw X, MulC) == C --> X == C /s MulC
1992  if (Mul->hasNoSignedWrap() && C.srem(*MulC).isZero()) {
1993  Constant *NewC = ConstantInt::get(Mul->getType(), C.sdiv(*MulC));
1994  return new ICmpInst(Pred, Mul->getOperand(0), NewC);
1995  }
1996  // (mul nuw X, MulC) == C --> X == C /u MulC
1997  if (Mul->hasNoUnsignedWrap() && C.urem(*MulC).isZero()) {
1998  Constant *NewC = ConstantInt::get(Mul->getType(), C.udiv(*MulC));
1999  return new ICmpInst(Pred, Mul->getOperand(0), NewC);
2000  }
2001  }
2002 
2003  return nullptr;
2004 }
2005 
2006 /// Fold icmp (shl 1, Y), C.
2008  const APInt &C) {
2009  Value *Y;
2010  if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
2011  return nullptr;
2012 
2013  Type *ShiftType = Shl->getType();
2014  unsigned TypeBits = C.getBitWidth();
2015  bool CIsPowerOf2 = C.isPowerOf2();
2016  ICmpInst::Predicate Pred = Cmp.getPredicate();
2017  if (Cmp.isUnsigned()) {
2018  // (1 << Y) pred C -> Y pred Log2(C)
2019  if (!CIsPowerOf2) {
2020  // (1 << Y) < 30 -> Y <= 4
2021  // (1 << Y) <= 30 -> Y <= 4
2022  // (1 << Y) >= 30 -> Y > 4
2023  // (1 << Y) > 30 -> Y > 4
2024  if (Pred == ICmpInst::ICMP_ULT)
2025  Pred = ICmpInst::ICMP_ULE;
2026  else if (Pred == ICmpInst::ICMP_UGE)
2027  Pred = ICmpInst::ICMP_UGT;
2028  }
2029 
2030  // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
2031  // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
2032  unsigned CLog2 = C.logBase2();
2033  if (CLog2 == TypeBits - 1) {
2034  if (Pred == ICmpInst::ICMP_UGE)
2035  Pred = ICmpInst::ICMP_EQ;
2036  else if (Pred == ICmpInst::ICMP_ULT)
2037  Pred = ICmpInst::ICMP_NE;
2038  }
2039  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
2040  } else if (Cmp.isSigned()) {
2041  Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
2042  if (C.isAllOnes()) {
2043  // (1 << Y) <= -1 -> Y == 31
2044  if (Pred == ICmpInst::ICMP_SLE)
2045  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2046 
2047  // (1 << Y) > -1 -> Y != 31
2048  if (Pred == ICmpInst::ICMP_SGT)
2049  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2050  } else if (!C) {
2051  // (1 << Y) < 0 -> Y == 31
2052  // (1 << Y) <= 0 -> Y == 31
2053  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
2054  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2055 
2056  // (1 << Y) >= 0 -> Y != 31
2057  // (1 << Y) > 0 -> Y != 31
2058  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
2059  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2060  }
2061  } else if (Cmp.isEquality() && CIsPowerOf2) {
2062  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
2063  }
2064 
2065  return nullptr;
2066 }
2067 
2068 /// Fold icmp (shl X, Y), C.
2070  BinaryOperator *Shl,
2071  const APInt &C) {
2072  const APInt *ShiftVal;
2073  if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
2074  return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
2075 
2076  const APInt *ShiftAmt;
2077  if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
2078  return foldICmpShlOne(Cmp, Shl, C);
2079 
2080  // Check that the shift amount is in range. If not, don't perform undefined
2081  // shifts. When the shift is visited, it will be simplified.
2082  unsigned TypeBits = C.getBitWidth();
2083  if (ShiftAmt->uge(TypeBits))
2084  return nullptr;
2085 
2086  ICmpInst::Predicate Pred = Cmp.getPredicate();
2087  Value *X = Shl->getOperand(0);
2088  Type *ShType = Shl->getType();
2089 
2090  // NSW guarantees that we are only shifting out sign bits from the high bits,
2091  // so we can ASHR the compare constant without needing a mask and eliminate
2092  // the shift.
2093  if (Shl->hasNoSignedWrap()) {
2094  if (Pred == ICmpInst::ICMP_SGT) {
2095  // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
2096  APInt ShiftedC = C.ashr(*ShiftAmt);
2097  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2098  }
2099  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2100  C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
2101  APInt ShiftedC = C.ashr(*ShiftAmt);
2102  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2103  }
2104  if (Pred == ICmpInst::ICMP_SLT) {
2105  // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
2106  // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2107  // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2108  // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2109  assert(!C.isMinSignedValue() && "Unexpected icmp slt");
2110  APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
2111  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2112  }
2113  // If this is a signed comparison to 0 and the shift is sign preserving,
2114  // use the shift LHS operand instead; isSignTest may change 'Pred', so only
2115  // do that if we're sure to not continue on in this function.
2116  if (isSignTest(Pred, C))
2117  return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
2118  }
2119 
2120  // NUW guarantees that we are only shifting out zero bits from the high bits,
2121  // so we can LSHR the compare constant without needing a mask and eliminate
2122  // the shift.
2123  if (Shl->hasNoUnsignedWrap()) {
2124  if (Pred == ICmpInst::ICMP_UGT) {
2125  // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2126  APInt ShiftedC = C.lshr(*ShiftAmt);
2127  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2128  }
2129  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2130  C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
2131  APInt ShiftedC = C.lshr(*ShiftAmt);
2132  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2133  }
2134  if (Pred == ICmpInst::ICMP_ULT) {
2135  // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2136  // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2137  // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2138  // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2139  assert(C.ugt(0) && "ult 0 should have been eliminated");
2140  APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
2141  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2142  }
2143  }
2144 
2145  if (Cmp.isEquality() && Shl->hasOneUse()) {
2146  // Strength-reduce the shift into an 'and'.
2148  ShType,
2149  APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2150  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2151  Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
2152  return new ICmpInst(Pred, And, LShrC);
2153  }
2154 
2155  // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2156  bool TrueIfSigned = false;
2157  if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
2158  // (X << 31) <s 0 --> (X & 1) != 0
2160  ShType,
2161  APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2162  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2163  return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2164  And, Constant::getNullValue(ShType));
2165  }
2166 
2167  // Simplify 'shl' inequality test into 'and' equality test.
2168  if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2169  // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2170  if ((C + 1).isPowerOf2() &&
2171  (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2172  Value *And = Builder.CreateAnd(X, (~C).lshr(ShiftAmt->getZExtValue()));
2173  return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2175  And, Constant::getNullValue(ShType));
2176  }
2177  // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2178  if (C.isPowerOf2() &&
2179  (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2180  Value *And =
2181  Builder.CreateAnd(X, (~(C - 1)).lshr(ShiftAmt->getZExtValue()));
2182  return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2184  And, Constant::getNullValue(ShType));
2185  }
2186  }
2187 
2188  // Transform (icmp pred iM (shl iM %v, N), C)
2189  // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2190  // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2191  // This enables us to get rid of the shift in favor of a trunc that may be
2192  // free on the target. It has the additional benefit of comparing to a
2193  // smaller constant that may be more target-friendly.
2194  unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2195  if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2196  DL.isLegalInteger(TypeBits - Amt)) {
2197  Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2198  if (auto *ShVTy = dyn_cast<VectorType>(ShType))
2199  TruncTy = VectorType::get(TruncTy, ShVTy->getElementCount());
2200  Constant *NewC =
2201  ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2202  return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2203  }
2204 
2205  return nullptr;
2206 }
2207 
2208 /// Fold icmp ({al}shr X, Y), C.
2210  BinaryOperator *Shr,
2211  const APInt &C) {
2212  // An exact shr only shifts out zero bits, so:
2213  // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2214  Value *X = Shr->getOperand(0);
2215  CmpInst::Predicate Pred = Cmp.getPredicate();
2216  if (Cmp.isEquality() && Shr->isExact() && C.isZero())
2217  return new ICmpInst(Pred, X, Cmp.getOperand(1));
2218 
2219  const APInt *ShiftVal;
2220  if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2221  return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2222 
2223  const APInt *ShiftAmt;
2224  if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2225  return nullptr;
2226 
2227  // Check that the shift amount is in range. If not, don't perform undefined
2228  // shifts. When the shift is visited it will be simplified.
2229  unsigned TypeBits = C.getBitWidth();
2230  unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2231  if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2232  return nullptr;
2233 
2234  bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2235  bool IsExact = Shr->isExact();
2236  Type *ShrTy = Shr->getType();
2237  // TODO: If we could guarantee that InstSimplify would handle all of the
2238  // constant-value-based preconditions in the folds below, then we could assert
2239  // those conditions rather than checking them. This is difficult because of
2240  // undef/poison (PR34838).
2241  if (IsAShr) {
2242  if (Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_ULT || IsExact) {
2243  // When ShAmtC can be shifted losslessly:
2244  // icmp PRED (ashr exact X, ShAmtC), C --> icmp PRED X, (C << ShAmtC)
2245  // icmp slt/ult (ashr X, ShAmtC), C --> icmp slt/ult X, (C << ShAmtC)
2246  APInt ShiftedC = C.shl(ShAmtVal);
2247  if (ShiftedC.ashr(ShAmtVal) == C)
2248  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2249  }
2250  if (Pred == CmpInst::ICMP_SGT) {
2251  // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2252  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2253  if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2254  (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2255  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2256  }
2257 
2258  // If the compare constant has significant bits above the lowest sign-bit,
2259  // then convert an unsigned cmp to a test of the sign-bit:
2260  // (ashr X, ShiftC) u> C --> X s< 0
2261  // (ashr X, ShiftC) u< C --> X s> -1
2262  if (C.getBitWidth() > 2 && C.getNumSignBits() <= ShAmtVal) {
2263  if (Pred == CmpInst::ICMP_UGT) {
2264  return new ICmpInst(CmpInst::ICMP_SLT, X,
2265  ConstantInt::getNullValue(ShrTy));
2266  }
2267  if (Pred == CmpInst::ICMP_ULT) {
2268  return new ICmpInst(CmpInst::ICMP_SGT, X,
2270  }
2271  }
2272  } else {
2273  if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2274  // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2275  // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2276  APInt ShiftedC = C.shl(ShAmtVal);
2277  if (ShiftedC.lshr(ShAmtVal) == C)
2278  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2279  }
2280  if (Pred == CmpInst::ICMP_UGT) {
2281  // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2282  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2283  if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2284  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2285  }
2286  }
2287 
2288  if (!Cmp.isEquality())
2289  return nullptr;
2290 
2291  // Handle equality comparisons of shift-by-constant.
2292 
2293  // If the comparison constant changes with the shift, the comparison cannot
2294  // succeed (bits of the comparison constant cannot match the shifted value).
2295  // This should be known by InstSimplify and already be folded to true/false.
2296  assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2297  (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2298  "Expected icmp+shr simplify did not occur.");
2299 
2300  // If the bits shifted out are known zero, compare the unshifted value:
2301  // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2302  if (Shr->isExact())
2303  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2304 
2305  if (C.isZero()) {
2306  // == 0 is u< 1.
2307  if (Pred == CmpInst::ICMP_EQ)
2308  return new ICmpInst(CmpInst::ICMP_ULT, X,
2309  ConstantInt::get(ShrTy, (C + 1).shl(ShAmtVal)));
2310  else
2311  return new ICmpInst(CmpInst::ICMP_UGT, X,
2312  ConstantInt::get(ShrTy, (C + 1).shl(ShAmtVal) - 1));
2313  }
2314 
2315  if (Shr->hasOneUse()) {
2316  // Canonicalize the shift into an 'and':
2317  // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2318  APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2319  Constant *Mask = ConstantInt::get(ShrTy, Val);
2320  Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2321  return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2322  }
2323 
2324  return nullptr;
2325 }
2326 
2328  BinaryOperator *SRem,
2329  const APInt &C) {
2330  // Match an 'is positive' or 'is negative' comparison of remainder by a
2331  // constant power-of-2 value:
2332  // (X % pow2C) sgt/slt 0
2333  const ICmpInst::Predicate Pred = Cmp.getPredicate();
2334  if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT)
2335  return nullptr;
2336 
2337  // TODO: The one-use check is standard because we do not typically want to
2338  // create longer instruction sequences, but this might be a special-case
2339  // because srem is not good for analysis or codegen.
2340  if (!SRem->hasOneUse())
2341  return nullptr;
2342 
2343  const APInt *DivisorC;
2344  if (!C.isZero() || !match(SRem->getOperand(1), m_Power2(DivisorC)))
2345  return nullptr;
2346 
2347  // Mask off the sign bit and the modulo bits (low-bits).
2348  Type *Ty = SRem->getType();
2349  APInt SignMask = APInt::getSignMask(Ty->getScalarSizeInBits());
2350  Constant *MaskC = ConstantInt::get(Ty, SignMask | (*DivisorC - 1));
2351  Value *And = Builder.CreateAnd(SRem->getOperand(0), MaskC);
2352 
2353  // For 'is positive?' check that the sign-bit is clear and at least 1 masked
2354  // bit is set. Example:
2355  // (i8 X % 32) s> 0 --> (X & 159) s> 0
2356  if (Pred == ICmpInst::ICMP_SGT)
2358 
2359  // For 'is negative?' check that the sign-bit is set and at least 1 masked
2360  // bit is set. Example:
2361  // (i16 X % 4) s< 0 --> (X & 32771) u> 32768
2362  return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, SignMask));
2363 }
2364 
2365 /// Fold icmp (udiv X, Y), C.
2367  BinaryOperator *UDiv,
2368  const APInt &C) {
2369  const APInt *C2;
2370  if (!match(UDiv->getOperand(0), m_APInt(C2)))
2371  return nullptr;
2372 
2373  assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2374 
2375  // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2376  Value *Y = UDiv->getOperand(1);
2377  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2378  assert(!C.isMaxValue() &&
2379  "icmp ugt X, UINT_MAX should have been simplified already.");
2380  return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2381  ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2382  }
2383 
2384  // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2385  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2386  assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2387  return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2388  ConstantInt::get(Y->getType(), C2->udiv(C)));
2389  }
2390 
2391  return nullptr;
2392 }
2393 
2394 /// Fold icmp ({su}div X, Y), C.
2396  BinaryOperator *Div,
2397  const APInt &C) {
2398  // Fold: icmp pred ([us]div X, C2), C -> range test
2399  // Fold this div into the comparison, producing a range check.
2400  // Determine, based on the divide type, what the range is being
2401  // checked. If there is an overflow on the low or high side, remember
2402  // it, otherwise compute the range [low, hi) bounding the new value.
2403  // See: InsertRangeTest above for the kinds of replacements possible.
2404  const APInt *C2;
2405  if (!match(Div->getOperand(1), m_APInt(C2)))
2406  return nullptr;
2407 
2408  // FIXME: If the operand types don't match the type of the divide
2409  // then don't attempt this transform. The code below doesn't have the
2410  // logic to deal with a signed divide and an unsigned compare (and
2411  // vice versa). This is because (x /s C2) <s C produces different
2412  // results than (x /s C2) <u C or (x /u C2) <s C or even
2413  // (x /u C2) <u C. Simply casting the operands and result won't
2414  // work. :( The if statement below tests that condition and bails
2415  // if it finds it.
2416  bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2417  if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2418  return nullptr;
2419 
2420  // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2421  // INT_MIN will also fail if the divisor is 1. Although folds of all these
2422  // division-by-constant cases should be present, we can not assert that they
2423  // have happened before we reach this icmp instruction.
2424  if (C2->isZero() || C2->isOne() || (DivIsSigned && C2->isAllOnes()))
2425  return nullptr;
2426 
2427  // Compute Prod = C * C2. We are essentially solving an equation of
2428  // form X / C2 = C. We solve for X by multiplying C2 and C.
2429  // By solving for X, we can turn this into a range check instead of computing
2430  // a divide.
2431  APInt Prod = C * *C2;
2432 
2433  // Determine if the product overflows by seeing if the product is not equal to
2434  // the divide. Make sure we do the same kind of divide as in the LHS
2435  // instruction that we're folding.
2436  bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2437 
2438  ICmpInst::Predicate Pred = Cmp.getPredicate();
2439 
2440  // If the division is known to be exact, then there is no remainder from the
2441  // divide, so the covered range size is unit, otherwise it is the divisor.
2442  APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2443 
2444  // Figure out the interval that is being checked. For example, a comparison
2445  // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2446  // Compute this interval based on the constants involved and the signedness of
2447  // the compare/divide. This computes a half-open interval, keeping track of
2448  // whether either value in the interval overflows. After analysis each
2449  // overflow variable is set to 0 if it's corresponding bound variable is valid
2450  // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2451  int LoOverflow = 0, HiOverflow = 0;
2452  APInt LoBound, HiBound;
2453 
2454  if (!DivIsSigned) { // udiv
2455  // e.g. X/5 op 3 --> [15, 20)
2456  LoBound = Prod;
2457  HiOverflow = LoOverflow = ProdOV;
2458  if (!HiOverflow) {
2459  // If this is not an exact divide, then many values in the range collapse
2460  // to the same result value.
2461  HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2462  }
2463  } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2464  if (C.isZero()) { // (X / pos) op 0
2465  // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2466  LoBound = -(RangeSize - 1);
2467  HiBound = RangeSize;
2468  } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2469  LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2470  HiOverflow = LoOverflow = ProdOV;
2471  if (!HiOverflow)
2472  HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2473  } else { // (X / pos) op neg
2474  // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2475  HiBound = Prod + 1;
2476  LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2477  if (!LoOverflow) {
2478  APInt DivNeg = -RangeSize;
2479  LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2480  }
2481  }
2482  } else if (C2->isNegative()) { // Divisor is < 0.
2483  if (Div->isExact())
2484  RangeSize.negate();
2485  if (C.isZero()) { // (X / neg) op 0
2486  // e.g. X/-5 op 0 --> [-4, 5)
2487  LoBound = RangeSize + 1;
2488  HiBound = -RangeSize;
2489  if (HiBound == *C2) { // -INTMIN = INTMIN
2490  HiOverflow = 1; // [INTMIN+1, overflow)
2491  HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2492  }
2493  } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2494  // e.g. X/-5 op 3 --> [-19, -14)
2495  HiBound = Prod + 1;
2496  HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2497  if (!LoOverflow)
2498  LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2499  } else { // (X / neg) op neg
2500  LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2501  LoOverflow = HiOverflow = ProdOV;
2502  if (!HiOverflow)
2503  HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2504  }
2505 
2506  // Dividing by a negative swaps the condition. LT <-> GT
2507  Pred = ICmpInst::getSwappedPredicate(Pred);
2508  }
2509 
2510  Value *X = Div->getOperand(0);
2511  switch (Pred) {
2512  default: llvm_unreachable("Unhandled icmp opcode!");
2513  case ICmpInst::ICMP_EQ:
2514  if (LoOverflow && HiOverflow)
2515  return replaceInstUsesWith(Cmp, Builder.getFalse());
2516  if (HiOverflow)
2517  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2519  ConstantInt::get(Div->getType(), LoBound));
2520  if (LoOverflow)
2521  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2523  ConstantInt::get(Div->getType(), HiBound));
2524  return replaceInstUsesWith(
2525  Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2526  case ICmpInst::ICMP_NE:
2527  if (LoOverflow && HiOverflow)
2528  return replaceInstUsesWith(Cmp, Builder.getTrue());
2529  if (HiOverflow)
2530  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2532  ConstantInt::get(Div->getType(), LoBound));
2533  if (LoOverflow)
2534  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2536  ConstantInt::get(Div->getType(), HiBound));
2537  return replaceInstUsesWith(Cmp,
2538  insertRangeTest(X, LoBound, HiBound,
2539  DivIsSigned, false));
2540  case ICmpInst::ICMP_ULT:
2541  case ICmpInst::ICMP_SLT:
2542  if (LoOverflow == +1) // Low bound is greater than input range.
2543  return replaceInstUsesWith(Cmp, Builder.getTrue());
2544  if (LoOverflow == -1) // Low bound is less than input range.
2545  return replaceInstUsesWith(Cmp, Builder.getFalse());
2546  return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2547  case ICmpInst::ICMP_UGT:
2548  case ICmpInst::ICMP_SGT:
2549  if (HiOverflow == +1) // High bound greater than input range.
2550  return replaceInstUsesWith(Cmp, Builder.getFalse());
2551  if (HiOverflow == -1) // High bound less than input range.
2552  return replaceInstUsesWith(Cmp, Builder.getTrue());
2553  if (Pred == ICmpInst::ICMP_UGT)
2554  return new ICmpInst(ICmpInst::ICMP_UGE, X,
2555  ConstantInt::get(Div->getType(), HiBound));
2556  return new ICmpInst(ICmpInst::ICMP_SGE, X,
2557  ConstantInt::get(Div->getType(), HiBound));
2558  }
2559 
2560  return nullptr;
2561 }
2562 
2563 /// Fold icmp (sub X, Y), C.
2565  BinaryOperator *Sub,
2566  const APInt &C) {
2567  Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2568  ICmpInst::Predicate Pred = Cmp.getPredicate();
2569  Type *Ty = Sub->getType();
2570 
2571  // (SubC - Y) == C) --> Y == (SubC - C)
2572  // (SubC - Y) != C) --> Y != (SubC - C)
2573  Constant *SubC;
2574  if (Cmp.isEquality() && match(X, m_ImmConstant(SubC))) {
2575  return new ICmpInst(Pred, Y,
2577  }
2578 
2579  // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2580  const APInt *C2;
2581  APInt SubResult;
2582  ICmpInst::Predicate SwappedPred = Cmp.getSwappedPredicate();
2583  bool HasNSW = Sub->hasNoSignedWrap();
2584  bool HasNUW = Sub->hasNoUnsignedWrap();
2585  if (match(X, m_APInt(C2)) &&
2586  ((Cmp.isUnsigned() && HasNUW) || (Cmp.isSigned() && HasNSW)) &&
2587  !subWithOverflow(SubResult, *C2, C, Cmp.isSigned()))
2588  return new ICmpInst(SwappedPred, Y, ConstantInt::get(Ty, SubResult));
2589 
2590  // The following transforms are only worth it if the only user of the subtract
2591  // is the icmp.
2592  // TODO: This is an artificial restriction for all of the transforms below
2593  // that only need a single replacement icmp.
2594  if (!Sub->hasOneUse())
2595  return nullptr;
2596 
2597  // X - Y == 0 --> X == Y.
2598  // X - Y != 0 --> X != Y.
2599  if (Cmp.isEquality() && C.isZero())
2600  return new ICmpInst(Pred, X, Y);
2601 
2602  if (Sub->hasNoSignedWrap()) {
2603  // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2604  if (Pred == ICmpInst::ICMP_SGT && C.isAllOnes())
2605  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2606 
2607  // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2608  if (Pred == ICmpInst::ICMP_SGT && C.isZero())
2609  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2610 
2611  // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2612  if (Pred == ICmpInst::ICMP_SLT && C.isZero())
2613  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2614 
2615  // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2616  if (Pred == ICmpInst::ICMP_SLT && C.isOne())
2617  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2618  }
2619 
2620  if (!match(X, m_APInt(C2)))
2621  return nullptr;
2622 
2623  // C2 - Y <u C -> (Y | (C - 1)) == C2
2624  // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2625  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2626  (*C2 & (C - 1)) == (C - 1))
2627  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2628 
2629  // C2 - Y >u C -> (Y | C) != C2
2630  // iff C2 & C == C and C + 1 is a power of 2
2631  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2632  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2633 
2634  // We have handled special cases that reduce.
2635  // Canonicalize any remaining sub to add as:
2636  // (C2 - Y) > C --> (Y + ~C2) < ~C
2637  Value *Add = Builder.CreateAdd(Y, ConstantInt::get(Ty, ~(*C2)), "notsub",
2638  HasNUW, HasNSW);
2639  return new ICmpInst(SwappedPred, Add, ConstantInt::get(Ty, ~C));
2640 }
2641 
2642 /// Fold icmp (add X, Y), C.
2644  BinaryOperator *Add,
2645  const APInt &C) {
2646  Value *Y = Add->getOperand(1);
2647  const APInt *C2;
2648  if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2649  return nullptr;
2650 
2651  // Fold icmp pred (add X, C2), C.
2652  Value *X = Add->getOperand(0);
2653  Type *Ty = Add->getType();
2654  const CmpInst::Predicate Pred = Cmp.getPredicate();
2655 
2656  // If the add does not wrap, we can always adjust the compare by subtracting
2657  // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2658  // are canonicalized to SGT/SLT/UGT/ULT.
2659  if ((Add->hasNoSignedWrap() &&
2660  (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2661  (Add->hasNoUnsignedWrap() &&
2662  (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2663  bool Overflow;
2664  APInt NewC =
2665  Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2666  // If there is overflow, the result must be true or false.
2667  // TODO: Can we assert there is no overflow because InstSimplify always
2668  // handles those cases?
2669  if (!Overflow)
2670  // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2671  return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2672  }
2673 
2674  auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2675  const APInt &Upper = CR.getUpper();
2676  const APInt &Lower = CR.getLower();
2677  if (Cmp.isSigned()) {
2678  if (Lower.isSignMask())
2679  return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2680  if (Upper.isSignMask())
2681  return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2682  } else {
2683  if (Lower.isMinValue())
2684  return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2685  if (Upper.isMinValue())
2686  return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2687  }
2688 
2689  // This set of folds is intentionally placed after folds that use no-wrapping
2690  // flags because those folds are likely better for later analysis/codegen.
2693 
2694  // Fold compare with offset to opposite sign compare if it eliminates offset:
2695  // (X + C2) >u C --> X <s -C2 (if C == C2 + SMAX)
2696  if (Pred == CmpInst::ICMP_UGT && C == *C2 + SMax)
2697  return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, -(*C2)));
2698 
2699  // (X + C2) <u C --> X >s ~C2 (if C == C2 + SMIN)
2700  if (Pred == CmpInst::ICMP_ULT && C == *C2 + SMin)
2701  return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantInt::get(Ty, ~(*C2)));
2702 
2703  // (X + C2) >s C --> X <u (SMAX - C) (if C == C2 - 1)
2704  if (Pred == CmpInst::ICMP_SGT && C == *C2 - 1)
2705  return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, SMax - C));
2706 
2707  // (X + C2) <s C --> X >u (C ^ SMAX) (if C == C2)
2708  if (Pred == CmpInst::ICMP_SLT && C == *C2)
2709  return new ICmpInst(ICmpInst::ICMP_UGT, X, ConstantInt::get(Ty, C ^ SMax));
2710 
2711  if (!Add->hasOneUse())
2712  return nullptr;
2713 
2714  // X+C <u C2 -> (X & -C2) == C
2715  // iff C & (C2-1) == 0
2716  // C2 is a power of 2
2717  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2718  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2719  ConstantExpr::getNeg(cast<Constant>(Y)));
2720 
2721  // X+C >u C2 -> (X & ~C2) != C
2722  // iff C & C2 == 0
2723  // C2+1 is a power of 2
2724  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2725  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2726  ConstantExpr::getNeg(cast<Constant>(Y)));
2727 
2728  // The range test idiom can use either ult or ugt. Arbitrarily canonicalize
2729  // to the ult form.
2730  // X+C2 >u C -> X+(C2-C-1) <u ~C
2731  if (Pred == ICmpInst::ICMP_UGT)
2732  return new ICmpInst(ICmpInst::ICMP_ULT,
2733  Builder.CreateAdd(X, ConstantInt::get(Ty, *C2 - C - 1)),
2734  ConstantInt::get(Ty, ~C));
2735 
2736  return nullptr;
2737 }
2738 
2740  Value *&RHS, ConstantInt *&Less,
2741  ConstantInt *&Equal,
2742  ConstantInt *&Greater) {
2743  // TODO: Generalize this to work with other comparison idioms or ensure
2744  // they get canonicalized into this form.
2745 
2746  // select i1 (a == b),
2747  // i32 Equal,
2748  // i32 (select i1 (a < b), i32 Less, i32 Greater)
2749  // where Equal, Less and Greater are placeholders for any three constants.
2750  ICmpInst::Predicate PredA;
2751  if (!match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) ||
2752  !ICmpInst::isEquality(PredA))
2753  return false;
2754  Value *EqualVal = SI->getTrueValue();
2755  Value *UnequalVal = SI->getFalseValue();
2756  // We still can get non-canonical predicate here, so canonicalize.
2757  if (PredA == ICmpInst::ICMP_NE)
2758  std::swap(EqualVal, UnequalVal);
2759  if (!match(EqualVal, m_ConstantInt(Equal)))
2760  return false;
2761  ICmpInst::Predicate PredB;
2762  Value *LHS2, *RHS2;
2763  if (!match(UnequalVal, m_Select(m_ICmp(PredB, m_Value(LHS2), m_Value(RHS2)),
2764  m_ConstantInt(Less), m_ConstantInt(Greater))))
2765  return false;
2766  // We can get predicate mismatch here, so canonicalize if possible:
2767  // First, ensure that 'LHS' match.
2768  if (LHS2 != LHS) {
2769  // x sgt y <--> y slt x
2770  std::swap(LHS2, RHS2);
2771  PredB = ICmpInst::getSwappedPredicate(PredB);
2772  }
2773  if (LHS2 != LHS)
2774  return false;
2775  // We also need to canonicalize 'RHS'.
2776  if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(RHS2)) {
2777  // x sgt C-1 <--> x sge C <--> not(x slt C)
2778  auto FlippedStrictness =
2780  PredB, cast<Constant>(RHS2));
2781  if (!FlippedStrictness)
2782  return false;
2783  assert(FlippedStrictness->first == ICmpInst::ICMP_SGE &&
2784  "basic correctness failure");
2785  RHS2 = FlippedStrictness->second;
2786  // And kind-of perform the result swap.
2787  std::swap(Less, Greater);
2788  PredB = ICmpInst::ICMP_SLT;
2789  }
2790  return PredB == ICmpInst::ICMP_SLT && RHS == RHS2;
2791 }
2792 
2794  SelectInst *Select,
2795  ConstantInt *C) {
2796 
2797  assert(C && "Cmp RHS should be a constant int!");
2798  // If we're testing a constant value against the result of a three way
2799  // comparison, the result can be expressed directly in terms of the
2800  // original values being compared. Note: We could possibly be more
2801  // aggressive here and remove the hasOneUse test. The original select is
2802  // really likely to simplify or sink when we remove a test of the result.
2803  Value *OrigLHS, *OrigRHS;
2804  ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2805  if (Cmp.hasOneUse() &&
2806  matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2807  C3GreaterThan)) {
2808  assert(C1LessThan && C2Equal && C3GreaterThan);
2809 
2810  bool TrueWhenLessThan =
2811  ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2812  ->isAllOnesValue();
2813  bool TrueWhenEqual =
2814  ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2815  ->isAllOnesValue();
2816  bool TrueWhenGreaterThan =
2817  ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2818  ->isAllOnesValue();
2819 
2820  // This generates the new instruction that will replace the original Cmp
2821  // Instruction. Instead of enumerating the various combinations when
2822  // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2823  // false, we rely on chaining of ORs and future passes of InstCombine to
2824  // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2825 
2826  // When none of the three constants satisfy the predicate for the RHS (C),
2827  // the entire original Cmp can be simplified to a false.
2828  Value *Cond = Builder.getFalse();
2829  if (TrueWhenLessThan)
2830  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
2831  OrigLHS, OrigRHS));
2832  if (TrueWhenEqual)
2833  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
2834  OrigLHS, OrigRHS));
2835  if (TrueWhenGreaterThan)
2836  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
2837  OrigLHS, OrigRHS));
2838 
2839  return replaceInstUsesWith(Cmp, Cond);
2840  }
2841  return nullptr;
2842 }
2843 
2845  auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
2846  if (!Bitcast)
2847  return nullptr;
2848 
2849  ICmpInst::Predicate Pred = Cmp.getPredicate();
2850  Value *Op1 = Cmp.getOperand(1);
2851  Value *BCSrcOp = Bitcast->getOperand(0);
2852 
2853  // Make sure the bitcast doesn't change the number of vector elements.
2854  if (Bitcast->getSrcTy()->getScalarSizeInBits() ==
2855  Bitcast->getDestTy()->getScalarSizeInBits()) {
2856  // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
2857  Value *X;
2858  if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
2859  // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
2860  // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
2861  // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
2862  // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
2863  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
2864  Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
2865  match(Op1, m_Zero()))
2866  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2867 
2868  // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
2869  if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
2870  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
2871 
2872  // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
2873  if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
2874  return new ICmpInst(Pred, X,
2875  ConstantInt::getAllOnesValue(X->getType()));
2876  }
2877 
2878  // Zero-equality checks are preserved through unsigned floating-point casts:
2879  // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
2880  // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
2881  if (match(BCSrcOp, m_UIToFP(m_Value(X))))
2882  if (Cmp.isEquality() && match(Op1, m_Zero()))
2883  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2884 
2885  // If this is a sign-bit test of a bitcast of a casted FP value, eliminate
2886  // the FP extend/truncate because that cast does not change the sign-bit.
2887  // This is true for all standard IEEE-754 types and the X86 80-bit type.
2888  // The sign-bit is always the most significant bit in those types.
2889  const APInt *C;
2890  bool TrueIfSigned;
2891  if (match(Op1, m_APInt(C)) && Bitcast->hasOneUse() &&
2892  InstCombiner::isSignBitCheck(Pred, *C, TrueIfSigned)) {
2893  if (match(BCSrcOp, m_FPExt(m_Value(X))) ||
2894  match(BCSrcOp, m_FPTrunc(m_Value(X)))) {
2895  // (bitcast (fpext/fptrunc X)) to iX) < 0 --> (bitcast X to iY) < 0
2896  // (bitcast (fpext/fptrunc X)) to iX) > -1 --> (bitcast X to iY) > -1
2897  Type *XType = X->getType();
2898 
2899  // We can't currently handle Power style floating point operations here.
2900  if (!(XType->isPPC_FP128Ty() || BCSrcOp->getType()->isPPC_FP128Ty())) {
2901 
2902  Type *NewType = Builder.getIntNTy(XType->getScalarSizeInBits());
2903  if (auto *XVTy = dyn_cast<VectorType>(XType))
2904  NewType = VectorType::get(NewType, XVTy->getElementCount());
2905  Value *NewBitcast = Builder.CreateBitCast(X, NewType);
2906  if (TrueIfSigned)
2907  return new ICmpInst(ICmpInst::ICMP_SLT, NewBitcast,
2908  ConstantInt::getNullValue(NewType));
2909  else
2910  return new ICmpInst(ICmpInst::ICMP_SGT, NewBitcast,
2911  ConstantInt::getAllOnesValue(NewType));
2912  }
2913  }
2914  }
2915  }
2916 
2917  // Test to see if the operands of the icmp are casted versions of other
2918  // values. If the ptr->ptr cast can be stripped off both arguments, do so.
2919  if (Bitcast->getType()->isPointerTy() &&
2920  (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2921  // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2922  // so eliminate it as well.
2923  if (auto *BC2 = dyn_cast<BitCastInst>(Op1))
2924  Op1 = BC2->getOperand(0);
2925 
2926  Op1 = Builder.CreateBitCast(Op1, BCSrcOp->getType());
2927  return new ICmpInst(Pred, BCSrcOp, Op1);
2928  }
2929 
2930  const APInt *C;
2931  if (!match(Cmp.getOperand(1), m_APInt(C)) ||
2932  !Bitcast->getType()->isIntegerTy() ||
2933  !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2934  return nullptr;
2935 
2936  // If this is checking if all elements of a vector compare are set or not,
2937  // invert the casted vector equality compare and test if all compare
2938  // elements are clear or not. Compare against zero is generally easier for
2939  // analysis and codegen.
2940  // icmp eq/ne (bitcast (not X) to iN), -1 --> icmp eq/ne (bitcast X to iN), 0
2941  // Example: are all elements equal? --> are zero elements not equal?
2942  // TODO: Try harder to reduce compare of 2 freely invertible operands?
2943  if (Cmp.isEquality() && C->isAllOnes() && Bitcast->hasOneUse() &&
2944  isFreeToInvert(BCSrcOp, BCSrcOp->hasOneUse())) {
2945  Type *ScalarTy = Bitcast->getType();
2946  Value *Cast = Builder.CreateBitCast(Builder.CreateNot(BCSrcOp), ScalarTy);
2947  return new ICmpInst(Pred, Cast, ConstantInt::getNullValue(ScalarTy));
2948  }
2949 
2950  // If this is checking if all elements of an extended vector are clear or not,
2951  // compare in a narrow type to eliminate the extend:
2952  // icmp eq/ne (bitcast (ext X) to iN), 0 --> icmp eq/ne (bitcast X to iM), 0
2953  Value *X;
2954  if (Cmp.isEquality() && C->isZero() && Bitcast->hasOneUse() &&
2955  match(BCSrcOp, m_ZExtOrSExt(m_Value(X)))) {
2956  if (auto *VecTy = dyn_cast<FixedVectorType>(X->getType())) {
2957  Type *NewType = Builder.getIntNTy(VecTy->getPrimitiveSizeInBits());
2958  Value *NewCast = Builder.CreateBitCast(X, NewType);
2959  return new ICmpInst(Pred, NewCast, ConstantInt::getNullValue(NewType));
2960  }
2961  }
2962 
2963  // Folding: icmp <pred> iN X, C
2964  // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2965  // and C is a splat of a K-bit pattern
2966  // and SC is a constant vector = <C', C', C', ..., C'>
2967  // Into:
2968  // %E = extractelement <M x iK> %vec, i32 C'
2969  // icmp <pred> iK %E, trunc(C)
2970  Value *Vec;
2972  if (match(BCSrcOp, m_Shuffle(m_Value(Vec), m_Undef(), m_Mask(Mask)))) {
2973  // Check whether every element of Mask is the same constant
2974  if (is_splat(Mask)) {
2975  auto *VecTy = cast<VectorType>(BCSrcOp->getType());
2976  auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2977  if (C->isSplat(EltTy->getBitWidth())) {
2978  // Fold the icmp based on the value of C
2979  // If C is M copies of an iK sized bit pattern,
2980  // then:
2981  // => %E = extractelement <N x iK> %vec, i32 Elem
2982  // icmp <pred> iK %SplatVal, <pattern>
2983  Value *Elem = Builder.getInt32(Mask[0]);
2984  Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2985  Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
2986  return new ICmpInst(Pred, Extract, NewC);
2987  }
2988  }
2989  }
2990  return nullptr;
2991 }
2992 
2993 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2994 /// where X is some kind of instruction.
2996  const APInt *C;
2997  if (!match(Cmp.getOperand(1), m_APInt(C)))
2998  return nullptr;
2999 
3000  if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
3001  switch (BO->getOpcode()) {
3002  case Instruction::Xor:
3003  if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
3004  return I;
3005  break;
3006  case Instruction::And:
3007  if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
3008  return I;
3009  break;
3010  case Instruction::Or:
3011  if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
3012  return I;
3013  break;
3014  case Instruction::Mul:
3015  if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
3016  return I;
3017  break;
3018  case Instruction::Shl:
3019  if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
3020  return I;
3021  break;
3022  case Instruction::LShr:
3023  case Instruction::AShr:
3024  if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
3025  return I;
3026  break;
3027  case Instruction::SRem:
3028  if (Instruction *I = foldICmpSRemConstant(Cmp, BO, *C))
3029  return I;
3030  break;
3031  case Instruction::UDiv:
3032  if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
3033  return I;
3035  case Instruction::SDiv:
3036  if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
3037  return I;
3038  break;
3039  case Instruction::Sub:
3040  if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
3041  return I;
3042  break;
3043  case Instruction::Add:
3044  if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
3045  return I;
3046  break;
3047  default:
3048  break;
3049  }
3050  // TODO: These folds could be refactored to be part of the above calls.
3051  if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
3052  return I;
3053  }
3054 
3055  // Match against CmpInst LHS being instructions other than binary operators.
3056 
3057  if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
3058  // For now, we only support constant integers while folding the
3059  // ICMP(SELECT)) pattern. We can extend this to support vector of integers
3060  // similar to the cases handled by binary ops above.
3061  if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
3062  if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
3063  return I;
3064  }
3065 
3066  if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
3067  if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
3068  return I;
3069  }
3070 
3071  if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
3072  if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
3073  return I;
3074 
3075  return nullptr;
3076 }
3077 
3078 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
3079 /// icmp eq/ne BO, C.
3081  ICmpInst &Cmp, BinaryOperator *BO, const APInt &C) {
3082  // TODO: Some of these folds could work with arbitrary constants, but this
3083  // function is limited to scalar and vector splat constants.
3084  if (!Cmp.isEquality())
3085  return nullptr;
3086 
3087  ICmpInst::Predicate Pred = Cmp.getPredicate();
3088  bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
3089  Constant *RHS = cast<Constant>(Cmp.getOperand(1));
3090  Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
3091 
3092  switch (BO->getOpcode()) {
3093  case Instruction::SRem:
3094  // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
3095  if (C.isZero() && BO->hasOneUse()) {
3096  const APInt *BOC;
3097  if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
3098  Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
3099  return new ICmpInst(Pred, NewRem,
3101  }
3102  }
3103  break;
3104  case Instruction::Add: {
3105  // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
3106  if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3107  if (BO->hasOneUse())
3108  return new ICmpInst(Pred, BOp0, ConstantExpr::getSub(RHS, BOC));
3109  } else if (C.isZero()) {
3110  // Replace ((add A, B) != 0) with (A != -B) if A or B is
3111  // efficiently invertible, or if the add has just this one use.
3112  if (Value *NegVal = dyn_castNegVal(BOp1))
3113  return new ICmpInst(Pred, BOp0, NegVal);
3114  if (Value *NegVal = dyn_castNegVal(BOp0))
3115  return new ICmpInst(Pred, NegVal, BOp1);
3116  if (BO->hasOneUse()) {
3117  Value *Neg = Builder.CreateNeg(BOp1);
3118  Neg->takeName(BO);
3119  return new ICmpInst(Pred, BOp0, Neg);
3120  }
3121  }
3122  break;
3123  }
3124  case Instruction::Xor:
3125  if (BO->hasOneUse()) {
3126  if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3127  // For the xor case, we can xor two constants together, eliminating
3128  // the explicit xor.
3129  return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
3130  } else if (C.isZero()) {
3131  // Replace ((xor A, B) != 0) with (A != B)
3132  return new ICmpInst(Pred, BOp0, BOp1);
3133  }
3134  }
3135  break;
3136  case Instruction::Or: {
3137  const APInt *BOC;
3138  if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
3139  // Comparing if all bits outside of a constant mask are set?
3140  // Replace (X | C) == -1 with (X & ~C) == ~C.
3141  // This removes the -1 constant.
3142  Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
3143  Value *And = Builder.CreateAnd(BOp0, NotBOC);
3144  return new ICmpInst(Pred, And, NotBOC);
3145  }
3146  break;
3147  }
3148  case Instruction::And: {
3149  const APInt *BOC;
3150  if (match(BOp1, m_APInt(BOC))) {
3151  // If we have ((X & C) == C), turn it into ((X & C) != 0).
3152  if (C == *BOC && C.isPowerOf2())
3153  return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
3155  }
3156  break;
3157  }
3158  case Instruction::UDiv:
3159  if (C.isZero()) {
3160  // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
3161  auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3162  return new ICmpInst(NewPred, BOp1, BOp0);
3163  }
3164  break;
3165  default:
3166  break;
3167  }
3168  return nullptr;
3169 }
3170 
3171 /// Fold an equality icmp with LLVM intrinsic and constant operand.
3173  ICmpInst &Cmp, IntrinsicInst *II, const APInt &C) {
3174  Type *Ty = II->getType();
3175  unsigned BitWidth = C.getBitWidth();
3176  const ICmpInst::Predicate Pred = Cmp.getPredicate();
3177 
3178  switch (II->getIntrinsicID()) {
3179  case Intrinsic::abs:
3180  // abs(A) == 0 -> A == 0
3181  // abs(A) == INT_MIN -> A == INT_MIN
3182  if (C.isZero() || C.isMinSignedValue())
3183  return new ICmpInst(Pred, II->getArgOperand(0), ConstantInt::get(Ty, C));
3184  break;
3185 
3186  case Intrinsic::bswap:
3187  // bswap(A) == C -> A == bswap(C)
3188  return new ICmpInst(Pred, II->getArgOperand(0),
3189  ConstantInt::get(Ty, C.byteSwap()));
3190 
3191  case Intrinsic::ctlz:
3192  case Intrinsic::cttz: {
3193  // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
3194  if (C == BitWidth)
3195  return new ICmpInst(Pred, II->getArgOperand(0),
3197 
3198  // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
3199  // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
3200  // Limit to one use to ensure we don't increase instruction count.
3201  unsigned Num = C.getLimitedValue(BitWidth);
3202  if (Num != BitWidth && II->hasOneUse()) {
3203  bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
3204  APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
3205  : APInt::getHighBitsSet(BitWidth, Num + 1);
3206  APInt Mask2 = IsTrailing
3208  : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3209  return new ICmpInst(Pred, Builder.CreateAnd(II->getArgOperand(0), Mask1),
3210  ConstantInt::get(Ty, Mask2));
3211  }
3212  break;
3213  }
3214 
3215  case Intrinsic::ctpop: {
3216  // popcount(A) == 0 -> A == 0 and likewise for !=
3217  // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
3218  bool IsZero = C.isZero();
3219  if (IsZero || C == BitWidth)
3220  return new ICmpInst(Pred, II->getArgOperand(0),
3221  IsZero ? Constant::getNullValue(Ty)
3223 
3224  break;
3225  }
3226 
3227  case Intrinsic::fshl:
3228  case Intrinsic::fshr:
3229  if (II->getArgOperand(0) == II->getArgOperand(1)) {
3230  // (rot X, ?) == 0/-1 --> X == 0/-1
3231  // TODO: This transform is safe to re-use undef elts in a vector, but
3232  // the constant value passed in by the caller doesn't allow that.
3233  if (C.isZero() || C.isAllOnes())
3234  return new ICmpInst(Pred, II->getArgOperand(0), Cmp.getOperand(1));
3235 
3236  const APInt *RotAmtC;
3237  // ror(X, RotAmtC) == C --> X == rol(C, RotAmtC)
3238  // rol(X, RotAmtC) == C --> X == ror(C, RotAmtC)
3239  if (match(II->getArgOperand(2), m_APInt(RotAmtC)))
3240  return new ICmpInst(Pred, II->getArgOperand(0),
3241  II->getIntrinsicID() == Intrinsic::fshl
3242  ? ConstantInt::get(Ty, C.rotr(*RotAmtC))
3243  : ConstantInt::get(Ty, C.rotl(*RotAmtC)));
3244  }
3245  break;
3246 
3247  case Intrinsic::uadd_sat: {
3248  // uadd.sat(a, b) == 0 -> (a | b) == 0
3249  if (C.isZero()) {
3250  Value *Or = Builder.CreateOr(II->getArgOperand(0), II->getArgOperand(1));
3251  return new ICmpInst(Pred, Or, Constant::getNullValue(Ty));
3252  }
3253  break;
3254  }
3255 
3256  case Intrinsic::usub_sat: {
3257  // usub.sat(a, b) == 0 -> a <= b
3258  if (C.isZero()) {
3259  ICmpInst::Predicate NewPred =
3261  return new ICmpInst(NewPred, II->getArgOperand(0), II->getArgOperand(1));
3262  }
3263  break;
3264  }
3265  default:
3266  break;
3267  }
3268 
3269  return nullptr;
3270 }
3271 
3272 /// Fold an icmp with LLVM intrinsics
3274  assert(Cmp.isEquality());
3275 
3276  ICmpInst::Predicate Pred = Cmp.getPredicate();
3277  Value *Op0 = Cmp.getOperand(0);
3278  Value *Op1 = Cmp.getOperand(1);
3279  const auto *IIOp0 = dyn_cast<IntrinsicInst>(Op0);
3280  const auto *IIOp1 = dyn_cast<IntrinsicInst>(Op1);
3281  if (!IIOp0 || !IIOp1 || IIOp0->getIntrinsicID() != IIOp1->getIntrinsicID())
3282  return nullptr;
3283 
3284  switch (IIOp0->getIntrinsicID()) {
3285  case Intrinsic::bswap:
3286  case Intrinsic::bitreverse:
3287  // If both operands are byte-swapped or bit-reversed, just compare the
3288  // original values.
3289  return new ICmpInst(Pred, IIOp0->getOperand(0), IIOp1->getOperand(0));
3290  case Intrinsic::fshl:
3291  case Intrinsic::fshr:
3292  // If both operands are rotated by same amount, just compare the
3293  // original values.
3294  if (IIOp0->getOperand(0) != IIOp0->getOperand(1))
3295  break;
3296  if (IIOp1->getOperand(0) != IIOp1->getOperand(1))
3297  break;
3298  if (IIOp0->getOperand(2) != IIOp1->getOperand(2))
3299  break;
3300  return new ICmpInst(Pred, IIOp0->getOperand(0), IIOp1->getOperand(0));
3301  default:
3302  break;
3303  }
3304 
3305  return nullptr;
3306 }
3307 
3308 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
3310  IntrinsicInst *II,
3311  const APInt &C) {
3312  if (Cmp.isEquality())
3313  return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
3314 
3315  Type *Ty = II->getType();
3316  unsigned BitWidth = C.getBitWidth();
3317  ICmpInst::Predicate Pred = Cmp.getPredicate();
3318  switch (II->getIntrinsicID()) {
3319  case Intrinsic::ctpop: {
3320  // (ctpop X > BitWidth - 1) --> X == -1
3321  Value *X = II->getArgOperand(0);
3322  if (C == BitWidth - 1 && Pred == ICmpInst::ICMP_UGT)
3323  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ, X,
3325  // (ctpop X < BitWidth) --> X != -1
3326  if (C == BitWidth && Pred == ICmpInst::ICMP_ULT)
3327  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE, X,
3329  break;
3330  }
3331  case Intrinsic::ctlz: {
3332  // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
3333  if (Pred == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3334  unsigned Num = C.getLimitedValue();
3335  APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3336  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
3337  II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3338  }
3339 
3340  // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
3341  if (Pred == ICmpInst::ICMP_ULT && C.uge(1) && C.ule(BitWidth)) {
3342  unsigned Num = C.getLimitedValue();
3343  APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
3344  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
3345  II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3346  }
3347  break;
3348  }
3349  case Intrinsic::cttz: {
3350  // Limit to one use to ensure we don't increase instruction count.
3351  if (!II->hasOneUse())
3352  return nullptr;
3353 
3354  // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
3355  if (Pred == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3356  APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
3357  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
3358  Builder.CreateAnd(II->getArgOperand(0), Mask),
3360  }
3361 
3362  // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
3363  if (Pred == ICmpInst::ICMP_ULT && C.uge(1) && C.ule(BitWidth)) {
3364  APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue());
3365  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
3366  Builder.CreateAnd(II->getArgOperand(0), Mask),
3368  }
3369  break;
3370  }
3371  default:
3372  break;
3373  }
3374 
3375  return nullptr;
3376 }
3377 
3378 /// Handle icmp with constant (but not simple integer constant) RHS.
3380  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3381  Constant *RHSC = dyn_cast<Constant>(Op1);
3382  Instruction *LHSI = dyn_cast<Instruction>(Op0);
3383  if (!RHSC || !LHSI)
3384  return nullptr;
3385 
3386  switch (LHSI->getOpcode()) {
3387  case Instruction::GetElementPtr:
3388  // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3389  if (RHSC->isNullValue() &&
3390  cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3391  return new ICmpInst(
3392  I.getPredicate(), LHSI->getOperand(0),
3394  break;
3395  case Instruction::PHI:
3396  // Only fold icmp into the PHI if the phi and icmp are in the same
3397  // block. If in the same block, we're encouraging jump threading. If
3398  // not, we are just pessimizing the code by making an i1 phi.
3399  if (LHSI->getParent() == I.getParent())
3400  if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
3401  return NV;
3402  break;
3403  case Instruction::Select: {
3404  // If either operand of the select is a constant, we can fold the
3405  // comparison into the select arms, which will cause one to be
3406  // constant folded and the select turned into a bitwise or.
3407  Value *Op1 = nullptr, *Op2 = nullptr;
3408  ConstantInt *CI = nullptr;
3409 
3410  auto SimplifyOp = [&](Value *V) {
3411  Value *Op = nullptr;
3412  if (Constant *C = dyn_cast<Constant>(V)) {
3413  Op = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3414  } else if (RHSC->isNullValue()) {
3415  // If null is being compared, check if it can be further simplified.
3416  Op = SimplifyICmpInst(I.getPredicate(), V, RHSC, SQ);
3417  }
3418  return Op;
3419  };
3420  Op1 = SimplifyOp(LHSI->getOperand(1));
3421  if (Op1)
3422  CI = dyn_cast<ConstantInt>(Op1);
3423 
3424  Op2 = SimplifyOp(LHSI->getOperand(2));
3425  if (Op2)
3426  CI = dyn_cast<ConstantInt>(Op2);
3427 
3428  // We only want to perform this transformation if it will not lead to
3429  // additional code. This is true if either both sides of the select
3430  // fold to a constant (in which case the icmp is replaced with a select
3431  // which will usually simplify) or this is the only user of the
3432  // select (in which case we are trading a select+icmp for a simpler
3433  // select+icmp) or all uses of the select can be replaced based on
3434  // dominance information ("Global cases").
3435  bool Transform = false;
3436  if (Op1 && Op2)
3437  Transform = true;
3438  else if (Op1 || Op2) {
3439  // Local case
3440  if (LHSI->hasOneUse())
3441  Transform = true;
3442  // Global cases
3443  else if (CI && !CI->isZero())
3444  // When Op1 is constant try replacing select with second operand.
3445  // Otherwise Op2 is constant and try replacing select with first
3446  // operand.
3447  Transform =
3448  replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
3449  }
3450  if (Transform) {
3451  if (!Op1)
3452  Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
3453  I.getName());
3454  if (!Op2)
3455  Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
3456  I.getName());
3457  return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3458  }
3459  break;
3460  }
3461  case Instruction::IntToPtr:
3462  // icmp pred inttoptr(X), null -> icmp pred X, 0
3463  if (RHSC->isNullValue() &&
3464  DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3465  return new ICmpInst(
3466  I.getPredicate(), LHSI->getOperand(0),
3468  break;
3469 
3470  case Instruction::Load:
3471  // Try to optimize things like "A[i] > 4" to index computations.
3472  if (GetElementPtrInst *GEP =
3473  dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3474  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3475  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3476  !cast<LoadInst>(LHSI)->isVolatile())
3477  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
3478  return Res;
3479  }
3480  break;
3481  }
3482 
3483  return nullptr;
3484 }
3485 
3486 /// Some comparisons can be simplified.
3487 /// In this case, we are looking for comparisons that look like
3488 /// a check for a lossy truncation.
3489 /// Folds:
3490 /// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
3491 /// Where Mask is some pattern that produces all-ones in low bits:
3492 /// (-1 >> y)
3493 /// ((-1 << y) >> y) <- non-canonical, has extra uses
3494 /// ~(-1 << y)
3495 /// ((1 << y) + (-1)) <- non-canonical, has extra uses
3496 /// The Mask can be a constant, too.
3497 /// For some predicates, the operands are commutative.
3498 /// For others, x can only be on a specific side.
3501  ICmpInst::Predicate SrcPred;
3502  Value *X, *M, *Y;
3503  auto m_VariableMask = m_CombineOr(
3505  m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
3508  auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
3509  if (!match(&I, m_c_ICmp(SrcPred,
3511  m_Deferred(X))))
3512  return nullptr;
3513 
3514  ICmpInst::Predicate DstPred;
3515  switch (SrcPred) {
3516  case ICmpInst::Predicate::ICMP_EQ:
3517  // x & (-1 >> y) == x -> x u<= (-1 >> y)
3518  DstPred = ICmpInst::Predicate::ICMP_ULE;
3519  break;
3520  case ICmpInst::Predicate::ICMP_NE:
3521  // x & (-1 >> y) != x -> x u> (-1 >> y)
3522  DstPred = ICmpInst::Predicate::ICMP_UGT;
3523  break;
3524  case ICmpInst::Predicate::ICMP_ULT:
3525  // x & (-1 >> y) u< x -> x u> (-1 >> y)
3526  // x u> x & (-1 >> y) -> x u> (-1 >> y)
3527  DstPred = ICmpInst::Predicate::ICMP_UGT;
3528  break;
3529  case ICmpInst::Predicate::ICMP_UGE:
3530  // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
3531  // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
3532  DstPred = ICmpInst::Predicate::ICMP_ULE;
3533  break;
3534  case ICmpInst::Predicate::ICMP_SLT:
3535  // x & (-1 >> y) s< x -> x s> (-1 >> y)
3536  // x s> x & (-1 >> y) -> x s> (-1 >> y)
3537  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3538  return nullptr;
3539  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3540  return nullptr;
3541  DstPred = ICmpInst::Predicate::ICMP_SGT;
3542  break;
3543  case ICmpInst::Predicate::ICMP_SGE:
3544  // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
3545  // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3546  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3547  return nullptr;
3548  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3549  return nullptr;
3550  DstPred = ICmpInst::Predicate::ICMP_SLE;
3551  break;
3552  case ICmpInst::Predicate::ICMP_SGT:
3553  case ICmpInst::Predicate::ICMP_SLE:
3554  return nullptr;
3555  case ICmpInst::Predicate::ICMP_UGT:
3556  case ICmpInst::Predicate::ICMP_ULE:
3557  llvm_unreachable("Instsimplify took care of commut. variant");
3558  break;
3559  default:
3560  llvm_unreachable("All possible folds are handled.");
3561  }
3562 
3563  // The mask value may be a vector constant that has undefined elements. But it
3564  // may not be safe to propagate those undefs into the new compare, so replace
3565  // those elements by copying an existing, defined, and safe scalar constant.
3566  Type *OpTy = M->getType();
3567  auto *VecC = dyn_cast<Constant>(M);
3568  auto *OpVTy = dyn_cast<FixedVectorType>(OpTy);
3569  if (OpVTy && VecC && VecC->containsUndefOrPoisonElement()) {
3570  Constant *SafeReplacementConstant = nullptr;
3571  for (unsigned i = 0, e = OpVTy->getNumElements(); i != e; ++i) {
3572  if (!isa<UndefValue>(VecC->getAggregateElement(i))) {
3573  SafeReplacementConstant = VecC->getAggregateElement(i);
3574  break;
3575  }
3576  }
3577  assert(SafeReplacementConstant && "Failed to find undef replacement");
3578  M = Constant::replaceUndefsWith(VecC, SafeReplacementConstant);
3579  }
3580 
3581  return Builder.CreateICmp(DstPred, X, M);
3582 }
3583 
3584 /// Some comparisons can be simplified.
3585 /// In this case, we are looking for comparisons that look like
3586 /// a check for a lossy signed truncation.
3587 /// Folds: (MaskedBits is a constant.)
3588 /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3589 /// Into:
3590 /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3591 /// Where KeptBits = bitwidth(%x) - MaskedBits
3592 static Value *
3595  ICmpInst::Predicate SrcPred;
3596  Value *X;
3597  const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3598  // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3599  if (!match(&I, m_c_ICmp(SrcPred,
3601  m_APInt(C1))),
3602  m_Deferred(X))))
3603  return nullptr;
3604 
3605  // Potential handling of non-splats: for each element:
3606  // * if both are undef, replace with constant 0.
3607  // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3608  // * if both are not undef, and are different, bailout.
3609  // * else, only one is undef, then pick the non-undef one.
3610 
3611  // The shift amount must be equal.
3612  if (*C0 != *C1)
3613  return nullptr;
3614  const APInt &MaskedBits = *C0;
3615  assert(MaskedBits != 0 && "shift by zero should be folded away already.");
3616 
3617  ICmpInst::Predicate DstPred;
3618  switch (SrcPred) {
3619  case ICmpInst::Predicate::ICMP_EQ:
3620  // ((%x << MaskedBits) a>> MaskedBits) == %x
3621  // =>
3622  // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3623  DstPred = ICmpInst::Predicate::ICMP_ULT;
3624  break;
3625  case ICmpInst::Predicate::ICMP_NE:
3626  // ((%x << MaskedBits) a>> MaskedBits) != %x
3627  // =>
3628  // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3629  DstPred = ICmpInst::Predicate::ICMP_UGE;
3630  break;
3631  // FIXME: are more folds possible?
3632  default:
3633  return nullptr;
3634  }
3635 
3636  auto *XType = X->getType();
3637  const unsigned XBitWidth = XType->getScalarSizeInBits();
3638  const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3639  assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
3640 
3641  // KeptBits = bitwidth(%x) - MaskedBits
3642  const APInt KeptBits = BitWidth - MaskedBits;
3643  assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
3644  // ICmpCst = (1 << KeptBits)
3645  const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3646  assert(ICmpCst.isPowerOf2());
3647  // AddCst = (1 << (KeptBits-1))
3648  const APInt AddCst = ICmpCst.lshr(1);
3649  assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
3650 
3651  // T0 = add %x, AddCst
3652  Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3653  // T1 = T0 DstPred ICmpCst
3654  Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3655 
3656  return T1;
3657 }
3658 
3659 // Given pattern:
3660 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3661 // we should move shifts to the same hand of 'and', i.e. rewrite as
3662 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3663 // We are only interested in opposite logical shifts here.
3664 // One of the shifts can be truncated.
3665 // If we can, we want to end up creating 'lshr' shift.
3666 static Value *
3669  if (!I.isEquality() || !match(I.getOperand(1), m_Zero()) ||
3670  !I.getOperand(0)->hasOneUse())
3671  return nullptr;
3672 
3673  auto m_AnyLogicalShift = m_LogicalShift(m_Value(), m_Value());
3674 
3675  // Look for an 'and' of two logical shifts, one of which may be truncated.
3676  // We use m_TruncOrSelf() on the RHS to correctly handle commutative case.
3677  Instruction *XShift, *MaybeTruncation, *YShift;
3678  if (!match(
3679  I.getOperand(0),
3680  m_c_And(m_CombineAnd(m_AnyLogicalShift, m_Instruction(XShift)),
3682  m_AnyLogicalShift, m_Instruction(YShift))),
3683  m_Instruction(MaybeTruncation)))))
3684  return nullptr;
3685 
3686  // We potentially looked past 'trunc', but only when matching YShift,
3687  // therefore YShift must have the widest type.
3688  Instruction *WidestShift = YShift;
3689  // Therefore XShift must have the shallowest type.
3690  // Or they both have identical types if there was no truncation.
3691  Instruction *NarrowestShift = XShift;
3692 
3693  Type *WidestTy = WidestShift->getType();
3694  Type *NarrowestTy = NarrowestShift->getType();
3695  assert(NarrowestTy == I.getOperand(0)->getType() &&
3696  "We did not look past any shifts while matching XShift though.");
3697  bool HadTrunc = WidestTy != I.getOperand(0)->getType();
3698 
3699  // If YShift is a 'lshr', swap the shifts around.
3700  if (match(YShift, m_LShr(m_Value(), m_Value())))
3701  std::swap(XShift, YShift);
3702 
3703  // The shifts must be in opposite directions.
3704  auto XShiftOpcode = XShift->getOpcode();
3705  if (XShiftOpcode == YShift->getOpcode())
3706  return nullptr; // Do not care about same-direction shifts here.
3707 
3708  Value *X, *XShAmt, *Y, *YShAmt;
3709  match(XShift, m_BinOp(m_Value(X), m_ZExtOrSelf(m_Value(XShAmt))));
3710  match(YShift, m_BinOp(m_Value(Y), m_ZExtOrSelf(m_Value(YShAmt))));
3711 
3712  // If one of the values being shifted is a constant, then we will end with
3713  // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not,
3714  // however, we will need to ensure that we won't increase instruction count.
3715  if (!isa<Constant>(X) && !isa<Constant>(Y)) {
3716  // At least one of the hands of the 'and' should be one-use shift.
3717  if (!match(I.getOperand(0),
3718  m_c_And(m_OneUse(m_AnyLogicalShift), m_Value())))
3719  return nullptr;
3720  if (HadTrunc) {
3721  // Due to the 'trunc', we will need to widen X. For that either the old
3722  // 'trunc' or the shift amt in the non-truncated shift should be one-use.
3723  if (!MaybeTruncation->hasOneUse() &&
3724  !NarrowestShift->getOperand(1)->hasOneUse())
3725  return nullptr;
3726  }
3727  }
3728 
3729  // We have two shift amounts from two different shifts. The types of those
3730  // shift amounts may not match. If that's the case let's bailout now.
3731  if (XShAmt->getType() != YShAmt->getType())
3732  return nullptr;
3733 
3734  // As input, we have the following pattern:
3735  // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3736  // We want to rewrite that as:
3737  // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3738  // While we know that originally (Q+K) would not overflow
3739  // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
3740  // shift amounts. so it may now overflow in smaller bitwidth.
3741  // To ensure that does not happen, we need to ensure that the total maximal
3742  // shift amount is still representable in that smaller bit width.
3743  unsigned MaximalPossibleTotalShiftAmount =
3744  (WidestTy->getScalarSizeInBits() - 1) +
3745  (NarrowestTy->getScalarSizeInBits() - 1);
3746  APInt MaximalRepresentableShiftAmount =
3748  if (MaximalRepresentableShiftAmount.ult(MaximalPossibleTotalShiftAmount))
3749  return nullptr;
3750 
3751  // Can we fold (XShAmt+YShAmt) ?
3752  auto *NewShAmt = dyn_cast_or_null<Constant>(
3753  SimplifyAddInst(XShAmt, YShAmt, /*isNSW=*/false,
3754  /*isNUW=*/false, SQ.getWithInstruction(&I)));
3755  if (!NewShAmt)
3756  return nullptr;
3757  NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, WidestTy);
3758  unsigned WidestBitWidth = WidestTy->getScalarSizeInBits();
3759 
3760  // Is the new shift amount smaller than the bit width?
3761  // FIXME: could also rely on ConstantRange.
3762  if (!match(NewShAmt,
3763  m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
3764  APInt(WidestBitWidth, WidestBitWidth))))
3765  return nullptr;
3766 
3767  // An extra legality check is needed if we had trunc-of-lshr.
3768  if (HadTrunc && match(WidestShift, m_LShr(m_Value(), m_Value()))) {
3769  auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ,
3770  WidestShift]() {
3771  // It isn't obvious whether it's worth it to analyze non-constants here.
3772  // Also, let's basically give up on non-splat cases, pessimizing vectors.
3773  // If *any* of these preconditions matches we can perform the fold.
3774  Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy()
3775  ? NewShAmt->getSplatValue()
3776  : NewShAmt;
3777  // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold.
3778  if (NewShAmtSplat &&
3779  (NewShAmtSplat->isNullValue() ||
3780  NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1))
3781  return true;
3782  // We consider *min* leading zeros so a single outlier
3783  // blocks the transform as opposed to allowing it.
3784  if (auto *C = dyn_cast<Constant>(NarrowestShift->getOperand(0))) {
3785  KnownBits Known = computeKnownBits(C, SQ.DL);
3786  unsigned MinLeadZero = Known.countMinLeadingZeros();
3787  // If the value being shifted has at most lowest bit set we can fold.
3788  unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3789  if (MaxActiveBits <= 1)
3790  return true;
3791  // Precondition: NewShAmt u<= countLeadingZeros(C)
3792  if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(MinLeadZero))
3793  return true;
3794  }
3795  if (auto *C = dyn_cast<Constant>(WidestShift->getOperand(0))) {
3796  KnownBits Known = computeKnownBits(C, SQ.DL);
3797  unsigned MinLeadZero = Known.countMinLeadingZeros();
3798  // If the value being shifted has at most lowest bit set we can fold.
3799  unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3800  if (MaxActiveBits <= 1)
3801  return true;
3802  // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C)
3803  if (NewShAmtSplat) {
3804  APInt AdjNewShAmt =
3805  (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger();
3806  if (AdjNewShAmt.ule(MinLeadZero))
3807  return true;
3808  }
3809  }
3810  return false; // Can't tell if it's ok.
3811  };
3812  if (!CanFold())
3813  return nullptr;
3814  }
3815 
3816  // All good, we can do this fold.
3817  X = Builder.CreateZExt(X, WidestTy);
3818  Y = Builder.CreateZExt(Y, WidestTy);
3819  // The shift is the same that was for X.
3820  Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
3821  ? Builder.CreateLShr(X, NewShAmt)
3822  : Builder.CreateShl(X, NewShAmt);
3823  Value *T1 = Builder.CreateAnd(T0, Y);
3824  return Builder.CreateICmp(I.getPredicate(), T1,
3825  Constant::getNullValue(WidestTy));
3826 }
3827 
3828 /// Fold
3829 /// (-1 u/ x) u< y
3830 /// ((x * y) ?/ x) != y
3831 /// to
3832 /// @llvm.?mul.with.overflow(x, y) plus extraction of overflow bit
3833 /// Note that the comparison is commutative, while inverted (u>=, ==) predicate
3834 /// will mean that we are looking for the opposite answer.
3836  ICmpInst::Predicate Pred;
3837  Value *X, *Y;
3838  Instruction *Mul;
3839  Instruction *Div;
3840  bool NeedNegation;
3841  // Look for: (-1 u/ x) u</u>= y
3842  if (!I.isEquality() &&
3843  match(&I, m_c_ICmp(Pred,
3845  m_Instruction(Div)),
3846  m_Value(Y)))) {
3847  Mul = nullptr;
3848 
3849  // Are we checking that overflow does not happen, or does happen?
3850  switch (Pred) {
3851  case ICmpInst::Predicate::ICMP_ULT:
3852  NeedNegation = false;
3853  break; // OK
3854  case ICmpInst::Predicate::ICMP_UGE:
3855  NeedNegation = true;
3856  break; // OK
3857  default:
3858  return nullptr; // Wrong predicate.
3859  }
3860  } else // Look for: ((x * y) / x) !=/== y
3861  if (I.isEquality() &&
3862  match(&I,
3863  m_c_ICmp(Pred, m_Value(Y),
3864  m_CombineAnd(
3866  m_Value(X)),
3867  m_Instruction(Mul)),
3868  m_Deferred(X))),
3869  m_Instruction(Div))))) {
3870  NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ;
3871  } else
3872  return nullptr;
3873 
3875  // If the pattern included (x * y), we'll want to insert new instructions
3876  // right before that original multiplication so that we can replace it.
3877  bool MulHadOtherUses = Mul && !Mul->hasOneUse();
3878  if (MulHadOtherUses)
3879  Builder.SetInsertPoint(Mul);
3880 
3881  Function *F = Intrinsic::getDeclaration(I.getModule(),
3882  Div->getOpcode() == Instruction::UDiv
3883  ? Intrinsic::umul_with_overflow
3884  : Intrinsic::smul_with_overflow,
3885  X->getType());
3886  CallInst *Call = Builder.CreateCall(F, {X, Y}, "mul");
3887 
3888  // If the multiplication was used elsewhere, to ensure that we don't leave
3889  // "duplicate" instructions, replace uses of that original multiplication
3890  // with the multiplication result from the with.overflow intrinsic.
3891  if (MulHadOtherUses)
3892  replaceInstUsesWith(*Mul, Builder.CreateExtractValue(Call, 0, "mul.val"));
3893 
3894  Value *Res = Builder.CreateExtractValue(Call, 1, "mul.ov");
3895  if (NeedNegation) // This technically increases instruction count.
3896  Res = Builder.CreateNot(Res, "mul.not.ov");
3897 
3898  // If we replaced the mul, erase it. Do this after all uses of Builder,
3899  // as the mul is used as insertion point.
3900  if (MulHadOtherUses)
3901  eraseInstFromFunction(*Mul);
3902 
3903  return Res;
3904 }
3905 
3907  CmpInst::Predicate Pred;
3908  Value *X;
3909  if (!match(&I, m_c_ICmp(Pred, m_NSWNeg(m_Value(X)), m_Deferred(X))))
3910  return nullptr;
3911 
3912  if (ICmpInst::isSigned(Pred))
3913  Pred = ICmpInst::getSwappedPredicate(Pred);
3914  else if (ICmpInst::isUnsigned(Pred))
3915  Pred = ICmpInst::getSignedPredicate(Pred);
3916  // else for equality-comparisons just keep the predicate.
3917 
3918  return ICmpInst::Create(Instruction::ICmp, Pred, X,
3919  Constant::getNullValue(X->getType()), I.getName());
3920 }
3921 
3922 /// Try to fold icmp (binop), X or icmp X, (binop).
3923 /// TODO: A large part of this logic is duplicated in InstSimplify's
3924 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3925 /// duplication.
3927  const SimplifyQuery &SQ) {
3928  const SimplifyQuery Q = SQ.getWithInstruction(&I);
3929  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3930 
3931  // Special logic for binary operators.
3932  BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3933  BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3934  if (!BO0 && !BO1)
3935  return nullptr;
3936 
3937  if (Instruction *NewICmp = foldICmpXNegX(I))
3938  return NewICmp;
3939 
3940  const CmpInst::Predicate Pred = I.getPredicate();
3941  Value *X;
3942 
3943  // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3944  // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X
3945  if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
3946  (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3947  return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3948  // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0
3949  if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
3950  (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3951  return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3952 
3953  {
3954  // Similar to above: an unsigned overflow comparison may use offset + mask:
3955  // ((Op1 + C) & C) u< Op1 --> Op1 != 0
3956  // ((Op1 + C) & C) u>= Op1 --> Op1 == 0
3957  // Op0 u> ((Op0 + C) & C) --> Op0 != 0
3958  // Op0 u<= ((Op0 + C) & C) --> Op0 == 0
3959  BinaryOperator *BO;
3960  const APInt *C;
3961  if ((Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE) &&
3962  match(Op0, m_And(m_BinOp(BO), m_LowBitMask(C))) &&
3964  CmpInst::Predicate NewPred =
3966  Constant *Zero = ConstantInt::getNullValue(Op1->getType());
3967  return new ICmpInst(NewPred, Op1, Zero);
3968  }
3969 
3970  if ((Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE) &&
3971  match(Op1, m_And(m_BinOp(BO), m_LowBitMask(C))) &&
3973  CmpInst::Predicate NewPred =
3975  Constant *Zero = ConstantInt::getNullValue(Op1->getType());
3976  return new ICmpInst(NewPred, Op0, Zero);
3977  }
3978  }
3979 
3980  bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3981  if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3982  NoOp0WrapProblem =
3983  ICmpInst::isEquality(Pred) ||
3984  (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3985  (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3986  if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3987  NoOp1WrapProblem =
3988  ICmpInst::isEquality(Pred) ||
3989  (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3990  (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3991 
3992  // Analyze the case when either Op0 or Op1 is an add instruction.
3993  // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3994  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3995  if (BO0 && BO0->getOpcode() == Instruction::Add) {
3996  A = BO0->getOperand(0);
3997  B = BO0->getOperand(1);
3998  }
3999  if (BO1 && BO1->getOpcode() == Instruction::Add) {
4000  C = BO1->getOperand(0);
4001  D = BO1->getOperand(1);
4002  }
4003 
4004  // icmp (A+B), A -> icmp B, 0 for equalities or if there is no overflow.
4005  // icmp (A+B), B -> icmp A, 0 for equalities or if there is no overflow.
4006  if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
4007  return new ICmpInst(Pred, A == Op1 ? B : A,
4008  Constant::getNullValue(Op1->getType()));
4009 
4010  // icmp C, (C+D) -> icmp 0, D for equalities or if there is no overflow.
4011  // icmp D, (C+D) -> icmp 0, C for equalities or if there is no overflow.
4012  if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
4013  return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
4014  C == Op0 ? D : C);
4015 
4016  // icmp (A+B), (A+D) -> icmp B, D for equalities or if there is no overflow.
4017  if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
4018  NoOp1WrapProblem) {
4019  // Determine Y and Z in the form icmp (X+Y), (X+Z).
4020  Value *Y, *Z;
4021  if (A == C) {
4022  // C + B == C + D -> B == D
4023  Y = B;
4024  Z = D;
4025  } else if (A == D) {
4026  // D + B == C + D -> B == C
4027  Y = B;
4028  Z = C;
4029  } else if (B == C) {
4030  // A + C == C + D -> A == D
4031  Y = A;
4032  Z = D;
4033  } else {
4034  assert(B == D);
4035  // A + D == C + D -> A == C
4036  Y = A;
4037  Z = C;
4038  }
4039  return new ICmpInst(Pred, Y, Z);
4040  }
4041 
4042  // icmp slt (A + -1), Op1 -> icmp sle A, Op1
4043  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
4044  match(B, m_AllOnes()))
4045  return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
4046 
4047  // icmp sge (A + -1), Op1 -> icmp sgt A, Op1
4048  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
4049  match(B, m_AllOnes()))
4050  return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
4051 
4052  // icmp sle (A + 1), Op1 -> icmp slt A, Op1
4053  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
4054  return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
4055 
4056  // icmp sgt (A + 1), Op1 -> icmp sge A, Op1
4057  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
4058  return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
4059 
4060  // icmp sgt Op0, (C + -1) -> icmp sge Op0, C
4061  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
4062  match(D, m_AllOnes()))
4063  return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
4064 
4065  // icmp sle Op0, (C + -1) -> icmp slt Op0, C
4066  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
4067  match(D, m_AllOnes()))
4068  return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
4069 
4070  // icmp sge Op0, (C + 1) -> icmp sgt Op0, C
4071  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
4072  return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
4073 
4074  // icmp slt Op0, (C + 1) -> icmp sle Op0, C
4075  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
4076  return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
4077 
4078  // TODO: The subtraction-related identities shown below also hold, but
4079  // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
4080  // wouldn't happen even if they were implemented.
4081  //
4082  // icmp ult (A - 1), Op1 -> icmp ule A, Op1
4083  // icmp uge (A - 1), Op1 -> icmp ugt A, Op1
4084  // icmp ugt Op0, (C - 1) -> icmp uge Op0, C
4085  // icmp ule Op0, (C - 1) -> icmp ult Op0, C
4086 
4087  // icmp ule (A + 1), Op0 -> icmp ult A, Op1
4088  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
4089  return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
4090 
4091  // icmp ugt (A + 1), Op0 -> icmp uge A, Op1
4092  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
4093  return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
4094 
4095  // icmp uge Op0, (C + 1) -> icmp ugt Op0, C
4096  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
4097  return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
4098 
4099  // icmp ult Op0, (C + 1) -> icmp ule Op0, C
4100  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
4101  return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
4102 
4103  // if C1 has greater magnitude than C2:
4104  // icmp (A + C1), (C + C2) -> icmp (A + C3), C
4105  // s.t. C3 = C1 - C2
4106  //
4107  // if C2 has greater magnitude than C1:
4108  // icmp (A + C1), (C + C2) -> icmp A, (C + C3)
4109  // s.t. C3 = C2 - C1
4110  if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
4111  (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
4112  if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
4113  if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
4114  const APInt &AP1 = C1->getValue();
4115  const APInt &AP2 = C2->getValue();
4116  if (AP1.isNegative() == AP2.isNegative()) {
4117  APInt AP1Abs = C1->getValue().abs();
4118  APInt AP2Abs = C2->getValue().abs();
4119  if (AP1Abs.uge(AP2Abs)) {
4120  ConstantInt *C3 = Builder.getInt(AP1 - AP2);
4121  bool HasNUW = BO0->hasNoUnsignedWrap() && C3->getValue().ule(AP1);
4122  bool HasNSW = BO0->hasNoSignedWrap();
4123  Value *NewAdd = Builder.CreateAdd(A, C3, "", HasNUW, HasNSW);
4124  return new ICmpInst(Pred, NewAdd, C);
4125  } else {
4126  ConstantInt *C3 = Builder.getInt(AP2 - AP1);
4127  bool HasNUW = BO1->hasNoUnsignedWrap() && C3->getValue().ule(AP2);
4128  bool HasNSW = BO1->hasNoSignedWrap();
4129  Value *NewAdd = Builder.CreateAdd(C, C3, "", HasNUW, HasNSW);
4130  return new ICmpInst(Pred, A, NewAdd);
4131  }
4132  }
4133  }
4134 
4135  // Analyze the case when either Op0 or Op1 is a sub instruction.
4136  // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
4137  A = nullptr;
4138  B = nullptr;
4139  C = nullptr;
4140  D = nullptr;
4141  if (BO0 && BO0->getOpcode() == Instruction::Sub) {
4142  A = BO0->getOperand(0);
4143  B = BO0->getOperand(1);
4144  }
4145  if (BO1 && BO1->getOpcode() == Instruction::Sub) {
4146  C = BO1->getOperand(0);
4147  D = BO1->getOperand(1);
4148  }
4149 
4150  // icmp (A-B), A -> icmp 0, B for equalities or if there is no overflow.
4151  if (A == Op1 && NoOp0WrapProblem)
4152  return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
4153  // icmp C, (C-D) -> icmp D, 0 for equalities or if there is no overflow.
4154  if (C == Op0 && NoOp1WrapProblem)
4155  return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
4156 
4157  // Convert sub-with-unsigned-overflow comparisons into a comparison of args.
4158  // (A - B) u>/u<= A --> B u>/u<= A
4159  if (A == Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
4160  return new ICmpInst(Pred, B, A);
4161  // C u</u>= (C - D) --> C u</u>= D
4162  if (C == Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
4163  return new ICmpInst(Pred, C, D);
4164  // (A - B) u>=/u< A --> B u>/u<= A iff B != 0
4165  if (A == Op1 && (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
4166  isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
4167  return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), B, A);
4168  // C u<=/u> (C - D) --> C u</u>= D iff B != 0
4169  if (C == Op0 && (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
4170  isKnownNonZero(D, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
4171  return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), C, D);
4172 
4173  // icmp (A-B), (C-B) -> icmp A, C for equalities or if there is no overflow.
4174  if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem)
4175  return new ICmpInst(Pred, A, C);
4176 
4177  // icmp (A-B), (A-D) -> icmp D, B for equalities or if there is no overflow.
4178  if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem)
4179  return new ICmpInst(Pred, D, B);
4180 
4181  // icmp (0-X) < cst --> x > -cst
4182  if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
4183  Value *X;
4184  if (match(BO0, m_Neg(m_Value(X))))
4185  if (Constant *RHSC = dyn_cast<Constant>(Op1))
4186  if (RHSC->isNotMinSignedValue())
4187  return new ICmpInst(I.getSwappedPredicate(), X,
4188  ConstantExpr::getNeg(RHSC));
4189  }
4190 
4191  {
4192  // Try to remove shared constant multiplier from equality comparison:
4193  // X * C == Y * C (with no overflowing/aliasing) --> X == Y
4194  Value *X, *Y;
4195  const APInt *C;
4196  if (match(Op0, m_Mul(m_Value(X), m_APInt(C))) && *C != 0 &&
4197  match(Op1, m_Mul(m_Value(Y), m_SpecificInt(*C))) && I.isEquality())
4198  if (!C->countTrailingZeros() ||
4199  (BO0 && BO1 && BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap()) ||
4200  (BO0 && BO1 && BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap()))
4201  return new ICmpInst(Pred, X, Y);
4202  }
4203 
4204  BinaryOperator *SRem = nullptr;
4205  // icmp (srem X, Y), Y
4206  if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
4207  SRem = BO0;
4208  // icmp Y, (srem X, Y)
4209  else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
4210  Op0 == BO1->getOperand(1))
4211  SRem = BO1;
4212  if (SRem) {
4213  // We don't check hasOneUse to avoid increasing register pressure because
4214  // the value we use is the same value this instruction was already using.
4215  switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
4216  default:
4217  break;
4218  case ICmpInst::ICMP_EQ:
4219  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4220  case ICmpInst::ICMP_NE:
4221  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4222  case ICmpInst::ICMP_SGT:
4223  case ICmpInst::ICMP_SGE:
4224  return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
4226  case ICmpInst::ICMP_SLT:
4227  case ICmpInst::ICMP_SLE:
4228  return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
4229  Constant::getNullValue(SRem->getType()));
4230  }
4231  }
4232 
4233  if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
4234  BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
4235  switch (BO0->getOpcode()) {
4236  default:
4237  break;
4238  case Instruction::Add:
4239  case Instruction::Sub:
4240  case Instruction::Xor: {
4241  if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
4242  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4243 
4244  const APInt *C;
4245  if (match(BO0->getOperand(1), m_APInt(C))) {
4246  // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
4247  if (C->isSignMask()) {
4248  ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
4249  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4250  }
4251 
4252  // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
4253  if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
4254  ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
4255  NewPred = I.getSwappedPredicate(NewPred);
4256  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4257  }
4258  }
4259  break;
4260  }
4261  case Instruction::Mul: {
4262  if (!I.isEquality())
4263  break;
4264 
4265  const APInt *C;
4266  if (match(BO0->getOperand(1), m_APInt(C)) && !C->isZero() &&
4267  !C->isOne()) {
4268  // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
4269  // Mask = -1 >> count-trailing-zeros(C).
4270  if (unsigned TZs = C->countTrailingZeros()) {
4272  BO0->getType(),
4273  APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
4274  Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
4275  Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
4276  return new ICmpInst(Pred, And1, And2);
4277  }
4278  }
4279  break;
4280  }
4281  case Instruction::UDiv:
4282  case Instruction::LShr:
4283  if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
4284  break;
4285  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4286 
4287  case Instruction::SDiv:
4288  if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
4289  break;
4290  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4291 
4292  case Instruction::AShr:
4293  if (!BO0->isExact() || !BO1->isExact())
4294  break;
4295  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4296 
4297  case Instruction::Shl: {
4298  bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
4299  bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
4300  if (!NUW && !NSW)
4301  break;
4302  if (!NSW && I.isSigned())
4303  break;
4304  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4305  }
4306  }
4307  }
4308 
4309  if (BO0) {
4310  // Transform A & (L - 1) `ult` L --> L != 0
4311  auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
4312  auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
4313 
4314  if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
4315  auto *Zero = Constant::getNullValue(BO0->getType());
4316  return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
4317  }
4318  }
4319 
4320  if (Value *V = foldMultiplicationOverflowCheck(I))
4321  return replaceInstUsesWith(I, V);
4322 
4324  return replaceInstUsesWith(I, V);
4325 
4327  return replaceInstUsesWith(I, V);
4328 
4330  return replaceInstUsesWith(I, V);
4331 
4332  return nullptr;
4333 }
4334 
4335 /// Fold icmp Pred min|max(X, Y), X.
4337  ICmpInst::Predicate Pred = Cmp.getPredicate();
4338  Value *Op0 = Cmp.getOperand(0);
4339  Value *X = Cmp.getOperand(1);
4340 
4341  // Canonicalize minimum or maximum operand to LHS of the icmp.
4342  if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
4343  match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
4344  match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
4345  match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
4346  std::swap(Op0, X);
4347  Pred = Cmp.getSwappedPredicate();
4348  }
4349 
4350  Value *Y;
4351  if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
4352  // smin(X, Y) == X --> X s<= Y
4353  // smin(X, Y) s>= X --> X s<= Y
4354  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
4355  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
4356 
4357  // smin(X, Y) != X --> X s> Y
4358  // smin(X, Y) s< X --> X s> Y
4359  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
4360  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
4361 
4362  // These cases should be handled in InstSimplify:
4363  // smin(X, Y) s<= X --> true
4364  // smin(X, Y) s> X --> false
4365  return nullptr;
4366  }
4367 
4368  if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
4369  // smax(X, Y) == X --> X s>= Y
4370  // smax(X, Y) s<= X --> X s>= Y
4371  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
4372  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
4373 
4374  // smax(X, Y) != X --> X s< Y
4375  // smax(X, Y) s> X --> X s< Y
4376  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
4377  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
4378 
4379  // These cases should be handled in InstSimplify:
4380  // smax(X, Y) s>= X --> true
4381  // smax(X, Y) s< X --> false
4382  return nullptr;
4383  }
4384 
4385  if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
4386  // umin(X, Y) == X --> X u<= Y
4387  // umin(X, Y) u>= X --> X u<= Y
4388  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
4389  return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
4390 
4391  // umin(X, Y) != X --> X u> Y
4392  // umin(X, Y) u< X --> X u> Y
4393  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
4394  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
4395 
4396  // These cases should be handled in InstSimplify:
4397  // umin(X, Y) u<= X --> true
4398  // umin(X, Y) u> X --> false
4399  return nullptr;
4400  }
4401 
4402  if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
4403  // umax(X, Y) == X --> X u>= Y
4404  // umax(X, Y) u<= X --> X u>= Y
4405  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
4406  return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
4407 
4408  // umax(X, Y) != X --> X u< Y
4409  // umax(X, Y) u> X --> X u< Y
4410  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
4411  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
4412 
4413  // These cases should be handled in InstSimplify:
4414  // umax(X, Y) u>= X --> true
4415  // umax(X, Y) u< X --> false
4416  return nullptr;
4417  }
4418 
4419  return nullptr;
4420 }
4421 
4423  if (!I.isEquality())
4424  return nullptr;
4425 
4426  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4427  const CmpInst::Predicate Pred = I.getPredicate();
4428  Value *A, *B, *C, *D;
4429  if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4430  if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
4431  Value *OtherVal = A == Op1 ? B : A;
4432  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4433  }
4434 
4435  if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
4436  // A^c1 == C^c2 --> A == C^(c1^c2)
4437  ConstantInt *C1, *C2;
4438  if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
4439  Op1->hasOneUse()) {
4440  Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
4441  Value *Xor = Builder.CreateXor(C, NC);
4442  return new ICmpInst(Pred, A, Xor);
4443  }
4444 
4445  // A^B == A^D -> B == D
4446  if (A == C)
4447  return new ICmpInst(Pred, B, D);
4448  if (A == D)
4449  return new ICmpInst(Pred, B, C);
4450  if (B == C)
4451  return new ICmpInst(Pred, A, D);
4452  if (B == D)
4453  return new ICmpInst(Pred, A, C);
4454  }
4455  }
4456 
4457  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
4458  // A == (A^B) -> B == 0
4459  Value *OtherVal = A == Op0 ? B : A;
4460  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4461  }
4462 
4463  // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
4464  if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
4465  match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
4466  Value *X = nullptr, *Y = nullptr, *Z = nullptr;
4467 
4468  if (A == C) {
4469  X = B;
4470  Y = D;
4471  Z = A;
4472  } else if (A == D) {
4473  X = B;
4474  Y = C;
4475  Z = A;
4476  } else if (B == C) {
4477  X = A;
4478  Y = D;
4479  Z = B;
4480  } else if (B == D) {
4481  X = A;
4482  Y = C;
4483  Z = B;
4484  }
4485 
4486  if (X) { // Build (X^Y) & Z
4487  Op1 = Builder.CreateXor(X, Y);
4488  Op1 = Builder.CreateAnd(Op1, Z);
4489  return new ICmpInst(Pred, Op1, Constant::getNullValue(Op1->getType()));
4490  }
4491  }
4492 
4493  {
4494  // Similar to above, but specialized for constant because invert is needed:
4495  // (X | C) == (Y | C) --> (X ^ Y) & ~C == 0
4496  Value *X, *Y;
4497  Constant *C;
4498  if (match(Op0, m_OneUse(m_Or(m_Value(X), m_Constant(C)))) &&
4499  match(Op1, m_OneUse(m_Or(m_Value(Y), m_Specific(C))))) {
4500  Value *Xor = Builder.CreateXor(X, Y);
4501  Value *And = Builder.CreateAnd(Xor, ConstantExpr::getNot(C));
4502  return new ICmpInst(Pred, And, Constant::getNullValue(And->getType()));
4503  }
4504  }
4505 
4506  // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
4507  // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
4508  ConstantInt *Cst1;
4509  if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
4510  match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
4511  (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
4512  match(Op1, m_ZExt(m_Value(A))))) {
4513  APInt Pow2 = Cst1->getValue() + 1;
4514  if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
4515  Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
4516  return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
4517  }
4518 
4519  // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
4520  // For lshr and ashr pairs.
4521  if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4522  match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
4523  (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4524  match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
4525  unsigned TypeBits = Cst1->getBitWidth();
4526  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4527  if (ShAmt < TypeBits && ShAmt != 0) {
4528  ICmpInst::Predicate NewPred =
4530  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4531  APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
4532  return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
4533  }
4534  }
4535 
4536  // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
4537  if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
4538  match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
4539  unsigned TypeBits = Cst1->getBitWidth();
4540  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4541  if (ShAmt < TypeBits && ShAmt != 0) {
4542  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4543  APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
4544  Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
4545  I.getName() + ".mask");
4546  return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
4547  }
4548  }
4549 
4550  // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
4551  // "icmp (and X, mask), cst"
4552  uint64_t ShAmt = 0;
4553  if (Op0->hasOneUse() &&
4554  match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
4555