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"
20 #include "llvm/IR/ConstantRange.h"
21 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/IntrinsicInst.h"
24 #include "llvm/IR/PatternMatch.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/KnownBits.h"
28 
29 using namespace llvm;
30 using namespace PatternMatch;
31 
32 #define DEBUG_TYPE "instcombine"
33 
34 // How many times is a select replaced by one of its operands?
35 STATISTIC(NumSel, "Number of select opts");
36 
37 
38 /// Compute Result = In1+In2, returning true if the result overflowed for this
39 /// type.
40 static bool addWithOverflow(APInt &Result, const APInt &In1,
41  const APInt &In2, bool IsSigned = false) {
42  bool Overflow;
43  if (IsSigned)
44  Result = In1.sadd_ov(In2, Overflow);
45  else
46  Result = In1.uadd_ov(In2, Overflow);
47 
48  return Overflow;
49 }
50 
51 /// Compute Result = In1-In2, returning true if the result overflowed for this
52 /// type.
53 static bool subWithOverflow(APInt &Result, const APInt &In1,
54  const APInt &In2, bool IsSigned = false) {
55  bool Overflow;
56  if (IsSigned)
57  Result = In1.ssub_ov(In2, Overflow);
58  else
59  Result = In1.usub_ov(In2, Overflow);
60 
61  return Overflow;
62 }
63 
64 /// Given an icmp instruction, return true if any use of this comparison is a
65 /// branch on sign bit comparison.
66 static bool hasBranchUse(ICmpInst &I) {
67  for (auto *U : I.users())
68  if (isa<BranchInst>(U))
69  return true;
70  return false;
71 }
72 
73 /// Returns true if the exploded icmp can be expressed as a signed comparison
74 /// to zero and updates the predicate accordingly.
75 /// The signedness of the comparison is preserved.
76 /// TODO: Refactor with decomposeBitTestICmp()?
77 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
78  if (!ICmpInst::isSigned(Pred))
79  return false;
80 
81  if (C.isNullValue())
82  return ICmpInst::isRelational(Pred);
83 
84  if (C.isOneValue()) {
85  if (Pred == ICmpInst::ICMP_SLT) {
86  Pred = ICmpInst::ICMP_SLE;
87  return true;
88  }
89  } else if (C.isAllOnesValue()) {
90  if (Pred == ICmpInst::ICMP_SGT) {
91  Pred = ICmpInst::ICMP_SGE;
92  return true;
93  }
94  }
95 
96  return false;
97 }
98 
99 /// This is called when we see this pattern:
100 /// cmp pred (load (gep GV, ...)), cmpcst
101 /// where GV is a global variable with a constant initializer. Try to simplify
102 /// this into some simple computation that does not need the load. For example
103 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
104 ///
105 /// If AndCst is non-null, then the loaded value is masked with that constant
106 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
107 Instruction *
109  GlobalVariable *GV, CmpInst &ICI,
110  ConstantInt *AndCst) {
111  Constant *Init = GV->getInitializer();
112  if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
113  return nullptr;
114 
115  uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
116  // Don't blow up on huge arrays.
117  if (ArrayElementCount > MaxArraySizeForCombine)
118  return nullptr;
119 
120  // There are many forms of this optimization we can handle, for now, just do
121  // the simple index into a single-dimensional array.
122  //
123  // Require: GEP GV, 0, i {{, constant indices}}
124  if (GEP->getNumOperands() < 3 ||
125  !isa<ConstantInt>(GEP->getOperand(1)) ||
126  !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
127  isa<Constant>(GEP->getOperand(2)))
128  return nullptr;
129 
130  // Check that indices after the variable are constants and in-range for the
131  // type they index. Collect the indices. This is typically for arrays of
132  // structs.
133  SmallVector<unsigned, 4> LaterIndices;
134 
135  Type *EltTy = Init->getType()->getArrayElementType();
136  for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
137  ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
138  if (!Idx) return nullptr; // Variable index.
139 
140  uint64_t IdxVal = Idx->getZExtValue();
141  if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
142 
143  if (StructType *STy = dyn_cast<StructType>(EltTy))
144  EltTy = STy->getElementType(IdxVal);
145  else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
146  if (IdxVal >= ATy->getNumElements()) return nullptr;
147  EltTy = ATy->getElementType();
148  } else {
149  return nullptr; // Unknown type.
150  }
151 
152  LaterIndices.push_back(IdxVal);
153  }
154 
155  enum { Overdefined = -3, Undefined = -2 };
156 
157  // Variables for our state machines.
158 
159  // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
160  // "i == 47 | i == 87", where 47 is the first index the condition is true for,
161  // and 87 is the second (and last) index. FirstTrueElement is -2 when
162  // undefined, otherwise set to the first true element. SecondTrueElement is
163  // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
164  int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
165 
166  // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
167  // form "i != 47 & i != 87". Same state transitions as for true elements.
168  int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
169 
170  /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
171  /// define a state machine that triggers for ranges of values that the index
172  /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
173  /// This is -2 when undefined, -3 when overdefined, and otherwise the last
174  /// index in the range (inclusive). We use -2 for undefined here because we
175  /// use relative comparisons and don't want 0-1 to match -1.
176  int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
177 
178  // MagicBitvector - This is a magic bitvector where we set a bit if the
179  // comparison is true for element 'i'. If there are 64 elements or less in
180  // the array, this will fully represent all the comparison results.
181  uint64_t MagicBitvector = 0;
182 
183  // Scan the array and see if one of our patterns matches.
184  Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
185  for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
186  Constant *Elt = Init->getAggregateElement(i);
187  if (!Elt) return nullptr;
188 
189  // If this is indexing an array of structures, get the structure element.
190  if (!LaterIndices.empty())
191  Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
192 
193  // If the element is masked, handle it.
194  if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
195 
196  // Find out if the comparison would be true or false for the i'th element.
198  CompareRHS, DL, &TLI);
199  // If the result is undef for this element, ignore it.
200  if (isa<UndefValue>(C)) {
201  // Extend range state machines to cover this element in case there is an
202  // undef in the middle of the range.
203  if (TrueRangeEnd == (int)i-1)
204  TrueRangeEnd = i;
205  if (FalseRangeEnd == (int)i-1)
206  FalseRangeEnd = i;
207  continue;
208  }
209 
210  // If we can't compute the result for any of the elements, we have to give
211  // up evaluating the entire conditional.
212  if (!isa<ConstantInt>(C)) return nullptr;
213 
214  // Otherwise, we know if the comparison is true or false for this element,
215  // update our state machines.
216  bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
217 
218  // State machine for single/double/range index comparison.
219  if (IsTrueForElt) {
220  // Update the TrueElement state machine.
221  if (FirstTrueElement == Undefined)
222  FirstTrueElement = TrueRangeEnd = i; // First true element.
223  else {
224  // Update double-compare state machine.
225  if (SecondTrueElement == Undefined)
226  SecondTrueElement = i;
227  else
228  SecondTrueElement = Overdefined;
229 
230  // Update range state machine.
231  if (TrueRangeEnd == (int)i-1)
232  TrueRangeEnd = i;
233  else
234  TrueRangeEnd = Overdefined;
235  }
236  } else {
237  // Update the FalseElement state machine.
238  if (FirstFalseElement == Undefined)
239  FirstFalseElement = FalseRangeEnd = i; // First false element.
240  else {
241  // Update double-compare state machine.
242  if (SecondFalseElement == Undefined)
243  SecondFalseElement = i;
244  else
245  SecondFalseElement = Overdefined;
246 
247  // Update range state machine.
248  if (FalseRangeEnd == (int)i-1)
249  FalseRangeEnd = i;
250  else
251  FalseRangeEnd = Overdefined;
252  }
253  }
254 
255  // If this element is in range, update our magic bitvector.
256  if (i < 64 && IsTrueForElt)
257  MagicBitvector |= 1ULL << i;
258 
259  // If all of our states become overdefined, bail out early. Since the
260  // predicate is expensive, only check it every 8 elements. This is only
261  // really useful for really huge arrays.
262  if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
263  SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
264  FalseRangeEnd == Overdefined)
265  return nullptr;
266  }
267 
268  // Now that we've scanned the entire array, emit our new comparison(s). We
269  // order the state machines in complexity of the generated code.
270  Value *Idx = GEP->getOperand(2);
271 
272  // If the index is larger than the pointer size of the target, truncate the
273  // index down like the GEP would do implicitly. We don't have to do this for
274  // an inbounds GEP because the index can't be out of range.
275  if (!GEP->isInBounds()) {
276  Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
277  unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
278  if (Idx->getType()->getPrimitiveSizeInBits().getFixedSize() > PtrSize)
279  Idx = Builder.CreateTrunc(Idx, IntPtrTy);
280  }
281 
282  // If inbounds keyword is not present, Idx * ElementSize can overflow.
283  // Let's assume that ElementSize is 2 and the wanted value is at offset 0.
284  // Then, there are two possible values for Idx to match offset 0:
285  // 0x00..00, 0x80..00.
286  // Emitting 'icmp eq Idx, 0' isn't correct in this case because the
287  // comparison is false if Idx was 0x80..00.
288  // We need to erase the highest countTrailingZeros(ElementSize) bits of Idx.
289  unsigned ElementSize =
290  DL.getTypeAllocSize(Init->getType()->getArrayElementType());
291  auto MaskIdx = [&](Value* Idx){
292  if (!GEP->isInBounds() && countTrailingZeros(ElementSize) != 0) {
293  Value *Mask = ConstantInt::get(Idx->getType(), -1);
294  Mask = Builder.CreateLShr(Mask, countTrailingZeros(ElementSize));
295  Idx = Builder.CreateAnd(Idx, Mask);
296  }
297  return Idx;
298  };
299 
300  // If the comparison is only true for one or two elements, emit direct
301  // comparisons.
302  if (SecondTrueElement != Overdefined) {
303  Idx = MaskIdx(Idx);
304  // None true -> false.
305  if (FirstTrueElement == Undefined)
306  return replaceInstUsesWith(ICI, Builder.getFalse());
307 
308  Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
309 
310  // True for one element -> 'i == 47'.
311  if (SecondTrueElement == Undefined)
312  return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
313 
314  // True for two elements -> 'i == 47 | i == 72'.
315  Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
316  Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
317  Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
318  return BinaryOperator::CreateOr(C1, C2);
319  }
320 
321  // If the comparison is only false for one or two elements, emit direct
322  // comparisons.
323  if (SecondFalseElement != Overdefined) {
324  Idx = MaskIdx(Idx);
325  // None false -> true.
326  if (FirstFalseElement == Undefined)
327  return replaceInstUsesWith(ICI, Builder.getTrue());
328 
329  Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
330 
331  // False for one element -> 'i != 47'.
332  if (SecondFalseElement == Undefined)
333  return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
334 
335  // False for two elements -> 'i != 47 & i != 72'.
336  Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
337  Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
338  Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
339  return BinaryOperator::CreateAnd(C1, C2);
340  }
341 
342  // If the comparison can be replaced with a range comparison for the elements
343  // where it is true, emit the range check.
344  if (TrueRangeEnd != Overdefined) {
345  assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
346  Idx = MaskIdx(Idx);
347 
348  // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
349  if (FirstTrueElement) {
350  Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
351  Idx = Builder.CreateAdd(Idx, Offs);
352  }
353 
354  Value *End = ConstantInt::get(Idx->getType(),
355  TrueRangeEnd-FirstTrueElement+1);
356  return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
357  }
358 
359  // False range check.
360  if (FalseRangeEnd != Overdefined) {
361  assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
362  Idx = MaskIdx(Idx);
363  // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
364  if (FirstFalseElement) {
365  Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
366  Idx = Builder.CreateAdd(Idx, Offs);
367  }
368 
369  Value *End = ConstantInt::get(Idx->getType(),
370  FalseRangeEnd-FirstFalseElement);
371  return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
372  }
373 
374  // If a magic bitvector captures the entire comparison state
375  // of this load, replace it with computation that does:
376  // ((magic_cst >> i) & 1) != 0
377  {
378  Type *Ty = nullptr;
379 
380  // Look for an appropriate type:
381  // - The type of Idx if the magic fits
382  // - The smallest fitting legal type
383  if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
384  Ty = Idx->getType();
385  else
386  Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
387 
388  if (Ty) {
389  Idx = MaskIdx(Idx);
390  Value *V = Builder.CreateIntCast(Idx, Ty, false);
391  V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
392  V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
393  return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
394  }
395  }
396 
397  return nullptr;
398 }
399 
400 /// Return a value that can be used to compare the *offset* implied by a GEP to
401 /// zero. For example, if we have &A[i], we want to return 'i' for
402 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
403 /// are involved. The above expression would also be legal to codegen as
404 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
405 /// This latter form is less amenable to optimization though, and we are allowed
406 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
407 ///
408 /// If we can't emit an optimized form for this expression, this returns null.
409 ///
411  const DataLayout &DL) {
413 
414  // Check to see if this gep only has a single variable index. If so, and if
415  // any constant indices are a multiple of its scale, then we can compute this
416  // in terms of the scale of the variable index. For example, if the GEP
417  // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
418  // because the expression will cross zero at the same point.
419  unsigned i, e = GEP->getNumOperands();
420  int64_t Offset = 0;
421  for (i = 1; i != e; ++i, ++GTI) {
422  if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
423  // Compute the aggregate offset of constant indices.
424  if (CI->isZero()) continue;
425 
426  // Handle a struct index, which adds its field offset to the pointer.
427  if (StructType *STy = GTI.getStructTypeOrNull()) {
428  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
429  } else {
430  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
431  Offset += Size*CI->getSExtValue();
432  }
433  } else {
434  // Found our variable index.
435  break;
436  }
437  }
438 
439  // If there are no variable indices, we must have a constant offset, just
440  // evaluate it the general way.
441  if (i == e) return nullptr;
442 
443  Value *VariableIdx = GEP->getOperand(i);
444  // Determine the scale factor of the variable element. For example, this is
445  // 4 if the variable index is into an array of i32.
446  uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
447 
448  // Verify that there are no other variable indices. If so, emit the hard way.
449  for (++i, ++GTI; i != e; ++i, ++GTI) {
450  ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
451  if (!CI) return nullptr;
452 
453  // Compute the aggregate offset of constant indices.
454  if (CI->isZero()) continue;
455 
456  // Handle a struct index, which adds its field offset to the pointer.
457  if (StructType *STy = GTI.getStructTypeOrNull()) {
458  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
459  } else {
460  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
461  Offset += Size*CI->getSExtValue();
462  }
463  }
464 
465  // Okay, we know we have a single variable index, which must be a
466  // pointer/array/vector index. If there is no offset, life is simple, return
467  // the index.
468  Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
469  unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
470  if (Offset == 0) {
471  // Cast to intptrty in case a truncation occurs. If an extension is needed,
472  // we don't need to bother extending: the extension won't affect where the
473  // computation crosses zero.
474  if (VariableIdx->getType()->getPrimitiveSizeInBits().getFixedSize() >
475  IntPtrWidth) {
476  VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
477  }
478  return VariableIdx;
479  }
480 
481  // Otherwise, there is an index. The computation we will do will be modulo
482  // the pointer size.
483  Offset = SignExtend64(Offset, IntPtrWidth);
484  VariableScale = SignExtend64(VariableScale, IntPtrWidth);
485 
486  // To do this transformation, any constant index must be a multiple of the
487  // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
488  // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
489  // multiple of the variable scale.
490  int64_t NewOffs = Offset / (int64_t)VariableScale;
491  if (Offset != NewOffs*(int64_t)VariableScale)
492  return nullptr;
493 
494  // Okay, we can do this evaluation. Start by converting the index to intptr.
495  if (VariableIdx->getType() != IntPtrTy)
496  VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
497  true /*Signed*/);
498  Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
499  return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
500 }
501 
502 /// Returns true if we can rewrite Start as a GEP with pointer Base
503 /// and some integer offset. The nodes that need to be re-written
504 /// for this transformation will be added to Explored.
505 static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
506  const DataLayout &DL,
507  SetVector<Value *> &Explored) {
508  SmallVector<Value *, 16> WorkList(1, Start);
509  Explored.insert(Base);
510 
511  // The following traversal gives us an order which can be used
512  // when doing the final transformation. Since in the final
513  // transformation we create the PHI replacement instructions first,
514  // we don't have to get them in any particular order.
515  //
516  // However, for other instructions we will have to traverse the
517  // operands of an instruction first, which means that we have to
518  // do a post-order traversal.
519  while (!WorkList.empty()) {
521 
522  while (!WorkList.empty()) {
523  if (Explored.size() >= 100)
524  return false;
525 
526  Value *V = WorkList.back();
527 
528  if (Explored.contains(V)) {
529  WorkList.pop_back();
530  continue;
531  }
532 
533  if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
534  !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
535  // We've found some value that we can't explore which is different from
536  // the base. Therefore we can't do this transformation.
537  return false;
538 
539  if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
540  auto *CI = cast<CastInst>(V);
541  if (!CI->isNoopCast(DL))
542  return false;
543 
544  if (Explored.count(CI->getOperand(0)) == 0)
545  WorkList.push_back(CI->getOperand(0));
546  }
547 
548  if (auto *GEP = dyn_cast<GEPOperator>(V)) {
549  // We're limiting the GEP to having one index. This will preserve
550  // the original pointer type. We could handle more cases in the
551  // future.
552  if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
553  GEP->getType() != Start->getType())
554  return false;
555 
556  if (Explored.count(GEP->getOperand(0)) == 0)
557  WorkList.push_back(GEP->getOperand(0));
558  }
559 
560  if (WorkList.back() == V) {
561  WorkList.pop_back();
562  // We've finished visiting this node, mark it as such.
563  Explored.insert(V);
564  }
565 
566  if (auto *PN = dyn_cast<PHINode>(V)) {
567  // We cannot transform PHIs on unsplittable basic blocks.
568  if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
569  return false;
570  Explored.insert(PN);
571  PHIs.insert(PN);
572  }
573  }
574 
575  // Explore the PHI nodes further.
576  for (auto *PN : PHIs)
577  for (Value *Op : PN->incoming_values())
578  if (Explored.count(Op) == 0)
579  WorkList.push_back(Op);
580  }
581 
582  // Make sure that we can do this. Since we can't insert GEPs in a basic
583  // block before a PHI node, we can't easily do this transformation if
584  // we have PHI node users of transformed instructions.
585  for (Value *Val : Explored) {
586  for (Value *Use : Val->uses()) {
587 
588  auto *PHI = dyn_cast<PHINode>(Use);
589  auto *Inst = dyn_cast<Instruction>(Val);
590 
591  if (Inst == Base || Inst == PHI || !Inst || !PHI ||
592  Explored.count(PHI) == 0)
593  continue;
594 
595  if (PHI->getParent() == Inst->getParent())
596  return false;
597  }
598  }
599  return true;
600 }
601 
602 // Sets the appropriate insert point on Builder where we can add
603 // a replacement Instruction for V (if that is possible).
605  bool Before = true) {
606  if (auto *PHI = dyn_cast<PHINode>(V)) {
607  Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
608  return;
609  }
610  if (auto *I = dyn_cast<Instruction>(V)) {
611  if (!Before)
612  I = &*std::next(I->getIterator());
613  Builder.SetInsertPoint(I);
614  return;
615  }
616  if (auto *A = dyn_cast<Argument>(V)) {
617  // Set the insertion point in the entry block.
618  BasicBlock &Entry = A->getParent()->getEntryBlock();
619  Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
620  return;
621  }
622  // Otherwise, this is a constant and we don't need to set a new
623  // insertion point.
624  assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
625 }
626 
627 /// Returns a re-written value of Start as an indexed GEP using Base as a
628 /// pointer.
630  const DataLayout &DL,
631  SetVector<Value *> &Explored) {
632  // Perform all the substitutions. This is a bit tricky because we can
633  // have cycles in our use-def chains.
634  // 1. Create the PHI nodes without any incoming values.
635  // 2. Create all the other values.
636  // 3. Add the edges for the PHI nodes.
637  // 4. Emit GEPs to get the original pointers.
638  // 5. Remove the original instructions.
639  Type *IndexType = IntegerType::get(
640  Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
641 
643  NewInsts[Base] = ConstantInt::getNullValue(IndexType);
644 
645  // Create the new PHI nodes, without adding any incoming values.
646  for (Value *Val : Explored) {
647  if (Val == Base)
648  continue;
649  // Create empty phi nodes. This avoids cyclic dependencies when creating
650  // the remaining instructions.
651  if (auto *PHI = dyn_cast<PHINode>(Val))
652  NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
653  PHI->getName() + ".idx", PHI);
654  }
655  IRBuilder<> Builder(Base->getContext());
656 
657  // Create all the other instructions.
658  for (Value *Val : Explored) {
659 
660  if (NewInsts.find(Val) != NewInsts.end())
661  continue;
662 
663  if (auto *CI = dyn_cast<CastInst>(Val)) {
664  // Don't get rid of the intermediate variable here; the store can grow
665  // the map which will invalidate the reference to the input value.
666  Value *V = NewInsts[CI->getOperand(0)];
667  NewInsts[CI] = V;
668  continue;
669  }
670  if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
671  Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
672  : GEP->getOperand(1);
674  // Indices might need to be sign extended. GEPs will magically do
675  // this, but we need to do it ourselves here.
676  if (Index->getType()->getScalarSizeInBits() !=
677  NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
678  Index = Builder.CreateSExtOrTrunc(
679  Index, NewInsts[GEP->getOperand(0)]->getType(),
680  GEP->getOperand(0)->getName() + ".sext");
681  }
682 
683  auto *Op = NewInsts[GEP->getOperand(0)];
684  if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
685  NewInsts[GEP] = Index;
686  else
687  NewInsts[GEP] = Builder.CreateNSWAdd(
688  Op, Index, GEP->getOperand(0)->getName() + ".add");
689  continue;
690  }
691  if (isa<PHINode>(Val))
692  continue;
693 
694  llvm_unreachable("Unexpected instruction type");
695  }
696 
697  // Add the incoming values to the PHI nodes.
698  for (Value *Val : Explored) {
699  if (Val == Base)
700  continue;
701  // All the instructions have been created, we can now add edges to the
702  // phi nodes.
703  if (auto *PHI = dyn_cast<PHINode>(Val)) {
704  PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
705  for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
706  Value *NewIncoming = PHI->getIncomingValue(I);
707 
708  if (NewInsts.find(NewIncoming) != NewInsts.end())
709  NewIncoming = NewInsts[NewIncoming];
710 
711  NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
712  }
713  }
714  }
715 
716  for (Value *Val : Explored) {
717  if (Val == Base)
718  continue;
719 
720  // Depending on the type, for external users we have to emit
721  // a GEP or a GEP + ptrtoint.
722  setInsertionPoint(Builder, Val, false);
723 
724  // If required, create an inttoptr instruction for Base.
725  Value *NewBase = Base;
726  if (!Base->getType()->isPointerTy())
727  NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
728  Start->getName() + "to.ptr");
729 
730  Value *GEP = Builder.CreateInBoundsGEP(
731  Start->getType()->getPointerElementType(), NewBase,
732  makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
733 
734  if (!Val->getType()->isPointerTy()) {
735  Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
736  Val->getName() + ".conv");
737  GEP = Cast;
738  }
739  Val->replaceAllUsesWith(GEP);
740  }
741 
742  return NewInsts[Start];
743 }
744 
745 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
746 /// the input Value as a constant indexed GEP. Returns a pair containing
747 /// the GEPs Pointer and Index.
748 static std::pair<Value *, Value *>
750  Type *IndexType = IntegerType::get(V->getContext(),
751  DL.getIndexTypeSizeInBits(V->getType()));
752 
754  while (true) {
755  if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
756  // We accept only inbouds GEPs here to exclude the possibility of
757  // overflow.
758  if (!GEP->isInBounds())
759  break;
760  if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
761  GEP->getType() == V->getType()) {
762  V = GEP->getOperand(0);
763  Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
765  Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
766  continue;
767  }
768  break;
769  }
770  if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
771  if (!CI->isNoopCast(DL))
772  break;
773  V = CI->getOperand(0);
774  continue;
775  }
776  if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
777  if (!CI->isNoopCast(DL))
778  break;
779  V = CI->getOperand(0);
780  continue;
781  }
782  break;
783  }
784  return {V, Index};
785 }
786 
787 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
788 /// We can look through PHIs, GEPs and casts in order to determine a common base
789 /// between GEPLHS and RHS.
792  const DataLayout &DL) {
793  // FIXME: Support vector of pointers.
794  if (GEPLHS->getType()->isVectorTy())
795  return nullptr;
796 
797  if (!GEPLHS->hasAllConstantIndices())
798  return nullptr;
799 
800  // Make sure the pointers have the same type.
801  if (GEPLHS->getType() != RHS->getType())
802  return nullptr;
803 
804  Value *PtrBase, *Index;
805  std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
806 
807  // The set of nodes that will take part in this transformation.
808  SetVector<Value *> Nodes;
809 
810  if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
811  return nullptr;
812 
813  // We know we can re-write this as
814  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
815  // Since we've only looked through inbouds GEPs we know that we
816  // can't have overflow on either side. We can therefore re-write
817  // this as:
818  // OFFSET1 cmp OFFSET2
819  Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
820 
821  // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
822  // GEP having PtrBase as the pointer base, and has returned in NewRHS the
823  // offset. Since Index is the offset of LHS to the base pointer, we will now
824  // compare the offsets instead of comparing the pointers.
825  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
826 }
827 
828 /// Fold comparisons between a GEP instruction and something else. At this point
829 /// we know that the GEP is on the LHS of the comparison.
832  Instruction &I) {
833  // Don't transform signed compares of GEPs into index compares. Even if the
834  // GEP is inbounds, the final add of the base pointer can have signed overflow
835  // and would change the result of the icmp.
836  // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
837  // the maximum signed value for the pointer type.
839  return nullptr;
840 
841  // Look through bitcasts and addrspacecasts. We do not however want to remove
842  // 0 GEPs.
843  if (!isa<GetElementPtrInst>(RHS))
844  RHS = RHS->stripPointerCasts();
845 
846  Value *PtrBase = GEPLHS->getOperand(0);
847  // FIXME: Support vector pointer GEPs.
848  if (PtrBase == RHS && GEPLHS->isInBounds() &&
849  !GEPLHS->getType()->isVectorTy()) {
850  // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
851  // This transformation (ignoring the base and scales) is valid because we
852  // know pointers can't overflow since the gep is inbounds. See if we can
853  // output an optimized form.
854  Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
855 
856  // If not, synthesize the offset the hard way.
857  if (!Offset)
858  Offset = EmitGEPOffset(GEPLHS);
860  Constant::getNullValue(Offset->getType()));
861  }
862 
863  if (GEPLHS->isInBounds() && ICmpInst::isEquality(Cond) &&
864  isa<Constant>(RHS) && cast<Constant>(RHS)->isNullValue() &&
865  !NullPointerIsDefined(I.getFunction(),
866  RHS->getType()->getPointerAddressSpace())) {
867  // For most address spaces, an allocation can't be placed at null, but null
868  // itself is treated as a 0 size allocation in the in bounds rules. Thus,
869  // the only valid inbounds address derived from null, is null itself.
870  // Thus, we have four cases to consider:
871  // 1) Base == nullptr, Offset == 0 -> inbounds, null
872  // 2) Base == nullptr, Offset != 0 -> poison as the result is out of bounds
873  // 3) Base != nullptr, Offset == (-base) -> poison (crossing allocations)
874  // 4) Base != nullptr, Offset != (-base) -> nonnull (and possibly poison)
875  //
876  // (Note if we're indexing a type of size 0, that simply collapses into one
877  // of the buckets above.)
878  //
879  // In general, we're allowed to make values less poison (i.e. remove
880  // sources of full UB), so in this case, we just select between the two
881  // non-poison cases (1 and 4 above).
882  //
883  // For vectors, we apply the same reasoning on a per-lane basis.
884  auto *Base = GEPLHS->getPointerOperand();
885  if (GEPLHS->getType()->isVectorTy() && Base->getType()->isPointerTy()) {
886  auto EC = cast<VectorType>(GEPLHS->getType())->getElementCount();
887  Base = Builder.CreateVectorSplat(EC, Base);
888  }
889  return new ICmpInst(Cond, Base,
891  cast<Constant>(RHS), Base->getType()));
892  } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
893  // If the base pointers are different, but the indices are the same, just
894  // compare the base pointer.
895  if (PtrBase != GEPRHS->getOperand(0)) {
896  bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
897  IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
898  GEPRHS->getOperand(0)->getType();
899  if (IndicesTheSame)
900  for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
901  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
902  IndicesTheSame = false;
903  break;
904  }
905 
906  // If all indices are the same, just compare the base pointers.
907  Type *BaseType = GEPLHS->getOperand(0)->getType();
908  if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
909  return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
910 
911  // If we're comparing GEPs with two base pointers that only differ in type
912  // and both GEPs have only constant indices or just one use, then fold
913  // the compare with the adjusted indices.
914  // FIXME: Support vector of pointers.
915  if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
916  (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
917  (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
918  PtrBase->stripPointerCasts() ==
919  GEPRHS->getOperand(0)->stripPointerCasts() &&
920  !GEPLHS->getType()->isVectorTy()) {
921  Value *LOffset = EmitGEPOffset(GEPLHS);
922  Value *ROffset = EmitGEPOffset(GEPRHS);
923 
924  // If we looked through an addrspacecast between different sized address
925  // spaces, the LHS and RHS pointers are different sized
926  // integers. Truncate to the smaller one.
927  Type *LHSIndexTy = LOffset->getType();
928  Type *RHSIndexTy = ROffset->getType();
929  if (LHSIndexTy != RHSIndexTy) {
930  if (LHSIndexTy->getPrimitiveSizeInBits().getFixedSize() <
931  RHSIndexTy->getPrimitiveSizeInBits().getFixedSize()) {
932  ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
933  } else
934  LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
935  }
936 
937  Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
938  LOffset, ROffset);
939  return replaceInstUsesWith(I, Cmp);
940  }
941 
942  // Otherwise, the base pointers are different and the indices are
943  // different. Try convert this to an indexed compare by looking through
944  // PHIs/casts.
945  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
946  }
947 
948  // If one of the GEPs has all zero indices, recurse.
949  // FIXME: Handle vector of pointers.
950  if (!GEPLHS->getType()->isVectorTy() && GEPLHS->hasAllZeroIndices())
951  return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
953 
954  // If the other GEP has all zero indices, recurse.
955  // FIXME: Handle vector of pointers.
956  if (!GEPRHS->getType()->isVectorTy() && GEPRHS->hasAllZeroIndices())
957  return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
958 
959  bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
960  if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
961  // If the GEPs only differ by one index, compare it.
962  unsigned NumDifferences = 0; // Keep track of # differences.
963  unsigned DiffOperand = 0; // The operand that differs.
964  for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
965  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
966  Type *LHSType = GEPLHS->getOperand(i)->getType();
967  Type *RHSType = GEPRHS->getOperand(i)->getType();
968  // FIXME: Better support for vector of pointers.
969  if (LHSType->getPrimitiveSizeInBits() !=
970  RHSType->getPrimitiveSizeInBits() ||
971  (GEPLHS->getType()->isVectorTy() &&
972  (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) {
973  // Irreconcilable differences.
974  NumDifferences = 2;
975  break;
976  }
977 
978  if (NumDifferences++) break;
979  DiffOperand = i;
980  }
981 
982  if (NumDifferences == 0) // SAME GEP?
983  return replaceInstUsesWith(I, // No comparison is needed here.
985 
986  else if (NumDifferences == 1 && GEPsInBounds) {
987  Value *LHSV = GEPLHS->getOperand(DiffOperand);
988  Value *RHSV = GEPRHS->getOperand(DiffOperand);
989  // Make sure we do a signed comparison here.
990  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
991  }
992  }
993 
994  // Only lower this if the icmp is the only user of the GEP or if we expect
995  // the result to fold to a constant!
996  if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
997  (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
998  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
999  Value *L = EmitGEPOffset(GEPLHS);
1000  Value *R = EmitGEPOffset(GEPRHS);
1001  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
1002  }
1003  }
1004 
1005  // Try convert this to an indexed compare by looking through PHIs/casts as a
1006  // last resort.
1007  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1008 }
1009 
1011  const AllocaInst *Alloca,
1012  const Value *Other) {
1013  assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1014 
1015  // It would be tempting to fold away comparisons between allocas and any
1016  // pointer not based on that alloca (e.g. an argument). However, even
1017  // though such pointers cannot alias, they can still compare equal.
1018  //
1019  // But LLVM doesn't specify where allocas get their memory, so if the alloca
1020  // doesn't escape we can argue that it's impossible to guess its value, and we
1021  // can therefore act as if any such guesses are wrong.
1022  //
1023  // The code below checks that the alloca doesn't escape, and that it's only
1024  // used in a comparison once (the current instruction). The
1025  // single-comparison-use condition ensures that we're trivially folding all
1026  // comparisons against the alloca consistently, and avoids the risk of
1027  // erroneously folding a comparison of the pointer with itself.
1028 
1029  unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1030 
1032  for (const Use &U : Alloca->uses()) {
1033  if (Worklist.size() >= MaxIter)
1034  return nullptr;
1035  Worklist.push_back(&U);
1036  }
1037 
1038  unsigned NumCmps = 0;
1039  while (!Worklist.empty()) {
1040  assert(Worklist.size() <= MaxIter);
1041  const Use *U = Worklist.pop_back_val();
1042  const Value *V = U->getUser();
1043  --MaxIter;
1044 
1045  if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1046  isa<SelectInst>(V)) {
1047  // Track the uses.
1048  } else if (isa<LoadInst>(V)) {
1049  // Loading from the pointer doesn't escape it.
1050  continue;
1051  } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1052  // Storing *to* the pointer is fine, but storing the pointer escapes it.
1053  if (SI->getValueOperand() == U->get())
1054  return nullptr;
1055  continue;
1056  } else if (isa<ICmpInst>(V)) {
1057  if (NumCmps++)
1058  return nullptr; // Found more than one cmp.
1059  continue;
1060  } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1061  switch (Intrin->getIntrinsicID()) {
1062  // These intrinsics don't escape or compare the pointer. Memset is safe
1063  // because we don't allow ptrtoint. Memcpy and memmove are safe because
1064  // we don't allow stores, so src cannot point to V.
1065  case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1066  case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1067  continue;
1068  default:
1069  return nullptr;
1070  }
1071  } else {
1072  return nullptr;
1073  }
1074  for (const Use &U : V->uses()) {
1075  if (Worklist.size() >= MaxIter)
1076  return nullptr;
1077  Worklist.push_back(&U);
1078  }
1079  }
1080 
1081  Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1082  return replaceInstUsesWith(
1083  ICI,
1085 }
1086 
1087 /// Fold "icmp pred (X+C), X".
1089  ICmpInst::Predicate Pred) {
1090  // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1091  // so the values can never be equal. Similarly for all other "or equals"
1092  // operators.
1093  assert(!!C && "C should not be zero!");
1094 
1095  // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1096  // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1097  // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1098  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1099  Constant *R = ConstantInt::get(X->getType(),
1100  APInt::getMaxValue(C.getBitWidth()) - C);
1101  return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1102  }
1103 
1104  // (X+1) >u X --> X <u (0-1) --> X != 255
1105  // (X+2) >u X --> X <u (0-2) --> X <u 254
1106  // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1107  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1108  return new ICmpInst(ICmpInst::ICMP_ULT, X,
1109  ConstantInt::get(X->getType(), -C));
1110 
1111  APInt SMax = APInt::getSignedMaxValue(C.getBitWidth());
1112 
1113  // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1114  // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1115  // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1116  // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1117  // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1118  // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1119  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1120  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1121  ConstantInt::get(X->getType(), SMax - C));
1122 
1123  // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1124  // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1125  // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1126  // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1127  // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1128  // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1129 
1130  assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1131  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1132  ConstantInt::get(X->getType(), SMax - (C - 1)));
1133 }
1134 
1135 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1136 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1137 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1139  const APInt &AP1,
1140  const APInt &AP2) {
1141  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1142 
1143  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1144  if (I.getPredicate() == I.ICMP_NE)
1145  Pred = CmpInst::getInversePredicate(Pred);
1146  return new ICmpInst(Pred, LHS, RHS);
1147  };
1148 
1149  // Don't bother doing any work for cases which InstSimplify handles.
1150  if (AP2.isNullValue())
1151  return nullptr;
1152 
1153  bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1154  if (IsAShr) {
1155  if (AP2.isAllOnesValue())
1156  return nullptr;
1157  if (AP2.isNegative() != AP1.isNegative())
1158  return nullptr;
1159  if (AP2.sgt(AP1))
1160  return nullptr;
1161  }
1162 
1163  if (!AP1)
1164  // 'A' must be large enough to shift out the highest set bit.
1165  return getICmp(I.ICMP_UGT, A,
1166  ConstantInt::get(A->getType(), AP2.logBase2()));
1167 
1168  if (AP1 == AP2)
1169  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1170 
1171  int Shift;
1172  if (IsAShr && AP1.isNegative())
1173  Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1174  else
1175  Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1176 
1177  if (Shift > 0) {
1178  if (IsAShr && AP1 == AP2.ashr(Shift)) {
1179  // There are multiple solutions if we are comparing against -1 and the LHS
1180  // of the ashr is not a power of two.
1181  if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1182  return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1183  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1184  } else if (AP1 == AP2.lshr(Shift)) {
1185  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1186  }
1187  }
1188 
1189  // Shifting const2 will never be equal to const1.
1190  // FIXME: This should always be handled by InstSimplify?
1191  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1192  return replaceInstUsesWith(I, TorF);
1193 }
1194 
1195 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1196 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1198  const APInt &AP1,
1199  const APInt &AP2) {
1200  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1201 
1202  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1203  if (I.getPredicate() == I.ICMP_NE)
1204  Pred = CmpInst::getInversePredicate(Pred);
1205  return new ICmpInst(Pred, LHS, RHS);
1206  };
1207 
1208  // Don't bother doing any work for cases which InstSimplify handles.
1209  if (AP2.isNullValue())
1210  return nullptr;
1211 
1212  unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1213 
1214  if (!AP1 && AP2TrailingZeros != 0)
1215  return getICmp(
1216  I.ICMP_UGE, A,
1217  ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1218 
1219  if (AP1 == AP2)
1220  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1221 
1222  // Get the distance between the lowest bits that are set.
1223  int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1224 
1225  if (Shift > 0 && AP2.shl(Shift) == AP1)
1226  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1227 
1228  // Shifting const2 will never be equal to const1.
1229  // FIXME: This should always be handled by InstSimplify?
1230  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1231  return replaceInstUsesWith(I, TorF);
1232 }
1233 
1234 /// The caller has matched a pattern of the form:
1235 /// I = icmp ugt (add (add A, B), CI2), CI1
1236 /// If this is of the form:
1237 /// sum = a + b
1238 /// if (sum+128 >u 255)
1239 /// Then replace it with llvm.sadd.with.overflow.i8.
1240 ///
1242  ConstantInt *CI2, ConstantInt *CI1,
1243  InstCombinerImpl &IC) {
1244  // The transformation we're trying to do here is to transform this into an
1245  // llvm.sadd.with.overflow. To do this, we have to replace the original add
1246  // with a narrower add, and discard the add-with-constant that is part of the
1247  // range check (if we can't eliminate it, this isn't profitable).
1248 
1249  // In order to eliminate the add-with-constant, the compare can be its only
1250  // use.
1251  Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1252  if (!AddWithCst->hasOneUse())
1253  return nullptr;
1254 
1255  // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1256  if (!CI2->getValue().isPowerOf2())
1257  return nullptr;
1258  unsigned NewWidth = CI2->getValue().countTrailingZeros();
1259  if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1260  return nullptr;
1261 
1262  // The width of the new add formed is 1 more than the bias.
1263  ++NewWidth;
1264 
1265  // Check to see that CI1 is an all-ones value with NewWidth bits.
1266  if (CI1->getBitWidth() == NewWidth ||
1267  CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1268  return nullptr;
1269 
1270  // This is only really a signed overflow check if the inputs have been
1271  // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1272  // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1273  unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1274  if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1275  IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1276  return nullptr;
1277 
1278  // In order to replace the original add with a narrower
1279  // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1280  // and truncates that discard the high bits of the add. Verify that this is
1281  // the case.
1282  Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1283  for (User *U : OrigAdd->users()) {
1284  if (U == AddWithCst)
1285  continue;
1286 
1287  // Only accept truncates for now. We would really like a nice recursive
1288  // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1289  // chain to see which bits of a value are actually demanded. If the
1290  // original add had another add which was then immediately truncated, we
1291  // could still do the transformation.
1292  TruncInst *TI = dyn_cast<TruncInst>(U);
1293  if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1294  return nullptr;
1295  }
1296 
1297  // If the pattern matches, truncate the inputs to the narrower type and
1298  // use the sadd_with_overflow intrinsic to efficiently compute both the
1299  // result and the overflow bit.
1300  Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1302  I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1303 
1305 
1306  // Put the new code above the original add, in case there are any uses of the
1307  // add between the add and the compare.
1308  Builder.SetInsertPoint(OrigAdd);
1309 
1310  Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1311  Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1312  CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1313  Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1314  Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1315 
1316  // The inner add was the result of the narrow add, zero extended to the
1317  // wider type. Replace it with the result computed by the intrinsic.
1318  IC.replaceInstUsesWith(*OrigAdd, ZExt);
1319  IC.eraseInstFromFunction(*OrigAdd);
1320 
1321  // The original icmp gets replaced with the overflow value.
1322  return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1323 }
1324 
1325 /// If we have:
1326 /// icmp eq/ne (urem/srem %x, %y), 0
1327 /// iff %y is a power-of-two, we can replace this with a bit test:
1328 /// icmp eq/ne (and %x, (add %y, -1)), 0
1330  // This fold is only valid for equality predicates.
1331  if (!I.isEquality())
1332  return nullptr;
1333  ICmpInst::Predicate Pred;
1334  Value *X, *Y, *Zero;
1335  if (!match(&I, m_ICmp(Pred, m_OneUse(m_IRem(m_Value(X), m_Value(Y))),
1336  m_CombineAnd(m_Zero(), m_Value(Zero)))))
1337  return nullptr;
1338  if (!isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, 0, &I))
1339  return nullptr;
1340  // This may increase instruction count, we don't enforce that Y is a constant.
1341  Value *Mask = Builder.CreateAdd(Y, Constant::getAllOnesValue(Y->getType()));
1342  Value *Masked = Builder.CreateAnd(X, Mask);
1343  return ICmpInst::Create(Instruction::ICmp, Pred, Masked, Zero);
1344 }
1345 
1346 /// Fold equality-comparison between zero and any (maybe truncated) right-shift
1347 /// by one-less-than-bitwidth into a sign test on the original value.
1349  Instruction *Val;
1350  ICmpInst::Predicate Pred;
1351  if (!I.isEquality() || !match(&I, m_ICmp(Pred, m_Instruction(Val), m_Zero())))
1352  return nullptr;
1353 
1354  Value *X;
1355  Type *XTy;
1356 
1357  Constant *C;
1358  if (match(Val, m_TruncOrSelf(m_Shr(m_Value(X), m_Constant(C))))) {
1359  XTy = X->getType();
1360  unsigned XBitWidth = XTy->getScalarSizeInBits();
1361  if (!match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1362  APInt(XBitWidth, XBitWidth - 1))))
1363  return nullptr;
1364  } else if (isa<BinaryOperator>(Val) &&
1365  (X = reassociateShiftAmtsOfTwoSameDirectionShifts(
1366  cast<BinaryOperator>(Val), SQ.getWithInstruction(Val),
1367  /*AnalyzeForSignBitExtraction=*/true))) {
1368  XTy = X->getType();
1369  } else
1370  return nullptr;
1371 
1372  return ICmpInst::Create(Instruction::ICmp,
1376 }
1377 
1378 // Handle icmp pred X, 0
1380  CmpInst::Predicate Pred = Cmp.getPredicate();
1381  if (!match(Cmp.getOperand(1), m_Zero()))
1382  return nullptr;
1383 
1384  // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1385  if (Pred == ICmpInst::ICMP_SGT) {
1386  Value *A, *B;
1387  SelectPatternResult SPR = matchSelectPattern(Cmp.getOperand(0), A, B);
1388  if (SPR.Flavor == SPF_SMIN) {
1389  if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1390  return new ICmpInst(Pred, B, Cmp.getOperand(1));
1391  if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1392  return new ICmpInst(Pred, A, Cmp.getOperand(1));
1393  }
1394  }
1395 
1396  if (Instruction *New = foldIRemByPowerOfTwoToBitTest(Cmp))
1397  return New;
1398 
1399  // Given:
1400  // icmp eq/ne (urem %x, %y), 0
1401  // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
1402  // icmp eq/ne %x, 0
1403  Value *X, *Y;
1404  if (match(Cmp.getOperand(0), m_URem(m_Value(X), m_Value(Y))) &&
1405  ICmpInst::isEquality(Pred)) {
1406  KnownBits XKnown = computeKnownBits(X, 0, &Cmp);
1407  KnownBits YKnown = computeKnownBits(Y, 0, &Cmp);
1408  if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
1409  return new ICmpInst(Pred, X, Cmp.getOperand(1));
1410  }
1411 
1412  return nullptr;
1413 }
1414 
1415 /// Fold icmp Pred X, C.
1416 /// TODO: This code structure does not make sense. The saturating add fold
1417 /// should be moved to some other helper and extended as noted below (it is also
1418 /// possible that code has been made unnecessary - do we canonicalize IR to
1419 /// overflow/saturating intrinsics or not?).
1421  // Match the following pattern, which is a common idiom when writing
1422  // overflow-safe integer arithmetic functions. The source performs an addition
1423  // in wider type and explicitly checks for overflow using comparisons against
1424  // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1425  //
1426  // TODO: This could probably be generalized to handle other overflow-safe
1427  // operations if we worked out the formulas to compute the appropriate magic
1428  // constants.
1429  //
1430  // sum = a + b
1431  // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1432  CmpInst::Predicate Pred = Cmp.getPredicate();
1433  Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1434  Value *A, *B;
1435  ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1436  if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1437  match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1438  if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1439  return Res;
1440 
1441  // icmp(phi(C1, C2, ...), C) -> phi(icmp(C1, C), icmp(C2, C), ...).
1442  Constant *C = dyn_cast<Constant>(Op1);
1443  if (!C || C->canTrap())
1444  return nullptr;
1445 
1446  if (auto *Phi = dyn_cast<PHINode>(Op0))
1447  if (all_of(Phi->operands(), [](Value *V) { return isa<Constant>(V); })) {
1448  Type *Ty = Cmp.getType();
1449  Builder.SetInsertPoint(Phi);
1450  PHINode *NewPhi =
1451  Builder.CreatePHI(Ty, Phi->getNumOperands());
1452  for (BasicBlock *Predecessor : predecessors(Phi->getParent())) {
1453  auto *Input =
1454  cast<Constant>(Phi->getIncomingValueForBlock(Predecessor));
1455  auto *BoolInput = ConstantExpr::getCompare(Pred, Input, C);
1456  NewPhi->addIncoming(BoolInput, Predecessor);
1457  }
1458  NewPhi->takeName(&Cmp);
1459  return replaceInstUsesWith(Cmp, NewPhi);
1460  }
1461 
1462  return nullptr;
1463 }
1464 
1465 /// Canonicalize icmp instructions based on dominating conditions.
1467  // This is a cheap/incomplete check for dominance - just match a single
1468  // predecessor with a conditional branch.
1469  BasicBlock *CmpBB = Cmp.getParent();
1470  BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1471  if (!DomBB)
1472  return nullptr;
1473 
1474  Value *DomCond;
1475  BasicBlock *TrueBB, *FalseBB;
1476  if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1477  return nullptr;
1478 
1479  assert((TrueBB == CmpBB || FalseBB == CmpBB) &&
1480  "Predecessor block does not point to successor?");
1481 
1482  // The branch should get simplified. Don't bother simplifying this condition.
1483  if (TrueBB == FalseBB)
1484  return nullptr;
1485 
1486  // Try to simplify this compare to T/F based on the dominating condition.
1487  Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1488  if (Imp)
1489  return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1490 
1491  CmpInst::Predicate Pred = Cmp.getPredicate();
1492  Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1493  ICmpInst::Predicate DomPred;
1494  const APInt *C, *DomC;
1495  if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1496  match(Y, m_APInt(C))) {
1497  // We have 2 compares of a variable with constants. Calculate the constant
1498  // ranges of those compares to see if we can transform the 2nd compare:
1499  // DomBB:
1500  // DomCond = icmp DomPred X, DomC
1501  // br DomCond, CmpBB, FalseBB
1502  // CmpBB:
1503  // Cmp = icmp Pred X, C
1505  ConstantRange DominatingCR =
1506  (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1508  CmpInst::getInversePredicate(DomPred), *DomC);
1509  ConstantRange Intersection = DominatingCR.intersectWith(CR);
1510  ConstantRange Difference = DominatingCR.difference(CR);
1511  if (Intersection.isEmptySet())
1512  return replaceInstUsesWith(Cmp, Builder.getFalse());
1513  if (Difference.isEmptySet())
1514  return replaceInstUsesWith(Cmp, Builder.getTrue());
1515 
1516  // Canonicalizing a sign bit comparison that gets used in a branch,
1517  // pessimizes codegen by generating branch on zero instruction instead
1518  // of a test and branch. So we avoid canonicalizing in such situations
1519  // because test and branch instruction has better branch displacement
1520  // than compare and branch instruction.
1521  bool UnusedBit;
1522  bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1523  if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1524  return nullptr;
1525 
1526  // Avoid an infinite loop with min/max canonicalization.
1527  // TODO: This will be unnecessary if we canonicalize to min/max intrinsics.
1528  if (Cmp.hasOneUse() &&
1529  match(Cmp.user_back(), m_MaxOrMin(m_Value(), m_Value())))
1530  return nullptr;
1531 
1532  if (const APInt *EqC = Intersection.getSingleElement())
1533  return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1534  if (const APInt *NeC = Difference.getSingleElement())
1535  return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1536  }
1537 
1538  return nullptr;
1539 }
1540 
1541 /// Fold icmp (trunc X, Y), C.
1543  TruncInst *Trunc,
1544  const APInt &C) {
1545  ICmpInst::Predicate Pred = Cmp.getPredicate();
1546  Value *X = Trunc->getOperand(0);
1547  if (C.isOneValue() && C.getBitWidth() > 1) {
1548  // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1549  Value *V = nullptr;
1550  if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1551  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1552  ConstantInt::get(V->getType(), 1));
1553  }
1554 
1555  unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1556  SrcBits = X->getType()->getScalarSizeInBits();
1557  if (Cmp.isEquality() && Trunc->hasOneUse()) {
1558  // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1559  // of the high bits truncated out of x are known.
1560  KnownBits Known = computeKnownBits(X, 0, &Cmp);
1561 
1562  // If all the high bits are known, we can do this xform.
1563  if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1564  // Pull in the high bits from known-ones set.
1565  APInt NewRHS = C.zext(SrcBits);
1566  NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1567  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1568  }
1569  }
1570 
1571  // Look through truncated right-shift of the sign-bit for a sign-bit check:
1572  // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] < 0 --> ShOp < 0
1573  // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] > -1 --> ShOp > -1
1574  Value *ShOp;
1575  const APInt *ShAmtC;
1576  bool TrueIfSigned;
1577  if (isSignBitCheck(Pred, C, TrueIfSigned) &&
1578  match(X, m_Shr(m_Value(ShOp), m_APInt(ShAmtC))) &&
1579  DstBits == SrcBits - ShAmtC->getZExtValue()) {
1580  return TrueIfSigned
1581  ? new ICmpInst(ICmpInst::ICMP_SLT, ShOp,
1582  ConstantInt::getNullValue(X->getType()))
1583  : new ICmpInst(ICmpInst::ICMP_SGT, ShOp,
1584  ConstantInt::getAllOnesValue(X->getType()));
1585  }
1586 
1587  return nullptr;
1588 }
1589 
1590 /// Fold icmp (xor X, Y), C.
1592  BinaryOperator *Xor,
1593  const APInt &C) {
1594  Value *X = Xor->getOperand(0);
1595  Value *Y = Xor->getOperand(1);
1596  const APInt *XorC;
1597  if (!match(Y, m_APInt(XorC)))
1598  return nullptr;
1599 
1600  // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1601  // fold the xor.
1602  ICmpInst::Predicate Pred = Cmp.getPredicate();
1603  bool TrueIfSigned = false;
1604  if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1605 
1606  // If the sign bit of the XorCst is not set, there is no change to
1607  // the operation, just stop using the Xor.
1608  if (!XorC->isNegative())
1609  return replaceOperand(Cmp, 0, X);
1610 
1611  // Emit the opposite comparison.
1612  if (TrueIfSigned)
1613  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1614  ConstantInt::getAllOnesValue(X->getType()));
1615  else
1616  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1617  ConstantInt::getNullValue(X->getType()));
1618  }
1619 
1620  if (Xor->hasOneUse()) {
1621  // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1622  if (!Cmp.isEquality() && XorC->isSignMask()) {
1623  Pred = Cmp.getFlippedSignednessPredicate();
1624  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1625  }
1626 
1627  // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1628  if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1629  Pred = Cmp.getFlippedSignednessPredicate();
1630  Pred = Cmp.getSwappedPredicate(Pred);
1631  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1632  }
1633  }
1634 
1635  // Mask constant magic can eliminate an 'xor' with unsigned compares.
1636  if (Pred == ICmpInst::ICMP_UGT) {
1637  // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1638  if (*XorC == ~C && (C + 1).isPowerOf2())
1639  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1640  // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1641  if (*XorC == C && (C + 1).isPowerOf2())
1642  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1643  }
1644  if (Pred == ICmpInst::ICMP_ULT) {
1645  // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1646  if (*XorC == -C && C.isPowerOf2())
1647  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1648  ConstantInt::get(X->getType(), ~C));
1649  // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1650  if (*XorC == C && (-C).isPowerOf2())
1651  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1652  ConstantInt::get(X->getType(), ~C));
1653  }
1654  return nullptr;
1655 }
1656 
1657 /// Fold icmp (and (sh X, Y), C2), C1.
1659  BinaryOperator *And,
1660  const APInt &C1,
1661  const APInt &C2) {
1662  BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1663  if (!Shift || !Shift->isShift())
1664  return nullptr;
1665 
1666  // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1667  // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1668  // code produced by the clang front-end, for bitfield access.
1669  // This seemingly simple opportunity to fold away a shift turns out to be
1670  // rather complicated. See PR17827 for details.
1671  unsigned ShiftOpcode = Shift->getOpcode();
1672  bool IsShl = ShiftOpcode == Instruction::Shl;
1673  const APInt *C3;
1674  if (match(Shift->getOperand(1), m_APInt(C3))) {
1675  APInt NewAndCst, NewCmpCst;
1676  bool AnyCmpCstBitsShiftedOut;
1677  if (ShiftOpcode == Instruction::Shl) {
1678  // For a left shift, we can fold if the comparison is not signed. We can
1679  // also fold a signed comparison if the mask value and comparison value
1680  // are not negative. These constraints may not be obvious, but we can
1681  // prove that they are correct using an SMT solver.
1682  if (Cmp.isSigned() && (C2.isNegative() || C1.isNegative()))
1683  return nullptr;
1684 
1685  NewCmpCst = C1.lshr(*C3);
1686  NewAndCst = C2.lshr(*C3);
1687  AnyCmpCstBitsShiftedOut = NewCmpCst.shl(*C3) != C1;
1688  } else if (ShiftOpcode == Instruction::LShr) {
1689  // For a logical right shift, we can fold if the comparison is not signed.
1690  // We can also fold a signed comparison if the shifted mask value and the
1691  // shifted comparison value are not negative. These constraints may not be
1692  // obvious, but we can prove that they are correct using an SMT solver.
1693  NewCmpCst = C1.shl(*C3);
1694  NewAndCst = C2.shl(*C3);
1695  AnyCmpCstBitsShiftedOut = NewCmpCst.lshr(*C3) != C1;
1696  if (Cmp.isSigned() && (NewAndCst.isNegative() || NewCmpCst.isNegative()))
1697  return nullptr;
1698  } else {
1699  // For an arithmetic shift, check that both constants don't use (in a
1700  // signed sense) the top bits being shifted out.
1701  assert(ShiftOpcode == Instruction::AShr && "Unknown shift opcode");
1702  NewCmpCst = C1.shl(*C3);
1703  NewAndCst = C2.shl(*C3);
1704  AnyCmpCstBitsShiftedOut = NewCmpCst.ashr(*C3) != C1;
1705  if (NewAndCst.ashr(*C3) != C2)
1706  return nullptr;
1707  }
1708 
1709  if (AnyCmpCstBitsShiftedOut) {
1710  // If we shifted bits out, the fold is not going to work out. As a
1711  // special case, check to see if this means that the result is always
1712  // true or false now.
1713  if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1714  return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1715  if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1716  return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1717  } else {
1718  Value *NewAnd = Builder.CreateAnd(
1719  Shift->getOperand(0), ConstantInt::get(And->getType(), NewAndCst));
1720  return new ICmpInst(Cmp.getPredicate(),
1721  NewAnd, ConstantInt::get(And->getType(), NewCmpCst));
1722  }
1723  }
1724 
1725  // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1726  // preferable because it allows the C2 << Y expression to be hoisted out of a
1727  // loop if Y is invariant and X is not.
1728  if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() &&
1729  !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1730  // Compute C2 << Y.
1731  Value *NewShift =
1732  IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1733  : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1734 
1735  // Compute X & (C2 << Y).
1736  Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1737  return replaceOperand(Cmp, 0, NewAnd);
1738  }
1739 
1740  return nullptr;
1741 }
1742 
1743 /// Fold icmp (and X, C2), C1.
1745  BinaryOperator *And,
1746  const APInt &C1) {
1747  bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1748 
1749  // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1750  // TODO: We canonicalize to the longer form for scalars because we have
1751  // better analysis/folds for icmp, and codegen may be better with icmp.
1752  if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isNullValue() &&
1753  match(And->getOperand(1), m_One()))
1754  return new TruncInst(And->getOperand(0), Cmp.getType());
1755 
1756  const APInt *C2;
1757  Value *X;
1758  if (!match(And, m_And(m_Value(X), m_APInt(C2))))
1759  return nullptr;
1760 
1761  // Don't perform the following transforms if the AND has multiple uses
1762  if (!And->hasOneUse())
1763  return nullptr;
1764 
1765  if (Cmp.isEquality() && C1.isNullValue()) {
1766  // Restrict this fold to single-use 'and' (PR10267).
1767  // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1768  if (C2->isSignMask()) {
1769  Constant *Zero = Constant::getNullValue(X->getType());
1770  auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1771  return new ICmpInst(NewPred, X, Zero);
1772  }
1773 
1774  // Restrict this fold only for single-use 'and' (PR10267).
1775  // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1776  if ((~(*C2) + 1).isPowerOf2()) {
1777  Constant *NegBOC =
1778  ConstantExpr::getNeg(cast<Constant>(And->getOperand(1)));
1779  auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1780  return new ICmpInst(NewPred, X, NegBOC);
1781  }
1782  }
1783 
1784  // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1785  // the input width without changing the value produced, eliminate the cast:
1786  //
1787  // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1788  //
1789  // We can do this transformation if the constants do not have their sign bits
1790  // set or if it is an equality comparison. Extending a relational comparison
1791  // when we're checking the sign bit would not work.
1792  Value *W;
1793  if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1794  (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1795  // TODO: Is this a good transform for vectors? Wider types may reduce
1796  // throughput. Should this transform be limited (even for scalars) by using
1797  // shouldChangeType()?
1798  if (!Cmp.getType()->isVectorTy()) {
1799  Type *WideType = W->getType();
1800  unsigned WideScalarBits = WideType->getScalarSizeInBits();
1801  Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1802  Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1803  Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1804  return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1805  }
1806  }
1807 
1808  if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1809  return I;
1810 
1811  // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1812  // (icmp pred (and A, (or (shl 1, B), 1), 0))
1813  //
1814  // iff pred isn't signed
1815  if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() &&
1816  match(And->getOperand(1), m_One())) {
1817  Constant *One = cast<Constant>(And->getOperand(1));
1818  Value *Or = And->getOperand(0);
1819  Value *A, *B, *LShr;
1820  if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1821  match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1822  unsigned UsesRemoved = 0;
1823  if (And->hasOneUse())
1824  ++UsesRemoved;
1825  if (Or->hasOneUse())
1826  ++UsesRemoved;
1827  if (LShr->hasOneUse())
1828  ++UsesRemoved;
1829 
1830  // Compute A & ((1 << B) | 1)
1831  Value *NewOr = nullptr;
1832  if (auto *C = dyn_cast<Constant>(B)) {
1833  if (UsesRemoved >= 1)
1834  NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1835  } else {
1836  if (UsesRemoved >= 3)
1837  NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1838  /*HasNUW=*/true),
1839  One, Or->getName());
1840  }
1841  if (NewOr) {
1842  Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1843  return replaceOperand(Cmp, 0, NewAnd);
1844  }
1845  }
1846  }
1847 
1848  return nullptr;
1849 }
1850 
1851 /// Fold icmp (and X, Y), C.
1853  BinaryOperator *And,
1854  const APInt &C) {
1855  if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1856  return I;
1857 
1858  const ICmpInst::Predicate Pred = Cmp.getPredicate();
1859  bool TrueIfNeg;
1860  if (isSignBitCheck(Pred, C, TrueIfNeg)) {
1861  // ((X - 1) & ~X) < 0 --> X == 0
1862  // ((X - 1) & ~X) >= 0 --> X != 0
1863  Value *X;
1864  if (match(And->getOperand(0), m_Add(m_Value(X), m_AllOnes())) &&
1865  match(And->getOperand(1), m_Not(m_Specific(X)))) {
1866  auto NewPred = TrueIfNeg ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE;
1867  return new ICmpInst(NewPred, X, ConstantInt::getNullValue(X->getType()));
1868  }
1869  }
1870 
1871  // TODO: These all require that Y is constant too, so refactor with the above.
1872 
1873  // Try to optimize things like "A[i] & 42 == 0" to index computations.
1874  Value *X = And->getOperand(0);
1875  Value *Y = And->getOperand(1);
1876  if (auto *LI = dyn_cast<LoadInst>(X))
1877  if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1878  if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1879  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1880  !LI->isVolatile() && isa<ConstantInt>(Y)) {
1881  ConstantInt *C2 = cast<ConstantInt>(Y);
1882  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1883  return Res;
1884  }
1885 
1886  if (!Cmp.isEquality())
1887  return nullptr;
1888 
1889  // X & -C == -C -> X > u ~C
1890  // X & -C != -C -> X <= u ~C
1891  // iff C is a power of 2
1892  if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1893  auto NewPred =
1895  return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1896  }
1897 
1898  // (X & C2) == 0 -> (trunc X) >= 0
1899  // (X & C2) != 0 -> (trunc X) < 0
1900  // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1901  const APInt *C2;
1902  if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) {
1903  int32_t ExactLogBase2 = C2->exactLogBase2();
1904  if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1905  Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1906  if (auto *AndVTy = dyn_cast<VectorType>(And->getType()))
1907  NTy = VectorType::get(NTy, AndVTy->getElementCount());
1908  Value *Trunc = Builder.CreateTrunc(X, NTy);
1909  auto NewPred =
1911  return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1912  }
1913  }
1914 
1915  return nullptr;
1916 }
1917 
1918 /// Fold icmp (or X, Y), C.
1920  BinaryOperator *Or,
1921  const APInt &C) {
1922  ICmpInst::Predicate Pred = Cmp.getPredicate();
1923  if (C.isOneValue()) {
1924  // icmp slt signum(V) 1 --> icmp slt V, 1
1925  Value *V = nullptr;
1926  if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1927  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1928  ConstantInt::get(V->getType(), 1));
1929  }
1930 
1931  Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
1932  const APInt *MaskC;
1933  if (match(OrOp1, m_APInt(MaskC)) && Cmp.isEquality()) {
1934  if (*MaskC == C && (C + 1).isPowerOf2()) {
1935  // X | C == C --> X <=u C
1936  // X | C != C --> X >u C
1937  // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1939  return new ICmpInst(Pred, OrOp0, OrOp1);
1940  }
1941 
1942  // More general: canonicalize 'equality with set bits mask' to
1943  // 'equality with clear bits mask'.
1944  // (X | MaskC) == C --> (X & ~MaskC) == C ^ MaskC
1945  // (X | MaskC) != C --> (X & ~MaskC) != C ^ MaskC
1946  if (Or->hasOneUse()) {
1947  Value *And = Builder.CreateAnd(OrOp0, ~(*MaskC));
1948  Constant *NewC = ConstantInt::get(Or->getType(), C ^ (*MaskC));
1949  return new ICmpInst(Pred, And, NewC);
1950  }
1951  }
1952 
1953  if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse())
1954  return nullptr;
1955 
1956  Value *P, *Q;
1957  if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1958  // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1959  // -> and (icmp eq P, null), (icmp eq Q, null).
1960  Value *CmpP =
1961  Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1962  Value *CmpQ =
1963  Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1964  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1965  return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1966  }
1967 
1968  // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1969  // a shorter form that has more potential to be folded even further.
1970  Value *X1, *X2, *X3, *X4;
1971  if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1972  match(OrOp1, m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1973  // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1974  // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1975  Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1976  Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1977  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1978  return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1979  }
1980 
1981  return nullptr;
1982 }
1983 
1984 /// Fold icmp (mul X, Y), C.
1986  BinaryOperator *Mul,
1987  const APInt &C) {
1988  const APInt *MulC;
1989  if (!match(Mul->getOperand(1), m_APInt(MulC)))
1990  return nullptr;
1991 
1992  // If this is a test of the sign bit and the multiply is sign-preserving with
1993  // a constant operand, use the multiply LHS operand instead.
1994  ICmpInst::Predicate Pred = Cmp.getPredicate();
1995  if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1996  if (MulC->isNegative())
1997  Pred = ICmpInst::getSwappedPredicate(Pred);
1998  return new ICmpInst(Pred, Mul->getOperand(0),
1999  Constant::getNullValue(Mul->getType()));
2000  }
2001 
2002  // If the multiply does not wrap, try to divide the compare constant by the
2003  // multiplication factor.
2004  if (Cmp.isEquality() && !MulC->isNullValue()) {
2005  // (mul nsw X, MulC) == C --> X == C /s MulC
2006  if (Mul->hasNoSignedWrap() && C.srem(*MulC).isNullValue()) {
2007  Constant *NewC = ConstantInt::get(Mul->getType(), C.sdiv(*MulC));
2008  return new ICmpInst(Pred, Mul->getOperand(0), NewC);
2009  }
2010  // (mul nuw X, MulC) == C --> X == C /u MulC
2011  if (Mul->hasNoUnsignedWrap() && C.urem(*MulC).isNullValue()) {
2012  Constant *NewC = ConstantInt::get(Mul->getType(), C.udiv(*MulC));
2013  return new ICmpInst(Pred, Mul->getOperand(0), NewC);
2014  }
2015  }
2016 
2017  return nullptr;
2018 }
2019 
2020 /// Fold icmp (shl 1, Y), C.
2022  const APInt &C) {
2023  Value *Y;
2024  if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
2025  return nullptr;
2026 
2027  Type *ShiftType = Shl->getType();
2028  unsigned TypeBits = C.getBitWidth();
2029  bool CIsPowerOf2 = C.isPowerOf2();
2030  ICmpInst::Predicate Pred = Cmp.getPredicate();
2031  if (Cmp.isUnsigned()) {
2032  // (1 << Y) pred C -> Y pred Log2(C)
2033  if (!CIsPowerOf2) {
2034  // (1 << Y) < 30 -> Y <= 4
2035  // (1 << Y) <= 30 -> Y <= 4
2036  // (1 << Y) >= 30 -> Y > 4
2037  // (1 << Y) > 30 -> Y > 4
2038  if (Pred == ICmpInst::ICMP_ULT)
2039  Pred = ICmpInst::ICMP_ULE;
2040  else if (Pred == ICmpInst::ICMP_UGE)
2041  Pred = ICmpInst::ICMP_UGT;
2042  }
2043 
2044  // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
2045  // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
2046  unsigned CLog2 = C.logBase2();
2047  if (CLog2 == TypeBits - 1) {
2048  if (Pred == ICmpInst::ICMP_UGE)
2049  Pred = ICmpInst::ICMP_EQ;
2050  else if (Pred == ICmpInst::ICMP_ULT)
2051  Pred = ICmpInst::ICMP_NE;
2052  }
2053  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
2054  } else if (Cmp.isSigned()) {
2055  Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
2056  if (C.isAllOnesValue()) {
2057  // (1 << Y) <= -1 -> Y == 31
2058  if (Pred == ICmpInst::ICMP_SLE)
2059  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2060 
2061  // (1 << Y) > -1 -> Y != 31
2062  if (Pred == ICmpInst::ICMP_SGT)
2063  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2064  } else if (!C) {
2065  // (1 << Y) < 0 -> Y == 31
2066  // (1 << Y) <= 0 -> Y == 31
2067  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
2068  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2069 
2070  // (1 << Y) >= 0 -> Y != 31
2071  // (1 << Y) > 0 -> Y != 31
2072  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
2073  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2074  }
2075  } else if (Cmp.isEquality() && CIsPowerOf2) {
2076  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
2077  }
2078 
2079  return nullptr;
2080 }
2081 
2082 /// Fold icmp (shl X, Y), C.
2084  BinaryOperator *Shl,
2085  const APInt &C) {
2086  const APInt *ShiftVal;
2087  if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
2088  return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
2089 
2090  const APInt *ShiftAmt;
2091  if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
2092  return foldICmpShlOne(Cmp, Shl, C);
2093 
2094  // Check that the shift amount is in range. If not, don't perform undefined
2095  // shifts. When the shift is visited, it will be simplified.
2096  unsigned TypeBits = C.getBitWidth();
2097  if (ShiftAmt->uge(TypeBits))
2098  return nullptr;
2099 
2100  ICmpInst::Predicate Pred = Cmp.getPredicate();
2101  Value *X = Shl->getOperand(0);
2102  Type *ShType = Shl->getType();
2103 
2104  // NSW guarantees that we are only shifting out sign bits from the high bits,
2105  // so we can ASHR the compare constant without needing a mask and eliminate
2106  // the shift.
2107  if (Shl->hasNoSignedWrap()) {
2108  if (Pred == ICmpInst::ICMP_SGT) {
2109  // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
2110  APInt ShiftedC = C.ashr(*ShiftAmt);
2111  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2112  }
2113  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2114  C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
2115  APInt ShiftedC = C.ashr(*ShiftAmt);
2116  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2117  }
2118  if (Pred == ICmpInst::ICMP_SLT) {
2119  // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
2120  // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2121  // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2122  // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2123  assert(!C.isMinSignedValue() && "Unexpected icmp slt");
2124  APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
2125  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2126  }
2127  // If this is a signed comparison to 0 and the shift is sign preserving,
2128  // use the shift LHS operand instead; isSignTest may change 'Pred', so only
2129  // do that if we're sure to not continue on in this function.
2130  if (isSignTest(Pred, C))
2131  return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
2132  }
2133 
2134  // NUW guarantees that we are only shifting out zero bits from the high bits,
2135  // so we can LSHR the compare constant without needing a mask and eliminate
2136  // the shift.
2137  if (Shl->hasNoUnsignedWrap()) {
2138  if (Pred == ICmpInst::ICMP_UGT) {
2139  // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2140  APInt ShiftedC = C.lshr(*ShiftAmt);
2141  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2142  }
2143  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2144  C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
2145  APInt ShiftedC = C.lshr(*ShiftAmt);
2146  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2147  }
2148  if (Pred == ICmpInst::ICMP_ULT) {
2149  // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2150  // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2151  // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2152  // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2153  assert(C.ugt(0) && "ult 0 should have been eliminated");
2154  APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
2155  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2156  }
2157  }
2158 
2159  if (Cmp.isEquality() && Shl->hasOneUse()) {
2160  // Strength-reduce the shift into an 'and'.
2162  ShType,
2163  APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2164  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2165  Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
2166  return new ICmpInst(Pred, And, LShrC);
2167  }
2168 
2169  // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2170  bool TrueIfSigned = false;
2171  if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
2172  // (X << 31) <s 0 --> (X & 1) != 0
2174  ShType,
2175  APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2176  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2177  return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2178  And, Constant::getNullValue(ShType));
2179  }
2180 
2181  // Simplify 'shl' inequality test into 'and' equality test.
2182  if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2183  // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2184  if ((C + 1).isPowerOf2() &&
2185  (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2186  Value *And = Builder.CreateAnd(X, (~C).lshr(ShiftAmt->getZExtValue()));
2187  return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2189  And, Constant::getNullValue(ShType));
2190  }
2191  // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2192  if (C.isPowerOf2() &&
2193  (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2194  Value *And =
2195  Builder.CreateAnd(X, (~(C - 1)).lshr(ShiftAmt->getZExtValue()));
2196  return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2198  And, Constant::getNullValue(ShType));
2199  }
2200  }
2201 
2202  // Transform (icmp pred iM (shl iM %v, N), C)
2203  // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2204  // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2205  // This enables us to get rid of the shift in favor of a trunc that may be
2206  // free on the target. It has the additional benefit of comparing to a
2207  // smaller constant that may be more target-friendly.
2208  unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2209  if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2210  DL.isLegalInteger(TypeBits - Amt)) {
2211  Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2212  if (auto *ShVTy = dyn_cast<VectorType>(ShType))
2213  TruncTy = VectorType::get(TruncTy, ShVTy->getElementCount());
2214  Constant *NewC =
2215  ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2216  return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2217  }
2218 
2219  return nullptr;
2220 }
2221 
2222 /// Fold icmp ({al}shr X, Y), C.
2224  BinaryOperator *Shr,
2225  const APInt &C) {
2226  // An exact shr only shifts out zero bits, so:
2227  // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2228  Value *X = Shr->getOperand(0);
2229  CmpInst::Predicate Pred = Cmp.getPredicate();
2230  if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2231  C.isNullValue())
2232  return new ICmpInst(Pred, X, Cmp.getOperand(1));
2233 
2234  const APInt *ShiftVal;
2235  if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2236  return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2237 
2238  const APInt *ShiftAmt;
2239  if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2240  return nullptr;
2241 
2242  // Check that the shift amount is in range. If not, don't perform undefined
2243  // shifts. When the shift is visited it will be simplified.
2244  unsigned TypeBits = C.getBitWidth();
2245  unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2246  if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2247  return nullptr;
2248 
2249  bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2250  bool IsExact = Shr->isExact();
2251  Type *ShrTy = Shr->getType();
2252  // TODO: If we could guarantee that InstSimplify would handle all of the
2253  // constant-value-based preconditions in the folds below, then we could assert
2254  // those conditions rather than checking them. This is difficult because of
2255  // undef/poison (PR34838).
2256  if (IsAShr) {
2257  if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2258  // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2259  // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2260  APInt ShiftedC = C.shl(ShAmtVal);
2261  if (ShiftedC.ashr(ShAmtVal) == C)
2262  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2263  }
2264  if (Pred == CmpInst::ICMP_SGT) {
2265  // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2266  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2267  if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2268  (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2269  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2270  }
2271 
2272  // If the compare constant has significant bits above the lowest sign-bit,
2273  // then convert an unsigned cmp to a test of the sign-bit:
2274  // (ashr X, ShiftC) u> C --> X s< 0
2275  // (ashr X, ShiftC) u< C --> X s> -1
2276  if (C.getBitWidth() > 2 && C.getNumSignBits() <= ShAmtVal) {
2277  if (Pred == CmpInst::ICMP_UGT) {
2278  return new ICmpInst(CmpInst::ICMP_SLT, X,
2279  ConstantInt::getNullValue(ShrTy));
2280  }
2281  if (Pred == CmpInst::ICMP_ULT) {
2282  return new ICmpInst(CmpInst::ICMP_SGT, X,
2284  }
2285  }
2286  } else {
2287  if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2288  // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2289  // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2290  APInt ShiftedC = C.shl(ShAmtVal);
2291  if (ShiftedC.lshr(ShAmtVal) == C)
2292  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2293  }
2294  if (Pred == CmpInst::ICMP_UGT) {
2295  // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2296  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2297  if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2298  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2299  }
2300  }
2301 
2302  if (!Cmp.isEquality())
2303  return nullptr;
2304 
2305  // Handle equality comparisons of shift-by-constant.
2306 
2307  // If the comparison constant changes with the shift, the comparison cannot
2308  // succeed (bits of the comparison constant cannot match the shifted value).
2309  // This should be known by InstSimplify and already be folded to true/false.
2310  assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2311  (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2312  "Expected icmp+shr simplify did not occur.");
2313 
2314  // If the bits shifted out are known zero, compare the unshifted value:
2315  // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2316  if (Shr->isExact())
2317  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2318 
2319  if (C.isNullValue()) {
2320  // == 0 is u< 1.
2321  if (Pred == CmpInst::ICMP_EQ)
2322  return new ICmpInst(CmpInst::ICMP_ULT, X,
2323  ConstantInt::get(ShrTy, (C + 1).shl(ShAmtVal)));
2324  else
2325  return new ICmpInst(CmpInst::ICMP_UGT, X,
2326  ConstantInt::get(ShrTy, (C + 1).shl(ShAmtVal) - 1));
2327  }
2328 
2329  if (Shr->hasOneUse()) {
2330  // Canonicalize the shift into an 'and':
2331  // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2332  APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2333  Constant *Mask = ConstantInt::get(ShrTy, Val);
2334  Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2335  return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2336  }
2337 
2338  return nullptr;
2339 }
2340 
2342  BinaryOperator *SRem,
2343  const APInt &C) {
2344  // Match an 'is positive' or 'is negative' comparison of remainder by a
2345  // constant power-of-2 value:
2346  // (X % pow2C) sgt/slt 0
2347  const ICmpInst::Predicate Pred = Cmp.getPredicate();
2348  if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT)
2349  return nullptr;
2350 
2351  // TODO: The one-use check is standard because we do not typically want to
2352  // create longer instruction sequences, but this might be a special-case
2353  // because srem is not good for analysis or codegen.
2354  if (!SRem->hasOneUse())
2355  return nullptr;
2356 
2357  const APInt *DivisorC;
2358  if (!C.isNullValue() || !match(SRem->getOperand(1), m_Power2(DivisorC)))
2359  return nullptr;
2360 
2361  // Mask off the sign bit and the modulo bits (low-bits).
2362  Type *Ty = SRem->getType();
2363  APInt SignMask = APInt::getSignMask(Ty->getScalarSizeInBits());
2364  Constant *MaskC = ConstantInt::get(Ty, SignMask | (*DivisorC - 1));
2365  Value *And = Builder.CreateAnd(SRem->getOperand(0), MaskC);
2366 
2367  // For 'is positive?' check that the sign-bit is clear and at least 1 masked
2368  // bit is set. Example:
2369  // (i8 X % 32) s> 0 --> (X & 159) s> 0
2370  if (Pred == ICmpInst::ICMP_SGT)
2372 
2373  // For 'is negative?' check that the sign-bit is set and at least 1 masked
2374  // bit is set. Example:
2375  // (i16 X % 4) s< 0 --> (X & 32771) u> 32768
2376  return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, SignMask));
2377 }
2378 
2379 /// Fold icmp (udiv X, Y), C.
2381  BinaryOperator *UDiv,
2382  const APInt &C) {
2383  const APInt *C2;
2384  if (!match(UDiv->getOperand(0), m_APInt(C2)))
2385  return nullptr;
2386 
2387  assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2388 
2389  // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2390  Value *Y = UDiv->getOperand(1);
2391  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2392  assert(!C.isMaxValue() &&
2393  "icmp ugt X, UINT_MAX should have been simplified already.");
2394  return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2395  ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2396  }
2397 
2398  // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2399  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2400  assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2401  return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2402  ConstantInt::get(Y->getType(), C2->udiv(C)));
2403  }
2404 
2405  return nullptr;
2406 }
2407 
2408 /// Fold icmp ({su}div X, Y), C.
2410  BinaryOperator *Div,
2411  const APInt &C) {
2412  // Fold: icmp pred ([us]div X, C2), C -> range test
2413  // Fold this div into the comparison, producing a range check.
2414  // Determine, based on the divide type, what the range is being
2415  // checked. If there is an overflow on the low or high side, remember
2416  // it, otherwise compute the range [low, hi) bounding the new value.
2417  // See: InsertRangeTest above for the kinds of replacements possible.
2418  const APInt *C2;
2419  if (!match(Div->getOperand(1), m_APInt(C2)))
2420  return nullptr;
2421 
2422  // FIXME: If the operand types don't match the type of the divide
2423  // then don't attempt this transform. The code below doesn't have the
2424  // logic to deal with a signed divide and an unsigned compare (and
2425  // vice versa). This is because (x /s C2) <s C produces different
2426  // results than (x /s C2) <u C or (x /u C2) <s C or even
2427  // (x /u C2) <u C. Simply casting the operands and result won't
2428  // work. :( The if statement below tests that condition and bails
2429  // if it finds it.
2430  bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2431  if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2432  return nullptr;
2433 
2434  // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2435  // INT_MIN will also fail if the divisor is 1. Although folds of all these
2436  // division-by-constant cases should be present, we can not assert that they
2437  // have happened before we reach this icmp instruction.
2438  if (C2->isNullValue() || C2->isOneValue() ||
2439  (DivIsSigned && C2->isAllOnesValue()))
2440  return nullptr;
2441 
2442  // Compute Prod = C * C2. We are essentially solving an equation of
2443  // form X / C2 = C. We solve for X by multiplying C2 and C.
2444  // By solving for X, we can turn this into a range check instead of computing
2445  // a divide.
2446  APInt Prod = C * *C2;
2447 
2448  // Determine if the product overflows by seeing if the product is not equal to
2449  // the divide. Make sure we do the same kind of divide as in the LHS
2450  // instruction that we're folding.
2451  bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2452 
2453  ICmpInst::Predicate Pred = Cmp.getPredicate();
2454 
2455  // If the division is known to be exact, then there is no remainder from the
2456  // divide, so the covered range size is unit, otherwise it is the divisor.
2457  APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2458 
2459  // Figure out the interval that is being checked. For example, a comparison
2460  // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2461  // Compute this interval based on the constants involved and the signedness of
2462  // the compare/divide. This computes a half-open interval, keeping track of
2463  // whether either value in the interval overflows. After analysis each
2464  // overflow variable is set to 0 if it's corresponding bound variable is valid
2465  // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2466  int LoOverflow = 0, HiOverflow = 0;
2467  APInt LoBound, HiBound;
2468 
2469  if (!DivIsSigned) { // udiv
2470  // e.g. X/5 op 3 --> [15, 20)
2471  LoBound = Prod;
2472  HiOverflow = LoOverflow = ProdOV;
2473  if (!HiOverflow) {
2474  // If this is not an exact divide, then many values in the range collapse
2475  // to the same result value.
2476  HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2477  }
2478  } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2479  if (C.isNullValue()) { // (X / pos) op 0
2480  // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2481  LoBound = -(RangeSize - 1);
2482  HiBound = RangeSize;
2483  } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2484  LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2485  HiOverflow = LoOverflow = ProdOV;
2486  if (!HiOverflow)
2487  HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2488  } else { // (X / pos) op neg
2489  // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2490  HiBound = Prod + 1;
2491  LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2492  if (!LoOverflow) {
2493  APInt DivNeg = -RangeSize;
2494  LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2495  }
2496  }
2497  } else if (C2->isNegative()) { // Divisor is < 0.
2498  if (Div->isExact())
2499  RangeSize.negate();
2500  if (C.isNullValue()) { // (X / neg) op 0
2501  // e.g. X/-5 op 0 --> [-4, 5)
2502  LoBound = RangeSize + 1;
2503  HiBound = -RangeSize;
2504  if (HiBound == *C2) { // -INTMIN = INTMIN
2505  HiOverflow = 1; // [INTMIN+1, overflow)
2506  HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2507  }
2508  } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2509  // e.g. X/-5 op 3 --> [-19, -14)
2510  HiBound = Prod + 1;
2511  HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2512  if (!LoOverflow)
2513  LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2514  } else { // (X / neg) op neg
2515  LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2516  LoOverflow = HiOverflow = ProdOV;
2517  if (!HiOverflow)
2518  HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2519  }
2520 
2521  // Dividing by a negative swaps the condition. LT <-> GT
2522  Pred = ICmpInst::getSwappedPredicate(Pred);
2523  }
2524 
2525  Value *X = Div->getOperand(0);
2526  switch (Pred) {
2527  default: llvm_unreachable("Unhandled icmp opcode!");
2528  case ICmpInst::ICMP_EQ:
2529  if (LoOverflow && HiOverflow)
2530  return replaceInstUsesWith(Cmp, Builder.getFalse());
2531  if (HiOverflow)
2532  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2534  ConstantInt::get(Div->getType(), LoBound));
2535  if (LoOverflow)
2536  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2538  ConstantInt::get(Div->getType(), HiBound));
2539  return replaceInstUsesWith(
2540  Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2541  case ICmpInst::ICMP_NE:
2542  if (LoOverflow && HiOverflow)
2543  return replaceInstUsesWith(Cmp, Builder.getTrue());
2544  if (HiOverflow)
2545  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2547  ConstantInt::get(Div->getType(), LoBound));
2548  if (LoOverflow)
2549  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2551  ConstantInt::get(Div->getType(), HiBound));
2552  return replaceInstUsesWith(Cmp,
2553  insertRangeTest(X, LoBound, HiBound,
2554  DivIsSigned, false));
2555  case ICmpInst::ICMP_ULT:
2556  case ICmpInst::ICMP_SLT:
2557  if (LoOverflow == +1) // Low bound is greater than input range.
2558  return replaceInstUsesWith(Cmp, Builder.getTrue());
2559  if (LoOverflow == -1) // Low bound is less than input range.
2560  return replaceInstUsesWith(Cmp, Builder.getFalse());
2561  return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2562  case ICmpInst::ICMP_UGT:
2563  case ICmpInst::ICMP_SGT:
2564  if (HiOverflow == +1) // High bound greater than input range.
2565  return replaceInstUsesWith(Cmp, Builder.getFalse());
2566  if (HiOverflow == -1) // High bound less than input range.
2567  return replaceInstUsesWith(Cmp, Builder.getTrue());
2568  if (Pred == ICmpInst::ICMP_UGT)
2569  return new ICmpInst(ICmpInst::ICMP_UGE, X,
2570  ConstantInt::get(Div->getType(), HiBound));
2571  return new ICmpInst(ICmpInst::ICMP_SGE, X,
2572  ConstantInt::get(Div->getType(), HiBound));
2573  }
2574 
2575  return nullptr;
2576 }
2577 
2578 /// Fold icmp (sub X, Y), C.
2580  BinaryOperator *Sub,
2581  const APInt &C) {
2582  Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2583  ICmpInst::Predicate Pred = Cmp.getPredicate();
2584  const APInt *C2;
2585  APInt SubResult;
2586 
2587  // icmp eq/ne (sub C, Y), C -> icmp eq/ne Y, 0
2588  if (match(X, m_APInt(C2)) && *C2 == C && Cmp.isEquality())
2589  return new ICmpInst(Cmp.getPredicate(), Y,
2590  ConstantInt::get(Y->getType(), 0));
2591 
2592  // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2593  if (match(X, m_APInt(C2)) &&
2594  ((Cmp.isUnsigned() && Sub->hasNoUnsignedWrap()) ||
2595  (Cmp.isSigned() && Sub->hasNoSignedWrap())) &&
2596  !subWithOverflow(SubResult, *C2, C, Cmp.isSigned()))
2597  return new ICmpInst(Cmp.getSwappedPredicate(), Y,
2598  ConstantInt::get(Y->getType(), SubResult));
2599 
2600  // The following transforms are only worth it if the only user of the subtract
2601  // is the icmp.
2602  if (!Sub->hasOneUse())
2603  return nullptr;
2604 
2605  if (Sub->hasNoSignedWrap()) {
2606  // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2607  if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue())
2608  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2609 
2610  // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2611  if (Pred == ICmpInst::ICMP_SGT && C.isNullValue())
2612  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2613 
2614  // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2615  if (Pred == ICmpInst::ICMP_SLT && C.isNullValue())
2616  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2617 
2618  // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2619  if (Pred == ICmpInst::ICMP_SLT && C.isOneValue())
2620  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2621  }
2622 
2623  if (!match(X, m_APInt(C2)))
2624  return nullptr;
2625 
2626  // C2 - Y <u C -> (Y | (C - 1)) == C2
2627  // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2628  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2629  (*C2 & (C - 1)) == (C - 1))
2630  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2631 
2632  // C2 - Y >u C -> (Y | C) != C2
2633  // iff C2 & C == C and C + 1 is a power of 2
2634  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2635  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2636 
2637  return nullptr;
2638 }
2639 
2640 /// Fold icmp (add X, Y), C.
2642  BinaryOperator *Add,
2643  const APInt &C) {
2644  Value *Y = Add->getOperand(1);
2645  const APInt *C2;
2646  if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2647  return nullptr;
2648 
2649  // Fold icmp pred (add X, C2), C.
2650  Value *X = Add->getOperand(0);
2651  Type *Ty = Add->getType();
2652  const CmpInst::Predicate Pred = Cmp.getPredicate();
2653 
2654  // If the add does not wrap, we can always adjust the compare by subtracting
2655  // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2656  // are canonicalized to SGT/SLT/UGT/ULT.
2657  if ((Add->hasNoSignedWrap() &&
2658  (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2659  (Add->hasNoUnsignedWrap() &&
2660  (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2661  bool Overflow;
2662  APInt NewC =
2663  Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2664  // If there is overflow, the result must be true or false.
2665  // TODO: Can we assert there is no overflow because InstSimplify always
2666  // handles those cases?
2667  if (!Overflow)
2668  // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2669  return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2670  }
2671 
2672  auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2673  const APInt &Upper = CR.getUpper();
2674  const APInt &Lower = CR.getLower();
2675  if (Cmp.isSigned()) {
2676  if (Lower.isSignMask())
2677  return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2678  if (Upper.isSignMask())
2679  return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2680  } else {
2681  if (Lower.isMinValue())
2682  return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2683  if (Upper.isMinValue())
2684  return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2685  }
2686 
2687  // This set of folds is intentionally placed after folds that use no-wrapping
2688  // flags because those folds are likely better for later analysis/codegen.
2691 
2692  // Fold compare with offset to opposite sign compare if it eliminates offset:
2693  // (X + C2) >u C --> X <s -C2 (if C == C2 + SMAX)
2694  if (Pred == CmpInst::ICMP_UGT && C == *C2 + SMax)
2695  return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, -(*C2)));
2696 
2697  // (X + C2) <u C --> X >s ~C2 (if C == C2 + SMIN)
2698  if (Pred == CmpInst::ICMP_ULT && C == *C2 + SMin)
2699  return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantInt::get(Ty, ~(*C2)));
2700 
2701  // (X + C2) >s C --> X <u (SMAX - C) (if C == C2 - 1)
2702  if (Pred == CmpInst::ICMP_SGT && C == *C2 - 1)
2703  return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, SMax - C));
2704 
2705  // (X + C2) <s C --> X >u (C ^ SMAX) (if C == C2)
2706  if (Pred == CmpInst::ICMP_SLT && C == *C2)
2707  return new ICmpInst(ICmpInst::ICMP_UGT, X, ConstantInt::get(Ty, C ^ SMax));
2708 
2709  if (!Add->hasOneUse())
2710  return nullptr;
2711 
2712  // X+C <u C2 -> (X & -C2) == C
2713  // iff C & (C2-1) == 0
2714  // C2 is a power of 2
2715  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2716  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2717  ConstantExpr::getNeg(cast<Constant>(Y)));
2718 
2719  // X+C >u C2 -> (X & ~C2) != C
2720  // iff C & C2 == 0
2721  // C2+1 is a power of 2
2722  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2723  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2724  ConstantExpr::getNeg(cast<Constant>(Y)));
2725 
2726  return nullptr;
2727 }
2728 
2730  Value *&RHS, ConstantInt *&Less,
2731  ConstantInt *&Equal,
2732  ConstantInt *&Greater) {
2733  // TODO: Generalize this to work with other comparison idioms or ensure
2734  // they get canonicalized into this form.
2735 
2736  // select i1 (a == b),
2737  // i32 Equal,
2738  // i32 (select i1 (a < b), i32 Less, i32 Greater)
2739  // where Equal, Less and Greater are placeholders for any three constants.
2740  ICmpInst::Predicate PredA;
2741  if (!match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) ||
2742  !ICmpInst::isEquality(PredA))
2743  return false;
2744  Value *EqualVal = SI->getTrueValue();
2745  Value *UnequalVal = SI->getFalseValue();
2746  // We still can get non-canonical predicate here, so canonicalize.
2747  if (PredA == ICmpInst::ICMP_NE)
2748  std::swap(EqualVal, UnequalVal);
2749  if (!match(EqualVal, m_ConstantInt(Equal)))
2750  return false;
2751  ICmpInst::Predicate PredB;
2752  Value *LHS2, *RHS2;
2753  if (!match(UnequalVal, m_Select(m_ICmp(PredB, m_Value(LHS2), m_Value(RHS2)),
2754  m_ConstantInt(Less), m_ConstantInt(Greater))))
2755  return false;
2756  // We can get predicate mismatch here, so canonicalize if possible:
2757  // First, ensure that 'LHS' match.
2758  if (LHS2 != LHS) {
2759  // x sgt y <--> y slt x
2760  std::swap(LHS2, RHS2);
2761  PredB = ICmpInst::getSwappedPredicate(PredB);
2762  }
2763  if (LHS2 != LHS)
2764  return false;
2765  // We also need to canonicalize 'RHS'.
2766  if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(RHS2)) {
2767  // x sgt C-1 <--> x sge C <--> not(x slt C)
2768  auto FlippedStrictness =
2770  PredB, cast<Constant>(RHS2));
2771  if (!FlippedStrictness)
2772  return false;
2773  assert(FlippedStrictness->first == ICmpInst::ICMP_SGE && "Sanity check");
2774  RHS2 = FlippedStrictness->second;
2775  // And kind-of perform the result swap.
2776  std::swap(Less, Greater);
2777  PredB = ICmpInst::ICMP_SLT;
2778  }
2779  return PredB == ICmpInst::ICMP_SLT && RHS == RHS2;
2780 }
2781 
2783  SelectInst *Select,
2784  ConstantInt *C) {
2785 
2786  assert(C && "Cmp RHS should be a constant int!");
2787  // If we're testing a constant value against the result of a three way
2788  // comparison, the result can be expressed directly in terms of the
2789  // original values being compared. Note: We could possibly be more
2790  // aggressive here and remove the hasOneUse test. The original select is
2791  // really likely to simplify or sink when we remove a test of the result.
2792  Value *OrigLHS, *OrigRHS;
2793  ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2794  if (Cmp.hasOneUse() &&
2795  matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2796  C3GreaterThan)) {
2797  assert(C1LessThan && C2Equal && C3GreaterThan);
2798 
2799  bool TrueWhenLessThan =
2800  ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2801  ->isAllOnesValue();
2802  bool TrueWhenEqual =
2803  ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2804  ->isAllOnesValue();
2805  bool TrueWhenGreaterThan =
2806  ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2807  ->isAllOnesValue();
2808 
2809  // This generates the new instruction that will replace the original Cmp
2810  // Instruction. Instead of enumerating the various combinations when
2811  // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2812  // false, we rely on chaining of ORs and future passes of InstCombine to
2813  // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2814 
2815  // When none of the three constants satisfy the predicate for the RHS (C),
2816  // the entire original Cmp can be simplified to a false.
2817  Value *Cond = Builder.getFalse();
2818  if (TrueWhenLessThan)
2819  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
2820  OrigLHS, OrigRHS));
2821  if (TrueWhenEqual)
2822  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
2823  OrigLHS, OrigRHS));
2824  if (TrueWhenGreaterThan)
2825  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
2826  OrigLHS, OrigRHS));
2827 
2828  return replaceInstUsesWith(Cmp, Cond);
2829  }
2830  return nullptr;
2831 }
2832 
2834  auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
2835  if (!Bitcast)
2836  return nullptr;
2837 
2838  ICmpInst::Predicate Pred = Cmp.getPredicate();
2839  Value *Op1 = Cmp.getOperand(1);
2840  Value *BCSrcOp = Bitcast->getOperand(0);
2841 
2842  // Make sure the bitcast doesn't change the number of vector elements.
2843  if (Bitcast->getSrcTy()->getScalarSizeInBits() ==
2844  Bitcast->getDestTy()->getScalarSizeInBits()) {
2845  // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
2846  Value *X;
2847  if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
2848  // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
2849  // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
2850  // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
2851  // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
2852  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
2853  Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
2854  match(Op1, m_Zero()))
2855  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2856 
2857  // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
2858  if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
2859  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
2860 
2861  // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
2862  if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
2863  return new ICmpInst(Pred, X,
2864  ConstantInt::getAllOnesValue(X->getType()));
2865  }
2866 
2867  // Zero-equality checks are preserved through unsigned floating-point casts:
2868  // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
2869  // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
2870  if (match(BCSrcOp, m_UIToFP(m_Value(X))))
2871  if (Cmp.isEquality() && match(Op1, m_Zero()))
2872  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2873 
2874  // If this is a sign-bit test of a bitcast of a casted FP value, eliminate
2875  // the FP extend/truncate because that cast does not change the sign-bit.
2876  // This is true for all standard IEEE-754 types and the X86 80-bit type.
2877  // The sign-bit is always the most significant bit in those types.
2878  const APInt *C;
2879  bool TrueIfSigned;
2880  if (match(Op1, m_APInt(C)) && Bitcast->hasOneUse() &&
2881  InstCombiner::isSignBitCheck(Pred, *C, TrueIfSigned)) {
2882  if (match(BCSrcOp, m_FPExt(m_Value(X))) ||
2883  match(BCSrcOp, m_FPTrunc(m_Value(X)))) {
2884  // (bitcast (fpext/fptrunc X)) to iX) < 0 --> (bitcast X to iY) < 0
2885  // (bitcast (fpext/fptrunc X)) to iX) > -1 --> (bitcast X to iY) > -1
2886  Type *XType = X->getType();
2887 
2888  // We can't currently handle Power style floating point operations here.
2889  if (!(XType->isPPC_FP128Ty() || BCSrcOp->getType()->isPPC_FP128Ty())) {
2890 
2891  Type *NewType = Builder.getIntNTy(XType->getScalarSizeInBits());
2892  if (auto *XVTy = dyn_cast<VectorType>(XType))
2893  NewType = VectorType::get(NewType, XVTy->getElementCount());
2894  Value *NewBitcast = Builder.CreateBitCast(X, NewType);
2895  if (TrueIfSigned)
2896  return new ICmpInst(ICmpInst::ICMP_SLT, NewBitcast,
2897  ConstantInt::getNullValue(NewType));
2898  else
2899  return new ICmpInst(ICmpInst::ICMP_SGT, NewBitcast,
2900  ConstantInt::getAllOnesValue(NewType));
2901  }
2902  }
2903  }
2904  }
2905 
2906  // Test to see if the operands of the icmp are casted versions of other
2907  // values. If the ptr->ptr cast can be stripped off both arguments, do so.
2908  if (Bitcast->getType()->isPointerTy() &&
2909  (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2910  // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2911  // so eliminate it as well.
2912  if (auto *BC2 = dyn_cast<BitCastInst>(Op1))
2913  Op1 = BC2->getOperand(0);
2914 
2915  Op1 = Builder.CreateBitCast(Op1, BCSrcOp->getType());
2916  return new ICmpInst(Pred, BCSrcOp, Op1);
2917  }
2918 
2919  const APInt *C;
2920  if (!match(Cmp.getOperand(1), m_APInt(C)) ||
2921  !Bitcast->getType()->isIntegerTy() ||
2922  !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2923  return nullptr;
2924 
2925  // If this is checking if all elements of a vector compare are set or not,
2926  // invert the casted vector equality compare and test if all compare
2927  // elements are clear or not. Compare against zero is generally easier for
2928  // analysis and codegen.
2929  // icmp eq/ne (bitcast (not X) to iN), -1 --> icmp eq/ne (bitcast X to iN), 0
2930  // Example: are all elements equal? --> are zero elements not equal?
2931  // TODO: Try harder to reduce compare of 2 freely invertible operands?
2932  if (Cmp.isEquality() && C->isAllOnesValue() && Bitcast->hasOneUse() &&
2933  isFreeToInvert(BCSrcOp, BCSrcOp->hasOneUse())) {
2934  Type *ScalarTy = Bitcast->getType();
2935  Value *Cast = Builder.CreateBitCast(Builder.CreateNot(BCSrcOp), ScalarTy);
2936  return new ICmpInst(Pred, Cast, ConstantInt::getNullValue(ScalarTy));
2937  }
2938 
2939  // If this is checking if all elements of an extended vector are clear or not,
2940  // compare in a narrow type to eliminate the extend:
2941  // icmp eq/ne (bitcast (ext X) to iN), 0 --> icmp eq/ne (bitcast X to iM), 0
2942  Value *X;
2943  if (Cmp.isEquality() && C->isNullValue() && Bitcast->hasOneUse() &&
2944  match(BCSrcOp, m_ZExtOrSExt(m_Value(X)))) {
2945  if (auto *VecTy = dyn_cast<FixedVectorType>(X->getType())) {
2946  Type *NewType = Builder.getIntNTy(VecTy->getPrimitiveSizeInBits());
2947  Value *NewCast = Builder.CreateBitCast(X, NewType);
2948  return new ICmpInst(Pred, NewCast, ConstantInt::getNullValue(NewType));
2949  }
2950  }
2951 
2952  // Folding: icmp <pred> iN X, C
2953  // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2954  // and C is a splat of a K-bit pattern
2955  // and SC is a constant vector = <C', C', C', ..., C'>
2956  // Into:
2957  // %E = extractelement <M x iK> %vec, i32 C'
2958  // icmp <pred> iK %E, trunc(C)
2959  Value *Vec;
2961  if (match(BCSrcOp, m_Shuffle(m_Value(Vec), m_Undef(), m_Mask(Mask)))) {
2962  // Check whether every element of Mask is the same constant
2963  if (is_splat(Mask)) {
2964  auto *VecTy = cast<VectorType>(BCSrcOp->getType());
2965  auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2966  if (C->isSplat(EltTy->getBitWidth())) {
2967  // Fold the icmp based on the value of C
2968  // If C is M copies of an iK sized bit pattern,
2969  // then:
2970  // => %E = extractelement <N x iK> %vec, i32 Elem
2971  // icmp <pred> iK %SplatVal, <pattern>
2972  Value *Elem = Builder.getInt32(Mask[0]);
2973  Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2974  Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
2975  return new ICmpInst(Pred, Extract, NewC);
2976  }
2977  }
2978  }
2979  return nullptr;
2980 }
2981 
2982 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2983 /// where X is some kind of instruction.
2985  const APInt *C;
2986  if (!match(Cmp.getOperand(1), m_APInt(C)))
2987  return nullptr;
2988 
2989  if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2990  switch (BO->getOpcode()) {
2991  case Instruction::Xor:
2992  if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
2993  return I;
2994  break;
2995  case Instruction::And:
2996  if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
2997  return I;
2998  break;
2999  case Instruction::Or:
3000  if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
3001  return I;
3002  break;
3003  case Instruction::Mul:
3004  if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
3005  return I;
3006  break;
3007  case Instruction::Shl:
3008  if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
3009  return I;
3010  break;
3011  case Instruction::LShr:
3012  case Instruction::AShr:
3013  if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
3014  return I;
3015  break;
3016  case Instruction::SRem:
3017  if (Instruction *I = foldICmpSRemConstant(Cmp, BO, *C))
3018  return I;
3019  break;
3020  case Instruction::UDiv:
3021  if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
3022  return I;
3024  case Instruction::SDiv:
3025  if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
3026  return I;
3027  break;
3028  case Instruction::Sub:
3029  if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
3030  return I;
3031  break;
3032  case Instruction::Add:
3033  if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
3034  return I;
3035  break;
3036  default:
3037  break;
3038  }
3039  // TODO: These folds could be refactored to be part of the above calls.
3040  if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
3041  return I;
3042  }
3043 
3044  // Match against CmpInst LHS being instructions other than binary operators.
3045 
3046  if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
3047  // For now, we only support constant integers while folding the
3048  // ICMP(SELECT)) pattern. We can extend this to support vector of integers
3049  // similar to the cases handled by binary ops above.
3050  if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
3051  if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
3052  return I;
3053  }
3054 
3055  if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
3056  if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
3057  return I;
3058  }
3059 
3060  if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
3061  if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
3062  return I;
3063 
3064  return nullptr;
3065 }
3066 
3067 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
3068 /// icmp eq/ne BO, C.
3070  ICmpInst &Cmp, BinaryOperator *BO, const APInt &C) {
3071  // TODO: Some of these folds could work with arbitrary constants, but this
3072  // function is limited to scalar and vector splat constants.
3073  if (!Cmp.isEquality())
3074  return nullptr;
3075 
3076  ICmpInst::Predicate Pred = Cmp.getPredicate();
3077  bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
3078  Constant *RHS = cast<Constant>(Cmp.getOperand(1));
3079  Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
3080 
3081  switch (BO->getOpcode()) {
3082  case Instruction::SRem:
3083  // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
3084  if (C.isNullValue() && BO->hasOneUse()) {
3085  const APInt *BOC;
3086  if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
3087  Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
3088  return new ICmpInst(Pred, NewRem,
3090  }
3091  }
3092  break;
3093  case Instruction::Add: {
3094  // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
3095  if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3096  if (BO->hasOneUse())
3097  return new ICmpInst(Pred, BOp0, ConstantExpr::getSub(RHS, BOC));
3098  } else if (C.isNullValue()) {
3099  // Replace ((add A, B) != 0) with (A != -B) if A or B is
3100  // efficiently invertible, or if the add has just this one use.
3101  if (Value *NegVal = dyn_castNegVal(BOp1))
3102  return new ICmpInst(Pred, BOp0, NegVal);
3103  if (Value *NegVal = dyn_castNegVal(BOp0))
3104  return new ICmpInst(Pred, NegVal, BOp1);
3105  if (BO->hasOneUse()) {
3106  Value *Neg = Builder.CreateNeg(BOp1);
3107  Neg->takeName(BO);
3108  return new ICmpInst(Pred, BOp0, Neg);
3109  }
3110  }
3111  break;
3112  }
3113  case Instruction::Xor:
3114  if (BO->hasOneUse()) {
3115  if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3116  // For the xor case, we can xor two constants together, eliminating
3117  // the explicit xor.
3118  return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
3119  } else if (C.isNullValue()) {
3120  // Replace ((xor A, B) != 0) with (A != B)
3121  return new ICmpInst(Pred, BOp0, BOp1);
3122  }
3123  }
3124  break;
3125  case Instruction::Sub:
3126  if (BO->hasOneUse()) {
3127  // Only check for constant LHS here, as constant RHS will be canonicalized
3128  // to add and use the fold above.
3129  if (Constant *BOC = dyn_cast<Constant>(BOp0)) {
3130  // Replace ((sub BOC, B) != C) with (B != BOC-C).
3131  return new ICmpInst(Pred, BOp1, ConstantExpr::getSub(BOC, RHS));
3132  } else if (C.isNullValue()) {
3133  // Replace ((sub A, B) != 0) with (A != B).
3134  return new ICmpInst(Pred, BOp0, BOp1);
3135  }
3136  }
3137  break;
3138  case Instruction::Or: {
3139  const APInt *BOC;
3140  if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
3141  // Comparing if all bits outside of a constant mask are set?
3142  // Replace (X | C) == -1 with (X & ~C) == ~C.
3143  // This removes the -1 constant.
3144  Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
3145  Value *And = Builder.CreateAnd(BOp0, NotBOC);
3146  return new ICmpInst(Pred, And, NotBOC);
3147  }
3148  break;
3149  }
3150  case Instruction::And: {
3151  const APInt *BOC;
3152  if (match(BOp1, m_APInt(BOC))) {
3153  // If we have ((X & C) == C), turn it into ((X & C) != 0).
3154  if (C == *BOC && C.isPowerOf2())
3155  return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
3156  BO, Constant::getNullValue(RHS->getType()));
3157  }
3158  break;
3159  }
3160  case Instruction::UDiv:
3161  if (C.isNullValue()) {
3162  // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
3163  auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3164  return new ICmpInst(NewPred, BOp1, BOp0);
3165  }
3166  break;
3167  default:
3168  break;
3169  }
3170  return nullptr;
3171 }
3172 
3173 /// Fold an equality icmp with LLVM intrinsic and constant operand.
3175  ICmpInst &Cmp, IntrinsicInst *II, const APInt &C) {
3176  Type *Ty = II->getType();
3177  unsigned BitWidth = C.getBitWidth();
3178  const ICmpInst::Predicate Pred = Cmp.getPredicate();
3179 
3180  switch (II->getIntrinsicID()) {
3181  case Intrinsic::abs:
3182  // abs(A) == 0 -> A == 0
3183  // abs(A) == INT_MIN -> A == INT_MIN
3184  if (C.isNullValue() || C.isMinSignedValue())
3185  return new ICmpInst(Pred, II->getArgOperand(0), ConstantInt::get(Ty, C));
3186  break;
3187 
3188  case Intrinsic::bswap:
3189  // bswap(A) == C -> A == bswap(C)
3190  return new ICmpInst(Pred, II->getArgOperand(0),
3191  ConstantInt::get(Ty, C.byteSwap()));
3192 
3193  case Intrinsic::ctlz:
3194  case Intrinsic::cttz: {
3195  // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
3196  if (C == BitWidth)
3197  return new ICmpInst(Pred, II->getArgOperand(0),
3199 
3200  // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
3201  // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
3202  // Limit to one use to ensure we don't increase instruction count.
3203  unsigned Num = C.getLimitedValue(BitWidth);
3204  if (Num != BitWidth && II->hasOneUse()) {
3205  bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
3206  APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
3207  : APInt::getHighBitsSet(BitWidth, Num + 1);
3208  APInt Mask2 = IsTrailing
3210  : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3211  return new ICmpInst(Pred, Builder.CreateAnd(II->getArgOperand(0), Mask1),
3212  ConstantInt::get(Ty, Mask2));
3213  }
3214  break;
3215  }
3216 
3217  case Intrinsic::ctpop: {
3218  // popcount(A) == 0 -> A == 0 and likewise for !=
3219  // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
3220  bool IsZero = C.isNullValue();
3221  if (IsZero || C == BitWidth)
3222  return new ICmpInst(Pred, II->getArgOperand(0),
3223  IsZero ? Constant::getNullValue(Ty)
3225 
3226  break;
3227  }
3228 
3229  case Intrinsic::fshl:
3230  case Intrinsic::fshr:
3231  if (II->getArgOperand(0) == II->getArgOperand(1)) {
3232  // (rot X, ?) == 0/-1 --> X == 0/-1
3233  // TODO: This transform is safe to re-use undef elts in a vector, but
3234  // the constant value passed in by the caller doesn't allow that.
3235  if (C.isNullValue() || C.isAllOnesValue())
3236  return new ICmpInst(Pred, II->getArgOperand(0), Cmp.getOperand(1));
3237 
3238  const APInt *RotAmtC;
3239  // ror(X, RotAmtC) == C --> X == rol(C, RotAmtC)
3240  // rol(X, RotAmtC) == C --> X == ror(C, RotAmtC)
3241  if (match(II->getArgOperand(2), m_APInt(RotAmtC)))
3242  return new ICmpInst(Pred, II->getArgOperand(0),
3243  II->getIntrinsicID() == Intrinsic::fshl
3244  ? ConstantInt::get(Ty, C.rotr(*RotAmtC))
3245  : ConstantInt::get(Ty, C.rotl(*RotAmtC)));
3246  }
3247  break;
3248 
3249  case Intrinsic::uadd_sat: {
3250  // uadd.sat(a, b) == 0 -> (a | b) == 0
3251  if (C.isNullValue()) {
3252  Value *Or = Builder.CreateOr(II->getArgOperand(0), II->getArgOperand(1));
3253  return new ICmpInst(Pred, Or, Constant::getNullValue(Ty));
3254  }
3255  break;
3256  }
3257 
3258  case Intrinsic::usub_sat: {
3259  // usub.sat(a, b) == 0 -> a <= b
3260  if (C.isNullValue()) {
3261  ICmpInst::Predicate NewPred =
3263  return new ICmpInst(NewPred, II->getArgOperand(0), II->getArgOperand(1));
3264  }
3265  break;
3266  }
3267  default:
3268  break;
3269  }
3270 
3271  return nullptr;
3272 }
3273 
3274 /// Fold an icmp with LLVM intrinsics
3276  assert(Cmp.isEquality());
3277 
3278  ICmpInst::Predicate Pred = Cmp.getPredicate();
3279  Value *Op0 = Cmp.getOperand(0);
3280  Value *Op1 = Cmp.getOperand(1);
3281  const auto *IIOp0 = dyn_cast<IntrinsicInst>(Op0);
3282  const auto *IIOp1 = dyn_cast<IntrinsicInst>(Op1);
3283  if (!IIOp0 || !IIOp1 || IIOp0->getIntrinsicID() != IIOp1->getIntrinsicID())
3284  return nullptr;
3285 
3286  switch (IIOp0->getIntrinsicID()) {
3287  case Intrinsic::bswap:
3288  case Intrinsic::bitreverse:
3289  // If both operands are byte-swapped or bit-reversed, just compare the
3290  // original values.
3291  return new ICmpInst(Pred, IIOp0->getOperand(0), IIOp1->getOperand(0));
3292  case Intrinsic::fshl:
3293  case Intrinsic::fshr:
3294  // If both operands are rotated by same amount, just compare the
3295  // original values.
3296  if (IIOp0->getOperand(0) != IIOp0->getOperand(1))
3297  break;
3298  if (IIOp1->getOperand(0) != IIOp1->getOperand(1))
3299  break;
3300  if (IIOp0->getOperand(2) != IIOp1->getOperand(2))
3301  break;
3302  return new ICmpInst(Pred, IIOp0->getOperand(0), IIOp1->getOperand(0));
3303  default:
3304  break;
3305  }
3306 
3307  return nullptr;
3308 }
3309 
3310 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
3312  IntrinsicInst *II,
3313  const APInt &C) {
3314  if (Cmp.isEquality())
3315  return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
3316 
3317  Type *Ty = II->getType();
3318  unsigned BitWidth = C.getBitWidth();
3319  ICmpInst::Predicate Pred = Cmp.getPredicate();
3320  switch (II->getIntrinsicID()) {
3321  case Intrinsic::ctpop: {
3322  // (ctpop X > BitWidth - 1) --> X == -1
3323  Value *X = II->getArgOperand(0);
3324  if (C == BitWidth - 1 && Pred == ICmpInst::ICMP_UGT)
3325  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ, X,
3327  // (ctpop X < BitWidth) --> X != -1
3328  if (C == BitWidth && Pred == ICmpInst::ICMP_ULT)
3329  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE, X,
3331  break;
3332  }
3333  case Intrinsic::ctlz: {
3334  // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
3335  if (Pred == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3336  unsigned Num = C.getLimitedValue();
3337  APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3338  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
3339  II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3340  }
3341 
3342  // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
3343  if (Pred == ICmpInst::ICMP_ULT && C.uge(1) && C.ule(BitWidth)) {
3344  unsigned Num = C.getLimitedValue();
3345  APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
3346  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
3347  II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3348  }
3349  break;
3350  }
3351  case Intrinsic::cttz: {
3352  // Limit to one use to ensure we don't increase instruction count.
3353  if (!II->hasOneUse())
3354  return nullptr;
3355 
3356  // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
3357  if (Pred == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3358  APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
3359  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
3360  Builder.CreateAnd(II->getArgOperand(0), Mask),
3362  }
3363 
3364  // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
3365  if (Pred == ICmpInst::ICMP_ULT && C.uge(1) && C.ule(BitWidth)) {
3366  APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue());
3367  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
3368  Builder.CreateAnd(II->getArgOperand(0), Mask),
3370  }
3371  break;
3372  }
3373  default:
3374  break;
3375  }
3376 
3377  return nullptr;
3378 }
3379 
3380 /// Handle icmp with constant (but not simple integer constant) RHS.
3382  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3383  Constant *RHSC = dyn_cast<Constant>(Op1);
3384  Instruction *LHSI = dyn_cast<Instruction>(Op0);
3385  if (!RHSC || !LHSI)
3386  return nullptr;
3387 
3388  switch (LHSI->getOpcode()) {
3389  case Instruction::GetElementPtr:
3390  // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3391  if (RHSC->isNullValue() &&
3392  cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3393  return new ICmpInst(
3394  I.getPredicate(), LHSI->getOperand(0),
3396  break;
3397  case Instruction::PHI:
3398  // Only fold icmp into the PHI if the phi and icmp are in the same
3399  // block. If in the same block, we're encouraging jump threading. If
3400  // not, we are just pessimizing the code by making an i1 phi.
3401  if (LHSI->getParent() == I.getParent())
3402  if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
3403  return NV;
3404  break;
3405  case Instruction::Select: {
3406  // If either operand of the select is a constant, we can fold the
3407  // comparison into the select arms, which will cause one to be
3408  // constant folded and the select turned into a bitwise or.
3409  Value *Op1 = nullptr, *Op2 = nullptr;
3410  ConstantInt *CI = nullptr;
3411 
3412  auto SimplifyOp = [&](Value *V) {
3413  Value *Op = nullptr;
3414  if (Constant *C = dyn_cast<Constant>(V)) {
3415  Op = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3416  } else if (RHSC->isNullValue()) {
3417  // If null is being compared, check if it can be further simplified.
3418  Op = SimplifyICmpInst(I.getPredicate(), V, RHSC, SQ);
3419  }
3420  return Op;
3421  };
3422  Op1 = SimplifyOp(LHSI->getOperand(1));
3423  if (Op1)
3424  CI = dyn_cast<ConstantInt>(Op1);
3425 
3426  Op2 = SimplifyOp(LHSI->getOperand(2));
3427  if (Op2)
3428  CI = dyn_cast<ConstantInt>(Op2);
3429 
3430  // We only want to perform this transformation if it will not lead to
3431  // additional code. This is true if either both sides of the select
3432  // fold to a constant (in which case the icmp is replaced with a select
3433  // which will usually simplify) or this is the only user of the
3434  // select (in which case we are trading a select+icmp for a simpler
3435  // select+icmp) or all uses of the select can be replaced based on
3436  // dominance information ("Global cases").
3437  bool Transform = false;
3438  if (Op1 && Op2)
3439  Transform = true;
3440  else if (Op1 || Op2) {
3441  // Local case
3442  if (LHSI->hasOneUse())
3443  Transform = true;
3444  // Global cases
3445  else if (CI && !CI->isZero())
3446  // When Op1 is constant try replacing select with second operand.
3447  // Otherwise Op2 is constant and try replacing select with first
3448  // operand.
3449  Transform =
3450  replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
3451  }
3452  if (Transform) {
3453  if (!Op1)
3454  Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
3455  I.getName());
3456  if (!Op2)
3457  Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
3458  I.getName());
3459  return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3460  }
3461  break;
3462  }
3463  case Instruction::IntToPtr:
3464  // icmp pred inttoptr(X), null -> icmp pred X, 0
3465  if (RHSC->isNullValue() &&
3466  DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3467  return new ICmpInst(
3468  I.getPredicate(), LHSI->getOperand(0),
3470  break;
3471 
3472  case Instruction::Load:
3473  // Try to optimize things like "A[i] > 4" to index computations.
3474  if (GetElementPtrInst *GEP =
3475  dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3476  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3477  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3478  !cast<LoadInst>(LHSI)->isVolatile())
3479  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
3480  return Res;
3481  }
3482  break;
3483  }
3484 
3485  return nullptr;
3486 }
3487 
3488 /// Some comparisons can be simplified.
3489 /// In this case, we are looking for comparisons that look like
3490 /// a check for a lossy truncation.
3491 /// Folds:
3492 /// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
3493 /// Where Mask is some pattern that produces all-ones in low bits:
3494 /// (-1 >> y)
3495 /// ((-1 << y) >> y) <- non-canonical, has extra uses
3496 /// ~(-1 << y)
3497 /// ((1 << y) + (-1)) <- non-canonical, has extra uses
3498 /// The Mask can be a constant, too.
3499 /// For some predicates, the operands are commutative.
3500 /// For others, x can only be on a specific side.
3503  ICmpInst::Predicate SrcPred;
3504  Value *X, *M, *Y;
3505  auto m_VariableMask = m_CombineOr(
3507  m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
3510  auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
3511  if (!match(&I, m_c_ICmp(SrcPred,
3513  m_Deferred(X))))
3514  return nullptr;
3515 
3516  ICmpInst::Predicate DstPred;
3517  switch (SrcPred) {
3518  case ICmpInst::Predicate::ICMP_EQ:
3519  // x & (-1 >> y) == x -> x u<= (-1 >> y)
3520  DstPred = ICmpInst::Predicate::ICMP_ULE;
3521  break;
3522  case ICmpInst::Predicate::ICMP_NE:
3523  // x & (-1 >> y) != x -> x u> (-1 >> y)
3524  DstPred = ICmpInst::Predicate::ICMP_UGT;
3525  break;
3526  case ICmpInst::Predicate::ICMP_ULT:
3527  // x & (-1 >> y) u< x -> x u> (-1 >> y)
3528  // x u> x & (-1 >> y) -> x u> (-1 >> y)
3529  DstPred = ICmpInst::Predicate::ICMP_UGT;
3530  break;
3531  case ICmpInst::Predicate::ICMP_UGE:
3532  // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
3533  // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
3534  DstPred = ICmpInst::Predicate::ICMP_ULE;
3535  break;
3536  case ICmpInst::Predicate::ICMP_SLT:
3537  // x & (-1 >> y) s< x -> x s> (-1 >> y)
3538  // x s> x & (-1 >> y) -> x s> (-1 >> y)
3539  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3540  return nullptr;
3541  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3542  return nullptr;
3543  DstPred = ICmpInst::Predicate::ICMP_SGT;
3544  break;
3545  case ICmpInst::Predicate::ICMP_SGE:
3546  // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
3547  // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3548  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3549  return nullptr;
3550  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3551  return nullptr;
3552  DstPred = ICmpInst::Predicate::ICMP_SLE;
3553  break;
3554  case ICmpInst::Predicate::ICMP_SGT:
3555  case ICmpInst::Predicate::ICMP_SLE:
3556  return nullptr;
3557  case ICmpInst::Predicate::ICMP_UGT:
3558  case ICmpInst::Predicate::ICMP_ULE:
3559  llvm_unreachable("Instsimplify took care of commut. variant");
3560  break;
3561  default:
3562  llvm_unreachable("All possible folds are handled.");
3563  }
3564 
3565  // The mask value may be a vector constant that has undefined elements. But it
3566  // may not be safe to propagate those undefs into the new compare, so replace
3567  // those elements by copying an existing, defined, and safe scalar constant.
3568  Type *OpTy = M->getType();
3569  auto *VecC = dyn_cast<Constant>(M);
3570  auto *OpVTy = dyn_cast<FixedVectorType>(OpTy);
3571  if (OpVTy && VecC && VecC->containsUndefOrPoisonElement()) {
3572  Constant *SafeReplacementConstant = nullptr;
3573  for (unsigned i = 0, e = OpVTy->getNumElements(); i != e; ++i) {
3574  if (!isa<UndefValue>(VecC->getAggregateElement(i))) {
3575  SafeReplacementConstant = VecC->getAggregateElement(i);
3576  break;
3577  }
3578  }
3579  assert(SafeReplacementConstant && "Failed to find undef replacement");
3580  M = Constant::replaceUndefsWith(VecC, SafeReplacementConstant);
3581  }
3582 
3583  return Builder.CreateICmp(DstPred, X, M);
3584 }
3585 
3586 /// Some comparisons can be simplified.
3587 /// In this case, we are looking for comparisons that look like
3588 /// a check for a lossy signed truncation.
3589 /// Folds: (MaskedBits is a constant.)
3590 /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3591 /// Into:
3592 /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3593 /// Where KeptBits = bitwidth(%x) - MaskedBits
3594 static Value *
3597  ICmpInst::Predicate SrcPred;
3598  Value *X;
3599  const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3600  // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3601  if (!match(&I, m_c_ICmp(SrcPred,
3603  m_APInt(C1))),
3604  m_Deferred(X))))
3605  return nullptr;
3606 
3607  // Potential handling of non-splats: for each element:
3608  // * if both are undef, replace with constant 0.
3609  // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3610  // * if both are not undef, and are different, bailout.
3611  // * else, only one is undef, then pick the non-undef one.
3612 
3613  // The shift amount must be equal.
3614  if (*C0 != *C1)
3615  return nullptr;
3616  const APInt &MaskedBits = *C0;
3617  assert(MaskedBits != 0 && "shift by zero should be folded away already.");
3618 
3619  ICmpInst::Predicate DstPred;
3620  switch (SrcPred) {
3621  case ICmpInst::Predicate::ICMP_EQ:
3622  // ((%x << MaskedBits) a>> MaskedBits) == %x
3623  // =>
3624  // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3625  DstPred = ICmpInst::Predicate::ICMP_ULT;
3626  break;
3627  case ICmpInst::Predicate::ICMP_NE:
3628  // ((%x << MaskedBits) a>> MaskedBits) != %x
3629  // =>
3630  // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3631  DstPred = ICmpInst::Predicate::ICMP_UGE;
3632  break;
3633  // FIXME: are more folds possible?
3634  default:
3635  return nullptr;
3636  }
3637 
3638  auto *XType = X->getType();
3639  const unsigned XBitWidth = XType->getScalarSizeInBits();
3640  const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3641  assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
3642 
3643  // KeptBits = bitwidth(%x) - MaskedBits
3644  const APInt KeptBits = BitWidth - MaskedBits;
3645  assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
3646  // ICmpCst = (1 << KeptBits)
3647  const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3648  assert(ICmpCst.isPowerOf2());
3649  // AddCst = (1 << (KeptBits-1))
3650  const APInt AddCst = ICmpCst.lshr(1);
3651  assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
3652 
3653  // T0 = add %x, AddCst
3654  Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3655  // T1 = T0 DstPred ICmpCst
3656  Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3657 
3658  return T1;
3659 }
3660 
3661 // Given pattern:
3662 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3663 // we should move shifts to the same hand of 'and', i.e. rewrite as
3664 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3665 // We are only interested in opposite logical shifts here.
3666 // One of the shifts can be truncated.
3667 // If we can, we want to end up creating 'lshr' shift.
3668 static Value *
3671  if (!I.isEquality() || !match(I.getOperand(1), m_Zero()) ||
3672  !I.getOperand(0)->hasOneUse())
3673  return nullptr;
3674 
3675  auto m_AnyLogicalShift = m_LogicalShift(m_Value(), m_Value());
3676 
3677  // Look for an 'and' of two logical shifts, one of which may be truncated.
3678  // We use m_TruncOrSelf() on the RHS to correctly handle commutative case.
3679  Instruction *XShift, *MaybeTruncation, *YShift;
3680  if (!match(
3681  I.getOperand(0),
3682  m_c_And(m_CombineAnd(m_AnyLogicalShift, m_Instruction(XShift)),
3684  m_AnyLogicalShift, m_Instruction(YShift))),
3685  m_Instruction(MaybeTruncation)))))
3686  return nullptr;
3687 
3688  // We potentially looked past 'trunc', but only when matching YShift,
3689  // therefore YShift must have the widest type.
3690  Instruction *WidestShift = YShift;
3691  // Therefore XShift must have the shallowest type.
3692  // Or they both have identical types if there was no truncation.
3693  Instruction *NarrowestShift = XShift;
3694 
3695  Type *WidestTy = WidestShift->getType();
3696  Type *NarrowestTy = NarrowestShift->getType();
3697  assert(NarrowestTy == I.getOperand(0)->getType() &&
3698  "We did not look past any shifts while matching XShift though.");
3699  bool HadTrunc = WidestTy != I.getOperand(0)->getType();
3700 
3701  // If YShift is a 'lshr', swap the shifts around.
3702  if (match(YShift, m_LShr(m_Value(), m_Value())))
3703  std::swap(XShift, YShift);
3704 
3705  // The shifts must be in opposite directions.
3706  auto XShiftOpcode = XShift->getOpcode();
3707  if (XShiftOpcode == YShift->getOpcode())
3708  return nullptr; // Do not care about same-direction shifts here.
3709 
3710  Value *X, *XShAmt, *Y, *YShAmt;
3711  match(XShift, m_BinOp(m_Value(X), m_ZExtOrSelf(m_Value(XShAmt))));
3712  match(YShift, m_BinOp(m_Value(Y), m_ZExtOrSelf(m_Value(YShAmt))));
3713 
3714  // If one of the values being shifted is a constant, then we will end with
3715  // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not,
3716  // however, we will need to ensure that we won't increase instruction count.
3717  if (!isa<Constant>(X) && !isa<Constant>(Y)) {
3718  // At least one of the hands of the 'and' should be one-use shift.
3719  if (!match(I.getOperand(0),
3720  m_c_And(m_OneUse(m_AnyLogicalShift), m_Value())))
3721  return nullptr;
3722  if (HadTrunc) {
3723  // Due to the 'trunc', we will need to widen X. For that either the old
3724  // 'trunc' or the shift amt in the non-truncated shift should be one-use.
3725  if (!MaybeTruncation->hasOneUse() &&
3726  !NarrowestShift->getOperand(1)->hasOneUse())
3727  return nullptr;
3728  }
3729  }
3730 
3731  // We have two shift amounts from two different shifts. The types of those
3732  // shift amounts may not match. If that's the case let's bailout now.
3733  if (XShAmt->getType() != YShAmt->getType())
3734  return nullptr;
3735 
3736  // As input, we have the following pattern:
3737  // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3738  // We want to rewrite that as:
3739  // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3740  // While we know that originally (Q+K) would not overflow
3741  // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
3742  // shift amounts. so it may now overflow in smaller bitwidth.
3743  // To ensure that does not happen, we need to ensure that the total maximal
3744  // shift amount is still representable in that smaller bit width.
3745  unsigned MaximalPossibleTotalShiftAmount =
3746  (WidestTy->getScalarSizeInBits() - 1) +
3747  (NarrowestTy->getScalarSizeInBits() - 1);
3748  APInt MaximalRepresentableShiftAmount =
3750  if (MaximalRepresentableShiftAmount.ult(MaximalPossibleTotalShiftAmount))
3751  return nullptr;
3752 
3753  // Can we fold (XShAmt+YShAmt) ?
3754  auto *NewShAmt = dyn_cast_or_null<Constant>(
3755  SimplifyAddInst(XShAmt, YShAmt, /*isNSW=*/false,
3756  /*isNUW=*/false, SQ.getWithInstruction(&I)));
3757  if (!NewShAmt)
3758  return nullptr;
3759  NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, WidestTy);
3760  unsigned WidestBitWidth = WidestTy->getScalarSizeInBits();
3761 
3762  // Is the new shift amount smaller than the bit width?
3763  // FIXME: could also rely on ConstantRange.
3764  if (!match(NewShAmt,
3765  m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
3766  APInt(WidestBitWidth, WidestBitWidth))))
3767  return nullptr;
3768 
3769  // An extra legality check is needed if we had trunc-of-lshr.
3770  if (HadTrunc && match(WidestShift, m_LShr(m_Value(), m_Value()))) {
3771  auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ,
3772  WidestShift]() {
3773  // It isn't obvious whether it's worth it to analyze non-constants here.
3774  // Also, let's basically give up on non-splat cases, pessimizing vectors.
3775  // If *any* of these preconditions matches we can perform the fold.
3776  Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy()
3777  ? NewShAmt->getSplatValue()
3778  : NewShAmt;
3779  // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold.
3780  if (NewShAmtSplat &&
3781  (NewShAmtSplat->isNullValue() ||
3782  NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1))
3783  return true;
3784  // We consider *min* leading zeros so a single outlier
3785  // blocks the transform as opposed to allowing it.
3786  if (auto *C = dyn_cast<Constant>(NarrowestShift->getOperand(0))) {
3787  KnownBits Known = computeKnownBits(C, SQ.DL);
3788  unsigned MinLeadZero = Known.countMinLeadingZeros();
3789  // If the value being shifted has at most lowest bit set we can fold.
3790  unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3791  if (MaxActiveBits <= 1)
3792  return true;
3793  // Precondition: NewShAmt u<= countLeadingZeros(C)
3794  if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(MinLeadZero))
3795  return true;
3796  }
3797  if (auto *C = dyn_cast<Constant>(WidestShift->getOperand(0))) {
3798  KnownBits Known = computeKnownBits(C, SQ.DL);
3799  unsigned MinLeadZero = Known.countMinLeadingZeros();
3800  // If the value being shifted has at most lowest bit set we can fold.
3801  unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3802  if (MaxActiveBits <= 1)
3803  return true;
3804  // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C)
3805  if (NewShAmtSplat) {
3806  APInt AdjNewShAmt =
3807  (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger();
3808  if (AdjNewShAmt.ule(MinLeadZero))
3809  return true;
3810  }
3811  }
3812  return false; // Can't tell if it's ok.
3813  };
3814  if (!CanFold())
3815  return nullptr;
3816  }
3817 
3818  // All good, we can do this fold.
3819  X = Builder.CreateZExt(X, WidestTy);
3820  Y = Builder.CreateZExt(Y, WidestTy);
3821  // The shift is the same that was for X.
3822  Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
3823  ? Builder.CreateLShr(X, NewShAmt)
3824  : Builder.CreateShl(X, NewShAmt);
3825  Value *T1 = Builder.CreateAnd(T0, Y);
3826  return Builder.CreateICmp(I.getPredicate(), T1,
3827  Constant::getNullValue(WidestTy));
3828 }
3829 
3830 /// Fold
3831 /// (-1 u/ x) u< y
3832 /// ((x * y) ?/ x) != y
3833 /// to
3834 /// @llvm.?mul.with.overflow(x, y) plus extraction of overflow bit
3835 /// Note that the comparison is commutative, while inverted (u>=, ==) predicate
3836 /// will mean that we are looking for the opposite answer.
3838  ICmpInst::Predicate Pred;
3839  Value *X, *Y;
3840  Instruction *Mul;
3841  Instruction *Div;
3842  bool NeedNegation;
3843  // Look for: (-1 u/ x) u</u>= y
3844  if (!I.isEquality() &&
3845  match(&I, m_c_ICmp(Pred,
3847  m_Instruction(Div)),
3848  m_Value(Y)))) {
3849  Mul = nullptr;
3850 
3851  // Are we checking that overflow does not happen, or does happen?
3852  switch (Pred) {
3853  case ICmpInst::Predicate::ICMP_ULT:
3854  NeedNegation = false;
3855  break; // OK
3856  case ICmpInst::Predicate::ICMP_UGE:
3857  NeedNegation = true;
3858  break; // OK
3859  default:
3860  return nullptr; // Wrong predicate.
3861  }
3862  } else // Look for: ((x * y) / x) !=/== y
3863  if (I.isEquality() &&
3864  match(&I,
3865  m_c_ICmp(Pred, m_Value(Y),
3866  m_CombineAnd(
3868  m_Value(X)),
3869  m_Instruction(Mul)),
3870  m_Deferred(X))),
3871  m_Instruction(Div))))) {
3872  NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ;
3873  } else
3874  return nullptr;
3875 
3877  // If the pattern included (x * y), we'll want to insert new instructions
3878  // right before that original multiplication so that we can replace it.
3879  bool MulHadOtherUses = Mul && !Mul->hasOneUse();
3880  if (MulHadOtherUses)
3881  Builder.SetInsertPoint(Mul);
3882 
3883  Function *F = Intrinsic::getDeclaration(I.getModule(),
3884  Div->getOpcode() == Instruction::UDiv
3885  ? Intrinsic::umul_with_overflow
3886  : Intrinsic::smul_with_overflow,
3887  X->getType());
3888  CallInst *Call = Builder.CreateCall(F, {X, Y}, "mul");
3889 
3890  // If the multiplication was used elsewhere, to ensure that we don't leave
3891  // "duplicate" instructions, replace uses of that original multiplication
3892  // with the multiplication result from the with.overflow intrinsic.
3893  if (MulHadOtherUses)
3894  replaceInstUsesWith(*Mul, Builder.CreateExtractValue(Call, 0, "mul.val"));
3895 
3896  Value *Res = Builder.CreateExtractValue(Call, 1, "mul.ov");
3897  if (NeedNegation) // This technically increases instruction count.
3898  Res = Builder.CreateNot(Res, "mul.not.ov");
3899 
3900  // If we replaced the mul, erase it. Do this after all uses of Builder,
3901  // as the mul is used as insertion point.
3902  if (MulHadOtherUses)
3903  eraseInstFromFunction(*Mul);
3904 
3905  return Res;
3906 }
3907 
3909  CmpInst::Predicate Pred;
3910  Value *X;
3911  if (!match(&I, m_c_ICmp(Pred, m_NSWNeg(m_Value(X)), m_Deferred(X))))
3912  return nullptr;
3913 
3914  if (ICmpInst::isSigned(Pred))
3915  Pred = ICmpInst::getSwappedPredicate(Pred);
3916  else if (ICmpInst::isUnsigned(Pred))
3917  Pred = ICmpInst::getSignedPredicate(Pred);
3918  // else for equality-comparisons just keep the predicate.
3919 
3920  return ICmpInst::Create(Instruction::ICmp, Pred, X,
3921  Constant::getNullValue(X->getType()), I.getName());
3922 }
3923 
3924 /// Try to fold icmp (binop), X or icmp X, (binop).
3925 /// TODO: A large part of this logic is duplicated in InstSimplify's
3926 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3927 /// duplication.
3929  const SimplifyQuery &SQ) {
3930  const SimplifyQuery Q = SQ.getWithInstruction(&I);
3931  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3932 
3933  // Special logic for binary operators.
3934  BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3935  BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3936  if (!BO0 && !BO1)
3937  return nullptr;
3938 
3939  if (Instruction *NewICmp = foldICmpXNegX(I))
3940  return NewICmp;
3941 
3942  const CmpInst::Predicate Pred = I.getPredicate();
3943  Value *X;
3944 
3945  // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3946  // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X
3947  if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
3948  (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3949  return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3950  // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0
3951  if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
3952  (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3953  return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3954 
3955  bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3956  if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3957  NoOp0WrapProblem =
3958  ICmpInst::isEquality(Pred) ||
3959  (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3960  (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3961  if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3962  NoOp1WrapProblem =
3963  ICmpInst::isEquality(Pred) ||
3964  (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3965  (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3966 
3967  // Analyze the case when either Op0 or Op1 is an add instruction.
3968  // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3969  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3970  if (BO0 && BO0->getOpcode() == Instruction::Add) {
3971  A = BO0->getOperand(0);
3972  B = BO0->getOperand(1);
3973  }
3974  if (BO1 && BO1->getOpcode() == Instruction::Add) {
3975  C = BO1->getOperand(0);
3976  D = BO1->getOperand(1);
3977  }
3978 
3979  // icmp (A+B), A -> icmp B, 0 for equalities or if there is no overflow.
3980  // icmp (A+B), B -> icmp A, 0 for equalities or if there is no overflow.
3981  if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3982  return new ICmpInst(Pred, A == Op1 ? B : A,
3983  Constant::getNullValue(Op1->getType()));
3984 
3985  // icmp C, (C+D) -> icmp 0, D for equalities or if there is no overflow.
3986  // icmp D, (C+D) -> icmp 0, C for equalities or if there is no overflow.
3987  if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3988  return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3989  C == Op0 ? D : C);
3990 
3991  // icmp (A+B), (A+D) -> icmp B, D for equalities or if there is no overflow.
3992  if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
3993  NoOp1WrapProblem) {
3994  // Determine Y and Z in the form icmp (X+Y), (X+Z).
3995  Value *Y, *Z;
3996  if (A == C) {
3997  // C + B == C + D -> B == D
3998  Y = B;
3999  Z = D;
4000  } else if (A == D) {
4001  // D + B == C + D -> B == C
4002  Y = B;
4003  Z = C;
4004  } else if (B == C) {
4005  // A + C == C + D -> A == D
4006  Y = A;
4007  Z = D;
4008  } else {
4009  assert(B == D);
4010  // A + D == C + D -> A == C
4011  Y = A;
4012  Z = C;
4013  }
4014  return new ICmpInst(Pred, Y, Z);
4015  }
4016 
4017  // icmp slt (A + -1), Op1 -> icmp sle A, Op1
4018  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
4019  match(B, m_AllOnes()))
4020  return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
4021 
4022  // icmp sge (A + -1), Op1 -> icmp sgt A, Op1
4023  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
4024  match(B, m_AllOnes()))
4025  return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
4026 
4027  // icmp sle (A + 1), Op1 -> icmp slt A, Op1
4028  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
4029  return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
4030 
4031  // icmp sgt (A + 1), Op1 -> icmp sge A, Op1
4032  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
4033  return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
4034 
4035  // icmp sgt Op0, (C + -1) -> icmp sge Op0, C
4036  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
4037  match(D, m_AllOnes()))
4038  return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
4039 
4040  // icmp sle Op0, (C + -1) -> icmp slt Op0, C
4041  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
4042  match(D, m_AllOnes()))
4043  return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
4044 
4045  // icmp sge Op0, (C + 1) -> icmp sgt Op0, C
4046  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
4047  return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
4048 
4049  // icmp slt Op0, (C + 1) -> icmp sle Op0, C
4050  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
4051  return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
4052 
4053  // TODO: The subtraction-related identities shown below also hold, but
4054  // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
4055  // wouldn't happen even if they were implemented.
4056  //
4057  // icmp ult (A - 1), Op1 -> icmp ule A, Op1
4058  // icmp uge (A - 1), Op1 -> icmp ugt A, Op1
4059  // icmp ugt Op0, (C - 1) -> icmp uge Op0, C
4060  // icmp ule Op0, (C - 1) -> icmp ult Op0, C
4061 
4062  // icmp ule (A + 1), Op0 -> icmp ult A, Op1
4063  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
4064  return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
4065 
4066  // icmp ugt (A + 1), Op0 -> icmp uge A, Op1
4067  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
4068  return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
4069 
4070  // icmp uge Op0, (C + 1) -> icmp ugt Op0, C
4071  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
4072  return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
4073 
4074  // icmp ult Op0, (C + 1) -> icmp ule Op0, C
4075  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
4076  return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
4077 
4078  // if C1 has greater magnitude than C2:
4079  // icmp (A + C1), (C + C2) -> icmp (A + C3), C
4080  // s.t. C3 = C1 - C2
4081  //
4082  // if C2 has greater magnitude than C1:
4083  // icmp (A + C1), (C + C2) -> icmp A, (C + C3)
4084  // s.t. C3 = C2 - C1
4085  if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
4086  (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
4087  if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
4088  if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
4089  const APInt &AP1 = C1->getValue();
4090  const APInt &AP2 = C2->getValue();
4091  if (AP1.isNegative() == AP2.isNegative()) {
4092  APInt AP1Abs = C1->getValue().abs();
4093  APInt AP2Abs = C2->getValue().abs();
4094  if (AP1Abs.uge(AP2Abs)) {
4095  ConstantInt *C3 = Builder.getInt(AP1 - AP2);
4096  bool HasNUW = BO0->hasNoUnsignedWrap() && C3->getValue().ule(AP1);
4097  bool HasNSW = BO0->hasNoSignedWrap();
4098  Value *NewAdd = Builder.CreateAdd(A, C3, "", HasNUW, HasNSW);
4099  return new ICmpInst(Pred, NewAdd, C);
4100  } else {
4101  ConstantInt *C3 = Builder.getInt(AP2 - AP1);
4102  bool HasNUW = BO1->hasNoUnsignedWrap() && C3->getValue().ule(AP2);
4103  bool HasNSW = BO1->hasNoSignedWrap();
4104  Value *NewAdd = Builder.CreateAdd(C, C3, "", HasNUW, HasNSW);
4105  return new ICmpInst(Pred, A, NewAdd);
4106  }
4107  }
4108  }
4109 
4110  // Analyze the case when either Op0 or Op1 is a sub instruction.
4111  // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
4112  A = nullptr;
4113  B = nullptr;
4114  C = nullptr;
4115  D = nullptr;
4116  if (BO0 && BO0->getOpcode() == Instruction::Sub) {
4117  A = BO0->getOperand(0);
4118  B = BO0->getOperand(1);
4119  }
4120  if (BO1 && BO1->getOpcode() == Instruction::Sub) {
4121  C = BO1->getOperand(0);
4122  D = BO1->getOperand(1);
4123  }
4124 
4125  // icmp (A-B), A -> icmp 0, B for equalities or if there is no overflow.
4126  if (A == Op1 && NoOp0WrapProblem)
4127  return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
4128  // icmp C, (C-D) -> icmp D, 0 for equalities or if there is no overflow.
4129  if (C == Op0 && NoOp1WrapProblem)
4130  return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
4131 
4132  // Convert sub-with-unsigned-overflow comparisons into a comparison of args.
4133  // (A - B) u>/u<= A --> B u>/u<= A
4134  if (A == Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
4135  return new ICmpInst(Pred, B, A);
4136  // C u</u>= (C - D) --> C u</u>= D
4137  if (C == Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
4138  return new ICmpInst(Pred, C, D);
4139  // (A - B) u>=/u< A --> B u>/u<= A iff B != 0
4140  if (A == Op1 && (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
4141  isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
4142  return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), B, A);
4143  // C u<=/u> (C - D) --> C u</u>= D iff B != 0
4144  if (C == Op0 && (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
4145  isKnownNonZero(D, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
4146  return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), C, D);
4147 
4148  // icmp (A-B), (C-B) -> icmp A, C for equalities or if there is no overflow.
4149  if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem)
4150  return new ICmpInst(Pred, A, C);
4151 
4152  // icmp (A-B), (A-D) -> icmp D, B for equalities or if there is no overflow.
4153  if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem)
4154  return new ICmpInst(Pred, D, B);
4155 
4156  // icmp (0-X) < cst --> x > -cst
4157  if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
4158  Value *X;
4159  if (match(BO0, m_Neg(m_Value(X))))
4160  if (Constant *RHSC = dyn_cast<Constant>(Op1))
4161  if (RHSC->isNotMinSignedValue())
4162  return new ICmpInst(I.getSwappedPredicate(), X,
4163  ConstantExpr::getNeg(RHSC));
4164  }
4165 
4166  {
4167  // Try to remove shared constant multiplier from equality comparison:
4168  // X * C == Y * C (with no overflowing/aliasing) --> X == Y
4169  Value *X, *Y;
4170  const APInt *C;
4171  if (match(Op0, m_Mul(m_Value(X), m_APInt(C))) && *C != 0 &&
4172  match(Op1, m_Mul(m_Value(Y), m_SpecificInt(*C))) && I.isEquality())
4173  if (!C->countTrailingZeros() ||
4174  (BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap()) ||
4175  (BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap()))
4176  return new ICmpInst(Pred, X, Y);
4177  }
4178 
4179  BinaryOperator *SRem = nullptr;
4180  // icmp (srem X, Y), Y
4181  if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
4182  SRem = BO0;
4183  // icmp Y, (srem X, Y)
4184  else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
4185  Op0 == BO1->getOperand(1))
4186  SRem = BO1;
4187  if (SRem) {
4188  // We don't check hasOneUse to avoid increasing register pressure because
4189  // the value we use is the same value this instruction was already using.
4190  switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
4191  default:
4192  break;
4193  case ICmpInst::ICMP_EQ:
4194  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4195  case ICmpInst::ICMP_NE:
4196  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4197  case ICmpInst::ICMP_SGT:
4198  case ICmpInst::ICMP_SGE:
4199  return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
4201  case ICmpInst::ICMP_SLT:
4202  case ICmpInst::ICMP_SLE:
4203  return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
4204  Constant::getNullValue(SRem->getType()));
4205  }
4206  }
4207 
4208  if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
4209  BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
4210  switch (BO0->getOpcode()) {
4211  default:
4212  break;
4213  case Instruction::Add:
4214  case Instruction::Sub:
4215  case Instruction::Xor: {
4216  if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
4217  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4218 
4219  const APInt *C;
4220  if (match(BO0->getOperand(1), m_APInt(C))) {
4221  // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
4222  if (C->isSignMask()) {
4223  ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
4224  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4225  }
4226 
4227  // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
4228  if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
4229  ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
4230  NewPred = I.getSwappedPredicate(NewPred);
4231  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4232  }
4233  }
4234  break;
4235  }
4236  case Instruction::Mul: {
4237  if (!I.isEquality())
4238  break;
4239 
4240  const APInt *C;
4241  if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
4242  !C->isOneValue()) {
4243  // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
4244  // Mask = -1 >> count-trailing-zeros(C).
4245  if (unsigned TZs = C->countTrailingZeros()) {
4247  BO0->getType(),
4248  APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
4249  Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
4250  Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
4251  return new ICmpInst(Pred, And1, And2);
4252  }
4253  }
4254  break;
4255  }
4256  case Instruction::UDiv:
4257  case Instruction::LShr:
4258  if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
4259  break;
4260  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4261 
4262  case Instruction::SDiv:
4263  if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
4264  break;
4265  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4266 
4267  case Instruction::AShr:
4268  if (!BO0->isExact() || !BO1->isExact())
4269  break;
4270  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4271 
4272  case Instruction::Shl: {
4273  bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
4274  bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
4275  if (!NUW && !NSW)
4276  break;
4277  if (!NSW && I.isSigned())
4278  break;
4279  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4280  }
4281  }
4282  }
4283 
4284  if (BO0) {
4285  // Transform A & (L - 1) `ult` L --> L != 0
4286  auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
4287  auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
4288 
4289  if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
4290  auto *Zero = Constant::getNullValue(BO0->getType());
4291  return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
4292  }
4293  }
4294 
4295  if (Value *V = foldMultiplicationOverflowCheck(I))
4296  return replaceInstUsesWith(I, V);
4297 
4299  return replaceInstUsesWith(I, V);
4300 
4302  return replaceInstUsesWith(I, V);
4303 
4305  return replaceInstUsesWith(I, V);
4306 
4307  return nullptr;
4308 }
4309 
4310 /// Fold icmp Pred min|max(X, Y), X.
4312  ICmpInst::Predicate Pred = Cmp.getPredicate();
4313  Value *Op0 = Cmp.getOperand(0);
4314  Value *X = Cmp.getOperand(1);
4315 
4316  // Canonicalize minimum or maximum operand to LHS of the icmp.
4317  if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
4318  match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
4319  match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
4320  match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
4321  std::swap(Op0, X);
4322  Pred = Cmp.getSwappedPredicate();
4323  }
4324 
4325  Value *Y;
4326  if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
4327  // smin(X, Y) == X --> X s<= Y
4328  // smin(X, Y) s>= X --> X s<= Y
4329  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
4330  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
4331 
4332  // smin(X, Y) != X --> X s> Y
4333  // smin(X, Y) s< X --> X s> Y
4334  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
4335  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
4336 
4337  // These cases should be handled in InstSimplify:
4338  // smin(X, Y) s<= X --> true
4339  // smin(X, Y) s> X --> false
4340  return nullptr;
4341  }
4342 
4343  if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
4344  // smax(X, Y) == X --> X s>= Y
4345  // smax(X, Y) s<= X --> X s>= Y
4346  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
4347  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
4348 
4349  // smax(X, Y) != X --> X s< Y
4350  // smax(X, Y) s> X --> X s< Y
4351  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
4352  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
4353 
4354  // These cases should be handled in InstSimplify:
4355  // smax(X, Y) s>= X --> true
4356  // smax(X, Y) s< X --> false
4357  return nullptr;
4358  }
4359 
4360  if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
4361  // umin(X, Y) == X --> X u<= Y
4362  // umin(X, Y) u>= X --> X u<= Y
4363  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
4364  return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
4365 
4366  // umin(X, Y) != X --> X u> Y
4367  // umin(X, Y) u< X --> X u> Y
4368  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
4369  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
4370 
4371  // These cases should be handled in InstSimplify:
4372  // umin(X, Y) u<= X --> true
4373  // umin(X, Y) u> X --> false
4374  return nullptr;
4375  }
4376 
4377  if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
4378  // umax(X, Y) == X --> X u>= Y
4379  // umax(X, Y) u<= X --> X u>= Y
4380  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
4381  return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
4382 
4383  // umax(X, Y) != X --> X u< Y
4384  // umax(X, Y) u> X --> X u< Y
4385  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
4386  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
4387 
4388  // These cases should be handled in InstSimplify:
4389  // umax(X, Y) u>= X --> true
4390  // umax(X, Y) u< X --> false
4391  return nullptr;
4392  }
4393 
4394  return nullptr;
4395 }
4396 
4398  if (!I.isEquality())
4399  return nullptr;
4400 
4401  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4402  const CmpInst::Predicate Pred = I.getPredicate();
4403  Value *A, *B, *C, *D;
4404  if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4405  if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
4406  Value *OtherVal = A == Op1 ? B : A;
4407  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4408  }
4409 
4410  if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
4411  // A^c1 == C^c2 --> A == C^(c1^c2)
4412  ConstantInt *C1, *C2;
4413  if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
4414  Op1->hasOneUse()) {
4415  Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
4416  Value *Xor = Builder.CreateXor(C, NC);
4417  return new ICmpInst(Pred, A, Xor);
4418  }
4419 
4420  // A^B == A^D -> B == D
4421  if (A == C)
4422  return new ICmpInst(Pred, B, D);
4423  if (A == D)
4424  return new ICmpInst(Pred, B, C);
4425  if (B == C)
4426  return new ICmpInst(Pred, A, D);
4427  if (B == D)
4428  return new ICmpInst(Pred, A, C);
4429  }
4430  }
4431 
4432  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
4433  // A == (A^B) -> B == 0
4434  Value *OtherVal = A == Op0 ? B : A;
4435  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4436  }
4437 
4438  // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
4439  if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
4440  match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
4441  Value *X = nullptr, *Y = nullptr, *Z = nullptr;
4442 
4443  if (A == C) {
4444  X = B;
4445  Y = D;
4446  Z = A;
4447  } else if (A == D) {
4448  X = B;
4449  Y = C;
4450  Z = A;
4451  } else if (B == C) {
4452  X = A;
4453  Y = D;
4454  Z = B;
4455  } else if (B == D) {
4456  X = A;
4457  Y = C;
4458  Z = B;
4459  }
4460 
4461  if (X) { // Build (X^Y) & Z
4462  Op1 = Builder.CreateXor(X, Y);
4463  Op1 = Builder.CreateAnd(Op1, Z);
4464  return new ICmpInst(Pred, Op1, Constant::getNullValue(Op1->getType()));
4465  }
4466  }
4467 
4468  {
4469  // Similar to above, but specialized for constant because invert is needed:
4470  // (X | C) == (Y | C) --> (X ^ Y) & ~C == 0
4471  Value *X, *Y;
4472  Constant *C;
4473  if (match(Op0, m_OneUse(m_Or(m_Value(X), m_Constant(C)))) &&
4474  match(Op1, m_OneUse(m_Or(m_Value(Y), m_Specific(C))))) {
4475  Value *Xor = Builder.CreateXor(X, Y);
4476  Value *And = Builder.CreateAnd(Xor, ConstantExpr::getNot(C));
4477  return new ICmpInst(Pred, And, Constant::getNullValue(And->getType()));
4478  }
4479  }
4480 
4481  // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
4482  // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
4483  ConstantInt *Cst1;
4484  if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
4485  match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
4486  (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
4487  match(Op1, m_ZExt(m_Value(A))))) {
4488  APInt Pow2 = Cst1->getValue() + 1;
4489  if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
4490  Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
4491  return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
4492  }
4493 
4494  // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
4495  // For lshr and ashr pairs.
4496  if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4497  match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
4498  (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4499  match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
4500  unsigned TypeBits = Cst1->getBitWidth();
4501  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4502  if (ShAmt < TypeBits && ShAmt != 0) {
4503  ICmpInst::Predicate NewPred =
4505  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4506  APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
4507  return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
4508  }
4509  }
4510 
4511  // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
4512  if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
4513  match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
4514  unsigned TypeBits = Cst1->getBitWidth();
4515  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4516  if (ShAmt < TypeBits && ShAmt != 0) {
4517  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4518  APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
4519  Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
4520  I.getName() + ".mask");
4521  return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
4522  }
4523  }
4524 
4525  // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
4526  // "icmp (and X, mask), cst"
4527  uint64_t ShAmt = 0;
4528  if (Op0->hasOneUse() &&
4529  match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
4530  match(Op1, m_ConstantInt(Cst1)) &&
4531  // Only do this when A has multiple uses. This is most important to do
4532  // when it exposes other optimizations.
4533  !A->hasOneUse()) {
4534  unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
4535 
4536  if (ShAmt < ASize) {
4537  APInt MaskV =
4539  MaskV <<= ShAmt;
4540 
4541  APInt CmpV = Cst1->getValue().zext(ASize);
4542  CmpV <<= ShAmt;
4543 
4544  Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
4545  return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
4546  }
4547  }
4548 
4550  return ICmp;
4551 
4552  // Canonicalize checking for a power-of-2-or-zero value:
4553  // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
4554  // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
4555  if (!match(Op0, m_OneUse(m_c_And(m_Add(m_Value(A), m_AllOnes()),
4556  m_Deferred(A)))) ||
4557  !match(Op1, m_ZeroInt()))
4558  A = nullptr;
4559 
4560  // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
4561  // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
4562  if (match(Op0, m_OneUse(m_c_And(m_Neg(m_Specific(Op1)), m_Specific(Op1)))))
4563  A = Op1;
4564  else if (match(Op1,
4565  m_OneUse(m_c_And(m_Neg(m_Specific(Op0)), m_Specific(Op0)))))
4566  A = Op0;
4567 
4568  if (A) {
4569  Type *Ty = A->getType();
4570  CallInst *CtPop = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, A);
4571  return Pred == ICmpInst::ICMP_EQ
4572  ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop, ConstantInt::get(Ty, 2))
4573  : new ICmpInst(ICmpInst::ICMP_UGT, CtPop, ConstantInt::get(Ty, 1));
4574  }
4575 
4576  return nullptr;
4577 }
4578 
4581