LLVM  10.0.0svn
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"
27 
28 using namespace llvm;
29 using namespace PatternMatch;
30 
31 #define DEBUG_TYPE "instcombine"
32 
33 // How many times is a select replaced by one of its operands?
34 STATISTIC(NumSel, "Number of select opts");
35 
36 
37 /// Compute Result = In1+In2, returning true if the result overflowed for this
38 /// type.
39 static bool addWithOverflow(APInt &Result, const APInt &In1,
40  const APInt &In2, bool IsSigned = false) {
41  bool Overflow;
42  if (IsSigned)
43  Result = In1.sadd_ov(In2, Overflow);
44  else
45  Result = In1.uadd_ov(In2, Overflow);
46 
47  return Overflow;
48 }
49 
50 /// Compute Result = In1-In2, returning true if the result overflowed for this
51 /// type.
52 static bool subWithOverflow(APInt &Result, const APInt &In1,
53  const APInt &In2, bool IsSigned = false) {
54  bool Overflow;
55  if (IsSigned)
56  Result = In1.ssub_ov(In2, Overflow);
57  else
58  Result = In1.usub_ov(In2, Overflow);
59 
60  return Overflow;
61 }
62 
63 /// Given an icmp instruction, return true if any use of this comparison is a
64 /// branch on sign bit comparison.
65 static bool hasBranchUse(ICmpInst &I) {
66  for (auto *U : I.users())
67  if (isa<BranchInst>(U))
68  return true;
69  return false;
70 }
71 
72 /// Given an exploded icmp instruction, return true if the comparison only
73 /// checks the sign bit. If it only checks the sign bit, set TrueIfSigned if the
74 /// result of the comparison is true when the input value is signed.
75 static bool isSignBitCheck(ICmpInst::Predicate Pred, const APInt &RHS,
76  bool &TrueIfSigned) {
77  switch (Pred) {
78  case ICmpInst::ICMP_SLT: // True if LHS s< 0
79  TrueIfSigned = true;
80  return RHS.isNullValue();
81  case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
82  TrueIfSigned = true;
83  return RHS.isAllOnesValue();
84  case ICmpInst::ICMP_SGT: // True if LHS s> -1
85  TrueIfSigned = false;
86  return RHS.isAllOnesValue();
87  case ICmpInst::ICMP_UGT:
88  // True if LHS u> RHS and RHS == high-bit-mask - 1
89  TrueIfSigned = true;
90  return RHS.isMaxSignedValue();
91  case ICmpInst::ICMP_UGE:
92  // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
93  TrueIfSigned = true;
94  return RHS.isSignMask();
95  default:
96  return false;
97  }
98 }
99 
100 /// Returns true if the exploded icmp can be expressed as a signed comparison
101 /// to zero and updates the predicate accordingly.
102 /// The signedness of the comparison is preserved.
103 /// TODO: Refactor with decomposeBitTestICmp()?
104 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
105  if (!ICmpInst::isSigned(Pred))
106  return false;
107 
108  if (C.isNullValue())
109  return ICmpInst::isRelational(Pred);
110 
111  if (C.isOneValue()) {
112  if (Pred == ICmpInst::ICMP_SLT) {
113  Pred = ICmpInst::ICMP_SLE;
114  return true;
115  }
116  } else if (C.isAllOnesValue()) {
117  if (Pred == ICmpInst::ICMP_SGT) {
118  Pred = ICmpInst::ICMP_SGE;
119  return true;
120  }
121  }
122 
123  return false;
124 }
125 
126 /// Given a signed integer type and a set of known zero and one bits, compute
127 /// the maximum and minimum values that could have the specified known zero and
128 /// known one bits, returning them in Min/Max.
129 /// TODO: Move to method on KnownBits struct?
131  APInt &Min, APInt &Max) {
132  assert(Known.getBitWidth() == Min.getBitWidth() &&
133  Known.getBitWidth() == Max.getBitWidth() &&
134  "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
135  APInt UnknownBits = ~(Known.Zero|Known.One);
136 
137  // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
138  // bit if it is unknown.
139  Min = Known.One;
140  Max = Known.One|UnknownBits;
141 
142  if (UnknownBits.isNegative()) { // Sign bit is unknown
143  Min.setSignBit();
144  Max.clearSignBit();
145  }
146 }
147 
148 /// Given an unsigned integer type and a set of known zero and one bits, compute
149 /// the maximum and minimum values that could have the specified known zero and
150 /// known one bits, returning them in Min/Max.
151 /// TODO: Move to method on KnownBits struct?
153  APInt &Min, APInt &Max) {
154  assert(Known.getBitWidth() == Min.getBitWidth() &&
155  Known.getBitWidth() == Max.getBitWidth() &&
156  "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
157  APInt UnknownBits = ~(Known.Zero|Known.One);
158 
159  // The minimum value is when the unknown bits are all zeros.
160  Min = Known.One;
161  // The maximum value is when the unknown bits are all ones.
162  Max = Known.One|UnknownBits;
163 }
164 
165 /// This is called when we see this pattern:
166 /// cmp pred (load (gep GV, ...)), cmpcst
167 /// where GV is a global variable with a constant initializer. Try to simplify
168 /// this into some simple computation that does not need the load. For example
169 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
170 ///
171 /// If AndCst is non-null, then the loaded value is masked with that constant
172 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
173 Instruction *InstCombiner::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
174  GlobalVariable *GV,
175  CmpInst &ICI,
176  ConstantInt *AndCst) {
177  Constant *Init = GV->getInitializer();
178  if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
179  return nullptr;
180 
181  uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
182  // Don't blow up on huge arrays.
183  if (ArrayElementCount > MaxArraySizeForCombine)
184  return nullptr;
185 
186  // There are many forms of this optimization we can handle, for now, just do
187  // the simple index into a single-dimensional array.
188  //
189  // Require: GEP GV, 0, i {{, constant indices}}
190  if (GEP->getNumOperands() < 3 ||
191  !isa<ConstantInt>(GEP->getOperand(1)) ||
192  !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
193  isa<Constant>(GEP->getOperand(2)))
194  return nullptr;
195 
196  // Check that indices after the variable are constants and in-range for the
197  // type they index. Collect the indices. This is typically for arrays of
198  // structs.
199  SmallVector<unsigned, 4> LaterIndices;
200 
201  Type *EltTy = Init->getType()->getArrayElementType();
202  for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
203  ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
204  if (!Idx) return nullptr; // Variable index.
205 
206  uint64_t IdxVal = Idx->getZExtValue();
207  if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
208 
209  if (StructType *STy = dyn_cast<StructType>(EltTy))
210  EltTy = STy->getElementType(IdxVal);
211  else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
212  if (IdxVal >= ATy->getNumElements()) return nullptr;
213  EltTy = ATy->getElementType();
214  } else {
215  return nullptr; // Unknown type.
216  }
217 
218  LaterIndices.push_back(IdxVal);
219  }
220 
221  enum { Overdefined = -3, Undefined = -2 };
222 
223  // Variables for our state machines.
224 
225  // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
226  // "i == 47 | i == 87", where 47 is the first index the condition is true for,
227  // and 87 is the second (and last) index. FirstTrueElement is -2 when
228  // undefined, otherwise set to the first true element. SecondTrueElement is
229  // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
230  int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
231 
232  // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
233  // form "i != 47 & i != 87". Same state transitions as for true elements.
234  int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
235 
236  /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
237  /// define a state machine that triggers for ranges of values that the index
238  /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
239  /// This is -2 when undefined, -3 when overdefined, and otherwise the last
240  /// index in the range (inclusive). We use -2 for undefined here because we
241  /// use relative comparisons and don't want 0-1 to match -1.
242  int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
243 
244  // MagicBitvector - This is a magic bitvector where we set a bit if the
245  // comparison is true for element 'i'. If there are 64 elements or less in
246  // the array, this will fully represent all the comparison results.
247  uint64_t MagicBitvector = 0;
248 
249  // Scan the array and see if one of our patterns matches.
250  Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
251  for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
252  Constant *Elt = Init->getAggregateElement(i);
253  if (!Elt) return nullptr;
254 
255  // If this is indexing an array of structures, get the structure element.
256  if (!LaterIndices.empty())
257  Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
258 
259  // If the element is masked, handle it.
260  if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
261 
262  // Find out if the comparison would be true or false for the i'th element.
264  CompareRHS, DL, &TLI);
265  // If the result is undef for this element, ignore it.
266  if (isa<UndefValue>(C)) {
267  // Extend range state machines to cover this element in case there is an
268  // undef in the middle of the range.
269  if (TrueRangeEnd == (int)i-1)
270  TrueRangeEnd = i;
271  if (FalseRangeEnd == (int)i-1)
272  FalseRangeEnd = i;
273  continue;
274  }
275 
276  // If we can't compute the result for any of the elements, we have to give
277  // up evaluating the entire conditional.
278  if (!isa<ConstantInt>(C)) return nullptr;
279 
280  // Otherwise, we know if the comparison is true or false for this element,
281  // update our state machines.
282  bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
283 
284  // State machine for single/double/range index comparison.
285  if (IsTrueForElt) {
286  // Update the TrueElement state machine.
287  if (FirstTrueElement == Undefined)
288  FirstTrueElement = TrueRangeEnd = i; // First true element.
289  else {
290  // Update double-compare state machine.
291  if (SecondTrueElement == Undefined)
292  SecondTrueElement = i;
293  else
294  SecondTrueElement = Overdefined;
295 
296  // Update range state machine.
297  if (TrueRangeEnd == (int)i-1)
298  TrueRangeEnd = i;
299  else
300  TrueRangeEnd = Overdefined;
301  }
302  } else {
303  // Update the FalseElement state machine.
304  if (FirstFalseElement == Undefined)
305  FirstFalseElement = FalseRangeEnd = i; // First false element.
306  else {
307  // Update double-compare state machine.
308  if (SecondFalseElement == Undefined)
309  SecondFalseElement = i;
310  else
311  SecondFalseElement = Overdefined;
312 
313  // Update range state machine.
314  if (FalseRangeEnd == (int)i-1)
315  FalseRangeEnd = i;
316  else
317  FalseRangeEnd = Overdefined;
318  }
319  }
320 
321  // If this element is in range, update our magic bitvector.
322  if (i < 64 && IsTrueForElt)
323  MagicBitvector |= 1ULL << i;
324 
325  // If all of our states become overdefined, bail out early. Since the
326  // predicate is expensive, only check it every 8 elements. This is only
327  // really useful for really huge arrays.
328  if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
329  SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
330  FalseRangeEnd == Overdefined)
331  return nullptr;
332  }
333 
334  // Now that we've scanned the entire array, emit our new comparison(s). We
335  // order the state machines in complexity of the generated code.
336  Value *Idx = GEP->getOperand(2);
337 
338  // If the index is larger than the pointer size of the target, truncate the
339  // index down like the GEP would do implicitly. We don't have to do this for
340  // an inbounds GEP because the index can't be out of range.
341  if (!GEP->isInBounds()) {
342  Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
343  unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
344  if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
345  Idx = Builder.CreateTrunc(Idx, IntPtrTy);
346  }
347 
348  // If the comparison is only true for one or two elements, emit direct
349  // comparisons.
350  if (SecondTrueElement != Overdefined) {
351  // None true -> false.
352  if (FirstTrueElement == Undefined)
353  return replaceInstUsesWith(ICI, Builder.getFalse());
354 
355  Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
356 
357  // True for one element -> 'i == 47'.
358  if (SecondTrueElement == Undefined)
359  return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
360 
361  // True for two elements -> 'i == 47 | i == 72'.
362  Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
363  Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
364  Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
365  return BinaryOperator::CreateOr(C1, C2);
366  }
367 
368  // If the comparison is only false for one or two elements, emit direct
369  // comparisons.
370  if (SecondFalseElement != Overdefined) {
371  // None false -> true.
372  if (FirstFalseElement == Undefined)
373  return replaceInstUsesWith(ICI, Builder.getTrue());
374 
375  Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
376 
377  // False for one element -> 'i != 47'.
378  if (SecondFalseElement == Undefined)
379  return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
380 
381  // False for two elements -> 'i != 47 & i != 72'.
382  Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
383  Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
384  Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
385  return BinaryOperator::CreateAnd(C1, C2);
386  }
387 
388  // If the comparison can be replaced with a range comparison for the elements
389  // where it is true, emit the range check.
390  if (TrueRangeEnd != Overdefined) {
391  assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
392 
393  // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
394  if (FirstTrueElement) {
395  Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
396  Idx = Builder.CreateAdd(Idx, Offs);
397  }
398 
399  Value *End = ConstantInt::get(Idx->getType(),
400  TrueRangeEnd-FirstTrueElement+1);
401  return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
402  }
403 
404  // False range check.
405  if (FalseRangeEnd != Overdefined) {
406  assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
407  // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
408  if (FirstFalseElement) {
409  Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
410  Idx = Builder.CreateAdd(Idx, Offs);
411  }
412 
413  Value *End = ConstantInt::get(Idx->getType(),
414  FalseRangeEnd-FirstFalseElement);
415  return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
416  }
417 
418  // If a magic bitvector captures the entire comparison state
419  // of this load, replace it with computation that does:
420  // ((magic_cst >> i) & 1) != 0
421  {
422  Type *Ty = nullptr;
423 
424  // Look for an appropriate type:
425  // - The type of Idx if the magic fits
426  // - The smallest fitting legal type
427  if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
428  Ty = Idx->getType();
429  else
430  Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
431 
432  if (Ty) {
433  Value *V = Builder.CreateIntCast(Idx, Ty, false);
434  V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
435  V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
436  return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
437  }
438  }
439 
440  return nullptr;
441 }
442 
443 /// Return a value that can be used to compare the *offset* implied by a GEP to
444 /// zero. For example, if we have &A[i], we want to return 'i' for
445 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
446 /// are involved. The above expression would also be legal to codegen as
447 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
448 /// This latter form is less amenable to optimization though, and we are allowed
449 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
450 ///
451 /// If we can't emit an optimized form for this expression, this returns null.
452 ///
454  const DataLayout &DL) {
456 
457  // Check to see if this gep only has a single variable index. If so, and if
458  // any constant indices are a multiple of its scale, then we can compute this
459  // in terms of the scale of the variable index. For example, if the GEP
460  // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
461  // because the expression will cross zero at the same point.
462  unsigned i, e = GEP->getNumOperands();
463  int64_t Offset = 0;
464  for (i = 1; i != e; ++i, ++GTI) {
465  if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
466  // Compute the aggregate offset of constant indices.
467  if (CI->isZero()) continue;
468 
469  // Handle a struct index, which adds its field offset to the pointer.
470  if (StructType *STy = GTI.getStructTypeOrNull()) {
471  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
472  } else {
473  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
474  Offset += Size*CI->getSExtValue();
475  }
476  } else {
477  // Found our variable index.
478  break;
479  }
480  }
481 
482  // If there are no variable indices, we must have a constant offset, just
483  // evaluate it the general way.
484  if (i == e) return nullptr;
485 
486  Value *VariableIdx = GEP->getOperand(i);
487  // Determine the scale factor of the variable element. For example, this is
488  // 4 if the variable index is into an array of i32.
489  uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
490 
491  // Verify that there are no other variable indices. If so, emit the hard way.
492  for (++i, ++GTI; i != e; ++i, ++GTI) {
493  ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
494  if (!CI) return nullptr;
495 
496  // Compute the aggregate offset of constant indices.
497  if (CI->isZero()) continue;
498 
499  // Handle a struct index, which adds its field offset to the pointer.
500  if (StructType *STy = GTI.getStructTypeOrNull()) {
501  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
502  } else {
503  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
504  Offset += Size*CI->getSExtValue();
505  }
506  }
507 
508  // Okay, we know we have a single variable index, which must be a
509  // pointer/array/vector index. If there is no offset, life is simple, return
510  // the index.
511  Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
512  unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
513  if (Offset == 0) {
514  // Cast to intptrty in case a truncation occurs. If an extension is needed,
515  // we don't need to bother extending: the extension won't affect where the
516  // computation crosses zero.
517  if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
518  VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
519  }
520  return VariableIdx;
521  }
522 
523  // Otherwise, there is an index. The computation we will do will be modulo
524  // the pointer size.
525  Offset = SignExtend64(Offset, IntPtrWidth);
526  VariableScale = SignExtend64(VariableScale, IntPtrWidth);
527 
528  // To do this transformation, any constant index must be a multiple of the
529  // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
530  // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
531  // multiple of the variable scale.
532  int64_t NewOffs = Offset / (int64_t)VariableScale;
533  if (Offset != NewOffs*(int64_t)VariableScale)
534  return nullptr;
535 
536  // Okay, we can do this evaluation. Start by converting the index to intptr.
537  if (VariableIdx->getType() != IntPtrTy)
538  VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
539  true /*Signed*/);
540  Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
541  return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
542 }
543 
544 /// Returns true if we can rewrite Start as a GEP with pointer Base
545 /// and some integer offset. The nodes that need to be re-written
546 /// for this transformation will be added to Explored.
547 static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
548  const DataLayout &DL,
549  SetVector<Value *> &Explored) {
550  SmallVector<Value *, 16> WorkList(1, Start);
551  Explored.insert(Base);
552 
553  // The following traversal gives us an order which can be used
554  // when doing the final transformation. Since in the final
555  // transformation we create the PHI replacement instructions first,
556  // we don't have to get them in any particular order.
557  //
558  // However, for other instructions we will have to traverse the
559  // operands of an instruction first, which means that we have to
560  // do a post-order traversal.
561  while (!WorkList.empty()) {
563 
564  while (!WorkList.empty()) {
565  if (Explored.size() >= 100)
566  return false;
567 
568  Value *V = WorkList.back();
569 
570  if (Explored.count(V) != 0) {
571  WorkList.pop_back();
572  continue;
573  }
574 
575  if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
576  !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
577  // We've found some value that we can't explore which is different from
578  // the base. Therefore we can't do this transformation.
579  return false;
580 
581  if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
582  auto *CI = dyn_cast<CastInst>(V);
583  if (!CI->isNoopCast(DL))
584  return false;
585 
586  if (Explored.count(CI->getOperand(0)) == 0)
587  WorkList.push_back(CI->getOperand(0));
588  }
589 
590  if (auto *GEP = dyn_cast<GEPOperator>(V)) {
591  // We're limiting the GEP to having one index. This will preserve
592  // the original pointer type. We could handle more cases in the
593  // future.
594  if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
595  GEP->getType() != Start->getType())
596  return false;
597 
598  if (Explored.count(GEP->getOperand(0)) == 0)
599  WorkList.push_back(GEP->getOperand(0));
600  }
601 
602  if (WorkList.back() == V) {
603  WorkList.pop_back();
604  // We've finished visiting this node, mark it as such.
605  Explored.insert(V);
606  }
607 
608  if (auto *PN = dyn_cast<PHINode>(V)) {
609  // We cannot transform PHIs on unsplittable basic blocks.
610  if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
611  return false;
612  Explored.insert(PN);
613  PHIs.insert(PN);
614  }
615  }
616 
617  // Explore the PHI nodes further.
618  for (auto *PN : PHIs)
619  for (Value *Op : PN->incoming_values())
620  if (Explored.count(Op) == 0)
621  WorkList.push_back(Op);
622  }
623 
624  // Make sure that we can do this. Since we can't insert GEPs in a basic
625  // block before a PHI node, we can't easily do this transformation if
626  // we have PHI node users of transformed instructions.
627  for (Value *Val : Explored) {
628  for (Value *Use : Val->uses()) {
629 
630  auto *PHI = dyn_cast<PHINode>(Use);
631  auto *Inst = dyn_cast<Instruction>(Val);
632 
633  if (Inst == Base || Inst == PHI || !Inst || !PHI ||
634  Explored.count(PHI) == 0)
635  continue;
636 
637  if (PHI->getParent() == Inst->getParent())
638  return false;
639  }
640  }
641  return true;
642 }
643 
644 // Sets the appropriate insert point on Builder where we can add
645 // a replacement Instruction for V (if that is possible).
646 static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
647  bool Before = true) {
648  if (auto *PHI = dyn_cast<PHINode>(V)) {
649  Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
650  return;
651  }
652  if (auto *I = dyn_cast<Instruction>(V)) {
653  if (!Before)
654  I = &*std::next(I->getIterator());
655  Builder.SetInsertPoint(I);
656  return;
657  }
658  if (auto *A = dyn_cast<Argument>(V)) {
659  // Set the insertion point in the entry block.
660  BasicBlock &Entry = A->getParent()->getEntryBlock();
661  Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
662  return;
663  }
664  // Otherwise, this is a constant and we don't need to set a new
665  // insertion point.
666  assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
667 }
668 
669 /// Returns a re-written value of Start as an indexed GEP using Base as a
670 /// pointer.
672  const DataLayout &DL,
673  SetVector<Value *> &Explored) {
674  // Perform all the substitutions. This is a bit tricky because we can
675  // have cycles in our use-def chains.
676  // 1. Create the PHI nodes without any incoming values.
677  // 2. Create all the other values.
678  // 3. Add the edges for the PHI nodes.
679  // 4. Emit GEPs to get the original pointers.
680  // 5. Remove the original instructions.
681  Type *IndexType = IntegerType::get(
682  Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
683 
685  NewInsts[Base] = ConstantInt::getNullValue(IndexType);
686 
687  // Create the new PHI nodes, without adding any incoming values.
688  for (Value *Val : Explored) {
689  if (Val == Base)
690  continue;
691  // Create empty phi nodes. This avoids cyclic dependencies when creating
692  // the remaining instructions.
693  if (auto *PHI = dyn_cast<PHINode>(Val))
694  NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
695  PHI->getName() + ".idx", PHI);
696  }
697  IRBuilder<> Builder(Base->getContext());
698 
699  // Create all the other instructions.
700  for (Value *Val : Explored) {
701 
702  if (NewInsts.find(Val) != NewInsts.end())
703  continue;
704 
705  if (auto *CI = dyn_cast<CastInst>(Val)) {
706  // Don't get rid of the intermediate variable here; the store can grow
707  // the map which will invalidate the reference to the input value.
708  Value *V = NewInsts[CI->getOperand(0)];
709  NewInsts[CI] = V;
710  continue;
711  }
712  if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
713  Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
714  : GEP->getOperand(1);
715  setInsertionPoint(Builder, GEP);
716  // Indices might need to be sign extended. GEPs will magically do
717  // this, but we need to do it ourselves here.
718  if (Index->getType()->getScalarSizeInBits() !=
719  NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
720  Index = Builder.CreateSExtOrTrunc(
721  Index, NewInsts[GEP->getOperand(0)]->getType(),
722  GEP->getOperand(0)->getName() + ".sext");
723  }
724 
725  auto *Op = NewInsts[GEP->getOperand(0)];
726  if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
727  NewInsts[GEP] = Index;
728  else
729  NewInsts[GEP] = Builder.CreateNSWAdd(
730  Op, Index, GEP->getOperand(0)->getName() + ".add");
731  continue;
732  }
733  if (isa<PHINode>(Val))
734  continue;
735 
736  llvm_unreachable("Unexpected instruction type");
737  }
738 
739  // Add the incoming values to the PHI nodes.
740  for (Value *Val : Explored) {
741  if (Val == Base)
742  continue;
743  // All the instructions have been created, we can now add edges to the
744  // phi nodes.
745  if (auto *PHI = dyn_cast<PHINode>(Val)) {
746  PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
747  for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
748  Value *NewIncoming = PHI->getIncomingValue(I);
749 
750  if (NewInsts.find(NewIncoming) != NewInsts.end())
751  NewIncoming = NewInsts[NewIncoming];
752 
753  NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
754  }
755  }
756  }
757 
758  for (Value *Val : Explored) {
759  if (Val == Base)
760  continue;
761 
762  // Depending on the type, for external users we have to emit
763  // a GEP or a GEP + ptrtoint.
764  setInsertionPoint(Builder, Val, false);
765 
766  // If required, create an inttoptr instruction for Base.
767  Value *NewBase = Base;
768  if (!Base->getType()->isPointerTy())
769  NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
770  Start->getName() + "to.ptr");
771 
772  Value *GEP = Builder.CreateInBoundsGEP(
773  Start->getType()->getPointerElementType(), NewBase,
774  makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
775 
776  if (!Val->getType()->isPointerTy()) {
777  Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
778  Val->getName() + ".conv");
779  GEP = Cast;
780  }
781  Val->replaceAllUsesWith(GEP);
782  }
783 
784  return NewInsts[Start];
785 }
786 
787 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
788 /// the input Value as a constant indexed GEP. Returns a pair containing
789 /// the GEPs Pointer and Index.
790 static std::pair<Value *, Value *>
792  Type *IndexType = IntegerType::get(V->getContext(),
794 
796  while (true) {
797  if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
798  // We accept only inbouds GEPs here to exclude the possibility of
799  // overflow.
800  if (!GEP->isInBounds())
801  break;
802  if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
803  GEP->getType() == V->getType()) {
804  V = GEP->getOperand(0);
805  Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
806  Index = ConstantExpr::getAdd(
807  Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
808  continue;
809  }
810  break;
811  }
812  if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
813  if (!CI->isNoopCast(DL))
814  break;
815  V = CI->getOperand(0);
816  continue;
817  }
818  if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
819  if (!CI->isNoopCast(DL))
820  break;
821  V = CI->getOperand(0);
822  continue;
823  }
824  break;
825  }
826  return {V, Index};
827 }
828 
829 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
830 /// We can look through PHIs, GEPs and casts in order to determine a common base
831 /// between GEPLHS and RHS.
833  ICmpInst::Predicate Cond,
834  const DataLayout &DL) {
835  // FIXME: Support vector of pointers.
836  if (GEPLHS->getType()->isVectorTy())
837  return nullptr;
838 
839  if (!GEPLHS->hasAllConstantIndices())
840  return nullptr;
841 
842  // Make sure the pointers have the same type.
843  if (GEPLHS->getType() != RHS->getType())
844  return nullptr;
845 
846  Value *PtrBase, *Index;
847  std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
848 
849  // The set of nodes that will take part in this transformation.
850  SetVector<Value *> Nodes;
851 
852  if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
853  return nullptr;
854 
855  // We know we can re-write this as
856  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
857  // Since we've only looked through inbouds GEPs we know that we
858  // can't have overflow on either side. We can therefore re-write
859  // this as:
860  // OFFSET1 cmp OFFSET2
861  Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
862 
863  // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
864  // GEP having PtrBase as the pointer base, and has returned in NewRHS the
865  // offset. Since Index is the offset of LHS to the base pointer, we will now
866  // compare the offsets instead of comparing the pointers.
867  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
868 }
869 
870 /// Fold comparisons between a GEP instruction and something else. At this point
871 /// we know that the GEP is on the LHS of the comparison.
872 Instruction *InstCombiner::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
873  ICmpInst::Predicate Cond,
874  Instruction &I) {
875  // Don't transform signed compares of GEPs into index compares. Even if the
876  // GEP is inbounds, the final add of the base pointer can have signed overflow
877  // and would change the result of the icmp.
878  // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
879  // the maximum signed value for the pointer type.
880  if (ICmpInst::isSigned(Cond))
881  return nullptr;
882 
883  // Look through bitcasts and addrspacecasts. We do not however want to remove
884  // 0 GEPs.
885  if (!isa<GetElementPtrInst>(RHS))
886  RHS = RHS->stripPointerCasts();
887 
888  Value *PtrBase = GEPLHS->getOperand(0);
889  // FIXME: Support vector pointer GEPs.
890  if (PtrBase == RHS && GEPLHS->isInBounds() &&
891  !GEPLHS->getType()->isVectorTy()) {
892  // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
893  // This transformation (ignoring the base and scales) is valid because we
894  // know pointers can't overflow since the gep is inbounds. See if we can
895  // output an optimized form.
896  Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
897 
898  // If not, synthesize the offset the hard way.
899  if (!Offset)
900  Offset = EmitGEPOffset(GEPLHS);
901  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
902  Constant::getNullValue(Offset->getType()));
903  }
904 
905  if (GEPLHS->isInBounds() && ICmpInst::isEquality(Cond) &&
906  isa<Constant>(RHS) && cast<Constant>(RHS)->isNullValue() &&
908  RHS->getType()->getPointerAddressSpace())) {
909  // For most address spaces, an allocation can't be placed at null, but null
910  // itself is treated as a 0 size allocation in the in bounds rules. Thus,
911  // the only valid inbounds address derived from null, is null itself.
912  // Thus, we have four cases to consider:
913  // 1) Base == nullptr, Offset == 0 -> inbounds, null
914  // 2) Base == nullptr, Offset != 0 -> poison as the result is out of bounds
915  // 3) Base != nullptr, Offset == (-base) -> poison (crossing allocations)
916  // 4) Base != nullptr, Offset != (-base) -> nonnull (and possibly poison)
917  //
918  // (Note if we're indexing a type of size 0, that simply collapses into one
919  // of the buckets above.)
920  //
921  // In general, we're allowed to make values less poison (i.e. remove
922  // sources of full UB), so in this case, we just select between the two
923  // non-poison cases (1 and 4 above).
924  //
925  // For vectors, we apply the same reasoning on a per-lane basis.
926  auto *Base = GEPLHS->getPointerOperand();
927  if (GEPLHS->getType()->isVectorTy() && Base->getType()->isPointerTy()) {
928  int NumElts = GEPLHS->getType()->getVectorNumElements();
929  Base = Builder.CreateVectorSplat(NumElts, Base);
930  }
931  return new ICmpInst(Cond, Base,
933  cast<Constant>(RHS), Base->getType()));
934  } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
935  // If the base pointers are different, but the indices are the same, just
936  // compare the base pointer.
937  if (PtrBase != GEPRHS->getOperand(0)) {
938  bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
939  IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
940  GEPRHS->getOperand(0)->getType();
941  if (IndicesTheSame)
942  for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
943  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
944  IndicesTheSame = false;
945  break;
946  }
947 
948  // If all indices are the same, just compare the base pointers.
949  Type *BaseType = GEPLHS->getOperand(0)->getType();
950  if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
951  return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
952 
953  // If we're comparing GEPs with two base pointers that only differ in type
954  // and both GEPs have only constant indices or just one use, then fold
955  // the compare with the adjusted indices.
956  // FIXME: Support vector of pointers.
957  if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
958  (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
959  (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
960  PtrBase->stripPointerCasts() ==
961  GEPRHS->getOperand(0)->stripPointerCasts() &&
962  !GEPLHS->getType()->isVectorTy()) {
963  Value *LOffset = EmitGEPOffset(GEPLHS);
964  Value *ROffset = EmitGEPOffset(GEPRHS);
965 
966  // If we looked through an addrspacecast between different sized address
967  // spaces, the LHS and RHS pointers are different sized
968  // integers. Truncate to the smaller one.
969  Type *LHSIndexTy = LOffset->getType();
970  Type *RHSIndexTy = ROffset->getType();
971  if (LHSIndexTy != RHSIndexTy) {
972  if (LHSIndexTy->getPrimitiveSizeInBits() <
973  RHSIndexTy->getPrimitiveSizeInBits()) {
974  ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
975  } else
976  LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
977  }
978 
979  Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
980  LOffset, ROffset);
981  return replaceInstUsesWith(I, Cmp);
982  }
983 
984  // Otherwise, the base pointers are different and the indices are
985  // different. Try convert this to an indexed compare by looking through
986  // PHIs/casts.
987  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
988  }
989 
990  // If one of the GEPs has all zero indices, recurse.
991  // FIXME: Handle vector of pointers.
992  if (!GEPLHS->getType()->isVectorTy() && GEPLHS->hasAllZeroIndices())
993  return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
995 
996  // If the other GEP has all zero indices, recurse.
997  // FIXME: Handle vector of pointers.
998  if (!GEPRHS->getType()->isVectorTy() && GEPRHS->hasAllZeroIndices())
999  return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
1000 
1001  bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
1002  if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
1003  // If the GEPs only differ by one index, compare it.
1004  unsigned NumDifferences = 0; // Keep track of # differences.
1005  unsigned DiffOperand = 0; // The operand that differs.
1006  for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
1007  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
1008  Type *LHSType = GEPLHS->getOperand(i)->getType();
1009  Type *RHSType = GEPRHS->getOperand(i)->getType();
1010  // FIXME: Better support for vector of pointers.
1011  if (LHSType->getPrimitiveSizeInBits() !=
1012  RHSType->getPrimitiveSizeInBits() ||
1013  (GEPLHS->getType()->isVectorTy() &&
1014  (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) {
1015  // Irreconcilable differences.
1016  NumDifferences = 2;
1017  break;
1018  }
1019 
1020  if (NumDifferences++) break;
1021  DiffOperand = i;
1022  }
1023 
1024  if (NumDifferences == 0) // SAME GEP?
1025  return replaceInstUsesWith(I, // No comparison is needed here.
1027 
1028  else if (NumDifferences == 1 && GEPsInBounds) {
1029  Value *LHSV = GEPLHS->getOperand(DiffOperand);
1030  Value *RHSV = GEPRHS->getOperand(DiffOperand);
1031  // Make sure we do a signed comparison here.
1032  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
1033  }
1034  }
1035 
1036  // Only lower this if the icmp is the only user of the GEP or if we expect
1037  // the result to fold to a constant!
1038  if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
1039  (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
1040  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
1041  Value *L = EmitGEPOffset(GEPLHS);
1042  Value *R = EmitGEPOffset(GEPRHS);
1043  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
1044  }
1045  }
1046 
1047  // Try convert this to an indexed compare by looking through PHIs/casts as a
1048  // last resort.
1049  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1050 }
1051 
1052 Instruction *InstCombiner::foldAllocaCmp(ICmpInst &ICI,
1053  const AllocaInst *Alloca,
1054  const Value *Other) {
1055  assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1056 
1057  // It would be tempting to fold away comparisons between allocas and any
1058  // pointer not based on that alloca (e.g. an argument). However, even
1059  // though such pointers cannot alias, they can still compare equal.
1060  //
1061  // But LLVM doesn't specify where allocas get their memory, so if the alloca
1062  // doesn't escape we can argue that it's impossible to guess its value, and we
1063  // can therefore act as if any such guesses are wrong.
1064  //
1065  // The code below checks that the alloca doesn't escape, and that it's only
1066  // used in a comparison once (the current instruction). The
1067  // single-comparison-use condition ensures that we're trivially folding all
1068  // comparisons against the alloca consistently, and avoids the risk of
1069  // erroneously folding a comparison of the pointer with itself.
1070 
1071  unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1072 
1074  for (const Use &U : Alloca->uses()) {
1075  if (Worklist.size() >= MaxIter)
1076  return nullptr;
1077  Worklist.push_back(&U);
1078  }
1079 
1080  unsigned NumCmps = 0;
1081  while (!Worklist.empty()) {
1082  assert(Worklist.size() <= MaxIter);
1083  const Use *U = Worklist.pop_back_val();
1084  const Value *V = U->getUser();
1085  --MaxIter;
1086 
1087  if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1088  isa<SelectInst>(V)) {
1089  // Track the uses.
1090  } else if (isa<LoadInst>(V)) {
1091  // Loading from the pointer doesn't escape it.
1092  continue;
1093  } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1094  // Storing *to* the pointer is fine, but storing the pointer escapes it.
1095  if (SI->getValueOperand() == U->get())
1096  return nullptr;
1097  continue;
1098  } else if (isa<ICmpInst>(V)) {
1099  if (NumCmps++)
1100  return nullptr; // Found more than one cmp.
1101  continue;
1102  } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1103  switch (Intrin->getIntrinsicID()) {
1104  // These intrinsics don't escape or compare the pointer. Memset is safe
1105  // because we don't allow ptrtoint. Memcpy and memmove are safe because
1106  // we don't allow stores, so src cannot point to V.
1107  case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1108  case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1109  continue;
1110  default:
1111  return nullptr;
1112  }
1113  } else {
1114  return nullptr;
1115  }
1116  for (const Use &U : V->uses()) {
1117  if (Worklist.size() >= MaxIter)
1118  return nullptr;
1119  Worklist.push_back(&U);
1120  }
1121  }
1122 
1123  Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1124  return replaceInstUsesWith(
1125  ICI,
1127 }
1128 
1129 /// Fold "icmp pred (X+C), X".
1130 Instruction *InstCombiner::foldICmpAddOpConst(Value *X, const APInt &C,
1131  ICmpInst::Predicate Pred) {
1132  // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1133  // so the values can never be equal. Similarly for all other "or equals"
1134  // operators.
1135  assert(!!C && "C should not be zero!");
1136 
1137  // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1138  // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1139  // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1140  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1141  Constant *R = ConstantInt::get(X->getType(),
1142  APInt::getMaxValue(C.getBitWidth()) - C);
1143  return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1144  }
1145 
1146  // (X+1) >u X --> X <u (0-1) --> X != 255
1147  // (X+2) >u X --> X <u (0-2) --> X <u 254
1148  // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1149  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1150  return new ICmpInst(ICmpInst::ICMP_ULT, X,
1151  ConstantInt::get(X->getType(), -C));
1152 
1154 
1155  // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1156  // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1157  // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1158  // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1159  // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1160  // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1161  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1162  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1163  ConstantInt::get(X->getType(), SMax - C));
1164 
1165  // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1166  // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1167  // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1168  // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1169  // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1170  // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1171 
1172  assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1173  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1174  ConstantInt::get(X->getType(), SMax - (C - 1)));
1175 }
1176 
1177 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1178 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1179 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1180 Instruction *InstCombiner::foldICmpShrConstConst(ICmpInst &I, Value *A,
1181  const APInt &AP1,
1182  const APInt &AP2) {
1183  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1184 
1185  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1186  if (I.getPredicate() == I.ICMP_NE)
1187  Pred = CmpInst::getInversePredicate(Pred);
1188  return new ICmpInst(Pred, LHS, RHS);
1189  };
1190 
1191  // Don't bother doing any work for cases which InstSimplify handles.
1192  if (AP2.isNullValue())
1193  return nullptr;
1194 
1195  bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1196  if (IsAShr) {
1197  if (AP2.isAllOnesValue())
1198  return nullptr;
1199  if (AP2.isNegative() != AP1.isNegative())
1200  return nullptr;
1201  if (AP2.sgt(AP1))
1202  return nullptr;
1203  }
1204 
1205  if (!AP1)
1206  // 'A' must be large enough to shift out the highest set bit.
1207  return getICmp(I.ICMP_UGT, A,
1208  ConstantInt::get(A->getType(), AP2.logBase2()));
1209 
1210  if (AP1 == AP2)
1211  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1212 
1213  int Shift;
1214  if (IsAShr && AP1.isNegative())
1215  Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1216  else
1217  Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1218 
1219  if (Shift > 0) {
1220  if (IsAShr && AP1 == AP2.ashr(Shift)) {
1221  // There are multiple solutions if we are comparing against -1 and the LHS
1222  // of the ashr is not a power of two.
1223  if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1224  return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1225  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1226  } else if (AP1 == AP2.lshr(Shift)) {
1227  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1228  }
1229  }
1230 
1231  // Shifting const2 will never be equal to const1.
1232  // FIXME: This should always be handled by InstSimplify?
1233  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1234  return replaceInstUsesWith(I, TorF);
1235 }
1236 
1237 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1238 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1239 Instruction *InstCombiner::foldICmpShlConstConst(ICmpInst &I, Value *A,
1240  const APInt &AP1,
1241  const APInt &AP2) {
1242  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1243 
1244  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1245  if (I.getPredicate() == I.ICMP_NE)
1246  Pred = CmpInst::getInversePredicate(Pred);
1247  return new ICmpInst(Pred, LHS, RHS);
1248  };
1249 
1250  // Don't bother doing any work for cases which InstSimplify handles.
1251  if (AP2.isNullValue())
1252  return nullptr;
1253 
1254  unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1255 
1256  if (!AP1 && AP2TrailingZeros != 0)
1257  return getICmp(
1258  I.ICMP_UGE, A,
1259  ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1260 
1261  if (AP1 == AP2)
1262  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1263 
1264  // Get the distance between the lowest bits that are set.
1265  int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1266 
1267  if (Shift > 0 && AP2.shl(Shift) == AP1)
1268  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1269 
1270  // Shifting const2 will never be equal to const1.
1271  // FIXME: This should always be handled by InstSimplify?
1272  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1273  return replaceInstUsesWith(I, TorF);
1274 }
1275 
1276 /// The caller has matched a pattern of the form:
1277 /// I = icmp ugt (add (add A, B), CI2), CI1
1278 /// If this is of the form:
1279 /// sum = a + b
1280 /// if (sum+128 >u 255)
1281 /// Then replace it with llvm.sadd.with.overflow.i8.
1282 ///
1284  ConstantInt *CI2, ConstantInt *CI1,
1285  InstCombiner &IC) {
1286  // The transformation we're trying to do here is to transform this into an
1287  // llvm.sadd.with.overflow. To do this, we have to replace the original add
1288  // with a narrower add, and discard the add-with-constant that is part of the
1289  // range check (if we can't eliminate it, this isn't profitable).
1290 
1291  // In order to eliminate the add-with-constant, the compare can be its only
1292  // use.
1293  Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1294  if (!AddWithCst->hasOneUse())
1295  return nullptr;
1296 
1297  // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1298  if (!CI2->getValue().isPowerOf2())
1299  return nullptr;
1300  unsigned NewWidth = CI2->getValue().countTrailingZeros();
1301  if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1302  return nullptr;
1303 
1304  // The width of the new add formed is 1 more than the bias.
1305  ++NewWidth;
1306 
1307  // Check to see that CI1 is an all-ones value with NewWidth bits.
1308  if (CI1->getBitWidth() == NewWidth ||
1309  CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1310  return nullptr;
1311 
1312  // This is only really a signed overflow check if the inputs have been
1313  // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1314  // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1315  unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1316  if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1317  IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1318  return nullptr;
1319 
1320  // In order to replace the original add with a narrower
1321  // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1322  // and truncates that discard the high bits of the add. Verify that this is
1323  // the case.
1324  Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1325  for (User *U : OrigAdd->users()) {
1326  if (U == AddWithCst)
1327  continue;
1328 
1329  // Only accept truncates for now. We would really like a nice recursive
1330  // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1331  // chain to see which bits of a value are actually demanded. If the
1332  // original add had another add which was then immediately truncated, we
1333  // could still do the transformation.
1334  TruncInst *TI = dyn_cast<TruncInst>(U);
1335  if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1336  return nullptr;
1337  }
1338 
1339  // If the pattern matches, truncate the inputs to the narrower type and
1340  // use the sadd_with_overflow intrinsic to efficiently compute both the
1341  // result and the overflow bit.
1342  Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1344  I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1345 
1346  InstCombiner::BuilderTy &Builder = IC.Builder;
1347 
1348  // Put the new code above the original add, in case there are any uses of the
1349  // add between the add and the compare.
1350  Builder.SetInsertPoint(OrigAdd);
1351 
1352  Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1353  Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1354  CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1355  Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1356  Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1357 
1358  // The inner add was the result of the narrow add, zero extended to the
1359  // wider type. Replace it with the result computed by the intrinsic.
1360  IC.replaceInstUsesWith(*OrigAdd, ZExt);
1361 
1362  // The original icmp gets replaced with the overflow value.
1363  return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1364 }
1365 
1366 /// If we have:
1367 /// icmp eq/ne (urem/srem %x, %y), 0
1368 /// iff %y is a power-of-two, we can replace this with a bit test:
1369 /// icmp eq/ne (and %x, (add %y, -1)), 0
1370 Instruction *InstCombiner::foldIRemByPowerOfTwoToBitTest(ICmpInst &I) {
1371  // This fold is only valid for equality predicates.
1372  if (!I.isEquality())
1373  return nullptr;
1374  ICmpInst::Predicate Pred;
1375  Value *X, *Y, *Zero;
1376  if (!match(&I, m_ICmp(Pred, m_OneUse(m_IRem(m_Value(X), m_Value(Y))),
1377  m_CombineAnd(m_Zero(), m_Value(Zero)))))
1378  return nullptr;
1379  if (!isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, 0, &I))
1380  return nullptr;
1381  // This may increase instruction count, we don't enforce that Y is a constant.
1382  Value *Mask = Builder.CreateAdd(Y, Constant::getAllOnesValue(Y->getType()));
1383  Value *Masked = Builder.CreateAnd(X, Mask);
1384  return ICmpInst::Create(Instruction::ICmp, Pred, Masked, Zero);
1385 }
1386 
1387 // Handle icmp pred X, 0
1388 Instruction *InstCombiner::foldICmpWithZero(ICmpInst &Cmp) {
1389  CmpInst::Predicate Pred = Cmp.getPredicate();
1390  if (!match(Cmp.getOperand(1), m_Zero()))
1391  return nullptr;
1392 
1393  // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1394  if (Pred == ICmpInst::ICMP_SGT) {
1395  Value *A, *B;
1397  if (SPR.Flavor == SPF_SMIN) {
1398  if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1399  return new ICmpInst(Pred, B, Cmp.getOperand(1));
1400  if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1401  return new ICmpInst(Pred, A, Cmp.getOperand(1));
1402  }
1403  }
1404 
1405  if (Instruction *New = foldIRemByPowerOfTwoToBitTest(Cmp))
1406  return New;
1407 
1408  // Given:
1409  // icmp eq/ne (urem %x, %y), 0
1410  // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
1411  // icmp eq/ne %x, 0
1412  Value *X, *Y;
1413  if (match(Cmp.getOperand(0), m_URem(m_Value(X), m_Value(Y))) &&
1414  ICmpInst::isEquality(Pred)) {
1415  KnownBits XKnown = computeKnownBits(X, 0, &Cmp);
1416  KnownBits YKnown = computeKnownBits(Y, 0, &Cmp);
1417  if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
1418  return new ICmpInst(Pred, X, Cmp.getOperand(1));
1419  }
1420 
1421  return nullptr;
1422 }
1423 
1424 /// Fold icmp Pred X, C.
1425 /// TODO: This code structure does not make sense. The saturating add fold
1426 /// should be moved to some other helper and extended as noted below (it is also
1427 /// possible that code has been made unnecessary - do we canonicalize IR to
1428 /// overflow/saturating intrinsics or not?).
1429 Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) {
1430  // Match the following pattern, which is a common idiom when writing
1431  // overflow-safe integer arithmetic functions. The source performs an addition
1432  // in wider type and explicitly checks for overflow using comparisons against
1433  // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1434  //
1435  // TODO: This could probably be generalized to handle other overflow-safe
1436  // operations if we worked out the formulas to compute the appropriate magic
1437  // constants.
1438  //
1439  // sum = a + b
1440  // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1441  CmpInst::Predicate Pred = Cmp.getPredicate();
1442  Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1443  Value *A, *B;
1444  ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1445  if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1446  match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1447  if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1448  return Res;
1449 
1450  return nullptr;
1451 }
1452 
1453 /// Canonicalize icmp instructions based on dominating conditions.
1454 Instruction *InstCombiner::foldICmpWithDominatingICmp(ICmpInst &Cmp) {
1455  // This is a cheap/incomplete check for dominance - just match a single
1456  // predecessor with a conditional branch.
1457  BasicBlock *CmpBB = Cmp.getParent();
1458  BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1459  if (!DomBB)
1460  return nullptr;
1461 
1462  Value *DomCond;
1463  BasicBlock *TrueBB, *FalseBB;
1464  if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1465  return nullptr;
1466 
1467  assert((TrueBB == CmpBB || FalseBB == CmpBB) &&
1468  "Predecessor block does not point to successor?");
1469 
1470  // The branch should get simplified. Don't bother simplifying this condition.
1471  if (TrueBB == FalseBB)
1472  return nullptr;
1473 
1474  // Try to simplify this compare to T/F based on the dominating condition.
1475  Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1476  if (Imp)
1477  return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1478 
1479  CmpInst::Predicate Pred = Cmp.getPredicate();
1480  Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1481  ICmpInst::Predicate DomPred;
1482  const APInt *C, *DomC;
1483  if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1484  match(Y, m_APInt(C))) {
1485  // We have 2 compares of a variable with constants. Calculate the constant
1486  // ranges of those compares to see if we can transform the 2nd compare:
1487  // DomBB:
1488  // DomCond = icmp DomPred X, DomC
1489  // br DomCond, CmpBB, FalseBB
1490  // CmpBB:
1491  // Cmp = icmp Pred X, C
1493  ConstantRange DominatingCR =
1494  (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1496  CmpInst::getInversePredicate(DomPred), *DomC);
1497  ConstantRange Intersection = DominatingCR.intersectWith(CR);
1498  ConstantRange Difference = DominatingCR.difference(CR);
1499  if (Intersection.isEmptySet())
1500  return replaceInstUsesWith(Cmp, Builder.getFalse());
1501  if (Difference.isEmptySet())
1502  return replaceInstUsesWith(Cmp, Builder.getTrue());
1503 
1504  // Canonicalizing a sign bit comparison that gets used in a branch,
1505  // pessimizes codegen by generating branch on zero instruction instead
1506  // of a test and branch. So we avoid canonicalizing in such situations
1507  // because test and branch instruction has better branch displacement
1508  // than compare and branch instruction.
1509  bool UnusedBit;
1510  bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1511  if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1512  return nullptr;
1513 
1514  if (const APInt *EqC = Intersection.getSingleElement())
1515  return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1516  if (const APInt *NeC = Difference.getSingleElement())
1517  return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1518  }
1519 
1520  return nullptr;
1521 }
1522 
1523 /// Fold icmp (trunc X, Y), C.
1524 Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
1525  TruncInst *Trunc,
1526  const APInt &C) {
1527  ICmpInst::Predicate Pred = Cmp.getPredicate();
1528  Value *X = Trunc->getOperand(0);
1529  if (C.isOneValue() && C.getBitWidth() > 1) {
1530  // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1531  Value *V = nullptr;
1532  if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1533  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1534  ConstantInt::get(V->getType(), 1));
1535  }
1536 
1537  if (Cmp.isEquality() && Trunc->hasOneUse()) {
1538  // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1539  // of the high bits truncated out of x are known.
1540  unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1541  SrcBits = X->getType()->getScalarSizeInBits();
1542  KnownBits Known = computeKnownBits(X, 0, &Cmp);
1543 
1544  // If all the high bits are known, we can do this xform.
1545  if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1546  // Pull in the high bits from known-ones set.
1547  APInt NewRHS = C.zext(SrcBits);
1548  NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1549  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1550  }
1551  }
1552 
1553  return nullptr;
1554 }
1555 
1556 /// Fold icmp (xor X, Y), C.
1557 Instruction *InstCombiner::foldICmpXorConstant(ICmpInst &Cmp,
1558  BinaryOperator *Xor,
1559  const APInt &C) {
1560  Value *X = Xor->getOperand(0);
1561  Value *Y = Xor->getOperand(1);
1562  const APInt *XorC;
1563  if (!match(Y, m_APInt(XorC)))
1564  return nullptr;
1565 
1566  // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1567  // fold the xor.
1568  ICmpInst::Predicate Pred = Cmp.getPredicate();
1569  bool TrueIfSigned = false;
1570  if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1571 
1572  // If the sign bit of the XorCst is not set, there is no change to
1573  // the operation, just stop using the Xor.
1574  if (!XorC->isNegative()) {
1575  Cmp.setOperand(0, X);
1576  Worklist.Add(Xor);
1577  return &Cmp;
1578  }
1579 
1580  // Emit the opposite comparison.
1581  if (TrueIfSigned)
1582  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1584  else
1585  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1587  }
1588 
1589  if (Xor->hasOneUse()) {
1590  // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1591  if (!Cmp.isEquality() && XorC->isSignMask()) {
1592  Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1593  : Cmp.getSignedPredicate();
1594  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1595  }
1596 
1597  // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1598  if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1599  Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1600  : Cmp.getSignedPredicate();
1601  Pred = Cmp.getSwappedPredicate(Pred);
1602  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1603  }
1604  }
1605 
1606  // Mask constant magic can eliminate an 'xor' with unsigned compares.
1607  if (Pred == ICmpInst::ICMP_UGT) {
1608  // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1609  if (*XorC == ~C && (C + 1).isPowerOf2())
1610  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1611  // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1612  if (*XorC == C && (C + 1).isPowerOf2())
1613  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1614  }
1615  if (Pred == ICmpInst::ICMP_ULT) {
1616  // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1617  if (*XorC == -C && C.isPowerOf2())
1618  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1619  ConstantInt::get(X->getType(), ~C));
1620  // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1621  if (*XorC == C && (-C).isPowerOf2())
1622  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1623  ConstantInt::get(X->getType(), ~C));
1624  }
1625  return nullptr;
1626 }
1627 
1628 /// Fold icmp (and (sh X, Y), C2), C1.
1629 Instruction *InstCombiner::foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
1630  const APInt &C1, const APInt &C2) {
1631  BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1632  if (!Shift || !Shift->isShift())
1633  return nullptr;
1634 
1635  // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1636  // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1637  // code produced by the clang front-end, for bitfield access.
1638  // This seemingly simple opportunity to fold away a shift turns out to be
1639  // rather complicated. See PR17827 for details.
1640  unsigned ShiftOpcode = Shift->getOpcode();
1641  bool IsShl = ShiftOpcode == Instruction::Shl;
1642  const APInt *C3;
1643  if (match(Shift->getOperand(1), m_APInt(C3))) {
1644  bool CanFold = false;
1645  if (ShiftOpcode == Instruction::Shl) {
1646  // For a left shift, we can fold if the comparison is not signed. We can
1647  // also fold a signed comparison if the mask value and comparison value
1648  // are not negative. These constraints may not be obvious, but we can
1649  // prove that they are correct using an SMT solver.
1650  if (!Cmp.isSigned() || (!C2.isNegative() && !C1.isNegative()))
1651  CanFold = true;
1652  } else {
1653  bool IsAshr = ShiftOpcode == Instruction::AShr;
1654  // For a logical right shift, we can fold if the comparison is not signed.
1655  // We can also fold a signed comparison if the shifted mask value and the
1656  // shifted comparison value are not negative. These constraints may not be
1657  // obvious, but we can prove that they are correct using an SMT solver.
1658  // For an arithmetic shift right we can do the same, if we ensure
1659  // the And doesn't use any bits being shifted in. Normally these would
1660  // be turned into lshr by SimplifyDemandedBits, but not if there is an
1661  // additional user.
1662  if (!IsAshr || (C2.shl(*C3).lshr(*C3) == C2)) {
1663  if (!Cmp.isSigned() ||
1664  (!C2.shl(*C3).isNegative() && !C1.shl(*C3).isNegative()))
1665  CanFold = true;
1666  }
1667  }
1668 
1669  if (CanFold) {
1670  APInt NewCst = IsShl ? C1.lshr(*C3) : C1.shl(*C3);
1671  APInt SameAsC1 = IsShl ? NewCst.shl(*C3) : NewCst.lshr(*C3);
1672  // Check to see if we are shifting out any of the bits being compared.
1673  if (SameAsC1 != C1) {
1674  // If we shifted bits out, the fold is not going to work out. As a
1675  // special case, check to see if this means that the result is always
1676  // true or false now.
1677  if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1678  return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1679  if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1680  return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1681  } else {
1682  Cmp.setOperand(1, ConstantInt::get(And->getType(), NewCst));
1683  APInt NewAndCst = IsShl ? C2.lshr(*C3) : C2.shl(*C3);
1684  And->setOperand(1, ConstantInt::get(And->getType(), NewAndCst));
1685  And->setOperand(0, Shift->getOperand(0));
1686  Worklist.Add(Shift); // Shift is dead.
1687  return &Cmp;
1688  }
1689  }
1690  }
1691 
1692  // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1693  // preferable because it allows the C2 << Y expression to be hoisted out of a
1694  // loop if Y is invariant and X is not.
1695  if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() &&
1696  !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1697  // Compute C2 << Y.
1698  Value *NewShift =
1699  IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1700  : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1701 
1702  // Compute X & (C2 << Y).
1703  Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1704  Cmp.setOperand(0, NewAnd);
1705  return &Cmp;
1706  }
1707 
1708  return nullptr;
1709 }
1710 
1711 /// Fold icmp (and X, C2), C1.
1712 Instruction *InstCombiner::foldICmpAndConstConst(ICmpInst &Cmp,
1713  BinaryOperator *And,
1714  const APInt &C1) {
1715  bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1716 
1717  // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1718  // TODO: We canonicalize to the longer form for scalars because we have
1719  // better analysis/folds for icmp, and codegen may be better with icmp.
1720  if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isNullValue() &&
1721  match(And->getOperand(1), m_One()))
1722  return new TruncInst(And->getOperand(0), Cmp.getType());
1723 
1724  const APInt *C2;
1725  Value *X;
1726  if (!match(And, m_And(m_Value(X), m_APInt(C2))))
1727  return nullptr;
1728 
1729  // Don't perform the following transforms if the AND has multiple uses
1730  if (!And->hasOneUse())
1731  return nullptr;
1732 
1733  if (Cmp.isEquality() && C1.isNullValue()) {
1734  // Restrict this fold to single-use 'and' (PR10267).
1735  // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1736  if (C2->isSignMask()) {
1737  Constant *Zero = Constant::getNullValue(X->getType());
1738  auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1739  return new ICmpInst(NewPred, X, Zero);
1740  }
1741 
1742  // Restrict this fold only for single-use 'and' (PR10267).
1743  // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1744  if ((~(*C2) + 1).isPowerOf2()) {
1745  Constant *NegBOC =
1746  ConstantExpr::getNeg(cast<Constant>(And->getOperand(1)));
1747  auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1748  return new ICmpInst(NewPred, X, NegBOC);
1749  }
1750  }
1751 
1752  // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1753  // the input width without changing the value produced, eliminate the cast:
1754  //
1755  // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1756  //
1757  // We can do this transformation if the constants do not have their sign bits
1758  // set or if it is an equality comparison. Extending a relational comparison
1759  // when we're checking the sign bit would not work.
1760  Value *W;
1761  if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1762  (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1763  // TODO: Is this a good transform for vectors? Wider types may reduce
1764  // throughput. Should this transform be limited (even for scalars) by using
1765  // shouldChangeType()?
1766  if (!Cmp.getType()->isVectorTy()) {
1767  Type *WideType = W->getType();
1768  unsigned WideScalarBits = WideType->getScalarSizeInBits();
1769  Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1770  Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1771  Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1772  return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1773  }
1774  }
1775 
1776  if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1777  return I;
1778 
1779  // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1780  // (icmp pred (and A, (or (shl 1, B), 1), 0))
1781  //
1782  // iff pred isn't signed
1783  if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() &&
1784  match(And->getOperand(1), m_One())) {
1785  Constant *One = cast<Constant>(And->getOperand(1));
1786  Value *Or = And->getOperand(0);
1787  Value *A, *B, *LShr;
1788  if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1789  match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1790  unsigned UsesRemoved = 0;
1791  if (And->hasOneUse())
1792  ++UsesRemoved;
1793  if (Or->hasOneUse())
1794  ++UsesRemoved;
1795  if (LShr->hasOneUse())
1796  ++UsesRemoved;
1797 
1798  // Compute A & ((1 << B) | 1)
1799  Value *NewOr = nullptr;
1800  if (auto *C = dyn_cast<Constant>(B)) {
1801  if (UsesRemoved >= 1)
1802  NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1803  } else {
1804  if (UsesRemoved >= 3)
1805  NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1806  /*HasNUW=*/true),
1807  One, Or->getName());
1808  }
1809  if (NewOr) {
1810  Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1811  Cmp.setOperand(0, NewAnd);
1812  return &Cmp;
1813  }
1814  }
1815  }
1816 
1817  return nullptr;
1818 }
1819 
1820 /// Fold icmp (and X, Y), C.
1821 Instruction *InstCombiner::foldICmpAndConstant(ICmpInst &Cmp,
1822  BinaryOperator *And,
1823  const APInt &C) {
1824  if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1825  return I;
1826 
1827  // TODO: These all require that Y is constant too, so refactor with the above.
1828 
1829  // Try to optimize things like "A[i] & 42 == 0" to index computations.
1830  Value *X = And->getOperand(0);
1831  Value *Y = And->getOperand(1);
1832  if (auto *LI = dyn_cast<LoadInst>(X))
1833  if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1834  if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1835  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1836  !LI->isVolatile() && isa<ConstantInt>(Y)) {
1837  ConstantInt *C2 = cast<ConstantInt>(Y);
1838  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1839  return Res;
1840  }
1841 
1842  if (!Cmp.isEquality())
1843  return nullptr;
1844 
1845  // X & -C == -C -> X > u ~C
1846  // X & -C != -C -> X <= u ~C
1847  // iff C is a power of 2
1848  if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1849  auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1851  return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1852  }
1853 
1854  // (X & C2) == 0 -> (trunc X) >= 0
1855  // (X & C2) != 0 -> (trunc X) < 0
1856  // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1857  const APInt *C2;
1858  if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) {
1859  int32_t ExactLogBase2 = C2->exactLogBase2();
1860  if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1861  Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1862  if (And->getType()->isVectorTy())
1863  NTy = VectorType::get(NTy, And->getType()->getVectorNumElements());
1864  Value *Trunc = Builder.CreateTrunc(X, NTy);
1865  auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1867  return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1868  }
1869  }
1870 
1871  return nullptr;
1872 }
1873 
1874 /// Fold icmp (or X, Y), C.
1875 Instruction *InstCombiner::foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
1876  const APInt &C) {
1877  ICmpInst::Predicate Pred = Cmp.getPredicate();
1878  if (C.isOneValue()) {
1879  // icmp slt signum(V) 1 --> icmp slt V, 1
1880  Value *V = nullptr;
1881  if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1882  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1883  ConstantInt::get(V->getType(), 1));
1884  }
1885 
1886  Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
1887  if (Cmp.isEquality() && Cmp.getOperand(1) == OrOp1) {
1888  // X | C == C --> X <=u C
1889  // X | C != C --> X >u C
1890  // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1891  if ((C + 1).isPowerOf2()) {
1893  return new ICmpInst(Pred, OrOp0, OrOp1);
1894  }
1895  // More general: are all bits outside of a mask constant set or not set?
1896  // X | C == C --> (X & ~C) == 0
1897  // X | C != C --> (X & ~C) != 0
1898  if (Or->hasOneUse()) {
1899  Value *A = Builder.CreateAnd(OrOp0, ~C);
1900  return new ICmpInst(Pred, A, ConstantInt::getNullValue(OrOp0->getType()));
1901  }
1902  }
1903 
1904  if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse())
1905  return nullptr;
1906 
1907  Value *P, *Q;
1908  if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1909  // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1910  // -> and (icmp eq P, null), (icmp eq Q, null).
1911  Value *CmpP =
1912  Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1913  Value *CmpQ =
1914  Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1915  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1916  return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1917  }
1918 
1919  // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1920  // a shorter form that has more potential to be folded even further.
1921  Value *X1, *X2, *X3, *X4;
1922  if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1923  match(OrOp1, m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1924  // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1925  // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1926  Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1927  Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1928  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1929  return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1930  }
1931 
1932  return nullptr;
1933 }
1934 
1935 /// Fold icmp (mul X, Y), C.
1936 Instruction *InstCombiner::foldICmpMulConstant(ICmpInst &Cmp,
1937  BinaryOperator *Mul,
1938  const APInt &C) {
1939  const APInt *MulC;
1940  if (!match(Mul->getOperand(1), m_APInt(MulC)))
1941  return nullptr;
1942 
1943  // If this is a test of the sign bit and the multiply is sign-preserving with
1944  // a constant operand, use the multiply LHS operand instead.
1945  ICmpInst::Predicate Pred = Cmp.getPredicate();
1946  if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1947  if (MulC->isNegative())
1948  Pred = ICmpInst::getSwappedPredicate(Pred);
1949  return new ICmpInst(Pred, Mul->getOperand(0),
1951  }
1952 
1953  return nullptr;
1954 }
1955 
1956 /// Fold icmp (shl 1, Y), C.
1958  const APInt &C) {
1959  Value *Y;
1960  if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
1961  return nullptr;
1962 
1963  Type *ShiftType = Shl->getType();
1964  unsigned TypeBits = C.getBitWidth();
1965  bool CIsPowerOf2 = C.isPowerOf2();
1966  ICmpInst::Predicate Pred = Cmp.getPredicate();
1967  if (Cmp.isUnsigned()) {
1968  // (1 << Y) pred C -> Y pred Log2(C)
1969  if (!CIsPowerOf2) {
1970  // (1 << Y) < 30 -> Y <= 4
1971  // (1 << Y) <= 30 -> Y <= 4
1972  // (1 << Y) >= 30 -> Y > 4
1973  // (1 << Y) > 30 -> Y > 4
1974  if (Pred == ICmpInst::ICMP_ULT)
1975  Pred = ICmpInst::ICMP_ULE;
1976  else if (Pred == ICmpInst::ICMP_UGE)
1977  Pred = ICmpInst::ICMP_UGT;
1978  }
1979 
1980  // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
1981  // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
1982  unsigned CLog2 = C.logBase2();
1983  if (CLog2 == TypeBits - 1) {
1984  if (Pred == ICmpInst::ICMP_UGE)
1985  Pred = ICmpInst::ICMP_EQ;
1986  else if (Pred == ICmpInst::ICMP_ULT)
1987  Pred = ICmpInst::ICMP_NE;
1988  }
1989  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
1990  } else if (Cmp.isSigned()) {
1991  Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
1992  if (C.isAllOnesValue()) {
1993  // (1 << Y) <= -1 -> Y == 31
1994  if (Pred == ICmpInst::ICMP_SLE)
1995  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1996 
1997  // (1 << Y) > -1 -> Y != 31
1998  if (Pred == ICmpInst::ICMP_SGT)
1999  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2000  } else if (!C) {
2001  // (1 << Y) < 0 -> Y == 31
2002  // (1 << Y) <= 0 -> Y == 31
2003  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
2004  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2005 
2006  // (1 << Y) >= 0 -> Y != 31
2007  // (1 << Y) > 0 -> Y != 31
2008  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
2009  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2010  }
2011  } else if (Cmp.isEquality() && CIsPowerOf2) {
2012  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
2013  }
2014 
2015  return nullptr;
2016 }
2017 
2018 /// Fold icmp (shl X, Y), C.
2019 Instruction *InstCombiner::foldICmpShlConstant(ICmpInst &Cmp,
2020  BinaryOperator *Shl,
2021  const APInt &C) {
2022  const APInt *ShiftVal;
2023  if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
2024  return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
2025 
2026  const APInt *ShiftAmt;
2027  if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
2028  return foldICmpShlOne(Cmp, Shl, C);
2029 
2030  // Check that the shift amount is in range. If not, don't perform undefined
2031  // shifts. When the shift is visited, it will be simplified.
2032  unsigned TypeBits = C.getBitWidth();
2033  if (ShiftAmt->uge(TypeBits))
2034  return nullptr;
2035 
2036  ICmpInst::Predicate Pred = Cmp.getPredicate();
2037  Value *X = Shl->getOperand(0);
2038  Type *ShType = Shl->getType();
2039 
2040  // NSW guarantees that we are only shifting out sign bits from the high bits,
2041  // so we can ASHR the compare constant without needing a mask and eliminate
2042  // the shift.
2043  if (Shl->hasNoSignedWrap()) {
2044  if (Pred == ICmpInst::ICMP_SGT) {
2045  // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
2046  APInt ShiftedC = C.ashr(*ShiftAmt);
2047  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2048  }
2049  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2050  C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
2051  APInt ShiftedC = C.ashr(*ShiftAmt);
2052  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2053  }
2054  if (Pred == ICmpInst::ICMP_SLT) {
2055  // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
2056  // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2057  // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2058  // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2059  assert(!C.isMinSignedValue() && "Unexpected icmp slt");
2060  APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
2061  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2062  }
2063  // If this is a signed comparison to 0 and the shift is sign preserving,
2064  // use the shift LHS operand instead; isSignTest may change 'Pred', so only
2065  // do that if we're sure to not continue on in this function.
2066  if (isSignTest(Pred, C))
2067  return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
2068  }
2069 
2070  // NUW guarantees that we are only shifting out zero bits from the high bits,
2071  // so we can LSHR the compare constant without needing a mask and eliminate
2072  // the shift.
2073  if (Shl->hasNoUnsignedWrap()) {
2074  if (Pred == ICmpInst::ICMP_UGT) {
2075  // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2076  APInt ShiftedC = C.lshr(*ShiftAmt);
2077  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2078  }
2079  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2080  C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
2081  APInt ShiftedC = C.lshr(*ShiftAmt);
2082  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2083  }
2084  if (Pred == ICmpInst::ICMP_ULT) {
2085  // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2086  // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2087  // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2088  // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2089  assert(C.ugt(0) && "ult 0 should have been eliminated");
2090  APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
2091  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2092  }
2093  }
2094 
2095  if (Cmp.isEquality() && Shl->hasOneUse()) {
2096  // Strength-reduce the shift into an 'and'.
2098  ShType,
2099  APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2100  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2101  Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
2102  return new ICmpInst(Pred, And, LShrC);
2103  }
2104 
2105  // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2106  bool TrueIfSigned = false;
2107  if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
2108  // (X << 31) <s 0 --> (X & 1) != 0
2110  ShType,
2111  APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2112  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2113  return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2114  And, Constant::getNullValue(ShType));
2115  }
2116 
2117  // Simplify 'shl' inequality test into 'and' equality test.
2118  if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2119  // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2120  if ((C + 1).isPowerOf2() &&
2121  (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2122  Value *And = Builder.CreateAnd(X, (~C).lshr(ShiftAmt->getZExtValue()));
2123  return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2125  And, Constant::getNullValue(ShType));
2126  }
2127  // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2128  if (C.isPowerOf2() &&
2129  (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2130  Value *And =
2131  Builder.CreateAnd(X, (~(C - 1)).lshr(ShiftAmt->getZExtValue()));
2132  return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2134  And, Constant::getNullValue(ShType));
2135  }
2136  }
2137 
2138  // Transform (icmp pred iM (shl iM %v, N), C)
2139  // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2140  // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2141  // This enables us to get rid of the shift in favor of a trunc that may be
2142  // free on the target. It has the additional benefit of comparing to a
2143  // smaller constant that may be more target-friendly.
2144  unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2145  if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2146  DL.isLegalInteger(TypeBits - Amt)) {
2147  Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2148  if (ShType->isVectorTy())
2149  TruncTy = VectorType::get(TruncTy, ShType->getVectorNumElements());
2150  Constant *NewC =
2151  ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2152  return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2153  }
2154 
2155  return nullptr;
2156 }
2157 
2158 /// Fold icmp ({al}shr X, Y), C.
2159 Instruction *InstCombiner::foldICmpShrConstant(ICmpInst &Cmp,
2160  BinaryOperator *Shr,
2161  const APInt &C) {
2162  // An exact shr only shifts out zero bits, so:
2163  // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2164  Value *X = Shr->getOperand(0);
2165  CmpInst::Predicate Pred = Cmp.getPredicate();
2166  if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2167  C.isNullValue())
2168  return new ICmpInst(Pred, X, Cmp.getOperand(1));
2169 
2170  const APInt *ShiftVal;
2171  if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2172  return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2173 
2174  const APInt *ShiftAmt;
2175  if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2176  return nullptr;
2177 
2178  // Check that the shift amount is in range. If not, don't perform undefined
2179  // shifts. When the shift is visited it will be simplified.
2180  unsigned TypeBits = C.getBitWidth();
2181  unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2182  if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2183  return nullptr;
2184 
2185  bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2186  bool IsExact = Shr->isExact();
2187  Type *ShrTy = Shr->getType();
2188  // TODO: If we could guarantee that InstSimplify would handle all of the
2189  // constant-value-based preconditions in the folds below, then we could assert
2190  // those conditions rather than checking them. This is difficult because of
2191  // undef/poison (PR34838).
2192  if (IsAShr) {
2193  if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2194  // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2195  // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2196  APInt ShiftedC = C.shl(ShAmtVal);
2197  if (ShiftedC.ashr(ShAmtVal) == C)
2198  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2199  }
2200  if (Pred == CmpInst::ICMP_SGT) {
2201  // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2202  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2203  if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2204  (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2205  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2206  }
2207  } else {
2208  if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2209  // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2210  // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2211  APInt ShiftedC = C.shl(ShAmtVal);
2212  if (ShiftedC.lshr(ShAmtVal) == C)
2213  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2214  }
2215  if (Pred == CmpInst::ICMP_UGT) {
2216  // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2217  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2218  if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2219  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2220  }
2221  }
2222 
2223  if (!Cmp.isEquality())
2224  return nullptr;
2225 
2226  // Handle equality comparisons of shift-by-constant.
2227 
2228  // If the comparison constant changes with the shift, the comparison cannot
2229  // succeed (bits of the comparison constant cannot match the shifted value).
2230  // This should be known by InstSimplify and already be folded to true/false.
2231  assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2232  (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2233  "Expected icmp+shr simplify did not occur.");
2234 
2235  // If the bits shifted out are known zero, compare the unshifted value:
2236  // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2237  if (Shr->isExact())
2238  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2239 
2240  if (Shr->hasOneUse()) {
2241  // Canonicalize the shift into an 'and':
2242  // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2243  APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2244  Constant *Mask = ConstantInt::get(ShrTy, Val);
2245  Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2246  return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2247  }
2248 
2249  return nullptr;
2250 }
2251 
2252 Instruction *InstCombiner::foldICmpSRemConstant(ICmpInst &Cmp,
2253  BinaryOperator *SRem,
2254  const APInt &C) {
2255  // Match an 'is positive' or 'is negative' comparison of remainder by a
2256  // constant power-of-2 value:
2257  // (X % pow2C) sgt/slt 0
2258  const ICmpInst::Predicate Pred = Cmp.getPredicate();
2259  if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT)
2260  return nullptr;
2261 
2262  // TODO: The one-use check is standard because we do not typically want to
2263  // create longer instruction sequences, but this might be a special-case
2264  // because srem is not good for analysis or codegen.
2265  if (!SRem->hasOneUse())
2266  return nullptr;
2267 
2268  const APInt *DivisorC;
2269  if (!C.isNullValue() || !match(SRem->getOperand(1), m_Power2(DivisorC)))
2270  return nullptr;
2271 
2272  // Mask off the sign bit and the modulo bits (low-bits).
2273  Type *Ty = SRem->getType();
2274  APInt SignMask = APInt::getSignMask(Ty->getScalarSizeInBits());
2275  Constant *MaskC = ConstantInt::get(Ty, SignMask | (*DivisorC - 1));
2276  Value *And = Builder.CreateAnd(SRem->getOperand(0), MaskC);
2277 
2278  // For 'is positive?' check that the sign-bit is clear and at least 1 masked
2279  // bit is set. Example:
2280  // (i8 X % 32) s> 0 --> (X & 159) s> 0
2281  if (Pred == ICmpInst::ICMP_SGT)
2283 
2284  // For 'is negative?' check that the sign-bit is set and at least 1 masked
2285  // bit is set. Example:
2286  // (i16 X % 4) s< 0 --> (X & 32771) u> 32768
2287  return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, SignMask));
2288 }
2289 
2290 /// Fold icmp (udiv X, Y), C.
2291 Instruction *InstCombiner::foldICmpUDivConstant(ICmpInst &Cmp,
2292  BinaryOperator *UDiv,
2293  const APInt &C) {
2294  const APInt *C2;
2295  if (!match(UDiv->getOperand(0), m_APInt(C2)))
2296  return nullptr;
2297 
2298  assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2299 
2300  // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2301  Value *Y = UDiv->getOperand(1);
2302  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2303  assert(!C.isMaxValue() &&
2304  "icmp ugt X, UINT_MAX should have been simplified already.");
2305  return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2306  ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2307  }
2308 
2309  // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2310  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2311  assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2312  return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2313  ConstantInt::get(Y->getType(), C2->udiv(C)));
2314  }
2315 
2316  return nullptr;
2317 }
2318 
2319 /// Fold icmp ({su}div X, Y), C.
2320 Instruction *InstCombiner::foldICmpDivConstant(ICmpInst &Cmp,
2321  BinaryOperator *Div,
2322  const APInt &C) {
2323  // Fold: icmp pred ([us]div X, C2), C -> range test
2324  // Fold this div into the comparison, producing a range check.
2325  // Determine, based on the divide type, what the range is being
2326  // checked. If there is an overflow on the low or high side, remember
2327  // it, otherwise compute the range [low, hi) bounding the new value.
2328  // See: InsertRangeTest above for the kinds of replacements possible.
2329  const APInt *C2;
2330  if (!match(Div->getOperand(1), m_APInt(C2)))
2331  return nullptr;
2332 
2333  // FIXME: If the operand types don't match the type of the divide
2334  // then don't attempt this transform. The code below doesn't have the
2335  // logic to deal with a signed divide and an unsigned compare (and
2336  // vice versa). This is because (x /s C2) <s C produces different
2337  // results than (x /s C2) <u C or (x /u C2) <s C or even
2338  // (x /u C2) <u C. Simply casting the operands and result won't
2339  // work. :( The if statement below tests that condition and bails
2340  // if it finds it.
2341  bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2342  if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2343  return nullptr;
2344 
2345  // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2346  // INT_MIN will also fail if the divisor is 1. Although folds of all these
2347  // division-by-constant cases should be present, we can not assert that they
2348  // have happened before we reach this icmp instruction.
2349  if (C2->isNullValue() || C2->isOneValue() ||
2350  (DivIsSigned && C2->isAllOnesValue()))
2351  return nullptr;
2352 
2353  // Compute Prod = C * C2. We are essentially solving an equation of
2354  // form X / C2 = C. We solve for X by multiplying C2 and C.
2355  // By solving for X, we can turn this into a range check instead of computing
2356  // a divide.
2357  APInt Prod = C * *C2;
2358 
2359  // Determine if the product overflows by seeing if the product is not equal to
2360  // the divide. Make sure we do the same kind of divide as in the LHS
2361  // instruction that we're folding.
2362  bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2363 
2364  ICmpInst::Predicate Pred = Cmp.getPredicate();
2365 
2366  // If the division is known to be exact, then there is no remainder from the
2367  // divide, so the covered range size is unit, otherwise it is the divisor.
2368  APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2369 
2370  // Figure out the interval that is being checked. For example, a comparison
2371  // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2372  // Compute this interval based on the constants involved and the signedness of
2373  // the compare/divide. This computes a half-open interval, keeping track of
2374  // whether either value in the interval overflows. After analysis each
2375  // overflow variable is set to 0 if it's corresponding bound variable is valid
2376  // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2377  int LoOverflow = 0, HiOverflow = 0;
2378  APInt LoBound, HiBound;
2379 
2380  if (!DivIsSigned) { // udiv
2381  // e.g. X/5 op 3 --> [15, 20)
2382  LoBound = Prod;
2383  HiOverflow = LoOverflow = ProdOV;
2384  if (!HiOverflow) {
2385  // If this is not an exact divide, then many values in the range collapse
2386  // to the same result value.
2387  HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2388  }
2389  } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2390  if (C.isNullValue()) { // (X / pos) op 0
2391  // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2392  LoBound = -(RangeSize - 1);
2393  HiBound = RangeSize;
2394  } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2395  LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2396  HiOverflow = LoOverflow = ProdOV;
2397  if (!HiOverflow)
2398  HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2399  } else { // (X / pos) op neg
2400  // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2401  HiBound = Prod + 1;
2402  LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2403  if (!LoOverflow) {
2404  APInt DivNeg = -RangeSize;
2405  LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2406  }
2407  }
2408  } else if (C2->isNegative()) { // Divisor is < 0.
2409  if (Div->isExact())
2410  RangeSize.negate();
2411  if (C.isNullValue()) { // (X / neg) op 0
2412  // e.g. X/-5 op 0 --> [-4, 5)
2413  LoBound = RangeSize + 1;
2414  HiBound = -RangeSize;
2415  if (HiBound == *C2) { // -INTMIN = INTMIN
2416  HiOverflow = 1; // [INTMIN+1, overflow)
2417  HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2418  }
2419  } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2420  // e.g. X/-5 op 3 --> [-19, -14)
2421  HiBound = Prod + 1;
2422  HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2423  if (!LoOverflow)
2424  LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2425  } else { // (X / neg) op neg
2426  LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2427  LoOverflow = HiOverflow = ProdOV;
2428  if (!HiOverflow)
2429  HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2430  }
2431 
2432  // Dividing by a negative swaps the condition. LT <-> GT
2433  Pred = ICmpInst::getSwappedPredicate(Pred);
2434  }
2435 
2436  Value *X = Div->getOperand(0);
2437  switch (Pred) {
2438  default: llvm_unreachable("Unhandled icmp opcode!");
2439  case ICmpInst::ICMP_EQ:
2440  if (LoOverflow && HiOverflow)
2441  return replaceInstUsesWith(Cmp, Builder.getFalse());
2442  if (HiOverflow)
2443  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2444  ICmpInst::ICMP_UGE, X,
2445  ConstantInt::get(Div->getType(), LoBound));
2446  if (LoOverflow)
2447  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2448  ICmpInst::ICMP_ULT, X,
2449  ConstantInt::get(Div->getType(), HiBound));
2450  return replaceInstUsesWith(
2451  Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2452  case ICmpInst::ICMP_NE:
2453  if (LoOverflow && HiOverflow)
2454  return replaceInstUsesWith(Cmp, Builder.getTrue());
2455  if (HiOverflow)
2456  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2457  ICmpInst::ICMP_ULT, X,
2458  ConstantInt::get(Div->getType(), LoBound));
2459  if (LoOverflow)
2460  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2461  ICmpInst::ICMP_UGE, X,
2462  ConstantInt::get(Div->getType(), HiBound));
2463  return replaceInstUsesWith(Cmp,
2464  insertRangeTest(X, LoBound, HiBound,
2465  DivIsSigned, false));
2466  case ICmpInst::ICMP_ULT:
2467  case ICmpInst::ICMP_SLT:
2468  if (LoOverflow == +1) // Low bound is greater than input range.
2469  return replaceInstUsesWith(Cmp, Builder.getTrue());
2470  if (LoOverflow == -1) // Low bound is less than input range.
2471  return replaceInstUsesWith(Cmp, Builder.getFalse());
2472  return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2473  case ICmpInst::ICMP_UGT:
2474  case ICmpInst::ICMP_SGT:
2475  if (HiOverflow == +1) // High bound greater than input range.
2476  return replaceInstUsesWith(Cmp, Builder.getFalse());
2477  if (HiOverflow == -1) // High bound less than input range.
2478  return replaceInstUsesWith(Cmp, Builder.getTrue());
2479  if (Pred == ICmpInst::ICMP_UGT)
2480  return new ICmpInst(ICmpInst::ICMP_UGE, X,
2481  ConstantInt::get(Div->getType(), HiBound));
2482  return new ICmpInst(ICmpInst::ICMP_SGE, X,
2483  ConstantInt::get(Div->getType(), HiBound));
2484  }
2485 
2486  return nullptr;
2487 }
2488 
2489 /// Fold icmp (sub X, Y), C.
2490 Instruction *InstCombiner::foldICmpSubConstant(ICmpInst &Cmp,
2491  BinaryOperator *Sub,
2492  const APInt &C) {
2493  Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2494  ICmpInst::Predicate Pred = Cmp.getPredicate();
2495  const APInt *C2;
2496  APInt SubResult;
2497 
2498  // icmp eq/ne (sub C, Y), C -> icmp eq/ne Y, 0
2499  if (match(X, m_APInt(C2)) && *C2 == C && Cmp.isEquality())
2500  return new ICmpInst(Cmp.getPredicate(), Y,
2501  ConstantInt::get(Y->getType(), 0));
2502 
2503  // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2504  if (match(X, m_APInt(C2)) &&
2505  ((Cmp.isUnsigned() && Sub->hasNoUnsignedWrap()) ||
2506  (Cmp.isSigned() && Sub->hasNoSignedWrap())) &&
2507  !subWithOverflow(SubResult, *C2, C, Cmp.isSigned()))
2508  return new ICmpInst(Cmp.getSwappedPredicate(), Y,
2509  ConstantInt::get(Y->getType(), SubResult));
2510 
2511  // The following transforms are only worth it if the only user of the subtract
2512  // is the icmp.
2513  if (!Sub->hasOneUse())
2514  return nullptr;
2515 
2516  if (Sub->hasNoSignedWrap()) {
2517  // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2518  if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue())
2519  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2520 
2521  // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2522  if (Pred == ICmpInst::ICMP_SGT && C.isNullValue())
2523  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2524 
2525  // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2526  if (Pred == ICmpInst::ICMP_SLT && C.isNullValue())
2527  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2528 
2529  // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2530  if (Pred == ICmpInst::ICMP_SLT && C.isOneValue())
2531  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2532  }
2533 
2534  if (!match(X, m_APInt(C2)))
2535  return nullptr;
2536 
2537  // C2 - Y <u C -> (Y | (C - 1)) == C2
2538  // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2539  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2540  (*C2 & (C - 1)) == (C - 1))
2541  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2542 
2543  // C2 - Y >u C -> (Y | C) != C2
2544  // iff C2 & C == C and C + 1 is a power of 2
2545  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2546  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2547 
2548  return nullptr;
2549 }
2550 
2551 /// Fold icmp (add X, Y), C.
2552 Instruction *InstCombiner::foldICmpAddConstant(ICmpInst &Cmp,
2554  const APInt &C) {
2555  Value *Y = Add->getOperand(1);
2556  const APInt *C2;
2557  if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2558  return nullptr;
2559 
2560  // Fold icmp pred (add X, C2), C.
2561  Value *X = Add->getOperand(0);
2562  Type *Ty = Add->getType();
2563  CmpInst::Predicate Pred = Cmp.getPredicate();
2564 
2565  if (!Add->hasOneUse())
2566  return nullptr;
2567 
2568  // If the add does not wrap, we can always adjust the compare by subtracting
2569  // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2570  // are canonicalized to SGT/SLT/UGT/ULT.
2571  if ((Add->hasNoSignedWrap() &&
2572  (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2573  (Add->hasNoUnsignedWrap() &&
2574  (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2575  bool Overflow;
2576  APInt NewC =
2577  Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2578  // If there is overflow, the result must be true or false.
2579  // TODO: Can we assert there is no overflow because InstSimplify always
2580  // handles those cases?
2581  if (!Overflow)
2582  // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2583  return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2584  }
2585 
2586  auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2587  const APInt &Upper = CR.getUpper();
2588  const APInt &Lower = CR.getLower();
2589  if (Cmp.isSigned()) {
2590  if (Lower.isSignMask())
2591  return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2592  if (Upper.isSignMask())
2593  return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2594  } else {
2595  if (Lower.isMinValue())
2596  return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2597  if (Upper.isMinValue())
2598  return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2599  }
2600 
2601  // X+C <u C2 -> (X & -C2) == C
2602  // iff C & (C2-1) == 0
2603  // C2 is a power of 2
2604  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2605  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2606  ConstantExpr::getNeg(cast<Constant>(Y)));
2607 
2608  // X+C >u C2 -> (X & ~C2) != C
2609  // iff C & C2 == 0
2610  // C2+1 is a power of 2
2611  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2612  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2613  ConstantExpr::getNeg(cast<Constant>(Y)));
2614 
2615  return nullptr;
2616 }
2617 
2618 bool InstCombiner::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2619  Value *&RHS, ConstantInt *&Less,
2620  ConstantInt *&Equal,
2621  ConstantInt *&Greater) {
2622  // TODO: Generalize this to work with other comparison idioms or ensure
2623  // they get canonicalized into this form.
2624 
2625  // select i1 (a == b),
2626  // i32 Equal,
2627  // i32 (select i1 (a < b), i32 Less, i32 Greater)
2628  // where Equal, Less and Greater are placeholders for any three constants.
2629  ICmpInst::Predicate PredA;
2630  if (!match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) ||
2631  !ICmpInst::isEquality(PredA))
2632  return false;
2633  Value *EqualVal = SI->getTrueValue();
2634  Value *UnequalVal = SI->getFalseValue();
2635  // We still can get non-canonical predicate here, so canonicalize.
2636  if (PredA == ICmpInst::ICMP_NE)
2637  std::swap(EqualVal, UnequalVal);
2638  if (!match(EqualVal, m_ConstantInt(Equal)))
2639  return false;
2640  ICmpInst::Predicate PredB;
2641  Value *LHS2, *RHS2;
2642  if (!match(UnequalVal, m_Select(m_ICmp(PredB, m_Value(LHS2), m_Value(RHS2)),
2643  m_ConstantInt(Less), m_ConstantInt(Greater))))
2644  return false;
2645  // We can get predicate mismatch here, so canonicalize if possible:
2646  // First, ensure that 'LHS' match.
2647  if (LHS2 != LHS) {
2648  // x sgt y <--> y slt x
2649  std::swap(LHS2, RHS2);
2650  PredB = ICmpInst::getSwappedPredicate(PredB);
2651  }
2652  if (LHS2 != LHS)
2653  return false;
2654  // We also need to canonicalize 'RHS'.
2655  if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(RHS2)) {
2656  // x sgt C-1 <--> x sge C <--> not(x slt C)
2657  auto FlippedStrictness =
2658  getFlippedStrictnessPredicateAndConstant(PredB, cast<Constant>(RHS2));
2659  if (!FlippedStrictness)
2660  return false;
2661  assert(FlippedStrictness->first == ICmpInst::ICMP_SGE && "Sanity check");
2662  RHS2 = FlippedStrictness->second;
2663  // And kind-of perform the result swap.
2664  std::swap(Less, Greater);
2665  PredB = ICmpInst::ICMP_SLT;
2666  }
2667  return PredB == ICmpInst::ICMP_SLT && RHS == RHS2;
2668 }
2669 
2670 Instruction *InstCombiner::foldICmpSelectConstant(ICmpInst &Cmp,
2671  SelectInst *Select,
2672  ConstantInt *C) {
2673 
2674  assert(C && "Cmp RHS should be a constant int!");
2675  // If we're testing a constant value against the result of a three way
2676  // comparison, the result can be expressed directly in terms of the
2677  // original values being compared. Note: We could possibly be more
2678  // aggressive here and remove the hasOneUse test. The original select is
2679  // really likely to simplify or sink when we remove a test of the result.
2680  Value *OrigLHS, *OrigRHS;
2681  ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2682  if (Cmp.hasOneUse() &&
2683  matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2684  C3GreaterThan)) {
2685  assert(C1LessThan && C2Equal && C3GreaterThan);
2686 
2687  bool TrueWhenLessThan =
2688  ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2689  ->isAllOnesValue();
2690  bool TrueWhenEqual =
2691  ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2692  ->isAllOnesValue();
2693  bool TrueWhenGreaterThan =
2694  ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2695  ->isAllOnesValue();
2696 
2697  // This generates the new instruction that will replace the original Cmp
2698  // Instruction. Instead of enumerating the various combinations when
2699  // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2700  // false, we rely on chaining of ORs and future passes of InstCombine to
2701  // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2702 
2703  // When none of the three constants satisfy the predicate for the RHS (C),
2704  // the entire original Cmp can be simplified to a false.
2705  Value *Cond = Builder.getFalse();
2706  if (TrueWhenLessThan)
2707  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
2708  OrigLHS, OrigRHS));
2709  if (TrueWhenEqual)
2710  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
2711  OrigLHS, OrigRHS));
2712  if (TrueWhenGreaterThan)
2713  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
2714  OrigLHS, OrigRHS));
2715 
2716  return replaceInstUsesWith(Cmp, Cond);
2717  }
2718  return nullptr;
2719 }
2720 
2722  InstCombiner::BuilderTy &Builder) {
2723  auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
2724  if (!Bitcast)
2725  return nullptr;
2726 
2727  ICmpInst::Predicate Pred = Cmp.getPredicate();
2728  Value *Op1 = Cmp.getOperand(1);
2729  Value *BCSrcOp = Bitcast->getOperand(0);
2730 
2731  // Make sure the bitcast doesn't change the number of vector elements.
2732  if (Bitcast->getSrcTy()->getScalarSizeInBits() ==
2733  Bitcast->getDestTy()->getScalarSizeInBits()) {
2734  // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
2735  Value *X;
2736  if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
2737  // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
2738  // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
2739  // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
2740  // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
2741  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
2742  Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
2743  match(Op1, m_Zero()))
2744  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2745 
2746  // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
2747  if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
2748  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
2749 
2750  // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
2751  if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
2752  return new ICmpInst(Pred, X,
2754  }
2755 
2756  // Zero-equality checks are preserved through unsigned floating-point casts:
2757  // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
2758  // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
2759  if (match(BCSrcOp, m_UIToFP(m_Value(X))))
2760  if (Cmp.isEquality() && match(Op1, m_Zero()))
2761  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2762  }
2763 
2764  // Test to see if the operands of the icmp are casted versions of other
2765  // values. If the ptr->ptr cast can be stripped off both arguments, do so.
2766  if (Bitcast->getType()->isPointerTy() &&
2767  (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2768  // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2769  // so eliminate it as well.
2770  if (auto *BC2 = dyn_cast<BitCastInst>(Op1))
2771  Op1 = BC2->getOperand(0);
2772 
2773  Op1 = Builder.CreateBitCast(Op1, BCSrcOp->getType());
2774  return new ICmpInst(Pred, BCSrcOp, Op1);
2775  }
2776 
2777  // Folding: icmp <pred> iN X, C
2778  // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2779  // and C is a splat of a K-bit pattern
2780  // and SC is a constant vector = <C', C', C', ..., C'>
2781  // Into:
2782  // %E = extractelement <M x iK> %vec, i32 C'
2783  // icmp <pred> iK %E, trunc(C)
2784  const APInt *C;
2785  if (!match(Cmp.getOperand(1), m_APInt(C)) ||
2786  !Bitcast->getType()->isIntegerTy() ||
2787  !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2788  return nullptr;
2789 
2790  Value *Vec;
2791  Constant *Mask;
2792  if (match(BCSrcOp,
2793  m_ShuffleVector(m_Value(Vec), m_Undef(), m_Constant(Mask)))) {
2794  // Check whether every element of Mask is the same constant
2795  if (auto *Elem = dyn_cast_or_null<ConstantInt>(Mask->getSplatValue())) {
2796  auto *VecTy = cast<VectorType>(BCSrcOp->getType());
2797  auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2798  if (C->isSplat(EltTy->getBitWidth())) {
2799  // Fold the icmp based on the value of C
2800  // If C is M copies of an iK sized bit pattern,
2801  // then:
2802  // => %E = extractelement <N x iK> %vec, i32 Elem
2803  // icmp <pred> iK %SplatVal, <pattern>
2804  Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2805  Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
2806  return new ICmpInst(Pred, Extract, NewC);
2807  }
2808  }
2809  }
2810  return nullptr;
2811 }
2812 
2813 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2814 /// where X is some kind of instruction.
2815 Instruction *InstCombiner::foldICmpInstWithConstant(ICmpInst &Cmp) {
2816  const APInt *C;
2817  if (!match(Cmp.getOperand(1), m_APInt(C)))
2818  return nullptr;
2819 
2820  if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2821  switch (BO->getOpcode()) {
2822  case Instruction::Xor:
2823  if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
2824  return I;
2825  break;
2826  case Instruction::And:
2827  if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
2828  return I;
2829  break;
2830  case Instruction::Or:
2831  if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
2832  return I;
2833  break;
2834  case Instruction::Mul:
2835  if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
2836  return I;
2837  break;
2838  case Instruction::Shl:
2839  if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
2840  return I;
2841  break;
2842  case Instruction::LShr:
2843  case Instruction::AShr:
2844  if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
2845  return I;
2846  break;
2847  case Instruction::SRem:
2848  if (Instruction *I = foldICmpSRemConstant(Cmp, BO, *C))
2849  return I;
2850  break;
2851  case Instruction::UDiv:
2852  if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
2853  return I;
2855  case Instruction::SDiv:
2856  if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
2857  return I;
2858  break;
2859  case Instruction::Sub:
2860  if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
2861  return I;
2862  break;
2863  case Instruction::Add:
2864  if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
2865  return I;
2866  break;
2867  default:
2868  break;
2869  }
2870  // TODO: These folds could be refactored to be part of the above calls.
2871  if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
2872  return I;
2873  }
2874 
2875  // Match against CmpInst LHS being instructions other than binary operators.
2876 
2877  if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
2878  // For now, we only support constant integers while folding the
2879  // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2880  // similar to the cases handled by binary ops above.
2881  if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2882  if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
2883  return I;
2884  }
2885 
2886  if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
2887  if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
2888  return I;
2889  }
2890 
2891  if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
2892  if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
2893  return I;
2894 
2895  return nullptr;
2896 }
2897 
2898 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2899 /// icmp eq/ne BO, C.
2900 Instruction *InstCombiner::foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
2901  BinaryOperator *BO,
2902  const APInt &C) {
2903  // TODO: Some of these folds could work with arbitrary constants, but this
2904  // function is limited to scalar and vector splat constants.
2905  if (!Cmp.isEquality())
2906  return nullptr;
2907 
2908  ICmpInst::Predicate Pred = Cmp.getPredicate();
2909  bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2910  Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2911  Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2912 
2913  switch (BO->getOpcode()) {
2914  case Instruction::SRem:
2915  // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2916  if (C.isNullValue() && BO->hasOneUse()) {
2917  const APInt *BOC;
2918  if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
2919  Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
2920  return new ICmpInst(Pred, NewRem,
2922  }
2923  }
2924  break;
2925  case Instruction::Add: {
2926  // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
2927  const APInt *BOC;
2928  if (match(BOp1, m_APInt(BOC))) {
2929  if (BO->hasOneUse()) {
2930  Constant *SubC = ConstantExpr::getSub(RHS, cast<Constant>(BOp1));
2931  return new ICmpInst(Pred, BOp0, SubC);
2932  }
2933  } else if (C.isNullValue()) {
2934  // Replace ((add A, B) != 0) with (A != -B) if A or B is
2935  // efficiently invertible, or if the add has just this one use.
2936  if (Value *NegVal = dyn_castNegVal(BOp1))
2937  return new ICmpInst(Pred, BOp0, NegVal);
2938  if (Value *NegVal = dyn_castNegVal(BOp0))
2939  return new ICmpInst(Pred, NegVal, BOp1);
2940  if (BO->hasOneUse()) {
2941  Value *Neg = Builder.CreateNeg(BOp1);
2942  Neg->takeName(BO);
2943  return new ICmpInst(Pred, BOp0, Neg);
2944  }
2945  }
2946  break;
2947  }
2948  case Instruction::Xor:
2949  if (BO->hasOneUse()) {
2950  if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
2951  // For the xor case, we can xor two constants together, eliminating
2952  // the explicit xor.
2953  return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
2954  } else if (C.isNullValue()) {
2955  // Replace ((xor A, B) != 0) with (A != B)
2956  return new ICmpInst(Pred, BOp0, BOp1);
2957  }
2958  }
2959  break;
2960  case Instruction::Sub:
2961  if (BO->hasOneUse()) {
2962  const APInt *BOC;
2963  if (match(BOp0, m_APInt(BOC))) {
2964  // Replace ((sub BOC, B) != C) with (B != BOC-C).
2965  Constant *SubC = ConstantExpr::getSub(cast<Constant>(BOp0), RHS);
2966  return new ICmpInst(Pred, BOp1, SubC);
2967  } else if (C.isNullValue()) {
2968  // Replace ((sub A, B) != 0) with (A != B).
2969  return new ICmpInst(Pred, BOp0, BOp1);
2970  }
2971  }
2972  break;
2973  case Instruction::Or: {
2974  const APInt *BOC;
2975  if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
2976  // Comparing if all bits outside of a constant mask are set?
2977  // Replace (X | C) == -1 with (X & ~C) == ~C.
2978  // This removes the -1 constant.
2979  Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
2980  Value *And = Builder.CreateAnd(BOp0, NotBOC);
2981  return new ICmpInst(Pred, And, NotBOC);
2982  }
2983  break;
2984  }
2985  case Instruction::And: {
2986  const APInt *BOC;
2987  if (match(BOp1, m_APInt(BOC))) {
2988  // If we have ((X & C) == C), turn it into ((X & C) != 0).
2989  if (C == *BOC && C.isPowerOf2())
2990  return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
2991  BO, Constant::getNullValue(RHS->getType()));
2992  }
2993  break;
2994  }
2995  case Instruction::Mul:
2996  if (C.isNullValue() && BO->hasNoSignedWrap()) {
2997  const APInt *BOC;
2998  if (match(BOp1, m_APInt(BOC)) && !BOC->isNullValue()) {
2999  // The trivial case (mul X, 0) is handled by InstSimplify.
3000  // General case : (mul X, C) != 0 iff X != 0
3001  // (mul X, C) == 0 iff X == 0
3002  return new ICmpInst(Pred, BOp0, Constant::getNullValue(RHS->getType()));
3003  }
3004  }
3005  break;
3006  case Instruction::UDiv:
3007  if (C.isNullValue()) {
3008  // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
3009  auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3010  return new ICmpInst(NewPred, BOp1, BOp0);
3011  }
3012  break;
3013  default:
3014  break;
3015  }
3016  return nullptr;
3017 }
3018 
3019 /// Fold an equality icmp with LLVM intrinsic and constant operand.
3020 Instruction *InstCombiner::foldICmpEqIntrinsicWithConstant(ICmpInst &Cmp,
3021  IntrinsicInst *II,
3022  const APInt &C) {
3023  Type *Ty = II->getType();
3024  unsigned BitWidth = C.getBitWidth();
3025  switch (II->getIntrinsicID()) {
3026  case Intrinsic::bswap:
3027  Worklist.Add(II);
3028  Cmp.setOperand(0, II->getArgOperand(0));
3029  Cmp.setOperand(1, ConstantInt::get(Ty, C.byteSwap()));
3030  return &Cmp;
3031 
3032  case Intrinsic::ctlz:
3033  case Intrinsic::cttz: {
3034  // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
3035  if (C == BitWidth) {
3036  Worklist.Add(II);
3037  Cmp.setOperand(0, II->getArgOperand(0));
3039  return &Cmp;
3040  }
3041 
3042  // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
3043  // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
3044  // Limit to one use to ensure we don't increase instruction count.
3045  unsigned Num = C.getLimitedValue(BitWidth);
3046  if (Num != BitWidth && II->hasOneUse()) {
3047  bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
3048  APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
3049  : APInt::getHighBitsSet(BitWidth, Num + 1);
3050  APInt Mask2 = IsTrailing
3051  ? APInt::getOneBitSet(BitWidth, Num)
3052  : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3053  Cmp.setOperand(0, Builder.CreateAnd(II->getArgOperand(0), Mask1));
3054  Cmp.setOperand(1, ConstantInt::get(Ty, Mask2));
3055  Worklist.Add(II);
3056  return &Cmp;
3057  }
3058  break;
3059  }
3060 
3061  case Intrinsic::ctpop: {
3062  // popcount(A) == 0 -> A == 0 and likewise for !=
3063  // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
3064  bool IsZero = C.isNullValue();
3065  if (IsZero || C == BitWidth) {
3066  Worklist.Add(II);
3067  Cmp.setOperand(0, II->getArgOperand(0));
3068  auto *NewOp =
3070  Cmp.setOperand(1, NewOp);
3071  return &Cmp;
3072  }
3073  break;
3074  }
3075  default:
3076  break;
3077  }
3078 
3079  return nullptr;
3080 }
3081 
3082 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
3083 Instruction *InstCombiner::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
3084  IntrinsicInst *II,
3085  const APInt &C) {
3086  if (Cmp.isEquality())
3087  return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
3088 
3089  Type *Ty = II->getType();
3090  unsigned BitWidth = C.getBitWidth();
3091  switch (II->getIntrinsicID()) {
3092  case Intrinsic::ctlz: {
3093  // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
3094  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3095  unsigned Num = C.getLimitedValue();
3096  APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3097  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
3098  II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3099  }
3100 
3101  // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
3102  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
3103  C.uge(1) && C.ule(BitWidth)) {
3104  unsigned Num = C.getLimitedValue();
3105  APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
3106  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
3107  II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3108  }
3109  break;
3110  }
3111  case Intrinsic::cttz: {
3112  // Limit to one use to ensure we don't increase instruction count.
3113  if (!II->hasOneUse())
3114  return nullptr;
3115 
3116  // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
3117  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3118  APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
3119  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
3120  Builder.CreateAnd(II->getArgOperand(0), Mask),
3122  }
3123 
3124  // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
3125  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT &&
3126  C.uge(1) && C.ule(BitWidth)) {
3128  return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
3129  Builder.CreateAnd(II->getArgOperand(0), Mask),
3131  }
3132  break;
3133  }
3134  default:
3135  break;
3136  }
3137 
3138  return nullptr;
3139 }
3140 
3141 /// Handle icmp with constant (but not simple integer constant) RHS.
3142 Instruction *InstCombiner::foldICmpInstWithConstantNotInt(ICmpInst &I) {
3143  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3144  Constant *RHSC = dyn_cast<Constant>(Op1);
3145  Instruction *LHSI = dyn_cast<Instruction>(Op0);
3146  if (!RHSC || !LHSI)
3147  return nullptr;
3148 
3149  switch (LHSI->getOpcode()) {
3150  case Instruction::GetElementPtr:
3151  // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3152  if (RHSC->isNullValue() &&
3153  cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3154  return new ICmpInst(
3155  I.getPredicate(), LHSI->getOperand(0),
3156  Constant::getNullValue(LHSI->getOperand(0)->getType()));
3157  break;
3158  case Instruction::PHI:
3159  // Only fold icmp into the PHI if the phi and icmp are in the same
3160  // block. If in the same block, we're encouraging jump threading. If
3161  // not, we are just pessimizing the code by making an i1 phi.
3162  if (LHSI->getParent() == I.getParent())
3163  if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
3164  return NV;
3165  break;
3166  case Instruction::Select: {
3167  // If either operand of the select is a constant, we can fold the
3168  // comparison into the select arms, which will cause one to be
3169  // constant folded and the select turned into a bitwise or.
3170  Value *Op1 = nullptr, *Op2 = nullptr;
3171  ConstantInt *CI = nullptr;
3172  if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
3173  Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3174  CI = dyn_cast<ConstantInt>(Op1);
3175  }
3176  if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
3177  Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3178  CI = dyn_cast<ConstantInt>(Op2);
3179  }
3180 
3181  // We only want to perform this transformation if it will not lead to
3182  // additional code. This is true if either both sides of the select
3183  // fold to a constant (in which case the icmp is replaced with a select
3184  // which will usually simplify) or this is the only user of the
3185  // select (in which case we are trading a select+icmp for a simpler
3186  // select+icmp) or all uses of the select can be replaced based on
3187  // dominance information ("Global cases").
3188  bool Transform = false;
3189  if (Op1 && Op2)
3190  Transform = true;
3191  else if (Op1 || Op2) {
3192  // Local case
3193  if (LHSI->hasOneUse())
3194  Transform = true;
3195  // Global cases
3196  else if (CI && !CI->isZero())
3197  // When Op1 is constant try replacing select with second operand.
3198  // Otherwise Op2 is constant and try replacing select with first
3199  // operand.
3200  Transform =
3201  replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
3202  }
3203  if (Transform) {
3204  if (!Op1)
3205  Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
3206  I.getName());
3207  if (!Op2)
3208  Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
3209  I.getName());
3210  return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3211  }
3212  break;
3213  }
3214  case Instruction::IntToPtr:
3215  // icmp pred inttoptr(X), null -> icmp pred X, 0
3216  if (RHSC->isNullValue() &&
3217  DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3218  return new ICmpInst(
3219  I.getPredicate(), LHSI->getOperand(0),
3220  Constant::getNullValue(LHSI->getOperand(0)->getType()));
3221  break;
3222 
3223  case Instruction::Load:
3224  // Try to optimize things like "A[i] > 4" to index computations.
3225  if (GetElementPtrInst *GEP =
3226  dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3227  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3228  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3229  !cast<LoadInst>(LHSI)->isVolatile())
3230  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
3231  return Res;
3232  }
3233  break;
3234  }
3235 
3236  return nullptr;
3237 }
3238 
3239 /// Some comparisons can be simplified.
3240 /// In this case, we are looking for comparisons that look like
3241 /// a check for a lossy truncation.
3242 /// Folds:
3243 /// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
3244 /// Where Mask is some pattern that produces all-ones in low bits:
3245 /// (-1 >> y)
3246 /// ((-1 << y) >> y) <- non-canonical, has extra uses
3247 /// ~(-1 << y)
3248 /// ((1 << y) + (-1)) <- non-canonical, has extra uses
3249 /// The Mask can be a constant, too.
3250 /// For some predicates, the operands are commutative.
3251 /// For others, x can only be on a specific side.
3253  InstCombiner::BuilderTy &Builder) {
3254  ICmpInst::Predicate SrcPred;
3255  Value *X, *M, *Y;
3256  auto m_VariableMask = m_CombineOr(
3258  m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
3260  m_LShr(m_Shl(m_AllOnes(), m_Value(Y)), m_Deferred(Y))));
3261  auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
3262  if (!match(&I, m_c_ICmp(SrcPred,
3263  m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)),
3264  m_Deferred(X))))
3265  return nullptr;
3266 
3267  ICmpInst::Predicate DstPred;
3268  switch (SrcPred) {
3269  case ICmpInst::Predicate::ICMP_EQ:
3270  // x & (-1 >> y) == x -> x u<= (-1 >> y)
3271  DstPred = ICmpInst::Predicate::ICMP_ULE;
3272  break;
3273  case ICmpInst::Predicate::ICMP_NE:
3274  // x & (-1 >> y) != x -> x u> (-1 >> y)
3275  DstPred = ICmpInst::Predicate::ICMP_UGT;
3276  break;
3277  case ICmpInst::Predicate::ICMP_UGT:
3278  // x u> x & (-1 >> y) -> x u> (-1 >> y)
3279  assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
3280  DstPred = ICmpInst::Predicate::ICMP_UGT;
3281  break;
3282  case ICmpInst::Predicate::ICMP_UGE:
3283  // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
3284  assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
3285  DstPred = ICmpInst::Predicate::ICMP_ULE;
3286  break;
3287  case ICmpInst::Predicate::ICMP_ULT:
3288  // x & (-1 >> y) u< x -> x u> (-1 >> y)
3289  assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
3290  DstPred = ICmpInst::Predicate::ICMP_UGT;
3291  break;
3292  case ICmpInst::Predicate::ICMP_ULE:
3293  // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
3294  assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
3295  DstPred = ICmpInst::Predicate::ICMP_ULE;
3296  break;
3297  case ICmpInst::Predicate::ICMP_SGT:
3298  // x s> x & (-1 >> y) -> x s> (-1 >> y)
3299  if (X != I.getOperand(0)) // X must be on LHS of comparison!
3300  return nullptr; // Ignore the other case.
3301  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3302  return nullptr;
3303  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3304  return nullptr;
3305  DstPred = ICmpInst::Predicate::ICMP_SGT;
3306  break;
3307  case ICmpInst::Predicate::ICMP_SGE:
3308  // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
3309  if (X != I.getOperand(1)) // X must be on RHS of comparison!
3310  return nullptr; // Ignore the other case.
3311  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3312  return nullptr;
3313  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3314  return nullptr;
3315  DstPred = ICmpInst::Predicate::ICMP_SLE;
3316  break;
3317  case ICmpInst::Predicate::ICMP_SLT:
3318  // x & (-1 >> y) s< x -> x s> (-1 >> y)
3319  if (X != I.getOperand(1)) // X must be on RHS of comparison!
3320  return nullptr; // Ignore the other case.
3321  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3322  return nullptr;
3323  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3324  return nullptr;
3325  DstPred = ICmpInst::Predicate::ICMP_SGT;
3326  break;
3327  case ICmpInst::Predicate::ICMP_SLE:
3328  // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3329  if (X != I.getOperand(0)) // X must be on LHS of comparison!
3330  return nullptr; // Ignore the other case.
3331  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3332  return nullptr;
3333  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3334  return nullptr;
3335  DstPred = ICmpInst::Predicate::ICMP_SLE;
3336  break;
3337  default:
3338  llvm_unreachable("All possible folds are handled.");
3339  }
3340 
3341  return Builder.CreateICmp(DstPred, X, M);
3342 }
3343 
3344 /// Some comparisons can be simplified.
3345 /// In this case, we are looking for comparisons that look like
3346 /// a check for a lossy signed truncation.
3347 /// Folds: (MaskedBits is a constant.)
3348 /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3349 /// Into:
3350 /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3351 /// Where KeptBits = bitwidth(%x) - MaskedBits
3352 static Value *
3354  InstCombiner::BuilderTy &Builder) {
3355  ICmpInst::Predicate SrcPred;
3356  Value *X;
3357  const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3358  // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3359  if (!match(&I, m_c_ICmp(SrcPred,
3360  m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
3361  m_APInt(C1))),
3362  m_Deferred(X))))
3363  return nullptr;
3364 
3365  // Potential handling of non-splats: for each element:
3366  // * if both are undef, replace with constant 0.
3367  // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3368  // * if both are not undef, and are different, bailout.
3369  // * else, only one is undef, then pick the non-undef one.
3370 
3371  // The shift amount must be equal.
3372  if (*C0 != *C1)
3373  return nullptr;
3374  const APInt &MaskedBits = *C0;
3375  assert(MaskedBits != 0 && "shift by zero should be folded away already.");
3376 
3377  ICmpInst::Predicate DstPred;
3378  switch (SrcPred) {
3379  case ICmpInst::Predicate::ICMP_EQ:
3380  // ((%x << MaskedBits) a>> MaskedBits) == %x
3381  // =>
3382  // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3383  DstPred = ICmpInst::Predicate::ICMP_ULT;
3384  break;
3385  case ICmpInst::Predicate::ICMP_NE:
3386  // ((%x << MaskedBits) a>> MaskedBits) != %x
3387  // =>
3388  // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3389  DstPred = ICmpInst::Predicate::ICMP_UGE;
3390  break;
3391  // FIXME: are more folds possible?
3392  default:
3393  return nullptr;
3394  }
3395 
3396  auto *XType = X->getType();
3397  const unsigned XBitWidth = XType->getScalarSizeInBits();
3398  const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3399  assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
3400 
3401  // KeptBits = bitwidth(%x) - MaskedBits
3402  const APInt KeptBits = BitWidth - MaskedBits;
3403  assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
3404  // ICmpCst = (1 << KeptBits)
3405  const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3406  assert(ICmpCst.isPowerOf2());
3407  // AddCst = (1 << (KeptBits-1))
3408  const APInt AddCst = ICmpCst.lshr(1);
3409  assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
3410 
3411  // T0 = add %x, AddCst
3412  Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3413  // T1 = T0 DstPred ICmpCst
3414  Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3415 
3416  return T1;
3417 }
3418 
3419 // Given pattern:
3420 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3421 // we should move shifts to the same hand of 'and', i.e. rewrite as
3422 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3423 // We are only interested in opposite logical shifts here.
3424 // One of the shifts can be truncated.
3425 // If we can, we want to end up creating 'lshr' shift.
3426 static Value *
3428  InstCombiner::BuilderTy &Builder) {
3429  if (!I.isEquality() || !match(I.getOperand(1), m_Zero()) ||
3430  !I.getOperand(0)->hasOneUse())
3431  return nullptr;
3432 
3433  auto m_AnyLogicalShift = m_LogicalShift(m_Value(), m_Value());
3434 
3435  // Look for an 'and' of two logical shifts, one of which may be truncated.
3436  // We use m_TruncOrSelf() on the RHS to correctly handle commutative case.
3437  Instruction *XShift, *MaybeTruncation, *YShift;
3438  if (!match(
3439  I.getOperand(0),
3440  m_c_And(m_CombineAnd(m_AnyLogicalShift, m_Instruction(XShift)),
3442  m_AnyLogicalShift, m_Instruction(YShift))),
3443  m_Instruction(MaybeTruncation)))))
3444  return nullptr;
3445 
3446  // We potentially looked past 'trunc', but only when matching YShift,
3447  // therefore YShift must have the widest type.
3448  Instruction *WidestShift = YShift;
3449  // Therefore XShift must have the shallowest type.
3450  // Or they both have identical types if there was no truncation.
3451  Instruction *NarrowestShift = XShift;
3452 
3453  Type *WidestTy = WidestShift->getType();
3454  assert(NarrowestShift->getType() == I.getOperand(0)->getType() &&
3455  "We did not look past any shifts while matching XShift though.");
3456  bool HadTrunc = WidestTy != I.getOperand(0)->getType();
3457 
3458  // If YShift is a 'lshr', swap the shifts around.
3459  if (match(YShift, m_LShr(m_Value(), m_Value())))
3460  std::swap(XShift, YShift);
3461 
3462  // The shifts must be in opposite directions.
3463  auto XShiftOpcode = XShift->getOpcode();
3464  if (XShiftOpcode == YShift->getOpcode())
3465  return nullptr; // Do not care about same-direction shifts here.
3466 
3467  Value *X, *XShAmt, *Y, *YShAmt;
3468  match(XShift, m_BinOp(m_Value(X), m_ZExtOrSelf(m_Value(XShAmt))));
3469  match(YShift, m_BinOp(m_Value(Y), m_ZExtOrSelf(m_Value(YShAmt))));
3470 
3471  // If one of the values being shifted is a constant, then we will end with
3472  // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not,
3473  // however, we will need to ensure that we won't increase instruction count.
3474  if (!isa<Constant>(X) && !isa<Constant>(Y)) {
3475  // At least one of the hands of the 'and' should be one-use shift.
3476  if (!match(I.getOperand(0),
3477  m_c_And(m_OneUse(m_AnyLogicalShift), m_Value())))
3478  return nullptr;
3479  if (HadTrunc) {
3480  // Due to the 'trunc', we will need to widen X. For that either the old
3481  // 'trunc' or the shift amt in the non-truncated shift should be one-use.
3482  if (!MaybeTruncation->hasOneUse() &&
3483  !NarrowestShift->getOperand(1)->hasOneUse())
3484  return nullptr;
3485  }
3486  }
3487 
3488  // We have two shift amounts from two different shifts. The types of those
3489  // shift amounts may not match. If that's the case let's bailout now.
3490  if (XShAmt->getType() != YShAmt->getType())
3491  return nullptr;
3492 
3493  // Can we fold (XShAmt+YShAmt) ?
3494  auto *NewShAmt = dyn_cast_or_null<Constant>(
3495  SimplifyAddInst(XShAmt, YShAmt, /*isNSW=*/false,
3496  /*isNUW=*/false, SQ.getWithInstruction(&I)));
3497  if (!NewShAmt)
3498  return nullptr;
3499  NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, WidestTy);
3500  unsigned WidestBitWidth = WidestTy->getScalarSizeInBits();
3501 
3502  // Is the new shift amount smaller than the bit width?
3503  // FIXME: could also rely on ConstantRange.
3504  if (!match(NewShAmt,
3505  m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
3506  APInt(WidestBitWidth, WidestBitWidth))))
3507  return nullptr;
3508 
3509  // An extra legality check is needed if we had trunc-of-lshr.
3510  if (HadTrunc && match(WidestShift, m_LShr(m_Value(), m_Value()))) {
3511  auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ,
3512  WidestShift]() {
3513  // It isn't obvious whether it's worth it to analyze non-constants here.
3514  // Also, let's basically give up on non-splat cases, pessimizing vectors.
3515  // If *any* of these preconditions matches we can perform the fold.
3516  Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy()
3517  ? NewShAmt->getSplatValue()
3518  : NewShAmt;
3519  // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold.
3520  if (NewShAmtSplat &&
3521  (NewShAmtSplat->isNullValue() ||
3522  NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1))
3523  return true;
3524  // We consider *min* leading zeros so a single outlier
3525  // blocks the transform as opposed to allowing it.
3526  if (auto *C = dyn_cast<Constant>(NarrowestShift->getOperand(0))) {
3527  KnownBits Known = computeKnownBits(C, SQ.DL);
3528  unsigned MinLeadZero = Known.countMinLeadingZeros();
3529  // If the value being shifted has at most lowest bit set we can fold.
3530  unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3531  if (MaxActiveBits <= 1)
3532  return true;
3533  // Precondition: NewShAmt u<= countLeadingZeros(C)
3534  if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(MinLeadZero))
3535  return true;
3536  }
3537  if (auto *C = dyn_cast<Constant>(WidestShift->getOperand(0))) {
3538  KnownBits Known = computeKnownBits(C, SQ.DL);
3539  unsigned MinLeadZero = Known.countMinLeadingZeros();
3540  // If the value being shifted has at most lowest bit set we can fold.
3541  unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3542  if (MaxActiveBits <= 1)
3543  return true;
3544  // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C)
3545  if (NewShAmtSplat) {
3546  APInt AdjNewShAmt =
3547  (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger();
3548  if (AdjNewShAmt.ule(MinLeadZero))
3549  return true;
3550  }
3551  }
3552  return false; // Can't tell if it's ok.
3553  };
3554  if (!CanFold())
3555  return nullptr;
3556  }
3557 
3558  // All good, we can do this fold.
3559  X = Builder.CreateZExt(X, WidestTy);
3560  Y = Builder.CreateZExt(Y, WidestTy);
3561  // The shift is the same that was for X.
3562  Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
3563  ? Builder.CreateLShr(X, NewShAmt)
3564  : Builder.CreateShl(X, NewShAmt);
3565  Value *T1 = Builder.CreateAnd(T0, Y);
3566  return Builder.CreateICmp(I.getPredicate(), T1,
3567  Constant::getNullValue(WidestTy));
3568 }
3569 
3570 /// Fold
3571 /// (-1 u/ x) u< y
3572 /// ((x * y) u/ x) != y
3573 /// to
3574 /// @llvm.umul.with.overflow(x, y) plus extraction of overflow bit
3575 /// Note that the comparison is commutative, while inverted (u>=, ==) predicate
3576 /// will mean that we are looking for the opposite answer.
3577 Value *InstCombiner::foldUnsignedMultiplicationOverflowCheck(ICmpInst &I) {
3578  ICmpInst::Predicate Pred;
3579  Value *X, *Y;
3580  Instruction *Mul;
3581  bool NeedNegation;
3582  // Look for: (-1 u/ x) u</u>= y
3583  if (!I.isEquality() &&
3584  match(&I, m_c_ICmp(Pred, m_OneUse(m_UDiv(m_AllOnes(), m_Value(X))),
3585  m_Value(Y)))) {
3586  Mul = nullptr;
3587  // Canonicalize as-if y was on RHS.
3588  if (I.getOperand(1) != Y)
3589  Pred = I.getSwappedPredicate();
3590 
3591  // Are we checking that overflow does not happen, or does happen?
3592  switch (Pred) {
3593  case ICmpInst::Predicate::ICMP_ULT:
3594  NeedNegation = false;
3595  break; // OK
3596  case ICmpInst::Predicate::ICMP_UGE:
3597  NeedNegation = true;
3598  break; // OK
3599  default:
3600  return nullptr; // Wrong predicate.
3601  }
3602  } else // Look for: ((x * y) u/ x) !=/== y
3603  if (I.isEquality() &&
3604  match(&I, m_c_ICmp(Pred, m_Value(Y),
3606  m_Value(X)),
3607  m_Instruction(Mul)),
3608  m_Deferred(X)))))) {
3609  NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ;
3610  } else
3611  return nullptr;
3612 
3613  BuilderTy::InsertPointGuard Guard(Builder);
3614  // If the pattern included (x * y), we'll want to insert new instructions
3615  // right before that original multiplication so that we can replace it.
3616  bool MulHadOtherUses = Mul && !Mul->hasOneUse();
3617  if (MulHadOtherUses)
3618  Builder.SetInsertPoint(Mul);
3619 
3621  I.getModule(), Intrinsic::umul_with_overflow, X->getType());
3622  CallInst *Call = Builder.CreateCall(F, {X, Y}, "umul");
3623 
3624  // If the multiplication was used elsewhere, to ensure that we don't leave
3625  // "duplicate" instructions, replace uses of that original multiplication
3626  // with the multiplication result from the with.overflow intrinsic.
3627  if (MulHadOtherUses)
3628  replaceInstUsesWith(*Mul, Builder.CreateExtractValue(Call, 0, "umul.val"));
3629 
3630  Value *Res = Builder.CreateExtractValue(Call, 1, "umul.ov");
3631  if (NeedNegation) // This technically increases instruction count.
3632  Res = Builder.CreateNot(Res, "umul.not.ov");
3633 
3634  return Res;
3635 }
3636 
3637 /// Try to fold icmp (binop), X or icmp X, (binop).
3638 /// TODO: A large part of this logic is duplicated in InstSimplify's
3639 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3640 /// duplication.
3641 Instruction *InstCombiner::foldICmpBinOp(ICmpInst &I) {
3642  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3643 
3644  // Special logic for binary operators.
3645  BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3646  BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3647  if (!BO0 && !BO1)
3648  return nullptr;
3649 
3650  const CmpInst::Predicate Pred = I.getPredicate();
3651  Value *X;
3652 
3653  // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3654  // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X
3655  if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
3656  (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3657  return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3658  // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0
3659  if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
3660  (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3661  return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3662 
3663  bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3664  if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3665  NoOp0WrapProblem =
3666  ICmpInst::isEquality(Pred) ||
3667  (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3668  (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3669  if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3670  NoOp1WrapProblem =
3671  ICmpInst::isEquality(Pred) ||
3672  (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3673  (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3674 
3675  // Analyze the case when either Op0 or Op1 is an add instruction.
3676  // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3677  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3678  if (BO0 && BO0->getOpcode() == Instruction::Add) {
3679  A = BO0->getOperand(0);
3680  B = BO0->getOperand(1);
3681  }
3682  if (BO1 && BO1->getOpcode() == Instruction::Add) {
3683  C = BO1->getOperand(0);
3684  D = BO1->getOperand(1);
3685  }
3686 
3687  // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3688  if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3689  return new ICmpInst(Pred, A == Op1 ? B : A,
3690  Constant::getNullValue(Op1->getType()));
3691 
3692  // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3693  if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3694  return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3695  C == Op0 ? D : C);
3696 
3697  // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
3698  if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
3699  NoOp1WrapProblem &&
3700  // Try not to increase register pressure.
3701  BO0->hasOneUse() && BO1->hasOneUse()) {
3702  // Determine Y and Z in the form icmp (X+Y), (X+Z).
3703  Value *Y, *Z;
3704  if (A == C) {
3705  // C + B == C + D -> B == D
3706  Y = B;
3707  Z = D;
3708  } else if (A == D) {
3709  // D + B == C + D -> B == C
3710  Y = B;
3711  Z = C;
3712  } else if (B == C) {
3713  // A + C == C + D -> A == D
3714  Y = A;
3715  Z = D;
3716  } else {
3717  assert(B == D);
3718  // A + D == C + D -> A == C
3719  Y = A;
3720  Z = C;
3721  }
3722  return new ICmpInst(Pred, Y, Z);
3723  }
3724 
3725  // icmp slt (X + -1), Y -> icmp sle X, Y
3726  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3727  match(B, m_AllOnes()))
3728  return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3729 
3730  // icmp sge (X + -1), Y -> icmp sgt X, Y
3731  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3732  match(B, m_AllOnes()))
3733  return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3734 
3735  // icmp sle (X + 1), Y -> icmp slt X, Y
3736  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
3737  return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3738 
3739  // icmp sgt (X + 1), Y -> icmp sge X, Y
3740  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
3741  return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3742 
3743  // icmp sgt X, (Y + -1) -> icmp sge X, Y
3744  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
3745  match(D, m_AllOnes()))
3746  return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
3747 
3748  // icmp sle X, (Y + -1) -> icmp slt X, Y
3749  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
3750  match(D, m_AllOnes()))
3751  return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
3752 
3753  // icmp sge X, (Y + 1) -> icmp sgt X, Y
3754  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
3755  return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
3756 
3757  // icmp slt X, (Y + 1) -> icmp sle X, Y
3758  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
3759  return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
3760 
3761  // TODO: The subtraction-related identities shown below also hold, but
3762  // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
3763  // wouldn't happen even if they were implemented.
3764  //
3765  // icmp ult (X - 1), Y -> icmp ule X, Y
3766  // icmp uge (X - 1), Y -> icmp ugt X, Y
3767  // icmp ugt X, (Y - 1) -> icmp uge X, Y
3768  // icmp ule X, (Y - 1) -> icmp ult X, Y
3769 
3770  // icmp ule (X + 1), Y -> icmp ult X, Y
3771  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
3772  return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3773 
3774  // icmp ugt (X + 1), Y -> icmp uge X, Y
3775  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
3776  return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3777 
3778  // icmp uge X, (Y + 1) -> icmp ugt X, Y
3779  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
3780  return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3781 
3782  // icmp ult X, (Y + 1) -> icmp ule X, Y
3783  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
3784  return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3785 
3786  // if C1 has greater magnitude than C2:
3787  // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
3788  // s.t. C3 = C1 - C2
3789  //
3790  // if C2 has greater magnitude than C1:
3791  // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
3792  // s.t. C3 = C2 - C1
3793  if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3794  (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3795  if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3796  if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3797  const APInt &AP1 = C1->getValue();
3798  const APInt &AP2 = C2->getValue();
3799  if (AP1.isNegative() == AP2.isNegative()) {
3800  APInt AP1Abs = C1->getValue().abs();
3801  APInt AP2Abs = C2->getValue().abs();
3802  if (AP1Abs.uge(AP2Abs)) {
3803  ConstantInt *C3 = Builder.getInt(AP1 - AP2);
3804  Value *NewAdd = Builder.CreateNSWAdd(A, C3);
3805  return new ICmpInst(Pred, NewAdd, C);
3806  } else {
3807  ConstantInt *C3 = Builder.getInt(AP2 - AP1);
3808  Value *NewAdd = Builder.CreateNSWAdd(C, C3);
3809  return new ICmpInst(Pred, A, NewAdd);
3810  }
3811  }
3812  }
3813 
3814  // Analyze the case when either Op0 or Op1 is a sub instruction.
3815  // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3816  A = nullptr;
3817  B = nullptr;
3818  C = nullptr;
3819  D = nullptr;
3820  if (BO0 && BO0->getOpcode() == Instruction::Sub) {
3821  A = BO0->getOperand(0);
3822  B = BO0->getOperand(1);
3823  }
3824  if (BO1 && BO1->getOpcode() == Instruction::Sub) {
3825  C = BO1->getOperand(0);
3826  D = BO1->getOperand(1);
3827  }
3828 
3829  // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
3830  if (A == Op1 && NoOp0WrapProblem)
3831  return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3832  // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
3833  if (C == Op0 && NoOp1WrapProblem)
3834  return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3835 
3836  // Convert sub-with-unsigned-overflow comparisons into a comparison of args.
3837  // (A - B) u>/u<= A --> B u>/u<= A
3838  if (A == Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3839  return new ICmpInst(Pred, B, A);
3840  // C u</u>= (C - D) --> C u</u>= D
3841  if (C == Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3842  return new ICmpInst(Pred, C, D);
3843 
3844  // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
3845  if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
3846  // Try not to increase register pressure.
3847  BO0->hasOneUse() && BO1->hasOneUse())
3848  return new ICmpInst(Pred, A, C);
3849  // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
3850  if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
3851  // Try not to increase register pressure.
3852  BO0->hasOneUse() && BO1->hasOneUse())
3853  return new ICmpInst(Pred, D, B);
3854 
3855  // icmp (0-X) < cst --> x > -cst
3856  if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3857  Value *X;
3858  if (match(BO0, m_Neg(m_Value(X))))
3859  if (Constant *RHSC = dyn_cast<Constant>(Op1))
3860  if (RHSC->isNotMinSignedValue())
3861  return new ICmpInst(I.getSwappedPredicate(), X,
3862  ConstantExpr::getNeg(RHSC));
3863  }
3864 
3865  BinaryOperator *SRem = nullptr;
3866  // icmp (srem X, Y), Y
3867  if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
3868  SRem = BO0;
3869  // icmp Y, (srem X, Y)
3870  else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3871  Op0 == BO1->getOperand(1))
3872  SRem = BO1;
3873  if (SRem) {
3874  // We don't check hasOneUse to avoid increasing register pressure because
3875  // the value we use is the same value this instruction was already using.
3876  switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3877  default:
3878  break;
3879  case ICmpInst::ICMP_EQ:
3880  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3881  case ICmpInst::ICMP_NE:
3882  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3883  case ICmpInst::ICMP_SGT:
3884  case ICmpInst::ICMP_SGE:
3885  return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3887  case ICmpInst::ICMP_SLT:
3888  case ICmpInst::ICMP_SLE:
3889  return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3890  Constant::getNullValue(SRem->getType()));
3891  }
3892  }
3893 
3894  if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
3895  BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
3896  switch (BO0->getOpcode()) {
3897  default:
3898  break;
3899  case Instruction::Add:
3900  case Instruction::Sub:
3901  case Instruction::Xor: {
3902  if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3903  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3904 
3905  const APInt *C;
3906  if (match(BO0->getOperand(1), m_APInt(C))) {
3907  // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
3908  if (C->isSignMask()) {
3909  ICmpInst::Predicate NewPred =
3911  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3912  }
3913 
3914  // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
3915  if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
3916  ICmpInst::Predicate NewPred =
3918  NewPred = I.getSwappedPredicate(NewPred);
3919  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3920  }
3921  }
3922  break;
3923  }
3924  case Instruction::Mul: {
3925  if (!I.isEquality())
3926  break;
3927 
3928  const APInt *C;
3929  if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
3930  !C->isOneValue()) {
3931  // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
3932  // Mask = -1 >> count-trailing-zeros(C).
3933  if (unsigned TZs = C->countTrailingZeros()) {
3935  BO0->getType(),
3936  APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
3937  Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
3938  Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
3939  return new ICmpInst(Pred, And1, And2);
3940  }
3941  // If there are no trailing zeros in the multiplier, just eliminate
3942  // the multiplies (no masking is needed):
3943  // icmp eq/ne (X * C), (Y * C) --> icmp eq/ne X, Y
3944  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3945  }
3946  break;
3947  }
3948  case Instruction::UDiv:
3949  case Instruction::LShr:
3950  if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
3951  break;
3952  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3953 
3954  case Instruction::SDiv:
3955  if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
3956  break;
3957  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3958 
3959  case Instruction::AShr:
3960  if (!BO0->isExact() || !BO1->isExact())
3961  break;
3962  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3963 
3964  case Instruction::Shl: {
3965  bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3966  bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3967  if (!NUW && !NSW)
3968  break;
3969  if (!NSW && I.isSigned())
3970  break;
3971  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3972  }
3973  }
3974  }
3975 
3976  if (BO0) {
3977  // Transform A & (L - 1) `ult` L --> L != 0
3978  auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
3979  auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
3980 
3981  if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
3982  auto *Zero = Constant::getNullValue(BO0->getType());
3983  return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
3984  }
3985  }
3986 
3987  if (Value *V = foldUnsignedMultiplicationOverflowCheck(I))
3988  return replaceInstUsesWith(I, V);
3989 
3990  if (Value *V = foldICmpWithLowBitMaskedVal(I, Builder))
3991  return replaceInstUsesWith(I, V);
3992 
3993  if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder))
3994  return replaceInstUsesWith(I, V);
3995 
3996  if (Value *V = foldShiftIntoShiftInAnotherHandOfAndInICmp(I, SQ, Builder))
3997  return replaceInstUsesWith(I, V);
3998 
3999  return nullptr;
4000 }
4001 
4002 /// Fold icmp Pred min|max(X, Y), X.
4004  ICmpInst::Predicate Pred = Cmp.getPredicate();
4005  Value *Op0 = Cmp.getOperand(0);
4006  Value *X = Cmp.getOperand(1);
4007 
4008  // Canonicalize minimum or maximum operand to LHS of the icmp.
4009  if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
4010  match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
4011  match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
4012  match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
4013  std::swap(Op0, X);
4014  Pred = Cmp.getSwappedPredicate();
4015  }
4016 
4017  Value *Y;
4018  if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
4019  // smin(X, Y) == X --> X s<= Y
4020  // smin(X, Y) s>= X --> X s<= Y
4021  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
4022  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
4023 
4024  // smin(X, Y) != X --> X s> Y
4025  // smin(X, Y) s< X --> X s> Y
4026  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
4027  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
4028 
4029  // These cases should be handled in InstSimplify:
4030  // smin(X, Y) s<= X --> true
4031  // smin(X, Y) s> X --> false
4032  return nullptr;
4033  }
4034 
4035  if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
4036  // smax(X, Y) == X --> X s>= Y
4037  // smax(X, Y) s<= X --> X s>= Y
4038  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
4039  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
4040 
4041  // smax(X, Y) != X --> X s< Y
4042  // smax(X, Y) s> X --> X s< Y
4043  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
4044  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
4045 
4046  // These cases should be handled in InstSimplify:
4047  // smax(X, Y) s>= X --> true
4048  // smax(X, Y) s< X --> false
4049  return nullptr;
4050  }
4051 
4052  if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
4053  // umin(X, Y) == X --> X u<= Y
4054  // umin(X, Y) u>= X --> X u<= Y
4055  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
4056  return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
4057 
4058  // umin(X, Y) != X --> X u> Y
4059  // umin(X, Y) u< X --> X u> Y
4060  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
4061  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
4062 
4063  // These cases should be handled in InstSimplify:
4064  // umin(X, Y) u<= X --> true
4065  // umin(X, Y) u> X --> false
4066  return nullptr;
4067  }
4068 
4069  if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
4070  // umax(X, Y) == X --> X u>= Y
4071  // umax(X, Y) u<= X --> X u>= Y
4072  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
4073  return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
4074 
4075  // umax(X, Y) != X --> X u< Y
4076  // umax(X, Y) u> X --> X u< Y
4077  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
4078  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
4079 
4080  // These cases should be handled in InstSimplify:
4081  // umax(X, Y) u>= X --> true
4082  // umax(X, Y) u< X --> false
4083  return nullptr;
4084  }
4085 
4086  return nullptr;
4087 }
4088 
4089 Instruction *InstCombiner::foldICmpEquality(ICmpInst &I) {
4090  if (!I.isEquality())
4091  return nullptr;
4092 
4093  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4094  const CmpInst::Predicate Pred = I.getPredicate();
4095  Value *A, *B, *C, *D;
4096  if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4097  if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
4098  Value *OtherVal = A == Op1 ? B : A;
4099  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4100  }
4101 
4102  if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
4103  // A^c1 == C^c2 --> A == C^(c1^c2)
4104  ConstantInt *C1, *C2;
4105  if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
4106  Op1->hasOneUse()) {
4107  Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
4108  Value *Xor = Builder.CreateXor(C, NC);
4109  return new ICmpInst(Pred, A, Xor);
4110  }
4111 
4112  // A^B == A^D -> B == D
4113  if (A == C)
4114  return new ICmpInst(Pred, B, D);
4115  if (A == D)
4116  return new ICmpInst(Pred, B, C);
4117  if (B == C)
4118  return new ICmpInst(Pred, A, D);
4119  if (B == D)
4120  return new ICmpInst(Pred, A, C);
4121  }
4122  }
4123 
4124  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
4125  // A == (A^B) -> B == 0
4126  Value *OtherVal = A == Op0 ? B : A;
4127  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4128  }
4129 
4130  // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
4131  if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
4132  match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
4133  Value *X = nullptr, *Y = nullptr, *Z = nullptr;
4134 
4135  if (A == C) {
4136  X = B;
4137  Y = D;
4138  Z = A;
4139  } else if (A == D) {
4140  X = B;
4141  Y = C;
4142  Z = A;
4143  } else if (B == C) {
4144  X = A;
4145  Y = D;
4146  Z = B;
4147  } else if (B == D) {
4148  X = A;
4149  Y = C;
4150  Z = B;
4151  }
4152 
4153  if (X) { // Build (X^Y) & Z
4154  Op1 = Builder.CreateXor(X, Y);
4155  Op1 = Builder.CreateAnd(Op1, Z);
4156  I.setOperand(0, Op1);
4157  I.setOperand(1, Constant::getNullValue(Op1->getType()));
4158  return &I;
4159  }
4160  }
4161 
4162  // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
4163  // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
4164  ConstantInt *Cst1;
4165  if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
4166  match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
4167  (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
4168  match(Op1, m_ZExt(m_Value(A))))) {
4169  APInt Pow2 = Cst1->getValue() + 1;
4170  if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
4171  Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
4172  return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
4173  }
4174 
4175  // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
4176  // For lshr and ashr pairs.
4177  if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4178  match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
4179  (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4180  match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
4181  unsigned TypeBits = Cst1->getBitWidth();
4182  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4183  if (ShAmt < TypeBits && ShAmt != 0) {
4184  ICmpInst::Predicate NewPred =
4186  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4187  APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
4188  return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
4189  }
4190  }
4191 
4192  // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
4193  if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
4194  match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
4195  unsigned TypeBits = Cst1->getBitWidth();
4196  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4197  if (ShAmt < TypeBits && ShAmt != 0) {
4198  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4199  APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
4200  Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
4201  I.getName() + ".mask");
4202  return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
4203  }
4204  }
4205 
4206  // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
4207  // "icmp (and X, mask), cst"
4208  uint64_t ShAmt = 0;
4209  if (Op0->hasOneUse() &&
4210  match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
4211  match(Op1, m_ConstantInt(Cst1)) &&
4212  // Only do this when A has multiple uses. This is most important to do
4213  // when it exposes other optimizations.
4214  !A->hasOneUse()) {
4215  unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
4216 
4217  if (ShAmt < ASize) {
4218  APInt MaskV =
4220  MaskV <<= ShAmt;
4221 
4222  APInt CmpV = Cst1->getValue().zext(ASize);
4223  CmpV <<= ShAmt;
4224 
4225  Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
4226  return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
4227  }
4228  }
4229 
4230  // If both operands are byte-swapped or bit-reversed, just compare the
4231  // original values.
4232  // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
4233  // and handle more intrinsics.
4234  if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
4235  (match(Op0, m_BitReverse(m_Value(A))) &&
4236  match(Op1, m_BitReverse(m_Value(B)))))
4237  return new ICmpInst(Pred, A, B);
4238 
4239  // Canonicalize checking for a power-of-2-or-zero value:
4240  // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
4241  // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
4242  if (!match(Op0, m_OneUse(m_c_And(m_Add(m_Value(A), m_AllOnes()),
4243  m_Deferred(A)))) ||
4244  !match(Op1, m_ZeroInt()))
4245  A = nullptr;
4246 
4247  // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
4248  // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
4249  if (match(Op0, m_OneUse(m_c_And(m_Neg(m_Specific(Op1)), m_Specific(Op1)))))
4250  A = Op1;
4251  else if (match(Op1,
4252  m_OneUse(m_c_And(m_Neg(m_Specific(Op0)), m_Specific(Op0)))))
4253  A = Op0;
4254 
4255  if (A) {
4256  Type *Ty = A->getType();
4257  CallInst *CtPop = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, A);
4258  return Pred == ICmpInst::ICMP_EQ
4259  ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop, ConstantInt::get(Ty, 2))
4260  : new ICmpInst(ICmpInst::ICMP_UGT, CtPop, ConstantInt::get(Ty, 1));
4261  }
4262 
4263  return nullptr;
4264 }
4265 
4267  InstCombiner::BuilderTy &Builder) {
4268  assert(isa<CastInst>(ICmp.getOperand(0)) && "Expected cast for operand 0");
4269  auto *CastOp0 = cast<CastInst>(ICmp.getOperand(0));
4270  Value *X;
4271  if (!match(CastOp0, m_ZExtOrSExt(m_Value(X))))
4272  return nullptr;
4273 
4274  bool IsSignedExt = CastOp0->getOpcode() == Instruction::SExt;
4275  bool IsSignedCmp = ICmp.isSigned();
4276  if (auto *CastOp1 = dyn_cast<CastInst>(ICmp.getOperand(1))) {
4277  // If the signedness of the two casts doesn't agree (i.e. one is a sext
4278  // and the other is a zext), then we can't handle this.
4279  // TODO: This is too strict. We can handle some predicates (equality?).
4280  if (CastOp0->getOpcode() != CastOp1->getOpcode())
4281  return nullptr;
4282 
4283  // Not an extension from the same type?
4284  Value *Y = CastOp1->getOperand(0);
4285  Type *XTy = X->getType(), *YTy = Y->getType();
4286  if (XTy != YTy) {
4287  // One of the casts must have one use because we are creating a new cast.
4288  if (!CastOp0->hasOneUse() && !CastOp1->hasOneUse())
4289  return nullptr;
4290  // Extend the narrower operand to the type of the wider operand.
4291  if (XTy->getScalarSizeInBits() < YTy->getScalarSizeInBits())
4292  X = Builder.CreateCast(CastOp0->getOpcode(), X, YTy);
4293  else if (YTy->getScalarSizeInBits() < XTy->getScalarSizeInBits())
4294  Y = Builder.CreateCast(CastOp0->getOpcode(), Y, XTy);
4295  else
4296  return nullptr;
4297  }
4298 
4299  // (zext X) == (zext Y) --> X == Y
4300  // (sext X) == (sext Y) --> X == Y
4301  if (ICmp.isEquality())
4302  return new ICmpInst(ICmp.getPredicate(), X, Y);
4303 
4304  // A signed comparison of sign extended values simplifies into a
4305  // signed comparison.
4306  if (IsSignedCmp && IsSignedExt)
4307  return new ICmpInst(ICmp.getPredicate(), X, Y);
4308 
4309  // The other three cases all fold into an unsigned comparison.
4310  return new ICmpInst(ICmp.getUnsignedPredicate(), X, Y);
4311  }
4312 
4313  // Below here, we are only folding a compare with constant.
4314  auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
4315  if (!C)
4316  return nullptr;
4317 
4318  // Compute the constant that would happen if we truncated to SrcTy then
4319  // re-extended to DestTy.
4320  Type *SrcTy = CastOp0->getSrcTy();
4321  Type *DestTy = CastOp0->getDestTy();
4322  Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
4323  Constant *Res2 = ConstantExpr::getCast(CastOp0->getOpcode(), Res1, DestTy);
4324 
4325  // If the re-extended constant didn't change...
4326  if (Res2 == C) {
4327  if (ICmp.isEquality())
4328  return new ICmpInst(ICmp.getPredicate(), X, Res1);
4329 
4330  // A signed comparison of sign extended values simplifies into a
4331  // signed comparison.
4332  if (IsSignedExt && IsSignedCmp)
4333  return new ICmpInst(ICmp.getPredicate(), X, Res1);
4334 
4335  // The other three cases all fold into an unsigned comparison.
4336  return new ICmpInst(ICmp.getUnsignedPredicate(), X, Res1);
4337  }
4338 
4339  // The re-extended constant changed, partly changed (in the case of a vector),
4340  // or could not be determined to be equal (in the case of a constant
4341  // expression), so the constant cannot be represented in the shorter type.
4342  // All the cases that fold to true or false will have already been handled
4343  // by SimplifyICmpInst, so only deal with the tricky case.
4344  if (IsSignedCmp || !IsSignedExt || !isa<ConstantInt>(C))
4345  return nullptr;
4346 
4347  // Is source op positive?
4348  // icmp ult (sext X), C --> icmp sgt X, -1
4349  if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
4350  return new ICmpInst(CmpInst::ICMP_SGT, X, Constant::getAllOnesValue(SrcTy));
4351 
4352  // Is source op negative?
4353  // icmp ugt (sext X), C --> icmp slt X, 0
4354  assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
4355  return new ICmpInst(CmpInst::ICMP_SLT, X, Constant::getNullValue(SrcTy));
4356 }
4357 
4358 /// Handle icmp (cast x), (cast or constant).
4359 Instruction *InstCombiner::foldICmpWithCastOp(ICmpInst &ICmp) {
4360  auto *CastOp0 = dyn_cast<CastInst>(ICmp.getOperand(0));
4361  if (!CastOp0)
4362  return nullptr;
4363  if (!isa<Constant>(ICmp.getOperand(1)) && !isa<CastInst>(ICmp.getOperand(1)))
4364  return nullptr;
4365 
4366  Value *Op0Src = CastOp0->getOperand(0);
4367  Type *SrcTy = CastOp0->getSrcTy();
4368  Type *DestTy = CastOp0->getDestTy();
4369 
4370  // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
4371  // integer type is the same size as the pointer type.
4372  auto CompatibleSizes = [&](Type *SrcTy, Type *DestTy) {
4373  if (isa<VectorType>(SrcTy)) {
4374  SrcTy = cast<VectorType>(SrcTy)->getElementType();
4375  DestTy = cast<VectorType>(DestTy)->getElementType();
4376  }
4377  return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
4378  };
4379  if (CastOp0->getOpcode() == Instruction::PtrToInt &&
4380  CompatibleSizes(SrcTy, DestTy)) {
4381  Value *NewOp1 = nullptr;
4382  if (auto *PtrToIntOp1 = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
4383  Value *PtrSrc = PtrToIntOp1->getOperand(0);
4384  if (PtrSrc->getType()->getPointerAddressSpace() ==
4385  Op0Src->getType()->getPointerAddressSpace()) {
4386  NewOp1 = PtrToIntOp1->getOperand(0);
4387  // If the pointer types don't match, insert a bitcast.
4388  if (Op0Src->getType() != NewOp1->getType())
4389  NewOp1 = Builder.CreateBitCast(NewOp1, Op0Src->getType());
4390  }
4391  } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
4392  NewOp1 = ConstantExpr::getIntToPtr(RHSC, SrcTy);
4393  }
4394 
4395  if (NewOp1)
4396  return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
4397  }
4398 
4399  return foldICmpWithZextOrSext(ICmp, Builder);
4400 }
4401 
4402 static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS) {
4403  switch (BinaryOp) {
4404  default:
4405  llvm_unreachable("Unsupported binary op");
4406  case Instruction::Add:
4407  case Instruction::Sub:
4408  return match(RHS, m_Zero());
4409  case Instruction::Mul:
4410  return match(RHS, m_One());
4411  }
4412 }
4413 
4415  Instruction::BinaryOps BinaryOp, bool IsSigned,
4416  Value *LHS, Value *RHS, Instruction *CxtI) const {
4417  switch (BinaryOp) {
4418  default:
4419  llvm_unreachable("Unsupported binary op");
4420  case Instruction::Add:
4421  if (IsSigned)
4422  return computeOverflowForSignedAdd(LHS, RHS, CxtI);
4423  else
4424  return computeOverflowForUnsignedAdd(LHS, RHS, CxtI);
4425  case Instruction::Sub:
4426  if (IsSigned)
4427  return computeOverflowForSignedSub(LHS, RHS, CxtI);
4428  else
4429  return computeOverflowForUnsignedSub(LHS, RHS, CxtI);
4430  case Instruction::Mul:
4431  if (IsSigned)
4432  return computeOverflowForSignedMul(LHS, RHS, CxtI);
4433  else
4434  return computeOverflowForUnsignedMul(LHS, RHS, CxtI);
4435  }
4436 }
4437 
4438 bool InstCombiner::OptimizeOverflowCheck(
4439  Instruction::BinaryOps BinaryOp, bool IsSigned, Value *LHS, Value *RHS,
4440  Instruction &OrigI, Value *&Result, Constant *&Overflow) {
4441  if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
4442  std::swap(LHS, RHS);
4443 
4444  // If the overflow check was an add followed by a compare, the insertion point
4445  // may be pointing to the compare. We want to insert the new instructions
4446  // before the add in case there are uses of the add between the add and the
4447  // compare.
4448  Builder.SetInsertPoint(&OrigI);
4449 
4450  if (isNeutralValue(BinaryOp, RHS)) {
4451  Result = LHS;
4452  Overflow = Builder.getFalse();
4453  return true;
4454  }
4455 
4456  switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, &OrigI)) {
4458  return false;
4461  Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4462  Result->takeName(&OrigI);
4463  Overflow = Builder.getTrue();
4464  return true;
4466  Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4467  Result->takeName(&OrigI);
4468  Overflow = Builder.getFalse();
4469  if (auto *Inst = dyn_cast<Instruction>(Result)) {
4470  if (IsSigned)
4471  Inst->setHasNoSignedWrap();
4472  else
4473  Inst->setHasNoUnsignedWrap();
4474  }
4475  return true;
4476  }
4477 
4478  llvm_unreachable("Unexpected overflow result");
4479 }
4480 
4481 /// Recognize and process idiom involving test for multiplication
4482 /// overflow.
4483 ///
4484 /// The caller has matched a pattern of the form:
4485 /// I = cmp u (mul(zext A, zext B), V
4486 /// The function checks if this is a test for overflow and if so replaces
4487 /// multiplication with call to 'mul.with.overflow' intrinsic.
4488 ///
4489 /// \param I Compare instruction.
4490 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
4491 /// the compare instruction. Must be of integer type.
4492 /// \param OtherVal The other argument of compare instruction.
4493 /// \returns Instruction which must replace the compare instruction, NULL if no
4494 /// replacement required.
4496  Value *OtherVal, InstCombiner &IC) {
4497  // Don't bother doing this transformation for pointers, don't do it for
4498  // vectors.
4499  if (!isa<IntegerType>(MulVal->getType()))
4500  return nullptr;
4501 
4502  assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
4503  assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
4504  auto *MulInstr = dyn_cast<Instruction>(MulVal);
4505  if (!MulInstr)
4506  return nullptr;
4507  assert(MulInstr->getOpcode() == Instruction::Mul);
4508 
4509  auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
4510  *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
4511  assert(LHS->getOpcode() == Instruction::ZExt);
4512  assert(RHS->getOpcode() == Instruction::ZExt);
4513  Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
4514 
4515  // Calculate type and width of the result produced by mul.with.overflow.
4516  Type *TyA = A->getType(), *TyB = B->getType();
4517  unsigned WidthA = TyA->getPrimitiveSizeInBits(),
4518  WidthB = TyB->getPrimitiveSizeInBits();
4519  unsigned MulWidth;
4520  Type *MulType;
4521  if (WidthB > WidthA) {
4522  MulWidth = WidthB;
4523  MulType = TyB;
4524  } else {
4525  MulWidth = WidthA;
4526  MulType = TyA;
4527  }
4528 
4529  // In order to replace the original mul with a narrower mul.with.overflow,
4530  // all uses must ignore upper bits of the product. The number of used low
4531  // bits must be not greater than the width of mul.with.overflow.
4532  if (MulVal->hasNUsesOrMore(2))
4533  for (User *U : MulVal->users()) {
4534  if (U == &I)
4535  continue;
4536  if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4537  // Check if truncation ignores bits above MulWidth.
4538  unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
4539  if (TruncWidth > MulWidth)
4540  return nullptr;
4541  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4542  // Check if AND ignores bits above MulWidth.
4543  if (BO->getOpcode() != Instruction::And)
4544  return nullptr;
4545  if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4546  const APInt &CVal = CI->getValue();
4547  if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
4548  return nullptr;
4549  } else {
4550  // In this case we could have the operand of the binary operation
4551  // being defined in another block, and performing the replacement
4552  // could break the dominance relation.
4553  return nullptr;
4554  }
4555  } else {
4556  // Other uses prohibit this transformation.
4557  return nullptr;
4558  }
4559  }
4560 
4561  // Recognize patterns
4562  switch (I.getPredicate()) {
4563  case ICmpInst::ICMP_EQ:
4564  case ICmpInst::ICMP_NE:
4565  // Recognize pattern:
4566  // mulval = mul(zext A, zext B)
4567  // cmp eq/neq mulval, zext trunc mulval
4568  if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
4569  if (Zext->hasOneUse()) {
4570  Value *ZextArg = Zext->getOperand(0);
4571  if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
4572  if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
4573  break; //Recognized
4574  }
4575 
4576  // Recognize pattern:
4577  // mulval = mul(zext A, zext B)
4578  // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
4579  ConstantInt *CI;
4580  Value *ValToMask;
4581  if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
4582  if (ValToMask != MulVal)
4583  return nullptr;
4584  const APInt &CVal = CI->getValue() + 1;
4585  if (CVal.isPowerOf2()) {
4586  unsigned MaskWidth = CVal.logBase2();
4587  if (MaskWidth == MulWidth)
4588  break; // Recognized
4589  }
4590  }
4591  return nullptr;
4592 
4593  case ICmpInst::ICMP_UGT:
4594  // Recognize pattern:
4595  // mulval = mul(zext A, zext B)
4596  // cmp ugt mulval, max
4597  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4598  APInt MaxVal = APInt::getMaxValue(MulWidth);
4599  MaxVal = MaxVal.zext(CI->getBitWidth());
4600  if (MaxVal.eq(CI->getValue()))
4601  break; // Recognized
4602  }
4603  return nullptr;
4604 
4605  case ICmpInst::ICMP_UGE:
4606  // Recognize pattern:
4607  // mulval = mul(zext A, zext B)
4608  // cmp uge mulval, max+1
4609  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4610  APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4611  if (MaxVal.eq(CI->getValue()))
4612  break; // Recognized
4613  }
4614  return nullptr;
4615 
4616  case ICmpInst::ICMP_ULE:
4617  // Recognize pattern:
4618  // mulval = mul(zext A, zext B)
4619  // cmp ule mulval, max
4620  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4621  APInt MaxVal = APInt::getMaxValue(MulWidth);
4622  MaxVal = MaxVal.zext(CI->getBitWidth());
4623  if (MaxVal.eq(CI->getValue()))
4624  break; // Recognized
4625  }
4626  return nullptr;
4627 
4628  case ICmpInst::ICMP_ULT:
4629  // Recognize pattern:
4630  // mulval = mul(zext A, zext B)
4631  // cmp ule mulval, max + 1
4632  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4633  APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4634  if (MaxVal.eq(CI->getValue()))
4635  break; // Recognized
4636  }
4637  return nullptr;
4638 
4639  default:
4640  return nullptr;
4641  }
4642 
4643  InstCombiner::BuilderTy &Builder = IC.Builder;
4644  Builder.SetInsertPoint(MulInstr);
4645 
4646  // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
4647  Value *MulA = A, *MulB = B;
4648  if (WidthA < MulWidth)
4649  MulA = Builder.CreateZExt(A, MulType);
4650  if (WidthB < MulWidth)
4651  MulB = Builder.CreateZExt(B, MulType);
4653  I.getModule(), Intrinsic::umul_with_overflow, MulType);
4654  CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
4655  IC.Worklist.Add(MulInstr);
4656 
4657  // If there are uses of mul result other than the comparison, we know that
4658  // they are truncation or binary AND. Change them to use result of
4659  // mul.with.overflow and adjust properly mask/size.
4660  if (MulVal->hasNUsesOrMore(2)) {
4661  Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
4662  for (auto UI = MulVal->user_begin(), UE = MulVal->user_end(); UI != UE;) {
4663  User *U = *UI++;
4664  if (U == &I || U == OtherVal)
4665  continue;
4666  if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4667  if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
4668  IC.replaceInstUsesWith(*TI, Mul);
4669  else
4670  TI->setOperand(0, Mul);
4671  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4672  assert(BO->getOpcode() == Instruction::And);
4673  // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
4674  ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
4675  APInt ShortMask = CI->getValue().trunc(MulWidth);
4676  Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
4677  Instruction *Zext =
4678  cast<Instruction>(Builder.CreateZExt(ShortAnd, BO->getType()));
4679  IC.Worklist.Add(Zext);
4680  IC.replaceInstUsesWith(*BO, Zext);
4681  } else {
4682  llvm_unreachable("Unexpected Binary operation");
4683  }
4684  IC.Worklist.Add(cast<Instruction>(U));
4685  }
4686  }
4687  if (isa<Instruction>(OtherVal))
4688  IC.Worklist.Add(cast<Instruction>(OtherVal));
4689 
4690  // The original icmp gets replaced with the overflow value, maybe inverted
4691  // depending on predicate.
4692  bool Inverse = false;
4693  switch (I.getPredicate()) {
4694  case ICmpInst::ICMP_NE:
4695  break;
4696  case ICmpInst::ICMP_EQ:
4697  Inverse = true;
4698  break;
4699  case ICmpInst::ICMP_UGT:
4700  case ICmpInst::ICMP_UGE:
4701  if (I.getOperand(0) == MulVal)
4702  break;
4703  Inverse = true;
4704  break;
4705  case ICmpInst::ICMP_ULT:
4706  case ICmpInst::ICMP_ULE:
4707  if (I.getOperand(1) == MulVal)
4708  break;
4709  Inverse = true;
4710  break;
4711  default:
4712  llvm_unreachable("Unexpected predicate");
4713  }
4714  if (Inverse) {
4715  Value *Res = Builder.CreateExtractValue(Call, 1);
4716  return BinaryOperator::CreateNot(Res);
4717  }
4718 
4719  return ExtractValueInst::Create(Call, 1);
4720 }
4721 
4722 /// When performing a comparison against a constant, it is possible that not all
4723 /// the bits in the LHS are demanded. This helper method computes the mask that
4724 /// IS demanded.
4725 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
4726  const APInt *RHS;
4727  if (!match(I.getOperand(1), m_APInt(RHS)))
4728  return APInt::getAllOnesValue(BitWidth);
4729 
4730  // If this is a normal comparison, it demands all bits. If it is a sign bit
4731  // comparison, it only demands the sign bit.
4732  bool UnusedBit;
4733  if (isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
4734  return APInt::getSignMask(BitWidth);
4735 
4736  switch (I.getPredicate()) {
4737  // For a UGT comparison, we don't care about any bits that
4738  // correspond to the trailing ones of the comparand. The value of these
4739  // bits doesn't impact the outcome of the comparison, because any value
4740  // greater than the RHS must differ in a bit higher than these due to carry.
4741  case ICmpInst::ICMP_UGT:
4742  return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
4743 
4744  // Similarly, for a ULT comparison, we don't care about the trailing zeros.
4745  // Any value less than the RHS must differ in a higher bit because of carries.
4746  case ICmpInst::ICMP_ULT:
4747  return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
4748 
4749  default:
4750  return APInt::getAllOnesValue(BitWidth);
4751  }
4752 }
4753 
4754 /// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst
4755 /// should be swapped.
4756 /// The decision is based on how many times these two operands are reused
4757 /// as subtract operands and their positions in those instructions.
4758 /// The rationale is that several architectures use the same instruction for
4759 /// both subtract and cmp. Thus, it is better if the order of those operands
4760 /// match.
4761 /// \return true if Op0 and Op1 should be swapped.
4762 static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) {
4763  // Filter out pointer values as those cannot appear directly in subtract.
4764  // FIXME: we may want to go through inttoptrs or bitcasts.
4765  if (Op0->getType()->isPointerTy())
4766  return false;
4767  // If a subtract already has the same operands as a compare, swapping would be
4768  // bad. If a subtract has the same operands as a compare but in reverse order,
4769  // then swapping is good.
4770  int GoodToSwap = 0;
4771  for (const User *U : Op0->users()) {
4772  if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0))))
4773  GoodToSwap++;
4774  else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1))))
4775  GoodToSwap--;
4776  }
4777  return GoodToSwap > 0;
4778 }
4779 
4780 /// Check that one use is in the same block as the definition and all
4781 /// other uses are in blocks dominated by a given block.
4782 ///
4783 /// \param DI Definition
4784 /// \param UI Use
4785 /// \param DB Block that must dominate all uses of \p DI outside
4786 /// the parent block
4787 /// \return true when \p UI is the only use of \p DI in the parent block
4788 /// and all other uses of \p DI are in blocks dominated by \p DB.
4789 ///
4791  const Instruction *UI,
4792  const BasicBlock *DB) const {
4793  assert(DI && UI && "Instruction not defined\n");
4794  // Ignore incomplete definitions.
4795  if (!DI->getParent())
4796  return false;
4797  // DI and UI must be in the same block.
4798  if (DI->getParent() != UI->getParent())
4799  return false;
4800  // Protect from self-referencing blocks.
4801  if (DI->getParent() == DB)
4802  return false;
4803  for (const User *U : DI->users()) {
4804  auto *Usr = cast<Instruction>(U);
4805  if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
4806  return false;
4807  }
4808  return true;
4809 }
4810 
4811 /// Return true when the instruction sequence within a block is select-cmp-br.
4812 static bool isChainSelectCmpBranch(const SelectInst *SI) {
4813  const BasicBlock *BB = SI->getParent();
4814  if (!BB)
4815  return false;
4816  auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
4817  if (!BI || BI->getNumSuccessors() != 2)
4818  return false;
4819  auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
4820  if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
4821  return false;
4822  return true;
4823 }
4824 
4825 /// True when a select result is replaced by one of its operands
4826 /// in select-icmp sequence. This will eventually result in the elimination
4827 /// of the select.
4828 ///
4829 /// \param SI Select instruction
4830 /// \param Icmp Compare instruction
4831 /// \param SIOpd Operand that replaces the select
4832 ///
4833 /// Notes:
4834 /// - The replacement is global and requires dominator information
4835 /// - The caller is responsible for the actual replacement
4836 ///
4837 /// Example:
4838 ///
4839 /// entry:
4840 /// %4 = select i1 %3, %C* %0, %C* null
4841 /// %5 = icmp eq %C* %4, null
4842 /// br i1 %5, label %9, label %7
4843 /// ...
4844 /// ; <label>:7 ; preds = %entry
4845 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
4846 /// ...
4847 ///
4848 /// can be transformed to
4849 ///
4850 /// %5 = icmp eq %C* %0, null
4851 /// %6 = select i1 %3, i1 %5, i1 true
4852 /// br i1 %6, label %9, label %7
4853 /// ...
4854 /// ; <label>:7 ; preds = %entry
4855 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
4856 ///
4857 /// Similar when the first operand of the select is a constant or/and
4858 /// the compare is for not equal rather than equal.
4859 ///
4860 /// NOTE: The function is only called when the select and compare constants
4861 /// are equal, the optimization can work only for EQ predicates. This is not a
4862 /// major restriction since a NE compare should be 'normalized' to an equal
4863 /// compare, which usually happens in the combiner and test case
4864 /// select-cmp-br.ll checks for it.
4866  const ICmpInst *Icmp,
4867  const unsigned SIOpd) {
4868  assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
4869  if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
4870  BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
4871  // The check for the single predecessor is not the best that can be
4872  // done. But it protects efficiently against cases like when SI's
4873  // home block has two successors, Succ and Succ1, and Succ1 predecessor
4874  // of Succ. Then SI can't be replaced by SIOpd because the use that gets
4875  // replaced can be reached on either path. So the uniqueness check
4876  // guarantees that the path all uses of SI (outside SI's parent) are on
4877  // is disjoint from all other paths out of SI. But that information
4878  // is more expensive to compute, and the trade-off here is in favor
4879  // of compile-time. It should also be noticed that we check for a single
4880  // predecessor and not only uniqueness. This to handle the situation when
4881  // Succ and Succ1 points to the same basic block.
4882  if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
4883  NumSel++;
4884  SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
4885  return true;
4886  }
4887  }
4888  return false;
4889 }
4890 
4891 /// Try to fold the comparison based on range information we can get by checking
4892 /// whether bits are known to be zero or one in the inputs.
4893 Instruction *InstCombiner::foldICmpUsingKnownBits(ICmpInst &I) {
4894  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4895  Type *Ty = Op0->getType();
4896  ICmpInst::Predicate Pred = I.getPredicate();
4897 
4898  // Get scalar or pointer size.
4899  unsigned BitWidth = Ty->isIntOrIntVectorTy()
4900  ? Ty->getScalarSizeInBits()
4901  : DL.getIndexTypeSizeInBits(Ty->getScalarType());
4902 
4903  if (!BitWidth)
4904  return nullptr;
4905 
4906  KnownBits Op0Known(BitWidth);
4907  KnownBits Op1Known(BitWidth);
4908 
4909  if (SimplifyDemandedBits(&I, 0,
4910  getDemandedBitsLHSMask(I, BitWidth),
4911  Op0Known, 0))
4912  return &I;
4913 
4914  if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth),
4915  Op1Known, 0))
4916  return &I;
4917 
4918  // Given the known and unknown bits, compute a range that the LHS could be
4919  // in. Compute the Min, Max and RHS values based on the known bits. For the
4920  // EQ and NE we use unsigned values.
4921  APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
4922  APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
4923  if (I.isSigned()) {
4924  computeSignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4925  computeSignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4926  } else {
4927  computeUnsignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4928  computeUnsignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4929  }
4930 
4931  // If Min and Max are known to be the same, then SimplifyDemandedBits figured
4932  // out that the LHS or RHS is a constant. Constant fold this now, so that
4933  // code below can assume that Min != Max.
4934  if (!isa<Constant>(Op0) && Op0Min == Op0Max)
4935  return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1);
4936  if (!isa<Constant>(Op1) && Op1Min == Op1Max)
4937  return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min));
4938 
4939  // Based on the range information we know about the LHS, see if we can
4940  // simplify this comparison. For example, (x&4) < 8 is always true.
4941  switch (Pred) {
4942  default:
4943  llvm_unreachable("Unknown icmp opcode!");
4944  case ICmpInst::ICMP_EQ:
4945  case ICmpInst::ICMP_NE: {
4946  if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) {
4947  return Pred == CmpInst::ICMP_EQ
4948  ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()))
4949  : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4950  }
4951 
4952  // If all bits are known zero except for one, then we know at most one bit
4953  // is set. If the comparison is against zero, then this is a check to see if
4954  // *that* bit is set.
4955  APInt Op0KnownZeroInverted = ~Op0Known.Zero;
4956  if (Op1Known.isZero()) {
4957  // If the LHS is an AND with the same constant, look through it.
4958  Value *LHS = nullptr;
4959  const APInt *LHSC;
4960  if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
4961  *LHSC != Op0KnownZeroInverted)
4962  LHS = Op0;
4963 
4964  Value *X;
4965  if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
4966  APInt ValToCheck = Op0KnownZeroInverted;
4967  Type *XTy = X->getType();
4968  if (ValToCheck.isPowerOf2()) {
4969  // ((1 << X) & 8) == 0 -> X != 3
4970  // ((1 << X) & 8) != 0 -> X == 3
4971  auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4972  auto NewPred = ICmpInst::getInversePredicate(Pred);
4973  return new ICmpInst(NewPred, X, CmpC);
4974  } else if ((++ValToCheck).isPowerOf2()) {
4975  // ((1 << X) & 7) == 0 -> X >= 3
4976  // ((1 << X) & 7) != 0 -> X < 3
4977  auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4978  auto NewPred =
4980  return new ICmpInst(NewPred, X, CmpC);
4981  }
4982  }
4983 
4984  // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
4985  const APInt *CI;
4986  if (Op0KnownZeroInverted.isOneValue() &&
4987  match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
4988  // ((8 >>u X) & 1) == 0 -> X != 3
4989  // ((8 >>u X) & 1) != 0 -> X == 3
4990  unsigned CmpVal = CI->countTrailingZeros();
4991  auto NewPred = ICmpInst::getInversePredicate(Pred);
4992  return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
4993  }
4994  }
4995  break;
4996  }
4997  case ICmpInst::ICMP_ULT: {
4998  if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
4999  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5000  if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
5001  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5002  if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
5003  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5004 
5005  const APInt *CmpC;
5006  if (match(Op1, m_APInt(CmpC))) {
5007  // A <u C -> A == C-1 if min(A)+1 == C
5008  if (*CmpC == Op0Min + 1)
5009  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5010  ConstantInt::get(Op1->getType(), *CmpC - 1));
5011  // X <u C --> X == 0, if the number of zero bits in the bottom of X
5012  // exceeds the log2 of C.
5013  if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
5014  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5015  Constant::getNullValue(Op1->getType()));
5016  }
5017  break;
5018  }
5019  case ICmpInst::ICMP_UGT: {
5020  if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
5021  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5022  if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
5023  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5024  if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
5025  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5026 
5027  const APInt *CmpC;
5028  if (match(Op1, m_APInt(CmpC))) {
5029  // A >u C -> A == C+1 if max(a)-1 == C
5030  if (*CmpC == Op0Max - 1)
5031  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5032  ConstantInt::get(Op1->getType(), *CmpC + 1));
5033  // X >u C --> X != 0, if the number of zero bits in the bottom of X
5034  // exceeds the log2 of C.
5035  if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
5036  return new ICmpInst(ICmpInst::ICMP_NE, Op0,
5037  Constant::getNullValue(Op1->getType()));
5038  }
5039  break;
5040  }
5041  case ICmpInst::ICMP_SLT: {
5042  if (Op0Max.