LLVM  8.0.0svn
InstCombineCompares.cpp
Go to the documentation of this file.
1 //===- InstCombineCompares.cpp --------------------------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the visitICmp and visitFCmp functions.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "InstCombineInternal.h"
15 #include "llvm/ADT/APSInt.h"
16 #include "llvm/ADT/SetVector.h"
17 #include "llvm/ADT/Statistic.h"
21 #include "llvm/IR/ConstantRange.h"
22 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/IntrinsicInst.h"
25 #include "llvm/IR/PatternMatch.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/KnownBits.h"
28 
29 using namespace llvm;
30 using namespace PatternMatch;
31 
32 #define DEBUG_TYPE "instcombine"
33 
34 // How many times is a select replaced by one of its operands?
35 STATISTIC(NumSel, "Number of select opts");
36 
37 
38 /// Compute Result = In1+In2, returning true if the result overflowed for this
39 /// type.
40 static bool addWithOverflow(APInt &Result, const APInt &In1,
41  const APInt &In2, bool IsSigned = false) {
42  bool Overflow;
43  if (IsSigned)
44  Result = In1.sadd_ov(In2, Overflow);
45  else
46  Result = In1.uadd_ov(In2, Overflow);
47 
48  return Overflow;
49 }
50 
51 /// Compute Result = In1-In2, returning true if the result overflowed for this
52 /// type.
53 static bool subWithOverflow(APInt &Result, const APInt &In1,
54  const APInt &In2, bool IsSigned = false) {
55  bool Overflow;
56  if (IsSigned)
57  Result = In1.ssub_ov(In2, Overflow);
58  else
59  Result = In1.usub_ov(In2, Overflow);
60 
61  return Overflow;
62 }
63 
64 /// Given an icmp instruction, return true if any use of this comparison is a
65 /// branch on sign bit comparison.
66 static bool hasBranchUse(ICmpInst &I) {
67  for (auto *U : I.users())
68  if (isa<BranchInst>(U))
69  return true;
70  return false;
71 }
72 
73 /// Given an exploded icmp instruction, return true if the comparison only
74 /// checks the sign bit. If it only checks the sign bit, set TrueIfSigned if the
75 /// result of the comparison is true when the input value is signed.
76 static bool isSignBitCheck(ICmpInst::Predicate Pred, const APInt &RHS,
77  bool &TrueIfSigned) {
78  switch (Pred) {
79  case ICmpInst::ICMP_SLT: // True if LHS s< 0
80  TrueIfSigned = true;
81  return RHS.isNullValue();
82  case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
83  TrueIfSigned = true;
84  return RHS.isAllOnesValue();
85  case ICmpInst::ICMP_SGT: // True if LHS s> -1
86  TrueIfSigned = false;
87  return RHS.isAllOnesValue();
88  case ICmpInst::ICMP_UGT:
89  // True if LHS u> RHS and RHS == high-bit-mask - 1
90  TrueIfSigned = true;
91  return RHS.isMaxSignedValue();
92  case ICmpInst::ICMP_UGE:
93  // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
94  TrueIfSigned = true;
95  return RHS.isSignMask();
96  default:
97  return false;
98  }
99 }
100 
101 /// Returns true if the exploded icmp can be expressed as a signed comparison
102 /// to zero and updates the predicate accordingly.
103 /// The signedness of the comparison is preserved.
104 /// TODO: Refactor with decomposeBitTestICmp()?
105 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
106  if (!ICmpInst::isSigned(Pred))
107  return false;
108 
109  if (C.isNullValue())
110  return ICmpInst::isRelational(Pred);
111 
112  if (C.isOneValue()) {
113  if (Pred == ICmpInst::ICMP_SLT) {
114  Pred = ICmpInst::ICMP_SLE;
115  return true;
116  }
117  } else if (C.isAllOnesValue()) {
118  if (Pred == ICmpInst::ICMP_SGT) {
119  Pred = ICmpInst::ICMP_SGE;
120  return true;
121  }
122  }
123 
124  return false;
125 }
126 
127 /// Given a signed integer type and a set of known zero and one bits, compute
128 /// the maximum and minimum values that could have the specified known zero and
129 /// known one bits, returning them in Min/Max.
130 /// TODO: Move to method on KnownBits struct?
132  APInt &Min, APInt &Max) {
133  assert(Known.getBitWidth() == Min.getBitWidth() &&
134  Known.getBitWidth() == Max.getBitWidth() &&
135  "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
136  APInt UnknownBits = ~(Known.Zero|Known.One);
137 
138  // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
139  // bit if it is unknown.
140  Min = Known.One;
141  Max = Known.One|UnknownBits;
142 
143  if (UnknownBits.isNegative()) { // Sign bit is unknown
144  Min.setSignBit();
145  Max.clearSignBit();
146  }
147 }
148 
149 /// Given an unsigned integer type and a set of known zero and one bits, compute
150 /// the maximum and minimum values that could have the specified known zero and
151 /// known one bits, returning them in Min/Max.
152 /// TODO: Move to method on KnownBits struct?
154  APInt &Min, APInt &Max) {
155  assert(Known.getBitWidth() == Min.getBitWidth() &&
156  Known.getBitWidth() == Max.getBitWidth() &&
157  "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
158  APInt UnknownBits = ~(Known.Zero|Known.One);
159 
160  // The minimum value is when the unknown bits are all zeros.
161  Min = Known.One;
162  // The maximum value is when the unknown bits are all ones.
163  Max = Known.One|UnknownBits;
164 }
165 
166 /// This is called when we see this pattern:
167 /// cmp pred (load (gep GV, ...)), cmpcst
168 /// where GV is a global variable with a constant initializer. Try to simplify
169 /// this into some simple computation that does not need the load. For example
170 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
171 ///
172 /// If AndCst is non-null, then the loaded value is masked with that constant
173 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
174 Instruction *InstCombiner::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
175  GlobalVariable *GV,
176  CmpInst &ICI,
177  ConstantInt *AndCst) {
178  Constant *Init = GV->getInitializer();
179  if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
180  return nullptr;
181 
182  uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
183  // Don't blow up on huge arrays.
184  if (ArrayElementCount > MaxArraySizeForCombine)
185  return nullptr;
186 
187  // There are many forms of this optimization we can handle, for now, just do
188  // the simple index into a single-dimensional array.
189  //
190  // Require: GEP GV, 0, i {{, constant indices}}
191  if (GEP->getNumOperands() < 3 ||
192  !isa<ConstantInt>(GEP->getOperand(1)) ||
193  !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
194  isa<Constant>(GEP->getOperand(2)))
195  return nullptr;
196 
197  // Check that indices after the variable are constants and in-range for the
198  // type they index. Collect the indices. This is typically for arrays of
199  // structs.
200  SmallVector<unsigned, 4> LaterIndices;
201 
202  Type *EltTy = Init->getType()->getArrayElementType();
203  for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
204  ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
205  if (!Idx) return nullptr; // Variable index.
206 
207  uint64_t IdxVal = Idx->getZExtValue();
208  if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
209 
210  if (StructType *STy = dyn_cast<StructType>(EltTy))
211  EltTy = STy->getElementType(IdxVal);
212  else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
213  if (IdxVal >= ATy->getNumElements()) return nullptr;
214  EltTy = ATy->getElementType();
215  } else {
216  return nullptr; // Unknown type.
217  }
218 
219  LaterIndices.push_back(IdxVal);
220  }
221 
222  enum { Overdefined = -3, Undefined = -2 };
223 
224  // Variables for our state machines.
225 
226  // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
227  // "i == 47 | i == 87", where 47 is the first index the condition is true for,
228  // and 87 is the second (and last) index. FirstTrueElement is -2 when
229  // undefined, otherwise set to the first true element. SecondTrueElement is
230  // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
231  int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
232 
233  // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
234  // form "i != 47 & i != 87". Same state transitions as for true elements.
235  int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
236 
237  /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
238  /// define a state machine that triggers for ranges of values that the index
239  /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
240  /// This is -2 when undefined, -3 when overdefined, and otherwise the last
241  /// index in the range (inclusive). We use -2 for undefined here because we
242  /// use relative comparisons and don't want 0-1 to match -1.
243  int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
244 
245  // MagicBitvector - This is a magic bitvector where we set a bit if the
246  // comparison is true for element 'i'. If there are 64 elements or less in
247  // the array, this will fully represent all the comparison results.
248  uint64_t MagicBitvector = 0;
249 
250  // Scan the array and see if one of our patterns matches.
251  Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
252  for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
253  Constant *Elt = Init->getAggregateElement(i);
254  if (!Elt) return nullptr;
255 
256  // If this is indexing an array of structures, get the structure element.
257  if (!LaterIndices.empty())
258  Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
259 
260  // If the element is masked, handle it.
261  if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
262 
263  // Find out if the comparison would be true or false for the i'th element.
265  CompareRHS, DL, &TLI);
266  // If the result is undef for this element, ignore it.
267  if (isa<UndefValue>(C)) {
268  // Extend range state machines to cover this element in case there is an
269  // undef in the middle of the range.
270  if (TrueRangeEnd == (int)i-1)
271  TrueRangeEnd = i;
272  if (FalseRangeEnd == (int)i-1)
273  FalseRangeEnd = i;
274  continue;
275  }
276 
277  // If we can't compute the result for any of the elements, we have to give
278  // up evaluating the entire conditional.
279  if (!isa<ConstantInt>(C)) return nullptr;
280 
281  // Otherwise, we know if the comparison is true or false for this element,
282  // update our state machines.
283  bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
284 
285  // State machine for single/double/range index comparison.
286  if (IsTrueForElt) {
287  // Update the TrueElement state machine.
288  if (FirstTrueElement == Undefined)
289  FirstTrueElement = TrueRangeEnd = i; // First true element.
290  else {
291  // Update double-compare state machine.
292  if (SecondTrueElement == Undefined)
293  SecondTrueElement = i;
294  else
295  SecondTrueElement = Overdefined;
296 
297  // Update range state machine.
298  if (TrueRangeEnd == (int)i-1)
299  TrueRangeEnd = i;
300  else
301  TrueRangeEnd = Overdefined;
302  }
303  } else {
304  // Update the FalseElement state machine.
305  if (FirstFalseElement == Undefined)
306  FirstFalseElement = FalseRangeEnd = i; // First false element.
307  else {
308  // Update double-compare state machine.
309  if (SecondFalseElement == Undefined)
310  SecondFalseElement = i;
311  else
312  SecondFalseElement = Overdefined;
313 
314  // Update range state machine.
315  if (FalseRangeEnd == (int)i-1)
316  FalseRangeEnd = i;
317  else
318  FalseRangeEnd = Overdefined;
319  }
320  }
321 
322  // If this element is in range, update our magic bitvector.
323  if (i < 64 && IsTrueForElt)
324  MagicBitvector |= 1ULL << i;
325 
326  // If all of our states become overdefined, bail out early. Since the
327  // predicate is expensive, only check it every 8 elements. This is only
328  // really useful for really huge arrays.
329  if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
330  SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
331  FalseRangeEnd == Overdefined)
332  return nullptr;
333  }
334 
335  // Now that we've scanned the entire array, emit our new comparison(s). We
336  // order the state machines in complexity of the generated code.
337  Value *Idx = GEP->getOperand(2);
338 
339  // If the index is larger than the pointer size of the target, truncate the
340  // index down like the GEP would do implicitly. We don't have to do this for
341  // an inbounds GEP because the index can't be out of range.
342  if (!GEP->isInBounds()) {
343  Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
344  unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
345  if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize)
346  Idx = Builder.CreateTrunc(Idx, IntPtrTy);
347  }
348 
349  // If the comparison is only true for one or two elements, emit direct
350  // comparisons.
351  if (SecondTrueElement != Overdefined) {
352  // None true -> false.
353  if (FirstTrueElement == Undefined)
354  return replaceInstUsesWith(ICI, Builder.getFalse());
355 
356  Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
357 
358  // True for one element -> 'i == 47'.
359  if (SecondTrueElement == Undefined)
360  return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
361 
362  // True for two elements -> 'i == 47 | i == 72'.
363  Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
364  Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
365  Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
366  return BinaryOperator::CreateOr(C1, C2);
367  }
368 
369  // If the comparison is only false for one or two elements, emit direct
370  // comparisons.
371  if (SecondFalseElement != Overdefined) {
372  // None false -> true.
373  if (FirstFalseElement == Undefined)
374  return replaceInstUsesWith(ICI, Builder.getTrue());
375 
376  Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
377 
378  // False for one element -> 'i != 47'.
379  if (SecondFalseElement == Undefined)
380  return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
381 
382  // False for two elements -> 'i != 47 & i != 72'.
383  Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
384  Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
385  Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
386  return BinaryOperator::CreateAnd(C1, C2);
387  }
388 
389  // If the comparison can be replaced with a range comparison for the elements
390  // where it is true, emit the range check.
391  if (TrueRangeEnd != Overdefined) {
392  assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
393 
394  // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
395  if (FirstTrueElement) {
396  Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
397  Idx = Builder.CreateAdd(Idx, Offs);
398  }
399 
400  Value *End = ConstantInt::get(Idx->getType(),
401  TrueRangeEnd-FirstTrueElement+1);
402  return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
403  }
404 
405  // False range check.
406  if (FalseRangeEnd != Overdefined) {
407  assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
408  // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
409  if (FirstFalseElement) {
410  Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
411  Idx = Builder.CreateAdd(Idx, Offs);
412  }
413 
414  Value *End = ConstantInt::get(Idx->getType(),
415  FalseRangeEnd-FirstFalseElement);
416  return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
417  }
418 
419  // If a magic bitvector captures the entire comparison state
420  // of this load, replace it with computation that does:
421  // ((magic_cst >> i) & 1) != 0
422  {
423  Type *Ty = nullptr;
424 
425  // Look for an appropriate type:
426  // - The type of Idx if the magic fits
427  // - The smallest fitting legal type
428  if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
429  Ty = Idx->getType();
430  else
431  Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
432 
433  if (Ty) {
434  Value *V = Builder.CreateIntCast(Idx, Ty, false);
435  V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
436  V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
437  return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
438  }
439  }
440 
441  return nullptr;
442 }
443 
444 /// Return a value that can be used to compare the *offset* implied by a GEP to
445 /// zero. For example, if we have &A[i], we want to return 'i' for
446 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
447 /// are involved. The above expression would also be legal to codegen as
448 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
449 /// This latter form is less amenable to optimization though, and we are allowed
450 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
451 ///
452 /// If we can't emit an optimized form for this expression, this returns null.
453 ///
455  const DataLayout &DL) {
457 
458  // Check to see if this gep only has a single variable index. If so, and if
459  // any constant indices are a multiple of its scale, then we can compute this
460  // in terms of the scale of the variable index. For example, if the GEP
461  // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
462  // because the expression will cross zero at the same point.
463  unsigned i, e = GEP->getNumOperands();
464  int64_t Offset = 0;
465  for (i = 1; i != e; ++i, ++GTI) {
466  if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
467  // Compute the aggregate offset of constant indices.
468  if (CI->isZero()) continue;
469 
470  // Handle a struct index, which adds its field offset to the pointer.
471  if (StructType *STy = GTI.getStructTypeOrNull()) {
472  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
473  } else {
474  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
475  Offset += Size*CI->getSExtValue();
476  }
477  } else {
478  // Found our variable index.
479  break;
480  }
481  }
482 
483  // If there are no variable indices, we must have a constant offset, just
484  // evaluate it the general way.
485  if (i == e) return nullptr;
486 
487  Value *VariableIdx = GEP->getOperand(i);
488  // Determine the scale factor of the variable element. For example, this is
489  // 4 if the variable index is into an array of i32.
490  uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
491 
492  // Verify that there are no other variable indices. If so, emit the hard way.
493  for (++i, ++GTI; i != e; ++i, ++GTI) {
494  ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
495  if (!CI) return nullptr;
496 
497  // Compute the aggregate offset of constant indices.
498  if (CI->isZero()) continue;
499 
500  // Handle a struct index, which adds its field offset to the pointer.
501  if (StructType *STy = GTI.getStructTypeOrNull()) {
502  Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
503  } else {
504  uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
505  Offset += Size*CI->getSExtValue();
506  }
507  }
508 
509  // Okay, we know we have a single variable index, which must be a
510  // pointer/array/vector index. If there is no offset, life is simple, return
511  // the index.
512  Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
513  unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
514  if (Offset == 0) {
515  // Cast to intptrty in case a truncation occurs. If an extension is needed,
516  // we don't need to bother extending: the extension won't affect where the
517  // computation crosses zero.
518  if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) {
519  VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
520  }
521  return VariableIdx;
522  }
523 
524  // Otherwise, there is an index. The computation we will do will be modulo
525  // the pointer size.
526  Offset = SignExtend64(Offset, IntPtrWidth);
527  VariableScale = SignExtend64(VariableScale, IntPtrWidth);
528 
529  // To do this transformation, any constant index must be a multiple of the
530  // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
531  // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
532  // multiple of the variable scale.
533  int64_t NewOffs = Offset / (int64_t)VariableScale;
534  if (Offset != NewOffs*(int64_t)VariableScale)
535  return nullptr;
536 
537  // Okay, we can do this evaluation. Start by converting the index to intptr.
538  if (VariableIdx->getType() != IntPtrTy)
539  VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
540  true /*Signed*/);
541  Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
542  return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
543 }
544 
545 /// Returns true if we can rewrite Start as a GEP with pointer Base
546 /// and some integer offset. The nodes that need to be re-written
547 /// for this transformation will be added to Explored.
548 static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
549  const DataLayout &DL,
550  SetVector<Value *> &Explored) {
551  SmallVector<Value *, 16> WorkList(1, Start);
552  Explored.insert(Base);
553 
554  // The following traversal gives us an order which can be used
555  // when doing the final transformation. Since in the final
556  // transformation we create the PHI replacement instructions first,
557  // we don't have to get them in any particular order.
558  //
559  // However, for other instructions we will have to traverse the
560  // operands of an instruction first, which means that we have to
561  // do a post-order traversal.
562  while (!WorkList.empty()) {
564 
565  while (!WorkList.empty()) {
566  if (Explored.size() >= 100)
567  return false;
568 
569  Value *V = WorkList.back();
570 
571  if (Explored.count(V) != 0) {
572  WorkList.pop_back();
573  continue;
574  }
575 
576  if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
577  !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
578  // We've found some value that we can't explore which is different from
579  // the base. Therefore we can't do this transformation.
580  return false;
581 
582  if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
583  auto *CI = dyn_cast<CastInst>(V);
584  if (!CI->isNoopCast(DL))
585  return false;
586 
587  if (Explored.count(CI->getOperand(0)) == 0)
588  WorkList.push_back(CI->getOperand(0));
589  }
590 
591  if (auto *GEP = dyn_cast<GEPOperator>(V)) {
592  // We're limiting the GEP to having one index. This will preserve
593  // the original pointer type. We could handle more cases in the
594  // future.
595  if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
596  GEP->getType() != Start->getType())
597  return false;
598 
599  if (Explored.count(GEP->getOperand(0)) == 0)
600  WorkList.push_back(GEP->getOperand(0));
601  }
602 
603  if (WorkList.back() == V) {
604  WorkList.pop_back();
605  // We've finished visiting this node, mark it as such.
606  Explored.insert(V);
607  }
608 
609  if (auto *PN = dyn_cast<PHINode>(V)) {
610  // We cannot transform PHIs on unsplittable basic blocks.
611  if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
612  return false;
613  Explored.insert(PN);
614  PHIs.insert(PN);
615  }
616  }
617 
618  // Explore the PHI nodes further.
619  for (auto *PN : PHIs)
620  for (Value *Op : PN->incoming_values())
621  if (Explored.count(Op) == 0)
622  WorkList.push_back(Op);
623  }
624 
625  // Make sure that we can do this. Since we can't insert GEPs in a basic
626  // block before a PHI node, we can't easily do this transformation if
627  // we have PHI node users of transformed instructions.
628  for (Value *Val : Explored) {
629  for (Value *Use : Val->uses()) {
630 
631  auto *PHI = dyn_cast<PHINode>(Use);
632  auto *Inst = dyn_cast<Instruction>(Val);
633 
634  if (Inst == Base || Inst == PHI || !Inst || !PHI ||
635  Explored.count(PHI) == 0)
636  continue;
637 
638  if (PHI->getParent() == Inst->getParent())
639  return false;
640  }
641  }
642  return true;
643 }
644 
645 // Sets the appropriate insert point on Builder where we can add
646 // a replacement Instruction for V (if that is possible).
647 static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
648  bool Before = true) {
649  if (auto *PHI = dyn_cast<PHINode>(V)) {
650  Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
651  return;
652  }
653  if (auto *I = dyn_cast<Instruction>(V)) {
654  if (!Before)
655  I = &*std::next(I->getIterator());
656  Builder.SetInsertPoint(I);
657  return;
658  }
659  if (auto *A = dyn_cast<Argument>(V)) {
660  // Set the insertion point in the entry block.
661  BasicBlock &Entry = A->getParent()->getEntryBlock();
662  Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
663  return;
664  }
665  // Otherwise, this is a constant and we don't need to set a new
666  // insertion point.
667  assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
668 }
669 
670 /// Returns a re-written value of Start as an indexed GEP using Base as a
671 /// pointer.
673  const DataLayout &DL,
674  SetVector<Value *> &Explored) {
675  // Perform all the substitutions. This is a bit tricky because we can
676  // have cycles in our use-def chains.
677  // 1. Create the PHI nodes without any incoming values.
678  // 2. Create all the other values.
679  // 3. Add the edges for the PHI nodes.
680  // 4. Emit GEPs to get the original pointers.
681  // 5. Remove the original instructions.
682  Type *IndexType = IntegerType::get(
683  Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
684 
686  NewInsts[Base] = ConstantInt::getNullValue(IndexType);
687 
688  // Create the new PHI nodes, without adding any incoming values.
689  for (Value *Val : Explored) {
690  if (Val == Base)
691  continue;
692  // Create empty phi nodes. This avoids cyclic dependencies when creating
693  // the remaining instructions.
694  if (auto *PHI = dyn_cast<PHINode>(Val))
695  NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
696  PHI->getName() + ".idx", PHI);
697  }
698  IRBuilder<> Builder(Base->getContext());
699 
700  // Create all the other instructions.
701  for (Value *Val : Explored) {
702 
703  if (NewInsts.find(Val) != NewInsts.end())
704  continue;
705 
706  if (auto *CI = dyn_cast<CastInst>(Val)) {
707  NewInsts[CI] = NewInsts[CI->getOperand(0)];
708  continue;
709  }
710  if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
711  Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
712  : GEP->getOperand(1);
713  setInsertionPoint(Builder, GEP);
714  // Indices might need to be sign extended. GEPs will magically do
715  // this, but we need to do it ourselves here.
716  if (Index->getType()->getScalarSizeInBits() !=
717  NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
718  Index = Builder.CreateSExtOrTrunc(
719  Index, NewInsts[GEP->getOperand(0)]->getType(),
720  GEP->getOperand(0)->getName() + ".sext");
721  }
722 
723  auto *Op = NewInsts[GEP->getOperand(0)];
724  if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
725  NewInsts[GEP] = Index;
726  else
727  NewInsts[GEP] = Builder.CreateNSWAdd(
728  Op, Index, GEP->getOperand(0)->getName() + ".add");
729  continue;
730  }
731  if (isa<PHINode>(Val))
732  continue;
733 
734  llvm_unreachable("Unexpected instruction type");
735  }
736 
737  // Add the incoming values to the PHI nodes.
738  for (Value *Val : Explored) {
739  if (Val == Base)
740  continue;
741  // All the instructions have been created, we can now add edges to the
742  // phi nodes.
743  if (auto *PHI = dyn_cast<PHINode>(Val)) {
744  PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
745  for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
746  Value *NewIncoming = PHI->getIncomingValue(I);
747 
748  if (NewInsts.find(NewIncoming) != NewInsts.end())
749  NewIncoming = NewInsts[NewIncoming];
750 
751  NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
752  }
753  }
754  }
755 
756  for (Value *Val : Explored) {
757  if (Val == Base)
758  continue;
759 
760  // Depending on the type, for external users we have to emit
761  // a GEP or a GEP + ptrtoint.
762  setInsertionPoint(Builder, Val, false);
763 
764  // If required, create an inttoptr instruction for Base.
765  Value *NewBase = Base;
766  if (!Base->getType()->isPointerTy())
767  NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
768  Start->getName() + "to.ptr");
769 
770  Value *GEP = Builder.CreateInBoundsGEP(
771  Start->getType()->getPointerElementType(), NewBase,
772  makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
773 
774  if (!Val->getType()->isPointerTy()) {
775  Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
776  Val->getName() + ".conv");
777  GEP = Cast;
778  }
779  Val->replaceAllUsesWith(GEP);
780  }
781 
782  return NewInsts[Start];
783 }
784 
785 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
786 /// the input Value as a constant indexed GEP. Returns a pair containing
787 /// the GEPs Pointer and Index.
788 static std::pair<Value *, Value *>
790  Type *IndexType = IntegerType::get(V->getContext(),
792 
794  while (true) {
795  if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
796  // We accept only inbouds GEPs here to exclude the possibility of
797  // overflow.
798  if (!GEP->isInBounds())
799  break;
800  if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
801  GEP->getType() == V->getType()) {
802  V = GEP->getOperand(0);
803  Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
804  Index = ConstantExpr::getAdd(
805  Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
806  continue;
807  }
808  break;
809  }
810  if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
811  if (!CI->isNoopCast(DL))
812  break;
813  V = CI->getOperand(0);
814  continue;
815  }
816  if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
817  if (!CI->isNoopCast(DL))
818  break;
819  V = CI->getOperand(0);
820  continue;
821  }
822  break;
823  }
824  return {V, Index};
825 }
826 
827 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
828 /// We can look through PHIs, GEPs and casts in order to determine a common base
829 /// between GEPLHS and RHS.
831  ICmpInst::Predicate Cond,
832  const DataLayout &DL) {
833  if (!GEPLHS->hasAllConstantIndices())
834  return nullptr;
835 
836  // Make sure the pointers have the same type.
837  if (GEPLHS->getType() != RHS->getType())
838  return nullptr;
839 
840  Value *PtrBase, *Index;
841  std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
842 
843  // The set of nodes that will take part in this transformation.
844  SetVector<Value *> Nodes;
845 
846  if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
847  return nullptr;
848 
849  // We know we can re-write this as
850  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
851  // Since we've only looked through inbouds GEPs we know that we
852  // can't have overflow on either side. We can therefore re-write
853  // this as:
854  // OFFSET1 cmp OFFSET2
855  Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
856 
857  // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
858  // GEP having PtrBase as the pointer base, and has returned in NewRHS the
859  // offset. Since Index is the offset of LHS to the base pointer, we will now
860  // compare the offsets instead of comparing the pointers.
861  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
862 }
863 
864 /// Fold comparisons between a GEP instruction and something else. At this point
865 /// we know that the GEP is on the LHS of the comparison.
866 Instruction *InstCombiner::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
867  ICmpInst::Predicate Cond,
868  Instruction &I) {
869  // Don't transform signed compares of GEPs into index compares. Even if the
870  // GEP is inbounds, the final add of the base pointer can have signed overflow
871  // and would change the result of the icmp.
872  // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
873  // the maximum signed value for the pointer type.
874  if (ICmpInst::isSigned(Cond))
875  return nullptr;
876 
877  // Look through bitcasts and addrspacecasts. We do not however want to remove
878  // 0 GEPs.
879  if (!isa<GetElementPtrInst>(RHS))
880  RHS = RHS->stripPointerCasts();
881 
882  Value *PtrBase = GEPLHS->getOperand(0);
883  if (PtrBase == RHS && GEPLHS->isInBounds()) {
884  // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
885  // This transformation (ignoring the base and scales) is valid because we
886  // know pointers can't overflow since the gep is inbounds. See if we can
887  // output an optimized form.
888  Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
889 
890  // If not, synthesize the offset the hard way.
891  if (!Offset)
892  Offset = EmitGEPOffset(GEPLHS);
893  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
894  Constant::getNullValue(Offset->getType()));
895  } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
896  // If the base pointers are different, but the indices are the same, just
897  // compare the base pointer.
898  if (PtrBase != GEPRHS->getOperand(0)) {
899  bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
900  IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
901  GEPRHS->getOperand(0)->getType();
902  if (IndicesTheSame)
903  for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
904  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
905  IndicesTheSame = false;
906  break;
907  }
908 
909  // If all indices are the same, just compare the base pointers.
910  Type *BaseType = GEPLHS->getOperand(0)->getType();
911  if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
912  return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
913 
914  // If we're comparing GEPs with two base pointers that only differ in type
915  // and both GEPs have only constant indices or just one use, then fold
916  // the compare with the adjusted indices.
917  if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
918  (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
919  (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
920  PtrBase->stripPointerCasts() ==
921  GEPRHS->getOperand(0)->stripPointerCasts()) {
922  Value *LOffset = EmitGEPOffset(GEPLHS);
923  Value *ROffset = EmitGEPOffset(GEPRHS);
924 
925  // If we looked through an addrspacecast between different sized address
926  // spaces, the LHS and RHS pointers are different sized
927  // integers. Truncate to the smaller one.
928  Type *LHSIndexTy = LOffset->getType();
929  Type *RHSIndexTy = ROffset->getType();
930  if (LHSIndexTy != RHSIndexTy) {
931  if (LHSIndexTy->getPrimitiveSizeInBits() <
932  RHSIndexTy->getPrimitiveSizeInBits()) {
933  ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
934  } else
935  LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
936  }
937 
938  Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
939  LOffset, ROffset);
940  return replaceInstUsesWith(I, Cmp);
941  }
942 
943  // Otherwise, the base pointers are different and the indices are
944  // different. Try convert this to an indexed compare by looking through
945  // PHIs/casts.
946  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
947  }
948 
949  // If one of the GEPs has all zero indices, recurse.
950  if (GEPLHS->hasAllZeroIndices())
951  return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
953 
954  // If the other GEP has all zero indices, recurse.
955  if (GEPRHS->hasAllZeroIndices())
956  return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
957 
958  bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
959  if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
960  // If the GEPs only differ by one index, compare it.
961  unsigned NumDifferences = 0; // Keep track of # differences.
962  unsigned DiffOperand = 0; // The operand that differs.
963  for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
964  if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
965  if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
966  GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
967  // Irreconcilable differences.
968  NumDifferences = 2;
969  break;
970  } else {
971  if (NumDifferences++) break;
972  DiffOperand = i;
973  }
974  }
975 
976  if (NumDifferences == 0) // SAME GEP?
977  return replaceInstUsesWith(I, // No comparison is needed here.
979 
980  else if (NumDifferences == 1 && GEPsInBounds) {
981  Value *LHSV = GEPLHS->getOperand(DiffOperand);
982  Value *RHSV = GEPRHS->getOperand(DiffOperand);
983  // Make sure we do a signed comparison here.
984  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
985  }
986  }
987 
988  // Only lower this if the icmp is the only user of the GEP or if we expect
989  // the result to fold to a constant!
990  if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
991  (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
992  // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
993  Value *L = EmitGEPOffset(GEPLHS);
994  Value *R = EmitGEPOffset(GEPRHS);
995  return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
996  }
997  }
998 
999  // Try convert this to an indexed compare by looking through PHIs/casts as a
1000  // last resort.
1001  return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1002 }
1003 
1004 Instruction *InstCombiner::foldAllocaCmp(ICmpInst &ICI,
1005  const AllocaInst *Alloca,
1006  const Value *Other) {
1007  assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1008 
1009  // It would be tempting to fold away comparisons between allocas and any
1010  // pointer not based on that alloca (e.g. an argument). However, even
1011  // though such pointers cannot alias, they can still compare equal.
1012  //
1013  // But LLVM doesn't specify where allocas get their memory, so if the alloca
1014  // doesn't escape we can argue that it's impossible to guess its value, and we
1015  // can therefore act as if any such guesses are wrong.
1016  //
1017  // The code below checks that the alloca doesn't escape, and that it's only
1018  // used in a comparison once (the current instruction). The
1019  // single-comparison-use condition ensures that we're trivially folding all
1020  // comparisons against the alloca consistently, and avoids the risk of
1021  // erroneously folding a comparison of the pointer with itself.
1022 
1023  unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1024 
1026  for (const Use &U : Alloca->uses()) {
1027  if (Worklist.size() >= MaxIter)
1028  return nullptr;
1029  Worklist.push_back(&U);
1030  }
1031 
1032  unsigned NumCmps = 0;
1033  while (!Worklist.empty()) {
1034  assert(Worklist.size() <= MaxIter);
1035  const Use *U = Worklist.pop_back_val();
1036  const Value *V = U->getUser();
1037  --MaxIter;
1038 
1039  if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1040  isa<SelectInst>(V)) {
1041  // Track the uses.
1042  } else if (isa<LoadInst>(V)) {
1043  // Loading from the pointer doesn't escape it.
1044  continue;
1045  } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1046  // Storing *to* the pointer is fine, but storing the pointer escapes it.
1047  if (SI->getValueOperand() == U->get())
1048  return nullptr;
1049  continue;
1050  } else if (isa<ICmpInst>(V)) {
1051  if (NumCmps++)
1052  return nullptr; // Found more than one cmp.
1053  continue;
1054  } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1055  switch (Intrin->getIntrinsicID()) {
1056  // These intrinsics don't escape or compare the pointer. Memset is safe
1057  // because we don't allow ptrtoint. Memcpy and memmove are safe because
1058  // we don't allow stores, so src cannot point to V.
1059  case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1060  case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1061  continue;
1062  default:
1063  return nullptr;
1064  }
1065  } else {
1066  return nullptr;
1067  }
1068  for (const Use &U : V->uses()) {
1069  if (Worklist.size() >= MaxIter)
1070  return nullptr;
1071  Worklist.push_back(&U);
1072  }
1073  }
1074 
1075  Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1076  return replaceInstUsesWith(
1077  ICI,
1079 }
1080 
1081 /// Fold "icmp pred (X+C), X".
1082 Instruction *InstCombiner::foldICmpAddOpConst(Value *X, const APInt &C,
1083  ICmpInst::Predicate Pred) {
1084  // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1085  // so the values can never be equal. Similarly for all other "or equals"
1086  // operators.
1087  assert(!!C && "C should not be zero!");
1088 
1089  // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1090  // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1091  // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1092  if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1093  Constant *R = ConstantInt::get(X->getType(),
1094  APInt::getMaxValue(C.getBitWidth()) - C);
1095  return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1096  }
1097 
1098  // (X+1) >u X --> X <u (0-1) --> X != 255
1099  // (X+2) >u X --> X <u (0-2) --> X <u 254
1100  // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1101  if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1102  return new ICmpInst(ICmpInst::ICMP_ULT, X,
1103  ConstantInt::get(X->getType(), -C));
1104 
1106 
1107  // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1108  // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1109  // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1110  // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1111  // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1112  // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1113  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1114  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1115  ConstantInt::get(X->getType(), SMax - C));
1116 
1117  // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1118  // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1119  // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1120  // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1121  // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1122  // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1123 
1124  assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1125  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1126  ConstantInt::get(X->getType(), SMax - (C - 1)));
1127 }
1128 
1129 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1130 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1131 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
1132 Instruction *InstCombiner::foldICmpShrConstConst(ICmpInst &I, Value *A,
1133  const APInt &AP1,
1134  const APInt &AP2) {
1135  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1136 
1137  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1138  if (I.getPredicate() == I.ICMP_NE)
1139  Pred = CmpInst::getInversePredicate(Pred);
1140  return new ICmpInst(Pred, LHS, RHS);
1141  };
1142 
1143  // Don't bother doing any work for cases which InstSimplify handles.
1144  if (AP2.isNullValue())
1145  return nullptr;
1146 
1147  bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1148  if (IsAShr) {
1149  if (AP2.isAllOnesValue())
1150  return nullptr;
1151  if (AP2.isNegative() != AP1.isNegative())
1152  return nullptr;
1153  if (AP2.sgt(AP1))
1154  return nullptr;
1155  }
1156 
1157  if (!AP1)
1158  // 'A' must be large enough to shift out the highest set bit.
1159  return getICmp(I.ICMP_UGT, A,
1160  ConstantInt::get(A->getType(), AP2.logBase2()));
1161 
1162  if (AP1 == AP2)
1163  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1164 
1165  int Shift;
1166  if (IsAShr && AP1.isNegative())
1167  Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1168  else
1169  Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1170 
1171  if (Shift > 0) {
1172  if (IsAShr && AP1 == AP2.ashr(Shift)) {
1173  // There are multiple solutions if we are comparing against -1 and the LHS
1174  // of the ashr is not a power of two.
1175  if (AP1.isAllOnesValue() && !AP2.isPowerOf2())
1176  return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1177  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1178  } else if (AP1 == AP2.lshr(Shift)) {
1179  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1180  }
1181  }
1182 
1183  // Shifting const2 will never be equal to const1.
1184  // FIXME: This should always be handled by InstSimplify?
1185  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1186  return replaceInstUsesWith(I, TorF);
1187 }
1188 
1189 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1190 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
1191 Instruction *InstCombiner::foldICmpShlConstConst(ICmpInst &I, Value *A,
1192  const APInt &AP1,
1193  const APInt &AP2) {
1194  assert(I.isEquality() && "Cannot fold icmp gt/lt");
1195 
1196  auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1197  if (I.getPredicate() == I.ICMP_NE)
1198  Pred = CmpInst::getInversePredicate(Pred);
1199  return new ICmpInst(Pred, LHS, RHS);
1200  };
1201 
1202  // Don't bother doing any work for cases which InstSimplify handles.
1203  if (AP2.isNullValue())
1204  return nullptr;
1205 
1206  unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1207 
1208  if (!AP1 && AP2TrailingZeros != 0)
1209  return getICmp(
1210  I.ICMP_UGE, A,
1211  ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1212 
1213  if (AP1 == AP2)
1214  return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1215 
1216  // Get the distance between the lowest bits that are set.
1217  int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1218 
1219  if (Shift > 0 && AP2.shl(Shift) == AP1)
1220  return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1221 
1222  // Shifting const2 will never be equal to const1.
1223  // FIXME: This should always be handled by InstSimplify?
1224  auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1225  return replaceInstUsesWith(I, TorF);
1226 }
1227 
1228 /// The caller has matched a pattern of the form:
1229 /// I = icmp ugt (add (add A, B), CI2), CI1
1230 /// If this is of the form:
1231 /// sum = a + b
1232 /// if (sum+128 >u 255)
1233 /// Then replace it with llvm.sadd.with.overflow.i8.
1234 ///
1236  ConstantInt *CI2, ConstantInt *CI1,
1237  InstCombiner &IC) {
1238  // The transformation we're trying to do here is to transform this into an
1239  // llvm.sadd.with.overflow. To do this, we have to replace the original add
1240  // with a narrower add, and discard the add-with-constant that is part of the
1241  // range check (if we can't eliminate it, this isn't profitable).
1242 
1243  // In order to eliminate the add-with-constant, the compare can be its only
1244  // use.
1245  Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1246  if (!AddWithCst->hasOneUse())
1247  return nullptr;
1248 
1249  // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1250  if (!CI2->getValue().isPowerOf2())
1251  return nullptr;
1252  unsigned NewWidth = CI2->getValue().countTrailingZeros();
1253  if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1254  return nullptr;
1255 
1256  // The width of the new add formed is 1 more than the bias.
1257  ++NewWidth;
1258 
1259  // Check to see that CI1 is an all-ones value with NewWidth bits.
1260  if (CI1->getBitWidth() == NewWidth ||
1261  CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1262  return nullptr;
1263 
1264  // This is only really a signed overflow check if the inputs have been
1265  // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1266  // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1267  unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1268  if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1269  IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1270  return nullptr;
1271 
1272  // In order to replace the original add with a narrower
1273  // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1274  // and truncates that discard the high bits of the add. Verify that this is
1275  // the case.
1276  Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1277  for (User *U : OrigAdd->users()) {
1278  if (U == AddWithCst)
1279  continue;
1280 
1281  // Only accept truncates for now. We would really like a nice recursive
1282  // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1283  // chain to see which bits of a value are actually demanded. If the
1284  // original add had another add which was then immediately truncated, we
1285  // could still do the transformation.
1286  TruncInst *TI = dyn_cast<TruncInst>(U);
1287  if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1288  return nullptr;
1289  }
1290 
1291  // If the pattern matches, truncate the inputs to the narrower type and
1292  // use the sadd_with_overflow intrinsic to efficiently compute both the
1293  // result and the overflow bit.
1294  Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1296  Intrinsic::sadd_with_overflow, NewType);
1297 
1298  InstCombiner::BuilderTy &Builder = IC.Builder;
1299 
1300  // Put the new code above the original add, in case there are any uses of the
1301  // add between the add and the compare.
1302  Builder.SetInsertPoint(OrigAdd);
1303 
1304  Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1305  Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1306  CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1307  Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1308  Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1309 
1310  // The inner add was the result of the narrow add, zero extended to the
1311  // wider type. Replace it with the result computed by the intrinsic.
1312  IC.replaceInstUsesWith(*OrigAdd, ZExt);
1313 
1314  // The original icmp gets replaced with the overflow value.
1315  return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1316 }
1317 
1318 // Handle (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1319 Instruction *InstCombiner::foldICmpWithZero(ICmpInst &Cmp) {
1320  CmpInst::Predicate Pred = Cmp.getPredicate();
1321  Value *X = Cmp.getOperand(0);
1322 
1323  if (match(Cmp.getOperand(1), m_Zero()) && Pred == ICmpInst::ICMP_SGT) {
1324  Value *A, *B;
1325  SelectPatternResult SPR = matchSelectPattern(X, A, B);
1326  if (SPR.Flavor == SPF_SMIN) {
1327  if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1328  return new ICmpInst(Pred, B, Cmp.getOperand(1));
1329  if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1330  return new ICmpInst(Pred, A, Cmp.getOperand(1));
1331  }
1332  }
1333  return nullptr;
1334 }
1335 
1336 /// Fold icmp Pred X, C.
1337 /// TODO: This code structure does not make sense. The saturating add fold
1338 /// should be moved to some other helper and extended as noted below (it is also
1339 /// possible that code has been made unnecessary - do we canonicalize IR to
1340 /// overflow/saturating intrinsics or not?).
1341 Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) {
1342  // Match the following pattern, which is a common idiom when writing
1343  // overflow-safe integer arithmetic functions. The source performs an addition
1344  // in wider type and explicitly checks for overflow using comparisons against
1345  // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1346  //
1347  // TODO: This could probably be generalized to handle other overflow-safe
1348  // operations if we worked out the formulas to compute the appropriate magic
1349  // constants.
1350  //
1351  // sum = a + b
1352  // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1353  CmpInst::Predicate Pred = Cmp.getPredicate();
1354  Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1355  Value *A, *B;
1356  ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1357  if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1358  match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1359  if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1360  return Res;
1361 
1362  return nullptr;
1363 }
1364 
1365 /// Canonicalize icmp instructions based on dominating conditions.
1366 Instruction *InstCombiner::foldICmpWithDominatingICmp(ICmpInst &Cmp) {
1367  // This is a cheap/incomplete check for dominance - just match a single
1368  // predecessor with a conditional branch.
1369  BasicBlock *CmpBB = Cmp.getParent();
1370  BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1371  if (!DomBB)
1372  return nullptr;
1373 
1374  Value *DomCond;
1375  BasicBlock *TrueBB, *FalseBB;
1376  if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1377  return nullptr;
1378 
1379  assert((TrueBB == CmpBB || FalseBB == CmpBB) &&
1380  "Predecessor block does not point to successor?");
1381 
1382  // The branch should get simplified. Don't bother simplifying this condition.
1383  if (TrueBB == FalseBB)
1384  return nullptr;
1385 
1386  // Try to simplify this compare to T/F based on the dominating condition.
1387  Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1388  if (Imp)
1389  return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1390 
1391  CmpInst::Predicate Pred = Cmp.getPredicate();
1392  Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1393  ICmpInst::Predicate DomPred;
1394  const APInt *C, *DomC;
1395  if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1396  match(Y, m_APInt(C))) {
1397  // We have 2 compares of a variable with constants. Calculate the constant
1398  // ranges of those compares to see if we can transform the 2nd compare:
1399  // DomBB:
1400  // DomCond = icmp DomPred X, DomC
1401  // br DomCond, CmpBB, FalseBB
1402  // CmpBB:
1403  // Cmp = icmp Pred X, C
1405  ConstantRange DominatingCR =
1406  (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1408  CmpInst::getInversePredicate(DomPred), *DomC);
1409  ConstantRange Intersection = DominatingCR.intersectWith(CR);
1410  ConstantRange Difference = DominatingCR.difference(CR);
1411  if (Intersection.isEmptySet())
1412  return replaceInstUsesWith(Cmp, Builder.getFalse());
1413  if (Difference.isEmptySet())
1414  return replaceInstUsesWith(Cmp, Builder.getTrue());
1415 
1416  // Canonicalizing a sign bit comparison that gets used in a branch,
1417  // pessimizes codegen by generating branch on zero instruction instead
1418  // of a test and branch. So we avoid canonicalizing in such situations
1419  // because test and branch instruction has better branch displacement
1420  // than compare and branch instruction.
1421  bool UnusedBit;
1422  bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1423  if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1424  return nullptr;
1425 
1426  if (const APInt *EqC = Intersection.getSingleElement())
1427  return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1428  if (const APInt *NeC = Difference.getSingleElement())
1429  return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1430  }
1431 
1432  return nullptr;
1433 }
1434 
1435 /// Fold icmp (trunc X, Y), C.
1436 Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp,
1437  TruncInst *Trunc,
1438  const APInt &C) {
1439  ICmpInst::Predicate Pred = Cmp.getPredicate();
1440  Value *X = Trunc->getOperand(0);
1441  if (C.isOneValue() && C.getBitWidth() > 1) {
1442  // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1443  Value *V = nullptr;
1444  if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1445  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1446  ConstantInt::get(V->getType(), 1));
1447  }
1448 
1449  if (Cmp.isEquality() && Trunc->hasOneUse()) {
1450  // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1451  // of the high bits truncated out of x are known.
1452  unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1453  SrcBits = X->getType()->getScalarSizeInBits();
1454  KnownBits Known = computeKnownBits(X, 0, &Cmp);
1455 
1456  // If all the high bits are known, we can do this xform.
1457  if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1458  // Pull in the high bits from known-ones set.
1459  APInt NewRHS = C.zext(SrcBits);
1460  NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1461  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1462  }
1463  }
1464 
1465  return nullptr;
1466 }
1467 
1468 /// Fold icmp (xor X, Y), C.
1469 Instruction *InstCombiner::foldICmpXorConstant(ICmpInst &Cmp,
1470  BinaryOperator *Xor,
1471  const APInt &C) {
1472  Value *X = Xor->getOperand(0);
1473  Value *Y = Xor->getOperand(1);
1474  const APInt *XorC;
1475  if (!match(Y, m_APInt(XorC)))
1476  return nullptr;
1477 
1478  // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1479  // fold the xor.
1480  ICmpInst::Predicate Pred = Cmp.getPredicate();
1481  bool TrueIfSigned = false;
1482  if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1483 
1484  // If the sign bit of the XorCst is not set, there is no change to
1485  // the operation, just stop using the Xor.
1486  if (!XorC->isNegative()) {
1487  Cmp.setOperand(0, X);
1488  Worklist.Add(Xor);
1489  return &Cmp;
1490  }
1491 
1492  // Emit the opposite comparison.
1493  if (TrueIfSigned)
1494  return new ICmpInst(ICmpInst::ICMP_SGT, X,
1496  else
1497  return new ICmpInst(ICmpInst::ICMP_SLT, X,
1499  }
1500 
1501  if (Xor->hasOneUse()) {
1502  // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1503  if (!Cmp.isEquality() && XorC->isSignMask()) {
1504  Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1505  : Cmp.getSignedPredicate();
1506  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1507  }
1508 
1509  // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1510  if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1511  Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate()
1512  : Cmp.getSignedPredicate();
1513  Pred = Cmp.getSwappedPredicate(Pred);
1514  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1515  }
1516  }
1517 
1518  // Mask constant magic can eliminate an 'xor' with unsigned compares.
1519  if (Pred == ICmpInst::ICMP_UGT) {
1520  // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1521  if (*XorC == ~C && (C + 1).isPowerOf2())
1522  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1523  // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1524  if (*XorC == C && (C + 1).isPowerOf2())
1525  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1526  }
1527  if (Pred == ICmpInst::ICMP_ULT) {
1528  // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1529  if (*XorC == -C && C.isPowerOf2())
1530  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1531  ConstantInt::get(X->getType(), ~C));
1532  // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1533  if (*XorC == C && (-C).isPowerOf2())
1534  return new ICmpInst(ICmpInst::ICMP_UGT, X,
1535  ConstantInt::get(X->getType(), ~C));
1536  }
1537  return nullptr;
1538 }
1539 
1540 /// Fold icmp (and (sh X, Y), C2), C1.
1541 Instruction *InstCombiner::foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And,
1542  const APInt &C1, const APInt &C2) {
1543  BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1544  if (!Shift || !Shift->isShift())
1545  return nullptr;
1546 
1547  // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1548  // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1549  // code produced by the clang front-end, for bitfield access.
1550  // This seemingly simple opportunity to fold away a shift turns out to be
1551  // rather complicated. See PR17827 for details.
1552  unsigned ShiftOpcode = Shift->getOpcode();
1553  bool IsShl = ShiftOpcode == Instruction::Shl;
1554  const APInt *C3;
1555  if (match(Shift->getOperand(1), m_APInt(C3))) {
1556  bool CanFold = false;
1557  if (ShiftOpcode == Instruction::Shl) {
1558  // For a left shift, we can fold if the comparison is not signed. We can
1559  // also fold a signed comparison if the mask value and comparison value
1560  // are not negative. These constraints may not be obvious, but we can
1561  // prove that they are correct using an SMT solver.
1562  if (!Cmp.isSigned() || (!C2.isNegative() && !C1.isNegative()))
1563  CanFold = true;
1564  } else {
1565  bool IsAshr = ShiftOpcode == Instruction::AShr;
1566  // For a logical right shift, we can fold if the comparison is not signed.
1567  // We can also fold a signed comparison if the shifted mask value and the
1568  // shifted comparison value are not negative. These constraints may not be
1569  // obvious, but we can prove that they are correct using an SMT solver.
1570  // For an arithmetic shift right we can do the same, if we ensure
1571  // the And doesn't use any bits being shifted in. Normally these would
1572  // be turned into lshr by SimplifyDemandedBits, but not if there is an
1573  // additional user.
1574  if (!IsAshr || (C2.shl(*C3).lshr(*C3) == C2)) {
1575  if (!Cmp.isSigned() ||
1576  (!C2.shl(*C3).isNegative() && !C1.shl(*C3).isNegative()))
1577  CanFold = true;
1578  }
1579  }
1580 
1581  if (CanFold) {
1582  APInt NewCst = IsShl ? C1.lshr(*C3) : C1.shl(*C3);
1583  APInt SameAsC1 = IsShl ? NewCst.shl(*C3) : NewCst.lshr(*C3);
1584  // Check to see if we are shifting out any of the bits being compared.
1585  if (SameAsC1 != C1) {
1586  // If we shifted bits out, the fold is not going to work out. As a
1587  // special case, check to see if this means that the result is always
1588  // true or false now.
1589  if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1590  return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1591  if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1592  return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1593  } else {
1594  Cmp.setOperand(1, ConstantInt::get(And->getType(), NewCst));
1595  APInt NewAndCst = IsShl ? C2.lshr(*C3) : C2.shl(*C3);
1596  And->setOperand(1, ConstantInt::get(And->getType(), NewAndCst));
1597  And->setOperand(0, Shift->getOperand(0));
1598  Worklist.Add(Shift); // Shift is dead.
1599  return &Cmp;
1600  }
1601  }
1602  }
1603 
1604  // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1605  // preferable because it allows the C2 << Y expression to be hoisted out of a
1606  // loop if Y is invariant and X is not.
1607  if (Shift->hasOneUse() && C1.isNullValue() && Cmp.isEquality() &&
1608  !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1609  // Compute C2 << Y.
1610  Value *NewShift =
1611  IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1612  : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1613 
1614  // Compute X & (C2 << Y).
1615  Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1616  Cmp.setOperand(0, NewAnd);
1617  return &Cmp;
1618  }
1619 
1620  return nullptr;
1621 }
1622 
1623 /// Fold icmp (and X, C2), C1.
1624 Instruction *InstCombiner::foldICmpAndConstConst(ICmpInst &Cmp,
1625  BinaryOperator *And,
1626  const APInt &C1) {
1627  // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1628  // TODO: We canonicalize to the longer form for scalars because we have
1629  // better analysis/folds for icmp, and codegen may be better with icmp.
1630  if (Cmp.getPredicate() == CmpInst::ICMP_NE && Cmp.getType()->isVectorTy() &&
1631  C1.isNullValue() && match(And->getOperand(1), m_One()))
1632  return new TruncInst(And->getOperand(0), Cmp.getType());
1633 
1634  const APInt *C2;
1635  if (!match(And->getOperand(1), m_APInt(C2)))
1636  return nullptr;
1637 
1638  if (!And->hasOneUse())
1639  return nullptr;
1640 
1641  // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1642  // the input width without changing the value produced, eliminate the cast:
1643  //
1644  // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1645  //
1646  // We can do this transformation if the constants do not have their sign bits
1647  // set or if it is an equality comparison. Extending a relational comparison
1648  // when we're checking the sign bit would not work.
1649  Value *W;
1650  if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1651  (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1652  // TODO: Is this a good transform for vectors? Wider types may reduce
1653  // throughput. Should this transform be limited (even for scalars) by using
1654  // shouldChangeType()?
1655  if (!Cmp.getType()->isVectorTy()) {
1656  Type *WideType = W->getType();
1657  unsigned WideScalarBits = WideType->getScalarSizeInBits();
1658  Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1659  Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1660  Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1661  return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1662  }
1663  }
1664 
1665  if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1666  return I;
1667 
1668  // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1669  // (icmp pred (and A, (or (shl 1, B), 1), 0))
1670  //
1671  // iff pred isn't signed
1672  if (!Cmp.isSigned() && C1.isNullValue() && And->getOperand(0)->hasOneUse() &&
1673  match(And->getOperand(1), m_One())) {
1674  Constant *One = cast<Constant>(And->getOperand(1));
1675  Value *Or = And->getOperand(0);
1676  Value *A, *B, *LShr;
1677  if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1678  match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1679  unsigned UsesRemoved = 0;
1680  if (And->hasOneUse())
1681  ++UsesRemoved;
1682  if (Or->hasOneUse())
1683  ++UsesRemoved;
1684  if (LShr->hasOneUse())
1685  ++UsesRemoved;
1686 
1687  // Compute A & ((1 << B) | 1)
1688  Value *NewOr = nullptr;
1689  if (auto *C = dyn_cast<Constant>(B)) {
1690  if (UsesRemoved >= 1)
1691  NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1692  } else {
1693  if (UsesRemoved >= 3)
1694  NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1695  /*HasNUW=*/true),
1696  One, Or->getName());
1697  }
1698  if (NewOr) {
1699  Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1700  Cmp.setOperand(0, NewAnd);
1701  return &Cmp;
1702  }
1703  }
1704  }
1705 
1706  return nullptr;
1707 }
1708 
1709 /// Fold icmp (and X, Y), C.
1710 Instruction *InstCombiner::foldICmpAndConstant(ICmpInst &Cmp,
1711  BinaryOperator *And,
1712  const APInt &C) {
1713  if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1714  return I;
1715 
1716  // TODO: These all require that Y is constant too, so refactor with the above.
1717 
1718  // Try to optimize things like "A[i] & 42 == 0" to index computations.
1719  Value *X = And->getOperand(0);
1720  Value *Y = And->getOperand(1);
1721  if (auto *LI = dyn_cast<LoadInst>(X))
1722  if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1723  if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1724  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1725  !LI->isVolatile() && isa<ConstantInt>(Y)) {
1726  ConstantInt *C2 = cast<ConstantInt>(Y);
1727  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1728  return Res;
1729  }
1730 
1731  if (!Cmp.isEquality())
1732  return nullptr;
1733 
1734  // X & -C == -C -> X > u ~C
1735  // X & -C != -C -> X <= u ~C
1736  // iff C is a power of 2
1737  if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1738  auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT
1740  return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1741  }
1742 
1743  // (X & C2) == 0 -> (trunc X) >= 0
1744  // (X & C2) != 0 -> (trunc X) < 0
1745  // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1746  const APInt *C2;
1747  if (And->hasOneUse() && C.isNullValue() && match(Y, m_APInt(C2))) {
1748  int32_t ExactLogBase2 = C2->exactLogBase2();
1749  if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1750  Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1751  if (And->getType()->isVectorTy())
1752  NTy = VectorType::get(NTy, And->getType()->getVectorNumElements());
1753  Value *Trunc = Builder.CreateTrunc(X, NTy);
1754  auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE
1756  return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1757  }
1758  }
1759 
1760  return nullptr;
1761 }
1762 
1763 /// Fold icmp (or X, Y), C.
1764 Instruction *InstCombiner::foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or,
1765  const APInt &C) {
1766  ICmpInst::Predicate Pred = Cmp.getPredicate();
1767  if (C.isOneValue()) {
1768  // icmp slt signum(V) 1 --> icmp slt V, 1
1769  Value *V = nullptr;
1770  if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1771  return new ICmpInst(ICmpInst::ICMP_SLT, V,
1772  ConstantInt::get(V->getType(), 1));
1773  }
1774 
1775  // X | C == C --> X <=u C
1776  // X | C != C --> X >u C
1777  // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1778  if (Cmp.isEquality() && Cmp.getOperand(1) == Or->getOperand(1) &&
1779  (C + 1).isPowerOf2()) {
1781  return new ICmpInst(Pred, Or->getOperand(0), Or->getOperand(1));
1782  }
1783 
1784  if (!Cmp.isEquality() || !C.isNullValue() || !Or->hasOneUse())
1785  return nullptr;
1786 
1787  Value *P, *Q;
1788  if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1789  // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1790  // -> and (icmp eq P, null), (icmp eq Q, null).
1791  Value *CmpP =
1792  Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1793  Value *CmpQ =
1794  Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1795  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1796  return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1797  }
1798 
1799  // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1800  // a shorter form that has more potential to be folded even further.
1801  Value *X1, *X2, *X3, *X4;
1802  if (match(Or->getOperand(0), m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1803  match(Or->getOperand(1), m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1804  // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1805  // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1806  Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1807  Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1808  auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1809  return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1810  }
1811 
1812  return nullptr;
1813 }
1814 
1815 /// Fold icmp (mul X, Y), C.
1816 Instruction *InstCombiner::foldICmpMulConstant(ICmpInst &Cmp,
1817  BinaryOperator *Mul,
1818  const APInt &C) {
1819  const APInt *MulC;
1820  if (!match(Mul->getOperand(1), m_APInt(MulC)))
1821  return nullptr;
1822 
1823  // If this is a test of the sign bit and the multiply is sign-preserving with
1824  // a constant operand, use the multiply LHS operand instead.
1825  ICmpInst::Predicate Pred = Cmp.getPredicate();
1826  if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
1827  if (MulC->isNegative())
1828  Pred = ICmpInst::getSwappedPredicate(Pred);
1829  return new ICmpInst(Pred, Mul->getOperand(0),
1831  }
1832 
1833  return nullptr;
1834 }
1835 
1836 /// Fold icmp (shl 1, Y), C.
1838  const APInt &C) {
1839  Value *Y;
1840  if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
1841  return nullptr;
1842 
1843  Type *ShiftType = Shl->getType();
1844  unsigned TypeBits = C.getBitWidth();
1845  bool CIsPowerOf2 = C.isPowerOf2();
1846  ICmpInst::Predicate Pred = Cmp.getPredicate();
1847  if (Cmp.isUnsigned()) {
1848  // (1 << Y) pred C -> Y pred Log2(C)
1849  if (!CIsPowerOf2) {
1850  // (1 << Y) < 30 -> Y <= 4
1851  // (1 << Y) <= 30 -> Y <= 4
1852  // (1 << Y) >= 30 -> Y > 4
1853  // (1 << Y) > 30 -> Y > 4
1854  if (Pred == ICmpInst::ICMP_ULT)
1855  Pred = ICmpInst::ICMP_ULE;
1856  else if (Pred == ICmpInst::ICMP_UGE)
1857  Pred = ICmpInst::ICMP_UGT;
1858  }
1859 
1860  // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
1861  // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
1862  unsigned CLog2 = C.logBase2();
1863  if (CLog2 == TypeBits - 1) {
1864  if (Pred == ICmpInst::ICMP_UGE)
1865  Pred = ICmpInst::ICMP_EQ;
1866  else if (Pred == ICmpInst::ICMP_ULT)
1867  Pred = ICmpInst::ICMP_NE;
1868  }
1869  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
1870  } else if (Cmp.isSigned()) {
1871  Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
1872  if (C.isAllOnesValue()) {
1873  // (1 << Y) <= -1 -> Y == 31
1874  if (Pred == ICmpInst::ICMP_SLE)
1875  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1876 
1877  // (1 << Y) > -1 -> Y != 31
1878  if (Pred == ICmpInst::ICMP_SGT)
1879  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1880  } else if (!C) {
1881  // (1 << Y) < 0 -> Y == 31
1882  // (1 << Y) <= 0 -> Y == 31
1883  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1884  return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
1885 
1886  // (1 << Y) >= 0 -> Y != 31
1887  // (1 << Y) > 0 -> Y != 31
1888  if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
1889  return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
1890  }
1891  } else if (Cmp.isEquality() && CIsPowerOf2) {
1892  return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
1893  }
1894 
1895  return nullptr;
1896 }
1897 
1898 /// Fold icmp (shl X, Y), C.
1899 Instruction *InstCombiner::foldICmpShlConstant(ICmpInst &Cmp,
1900  BinaryOperator *Shl,
1901  const APInt &C) {
1902  const APInt *ShiftVal;
1903  if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
1904  return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
1905 
1906  const APInt *ShiftAmt;
1907  if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
1908  return foldICmpShlOne(Cmp, Shl, C);
1909 
1910  // Check that the shift amount is in range. If not, don't perform undefined
1911  // shifts. When the shift is visited, it will be simplified.
1912  unsigned TypeBits = C.getBitWidth();
1913  if (ShiftAmt->uge(TypeBits))
1914  return nullptr;
1915 
1916  ICmpInst::Predicate Pred = Cmp.getPredicate();
1917  Value *X = Shl->getOperand(0);
1918  Type *ShType = Shl->getType();
1919 
1920  // NSW guarantees that we are only shifting out sign bits from the high bits,
1921  // so we can ASHR the compare constant without needing a mask and eliminate
1922  // the shift.
1923  if (Shl->hasNoSignedWrap()) {
1924  if (Pred == ICmpInst::ICMP_SGT) {
1925  // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
1926  APInt ShiftedC = C.ashr(*ShiftAmt);
1927  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1928  }
1929  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
1930  C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
1931  APInt ShiftedC = C.ashr(*ShiftAmt);
1932  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1933  }
1934  if (Pred == ICmpInst::ICMP_SLT) {
1935  // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
1936  // (X << S) <=s C is equiv to X <=s (C >> S) for all C
1937  // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
1938  // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
1939  assert(!C.isMinSignedValue() && "Unexpected icmp slt");
1940  APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
1941  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1942  }
1943  // If this is a signed comparison to 0 and the shift is sign preserving,
1944  // use the shift LHS operand instead; isSignTest may change 'Pred', so only
1945  // do that if we're sure to not continue on in this function.
1946  if (isSignTest(Pred, C))
1947  return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
1948  }
1949 
1950  // NUW guarantees that we are only shifting out zero bits from the high bits,
1951  // so we can LSHR the compare constant without needing a mask and eliminate
1952  // the shift.
1953  if (Shl->hasNoUnsignedWrap()) {
1954  if (Pred == ICmpInst::ICMP_UGT) {
1955  // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
1956  APInt ShiftedC = C.lshr(*ShiftAmt);
1957  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1958  }
1959  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
1960  C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
1961  APInt ShiftedC = C.lshr(*ShiftAmt);
1962  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1963  }
1964  if (Pred == ICmpInst::ICMP_ULT) {
1965  // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
1966  // (X << S) <=u C is equiv to X <=u (C >> S) for all C
1967  // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
1968  // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
1969  assert(C.ugt(0) && "ult 0 should have been eliminated");
1970  APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
1971  return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
1972  }
1973  }
1974 
1975  if (Cmp.isEquality() && Shl->hasOneUse()) {
1976  // Strength-reduce the shift into an 'and'.
1978  ShType,
1979  APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
1980  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
1981  Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
1982  return new ICmpInst(Pred, And, LShrC);
1983  }
1984 
1985  // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1986  bool TrueIfSigned = false;
1987  if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
1988  // (X << 31) <s 0 --> (X & 1) != 0
1990  ShType,
1991  APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
1992  Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
1993  return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1994  And, Constant::getNullValue(ShType));
1995  }
1996 
1997  // Transform (icmp pred iM (shl iM %v, N), C)
1998  // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
1999  // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2000  // This enables us to get rid of the shift in favor of a trunc that may be
2001  // free on the target. It has the additional benefit of comparing to a
2002  // smaller constant that may be more target-friendly.
2003  unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2004  if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2005  DL.isLegalInteger(TypeBits - Amt)) {
2006  Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2007  if (ShType->isVectorTy())
2008  TruncTy = VectorType::get(TruncTy, ShType->getVectorNumElements());
2009  Constant *NewC =
2010  ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2011  return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2012  }
2013 
2014  return nullptr;
2015 }
2016 
2017 /// Fold icmp ({al}shr X, Y), C.
2018 Instruction *InstCombiner::foldICmpShrConstant(ICmpInst &Cmp,
2019  BinaryOperator *Shr,
2020  const APInt &C) {
2021  // An exact shr only shifts out zero bits, so:
2022  // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2023  Value *X = Shr->getOperand(0);
2024  CmpInst::Predicate Pred = Cmp.getPredicate();
2025  if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() &&
2026  C.isNullValue())
2027  return new ICmpInst(Pred, X, Cmp.getOperand(1));
2028 
2029  const APInt *ShiftVal;
2030  if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2031  return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2032 
2033  const APInt *ShiftAmt;
2034  if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2035  return nullptr;
2036 
2037  // Check that the shift amount is in range. If not, don't perform undefined
2038  // shifts. When the shift is visited it will be simplified.
2039  unsigned TypeBits = C.getBitWidth();
2040  unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2041  if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2042  return nullptr;
2043 
2044  bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2045  bool IsExact = Shr->isExact();
2046  Type *ShrTy = Shr->getType();
2047  // TODO: If we could guarantee that InstSimplify would handle all of the
2048  // constant-value-based preconditions in the folds below, then we could assert
2049  // those conditions rather than checking them. This is difficult because of
2050  // undef/poison (PR34838).
2051  if (IsAShr) {
2052  if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2053  // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2054  // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2055  APInt ShiftedC = C.shl(ShAmtVal);
2056  if (ShiftedC.ashr(ShAmtVal) == C)
2057  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2058  }
2059  if (Pred == CmpInst::ICMP_SGT) {
2060  // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2061  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2062  if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2063  (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2064  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2065  }
2066  } else {
2067  if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2068  // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2069  // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2070  APInt ShiftedC = C.shl(ShAmtVal);
2071  if (ShiftedC.lshr(ShAmtVal) == C)
2072  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2073  }
2074  if (Pred == CmpInst::ICMP_UGT) {
2075  // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2076  APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2077  if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2078  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2079  }
2080  }
2081 
2082  if (!Cmp.isEquality())
2083  return nullptr;
2084 
2085  // Handle equality comparisons of shift-by-constant.
2086 
2087  // If the comparison constant changes with the shift, the comparison cannot
2088  // succeed (bits of the comparison constant cannot match the shifted value).
2089  // This should be known by InstSimplify and already be folded to true/false.
2090  assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2091  (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2092  "Expected icmp+shr simplify did not occur.");
2093 
2094  // If the bits shifted out are known zero, compare the unshifted value:
2095  // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2096  if (Shr->isExact())
2097  return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2098 
2099  if (Shr->hasOneUse()) {
2100  // Canonicalize the shift into an 'and':
2101  // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2102  APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2103  Constant *Mask = ConstantInt::get(ShrTy, Val);
2104  Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2105  return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2106  }
2107 
2108  return nullptr;
2109 }
2110 
2111 /// Fold icmp (udiv X, Y), C.
2112 Instruction *InstCombiner::foldICmpUDivConstant(ICmpInst &Cmp,
2113  BinaryOperator *UDiv,
2114  const APInt &C) {
2115  const APInt *C2;
2116  if (!match(UDiv->getOperand(0), m_APInt(C2)))
2117  return nullptr;
2118 
2119  assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2120 
2121  // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2122  Value *Y = UDiv->getOperand(1);
2123  if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2124  assert(!C.isMaxValue() &&
2125  "icmp ugt X, UINT_MAX should have been simplified already.");
2126  return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2127  ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2128  }
2129 
2130  // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2131  if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2132  assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2133  return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2134  ConstantInt::get(Y->getType(), C2->udiv(C)));
2135  }
2136 
2137  return nullptr;
2138 }
2139 
2140 /// Fold icmp ({su}div X, Y), C.
2141 Instruction *InstCombiner::foldICmpDivConstant(ICmpInst &Cmp,
2142  BinaryOperator *Div,
2143  const APInt &C) {
2144  // Fold: icmp pred ([us]div X, C2), C -> range test
2145  // Fold this div into the comparison, producing a range check.
2146  // Determine, based on the divide type, what the range is being
2147  // checked. If there is an overflow on the low or high side, remember
2148  // it, otherwise compute the range [low, hi) bounding the new value.
2149  // See: InsertRangeTest above for the kinds of replacements possible.
2150  const APInt *C2;
2151  if (!match(Div->getOperand(1), m_APInt(C2)))
2152  return nullptr;
2153 
2154  // FIXME: If the operand types don't match the type of the divide
2155  // then don't attempt this transform. The code below doesn't have the
2156  // logic to deal with a signed divide and an unsigned compare (and
2157  // vice versa). This is because (x /s C2) <s C produces different
2158  // results than (x /s C2) <u C or (x /u C2) <s C or even
2159  // (x /u C2) <u C. Simply casting the operands and result won't
2160  // work. :( The if statement below tests that condition and bails
2161  // if it finds it.
2162  bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2163  if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2164  return nullptr;
2165 
2166  // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2167  // INT_MIN will also fail if the divisor is 1. Although folds of all these
2168  // division-by-constant cases should be present, we can not assert that they
2169  // have happened before we reach this icmp instruction.
2170  if (C2->isNullValue() || C2->isOneValue() ||
2171  (DivIsSigned && C2->isAllOnesValue()))
2172  return nullptr;
2173 
2174  // Compute Prod = C * C2. We are essentially solving an equation of
2175  // form X / C2 = C. We solve for X by multiplying C2 and C.
2176  // By solving for X, we can turn this into a range check instead of computing
2177  // a divide.
2178  APInt Prod = C * *C2;
2179 
2180  // Determine if the product overflows by seeing if the product is not equal to
2181  // the divide. Make sure we do the same kind of divide as in the LHS
2182  // instruction that we're folding.
2183  bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2184 
2185  ICmpInst::Predicate Pred = Cmp.getPredicate();
2186 
2187  // If the division is known to be exact, then there is no remainder from the
2188  // divide, so the covered range size is unit, otherwise it is the divisor.
2189  APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2190 
2191  // Figure out the interval that is being checked. For example, a comparison
2192  // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2193  // Compute this interval based on the constants involved and the signedness of
2194  // the compare/divide. This computes a half-open interval, keeping track of
2195  // whether either value in the interval overflows. After analysis each
2196  // overflow variable is set to 0 if it's corresponding bound variable is valid
2197  // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2198  int LoOverflow = 0, HiOverflow = 0;
2199  APInt LoBound, HiBound;
2200 
2201  if (!DivIsSigned) { // udiv
2202  // e.g. X/5 op 3 --> [15, 20)
2203  LoBound = Prod;
2204  HiOverflow = LoOverflow = ProdOV;
2205  if (!HiOverflow) {
2206  // If this is not an exact divide, then many values in the range collapse
2207  // to the same result value.
2208  HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2209  }
2210  } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2211  if (C.isNullValue()) { // (X / pos) op 0
2212  // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2213  LoBound = -(RangeSize - 1);
2214  HiBound = RangeSize;
2215  } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2216  LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2217  HiOverflow = LoOverflow = ProdOV;
2218  if (!HiOverflow)
2219  HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2220  } else { // (X / pos) op neg
2221  // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2222  HiBound = Prod + 1;
2223  LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2224  if (!LoOverflow) {
2225  APInt DivNeg = -RangeSize;
2226  LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2227  }
2228  }
2229  } else if (C2->isNegative()) { // Divisor is < 0.
2230  if (Div->isExact())
2231  RangeSize.negate();
2232  if (C.isNullValue()) { // (X / neg) op 0
2233  // e.g. X/-5 op 0 --> [-4, 5)
2234  LoBound = RangeSize + 1;
2235  HiBound = -RangeSize;
2236  if (HiBound == *C2) { // -INTMIN = INTMIN
2237  HiOverflow = 1; // [INTMIN+1, overflow)
2238  HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2239  }
2240  } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2241  // e.g. X/-5 op 3 --> [-19, -14)
2242  HiBound = Prod + 1;
2243  HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2244  if (!LoOverflow)
2245  LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2246  } else { // (X / neg) op neg
2247  LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2248  LoOverflow = HiOverflow = ProdOV;
2249  if (!HiOverflow)
2250  HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2251  }
2252 
2253  // Dividing by a negative swaps the condition. LT <-> GT
2254  Pred = ICmpInst::getSwappedPredicate(Pred);
2255  }
2256 
2257  Value *X = Div->getOperand(0);
2258  switch (Pred) {
2259  default: llvm_unreachable("Unhandled icmp opcode!");
2260  case ICmpInst::ICMP_EQ:
2261  if (LoOverflow && HiOverflow)
2262  return replaceInstUsesWith(Cmp, Builder.getFalse());
2263  if (HiOverflow)
2264  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2265  ICmpInst::ICMP_UGE, X,
2266  ConstantInt::get(Div->getType(), LoBound));
2267  if (LoOverflow)
2268  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2269  ICmpInst::ICMP_ULT, X,
2270  ConstantInt::get(Div->getType(), HiBound));
2271  return replaceInstUsesWith(
2272  Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2273  case ICmpInst::ICMP_NE:
2274  if (LoOverflow && HiOverflow)
2275  return replaceInstUsesWith(Cmp, Builder.getTrue());
2276  if (HiOverflow)
2277  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2278  ICmpInst::ICMP_ULT, X,
2279  ConstantInt::get(Div->getType(), LoBound));
2280  if (LoOverflow)
2281  return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2282  ICmpInst::ICMP_UGE, X,
2283  ConstantInt::get(Div->getType(), HiBound));
2284  return replaceInstUsesWith(Cmp,
2285  insertRangeTest(X, LoBound, HiBound,
2286  DivIsSigned, false));
2287  case ICmpInst::ICMP_ULT:
2288  case ICmpInst::ICMP_SLT:
2289  if (LoOverflow == +1) // Low bound is greater than input range.
2290  return replaceInstUsesWith(Cmp, Builder.getTrue());
2291  if (LoOverflow == -1) // Low bound is less than input range.
2292  return replaceInstUsesWith(Cmp, Builder.getFalse());
2293  return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2294  case ICmpInst::ICMP_UGT:
2295  case ICmpInst::ICMP_SGT:
2296  if (HiOverflow == +1) // High bound greater than input range.
2297  return replaceInstUsesWith(Cmp, Builder.getFalse());
2298  if (HiOverflow == -1) // High bound less than input range.
2299  return replaceInstUsesWith(Cmp, Builder.getTrue());
2300  if (Pred == ICmpInst::ICMP_UGT)
2301  return new ICmpInst(ICmpInst::ICMP_UGE, X,
2302  ConstantInt::get(Div->getType(), HiBound));
2303  return new ICmpInst(ICmpInst::ICMP_SGE, X,
2304  ConstantInt::get(Div->getType(), HiBound));
2305  }
2306 
2307  return nullptr;
2308 }
2309 
2310 /// Fold icmp (sub X, Y), C.
2311 Instruction *InstCombiner::foldICmpSubConstant(ICmpInst &Cmp,
2312  BinaryOperator *Sub,
2313  const APInt &C) {
2314  Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2315  ICmpInst::Predicate Pred = Cmp.getPredicate();
2316 
2317  // The following transforms are only worth it if the only user of the subtract
2318  // is the icmp.
2319  if (!Sub->hasOneUse())
2320  return nullptr;
2321 
2322  if (Sub->hasNoSignedWrap()) {
2323  // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2324  if (Pred == ICmpInst::ICMP_SGT && C.isAllOnesValue())
2325  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2326 
2327  // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2328  if (Pred == ICmpInst::ICMP_SGT && C.isNullValue())
2329  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2330 
2331  // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2332  if (Pred == ICmpInst::ICMP_SLT && C.isNullValue())
2333  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2334 
2335  // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2336  if (Pred == ICmpInst::ICMP_SLT && C.isOneValue())
2337  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2338  }
2339 
2340  const APInt *C2;
2341  if (!match(X, m_APInt(C2)))
2342  return nullptr;
2343 
2344  // C2 - Y <u C -> (Y | (C - 1)) == C2
2345  // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2346  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2347  (*C2 & (C - 1)) == (C - 1))
2348  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2349 
2350  // C2 - Y >u C -> (Y | C) != C2
2351  // iff C2 & C == C and C + 1 is a power of 2
2352  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2353  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2354 
2355  return nullptr;
2356 }
2357 
2358 /// Fold icmp (add X, Y), C.
2359 Instruction *InstCombiner::foldICmpAddConstant(ICmpInst &Cmp,
2361  const APInt &C) {
2362  Value *Y = Add->getOperand(1);
2363  const APInt *C2;
2364  if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2365  return nullptr;
2366 
2367  // Fold icmp pred (add X, C2), C.
2368  Value *X = Add->getOperand(0);
2369  Type *Ty = Add->getType();
2370  CmpInst::Predicate Pred = Cmp.getPredicate();
2371 
2372  if (!Add->hasOneUse())
2373  return nullptr;
2374 
2375  // If the add does not wrap, we can always adjust the compare by subtracting
2376  // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2377  // are canonicalized to SGT/SLT/UGT/ULT.
2378  if ((Add->hasNoSignedWrap() &&
2379  (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2380  (Add->hasNoUnsignedWrap() &&
2381  (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2382  bool Overflow;
2383  APInt NewC =
2384  Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2385  // If there is overflow, the result must be true or false.
2386  // TODO: Can we assert there is no overflow because InstSimplify always
2387  // handles those cases?
2388  if (!Overflow)
2389  // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2390  return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2391  }
2392 
2393  auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2394  const APInt &Upper = CR.getUpper();
2395  const APInt &Lower = CR.getLower();
2396  if (Cmp.isSigned()) {
2397  if (Lower.isSignMask())
2398  return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2399  if (Upper.isSignMask())
2400  return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2401  } else {
2402  if (Lower.isMinValue())
2403  return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2404  if (Upper.isMinValue())
2405  return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2406  }
2407 
2408  // X+C <u C2 -> (X & -C2) == C
2409  // iff C & (C2-1) == 0
2410  // C2 is a power of 2
2411  if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2412  return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2413  ConstantExpr::getNeg(cast<Constant>(Y)));
2414 
2415  // X+C >u C2 -> (X & ~C2) != C
2416  // iff C & C2 == 0
2417  // C2+1 is a power of 2
2418  if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2419  return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2420  ConstantExpr::getNeg(cast<Constant>(Y)));
2421 
2422  return nullptr;
2423 }
2424 
2425 bool InstCombiner::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2426  Value *&RHS, ConstantInt *&Less,
2427  ConstantInt *&Equal,
2428  ConstantInt *&Greater) {
2429  // TODO: Generalize this to work with other comparison idioms or ensure
2430  // they get canonicalized into this form.
2431 
2432  // select i1 (a == b), i32 Equal, i32 (select i1 (a < b), i32 Less, i32
2433  // Greater), where Equal, Less and Greater are placeholders for any three
2434  // constants.
2435  ICmpInst::Predicate PredA, PredB;
2436  if (match(SI->getTrueValue(), m_ConstantInt(Equal)) &&
2437  match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) &&
2438  PredA == ICmpInst::ICMP_EQ &&
2439  match(SI->getFalseValue(),
2440  m_Select(m_ICmp(PredB, m_Specific(LHS), m_Specific(RHS)),
2441  m_ConstantInt(Less), m_ConstantInt(Greater))) &&
2442  PredB == ICmpInst::ICMP_SLT) {
2443  return true;
2444  }
2445  return false;
2446 }
2447 
2448 Instruction *InstCombiner::foldICmpSelectConstant(ICmpInst &Cmp,
2449  SelectInst *Select,
2450  ConstantInt *C) {
2451 
2452  assert(C && "Cmp RHS should be a constant int!");
2453  // If we're testing a constant value against the result of a three way
2454  // comparison, the result can be expressed directly in terms of the
2455  // original values being compared. Note: We could possibly be more
2456  // aggressive here and remove the hasOneUse test. The original select is
2457  // really likely to simplify or sink when we remove a test of the result.
2458  Value *OrigLHS, *OrigRHS;
2459  ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2460  if (Cmp.hasOneUse() &&
2461  matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2462  C3GreaterThan)) {
2463  assert(C1LessThan && C2Equal && C3GreaterThan);
2464 
2465  bool TrueWhenLessThan =
2466  ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2467  ->isAllOnesValue();
2468  bool TrueWhenEqual =
2469  ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2470  ->isAllOnesValue();
2471  bool TrueWhenGreaterThan =
2472  ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2473  ->isAllOnesValue();
2474 
2475  // This generates the new instruction that will replace the original Cmp
2476  // Instruction. Instead of enumerating the various combinations when
2477  // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2478  // false, we rely on chaining of ORs and future passes of InstCombine to
2479  // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2480 
2481  // When none of the three constants satisfy the predicate for the RHS (C),
2482  // the entire original Cmp can be simplified to a false.
2483  Value *Cond = Builder.getFalse();
2484  if (TrueWhenLessThan)
2485  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT, OrigLHS, OrigRHS));
2486  if (TrueWhenEqual)
2487  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ, OrigLHS, OrigRHS));
2488  if (TrueWhenGreaterThan)
2489  Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT, OrigLHS, OrigRHS));
2490 
2491  return replaceInstUsesWith(Cmp, Cond);
2492  }
2493  return nullptr;
2494 }
2495 
2496 Instruction *InstCombiner::foldICmpBitCastConstant(ICmpInst &Cmp,
2498  const APInt &C) {
2499  // Folding: icmp <pred> iN X, C
2500  // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2501  // and C is a splat of a K-bit pattern
2502  // and SC is a constant vector = <C', C', C', ..., C'>
2503  // Into:
2504  // %E = extractelement <M x iK> %vec, i32 C'
2505  // icmp <pred> iK %E, trunc(C)
2506  if (!Bitcast->getType()->isIntegerTy() ||
2507  !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2508  return nullptr;
2509 
2510  Value *BCIOp = Bitcast->getOperand(0);
2511  Value *Vec = nullptr; // 1st vector arg of the shufflevector
2512  Constant *Mask = nullptr; // Mask arg of the shufflevector
2513  if (match(BCIOp,
2514  m_ShuffleVector(m_Value(Vec), m_Undef(), m_Constant(Mask)))) {
2515  // Check whether every element of Mask is the same constant
2516  if (auto *Elem = dyn_cast_or_null<ConstantInt>(Mask->getSplatValue())) {
2517  auto *VecTy = cast<VectorType>(BCIOp->getType());
2518  auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2519  auto Pred = Cmp.getPredicate();
2520  if (C.isSplat(EltTy->getBitWidth())) {
2521  // Fold the icmp based on the value of C
2522  // If C is M copies of an iK sized bit pattern,
2523  // then:
2524  // => %E = extractelement <N x iK> %vec, i32 Elem
2525  // icmp <pred> iK %SplatVal, <pattern>
2526  Value *Extract = Builder.CreateExtractElement(Vec, Elem);
2527  Value *NewC = ConstantInt::get(EltTy, C.trunc(EltTy->getBitWidth()));
2528  return new ICmpInst(Pred, Extract, NewC);
2529  }
2530  }
2531  }
2532  return nullptr;
2533 }
2534 
2535 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
2536 /// where X is some kind of instruction.
2537 Instruction *InstCombiner::foldICmpInstWithConstant(ICmpInst &Cmp) {
2538  const APInt *C;
2539  if (!match(Cmp.getOperand(1), m_APInt(C)))
2540  return nullptr;
2541 
2542  if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
2543  switch (BO->getOpcode()) {
2544  case Instruction::Xor:
2545  if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
2546  return I;
2547  break;
2548  case Instruction::And:
2549  if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
2550  return I;
2551  break;
2552  case Instruction::Or:
2553  if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
2554  return I;
2555  break;
2556  case Instruction::Mul:
2557  if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
2558  return I;
2559  break;
2560  case Instruction::Shl:
2561  if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
2562  return I;
2563  break;
2564  case Instruction::LShr:
2565  case Instruction::AShr:
2566  if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
2567  return I;
2568  break;
2569  case Instruction::UDiv:
2570  if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
2571  return I;
2573  case Instruction::SDiv:
2574  if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
2575  return I;
2576  break;
2577  case Instruction::Sub:
2578  if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
2579  return I;
2580  break;
2581  case Instruction::Add:
2582  if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
2583  return I;
2584  break;
2585  default:
2586  break;
2587  }
2588  // TODO: These folds could be refactored to be part of the above calls.
2589  if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
2590  return I;
2591  }
2592 
2593  // Match against CmpInst LHS being instructions other than binary operators.
2594 
2595  if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
2596  // For now, we only support constant integers while folding the
2597  // ICMP(SELECT)) pattern. We can extend this to support vector of integers
2598  // similar to the cases handled by binary ops above.
2599  if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
2600  if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
2601  return I;
2602  }
2603 
2604  if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
2605  if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
2606  return I;
2607  }
2608 
2609  if (auto *BCI = dyn_cast<BitCastInst>(Cmp.getOperand(0))) {
2610  if (Instruction *I = foldICmpBitCastConstant(Cmp, BCI, *C))
2611  return I;
2612  }
2613 
2614  if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, *C))
2615  return I;
2616 
2617  return nullptr;
2618 }
2619 
2620 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
2621 /// icmp eq/ne BO, C.
2622 Instruction *InstCombiner::foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp,
2623  BinaryOperator *BO,
2624  const APInt &C) {
2625  // TODO: Some of these folds could work with arbitrary constants, but this
2626  // function is limited to scalar and vector splat constants.
2627  if (!Cmp.isEquality())
2628  return nullptr;
2629 
2630  ICmpInst::Predicate Pred = Cmp.getPredicate();
2631  bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
2632  Constant *RHS = cast<Constant>(Cmp.getOperand(1));
2633  Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
2634 
2635  switch (BO->getOpcode()) {
2636  case Instruction::SRem:
2637  // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
2638  if (C.isNullValue() && BO->hasOneUse()) {
2639  const APInt *BOC;
2640  if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
2641  Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
2642  return new ICmpInst(Pred, NewRem,
2644  }
2645  }
2646  break;
2647  case Instruction::Add: {
2648  // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
2649  const APInt *BOC;
2650  if (match(BOp1, m_APInt(BOC))) {
2651  if (BO->hasOneUse()) {
2652  Constant *SubC = ConstantExpr::getSub(RHS, cast<Constant>(BOp1));
2653  return new ICmpInst(Pred, BOp0, SubC);
2654  }
2655  } else if (C.isNullValue()) {
2656  // Replace ((add A, B) != 0) with (A != -B) if A or B is
2657  // efficiently invertible, or if the add has just this one use.
2658  if (Value *NegVal = dyn_castNegVal(BOp1))
2659  return new ICmpInst(Pred, BOp0, NegVal);
2660  if (Value *NegVal = dyn_castNegVal(BOp0))
2661  return new ICmpInst(Pred, NegVal, BOp1);
2662  if (BO->hasOneUse()) {
2663  Value *Neg = Builder.CreateNeg(BOp1);
2664  Neg->takeName(BO);
2665  return new ICmpInst(Pred, BOp0, Neg);
2666  }
2667  }
2668  break;
2669  }
2670  case Instruction::Xor:
2671  if (BO->hasOneUse()) {
2672  if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
2673  // For the xor case, we can xor two constants together, eliminating
2674  // the explicit xor.
2675  return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
2676  } else if (C.isNullValue()) {
2677  // Replace ((xor A, B) != 0) with (A != B)
2678  return new ICmpInst(Pred, BOp0, BOp1);
2679  }
2680  }
2681  break;
2682  case Instruction::Sub:
2683  if (BO->hasOneUse()) {
2684  const APInt *BOC;
2685  if (match(BOp0, m_APInt(BOC))) {
2686  // Replace ((sub BOC, B) != C) with (B != BOC-C).
2687  Constant *SubC = ConstantExpr::getSub(cast<Constant>(BOp0), RHS);
2688  return new ICmpInst(Pred, BOp1, SubC);
2689  } else if (C.isNullValue()) {
2690  // Replace ((sub A, B) != 0) with (A != B).
2691  return new ICmpInst(Pred, BOp0, BOp1);
2692  }
2693  }
2694  break;
2695  case Instruction::Or: {
2696  const APInt *BOC;
2697  if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
2698  // Comparing if all bits outside of a constant mask are set?
2699  // Replace (X | C) == -1 with (X & ~C) == ~C.
2700  // This removes the -1 constant.
2701  Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
2702  Value *And = Builder.CreateAnd(BOp0, NotBOC);
2703  return new ICmpInst(Pred, And, NotBOC);
2704  }
2705  break;
2706  }
2707  case Instruction::And: {
2708  const APInt *BOC;
2709  if (match(BOp1, m_APInt(BOC))) {
2710  // If we have ((X & C) == C), turn it into ((X & C) != 0).
2711  if (C == *BOC && C.isPowerOf2())
2712  return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
2713  BO, Constant::getNullValue(RHS->getType()));
2714 
2715  // Don't perform the following transforms if the AND has multiple uses
2716  if (!BO->hasOneUse())
2717  break;
2718 
2719  // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
2720  if (BOC->isSignMask()) {
2721  Constant *Zero = Constant::getNullValue(BOp0->getType());
2722  auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
2723  return new ICmpInst(NewPred, BOp0, Zero);
2724  }
2725 
2726  // ((X & ~7) == 0) --> X < 8
2727  if (C.isNullValue() && (~(*BOC) + 1).isPowerOf2()) {
2728  Constant *NegBOC = ConstantExpr::getNeg(cast<Constant>(BOp1));
2729  auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
2730  return new ICmpInst(NewPred, BOp0, NegBOC);
2731  }
2732  }
2733  break;
2734  }
2735  case Instruction::Mul:
2736  if (C.isNullValue() && BO->hasNoSignedWrap()) {
2737  const APInt *BOC;
2738  if (match(BOp1, m_APInt(BOC)) && !BOC->isNullValue()) {
2739  // The trivial case (mul X, 0) is handled by InstSimplify.
2740  // General case : (mul X, C) != 0 iff X != 0
2741  // (mul X, C) == 0 iff X == 0
2742  return new ICmpInst(Pred, BOp0, Constant::getNullValue(RHS->getType()));
2743  }
2744  }
2745  break;
2746  case Instruction::UDiv:
2747  if (C.isNullValue()) {
2748  // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
2749  auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
2750  return new ICmpInst(NewPred, BOp1, BOp0);
2751  }
2752  break;
2753  default:
2754  break;
2755  }
2756  return nullptr;
2757 }
2758 
2759 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
2760 Instruction *InstCombiner::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
2761  const APInt &C) {
2763  if (!II || !Cmp.isEquality())
2764  return nullptr;
2765 
2766  // Handle icmp {eq|ne} <intrinsic>, Constant.
2767  Type *Ty = II->getType();
2768  switch (II->getIntrinsicID()) {
2769  case Intrinsic::bswap:
2770  Worklist.Add(II);
2771  Cmp.setOperand(0, II->getArgOperand(0));
2772  Cmp.setOperand(1, ConstantInt::get(Ty, C.byteSwap()));
2773  return &Cmp;
2774 
2775  case Intrinsic::ctlz:
2776  case Intrinsic::cttz:
2777  // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
2778  if (C == C.getBitWidth()) {
2779  Worklist.Add(II);
2780  Cmp.setOperand(0, II->getArgOperand(0));
2782  return &Cmp;
2783  }
2784  break;
2785 
2786  case Intrinsic::ctpop: {
2787  // popcount(A) == 0 -> A == 0 and likewise for !=
2788  // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
2789  bool IsZero = C.isNullValue();
2790  if (IsZero || C == C.getBitWidth()) {
2791  Worklist.Add(II);
2792  Cmp.setOperand(0, II->getArgOperand(0));
2793  auto *NewOp =
2795  Cmp.setOperand(1, NewOp);
2796  return &Cmp;
2797  }
2798  break;
2799  }
2800  default:
2801  break;
2802  }
2803 
2804  return nullptr;
2805 }
2806 
2807 /// Handle icmp with constant (but not simple integer constant) RHS.
2808 Instruction *InstCombiner::foldICmpInstWithConstantNotInt(ICmpInst &I) {
2809  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2810  Constant *RHSC = dyn_cast<Constant>(Op1);
2811  Instruction *LHSI = dyn_cast<Instruction>(Op0);
2812  if (!RHSC || !LHSI)
2813  return nullptr;
2814 
2815  switch (LHSI->getOpcode()) {
2816  case Instruction::GetElementPtr:
2817  // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2818  if (RHSC->isNullValue() &&
2819  cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2820  return new ICmpInst(
2821  I.getPredicate(), LHSI->getOperand(0),
2822  Constant::getNullValue(LHSI->getOperand(0)->getType()));
2823  break;
2824  case Instruction::PHI:
2825  // Only fold icmp into the PHI if the phi and icmp are in the same
2826  // block. If in the same block, we're encouraging jump threading. If
2827  // not, we are just pessimizing the code by making an i1 phi.
2828  if (LHSI->getParent() == I.getParent())
2829  if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
2830  return NV;
2831  break;
2832  case Instruction::Select: {
2833  // If either operand of the select is a constant, we can fold the
2834  // comparison into the select arms, which will cause one to be
2835  // constant folded and the select turned into a bitwise or.
2836  Value *Op1 = nullptr, *Op2 = nullptr;
2837  ConstantInt *CI = nullptr;
2838  if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2839  Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2840  CI = dyn_cast<ConstantInt>(Op1);
2841  }
2842  if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2843  Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2844  CI = dyn_cast<ConstantInt>(Op2);
2845  }
2846 
2847  // We only want to perform this transformation if it will not lead to
2848  // additional code. This is true if either both sides of the select
2849  // fold to a constant (in which case the icmp is replaced with a select
2850  // which will usually simplify) or this is the only user of the
2851  // select (in which case we are trading a select+icmp for a simpler
2852  // select+icmp) or all uses of the select can be replaced based on
2853  // dominance information ("Global cases").
2854  bool Transform = false;
2855  if (Op1 && Op2)
2856  Transform = true;
2857  else if (Op1 || Op2) {
2858  // Local case
2859  if (LHSI->hasOneUse())
2860  Transform = true;
2861  // Global cases
2862  else if (CI && !CI->isZero())
2863  // When Op1 is constant try replacing select with second operand.
2864  // Otherwise Op2 is constant and try replacing select with first
2865  // operand.
2866  Transform =
2867  replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
2868  }
2869  if (Transform) {
2870  if (!Op1)
2871  Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
2872  I.getName());
2873  if (!Op2)
2874  Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
2875  I.getName());
2876  return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2877  }
2878  break;
2879  }
2880  case Instruction::IntToPtr:
2881  // icmp pred inttoptr(X), null -> icmp pred X, 0
2882  if (RHSC->isNullValue() &&
2883  DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
2884  return new ICmpInst(
2885  I.getPredicate(), LHSI->getOperand(0),
2886  Constant::getNullValue(LHSI->getOperand(0)->getType()));
2887  break;
2888 
2889  case Instruction::Load:
2890  // Try to optimize things like "A[i] > 4" to index computations.
2891  if (GetElementPtrInst *GEP =
2892  dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2893  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2894  if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2895  !cast<LoadInst>(LHSI)->isVolatile())
2896  if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
2897  return Res;
2898  }
2899  break;
2900  }
2901 
2902  return nullptr;
2903 }
2904 
2905 /// Some comparisons can be simplified.
2906 /// In this case, we are looking for comparisons that look like
2907 /// a check for a lossy truncation.
2908 /// Folds:
2909 /// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
2910 /// Where Mask is some pattern that produces all-ones in low bits:
2911 /// (-1 >> y)
2912 /// ((-1 << y) >> y) <- non-canonical, has extra uses
2913 /// ~(-1 << y)
2914 /// ((1 << y) + (-1)) <- non-canonical, has extra uses
2915 /// The Mask can be a constant, too.
2916 /// For some predicates, the operands are commutative.
2917 /// For others, x can only be on a specific side.
2919  InstCombiner::BuilderTy &Builder) {
2920  ICmpInst::Predicate SrcPred;
2921  Value *X, *M, *Y;
2922  auto m_VariableMask = m_CombineOr(
2924  m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
2926  m_LShr(m_Shl(m_AllOnes(), m_Value(Y)), m_Deferred(Y))));
2927  auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
2928  if (!match(&I, m_c_ICmp(SrcPred,
2929  m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)),
2930  m_Deferred(X))))
2931  return nullptr;
2932 
2933  ICmpInst::Predicate DstPred;
2934  switch (SrcPred) {
2935  case ICmpInst::Predicate::ICMP_EQ:
2936  // x & (-1 >> y) == x -> x u<= (-1 >> y)
2937  DstPred = ICmpInst::Predicate::ICMP_ULE;
2938  break;
2939  case ICmpInst::Predicate::ICMP_NE:
2940  // x & (-1 >> y) != x -> x u> (-1 >> y)
2941  DstPred = ICmpInst::Predicate::ICMP_UGT;
2942  break;
2943  case ICmpInst::Predicate::ICMP_UGT:
2944  // x u> x & (-1 >> y) -> x u> (-1 >> y)
2945  assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
2946  DstPred = ICmpInst::Predicate::ICMP_UGT;
2947  break;
2948  case ICmpInst::Predicate::ICMP_UGE:
2949  // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
2950  assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
2951  DstPred = ICmpInst::Predicate::ICMP_ULE;
2952  break;
2953  case ICmpInst::Predicate::ICMP_ULT:
2954  // x & (-1 >> y) u< x -> x u> (-1 >> y)
2955  assert(X == I.getOperand(1) && "instsimplify took care of commut. variant");
2956  DstPred = ICmpInst::Predicate::ICMP_UGT;
2957  break;
2958  case ICmpInst::Predicate::ICMP_ULE:
2959  // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
2960  assert(X == I.getOperand(0) && "instsimplify took care of commut. variant");
2961  DstPred = ICmpInst::Predicate::ICMP_ULE;
2962  break;
2963  case ICmpInst::Predicate::ICMP_SGT:
2964  // x s> x & (-1 >> y) -> x s> (-1 >> y)
2965  if (X != I.getOperand(0)) // X must be on LHS of comparison!
2966  return nullptr; // Ignore the other case.
2967  DstPred = ICmpInst::Predicate::ICMP_SGT;
2968  break;
2969  case ICmpInst::Predicate::ICMP_SGE:
2970  // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
2971  if (X != I.getOperand(1)) // X must be on RHS of comparison!
2972  return nullptr; // Ignore the other case.
2973  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
2974  return nullptr;
2975  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
2976  return nullptr;
2977  DstPred = ICmpInst::Predicate::ICMP_SLE;
2978  break;
2979  case ICmpInst::Predicate::ICMP_SLT:
2980  // x & (-1 >> y) s< x -> x s> (-1 >> y)
2981  if (X != I.getOperand(1)) // X must be on RHS of comparison!
2982  return nullptr; // Ignore the other case.
2983  if (!match(M, m_Constant())) // Can not do this fold with non-constant.
2984  return nullptr;
2985  if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
2986  return nullptr;
2987  DstPred = ICmpInst::Predicate::ICMP_SGT;
2988  break;
2989  case ICmpInst::Predicate::ICMP_SLE:
2990  // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
2991  if (X != I.getOperand(0)) // X must be on LHS of comparison!
2992  return nullptr; // Ignore the other case.
2993  DstPred = ICmpInst::Predicate::ICMP_SLE;
2994  break;
2995  default:
2996  llvm_unreachable("All possible folds are handled.");
2997  }
2998 
2999  return Builder.CreateICmp(DstPred, X, M);
3000 }
3001 
3002 /// Some comparisons can be simplified.
3003 /// In this case, we are looking for comparisons that look like
3004 /// a check for a lossy signed truncation.
3005 /// Folds: (MaskedBits is a constant.)
3006 /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3007 /// Into:
3008 /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3009 /// Where KeptBits = bitwidth(%x) - MaskedBits
3010 static Value *
3012  InstCombiner::BuilderTy &Builder) {
3013  ICmpInst::Predicate SrcPred;
3014  Value *X;
3015  const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3016  // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3017  if (!match(&I, m_c_ICmp(SrcPred,
3018  m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
3019  m_APInt(C1))),
3020  m_Deferred(X))))
3021  return nullptr;
3022 
3023  // Potential handling of non-splats: for each element:
3024  // * if both are undef, replace with constant 0.
3025  // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3026  // * if both are not undef, and are different, bailout.
3027  // * else, only one is undef, then pick the non-undef one.
3028 
3029  // The shift amount must be equal.
3030  if (*C0 != *C1)
3031  return nullptr;
3032  const APInt &MaskedBits = *C0;
3033  assert(MaskedBits != 0 && "shift by zero should be folded away already.");
3034 
3035  ICmpInst::Predicate DstPred;
3036  switch (SrcPred) {
3037  case ICmpInst::Predicate::ICMP_EQ:
3038  // ((%x << MaskedBits) a>> MaskedBits) == %x
3039  // =>
3040  // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3041  DstPred = ICmpInst::Predicate::ICMP_ULT;
3042  break;
3043  case ICmpInst::Predicate::ICMP_NE:
3044  // ((%x << MaskedBits) a>> MaskedBits) != %x
3045  // =>
3046  // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3047  DstPred = ICmpInst::Predicate::ICMP_UGE;
3048  break;
3049  // FIXME: are more folds possible?
3050  default:
3051  return nullptr;
3052  }
3053 
3054  auto *XType = X->getType();
3055  const unsigned XBitWidth = XType->getScalarSizeInBits();
3056  const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3057  assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
3058 
3059  // KeptBits = bitwidth(%x) - MaskedBits
3060  const APInt KeptBits = BitWidth - MaskedBits;
3061  assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
3062  // ICmpCst = (1 << KeptBits)
3063  const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3064  assert(ICmpCst.isPowerOf2());
3065  // AddCst = (1 << (KeptBits-1))
3066  const APInt AddCst = ICmpCst.lshr(1);
3067  assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
3068 
3069  // T0 = add %x, AddCst
3070  Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3071  // T1 = T0 DstPred ICmpCst
3072  Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3073 
3074  return T1;
3075 }
3076 
3077 /// Try to fold icmp (binop), X or icmp X, (binop).
3078 /// TODO: A large part of this logic is duplicated in InstSimplify's
3079 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3080 /// duplication.
3081 Instruction *InstCombiner::foldICmpBinOp(ICmpInst &I) {
3082  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3083 
3084  // Special logic for binary operators.
3085  BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3086  BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3087  if (!BO0 && !BO1)
3088  return nullptr;
3089 
3090  const CmpInst::Predicate Pred = I.getPredicate();
3091  Value *X;
3092 
3093  // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3094  // (Op1 + X) <u Op1 --> ~Op1 <u X
3095  // Op0 >u (Op0 + X) --> X >u ~Op0
3096  if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
3097  Pred == ICmpInst::ICMP_ULT)
3098  return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3099  if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
3100  Pred == ICmpInst::ICMP_UGT)
3101  return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3102 
3103  bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3104  if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3105  NoOp0WrapProblem =
3106  ICmpInst::isEquality(Pred) ||
3107  (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3108  (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3109  if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3110  NoOp1WrapProblem =
3111  ICmpInst::isEquality(Pred) ||
3112  (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3113  (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3114 
3115  // Analyze the case when either Op0 or Op1 is an add instruction.
3116  // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3117  Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3118  if (BO0 && BO0->getOpcode() == Instruction::Add) {
3119  A = BO0->getOperand(0);
3120  B = BO0->getOperand(1);
3121  }
3122  if (BO1 && BO1->getOpcode() == Instruction::Add) {
3123  C = BO1->getOperand(0);
3124  D = BO1->getOperand(1);
3125  }
3126 
3127  // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
3128  if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3129  return new ICmpInst(Pred, A == Op1 ? B : A,
3130  Constant::getNullValue(Op1->getType()));
3131 
3132  // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
3133  if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
3134  return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
3135  C == Op0 ? D : C);
3136 
3137  // icmp (X+Y), (X+Z) -> icmp Y, Z for equalities or if there is no overflow.
3138  if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
3139  NoOp1WrapProblem &&
3140  // Try not to increase register pressure.
3141  BO0->hasOneUse() && BO1->hasOneUse()) {
3142  // Determine Y and Z in the form icmp (X+Y), (X+Z).
3143  Value *Y, *Z;
3144  if (A == C) {
3145  // C + B == C + D -> B == D
3146  Y = B;
3147  Z = D;
3148  } else if (A == D) {
3149  // D + B == C + D -> B == C
3150  Y = B;
3151  Z = C;
3152  } else if (B == C) {
3153  // A + C == C + D -> A == D
3154  Y = A;
3155  Z = D;
3156  } else {
3157  assert(B == D);
3158  // A + D == C + D -> A == C
3159  Y = A;
3160  Z = C;
3161  }
3162  return new ICmpInst(Pred, Y, Z);
3163  }
3164 
3165  // icmp slt (X + -1), Y -> icmp sle X, Y
3166  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
3167  match(B, m_AllOnes()))
3168  return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
3169 
3170  // icmp sge (X + -1), Y -> icmp sgt X, Y
3171  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
3172  match(B, m_AllOnes()))
3173  return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
3174 
3175  // icmp sle (X + 1), Y -> icmp slt X, Y
3176  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
3177  return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
3178 
3179  // icmp sgt (X + 1), Y -> icmp sge X, Y
3180  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
3181  return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
3182 
3183  // icmp sgt X, (Y + -1) -> icmp sge X, Y
3184  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
3185  match(D, m_AllOnes()))
3186  return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
3187 
3188  // icmp sle X, (Y + -1) -> icmp slt X, Y
3189  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
3190  match(D, m_AllOnes()))
3191  return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
3192 
3193  // icmp sge X, (Y + 1) -> icmp sgt X, Y
3194  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
3195  return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
3196 
3197  // icmp slt X, (Y + 1) -> icmp sle X, Y
3198  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
3199  return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
3200 
3201  // TODO: The subtraction-related identities shown below also hold, but
3202  // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
3203  // wouldn't happen even if they were implemented.
3204  //
3205  // icmp ult (X - 1), Y -> icmp ule X, Y
3206  // icmp uge (X - 1), Y -> icmp ugt X, Y
3207  // icmp ugt X, (Y - 1) -> icmp uge X, Y
3208  // icmp ule X, (Y - 1) -> icmp ult X, Y
3209 
3210  // icmp ule (X + 1), Y -> icmp ult X, Y
3211  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
3212  return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
3213 
3214  // icmp ugt (X + 1), Y -> icmp uge X, Y
3215  if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
3216  return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
3217 
3218  // icmp uge X, (Y + 1) -> icmp ugt X, Y
3219  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
3220  return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
3221 
3222  // icmp ult X, (Y + 1) -> icmp ule X, Y
3223  if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
3224  return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
3225 
3226  // if C1 has greater magnitude than C2:
3227  // icmp (X + C1), (Y + C2) -> icmp (X + C3), Y
3228  // s.t. C3 = C1 - C2
3229  //
3230  // if C2 has greater magnitude than C1:
3231  // icmp (X + C1), (Y + C2) -> icmp X, (Y + C3)
3232  // s.t. C3 = C2 - C1
3233  if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
3234  (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
3235  if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
3236  if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
3237  const APInt &AP1 = C1->getValue();
3238  const APInt &AP2 = C2->getValue();
3239  if (AP1.isNegative() == AP2.isNegative()) {
3240  APInt AP1Abs = C1->getValue().abs();
3241  APInt AP2Abs = C2->getValue().abs();
3242  if (AP1Abs.uge(AP2Abs)) {
3243  ConstantInt *C3 = Builder.getInt(AP1 - AP2);
3244  Value *NewAdd = Builder.CreateNSWAdd(A, C3);
3245  return new ICmpInst(Pred, NewAdd, C);
3246  } else {
3247  ConstantInt *C3 = Builder.getInt(AP2 - AP1);
3248  Value *NewAdd = Builder.CreateNSWAdd(C, C3);
3249  return new ICmpInst(Pred, A, NewAdd);
3250  }
3251  }
3252  }
3253 
3254  // Analyze the case when either Op0 or Op1 is a sub instruction.
3255  // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
3256  A = nullptr;
3257  B = nullptr;
3258  C = nullptr;
3259  D = nullptr;
3260  if (BO0 && BO0->getOpcode() == Instruction::Sub) {
3261  A = BO0->getOperand(0);
3262  B = BO0->getOperand(1);
3263  }
3264  if (BO1 && BO1->getOpcode() == Instruction::Sub) {
3265  C = BO1->getOperand(0);
3266  D = BO1->getOperand(1);
3267  }
3268 
3269  // icmp (X-Y), X -> icmp 0, Y for equalities or if there is no overflow.
3270  if (A == Op1 && NoOp0WrapProblem)
3271  return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
3272  // icmp X, (X-Y) -> icmp Y, 0 for equalities or if there is no overflow.
3273  if (C == Op0 && NoOp1WrapProblem)
3274  return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
3275 
3276  // (A - B) >u A --> A <u B
3277  if (A == Op1 && Pred == ICmpInst::ICMP_UGT)
3278  return new ICmpInst(ICmpInst::ICMP_ULT, A, B);
3279  // C <u (C - D) --> C <u D
3280  if (C == Op0 && Pred == ICmpInst::ICMP_ULT)
3281  return new ICmpInst(ICmpInst::ICMP_ULT, C, D);
3282 
3283  // icmp (Y-X), (Z-X) -> icmp Y, Z for equalities or if there is no overflow.
3284  if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem &&
3285  // Try not to increase register pressure.
3286  BO0->hasOneUse() && BO1->hasOneUse())
3287  return new ICmpInst(Pred, A, C);
3288  // icmp (X-Y), (X-Z) -> icmp Z, Y for equalities or if there is no overflow.
3289  if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem &&
3290  // Try not to increase register pressure.
3291  BO0->hasOneUse() && BO1->hasOneUse())
3292  return new ICmpInst(Pred, D, B);
3293 
3294  // icmp (0-X) < cst --> x > -cst
3295  if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
3296  Value *X;
3297  if (match(BO0, m_Neg(m_Value(X))))
3298  if (Constant *RHSC = dyn_cast<Constant>(Op1))
3299  if (RHSC->isNotMinSignedValue())
3300  return new ICmpInst(I.getSwappedPredicate(), X,
3301  ConstantExpr::getNeg(RHSC));
3302  }
3303 
3304  BinaryOperator *SRem = nullptr;
3305  // icmp (srem X, Y), Y
3306  if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
3307  SRem = BO0;
3308  // icmp Y, (srem X, Y)
3309  else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
3310  Op0 == BO1->getOperand(1))
3311  SRem = BO1;
3312  if (SRem) {
3313  // We don't check hasOneUse to avoid increasing register pressure because
3314  // the value we use is the same value this instruction was already using.
3315  switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
3316  default:
3317  break;
3318  case ICmpInst::ICMP_EQ:
3319  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
3320  case ICmpInst::ICMP_NE:
3321  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
3322  case ICmpInst::ICMP_SGT:
3323  case ICmpInst::ICMP_SGE:
3324  return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
3326  case ICmpInst::ICMP_SLT:
3327  case ICmpInst::ICMP_SLE:
3328  return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
3329  Constant::getNullValue(SRem->getType()));
3330  }
3331  }
3332 
3333  if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
3334  BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
3335  switch (BO0->getOpcode()) {
3336  default:
3337  break;
3338  case Instruction::Add:
3339  case Instruction::Sub:
3340  case Instruction::Xor: {
3341  if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
3342  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3343 
3344  const APInt *C;
3345  if (match(BO0->getOperand(1), m_APInt(C))) {
3346  // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
3347  if (C->isSignMask()) {
3348  ICmpInst::Predicate NewPred =
3350  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3351  }
3352 
3353  // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
3354  if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
3355  ICmpInst::Predicate NewPred =
3357  NewPred = I.getSwappedPredicate(NewPred);
3358  return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
3359  }
3360  }
3361  break;
3362  }
3363  case Instruction::Mul: {
3364  if (!I.isEquality())
3365  break;
3366 
3367  const APInt *C;
3368  if (match(BO0->getOperand(1), m_APInt(C)) && !C->isNullValue() &&
3369  !C->isOneValue()) {
3370  // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
3371  // Mask = -1 >> count-trailing-zeros(C).
3372  if (unsigned TZs = C->countTrailingZeros()) {
3374  BO0->getType(),
3375  APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
3376  Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
3377  Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
3378  return new ICmpInst(Pred, And1, And2);
3379  }
3380  // If there are no trailing zeros in the multiplier, just eliminate
3381  // the multiplies (no masking is needed):
3382  // icmp eq/ne (X * C), (Y * C) --> icmp eq/ne X, Y
3383  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3384  }
3385  break;
3386  }
3387  case Instruction::UDiv:
3388  case Instruction::LShr:
3389  if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
3390  break;
3391  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3392 
3393  case Instruction::SDiv:
3394  if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
3395  break;
3396  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3397 
3398  case Instruction::AShr:
3399  if (!BO0->isExact() || !BO1->isExact())
3400  break;
3401  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3402 
3403  case Instruction::Shl: {
3404  bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
3405  bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
3406  if (!NUW && !NSW)
3407  break;
3408  if (!NSW && I.isSigned())
3409  break;
3410  return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
3411  }
3412  }
3413  }
3414 
3415  if (BO0) {
3416  // Transform A & (L - 1) `ult` L --> L != 0
3417  auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
3418  auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
3419 
3420  if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
3421  auto *Zero = Constant::getNullValue(BO0->getType());
3422  return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
3423  }
3424  }
3425 
3426  if (Value *V = foldICmpWithLowBitMaskedVal(I, Builder))
3427  return replaceInstUsesWith(I, V);
3428 
3429  if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder))
3430  return replaceInstUsesWith(I, V);
3431 
3432  return nullptr;
3433 }
3434 
3435 /// Fold icmp Pred min|max(X, Y), X.
3437  ICmpInst::Predicate Pred = Cmp.getPredicate();
3438  Value *Op0 = Cmp.getOperand(0);
3439  Value *X = Cmp.getOperand(1);
3440 
3441  // Canonicalize minimum or maximum operand to LHS of the icmp.
3442  if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
3443  match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
3444  match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
3445  match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
3446  std::swap(Op0, X);
3447  Pred = Cmp.getSwappedPredicate();
3448  }
3449 
3450  Value *Y;
3451  if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
3452  // smin(X, Y) == X --> X s<= Y
3453  // smin(X, Y) s>= X --> X s<= Y
3454  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
3455  return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
3456 
3457  // smin(X, Y) != X --> X s> Y
3458  // smin(X, Y) s< X --> X s> Y
3459  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
3460  return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
3461 
3462  // These cases should be handled in InstSimplify:
3463  // smin(X, Y) s<= X --> true
3464  // smin(X, Y) s> X --> false
3465  return nullptr;
3466  }
3467 
3468  if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
3469  // smax(X, Y) == X --> X s>= Y
3470  // smax(X, Y) s<= X --> X s>= Y
3471  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
3472  return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
3473 
3474  // smax(X, Y) != X --> X s< Y
3475  // smax(X, Y) s> X --> X s< Y
3476  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
3477  return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
3478 
3479  // These cases should be handled in InstSimplify:
3480  // smax(X, Y) s>= X --> true
3481  // smax(X, Y) s< X --> false
3482  return nullptr;
3483  }
3484 
3485  if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
3486  // umin(X, Y) == X --> X u<= Y
3487  // umin(X, Y) u>= X --> X u<= Y
3488  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
3489  return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
3490 
3491  // umin(X, Y) != X --> X u> Y
3492  // umin(X, Y) u< X --> X u> Y
3493  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
3494  return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
3495 
3496  // These cases should be handled in InstSimplify:
3497  // umin(X, Y) u<= X --> true
3498  // umin(X, Y) u> X --> false
3499  return nullptr;
3500  }
3501 
3502  if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
3503  // umax(X, Y) == X --> X u>= Y
3504  // umax(X, Y) u<= X --> X u>= Y
3505  if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
3506  return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
3507 
3508  // umax(X, Y) != X --> X u< Y
3509  // umax(X, Y) u> X --> X u< Y
3510  if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
3511  return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
3512 
3513  // These cases should be handled in InstSimplify:
3514  // umax(X, Y) u>= X --> true
3515  // umax(X, Y) u< X --> false
3516  return nullptr;
3517  }
3518 
3519  return nullptr;
3520 }
3521 
3522 Instruction *InstCombiner::foldICmpEquality(ICmpInst &I) {
3523  if (!I.isEquality())
3524  return nullptr;
3525 
3526  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3527  const CmpInst::Predicate Pred = I.getPredicate();
3528  Value *A, *B, *C, *D;
3529  if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3530  if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
3531  Value *OtherVal = A == Op1 ? B : A;
3532  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3533  }
3534 
3535  if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
3536  // A^c1 == C^c2 --> A == C^(c1^c2)
3537  ConstantInt *C1, *C2;
3538  if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
3539  Op1->hasOneUse()) {
3540  Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
3541  Value *Xor = Builder.CreateXor(C, NC);
3542  return new ICmpInst(Pred, A, Xor);
3543  }
3544 
3545  // A^B == A^D -> B == D
3546  if (A == C)
3547  return new ICmpInst(Pred, B, D);
3548  if (A == D)
3549  return new ICmpInst(Pred, B, C);
3550  if (B == C)
3551  return new ICmpInst(Pred, A, D);
3552  if (B == D)
3553  return new ICmpInst(Pred, A, C);
3554  }
3555  }
3556 
3557  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
3558  // A == (A^B) -> B == 0
3559  Value *OtherVal = A == Op0 ? B : A;
3560  return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
3561  }
3562 
3563  // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
3564  if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
3565  match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
3566  Value *X = nullptr, *Y = nullptr, *Z = nullptr;
3567 
3568  if (A == C) {
3569  X = B;
3570  Y = D;
3571  Z = A;
3572  } else if (A == D) {
3573  X = B;
3574  Y = C;
3575  Z = A;
3576  } else if (B == C) {
3577  X = A;
3578  Y = D;
3579  Z = B;
3580  } else if (B == D) {
3581  X = A;
3582  Y = C;
3583  Z = B;
3584  }
3585 
3586  if (X) { // Build (X^Y) & Z
3587  Op1 = Builder.CreateXor(X, Y);
3588  Op1 = Builder.CreateAnd(Op1, Z);
3589  I.setOperand(0, Op1);
3590  I.setOperand(1, Constant::getNullValue(Op1->getType()));
3591  return &I;
3592  }
3593  }
3594 
3595  // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
3596  // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
3597  ConstantInt *Cst1;
3598  if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
3599  match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
3600  (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
3601  match(Op1, m_ZExt(m_Value(A))))) {
3602  APInt Pow2 = Cst1->getValue() + 1;
3603  if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
3604  Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
3605  return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
3606  }
3607 
3608  // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
3609  // For lshr and ashr pairs.
3610  if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3611  match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
3612  (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
3613  match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
3614  unsigned TypeBits = Cst1->getBitWidth();
3615  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3616  if (ShAmt < TypeBits && ShAmt != 0) {
3617  ICmpInst::Predicate NewPred =
3619  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
3620  APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
3621  return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
3622  }
3623  }
3624 
3625  // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
3626  if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
3627  match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
3628  unsigned TypeBits = Cst1->getBitWidth();
3629  unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
3630  if (ShAmt < TypeBits && ShAmt != 0) {
3631  Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
3632  APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
3633  Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
3634  I.getName() + ".mask");
3635  return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
3636  }
3637  }
3638 
3639  // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
3640  // "icmp (and X, mask), cst"
3641  uint64_t ShAmt = 0;
3642  if (Op0->hasOneUse() &&
3643  match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
3644  match(Op1, m_ConstantInt(Cst1)) &&
3645  // Only do this when A has multiple uses. This is most important to do
3646  // when it exposes other optimizations.
3647  !A->hasOneUse()) {
3648  unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
3649 
3650  if (ShAmt < ASize) {
3651  APInt MaskV =
3653  MaskV <<= ShAmt;
3654 
3655  APInt CmpV = Cst1->getValue().zext(ASize);
3656  CmpV <<= ShAmt;
3657 
3658  Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
3659  return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
3660  }
3661  }
3662 
3663  // If both operands are byte-swapped or bit-reversed, just compare the
3664  // original values.
3665  // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
3666  // and handle more intrinsics.
3667  if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
3668  (match(Op0, m_BitReverse(m_Value(A))) &&
3669  match(Op1, m_BitReverse(m_Value(B)))))
3670  return new ICmpInst(Pred, A, B);
3671 
3672  return nullptr;
3673 }
3674 
3675 /// Handle icmp (cast x to y), (cast/cst). We only handle extending casts so
3676 /// far.
3677 Instruction *InstCombiner::foldICmpWithCastAndCast(ICmpInst &ICmp) {
3678  const CastInst *LHSCI = cast<CastInst>(ICmp.getOperand(0));
3679  Value *LHSCIOp = LHSCI->getOperand(0);
3680  Type *SrcTy = LHSCIOp->getType();
3681  Type *DestTy = LHSCI->getType();
3682  Value *RHSCIOp;
3683 
3684  // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
3685  // integer type is the same size as the pointer type.
3686  const auto& CompatibleSizes = [&](Type* SrcTy, Type* DestTy) -> bool {
3687  if (isa<VectorType>(SrcTy)) {
3688  SrcTy = cast<VectorType>(SrcTy)->getElementType();
3689  DestTy = cast<VectorType>(DestTy)->getElementType();
3690  }
3691  return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
3692  };
3693  if (LHSCI->getOpcode() == Instruction::PtrToInt &&
3694  CompatibleSizes(SrcTy, DestTy)) {
3695  Value *RHSOp = nullptr;
3696  if (auto *RHSC = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
3697  Value *RHSCIOp = RHSC->getOperand(0);
3698  if (RHSCIOp->getType()->getPointerAddressSpace() ==
3699  LHSCIOp->getType()->getPointerAddressSpace()) {
3700  RHSOp = RHSC->getOperand(0);
3701  // If the pointer types don't match, insert a bitcast.
3702  if (LHSCIOp->getType() != RHSOp->getType())
3703  RHSOp = Builder.CreateBitCast(RHSOp, LHSCIOp->getType());
3704  }
3705  } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
3706  RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
3707  }
3708 
3709  if (RHSOp)
3710  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSOp);
3711  }
3712 
3713  // The code below only handles extension cast instructions, so far.
3714  // Enforce this.
3715  if (LHSCI->getOpcode() != Instruction::ZExt &&
3716  LHSCI->getOpcode() != Instruction::SExt)
3717  return nullptr;
3718 
3719  bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
3720  bool isSignedCmp = ICmp.isSigned();
3721 
3722  if (auto *CI = dyn_cast<CastInst>(ICmp.getOperand(1))) {
3723  // Not an extension from the same type?
3724  RHSCIOp = CI->getOperand(0);
3725  if (RHSCIOp->getType() != LHSCIOp->getType())
3726  return nullptr;
3727 
3728  // If the signedness of the two casts doesn't agree (i.e. one is a sext
3729  // and the other is a zext), then we can't handle this.
3730  if (CI->getOpcode() != LHSCI->getOpcode())
3731  return nullptr;
3732 
3733  // Deal with equality cases early.
3734  if (ICmp.isEquality())
3735  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3736 
3737  // A signed comparison of sign extended values simplifies into a
3738  // signed comparison.
3739  if (isSignedCmp && isSignedExt)
3740  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp);
3741 
3742  // The other three cases all fold into an unsigned comparison.
3743  return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
3744  }
3745 
3746  // If we aren't dealing with a constant on the RHS, exit early.
3747  auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
3748  if (!C)
3749  return nullptr;
3750 
3751  // Compute the constant that would happen if we truncated to SrcTy then
3752  // re-extended to DestTy.
3753  Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
3754  Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
3755 
3756  // If the re-extended constant didn't change...
3757  if (Res2 == C) {
3758  // Deal with equality cases early.
3759  if (ICmp.isEquality())
3760  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3761 
3762  // A signed comparison of sign extended values simplifies into a
3763  // signed comparison.
3764  if (isSignedExt && isSignedCmp)
3765  return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1);
3766 
3767  // The other three cases all fold into an unsigned comparison.
3768  return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, Res1);
3769  }
3770 
3771  // The re-extended constant changed, partly changed (in the case of a vector),
3772  // or could not be determined to be equal (in the case of a constant
3773  // expression), so the constant cannot be represented in the shorter type.
3774  // Consequently, we cannot emit a simple comparison.
3775  // All the cases that fold to true or false will have already been handled
3776  // by SimplifyICmpInst, so only deal with the tricky case.
3777 
3778  if (isSignedCmp || !isSignedExt || !isa<ConstantInt>(C))
3779  return nullptr;
3780 
3781  // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
3782  // should have been folded away previously and not enter in here.
3783 
3784  // We're performing an unsigned comp with a sign extended value.
3785  // This is true if the input is >= 0. [aka >s -1]
3786  Constant *NegOne = Constant::getAllOnesValue(SrcTy);
3787  Value *Result = Builder.CreateICmpSGT(LHSCIOp, NegOne, ICmp.getName());
3788 
3789  // Finally, return the value computed.
3790  if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
3791  return replaceInstUsesWith(ICmp, Result);
3792 
3793  assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
3794  return BinaryOperator::CreateNot(Result);
3795 }
3796 
3797 bool InstCombiner::OptimizeOverflowCheck(OverflowCheckFlavor OCF, Value *LHS,
3798  Value *RHS, Instruction &OrigI,
3799  Value *&Result, Constant *&Overflow) {
3800  if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
3801  std::swap(LHS, RHS);
3802 
3803  auto SetResult = [&](Value *OpResult, Constant *OverflowVal, bool ReuseName) {
3804  Result = OpResult;
3805  Overflow = OverflowVal;
3806  if (ReuseName)
3807  Result->takeName(&OrigI);
3808  return true;
3809  };
3810 
3811  // If the overflow check was an add followed by a compare, the insertion point
3812  // may be pointing to the compare. We want to insert the new instructions
3813  // before the add in case there are uses of the add between the add and the
3814  // compare.
3815  Builder.SetInsertPoint(&OrigI);
3816 
3817  switch (OCF) {
3818  case OCF_INVALID:
3819  llvm_unreachable("bad overflow check kind!");
3820 
3821  case OCF_UNSIGNED_ADD: {
3822  OverflowResult OR = computeOverflowForUnsignedAdd(LHS, RHS, &OrigI);
3824  return SetResult(Builder.CreateNUWAdd(LHS, RHS), Builder.getFalse(),
3825  true);
3826 
3828  return SetResult(Builder.CreateAdd(LHS, RHS), Builder.getTrue(), true);
3829 
3830  // Fall through uadd into sadd
3832  }
3833  case OCF_SIGNED_ADD: {
3834  // X + 0 -> {X, false}
3835  if (match(RHS, m_Zero()))
3836  return SetResult(LHS, Builder.getFalse(), false);
3837 
3838  // We can strength reduce this signed add into a regular add if we can prove
3839  // that it will never overflow.
3840  if (OCF == OCF_SIGNED_ADD)
3841  if (willNotOverflowSignedAdd(LHS, RHS, OrigI))
3842  return SetResult(Builder.CreateNSWAdd(LHS, RHS), Builder.getFalse(),
3843  true);
3844  break;
3845  }
3846 
3847  case OCF_UNSIGNED_SUB:
3848  case OCF_SIGNED_SUB: {
3849  // X - 0 -> {X, false}
3850  if (match(RHS, m_Zero()))
3851  return SetResult(LHS, Builder.getFalse(), false);
3852 
3853  if (OCF == OCF_SIGNED_SUB) {
3854  if (willNotOverflowSignedSub(LHS, RHS, OrigI))
3855  return SetResult(Builder.CreateNSWSub(LHS, RHS), Builder.getFalse(),
3856  true);
3857  } else {
3858  if (willNotOverflowUnsignedSub(LHS, RHS, OrigI))
3859  return SetResult(Builder.CreateNUWSub(LHS, RHS), Builder.getFalse(),
3860  true);
3861  }
3862  break;
3863  }
3864 
3865  case OCF_UNSIGNED_MUL: {
3866  OverflowResult OR = computeOverflowForUnsignedMul(LHS, RHS, &OrigI);
3868  return SetResult(Builder.CreateNUWMul(LHS, RHS), Builder.getFalse(),
3869  true);
3871  return SetResult(Builder.CreateMul(LHS, RHS), Builder.getTrue(), true);
3873  }
3874  case OCF_SIGNED_MUL:
3875  // X * undef -> undef
3876  if (isa<UndefValue>(RHS))
3877  return SetResult(RHS, UndefValue::get(Builder.getInt1Ty()), false);
3878 
3879  // X * 0 -> {0, false}
3880  if (match(RHS, m_Zero()))
3881  return SetResult(RHS, Builder.getFalse(), false);
3882 
3883  // X * 1 -> {X, false}
3884  if (match(RHS, m_One()))
3885  return SetResult(LHS, Builder.getFalse(), false);
3886 
3887  if (OCF == OCF_SIGNED_MUL)
3888  if (willNotOverflowSignedMul(LHS, RHS, OrigI))
3889  return SetResult(Builder.CreateNSWMul(LHS, RHS), Builder.getFalse(),
3890  true);
3891  break;
3892  }
3893 
3894  return false;
3895 }
3896 
3897 /// Recognize and process idiom involving test for multiplication
3898 /// overflow.
3899 ///
3900 /// The caller has matched a pattern of the form:
3901 /// I = cmp u (mul(zext A, zext B), V
3902 /// The function checks if this is a test for overflow and if so replaces
3903 /// multiplication with call to 'mul.with.overflow' intrinsic.
3904 ///
3905 /// \param I Compare instruction.
3906 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
3907 /// the compare instruction. Must be of integer type.
3908 /// \param OtherVal The other argument of compare instruction.
3909 /// \returns Instruction which must replace the compare instruction, NULL if no
3910 /// replacement required.
3912  Value *OtherVal, InstCombiner &IC) {
3913  // Don't bother doing this transformation for pointers, don't do it for
3914  // vectors.
3915  if (!isa<IntegerType>(MulVal->getType()))
3916  return nullptr;
3917 
3918  assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
3919  assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
3920  auto *MulInstr = dyn_cast<Instruction>(MulVal);
3921  if (!MulInstr)
3922  return nullptr;
3923  assert(MulInstr->getOpcode() == Instruction::Mul);
3924 
3925  auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
3926  *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
3927  assert(LHS->getOpcode() == Instruction::ZExt);
3928  assert(RHS->getOpcode() == Instruction::ZExt);
3929  Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
3930 
3931  // Calculate type and width of the result produced by mul.with.overflow.
3932  Type *TyA = A->getType(), *TyB = B->getType();
3933  unsigned WidthA = TyA->getPrimitiveSizeInBits(),
3934  WidthB = TyB->getPrimitiveSizeInBits();
3935  unsigned MulWidth;
3936  Type *MulType;
3937  if (WidthB > WidthA) {
3938  MulWidth = WidthB;
3939  MulType = TyB;
3940  } else {
3941  MulWidth = WidthA;
3942  MulType = TyA;
3943  }
3944 
3945  // In order to replace the original mul with a narrower mul.with.overflow,
3946  // all uses must ignore upper bits of the product. The number of used low
3947  // bits must be not greater than the width of mul.with.overflow.
3948  if (MulVal->hasNUsesOrMore(2))
3949  for (User *U : MulVal->users()) {
3950  if (U == &I)
3951  continue;
3952  if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
3953  // Check if truncation ignores bits above MulWidth.
3954  unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
3955  if (TruncWidth > MulWidth)
3956  return nullptr;
3957  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
3958  // Check if AND ignores bits above MulWidth.
3959  if (BO->getOpcode() != Instruction::And)
3960  return nullptr;
3961  if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
3962  const APInt &CVal = CI->getValue();
3963  if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
3964  return nullptr;
3965  } else {
3966  // In this case we could have the operand of the binary operation
3967  // being defined in another block, and performing the replacement
3968  // could break the dominance relation.
3969  return nullptr;
3970  }
3971  } else {
3972  // Other uses prohibit this transformation.
3973  return nullptr;
3974  }
3975  }
3976 
3977  // Recognize patterns
3978  switch (I.getPredicate()) {
3979  case ICmpInst::ICMP_EQ:
3980  case ICmpInst::ICMP_NE:
3981  // Recognize pattern:
3982  // mulval = mul(zext A, zext B)
3983  // cmp eq/neq mulval, zext trunc mulval
3984  if (ZExtInst *Zext = dyn_cast<ZExtInst>(OtherVal))
3985  if (Zext->hasOneUse()) {
3986  Value *ZextArg = Zext->getOperand(0);
3987  if (TruncInst *Trunc = dyn_cast<TruncInst>(ZextArg))
3988  if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth)
3989  break; //Recognized
3990  }
3991 
3992  // Recognize pattern:
3993  // mulval = mul(zext A, zext B)
3994  // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
3995  ConstantInt *CI;
3996  Value *ValToMask;
3997  if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
3998  if (ValToMask != MulVal)
3999  return nullptr;
4000  const APInt &CVal = CI->getValue() + 1;
4001  if (CVal.isPowerOf2()) {
4002  unsigned MaskWidth = CVal.logBase2();
4003  if (MaskWidth == MulWidth)
4004  break; // Recognized
4005  }
4006  }
4007  return nullptr;
4008 
4009  case ICmpInst::ICMP_UGT:
4010  // Recognize pattern:
4011  // mulval = mul(zext A, zext B)
4012  // cmp ugt mulval, max
4013  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4014  APInt MaxVal = APInt::getMaxValue(MulWidth);
4015  MaxVal = MaxVal.zext(CI->getBitWidth());
4016  if (MaxVal.eq(CI->getValue()))
4017  break; // Recognized
4018  }
4019  return nullptr;
4020 
4021  case ICmpInst::ICMP_UGE:
4022  // Recognize pattern:
4023  // mulval = mul(zext A, zext B)
4024  // cmp uge mulval, max+1
4025  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4026  APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4027  if (MaxVal.eq(CI->getValue()))
4028  break; // Recognized
4029  }
4030  return nullptr;
4031 
4032  case ICmpInst::ICMP_ULE:
4033  // Recognize pattern:
4034  // mulval = mul(zext A, zext B)
4035  // cmp ule mulval, max
4036  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4037  APInt MaxVal = APInt::getMaxValue(MulWidth);
4038  MaxVal = MaxVal.zext(CI->getBitWidth());
4039  if (MaxVal.eq(CI->getValue()))
4040  break; // Recognized
4041  }
4042  return nullptr;
4043 
4044  case ICmpInst::ICMP_ULT:
4045  // Recognize pattern:
4046  // mulval = mul(zext A, zext B)
4047  // cmp ule mulval, max + 1
4048  if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4049  APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4050  if (MaxVal.eq(CI->getValue()))
4051  break; // Recognized
4052  }
4053  return nullptr;
4054 
4055  default:
4056  return nullptr;
4057  }
4058 
4059  InstCombiner::BuilderTy &Builder = IC.Builder;
4060  Builder.SetInsertPoint(MulInstr);
4061 
4062  // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
4063  Value *MulA = A, *MulB = B;
4064  if (WidthA < MulWidth)
4065  MulA = Builder.CreateZExt(A, MulType);
4066  if (WidthB < MulWidth)
4067  MulB = Builder.CreateZExt(B, MulType);
4069  Intrinsic::umul_with_overflow, MulType);
4070  CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
4071  IC.Worklist.Add(MulInstr);
4072 
4073  // If there are uses of mul result other than the comparison, we know that
4074  // they are truncation or binary AND. Change them to use result of
4075  // mul.with.overflow and adjust properly mask/size.
4076  if (MulVal->hasNUsesOrMore(2)) {
4077  Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
4078  for (auto UI = MulVal->user_begin(), UE = MulVal->user_end(); UI != UE;) {
4079  User *U = *UI++;
4080  if (U == &I || U == OtherVal)
4081  continue;
4082  if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4083  if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
4084  IC.replaceInstUsesWith(*TI, Mul);
4085  else
4086  TI->setOperand(0, Mul);
4087  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4088  assert(BO->getOpcode() == Instruction::And);
4089  // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
4090  ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
4091  APInt ShortMask = CI->getValue().trunc(MulWidth);
4092  Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
4093  Instruction *Zext =
4094  cast<Instruction>(Builder.CreateZExt(ShortAnd, BO->getType()));
4095  IC.Worklist.Add(Zext);
4096  IC.replaceInstUsesWith(*BO, Zext);
4097  } else {
4098  llvm_unreachable("Unexpected Binary operation");
4099  }
4100  IC.Worklist.Add(cast<Instruction>(U));
4101  }
4102  }
4103  if (isa<Instruction>(OtherVal))
4104  IC.Worklist.Add(cast<Instruction>(OtherVal));
4105 
4106  // The original icmp gets replaced with the overflow value, maybe inverted
4107  // depending on predicate.
4108  bool Inverse = false;
4109  switch (I.getPredicate()) {
4110  case ICmpInst::ICMP_NE:
4111  break;
4112  case ICmpInst::ICMP_EQ:
4113  Inverse = true;
4114  break;
4115  case ICmpInst::ICMP_UGT:
4116  case ICmpInst::ICMP_UGE:
4117  if (I.getOperand(0) == MulVal)
4118  break;
4119  Inverse = true;
4120  break;
4121  case ICmpInst::ICMP_ULT:
4122  case ICmpInst::ICMP_ULE:
4123  if (I.getOperand(1) == MulVal)
4124  break;
4125  Inverse = true;
4126  break;
4127  default:
4128  llvm_unreachable("Unexpected predicate");
4129  }
4130  if (Inverse) {
4131  Value *Res = Builder.CreateExtractValue(Call, 1);
4132  return BinaryOperator::CreateNot(Res);
4133  }
4134 
4135  return ExtractValueInst::Create(Call, 1);
4136 }
4137 
4138 /// When performing a comparison against a constant, it is possible that not all
4139 /// the bits in the LHS are demanded. This helper method computes the mask that
4140 /// IS demanded.
4141 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
4142  const APInt *RHS;
4143  if (!match(I.getOperand(1), m_APInt(RHS)))
4144  return APInt::getAllOnesValue(BitWidth);
4145 
4146  // If this is a normal comparison, it demands all bits. If it is a sign bit
4147  // comparison, it only demands the sign bit.
4148  bool UnusedBit;
4149  if (isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
4150  return APInt::getSignMask(BitWidth);
4151 
4152  switch (I.getPredicate()) {
4153  // For a UGT comparison, we don't care about any bits that
4154  // correspond to the trailing ones of the comparand. The value of these
4155  // bits doesn't impact the outcome of the comparison, because any value
4156  // greater than the RHS must differ in a bit higher than these due to carry.
4157  case ICmpInst::ICMP_UGT:
4158  return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
4159 
4160  // Similarly, for a ULT comparison, we don't care about the trailing zeros.
4161  // Any value less than the RHS must differ in a higher bit because of carries.
4162  case ICmpInst::ICMP_ULT:
4163  return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
4164 
4165  default:
4166  return APInt::getAllOnesValue(BitWidth);
4167  }
4168 }
4169 
4170 /// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst
4171 /// should be swapped.
4172 /// The decision is based on how many times these two operands are reused
4173 /// as subtract operands and their positions in those instructions.
4174 /// The rationale is that several architectures use the same instruction for
4175 /// both subtract and cmp. Thus, it is better if the order of those operands
4176 /// match.
4177 /// \return true if Op0 and Op1 should be swapped.
4178 static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) {
4179  // Filter out pointer values as those cannot appear directly in subtract.
4180  // FIXME: we may want to go through inttoptrs or bitcasts.
4181  if (Op0->getType()->isPointerTy())
4182  return false;
4183  // If a subtract already has the same operands as a compare, swapping would be
4184  // bad. If a subtract has the same operands as a compare but in reverse order,
4185  // then swapping is good.
4186  int GoodToSwap = 0;
4187  for (const User *U : Op0->users()) {
4188  if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0))))
4189  GoodToSwap++;
4190  else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1))))
4191  GoodToSwap--;
4192  }
4193  return GoodToSwap > 0;
4194 }
4195 
4196 /// Check that one use is in the same block as the definition and all
4197 /// other uses are in blocks dominated by a given block.
4198 ///
4199 /// \param DI Definition
4200 /// \param UI Use
4201 /// \param DB Block that must dominate all uses of \p DI outside
4202 /// the parent block
4203 /// \return true when \p UI is the only use of \p DI in the parent block
4204 /// and all other uses of \p DI are in blocks dominated by \p DB.
4205 ///
4207  const Instruction *UI,
4208  const BasicBlock *DB) const {
4209  assert(DI && UI && "Instruction not defined\n");
4210  // Ignore incomplete definitions.
4211  if (!DI->getParent())
4212  return false;
4213  // DI and UI must be in the same block.
4214  if (DI->getParent() != UI->getParent())
4215  return false;
4216  // Protect from self-referencing blocks.
4217  if (DI->getParent() == DB)
4218  return false;
4219  for (const User *U : DI->users()) {
4220  auto *Usr = cast<Instruction>(U);
4221  if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
4222  return false;
4223  }
4224  return true;
4225 }
4226 
4227 /// Return true when the instruction sequence within a block is select-cmp-br.
4228 static bool isChainSelectCmpBranch(const SelectInst *SI) {
4229  const BasicBlock *BB = SI->getParent();
4230  if (!BB)
4231  return false;
4232  auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
4233  if (!BI || BI->getNumSuccessors() != 2)
4234  return false;
4235  auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
4236  if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
4237  return false;
4238  return true;
4239 }
4240 
4241 /// True when a select result is replaced by one of its operands
4242 /// in select-icmp sequence. This will eventually result in the elimination
4243 /// of the select.
4244 ///
4245 /// \param SI Select instruction
4246 /// \param Icmp Compare instruction
4247 /// \param SIOpd Operand that replaces the select
4248 ///
4249 /// Notes:
4250 /// - The replacement is global and requires dominator information
4251 /// - The caller is responsible for the actual replacement
4252 ///
4253 /// Example:
4254 ///
4255 /// entry:
4256 /// %4 = select i1 %3, %C* %0, %C* null
4257 /// %5 = icmp eq %C* %4, null
4258 /// br i1 %5, label %9, label %7
4259 /// ...
4260 /// ; <label>:7 ; preds = %entry
4261 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
4262 /// ...
4263 ///
4264 /// can be transformed to
4265 ///
4266 /// %5 = icmp eq %C* %0, null
4267 /// %6 = select i1 %3, i1 %5, i1 true
4268 /// br i1 %6, label %9, label %7
4269 /// ...
4270 /// ; <label>:7 ; preds = %entry
4271 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
4272 ///
4273 /// Similar when the first operand of the select is a constant or/and
4274 /// the compare is for not equal rather than equal.
4275 ///
4276 /// NOTE: The function is only called when the select and compare constants
4277 /// are equal, the optimization can work only for EQ predicates. This is not a
4278 /// major restriction since a NE compare should be 'normalized' to an equal
4279 /// compare, which usually happens in the combiner and test case
4280 /// select-cmp-br.ll checks for it.
4282  const ICmpInst *Icmp,
4283  const unsigned SIOpd) {
4284  assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
4285  if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
4286  BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
4287  // The check for the single predecessor is not the best that can be
4288  // done. But it protects efficiently against cases like when SI's
4289  // home block has two successors, Succ and Succ1, and Succ1 predecessor
4290  // of Succ. Then SI can't be replaced by SIOpd because the use that gets
4291  // replaced can be reached on either path. So the uniqueness check
4292  // guarantees that the path all uses of SI (outside SI's parent) are on
4293  // is disjoint from all other paths out of SI. But that information
4294  // is more expensive to compute, and the trade-off here is in favor
4295  // of compile-time. It should also be noticed that we check for a single
4296  // predecessor and not only uniqueness. This to handle the situation when
4297  // Succ and Succ1 points to the same basic block.
4298  if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
4299  NumSel++;
4300  SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
4301  return true;
4302  }
4303  }
4304  return false;
4305 }
4306 
4307 /// Try to fold the comparison based on range information we can get by checking
4308 /// whether bits are known to be zero or one in the inputs.
4309 Instruction *InstCombiner::foldICmpUsingKnownBits(ICmpInst &I) {
4310  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4311  Type *Ty = Op0->getType();
4312  ICmpInst::Predicate Pred = I.getPredicate();
4313 
4314  // Get scalar or pointer size.
4315  unsigned BitWidth = Ty->isIntOrIntVectorTy()
4316  ? Ty->getScalarSizeInBits()
4317  : DL.getIndexTypeSizeInBits(Ty->getScalarType());
4318 
4319  if (!BitWidth)
4320  return nullptr;
4321 
4322  KnownBits Op0Known(BitWidth);
4323  KnownBits Op1Known(BitWidth);
4324 
4325  if (SimplifyDemandedBits(&I, 0,
4326  getDemandedBitsLHSMask(I, BitWidth),
4327  Op0Known, 0))
4328  return &I;
4329 
4330  if (SimplifyDemandedBits(&I, 1, APInt::getAllOnesValue(BitWidth),
4331  Op1Known, 0))
4332  return &I;
4333 
4334  // Given the known and unknown bits, compute a range that the LHS could be
4335  // in. Compute the Min, Max and RHS values based on the known bits. For the
4336  // EQ and NE we use unsigned values.
4337  APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
4338  APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
4339  if (I.isSigned()) {
4340  computeSignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4341  computeSignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4342  } else {
4343  computeUnsignedMinMaxValuesFromKnownBits(Op0Known, Op0Min, Op0Max);
4344  computeUnsignedMinMaxValuesFromKnownBits(Op1Known, Op1Min, Op1Max);
4345  }
4346 
4347  // If Min and Max are known to be the same, then SimplifyDemandedBits figured
4348  // out that the LHS or RHS is a constant. Constant fold this now, so that
4349  // code below can assume that Min != Max.
4350  if (!isa<Constant>(Op0) && Op0Min == Op0Max)
4351  return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1);
4352  if (!isa<Constant>(Op1) && Op1Min == Op1Max)
4353  return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min));
4354 
4355  // Based on the range information we know about the LHS, see if we can
4356  // simplify this comparison. For example, (x&4) < 8 is always true.
4357  switch (Pred) {
4358  default:
4359  llvm_unreachable("Unknown icmp opcode!");
4360  case ICmpInst::ICMP_EQ:
4361  case ICmpInst::ICMP_NE: {
4362  if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max)) {
4363  return Pred == CmpInst::ICMP_EQ
4364  ? replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()))
4365  : replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4366  }
4367 
4368  // If all bits are known zero except for one, then we know at most one bit
4369  // is set. If the comparison is against zero, then this is a check to see if
4370  // *that* bit is set.
4371  APInt Op0KnownZeroInverted = ~Op0Known.Zero;
4372  if (Op1Known.isZero()) {
4373  // If the LHS is an AND with the same constant, look through it.
4374  Value *LHS = nullptr;
4375  const APInt *LHSC;
4376  if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
4377  *LHSC != Op0KnownZeroInverted)
4378  LHS = Op0;
4379 
4380  Value *X;
4381  if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
4382  APInt ValToCheck = Op0KnownZeroInverted;
4383  Type *XTy = X->getType();
4384  if (ValToCheck.isPowerOf2()) {
4385  // ((1 << X) & 8) == 0 -> X != 3
4386  // ((1 << X) & 8) != 0 -> X == 3
4387  auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4388  auto NewPred = ICmpInst::getInversePredicate(Pred);
4389  return new ICmpInst(NewPred, X, CmpC);
4390  } else if ((++ValToCheck).isPowerOf2()) {
4391  // ((1 << X) & 7) == 0 -> X >= 3
4392  // ((1 << X) & 7) != 0 -> X < 3
4393  auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
4394  auto NewPred =
4396  return new ICmpInst(NewPred, X, CmpC);
4397  }
4398  }
4399 
4400  // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
4401  const APInt *CI;
4402  if (Op0KnownZeroInverted.isOneValue() &&
4403  match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
4404  // ((8 >>u X) & 1) == 0 -> X != 3
4405  // ((8 >>u X) & 1) != 0 -> X == 3
4406  unsigned CmpVal = CI->countTrailingZeros();
4407  auto NewPred = ICmpInst::getInversePredicate(Pred);
4408  return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
4409  }
4410  }
4411  break;
4412  }
4413  case ICmpInst::ICMP_ULT: {
4414  if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
4415  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4416  if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
4417  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4418  if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
4419  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4420 
4421  const APInt *CmpC;
4422  if (match(Op1, m_APInt(CmpC))) {
4423  // A <u C -> A == C-1 if min(A)+1 == C
4424  if (*CmpC == Op0Min + 1)
4425  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4426  ConstantInt::get(Op1->getType(), *CmpC - 1));
4427  // X <u C --> X == 0, if the number of zero bits in the bottom of X
4428  // exceeds the log2 of C.
4429  if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
4430  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4431  Constant::getNullValue(Op1->getType()));
4432  }
4433  break;
4434  }
4435  case ICmpInst::ICMP_UGT: {
4436  if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
4437  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4438  if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
4439  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4440  if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
4441  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4442 
4443  const APInt *CmpC;
4444  if (match(Op1, m_APInt(CmpC))) {
4445  // A >u C -> A == C+1 if max(a)-1 == C
4446  if (*CmpC == Op0Max - 1)
4447  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4448  ConstantInt::get(Op1->getType(), *CmpC + 1));
4449  // X >u C --> X != 0, if the number of zero bits in the bottom of X
4450  // exceeds the log2 of C.
4451  if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
4452  return new ICmpInst(ICmpInst::ICMP_NE, Op0,
4453  Constant::getNullValue(Op1->getType()));
4454  }
4455  break;
4456  }
4457  case ICmpInst::ICMP_SLT: {
4458  if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
4459  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4460  if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
4461  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4462  if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
4463  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4464  const APInt *CmpC;
4465  if (match(Op1, m_APInt(CmpC))) {
4466  if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
4467  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4468  ConstantInt::get(Op1->getType(), *CmpC - 1));
4469  }
4470  break;
4471  }
4472  case ICmpInst::ICMP_SGT: {
4473  if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
4474  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4475  if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
4476  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4477  if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
4478  return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4479  const APInt *CmpC;
4480  if (match(Op1, m_APInt(CmpC))) {
4481  if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
4482  return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
4483  ConstantInt::get(Op1->getType(), *CmpC + 1));
4484  }
4485  break;
4486  }
4487  case ICmpInst::ICMP_SGE:
4488  assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
4489  if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
4490  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4491  if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
4492  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4493  if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
4494  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4495  break;
4496  case ICmpInst::ICMP_SLE:
4497  assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
4498  if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
4499  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4500  if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
4501  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4502  if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
4503  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4504  break;
4505  case ICmpInst::ICMP_UGE:
4506  assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
4507  if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
4508  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4509  if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
4510  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4511  if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
4512  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4513  break;
4514  case ICmpInst::ICMP_ULE:
4515  assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
4516  if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
4517  return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4518  if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
4519  return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4520  if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
4521  return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4522  break;
4523  }
4524 
4525  // Turn a signed comparison into an unsigned one if both operands are known to
4526  // have the same sign.
4527  if (I.isSigned() &&
4528  ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
4529  (Op0Known.One.isNegative() && Op1Known.One.isNegative())))
4530  return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
4531 
4532  return nullptr;
4533 }
4534 
4535 /// If we have an icmp le or icmp ge instruction with a constant operand, turn
4536 /// it into the appropriate icmp lt or icmp gt instruction. This transform
4537 /// allows them to be folded in visitICmpInst.
4539  ICmpInst::Predicate Pred = I.getPredicate();
4540  if (Pred != ICmpInst::ICMP_SLE && Pred != ICmpInst::ICMP_SGE &&
4541  Pred != ICmpInst::ICMP_ULE && Pred != ICmpInst::ICMP_UGE)
4542  return nullptr;
4543 
4544  Value *Op0 = I.getOperand(0);
4545  Value *Op1 = I.getOperand(1);
4546  auto *Op1C = dyn_cast<Constant>(Op1);
4547  if (!Op1C)
4548  return nullptr;
4549 
4550  // Check if the constant operand can be safely incremented/decremented without
4551  // overflowing/underflowing. For scalars, SimplifyICmpInst has already handled
4552  // the edge cases for us, so we just assert on them. For vectors, we must
4553  // handle the edge cases.
4554  Type *Op1Type = Op1->getType();
4555  bool IsSigned = I.isSigned();
4556  bool IsLE = (Pred == ICmpInst::ICMP_SLE || Pred == ICmpInst::ICMP_ULE);
4557  auto *CI = dyn_cast<ConstantInt>(Op1C);
4558  if (CI) {
4559  // A <= MAX -> TRUE ; A >= MIN -> TRUE
4560  assert(IsLE ? !CI->isMaxValue(IsSigned) : !CI->isMinValue(IsSigned));
4561  } else if (Op1Type->isVectorTy()) {
4562  // TODO? If the edge cases for vectors were guaranteed to be handled as they
4563  // are for scalar, we could remove the min/max checks. However, to do that,
4564  // we would have to use insertelement/shufflevector to replace edge values.
4565  unsigned NumElts = Op1Type->getVectorNumElements();
4566  for (unsigned i = 0; i != NumElts; ++i) {
4567  Constant *Elt = Op1C->getAggregateElement(i);
4568  if (!Elt)
4569  return nullptr;
4570 
4571  if (isa<UndefValue>(Elt))
4572  continue;
4573 
4574  // Bail out if we can't determine if this constant is min/max or if we
4575  // know that this constant is min/max.
4576  auto *CI = dyn_cast<ConstantInt>(Elt);
4577  if (!CI || (IsLE ? CI->isMaxValue(IsSigned) : CI->isMinValue(IsSigned)))
4578  return nullptr;
4579  }
4580  } else {
4581  // ConstantExpr?
4582  return nullptr;
4583  }
4584 
4585  // Increment or decrement the constant and set the new comparison predicate:
4586  // ULE -> ULT ; UGE -> UGT ; SLE -> SLT ; SGE -> SGT
4587  Constant *OneOrNegOne = ConstantInt::get(Op1Type, IsLE ? 1 : -1, true);
4589  NewPred = IsSigned ? ICmpInst::getSignedPredicate(NewPred) : NewPred;
4590  return new ICmpInst(NewPred, Op0, ConstantExpr::getAdd(Op1C, OneOrNegOne));
4591 }
4592 
4593 /// Integer compare with boolean values can always be turned into bitwise ops.
4595  InstCombiner::BuilderTy &Builder) {
4596  Value *A = I.getOperand(0), *B = I.getOperand(1);
4597  assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only");
4598 
4599  // A boolean compared to true/false can be simplified to Op0/true/false in
4600  // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
4601  // Cases not handled by InstSimplify are always 'not' of Op0.
4602  if (match(B, m_Zero())) {
4603  switch (I.getPredicate()) {
4604  case CmpInst::ICMP_EQ: // A == 0 -> !A
4605  case CmpInst::ICMP_ULE: // A <=u 0 -> !A
4606  case CmpInst::ICMP_SGE: // A >=s 0 -> !A
4607  return BinaryOperator::CreateNot(A);
4608  default:
4609  llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4610  }
4611  } else if (match(B, m_One())) {
4612  switch (I.getPredicate()) {
4613  case CmpInst::ICMP_NE: // A != 1 -> !A
4614  case CmpInst::ICMP_ULT: // A <u 1 -> !A
4615  case CmpInst::ICMP_SGT: // A >s -1 -> !A
4616  return BinaryOperator::CreateNot(A);
4617  default:
4618  llvm_unreachable("ICmp i1 X, C not simplified as expected.");
4619  }
4620  }
4621 
4622  switch (I.getPredicate()) {
4623  default:
4624  llvm_unreachable("Invalid icmp instruction!");
4625  case ICmpInst::ICMP_EQ:
4626  // icmp eq i1 A, B -> ~(A ^ B)
4627  return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
4628 
4629  case ICmpInst::ICMP_NE:
4630  // icmp ne i1 A, B -> A ^ B
4631  return BinaryOperator::CreateXor(A, B);
4632 
4633  case ICmpInst::ICMP_UGT:
4634  // icmp ugt -> icmp ult
4635  std::swap(A, B);
4637  case ICmpInst::ICMP_ULT:
4638  // icmp ult i1 A, B -> ~A & B
4639  return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
4640 
4641  case ICmpInst::ICMP_SGT:
4642  // icmp sgt -> icmp slt
4643  std::swap(A, B);
4645  case ICmpInst::ICMP_SLT:
4646  // icmp slt i1 A, B -> A & ~B
4647  return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
4648 
4649  case ICmpInst::ICMP_UGE:
4650  // icmp uge -> icmp ule
4651  std::swap(A, B);
4653  case ICmpInst::ICMP_ULE:
4654  // icmp ule i1 A, B -> ~A | B
4655  return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
4656 
4657  case ICmpInst::ICMP_SGE:
4658  // icmp sge -> icmp sle
4659  std::swap(A, B);
4661  case ICmpInst::ICMP_SLE:
4662  // icmp sle i1 A, B -> A | ~B
4663  return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
4664  }
4665 }
4666 
4667 // Transform pattern like:
4668 // (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X
4669 // (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X
4670 // Into:
4671 // (X l>> Y) != 0
4672 // (X l>> Y) == 0
4674  InstCombiner::BuilderTy &Builder) {
4675  ICmpInst::Predicate Pred, NewPred;
4676  Value *X, *Y;
4677  if (match(&Cmp,
4678  m_c_ICmp(Pred, m_OneUse(m_Shl(m_One(), m_Value(Y))), m_Value(X)))) {
4679  // We want X to be the icmp's second operand, so swap predicate if it isn't.
4680  if (Cmp.getOperand(0) == X)
4681  Pred = Cmp.getSwappedPredicate();
4682 
4683  switch (Pred) {
4684  case ICmpInst::ICMP_ULE:
4685  NewPred = ICmpInst::ICMP_NE;
4686  break;
4687  case ICmpInst::ICMP_UGT:
4688  NewPred = ICmpInst::ICMP_EQ;
4689  break;
4690  default:
4691  return nullptr;
4692  }
4693  } else if (match(&Cmp, m_c_ICmp(Pred,
4695  m_Not(m_Shl(m_AllOnes(), m_Value(Y))),
4696  m_Add(m_Shl(m_One(), m_Value(Y)),
4697  m_AllOnes()))),
4698  m_Value(X)))) {
4699  // The variant with 'add' is not canonical, (the variant with 'not' is)
4700  // we only get it because it has extra uses, and can't be canonicalized,
4701 
4702  // We want X to be the icmp's second operand, so swap predicate if it isn't.
4703  if (Cmp.getOperand(0) == X)
4704  Pred = Cmp.getSwappedPredicate();
4705 
4706  switch (Pred) {
4707  case ICmpInst::ICMP_ULT:
4708  NewPred = ICmpInst::ICMP_NE;
4709  break;
4710  case ICmpInst::ICMP_UGE:
4711  NewPred = ICmpInst::ICMP_EQ;
4712  break;
4713  default:
4714  return nullptr;
4715  }
4716  } else
4717  return nullptr;
4718 
4719  Value *NewX = Builder.CreateLShr(X, Y, X->getName() + ".highbits");
4720  Constant *Zero = Constant::getNullValue(NewX->getType());
4721  return CmpInst::Create(Instruction::ICmp, NewPred, NewX, Zero);
4722 }
4723 
4725  InstCombiner::BuilderTy &Builder) {
4726  // If both arguments of the cmp are shuffles that use the same mask and
4727  // shuffle within a single vector, move the shuffle after the cmp.
4728  Value *LHS = Cmp.getOperand(0), *RHS = Cmp.getOperand(1);
4729  Value *V1, *V2;
4730  Constant *M;
4731  if (match(LHS, m_ShuffleVector(m_Value(V1), m_Undef(), m_Constant(M))) &&
4732  match(RHS, m_ShuffleVector(m_Value(V2), m_Undef(), m_Specific(M))) &&
4733  V1->getType() == V2->getType() &&
4734  (LHS->hasOneUse() || RHS->hasOneUse())) {
4735  // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M
4737  Value *NewCmp = isa<ICmpInst>(Cmp) ? Builder.CreateICmp(P, V1, V2)
4738  : Builder.CreateFCmp(P, V1, V2);
4739  return new ShuffleVectorInst(NewCmp, UndefValue::get(NewCmp->getType()), M);
4740  }
4741  return nullptr;
4742 }
4743 
4745  bool Changed = false;
4746  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4747  unsigned Op0Cplxity = getComplexity(Op0);
4748  unsigned Op1Cplxity = getComplexity(Op1);
4749 
4750  /// Orders the operands of the compare so that they are listed from most
4751  /// complex to least complex. This puts constants before unary operators,
4752  /// before binary operators.
4753  if (Op0Cplxity < Op1Cplxity ||
4754  (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) {
4755  I.swapOperands();
4756  std::swap(Op0, Op1);
4757  Changed = true;
4758  }
4759 
4760  if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1,
4761  SQ.getWithInstruction(&I)))
4762  return replaceInstUsesWith(I, V);
4763 
4764  // Comparing -val or val with non-zero is the same as just comparing val
4765  // ie, abs(val) != 0 -> val != 0
4766  if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
4767  Value *Cond, *SelectTrue, *SelectFalse;
4768  if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
4769  m_Value(SelectFalse)))) {
4770  if (Value *V = dyn_castNegVal(SelectTrue)) {
4771  if (V == SelectFalse)
4772  return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4773  }
4774  else if (Value *V = dyn_castNegVal(SelectFalse)) {
4775  if (V == SelectTrue)
4776  return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
4777  }
4778  }
4779  }
4780 
4781  if (Op0->getType()->isIntOrIntVectorTy(1))
4782  if (Instruction *Res = canonicalizeICmpBool(I, Builder))
4783  return Res;
4784 
4785  if (ICmpInst *NewICmp = canonicalizeCmpWithConstant(I))
4786  return NewICmp;
4787 
4788  if (Instruction *Res = foldICmpWithConstant(I))
4789  return Res;
4790 
4791  if (Instruction *Res = foldICmpWithDominatingICmp(I))
4792  return Res;
4793 
4794  if (Instruction *Res = foldICmpUsingKnownBits(I))
4795  return Res;
4796 
4797  // Test if the ICmpInst instruction is used exclusively by a select as
4798  // part of a minimum or maximum operation. If so, refrain from doing
4799  // any other folding. This helps out other analyses which understand
4800  // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
4801  // and CodeGen. And in this case, at least one of the comparison
4802  // operands has at least one user besides the compare (the select),
4803  // which would often largely negate the benefit of folding anyway.
4804  //
4805  // Do the same for the other patterns recognized by matchSelectPattern.
4806  if (I.hasOneUse())
4807  if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
4808  Value *A, *B;
4809  SelectPatternResult SPR = matchSelectPattern(SI, A, B);
4810  if (SPR.Flavor != SPF_UNKNOWN)
4811  return nullptr;
4812  }
4813 
4814  // Do this after checking for min/max to prevent infinite looping.
4815  if (Instruction *Res = foldICmpWithZero(I))
4816  return Res;
4817 
4818  // FIXME: We only do this after checking for min/max to prevent infinite
4819  // looping caused by a reverse canonicalization of these patterns for min/max.
4820  // FIXME: The organization of folds is a mess. These would naturally go into
4821  // canonicalizeCmpWithConstant(), but we can't move all of the above folds
4822  // down here after the min/max restriction.
4823  ICmpInst::Predicate Pred = I.getPredicate();
4824  const APInt *C;
4825  if (match(Op1, m_APInt(C))) {
4826  // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set
4827  if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) {
4828  Constant *Zero = Constant::getNullValue(Op0->getType());
4829  return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero);
4830  }
4831 
4832  // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear
4833  if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) {
4834  Constant *AllOnes = Constant::getAllOnesValue(Op0->getType());
4835  return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
4836  }
4837  }
4838 
4839  if (Instruction *Res = foldICmpInstWithConstant(I))
4840  return Res;
4841 
4842  if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
4843  return Res;
4844 
4845  // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
4846  if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
4847  if (Instruction *NI = foldGEPICmp(GEP, Op1, I.getPredicate(), I))
4848  return NI;
4849  if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
4850  if (Instruction *NI = foldGEPICmp(GEP, Op0,
4852  return NI;
4853 
4854  // Try to optimize equality comparisons against alloca-based pointers.
4855  if (Op0->getType()->isPointerTy() && I.isEquality()) {
4856  assert(Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?");
4857  if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op0, DL)))
4858  if (Instruction *New = foldAllocaCmp(I, Alloca, Op1))
4859  return New;
4860  if (auto *Alloca = dyn_cast<AllocaInst>(GetUnderlyingObject(Op1, DL)))
4861  if (Instruction *New = foldAllocaCmp(I, Alloca, Op0))
4862  return New;
4863  }
4864 
4865  // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
4866  Value *X;
4867  if (match(Op0, m_BitCast(m_SIToFP(m_Value(X))))) {
4868  // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
4869  // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
4870  // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
4871  // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
4872  if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
4873  Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
4874  match(Op1, m_Zero()))
4875  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
4876 
4877  // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
4878  if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
4879  return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
4880 
4881  // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
4882  if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
4883  return new ICmpInst(Pred, X, ConstantInt::getAllOnesValue(X->getType()));
4884  }
4885 
4886  // Zero-equality checks are preserved through unsigned floating-point casts:
4887  // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
4888  // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
4889  if (match(Op0, m_BitCast(m_UIToFP(m_Value(X)))))
4890  if (I.isEquality() && match(Op1, m_Zero()))
4891  return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
4892 
4893  // Test to see if the operands of the icmp are casted versions of other
4894  // values. If the ptr->ptr cast can be stripped off both arguments, we do so
4895  // now.
4896  if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
4897  if (Op0->getType()->isPointerTy() &&
4898  (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
4899  // We keep moving the cast from the left operand over to the right
4900  // operand, where it can often be eliminated completely.
4901  Op0 = CI->getOperand(0);
4902 
4903  // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
4904  // so eliminate it as well.
4905  if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
4906  Op1 = CI2->getOperand(0);
4907 
4908  // If Op1 is a constant, we can fold the cast into the constant.
4909  if (Op0->getType() != Op1->getType()) {
4910  if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
4911  Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
4912  } else {
4913  // Otherwise, cast the RHS right before the icmp
4914  Op1 = Builder.CreateBitCast(Op1, Op0->getType());
4915  }
4916  }
4917  return new ICmpInst(I.getPredicate(), Op0, Op1);
4918  }
4919  }
4920 
4921  if (isa<CastInst>(Op0)) {
4922  // Handle the special case of: icmp (cast bool to X), <cst>
4923  // This comes up when you have code like
4924  // int X = A < B;
4925  // if (X) ...
4926  // For generality, we handle any zero-extension of any operand comparison
4927  // with a constant or another cast from the same type.
4928  if (isa<Constant>(Op1) || isa<CastInst>(Op1))
4929  if (Instruction *R = foldICmpWithCastAndCast(I))
4930  return R;
4931  }
4932 
4933  if (Instruction *Res = foldICmpBinOp(I))
4934  return Res;
4935 
4936  if (Instruction *Res = foldICmpWithMinMax(I))
4937  return Res;
4938 
4939  {
4940  Value *A, *B;
4941  // Transform (A & ~B) == 0 --> (A & B) != 0
4942  // and (A & ~B) != 0 --> (A & B) == 0
4943  // if A is a power of 2.
4944  if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
4945  match(Op1, m_Zero()) &&
4946  isKnownToBeAPowerOfTwo(A, false, 0, &I) && I.isEquality())
4947  return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(A, B),
4948  Op1);
4949 
4950  // ~X < ~Y --> Y < X
4951  // ~X < C --> X > ~C
4952  if (match(Op0, m_Not(m_Value(A)))) {
4953  if (match(Op1, m_Not(m_Value(B))))
4954  return new ICmpInst(I.getPredicate(), B, A);
4955 
4956  const APInt *C;
4957  if (match(Op1, m_APInt(C)))
4958  return new ICmpInst(I.getSwappedPredicate(), A,
4959  ConstantInt::get(Op1->getType(), ~(*C)));
4960  }
4961 
4962  Instruction *AddI = nullptr;
4963  if (match(&I, m_UAddWithOverflow(m_Value(A), m_Value(B),
4964  m_Instruction(AddI))) &&
4965  isa<IntegerType>(A->getType())) {
4966  Value *Result;
4967  Constant *Overflow;
4968  if (OptimizeOverflowCheck(OCF_UNSIGNED_ADD, A, B, *AddI, Result,
4969  Overflow)) {
4970  replaceInstUsesWith(*AddI, Result);
4971  return replaceInstUsesWith(I, Overflow);
4972  }
4973  }
4974 
4975  // (zext a) * (zext b) --> llvm.umul.with.overflow.
4976  if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
4977  if (Instruction *R = processUMulZExtIdiom(I, Op0, Op1, *this))
4978  return R;
4979  }
4980  if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
4981  if (Instruction *R = processUMulZExtIdiom(I, Op1, Op0, *this))
4982  return R;
4983  }
4984  }
4985 
4986  if (Instruction *Res = foldICmpEquality(I))
4987  return Res;
4988 
4989  // The 'cmpxchg' instruction returns an aggregate containing the old value and
4990  // an i1 which indicates whether or not we successfully did the swap.
4991  //
4992  // Replace comparisons between the old value and the expected value with the
4993  // indicator that 'cmpxchg' returns.
4994  //
4995  // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
4996  // spuriously fail. In those cases, the old value may equal the expected
4997  // value but it is possible for the swap to not occur.
4998  if (I.getPredicate() == ICmpInst::ICMP_EQ)
4999  if (auto *EVI = dyn_cast<ExtractValueInst>(Op0))
5000  if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand()))
5001  if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
5002  !ACXI->isWeak())
5003  return ExtractValueInst::Create(ACXI, 1);
5004 
5005  {
5006  Value *X;
5007  const APInt *C;
5008  // icmp X+Cst, X
5009  if (match(Op0, m_Add(m_Value(X), m_APInt(C))) && Op1 == X)
5010  return foldICmpAddOpConst(X, *C, I.getPredicate());
5011 
5012  // icmp X, X+Cst
5013  if (match(Op1, m_Add(m_Value(X), m_APInt(C))) && Op0 == X)
5014  return foldICmpAddOpConst(X, *C, I.getSwappedPredicate());
5015  }
5016 
5017  if (Instruction *Res = foldICmpWithHighBitMask(I, Builder))
5018  return Res;
5019 
5020  if (I.getType()->isVectorTy())
5021  if (Instruction *Res = foldVectorCmp(I, Builder))
5022  return Res;
5023 
5024  return Changed ? &I : nullptr;
5025 }
5026 
5027 /// Fold fcmp ([us]itofp x, cst) if possible.
5028 Instruction *InstCombiner::foldFCmpIntToFPConst(FCmpInst &I, Instruction *LHSI,
5029  Constant *RHSC) {
5030  if (!isa<ConstantFP>(RHSC)) return nullptr;
5031  const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5032 
5033  // Get the width of the mantissa. We don't want to hack on conversions that
5034  // might lose information from the integer, e.g. "i64 -> float"
5035  int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5036  if (MantissaWidth == -1) return nullptr; // Unknown.
5037 
5038  IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5039 
5040  bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5041 
5042  if (I.isEquality()) {
5044  bool IsExact = false;
5045