1 //===- InstCombineCompares.cpp --------------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the visitICmp and visitFCmp functions.
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "InstCombineInternal.h"
14 #include "llvm/ADT/APSInt.h"
15 #include "llvm/ADT/SetVector.h"
16 #include "llvm/ADT/Statistic.h"
17 #include "llvm/Analysis/ConstantFolding.h"
18 #include "llvm/Analysis/InstructionSimplify.h"
19 #include "llvm/Analysis/TargetLibraryInfo.h"
20 #include "llvm/IR/ConstantRange.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/GetElementPtrTypeIterator.h"
23 #include "llvm/IR/IntrinsicInst.h"
24 #include "llvm/IR/PatternMatch.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/KnownBits.h"
27 #include "llvm/Transforms/InstCombine/InstCombiner.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.
addWithOverflow(APInt & Result,const APInt & In1,const APInt & In2,bool IsSigned=false)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.
subWithOverflow(APInt & Result,const APInt & In1,const APInt & In2,bool IsSigned=false)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.
hasBranchUse(ICmpInst & I)66 static bool hasBranchUse(ICmpInst &I) {
67 for (auto *U : I.users())
68 if (isa<BranchInst>(U))
69 return true;
70 return false;
71 }
72
73 /// Returns true if the exploded icmp can be expressed as a signed comparison
74 /// to zero and updates the predicate accordingly.
75 /// The signedness of the comparison is preserved.
76 /// TODO: Refactor with decomposeBitTestICmp()?
isSignTest(ICmpInst::Predicate & Pred,const APInt & C)77 static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) {
78 if (!ICmpInst::isSigned(Pred))
79 return false;
80
81 if (C.isZero())
82 return ICmpInst::isRelational(Pred);
83
84 if (C.isOne()) {
85 if (Pred == ICmpInst::ICMP_SLT) {
86 Pred = ICmpInst::ICMP_SLE;
87 return true;
88 }
89 } else if (C.isAllOnes()) {
90 if (Pred == ICmpInst::ICMP_SGT) {
91 Pred = ICmpInst::ICMP_SGE;
92 return true;
93 }
94 }
95
96 return false;
97 }
98
99 /// This is called when we see this pattern:
100 /// cmp pred (load (gep GV, ...)), cmpcst
101 /// where GV is a global variable with a constant initializer. Try to simplify
102 /// this into some simple computation that does not need the load. For example
103 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
104 ///
105 /// If AndCst is non-null, then the loaded value is masked with that constant
106 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
107 Instruction *
foldCmpLoadFromIndexedGlobal(GetElementPtrInst * GEP,GlobalVariable * GV,CmpInst & ICI,ConstantInt * AndCst)108 InstCombinerImpl::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP,
109 GlobalVariable *GV, CmpInst &ICI,
110 ConstantInt *AndCst) {
111 Constant *Init = GV->getInitializer();
112 if (!isa<ConstantArray>(Init) && !isa<ConstantDataArray>(Init))
113 return nullptr;
114
115 uint64_t ArrayElementCount = Init->getType()->getArrayNumElements();
116 // Don't blow up on huge arrays.
117 if (ArrayElementCount > MaxArraySizeForCombine)
118 return nullptr;
119
120 // There are many forms of this optimization we can handle, for now, just do
121 // the simple index into a single-dimensional array.
122 //
123 // Require: GEP GV, 0, i {{, constant indices}}
124 if (GEP->getNumOperands() < 3 ||
125 !isa<ConstantInt>(GEP->getOperand(1)) ||
126 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
127 isa<Constant>(GEP->getOperand(2)))
128 return nullptr;
129
130 // Check that indices after the variable are constants and in-range for the
131 // type they index. Collect the indices. This is typically for arrays of
132 // structs.
133 SmallVector<unsigned, 4> LaterIndices;
134
135 Type *EltTy = Init->getType()->getArrayElementType();
136 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
137 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
138 if (!Idx) return nullptr; // Variable index.
139
140 uint64_t IdxVal = Idx->getZExtValue();
141 if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index.
142
143 if (StructType *STy = dyn_cast<StructType>(EltTy))
144 EltTy = STy->getElementType(IdxVal);
145 else if (ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
146 if (IdxVal >= ATy->getNumElements()) return nullptr;
147 EltTy = ATy->getElementType();
148 } else {
149 return nullptr; // Unknown type.
150 }
151
152 LaterIndices.push_back(IdxVal);
153 }
154
155 enum { Overdefined = -3, Undefined = -2 };
156
157 // Variables for our state machines.
158
159 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
160 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
161 // and 87 is the second (and last) index. FirstTrueElement is -2 when
162 // undefined, otherwise set to the first true element. SecondTrueElement is
163 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
164 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
165
166 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
167 // form "i != 47 & i != 87". Same state transitions as for true elements.
168 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
169
170 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
171 /// define a state machine that triggers for ranges of values that the index
172 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
173 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
174 /// index in the range (inclusive). We use -2 for undefined here because we
175 /// use relative comparisons and don't want 0-1 to match -1.
176 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
177
178 // MagicBitvector - This is a magic bitvector where we set a bit if the
179 // comparison is true for element 'i'. If there are 64 elements or less in
180 // the array, this will fully represent all the comparison results.
181 uint64_t MagicBitvector = 0;
182
183 // Scan the array and see if one of our patterns matches.
184 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
185 for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) {
186 Constant *Elt = Init->getAggregateElement(i);
187 if (!Elt) return nullptr;
188
189 // If this is indexing an array of structures, get the structure element.
190 if (!LaterIndices.empty())
191 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices);
192
193 // If the element is masked, handle it.
194 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
195
196 // Find out if the comparison would be true or false for the i'th element.
197 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
198 CompareRHS, DL, &TLI);
199 // If the result is undef for this element, ignore it.
200 if (isa<UndefValue>(C)) {
201 // Extend range state machines to cover this element in case there is an
202 // undef in the middle of the range.
203 if (TrueRangeEnd == (int)i-1)
204 TrueRangeEnd = i;
205 if (FalseRangeEnd == (int)i-1)
206 FalseRangeEnd = i;
207 continue;
208 }
209
210 // If we can't compute the result for any of the elements, we have to give
211 // up evaluating the entire conditional.
212 if (!isa<ConstantInt>(C)) return nullptr;
213
214 // Otherwise, we know if the comparison is true or false for this element,
215 // update our state machines.
216 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
217
218 // State machine for single/double/range index comparison.
219 if (IsTrueForElt) {
220 // Update the TrueElement state machine.
221 if (FirstTrueElement == Undefined)
222 FirstTrueElement = TrueRangeEnd = i; // First true element.
223 else {
224 // Update double-compare state machine.
225 if (SecondTrueElement == Undefined)
226 SecondTrueElement = i;
227 else
228 SecondTrueElement = Overdefined;
229
230 // Update range state machine.
231 if (TrueRangeEnd == (int)i-1)
232 TrueRangeEnd = i;
233 else
234 TrueRangeEnd = Overdefined;
235 }
236 } else {
237 // Update the FalseElement state machine.
238 if (FirstFalseElement == Undefined)
239 FirstFalseElement = FalseRangeEnd = i; // First false element.
240 else {
241 // Update double-compare state machine.
242 if (SecondFalseElement == Undefined)
243 SecondFalseElement = i;
244 else
245 SecondFalseElement = Overdefined;
246
247 // Update range state machine.
248 if (FalseRangeEnd == (int)i-1)
249 FalseRangeEnd = i;
250 else
251 FalseRangeEnd = Overdefined;
252 }
253 }
254
255 // If this element is in range, update our magic bitvector.
256 if (i < 64 && IsTrueForElt)
257 MagicBitvector |= 1ULL << i;
258
259 // If all of our states become overdefined, bail out early. Since the
260 // predicate is expensive, only check it every 8 elements. This is only
261 // really useful for really huge arrays.
262 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
263 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
264 FalseRangeEnd == Overdefined)
265 return nullptr;
266 }
267
268 // Now that we've scanned the entire array, emit our new comparison(s). We
269 // order the state machines in complexity of the generated code.
270 Value *Idx = GEP->getOperand(2);
271
272 // If the index is larger than the pointer size of the target, truncate the
273 // index down like the GEP would do implicitly. We don't have to do this for
274 // an inbounds GEP because the index can't be out of range.
275 if (!GEP->isInBounds()) {
276 Type *IntPtrTy = DL.getIntPtrType(GEP->getType());
277 unsigned PtrSize = IntPtrTy->getIntegerBitWidth();
278 if (Idx->getType()->getPrimitiveSizeInBits().getFixedSize() > PtrSize)
279 Idx = Builder.CreateTrunc(Idx, IntPtrTy);
280 }
281
282 // If inbounds keyword is not present, Idx * ElementSize can overflow.
283 // Let's assume that ElementSize is 2 and the wanted value is at offset 0.
284 // Then, there are two possible values for Idx to match offset 0:
285 // 0x00..00, 0x80..00.
286 // Emitting 'icmp eq Idx, 0' isn't correct in this case because the
287 // comparison is false if Idx was 0x80..00.
288 // We need to erase the highest countTrailingZeros(ElementSize) bits of Idx.
289 unsigned ElementSize =
290 DL.getTypeAllocSize(Init->getType()->getArrayElementType());
291 auto MaskIdx = [&](Value* Idx){
292 if (!GEP->isInBounds() && countTrailingZeros(ElementSize) != 0) {
293 Value *Mask = ConstantInt::get(Idx->getType(), -1);
294 Mask = Builder.CreateLShr(Mask, countTrailingZeros(ElementSize));
295 Idx = Builder.CreateAnd(Idx, Mask);
296 }
297 return Idx;
298 };
299
300 // If the comparison is only true for one or two elements, emit direct
301 // comparisons.
302 if (SecondTrueElement != Overdefined) {
303 Idx = MaskIdx(Idx);
304 // None true -> false.
305 if (FirstTrueElement == Undefined)
306 return replaceInstUsesWith(ICI, Builder.getFalse());
307
308 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
309
310 // True for one element -> 'i == 47'.
311 if (SecondTrueElement == Undefined)
312 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
313
314 // True for two elements -> 'i == 47 | i == 72'.
315 Value *C1 = Builder.CreateICmpEQ(Idx, FirstTrueIdx);
316 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
317 Value *C2 = Builder.CreateICmpEQ(Idx, SecondTrueIdx);
318 return BinaryOperator::CreateOr(C1, C2);
319 }
320
321 // If the comparison is only false for one or two elements, emit direct
322 // comparisons.
323 if (SecondFalseElement != Overdefined) {
324 Idx = MaskIdx(Idx);
325 // None false -> true.
326 if (FirstFalseElement == Undefined)
327 return replaceInstUsesWith(ICI, Builder.getTrue());
328
329 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
330
331 // False for one element -> 'i != 47'.
332 if (SecondFalseElement == Undefined)
333 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
334
335 // False for two elements -> 'i != 47 & i != 72'.
336 Value *C1 = Builder.CreateICmpNE(Idx, FirstFalseIdx);
337 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
338 Value *C2 = Builder.CreateICmpNE(Idx, SecondFalseIdx);
339 return BinaryOperator::CreateAnd(C1, C2);
340 }
341
342 // If the comparison can be replaced with a range comparison for the elements
343 // where it is true, emit the range check.
344 if (TrueRangeEnd != Overdefined) {
345 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
346 Idx = MaskIdx(Idx);
347
348 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
349 if (FirstTrueElement) {
350 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
351 Idx = Builder.CreateAdd(Idx, Offs);
352 }
353
354 Value *End = ConstantInt::get(Idx->getType(),
355 TrueRangeEnd-FirstTrueElement+1);
356 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
357 }
358
359 // False range check.
360 if (FalseRangeEnd != Overdefined) {
361 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
362 Idx = MaskIdx(Idx);
363 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
364 if (FirstFalseElement) {
365 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
366 Idx = Builder.CreateAdd(Idx, Offs);
367 }
368
369 Value *End = ConstantInt::get(Idx->getType(),
370 FalseRangeEnd-FirstFalseElement);
371 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
372 }
373
374 // If a magic bitvector captures the entire comparison state
375 // of this load, replace it with computation that does:
376 // ((magic_cst >> i) & 1) != 0
377 {
378 Type *Ty = nullptr;
379
380 // Look for an appropriate type:
381 // - The type of Idx if the magic fits
382 // - The smallest fitting legal type
383 if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth())
384 Ty = Idx->getType();
385 else
386 Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount);
387
388 if (Ty) {
389 Idx = MaskIdx(Idx);
390 Value *V = Builder.CreateIntCast(Idx, Ty, false);
391 V = Builder.CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
392 V = Builder.CreateAnd(ConstantInt::get(Ty, 1), V);
393 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
394 }
395 }
396
397 return nullptr;
398 }
399
400 /// Return a value that can be used to compare the *offset* implied by a GEP to
401 /// zero. For example, if we have &A[i], we want to return 'i' for
402 /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales
403 /// are involved. The above expression would also be legal to codegen as
404 /// "icmp ne (i*4), 0" (assuming A is a pointer to i32).
405 /// This latter form is less amenable to optimization though, and we are allowed
406 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
407 ///
408 /// If we can't emit an optimized form for this expression, this returns null.
409 ///
evaluateGEPOffsetExpression(User * GEP,InstCombinerImpl & IC,const DataLayout & DL)410 static Value *evaluateGEPOffsetExpression(User *GEP, InstCombinerImpl &IC,
411 const DataLayout &DL) {
412 gep_type_iterator GTI = gep_type_begin(GEP);
413
414 // Check to see if this gep only has a single variable index. If so, and if
415 // any constant indices are a multiple of its scale, then we can compute this
416 // in terms of the scale of the variable index. For example, if the GEP
417 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
418 // because the expression will cross zero at the same point.
419 unsigned i, e = GEP->getNumOperands();
420 int64_t Offset = 0;
421 for (i = 1; i != e; ++i, ++GTI) {
422 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
423 // Compute the aggregate offset of constant indices.
424 if (CI->isZero()) continue;
425
426 // Handle a struct index, which adds its field offset to the pointer.
427 if (StructType *STy = GTI.getStructTypeOrNull()) {
428 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
429 } else {
430 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
431 Offset += Size*CI->getSExtValue();
432 }
433 } else {
434 // Found our variable index.
435 break;
436 }
437 }
438
439 // If there are no variable indices, we must have a constant offset, just
440 // evaluate it the general way.
441 if (i == e) return nullptr;
442
443 Value *VariableIdx = GEP->getOperand(i);
444 // Determine the scale factor of the variable element. For example, this is
445 // 4 if the variable index is into an array of i32.
446 uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType());
447
448 // Verify that there are no other variable indices. If so, emit the hard way.
449 for (++i, ++GTI; i != e; ++i, ++GTI) {
450 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
451 if (!CI) return nullptr;
452
453 // Compute the aggregate offset of constant indices.
454 if (CI->isZero()) continue;
455
456 // Handle a struct index, which adds its field offset to the pointer.
457 if (StructType *STy = GTI.getStructTypeOrNull()) {
458 Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
459 } else {
460 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
461 Offset += Size*CI->getSExtValue();
462 }
463 }
464
465 // Okay, we know we have a single variable index, which must be a
466 // pointer/array/vector index. If there is no offset, life is simple, return
467 // the index.
468 Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType());
469 unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth();
470 if (Offset == 0) {
471 // Cast to intptrty in case a truncation occurs. If an extension is needed,
472 // we don't need to bother extending: the extension won't affect where the
473 // computation crosses zero.
474 if (VariableIdx->getType()->getPrimitiveSizeInBits().getFixedSize() >
475 IntPtrWidth) {
476 VariableIdx = IC.Builder.CreateTrunc(VariableIdx, IntPtrTy);
477 }
478 return VariableIdx;
479 }
480
481 // Otherwise, there is an index. The computation we will do will be modulo
482 // the pointer size.
483 Offset = SignExtend64(Offset, IntPtrWidth);
484 VariableScale = SignExtend64(VariableScale, IntPtrWidth);
485
486 // To do this transformation, any constant index must be a multiple of the
487 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
488 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
489 // multiple of the variable scale.
490 int64_t NewOffs = Offset / (int64_t)VariableScale;
491 if (Offset != NewOffs*(int64_t)VariableScale)
492 return nullptr;
493
494 // Okay, we can do this evaluation. Start by converting the index to intptr.
495 if (VariableIdx->getType() != IntPtrTy)
496 VariableIdx = IC.Builder.CreateIntCast(VariableIdx, IntPtrTy,
497 true /*Signed*/);
498 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
499 return IC.Builder.CreateAdd(VariableIdx, OffsetVal, "offset");
500 }
501
502 /// Returns true if we can rewrite Start as a GEP with pointer Base
503 /// and some integer offset. The nodes that need to be re-written
504 /// for this transformation will be added to Explored.
canRewriteGEPAsOffset(Value * Start,Value * Base,const DataLayout & DL,SetVector<Value * > & Explored)505 static bool canRewriteGEPAsOffset(Value *Start, Value *Base,
506 const DataLayout &DL,
507 SetVector<Value *> &Explored) {
508 SmallVector<Value *, 16> WorkList(1, Start);
509 Explored.insert(Base);
510
511 // The following traversal gives us an order which can be used
512 // when doing the final transformation. Since in the final
513 // transformation we create the PHI replacement instructions first,
514 // we don't have to get them in any particular order.
515 //
516 // However, for other instructions we will have to traverse the
517 // operands of an instruction first, which means that we have to
518 // do a post-order traversal.
519 while (!WorkList.empty()) {
520 SetVector<PHINode *> PHIs;
521
522 while (!WorkList.empty()) {
523 if (Explored.size() >= 100)
524 return false;
525
526 Value *V = WorkList.back();
527
528 if (Explored.contains(V)) {
529 WorkList.pop_back();
530 continue;
531 }
532
533 if (!isa<IntToPtrInst>(V) && !isa<PtrToIntInst>(V) &&
534 !isa<GetElementPtrInst>(V) && !isa<PHINode>(V))
535 // We've found some value that we can't explore which is different from
536 // the base. Therefore we can't do this transformation.
537 return false;
538
539 if (isa<IntToPtrInst>(V) || isa<PtrToIntInst>(V)) {
540 auto *CI = cast<CastInst>(V);
541 if (!CI->isNoopCast(DL))
542 return false;
543
544 if (Explored.count(CI->getOperand(0)) == 0)
545 WorkList.push_back(CI->getOperand(0));
546 }
547
548 if (auto *GEP = dyn_cast<GEPOperator>(V)) {
549 // We're limiting the GEP to having one index. This will preserve
550 // the original pointer type. We could handle more cases in the
551 // future.
552 if (GEP->getNumIndices() != 1 || !GEP->isInBounds() ||
553 GEP->getType() != Start->getType())
554 return false;
555
556 if (Explored.count(GEP->getOperand(0)) == 0)
557 WorkList.push_back(GEP->getOperand(0));
558 }
559
560 if (WorkList.back() == V) {
561 WorkList.pop_back();
562 // We've finished visiting this node, mark it as such.
563 Explored.insert(V);
564 }
565
566 if (auto *PN = dyn_cast<PHINode>(V)) {
567 // We cannot transform PHIs on unsplittable basic blocks.
568 if (isa<CatchSwitchInst>(PN->getParent()->getTerminator()))
569 return false;
570 Explored.insert(PN);
571 PHIs.insert(PN);
572 }
573 }
574
575 // Explore the PHI nodes further.
576 for (auto *PN : PHIs)
577 for (Value *Op : PN->incoming_values())
578 if (Explored.count(Op) == 0)
579 WorkList.push_back(Op);
580 }
581
582 // Make sure that we can do this. Since we can't insert GEPs in a basic
583 // block before a PHI node, we can't easily do this transformation if
584 // we have PHI node users of transformed instructions.
585 for (Value *Val : Explored) {
586 for (Value *Use : Val->uses()) {
587
588 auto *PHI = dyn_cast<PHINode>(Use);
589 auto *Inst = dyn_cast<Instruction>(Val);
590
591 if (Inst == Base || Inst == PHI || !Inst || !PHI ||
592 Explored.count(PHI) == 0)
593 continue;
594
595 if (PHI->getParent() == Inst->getParent())
596 return false;
597 }
598 }
599 return true;
600 }
601
602 // Sets the appropriate insert point on Builder where we can add
603 // a replacement Instruction for V (if that is possible).
setInsertionPoint(IRBuilder<> & Builder,Value * V,bool Before=true)604 static void setInsertionPoint(IRBuilder<> &Builder, Value *V,
605 bool Before = true) {
606 if (auto *PHI = dyn_cast<PHINode>(V)) {
607 Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt());
608 return;
609 }
610 if (auto *I = dyn_cast<Instruction>(V)) {
611 if (!Before)
612 I = &*std::next(I->getIterator());
613 Builder.SetInsertPoint(I);
614 return;
615 }
616 if (auto *A = dyn_cast<Argument>(V)) {
617 // Set the insertion point in the entry block.
618 BasicBlock &Entry = A->getParent()->getEntryBlock();
619 Builder.SetInsertPoint(&*Entry.getFirstInsertionPt());
620 return;
621 }
622 // Otherwise, this is a constant and we don't need to set a new
623 // insertion point.
624 assert(isa<Constant>(V) && "Setting insertion point for unknown value!");
625 }
626
627 /// Returns a re-written value of Start as an indexed GEP using Base as a
628 /// pointer.
rewriteGEPAsOffset(Value * Start,Value * Base,const DataLayout & DL,SetVector<Value * > & Explored)629 static Value *rewriteGEPAsOffset(Value *Start, Value *Base,
630 const DataLayout &DL,
631 SetVector<Value *> &Explored) {
632 // Perform all the substitutions. This is a bit tricky because we can
633 // have cycles in our use-def chains.
634 // 1. Create the PHI nodes without any incoming values.
635 // 2. Create all the other values.
636 // 3. Add the edges for the PHI nodes.
637 // 4. Emit GEPs to get the original pointers.
638 // 5. Remove the original instructions.
639 Type *IndexType = IntegerType::get(
640 Base->getContext(), DL.getIndexTypeSizeInBits(Start->getType()));
641
642 DenseMap<Value *, Value *> NewInsts;
643 NewInsts[Base] = ConstantInt::getNullValue(IndexType);
644
645 // Create the new PHI nodes, without adding any incoming values.
646 for (Value *Val : Explored) {
647 if (Val == Base)
648 continue;
649 // Create empty phi nodes. This avoids cyclic dependencies when creating
650 // the remaining instructions.
651 if (auto *PHI = dyn_cast<PHINode>(Val))
652 NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(),
653 PHI->getName() + ".idx", PHI);
654 }
655 IRBuilder<> Builder(Base->getContext());
656
657 // Create all the other instructions.
658 for (Value *Val : Explored) {
659
660 if (NewInsts.find(Val) != NewInsts.end())
661 continue;
662
663 if (auto *CI = dyn_cast<CastInst>(Val)) {
664 // Don't get rid of the intermediate variable here; the store can grow
665 // the map which will invalidate the reference to the input value.
666 Value *V = NewInsts[CI->getOperand(0)];
667 NewInsts[CI] = V;
668 continue;
669 }
670 if (auto *GEP = dyn_cast<GEPOperator>(Val)) {
671 Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)]
672 : GEP->getOperand(1);
673 setInsertionPoint(Builder, GEP);
674 // Indices might need to be sign extended. GEPs will magically do
675 // this, but we need to do it ourselves here.
676 if (Index->getType()->getScalarSizeInBits() !=
677 NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) {
678 Index = Builder.CreateSExtOrTrunc(
679 Index, NewInsts[GEP->getOperand(0)]->getType(),
680 GEP->getOperand(0)->getName() + ".sext");
681 }
682
683 auto *Op = NewInsts[GEP->getOperand(0)];
684 if (isa<ConstantInt>(Op) && cast<ConstantInt>(Op)->isZero())
685 NewInsts[GEP] = Index;
686 else
687 NewInsts[GEP] = Builder.CreateNSWAdd(
688 Op, Index, GEP->getOperand(0)->getName() + ".add");
689 continue;
690 }
691 if (isa<PHINode>(Val))
692 continue;
693
694 llvm_unreachable("Unexpected instruction type");
695 }
696
697 // Add the incoming values to the PHI nodes.
698 for (Value *Val : Explored) {
699 if (Val == Base)
700 continue;
701 // All the instructions have been created, we can now add edges to the
702 // phi nodes.
703 if (auto *PHI = dyn_cast<PHINode>(Val)) {
704 PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]);
705 for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
706 Value *NewIncoming = PHI->getIncomingValue(I);
707
708 if (NewInsts.find(NewIncoming) != NewInsts.end())
709 NewIncoming = NewInsts[NewIncoming];
710
711 NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I));
712 }
713 }
714 }
715
716 for (Value *Val : Explored) {
717 if (Val == Base)
718 continue;
719
720 // Depending on the type, for external users we have to emit
721 // a GEP or a GEP + ptrtoint.
722 setInsertionPoint(Builder, Val, false);
723
724 // If required, create an inttoptr instruction for Base.
725 Value *NewBase = Base;
726 if (!Base->getType()->isPointerTy())
727 NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(),
728 Start->getName() + "to.ptr");
729
730 Value *GEP = Builder.CreateInBoundsGEP(
731 Start->getType()->getPointerElementType(), NewBase,
732 makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr");
733
734 if (!Val->getType()->isPointerTy()) {
735 Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(),
736 Val->getName() + ".conv");
737 GEP = Cast;
738 }
739 Val->replaceAllUsesWith(GEP);
740 }
741
742 return NewInsts[Start];
743 }
744
745 /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express
746 /// the input Value as a constant indexed GEP. Returns a pair containing
747 /// the GEPs Pointer and Index.
748 static std::pair<Value *, Value *>
getAsConstantIndexedAddress(Value * V,const DataLayout & DL)749 getAsConstantIndexedAddress(Value *V, const DataLayout &DL) {
750 Type *IndexType = IntegerType::get(V->getContext(),
751 DL.getIndexTypeSizeInBits(V->getType()));
752
753 Constant *Index = ConstantInt::getNullValue(IndexType);
754 while (true) {
755 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
756 // We accept only inbouds GEPs here to exclude the possibility of
757 // overflow.
758 if (!GEP->isInBounds())
759 break;
760 if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 &&
761 GEP->getType() == V->getType()) {
762 V = GEP->getOperand(0);
763 Constant *GEPIndex = static_cast<Constant *>(GEP->getOperand(1));
764 Index = ConstantExpr::getAdd(
765 Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType));
766 continue;
767 }
768 break;
769 }
770 if (auto *CI = dyn_cast<IntToPtrInst>(V)) {
771 if (!CI->isNoopCast(DL))
772 break;
773 V = CI->getOperand(0);
774 continue;
775 }
776 if (auto *CI = dyn_cast<PtrToIntInst>(V)) {
777 if (!CI->isNoopCast(DL))
778 break;
779 V = CI->getOperand(0);
780 continue;
781 }
782 break;
783 }
784 return {V, Index};
785 }
786
787 /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant.
788 /// We can look through PHIs, GEPs and casts in order to determine a common base
789 /// between GEPLHS and RHS.
transformToIndexedCompare(GEPOperator * GEPLHS,Value * RHS,ICmpInst::Predicate Cond,const DataLayout & DL)790 static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS,
791 ICmpInst::Predicate Cond,
792 const DataLayout &DL) {
793 // FIXME: Support vector of pointers.
794 if (GEPLHS->getType()->isVectorTy())
795 return nullptr;
796
797 if (!GEPLHS->hasAllConstantIndices())
798 return nullptr;
799
800 // Make sure the pointers have the same type.
801 if (GEPLHS->getType() != RHS->getType())
802 return nullptr;
803
804 Value *PtrBase, *Index;
805 std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL);
806
807 // The set of nodes that will take part in this transformation.
808 SetVector<Value *> Nodes;
809
810 if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes))
811 return nullptr;
812
813 // We know we can re-write this as
814 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)
815 // Since we've only looked through inbouds GEPs we know that we
816 // can't have overflow on either side. We can therefore re-write
817 // this as:
818 // OFFSET1 cmp OFFSET2
819 Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes);
820
821 // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written
822 // GEP having PtrBase as the pointer base, and has returned in NewRHS the
823 // offset. Since Index is the offset of LHS to the base pointer, we will now
824 // compare the offsets instead of comparing the pointers.
825 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS);
826 }
827
828 /// Fold comparisons between a GEP instruction and something else. At this point
829 /// we know that the GEP is on the LHS of the comparison.
foldGEPICmp(GEPOperator * GEPLHS,Value * RHS,ICmpInst::Predicate Cond,Instruction & I)830 Instruction *InstCombinerImpl::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
831 ICmpInst::Predicate Cond,
832 Instruction &I) {
833 // Don't transform signed compares of GEPs into index compares. Even if the
834 // GEP is inbounds, the final add of the base pointer can have signed overflow
835 // and would change the result of the icmp.
836 // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
837 // the maximum signed value for the pointer type.
838 if (ICmpInst::isSigned(Cond))
839 return nullptr;
840
841 // Look through bitcasts and addrspacecasts. We do not however want to remove
842 // 0 GEPs.
843 if (!isa<GetElementPtrInst>(RHS))
844 RHS = RHS->stripPointerCasts();
845
846 Value *PtrBase = GEPLHS->getOperand(0);
847 // FIXME: Support vector pointer GEPs.
848 if (PtrBase == RHS && GEPLHS->isInBounds() &&
849 !GEPLHS->getType()->isVectorTy()) {
850 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
851 // This transformation (ignoring the base and scales) is valid because we
852 // know pointers can't overflow since the gep is inbounds. See if we can
853 // output an optimized form.
854 Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL);
855
856 // If not, synthesize the offset the hard way.
857 if (!Offset)
858 Offset = EmitGEPOffset(GEPLHS);
859 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
860 Constant::getNullValue(Offset->getType()));
861 }
862
863 if (GEPLHS->isInBounds() && ICmpInst::isEquality(Cond) &&
864 isa<Constant>(RHS) && cast<Constant>(RHS)->isNullValue() &&
865 !NullPointerIsDefined(I.getFunction(),
866 RHS->getType()->getPointerAddressSpace())) {
867 // For most address spaces, an allocation can't be placed at null, but null
868 // itself is treated as a 0 size allocation in the in bounds rules. Thus,
869 // the only valid inbounds address derived from null, is null itself.
870 // Thus, we have four cases to consider:
871 // 1) Base == nullptr, Offset == 0 -> inbounds, null
872 // 2) Base == nullptr, Offset != 0 -> poison as the result is out of bounds
873 // 3) Base != nullptr, Offset == (-base) -> poison (crossing allocations)
874 // 4) Base != nullptr, Offset != (-base) -> nonnull (and possibly poison)
875 //
876 // (Note if we're indexing a type of size 0, that simply collapses into one
877 // of the buckets above.)
878 //
879 // In general, we're allowed to make values less poison (i.e. remove
880 // sources of full UB), so in this case, we just select between the two
881 // non-poison cases (1 and 4 above).
882 //
883 // For vectors, we apply the same reasoning on a per-lane basis.
884 auto *Base = GEPLHS->getPointerOperand();
885 if (GEPLHS->getType()->isVectorTy() && Base->getType()->isPointerTy()) {
886 auto EC = cast<VectorType>(GEPLHS->getType())->getElementCount();
887 Base = Builder.CreateVectorSplat(EC, Base);
888 }
889 return new ICmpInst(Cond, Base,
890 ConstantExpr::getPointerBitCastOrAddrSpaceCast(
891 cast<Constant>(RHS), Base->getType()));
892 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
893 // If the base pointers are different, but the indices are the same, just
894 // compare the base pointer.
895 if (PtrBase != GEPRHS->getOperand(0)) {
896 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
897 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
898 GEPRHS->getOperand(0)->getType();
899 if (IndicesTheSame)
900 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
901 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
902 IndicesTheSame = false;
903 break;
904 }
905
906 // If all indices are the same, just compare the base pointers.
907 Type *BaseType = GEPLHS->getOperand(0)->getType();
908 if (IndicesTheSame && CmpInst::makeCmpResultType(BaseType) == I.getType())
909 return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
910
911 // If we're comparing GEPs with two base pointers that only differ in type
912 // and both GEPs have only constant indices or just one use, then fold
913 // the compare with the adjusted indices.
914 // FIXME: Support vector of pointers.
915 if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
916 (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
917 (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
918 PtrBase->stripPointerCasts() ==
919 GEPRHS->getOperand(0)->stripPointerCasts() &&
920 !GEPLHS->getType()->isVectorTy()) {
921 Value *LOffset = EmitGEPOffset(GEPLHS);
922 Value *ROffset = EmitGEPOffset(GEPRHS);
923
924 // If we looked through an addrspacecast between different sized address
925 // spaces, the LHS and RHS pointers are different sized
926 // integers. Truncate to the smaller one.
927 Type *LHSIndexTy = LOffset->getType();
928 Type *RHSIndexTy = ROffset->getType();
929 if (LHSIndexTy != RHSIndexTy) {
930 if (LHSIndexTy->getPrimitiveSizeInBits().getFixedSize() <
931 RHSIndexTy->getPrimitiveSizeInBits().getFixedSize()) {
932 ROffset = Builder.CreateTrunc(ROffset, LHSIndexTy);
933 } else
934 LOffset = Builder.CreateTrunc(LOffset, RHSIndexTy);
935 }
936
937 Value *Cmp = Builder.CreateICmp(ICmpInst::getSignedPredicate(Cond),
938 LOffset, ROffset);
939 return replaceInstUsesWith(I, Cmp);
940 }
941
942 // Otherwise, the base pointers are different and the indices are
943 // different. Try convert this to an indexed compare by looking through
944 // PHIs/casts.
945 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
946 }
947
948 // If one of the GEPs has all zero indices, recurse.
949 // FIXME: Handle vector of pointers.
950 if (!GEPLHS->getType()->isVectorTy() && GEPLHS->hasAllZeroIndices())
951 return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
952 ICmpInst::getSwappedPredicate(Cond), I);
953
954 // If the other GEP has all zero indices, recurse.
955 // FIXME: Handle vector of pointers.
956 if (!GEPRHS->getType()->isVectorTy() && GEPRHS->hasAllZeroIndices())
957 return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
958
959 bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds();
960 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
961 // If the GEPs only differ by one index, compare it.
962 unsigned NumDifferences = 0; // Keep track of # differences.
963 unsigned DiffOperand = 0; // The operand that differs.
964 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
965 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
966 Type *LHSType = GEPLHS->getOperand(i)->getType();
967 Type *RHSType = GEPRHS->getOperand(i)->getType();
968 // FIXME: Better support for vector of pointers.
969 if (LHSType->getPrimitiveSizeInBits() !=
970 RHSType->getPrimitiveSizeInBits() ||
971 (GEPLHS->getType()->isVectorTy() &&
972 (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) {
973 // Irreconcilable differences.
974 NumDifferences = 2;
975 break;
976 }
977
978 if (NumDifferences++) break;
979 DiffOperand = i;
980 }
981
982 if (NumDifferences == 0) // SAME GEP?
983 return replaceInstUsesWith(I, // No comparison is needed here.
984 ConstantInt::get(I.getType(), ICmpInst::isTrueWhenEqual(Cond)));
985
986 else if (NumDifferences == 1 && GEPsInBounds) {
987 Value *LHSV = GEPLHS->getOperand(DiffOperand);
988 Value *RHSV = GEPRHS->getOperand(DiffOperand);
989 // Make sure we do a signed comparison here.
990 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
991 }
992 }
993
994 // Only lower this if the icmp is the only user of the GEP or if we expect
995 // the result to fold to a constant!
996 if (GEPsInBounds && (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
997 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
998 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
999 Value *L = EmitGEPOffset(GEPLHS);
1000 Value *R = EmitGEPOffset(GEPRHS);
1001 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
1002 }
1003 }
1004
1005 // Try convert this to an indexed compare by looking through PHIs/casts as a
1006 // last resort.
1007 return transformToIndexedCompare(GEPLHS, RHS, Cond, DL);
1008 }
1009
foldAllocaCmp(ICmpInst & ICI,const AllocaInst * Alloca,const Value * Other)1010 Instruction *InstCombinerImpl::foldAllocaCmp(ICmpInst &ICI,
1011 const AllocaInst *Alloca,
1012 const Value *Other) {
1013 assert(ICI.isEquality() && "Cannot fold non-equality comparison.");
1014
1015 // It would be tempting to fold away comparisons between allocas and any
1016 // pointer not based on that alloca (e.g. an argument). However, even
1017 // though such pointers cannot alias, they can still compare equal.
1018 //
1019 // But LLVM doesn't specify where allocas get their memory, so if the alloca
1020 // doesn't escape we can argue that it's impossible to guess its value, and we
1021 // can therefore act as if any such guesses are wrong.
1022 //
1023 // The code below checks that the alloca doesn't escape, and that it's only
1024 // used in a comparison once (the current instruction). The
1025 // single-comparison-use condition ensures that we're trivially folding all
1026 // comparisons against the alloca consistently, and avoids the risk of
1027 // erroneously folding a comparison of the pointer with itself.
1028
1029 unsigned MaxIter = 32; // Break cycles and bound to constant-time.
1030
1031 SmallVector<const Use *, 32> Worklist;
1032 for (const Use &U : Alloca->uses()) {
1033 if (Worklist.size() >= MaxIter)
1034 return nullptr;
1035 Worklist.push_back(&U);
1036 }
1037
1038 unsigned NumCmps = 0;
1039 while (!Worklist.empty()) {
1040 assert(Worklist.size() <= MaxIter);
1041 const Use *U = Worklist.pop_back_val();
1042 const Value *V = U->getUser();
1043 --MaxIter;
1044
1045 if (isa<BitCastInst>(V) || isa<GetElementPtrInst>(V) || isa<PHINode>(V) ||
1046 isa<SelectInst>(V)) {
1047 // Track the uses.
1048 } else if (isa<LoadInst>(V)) {
1049 // Loading from the pointer doesn't escape it.
1050 continue;
1051 } else if (const auto *SI = dyn_cast<StoreInst>(V)) {
1052 // Storing *to* the pointer is fine, but storing the pointer escapes it.
1053 if (SI->getValueOperand() == U->get())
1054 return nullptr;
1055 continue;
1056 } else if (isa<ICmpInst>(V)) {
1057 if (NumCmps++)
1058 return nullptr; // Found more than one cmp.
1059 continue;
1060 } else if (const auto *Intrin = dyn_cast<IntrinsicInst>(V)) {
1061 switch (Intrin->getIntrinsicID()) {
1062 // These intrinsics don't escape or compare the pointer. Memset is safe
1063 // because we don't allow ptrtoint. Memcpy and memmove are safe because
1064 // we don't allow stores, so src cannot point to V.
1065 case Intrinsic::lifetime_start: case Intrinsic::lifetime_end:
1066 case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset:
1067 continue;
1068 default:
1069 return nullptr;
1070 }
1071 } else {
1072 return nullptr;
1073 }
1074 for (const Use &U : V->uses()) {
1075 if (Worklist.size() >= MaxIter)
1076 return nullptr;
1077 Worklist.push_back(&U);
1078 }
1079 }
1080
1081 Type *CmpTy = CmpInst::makeCmpResultType(Other->getType());
1082 return replaceInstUsesWith(
1083 ICI,
1084 ConstantInt::get(CmpTy, !CmpInst::isTrueWhenEqual(ICI.getPredicate())));
1085 }
1086
1087 /// Fold "icmp pred (X+C), X".
foldICmpAddOpConst(Value * X,const APInt & C,ICmpInst::Predicate Pred)1088 Instruction *InstCombinerImpl::foldICmpAddOpConst(Value *X, const APInt &C,
1089 ICmpInst::Predicate Pred) {
1090 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
1091 // so the values can never be equal. Similarly for all other "or equals"
1092 // operators.
1093 assert(!!C && "C should not be zero!");
1094
1095 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
1096 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
1097 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
1098 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
1099 Constant *R = ConstantInt::get(X->getType(),
1100 APInt::getMaxValue(C.getBitWidth()) - C);
1101 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
1102 }
1103
1104 // (X+1) >u X --> X <u (0-1) --> X != 255
1105 // (X+2) >u X --> X <u (0-2) --> X <u 254
1106 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
1107 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
1108 return new ICmpInst(ICmpInst::ICMP_ULT, X,
1109 ConstantInt::get(X->getType(), -C));
1110
1111 APInt SMax = APInt::getSignedMaxValue(C.getBitWidth());
1112
1113 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
1114 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
1115 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
1116 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
1117 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
1118 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
1119 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
1120 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1121 ConstantInt::get(X->getType(), SMax - C));
1122
1123 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
1124 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
1125 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
1126 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
1127 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
1128 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
1129
1130 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
1131 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1132 ConstantInt::get(X->getType(), SMax - (C - 1)));
1133 }
1134
1135 /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" ->
1136 /// (icmp eq/ne A, Log2(AP2/AP1)) ->
1137 /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)).
foldICmpShrConstConst(ICmpInst & I,Value * A,const APInt & AP1,const APInt & AP2)1138 Instruction *InstCombinerImpl::foldICmpShrConstConst(ICmpInst &I, Value *A,
1139 const APInt &AP1,
1140 const APInt &AP2) {
1141 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1142
1143 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1144 if (I.getPredicate() == I.ICMP_NE)
1145 Pred = CmpInst::getInversePredicate(Pred);
1146 return new ICmpInst(Pred, LHS, RHS);
1147 };
1148
1149 // Don't bother doing any work for cases which InstSimplify handles.
1150 if (AP2.isZero())
1151 return nullptr;
1152
1153 bool IsAShr = isa<AShrOperator>(I.getOperand(0));
1154 if (IsAShr) {
1155 if (AP2.isAllOnes())
1156 return nullptr;
1157 if (AP2.isNegative() != AP1.isNegative())
1158 return nullptr;
1159 if (AP2.sgt(AP1))
1160 return nullptr;
1161 }
1162
1163 if (!AP1)
1164 // 'A' must be large enough to shift out the highest set bit.
1165 return getICmp(I.ICMP_UGT, A,
1166 ConstantInt::get(A->getType(), AP2.logBase2()));
1167
1168 if (AP1 == AP2)
1169 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1170
1171 int Shift;
1172 if (IsAShr && AP1.isNegative())
1173 Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes();
1174 else
1175 Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros();
1176
1177 if (Shift > 0) {
1178 if (IsAShr && AP1 == AP2.ashr(Shift)) {
1179 // There are multiple solutions if we are comparing against -1 and the LHS
1180 // of the ashr is not a power of two.
1181 if (AP1.isAllOnes() && !AP2.isPowerOf2())
1182 return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift));
1183 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1184 } else if (AP1 == AP2.lshr(Shift)) {
1185 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1186 }
1187 }
1188
1189 // Shifting const2 will never be equal to const1.
1190 // FIXME: This should always be handled by InstSimplify?
1191 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1192 return replaceInstUsesWith(I, TorF);
1193 }
1194
1195 /// Handle "(icmp eq/ne (shl AP2, A), AP1)" ->
1196 /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)).
foldICmpShlConstConst(ICmpInst & I,Value * A,const APInt & AP1,const APInt & AP2)1197 Instruction *InstCombinerImpl::foldICmpShlConstConst(ICmpInst &I, Value *A,
1198 const APInt &AP1,
1199 const APInt &AP2) {
1200 assert(I.isEquality() && "Cannot fold icmp gt/lt");
1201
1202 auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) {
1203 if (I.getPredicate() == I.ICMP_NE)
1204 Pred = CmpInst::getInversePredicate(Pred);
1205 return new ICmpInst(Pred, LHS, RHS);
1206 };
1207
1208 // Don't bother doing any work for cases which InstSimplify handles.
1209 if (AP2.isZero())
1210 return nullptr;
1211
1212 unsigned AP2TrailingZeros = AP2.countTrailingZeros();
1213
1214 if (!AP1 && AP2TrailingZeros != 0)
1215 return getICmp(
1216 I.ICMP_UGE, A,
1217 ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros));
1218
1219 if (AP1 == AP2)
1220 return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType()));
1221
1222 // Get the distance between the lowest bits that are set.
1223 int Shift = AP1.countTrailingZeros() - AP2TrailingZeros;
1224
1225 if (Shift > 0 && AP2.shl(Shift) == AP1)
1226 return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift));
1227
1228 // Shifting const2 will never be equal to const1.
1229 // FIXME: This should always be handled by InstSimplify?
1230 auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE);
1231 return replaceInstUsesWith(I, TorF);
1232 }
1233
1234 /// The caller has matched a pattern of the form:
1235 /// I = icmp ugt (add (add A, B), CI2), CI1
1236 /// If this is of the form:
1237 /// sum = a + b
1238 /// if (sum+128 >u 255)
1239 /// Then replace it with llvm.sadd.with.overflow.i8.
1240 ///
processUGT_ADDCST_ADD(ICmpInst & I,Value * A,Value * B,ConstantInt * CI2,ConstantInt * CI1,InstCombinerImpl & IC)1241 static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1242 ConstantInt *CI2, ConstantInt *CI1,
1243 InstCombinerImpl &IC) {
1244 // The transformation we're trying to do here is to transform this into an
1245 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1246 // with a narrower add, and discard the add-with-constant that is part of the
1247 // range check (if we can't eliminate it, this isn't profitable).
1248
1249 // In order to eliminate the add-with-constant, the compare can be its only
1250 // use.
1251 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1252 if (!AddWithCst->hasOneUse())
1253 return nullptr;
1254
1255 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1256 if (!CI2->getValue().isPowerOf2())
1257 return nullptr;
1258 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1259 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31)
1260 return nullptr;
1261
1262 // The width of the new add formed is 1 more than the bias.
1263 ++NewWidth;
1264
1265 // Check to see that CI1 is an all-ones value with NewWidth bits.
1266 if (CI1->getBitWidth() == NewWidth ||
1267 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1268 return nullptr;
1269
1270 // This is only really a signed overflow check if the inputs have been
1271 // sign-extended; check for that condition. For example, if CI2 is 2^31 and
1272 // the operands of the add are 64 bits wide, we need at least 33 sign bits.
1273 unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1;
1274 if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits ||
1275 IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits)
1276 return nullptr;
1277
1278 // In order to replace the original add with a narrower
1279 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1280 // and truncates that discard the high bits of the add. Verify that this is
1281 // the case.
1282 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1283 for (User *U : OrigAdd->users()) {
1284 if (U == AddWithCst)
1285 continue;
1286
1287 // Only accept truncates for now. We would really like a nice recursive
1288 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1289 // chain to see which bits of a value are actually demanded. If the
1290 // original add had another add which was then immediately truncated, we
1291 // could still do the transformation.
1292 TruncInst *TI = dyn_cast<TruncInst>(U);
1293 if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth)
1294 return nullptr;
1295 }
1296
1297 // If the pattern matches, truncate the inputs to the narrower type and
1298 // use the sadd_with_overflow intrinsic to efficiently compute both the
1299 // result and the overflow bit.
1300 Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1301 Function *F = Intrinsic::getDeclaration(
1302 I.getModule(), Intrinsic::sadd_with_overflow, NewType);
1303
1304 InstCombiner::BuilderTy &Builder = IC.Builder;
1305
1306 // Put the new code above the original add, in case there are any uses of the
1307 // add between the add and the compare.
1308 Builder.SetInsertPoint(OrigAdd);
1309
1310 Value *TruncA = Builder.CreateTrunc(A, NewType, A->getName() + ".trunc");
1311 Value *TruncB = Builder.CreateTrunc(B, NewType, B->getName() + ".trunc");
1312 CallInst *Call = Builder.CreateCall(F, {TruncA, TruncB}, "sadd");
1313 Value *Add = Builder.CreateExtractValue(Call, 0, "sadd.result");
1314 Value *ZExt = Builder.CreateZExt(Add, OrigAdd->getType());
1315
1316 // The inner add was the result of the narrow add, zero extended to the
1317 // wider type. Replace it with the result computed by the intrinsic.
1318 IC.replaceInstUsesWith(*OrigAdd, ZExt);
1319 IC.eraseInstFromFunction(*OrigAdd);
1320
1321 // The original icmp gets replaced with the overflow value.
1322 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1323 }
1324
1325 /// If we have:
1326 /// icmp eq/ne (urem/srem %x, %y), 0
1327 /// iff %y is a power-of-two, we can replace this with a bit test:
1328 /// icmp eq/ne (and %x, (add %y, -1)), 0
foldIRemByPowerOfTwoToBitTest(ICmpInst & I)1329 Instruction *InstCombinerImpl::foldIRemByPowerOfTwoToBitTest(ICmpInst &I) {
1330 // This fold is only valid for equality predicates.
1331 if (!I.isEquality())
1332 return nullptr;
1333 ICmpInst::Predicate Pred;
1334 Value *X, *Y, *Zero;
1335 if (!match(&I, m_ICmp(Pred, m_OneUse(m_IRem(m_Value(X), m_Value(Y))),
1336 m_CombineAnd(m_Zero(), m_Value(Zero)))))
1337 return nullptr;
1338 if (!isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, 0, &I))
1339 return nullptr;
1340 // This may increase instruction count, we don't enforce that Y is a constant.
1341 Value *Mask = Builder.CreateAdd(Y, Constant::getAllOnesValue(Y->getType()));
1342 Value *Masked = Builder.CreateAnd(X, Mask);
1343 return ICmpInst::Create(Instruction::ICmp, Pred, Masked, Zero);
1344 }
1345
1346 /// Fold equality-comparison between zero and any (maybe truncated) right-shift
1347 /// by one-less-than-bitwidth into a sign test on the original value.
foldSignBitTest(ICmpInst & I)1348 Instruction *InstCombinerImpl::foldSignBitTest(ICmpInst &I) {
1349 Instruction *Val;
1350 ICmpInst::Predicate Pred;
1351 if (!I.isEquality() || !match(&I, m_ICmp(Pred, m_Instruction(Val), m_Zero())))
1352 return nullptr;
1353
1354 Value *X;
1355 Type *XTy;
1356
1357 Constant *C;
1358 if (match(Val, m_TruncOrSelf(m_Shr(m_Value(X), m_Constant(C))))) {
1359 XTy = X->getType();
1360 unsigned XBitWidth = XTy->getScalarSizeInBits();
1361 if (!match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
1362 APInt(XBitWidth, XBitWidth - 1))))
1363 return nullptr;
1364 } else if (isa<BinaryOperator>(Val) &&
1365 (X = reassociateShiftAmtsOfTwoSameDirectionShifts(
1366 cast<BinaryOperator>(Val), SQ.getWithInstruction(Val),
1367 /*AnalyzeForSignBitExtraction=*/true))) {
1368 XTy = X->getType();
1369 } else
1370 return nullptr;
1371
1372 return ICmpInst::Create(Instruction::ICmp,
1373 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_SGE
1374 : ICmpInst::ICMP_SLT,
1375 X, ConstantInt::getNullValue(XTy));
1376 }
1377
1378 // Handle icmp pred X, 0
foldICmpWithZero(ICmpInst & Cmp)1379 Instruction *InstCombinerImpl::foldICmpWithZero(ICmpInst &Cmp) {
1380 CmpInst::Predicate Pred = Cmp.getPredicate();
1381 if (!match(Cmp.getOperand(1), m_Zero()))
1382 return nullptr;
1383
1384 // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
1385 if (Pred == ICmpInst::ICMP_SGT) {
1386 Value *A, *B;
1387 SelectPatternResult SPR = matchSelectPattern(Cmp.getOperand(0), A, B);
1388 if (SPR.Flavor == SPF_SMIN) {
1389 if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
1390 return new ICmpInst(Pred, B, Cmp.getOperand(1));
1391 if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
1392 return new ICmpInst(Pred, A, Cmp.getOperand(1));
1393 }
1394 }
1395
1396 if (Instruction *New = foldIRemByPowerOfTwoToBitTest(Cmp))
1397 return New;
1398
1399 // Given:
1400 // icmp eq/ne (urem %x, %y), 0
1401 // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem':
1402 // icmp eq/ne %x, 0
1403 Value *X, *Y;
1404 if (match(Cmp.getOperand(0), m_URem(m_Value(X), m_Value(Y))) &&
1405 ICmpInst::isEquality(Pred)) {
1406 KnownBits XKnown = computeKnownBits(X, 0, &Cmp);
1407 KnownBits YKnown = computeKnownBits(Y, 0, &Cmp);
1408 if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2)
1409 return new ICmpInst(Pred, X, Cmp.getOperand(1));
1410 }
1411
1412 return nullptr;
1413 }
1414
1415 /// Fold icmp Pred X, C.
1416 /// TODO: This code structure does not make sense. The saturating add fold
1417 /// should be moved to some other helper and extended as noted below (it is also
1418 /// possible that code has been made unnecessary - do we canonicalize IR to
1419 /// overflow/saturating intrinsics or not?).
foldICmpWithConstant(ICmpInst & Cmp)1420 Instruction *InstCombinerImpl::foldICmpWithConstant(ICmpInst &Cmp) {
1421 // Match the following pattern, which is a common idiom when writing
1422 // overflow-safe integer arithmetic functions. The source performs an addition
1423 // in wider type and explicitly checks for overflow using comparisons against
1424 // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic.
1425 //
1426 // TODO: This could probably be generalized to handle other overflow-safe
1427 // operations if we worked out the formulas to compute the appropriate magic
1428 // constants.
1429 //
1430 // sum = a + b
1431 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1432 CmpInst::Predicate Pred = Cmp.getPredicate();
1433 Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
1434 Value *A, *B;
1435 ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1436 if (Pred == ICmpInst::ICMP_UGT && match(Op1, m_ConstantInt(CI)) &&
1437 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1438 if (Instruction *Res = processUGT_ADDCST_ADD(Cmp, A, B, CI2, CI, *this))
1439 return Res;
1440
1441 // icmp(phi(C1, C2, ...), C) -> phi(icmp(C1, C), icmp(C2, C), ...).
1442 Constant *C = dyn_cast<Constant>(Op1);
1443 if (!C || C->canTrap())
1444 return nullptr;
1445
1446 if (auto *Phi = dyn_cast<PHINode>(Op0))
1447 if (all_of(Phi->operands(), [](Value *V) { return isa<Constant>(V); })) {
1448 Type *Ty = Cmp.getType();
1449 Builder.SetInsertPoint(Phi);
1450 PHINode *NewPhi =
1451 Builder.CreatePHI(Ty, Phi->getNumOperands());
1452 for (BasicBlock *Predecessor : predecessors(Phi->getParent())) {
1453 auto *Input =
1454 cast<Constant>(Phi->getIncomingValueForBlock(Predecessor));
1455 auto *BoolInput = ConstantExpr::getCompare(Pred, Input, C);
1456 NewPhi->addIncoming(BoolInput, Predecessor);
1457 }
1458 NewPhi->takeName(&Cmp);
1459 return replaceInstUsesWith(Cmp, NewPhi);
1460 }
1461
1462 return nullptr;
1463 }
1464
1465 /// Canonicalize icmp instructions based on dominating conditions.
foldICmpWithDominatingICmp(ICmpInst & Cmp)1466 Instruction *InstCombinerImpl::foldICmpWithDominatingICmp(ICmpInst &Cmp) {
1467 // This is a cheap/incomplete check for dominance - just match a single
1468 // predecessor with a conditional branch.
1469 BasicBlock *CmpBB = Cmp.getParent();
1470 BasicBlock *DomBB = CmpBB->getSinglePredecessor();
1471 if (!DomBB)
1472 return nullptr;
1473
1474 Value *DomCond;
1475 BasicBlock *TrueBB, *FalseBB;
1476 if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB)))
1477 return nullptr;
1478
1479 assert((TrueBB == CmpBB || FalseBB == CmpBB) &&
1480 "Predecessor block does not point to successor?");
1481
1482 // The branch should get simplified. Don't bother simplifying this condition.
1483 if (TrueBB == FalseBB)
1484 return nullptr;
1485
1486 // Try to simplify this compare to T/F based on the dominating condition.
1487 Optional<bool> Imp = isImpliedCondition(DomCond, &Cmp, DL, TrueBB == CmpBB);
1488 if (Imp)
1489 return replaceInstUsesWith(Cmp, ConstantInt::get(Cmp.getType(), *Imp));
1490
1491 CmpInst::Predicate Pred = Cmp.getPredicate();
1492 Value *X = Cmp.getOperand(0), *Y = Cmp.getOperand(1);
1493 ICmpInst::Predicate DomPred;
1494 const APInt *C, *DomC;
1495 if (match(DomCond, m_ICmp(DomPred, m_Specific(X), m_APInt(DomC))) &&
1496 match(Y, m_APInt(C))) {
1497 // We have 2 compares of a variable with constants. Calculate the constant
1498 // ranges of those compares to see if we can transform the 2nd compare:
1499 // DomBB:
1500 // DomCond = icmp DomPred X, DomC
1501 // br DomCond, CmpBB, FalseBB
1502 // CmpBB:
1503 // Cmp = icmp Pred X, C
1504 ConstantRange CR = ConstantRange::makeExactICmpRegion(Pred, *C);
1505 ConstantRange DominatingCR =
1506 (CmpBB == TrueBB) ? ConstantRange::makeExactICmpRegion(DomPred, *DomC)
1507 : ConstantRange::makeExactICmpRegion(
1508 CmpInst::getInversePredicate(DomPred), *DomC);
1509 ConstantRange Intersection = DominatingCR.intersectWith(CR);
1510 ConstantRange Difference = DominatingCR.difference(CR);
1511 if (Intersection.isEmptySet())
1512 return replaceInstUsesWith(Cmp, Builder.getFalse());
1513 if (Difference.isEmptySet())
1514 return replaceInstUsesWith(Cmp, Builder.getTrue());
1515
1516 // Canonicalizing a sign bit comparison that gets used in a branch,
1517 // pessimizes codegen by generating branch on zero instruction instead
1518 // of a test and branch. So we avoid canonicalizing in such situations
1519 // because test and branch instruction has better branch displacement
1520 // than compare and branch instruction.
1521 bool UnusedBit;
1522 bool IsSignBit = isSignBitCheck(Pred, *C, UnusedBit);
1523 if (Cmp.isEquality() || (IsSignBit && hasBranchUse(Cmp)))
1524 return nullptr;
1525
1526 // Avoid an infinite loop with min/max canonicalization.
1527 // TODO: This will be unnecessary if we canonicalize to min/max intrinsics.
1528 if (Cmp.hasOneUse() &&
1529 match(Cmp.user_back(), m_MaxOrMin(m_Value(), m_Value())))
1530 return nullptr;
1531
1532 if (const APInt *EqC = Intersection.getSingleElement())
1533 return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(*EqC));
1534 if (const APInt *NeC = Difference.getSingleElement())
1535 return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(*NeC));
1536 }
1537
1538 return nullptr;
1539 }
1540
1541 /// Fold icmp (trunc X, Y), C.
foldICmpTruncConstant(ICmpInst & Cmp,TruncInst * Trunc,const APInt & C)1542 Instruction *InstCombinerImpl::foldICmpTruncConstant(ICmpInst &Cmp,
1543 TruncInst *Trunc,
1544 const APInt &C) {
1545 ICmpInst::Predicate Pred = Cmp.getPredicate();
1546 Value *X = Trunc->getOperand(0);
1547 if (C.isOne() && C.getBitWidth() > 1) {
1548 // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1
1549 Value *V = nullptr;
1550 if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V))))
1551 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1552 ConstantInt::get(V->getType(), 1));
1553 }
1554
1555 unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
1556 SrcBits = X->getType()->getScalarSizeInBits();
1557 if (Cmp.isEquality() && Trunc->hasOneUse()) {
1558 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
1559 // of the high bits truncated out of x are known.
1560 KnownBits Known = computeKnownBits(X, 0, &Cmp);
1561
1562 // If all the high bits are known, we can do this xform.
1563 if ((Known.Zero | Known.One).countLeadingOnes() >= SrcBits - DstBits) {
1564 // Pull in the high bits from known-ones set.
1565 APInt NewRHS = C.zext(SrcBits);
1566 NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
1567 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS));
1568 }
1569 }
1570
1571 // Look through truncated right-shift of the sign-bit for a sign-bit check:
1572 // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] < 0 --> ShOp < 0
1573 // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] > -1 --> ShOp > -1
1574 Value *ShOp;
1575 const APInt *ShAmtC;
1576 bool TrueIfSigned;
1577 if (isSignBitCheck(Pred, C, TrueIfSigned) &&
1578 match(X, m_Shr(m_Value(ShOp), m_APInt(ShAmtC))) &&
1579 DstBits == SrcBits - ShAmtC->getZExtValue()) {
1580 return TrueIfSigned
1581 ? new ICmpInst(ICmpInst::ICMP_SLT, ShOp,
1582 ConstantInt::getNullValue(X->getType()))
1583 : new ICmpInst(ICmpInst::ICMP_SGT, ShOp,
1584 ConstantInt::getAllOnesValue(X->getType()));
1585 }
1586
1587 return nullptr;
1588 }
1589
1590 /// Fold icmp (xor X, Y), C.
foldICmpXorConstant(ICmpInst & Cmp,BinaryOperator * Xor,const APInt & C)1591 Instruction *InstCombinerImpl::foldICmpXorConstant(ICmpInst &Cmp,
1592 BinaryOperator *Xor,
1593 const APInt &C) {
1594 Value *X = Xor->getOperand(0);
1595 Value *Y = Xor->getOperand(1);
1596 const APInt *XorC;
1597 if (!match(Y, m_APInt(XorC)))
1598 return nullptr;
1599
1600 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1601 // fold the xor.
1602 ICmpInst::Predicate Pred = Cmp.getPredicate();
1603 bool TrueIfSigned = false;
1604 if (isSignBitCheck(Cmp.getPredicate(), C, TrueIfSigned)) {
1605
1606 // If the sign bit of the XorCst is not set, there is no change to
1607 // the operation, just stop using the Xor.
1608 if (!XorC->isNegative())
1609 return replaceOperand(Cmp, 0, X);
1610
1611 // Emit the opposite comparison.
1612 if (TrueIfSigned)
1613 return new ICmpInst(ICmpInst::ICMP_SGT, X,
1614 ConstantInt::getAllOnesValue(X->getType()));
1615 else
1616 return new ICmpInst(ICmpInst::ICMP_SLT, X,
1617 ConstantInt::getNullValue(X->getType()));
1618 }
1619
1620 if (Xor->hasOneUse()) {
1621 // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask))
1622 if (!Cmp.isEquality() && XorC->isSignMask()) {
1623 Pred = Cmp.getFlippedSignednessPredicate();
1624 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1625 }
1626
1627 // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask))
1628 if (!Cmp.isEquality() && XorC->isMaxSignedValue()) {
1629 Pred = Cmp.getFlippedSignednessPredicate();
1630 Pred = Cmp.getSwappedPredicate(Pred);
1631 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), C ^ *XorC));
1632 }
1633 }
1634
1635 // Mask constant magic can eliminate an 'xor' with unsigned compares.
1636 if (Pred == ICmpInst::ICMP_UGT) {
1637 // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2)
1638 if (*XorC == ~C && (C + 1).isPowerOf2())
1639 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
1640 // (xor X, C) >u C --> X >u C (when C+1 is a power of 2)
1641 if (*XorC == C && (C + 1).isPowerOf2())
1642 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
1643 }
1644 if (Pred == ICmpInst::ICMP_ULT) {
1645 // (xor X, -C) <u C --> X >u ~C (when C is a power of 2)
1646 if (*XorC == -C && C.isPowerOf2())
1647 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1648 ConstantInt::get(X->getType(), ~C));
1649 // (xor X, C) <u C --> X >u ~C (when -C is a power of 2)
1650 if (*XorC == C && (-C).isPowerOf2())
1651 return new ICmpInst(ICmpInst::ICMP_UGT, X,
1652 ConstantInt::get(X->getType(), ~C));
1653 }
1654 return nullptr;
1655 }
1656
1657 /// Fold icmp (and (sh X, Y), C2), C1.
foldICmpAndShift(ICmpInst & Cmp,BinaryOperator * And,const APInt & C1,const APInt & C2)1658 Instruction *InstCombinerImpl::foldICmpAndShift(ICmpInst &Cmp,
1659 BinaryOperator *And,
1660 const APInt &C1,
1661 const APInt &C2) {
1662 BinaryOperator *Shift = dyn_cast<BinaryOperator>(And->getOperand(0));
1663 if (!Shift || !Shift->isShift())
1664 return nullptr;
1665
1666 // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could
1667 // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in
1668 // code produced by the clang front-end, for bitfield access.
1669 // This seemingly simple opportunity to fold away a shift turns out to be
1670 // rather complicated. See PR17827 for details.
1671 unsigned ShiftOpcode = Shift->getOpcode();
1672 bool IsShl = ShiftOpcode == Instruction::Shl;
1673 const APInt *C3;
1674 if (match(Shift->getOperand(1), m_APInt(C3))) {
1675 APInt NewAndCst, NewCmpCst;
1676 bool AnyCmpCstBitsShiftedOut;
1677 if (ShiftOpcode == Instruction::Shl) {
1678 // For a left shift, we can fold if the comparison is not signed. We can
1679 // also fold a signed comparison if the mask value and comparison value
1680 // are not negative. These constraints may not be obvious, but we can
1681 // prove that they are correct using an SMT solver.
1682 if (Cmp.isSigned() && (C2.isNegative() || C1.isNegative()))
1683 return nullptr;
1684
1685 NewCmpCst = C1.lshr(*C3);
1686 NewAndCst = C2.lshr(*C3);
1687 AnyCmpCstBitsShiftedOut = NewCmpCst.shl(*C3) != C1;
1688 } else if (ShiftOpcode == Instruction::LShr) {
1689 // For a logical right shift, we can fold if the comparison is not signed.
1690 // We can also fold a signed comparison if the shifted mask value and the
1691 // shifted comparison value are not negative. These constraints may not be
1692 // obvious, but we can prove that they are correct using an SMT solver.
1693 NewCmpCst = C1.shl(*C3);
1694 NewAndCst = C2.shl(*C3);
1695 AnyCmpCstBitsShiftedOut = NewCmpCst.lshr(*C3) != C1;
1696 if (Cmp.isSigned() && (NewAndCst.isNegative() || NewCmpCst.isNegative()))
1697 return nullptr;
1698 } else {
1699 // For an arithmetic shift, check that both constants don't use (in a
1700 // signed sense) the top bits being shifted out.
1701 assert(ShiftOpcode == Instruction::AShr && "Unknown shift opcode");
1702 NewCmpCst = C1.shl(*C3);
1703 NewAndCst = C2.shl(*C3);
1704 AnyCmpCstBitsShiftedOut = NewCmpCst.ashr(*C3) != C1;
1705 if (NewAndCst.ashr(*C3) != C2)
1706 return nullptr;
1707 }
1708
1709 if (AnyCmpCstBitsShiftedOut) {
1710 // If we shifted bits out, the fold is not going to work out. As a
1711 // special case, check to see if this means that the result is always
1712 // true or false now.
1713 if (Cmp.getPredicate() == ICmpInst::ICMP_EQ)
1714 return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType()));
1715 if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1716 return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType()));
1717 } else {
1718 Value *NewAnd = Builder.CreateAnd(
1719 Shift->getOperand(0), ConstantInt::get(And->getType(), NewAndCst));
1720 return new ICmpInst(Cmp.getPredicate(),
1721 NewAnd, ConstantInt::get(And->getType(), NewCmpCst));
1722 }
1723 }
1724
1725 // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is
1726 // preferable because it allows the C2 << Y expression to be hoisted out of a
1727 // loop if Y is invariant and X is not.
1728 if (Shift->hasOneUse() && C1.isZero() && Cmp.isEquality() &&
1729 !Shift->isArithmeticShift() && !isa<Constant>(Shift->getOperand(0))) {
1730 // Compute C2 << Y.
1731 Value *NewShift =
1732 IsShl ? Builder.CreateLShr(And->getOperand(1), Shift->getOperand(1))
1733 : Builder.CreateShl(And->getOperand(1), Shift->getOperand(1));
1734
1735 // Compute X & (C2 << Y).
1736 Value *NewAnd = Builder.CreateAnd(Shift->getOperand(0), NewShift);
1737 return replaceOperand(Cmp, 0, NewAnd);
1738 }
1739
1740 return nullptr;
1741 }
1742
1743 /// Fold icmp (and X, C2), C1.
foldICmpAndConstConst(ICmpInst & Cmp,BinaryOperator * And,const APInt & C1)1744 Instruction *InstCombinerImpl::foldICmpAndConstConst(ICmpInst &Cmp,
1745 BinaryOperator *And,
1746 const APInt &C1) {
1747 bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE;
1748
1749 // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1
1750 // TODO: We canonicalize to the longer form for scalars because we have
1751 // better analysis/folds for icmp, and codegen may be better with icmp.
1752 if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isZero() &&
1753 match(And->getOperand(1), m_One()))
1754 return new TruncInst(And->getOperand(0), Cmp.getType());
1755
1756 const APInt *C2;
1757 Value *X;
1758 if (!match(And, m_And(m_Value(X), m_APInt(C2))))
1759 return nullptr;
1760
1761 // Don't perform the following transforms if the AND has multiple uses
1762 if (!And->hasOneUse())
1763 return nullptr;
1764
1765 if (Cmp.isEquality() && C1.isZero()) {
1766 // Restrict this fold to single-use 'and' (PR10267).
1767 // Replace (and X, (1 << size(X)-1) != 0) with X s< 0
1768 if (C2->isSignMask()) {
1769 Constant *Zero = Constant::getNullValue(X->getType());
1770 auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1771 return new ICmpInst(NewPred, X, Zero);
1772 }
1773
1774 // Restrict this fold only for single-use 'and' (PR10267).
1775 // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two.
1776 if ((~(*C2) + 1).isPowerOf2()) {
1777 Constant *NegBOC =
1778 ConstantExpr::getNeg(cast<Constant>(And->getOperand(1)));
1779 auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1780 return new ICmpInst(NewPred, X, NegBOC);
1781 }
1782 }
1783
1784 // If the LHS is an 'and' of a truncate and we can widen the and/compare to
1785 // the input width without changing the value produced, eliminate the cast:
1786 //
1787 // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1'
1788 //
1789 // We can do this transformation if the constants do not have their sign bits
1790 // set or if it is an equality comparison. Extending a relational comparison
1791 // when we're checking the sign bit would not work.
1792 Value *W;
1793 if (match(And->getOperand(0), m_OneUse(m_Trunc(m_Value(W)))) &&
1794 (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) {
1795 // TODO: Is this a good transform for vectors? Wider types may reduce
1796 // throughput. Should this transform be limited (even for scalars) by using
1797 // shouldChangeType()?
1798 if (!Cmp.getType()->isVectorTy()) {
1799 Type *WideType = W->getType();
1800 unsigned WideScalarBits = WideType->getScalarSizeInBits();
1801 Constant *ZextC1 = ConstantInt::get(WideType, C1.zext(WideScalarBits));
1802 Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits));
1803 Value *NewAnd = Builder.CreateAnd(W, ZextC2, And->getName());
1804 return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1);
1805 }
1806 }
1807
1808 if (Instruction *I = foldICmpAndShift(Cmp, And, C1, *C2))
1809 return I;
1810
1811 // (icmp pred (and (or (lshr A, B), A), 1), 0) -->
1812 // (icmp pred (and A, (or (shl 1, B), 1), 0))
1813 //
1814 // iff pred isn't signed
1815 if (!Cmp.isSigned() && C1.isZero() && And->getOperand(0)->hasOneUse() &&
1816 match(And->getOperand(1), m_One())) {
1817 Constant *One = cast<Constant>(And->getOperand(1));
1818 Value *Or = And->getOperand(0);
1819 Value *A, *B, *LShr;
1820 if (match(Or, m_Or(m_Value(LShr), m_Value(A))) &&
1821 match(LShr, m_LShr(m_Specific(A), m_Value(B)))) {
1822 unsigned UsesRemoved = 0;
1823 if (And->hasOneUse())
1824 ++UsesRemoved;
1825 if (Or->hasOneUse())
1826 ++UsesRemoved;
1827 if (LShr->hasOneUse())
1828 ++UsesRemoved;
1829
1830 // Compute A & ((1 << B) | 1)
1831 Value *NewOr = nullptr;
1832 if (auto *C = dyn_cast<Constant>(B)) {
1833 if (UsesRemoved >= 1)
1834 NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One);
1835 } else {
1836 if (UsesRemoved >= 3)
1837 NewOr = Builder.CreateOr(Builder.CreateShl(One, B, LShr->getName(),
1838 /*HasNUW=*/true),
1839 One, Or->getName());
1840 }
1841 if (NewOr) {
1842 Value *NewAnd = Builder.CreateAnd(A, NewOr, And->getName());
1843 return replaceOperand(Cmp, 0, NewAnd);
1844 }
1845 }
1846 }
1847
1848 return nullptr;
1849 }
1850
1851 /// Fold icmp (and X, Y), C.
foldICmpAndConstant(ICmpInst & Cmp,BinaryOperator * And,const APInt & C)1852 Instruction *InstCombinerImpl::foldICmpAndConstant(ICmpInst &Cmp,
1853 BinaryOperator *And,
1854 const APInt &C) {
1855 if (Instruction *I = foldICmpAndConstConst(Cmp, And, C))
1856 return I;
1857
1858 const ICmpInst::Predicate Pred = Cmp.getPredicate();
1859 bool TrueIfNeg;
1860 if (isSignBitCheck(Pred, C, TrueIfNeg)) {
1861 // ((X - 1) & ~X) < 0 --> X == 0
1862 // ((X - 1) & ~X) >= 0 --> X != 0
1863 Value *X;
1864 if (match(And->getOperand(0), m_Add(m_Value(X), m_AllOnes())) &&
1865 match(And->getOperand(1), m_Not(m_Specific(X)))) {
1866 auto NewPred = TrueIfNeg ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE;
1867 return new ICmpInst(NewPred, X, ConstantInt::getNullValue(X->getType()));
1868 }
1869 }
1870
1871 // TODO: These all require that Y is constant too, so refactor with the above.
1872
1873 // Try to optimize things like "A[i] & 42 == 0" to index computations.
1874 Value *X = And->getOperand(0);
1875 Value *Y = And->getOperand(1);
1876 if (auto *LI = dyn_cast<LoadInst>(X))
1877 if (auto *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1878 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1879 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1880 !LI->isVolatile() && isa<ConstantInt>(Y)) {
1881 ConstantInt *C2 = cast<ConstantInt>(Y);
1882 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2))
1883 return Res;
1884 }
1885
1886 if (!Cmp.isEquality())
1887 return nullptr;
1888
1889 // X & -C == -C -> X > u ~C
1890 // X & -C != -C -> X <= u ~C
1891 // iff C is a power of 2
1892 if (Cmp.getOperand(1) == Y && (-C).isPowerOf2()) {
1893 auto NewPred =
1894 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT : CmpInst::ICMP_ULE;
1895 return new ICmpInst(NewPred, X, SubOne(cast<Constant>(Cmp.getOperand(1))));
1896 }
1897
1898 // (X & C2) == 0 -> (trunc X) >= 0
1899 // (X & C2) != 0 -> (trunc X) < 0
1900 // iff C2 is a power of 2 and it masks the sign bit of a legal integer type.
1901 const APInt *C2;
1902 if (And->hasOneUse() && C.isZero() && match(Y, m_APInt(C2))) {
1903 int32_t ExactLogBase2 = C2->exactLogBase2();
1904 if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) {
1905 Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1);
1906 if (auto *AndVTy = dyn_cast<VectorType>(And->getType()))
1907 NTy = VectorType::get(NTy, AndVTy->getElementCount());
1908 Value *Trunc = Builder.CreateTrunc(X, NTy);
1909 auto NewPred =
1910 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE : CmpInst::ICMP_SLT;
1911 return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy));
1912 }
1913 }
1914
1915 return nullptr;
1916 }
1917
1918 /// Fold icmp (or X, Y), C.
foldICmpOrConstant(ICmpInst & Cmp,BinaryOperator * Or,const APInt & C)1919 Instruction *InstCombinerImpl::foldICmpOrConstant(ICmpInst &Cmp,
1920 BinaryOperator *Or,
1921 const APInt &C) {
1922 ICmpInst::Predicate Pred = Cmp.getPredicate();
1923 if (C.isOne()) {
1924 // icmp slt signum(V) 1 --> icmp slt V, 1
1925 Value *V = nullptr;
1926 if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V))))
1927 return new ICmpInst(ICmpInst::ICMP_SLT, V,
1928 ConstantInt::get(V->getType(), 1));
1929 }
1930
1931 Value *OrOp0 = Or->getOperand(0), *OrOp1 = Or->getOperand(1);
1932 const APInt *MaskC;
1933 if (match(OrOp1, m_APInt(MaskC)) && Cmp.isEquality()) {
1934 if (*MaskC == C && (C + 1).isPowerOf2()) {
1935 // X | C == C --> X <=u C
1936 // X | C != C --> X >u C
1937 // iff C+1 is a power of 2 (C is a bitmask of the low bits)
1938 Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT;
1939 return new ICmpInst(Pred, OrOp0, OrOp1);
1940 }
1941
1942 // More general: canonicalize 'equality with set bits mask' to
1943 // 'equality with clear bits mask'.
1944 // (X | MaskC) == C --> (X & ~MaskC) == C ^ MaskC
1945 // (X | MaskC) != C --> (X & ~MaskC) != C ^ MaskC
1946 if (Or->hasOneUse()) {
1947 Value *And = Builder.CreateAnd(OrOp0, ~(*MaskC));
1948 Constant *NewC = ConstantInt::get(Or->getType(), C ^ (*MaskC));
1949 return new ICmpInst(Pred, And, NewC);
1950 }
1951 }
1952
1953 // (X | (X-1)) s< 0 --> X < 1
1954 // (X | (X-1)) s> -1 --> X > 0
1955 Value *X;
1956 bool TrueIfSigned;
1957 if (isSignBitCheck(Pred, C, TrueIfSigned) &&
1958 match(Or, m_c_Or(m_Add(m_Value(X), m_AllOnes()), m_Deferred(X)))) {
1959 auto NewPred = TrueIfSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGT;
1960 Constant *NewC = ConstantInt::get(X->getType(), TrueIfSigned ? 1 : 0);
1961 return new ICmpInst(NewPred, X, NewC);
1962 }
1963
1964 if (!Cmp.isEquality() || !C.isZero() || !Or->hasOneUse())
1965 return nullptr;
1966
1967 Value *P, *Q;
1968 if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1969 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1970 // -> and (icmp eq P, null), (icmp eq Q, null).
1971 Value *CmpP =
1972 Builder.CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType()));
1973 Value *CmpQ =
1974 Builder.CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType()));
1975 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1976 return BinaryOperator::Create(BOpc, CmpP, CmpQ);
1977 }
1978
1979 // Are we using xors to bitwise check for a pair of (in)equalities? Convert to
1980 // a shorter form that has more potential to be folded even further.
1981 Value *X1, *X2, *X3, *X4;
1982 if (match(OrOp0, m_OneUse(m_Xor(m_Value(X1), m_Value(X2)))) &&
1983 match(OrOp1, m_OneUse(m_Xor(m_Value(X3), m_Value(X4))))) {
1984 // ((X1 ^ X2) || (X3 ^ X4)) == 0 --> (X1 == X2) && (X3 == X4)
1985 // ((X1 ^ X2) || (X3 ^ X4)) != 0 --> (X1 != X2) || (X3 != X4)
1986 Value *Cmp12 = Builder.CreateICmp(Pred, X1, X2);
1987 Value *Cmp34 = Builder.CreateICmp(Pred, X3, X4);
1988 auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or;
1989 return BinaryOperator::Create(BOpc, Cmp12, Cmp34);
1990 }
1991
1992 return nullptr;
1993 }
1994
1995 /// Fold icmp (mul X, Y), C.
foldICmpMulConstant(ICmpInst & Cmp,BinaryOperator * Mul,const APInt & C)1996 Instruction *InstCombinerImpl::foldICmpMulConstant(ICmpInst &Cmp,
1997 BinaryOperator *Mul,
1998 const APInt &C) {
1999 const APInt *MulC;
2000 if (!match(Mul->getOperand(1), m_APInt(MulC)))
2001 return nullptr;
2002
2003 // If this is a test of the sign bit and the multiply is sign-preserving with
2004 // a constant operand, use the multiply LHS operand instead.
2005 ICmpInst::Predicate Pred = Cmp.getPredicate();
2006 if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) {
2007 if (MulC->isNegative())
2008 Pred = ICmpInst::getSwappedPredicate(Pred);
2009 return new ICmpInst(Pred, Mul->getOperand(0),
2010 Constant::getNullValue(Mul->getType()));
2011 }
2012
2013 // If the multiply does not wrap, try to divide the compare constant by the
2014 // multiplication factor.
2015 if (Cmp.isEquality() && !MulC->isZero()) {
2016 // (mul nsw X, MulC) == C --> X == C /s MulC
2017 if (Mul->hasNoSignedWrap() && C.srem(*MulC).isZero()) {
2018 Constant *NewC = ConstantInt::get(Mul->getType(), C.sdiv(*MulC));
2019 return new ICmpInst(Pred, Mul->getOperand(0), NewC);
2020 }
2021 // (mul nuw X, MulC) == C --> X == C /u MulC
2022 if (Mul->hasNoUnsignedWrap() && C.urem(*MulC).isZero()) {
2023 Constant *NewC = ConstantInt::get(Mul->getType(), C.udiv(*MulC));
2024 return new ICmpInst(Pred, Mul->getOperand(0), NewC);
2025 }
2026 }
2027
2028 return nullptr;
2029 }
2030
2031 /// Fold icmp (shl 1, Y), C.
foldICmpShlOne(ICmpInst & Cmp,Instruction * Shl,const APInt & C)2032 static Instruction *foldICmpShlOne(ICmpInst &Cmp, Instruction *Shl,
2033 const APInt &C) {
2034 Value *Y;
2035 if (!match(Shl, m_Shl(m_One(), m_Value(Y))))
2036 return nullptr;
2037
2038 Type *ShiftType = Shl->getType();
2039 unsigned TypeBits = C.getBitWidth();
2040 bool CIsPowerOf2 = C.isPowerOf2();
2041 ICmpInst::Predicate Pred = Cmp.getPredicate();
2042 if (Cmp.isUnsigned()) {
2043 // (1 << Y) pred C -> Y pred Log2(C)
2044 if (!CIsPowerOf2) {
2045 // (1 << Y) < 30 -> Y <= 4
2046 // (1 << Y) <= 30 -> Y <= 4
2047 // (1 << Y) >= 30 -> Y > 4
2048 // (1 << Y) > 30 -> Y > 4
2049 if (Pred == ICmpInst::ICMP_ULT)
2050 Pred = ICmpInst::ICMP_ULE;
2051 else if (Pred == ICmpInst::ICMP_UGE)
2052 Pred = ICmpInst::ICMP_UGT;
2053 }
2054
2055 // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31
2056 // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31
2057 unsigned CLog2 = C.logBase2();
2058 if (CLog2 == TypeBits - 1) {
2059 if (Pred == ICmpInst::ICMP_UGE)
2060 Pred = ICmpInst::ICMP_EQ;
2061 else if (Pred == ICmpInst::ICMP_ULT)
2062 Pred = ICmpInst::ICMP_NE;
2063 }
2064 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2));
2065 } else if (Cmp.isSigned()) {
2066 Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1);
2067 if (C.isAllOnes()) {
2068 // (1 << Y) <= -1 -> Y == 31
2069 if (Pred == ICmpInst::ICMP_SLE)
2070 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2071
2072 // (1 << Y) > -1 -> Y != 31
2073 if (Pred == ICmpInst::ICMP_SGT)
2074 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2075 } else if (!C) {
2076 // (1 << Y) < 0 -> Y == 31
2077 // (1 << Y) <= 0 -> Y == 31
2078 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
2079 return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne);
2080
2081 // (1 << Y) >= 0 -> Y != 31
2082 // (1 << Y) > 0 -> Y != 31
2083 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
2084 return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne);
2085 }
2086 } else if (Cmp.isEquality() && CIsPowerOf2) {
2087 return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C.logBase2()));
2088 }
2089
2090 return nullptr;
2091 }
2092
2093 /// Fold icmp (shl X, Y), C.
foldICmpShlConstant(ICmpInst & Cmp,BinaryOperator * Shl,const APInt & C)2094 Instruction *InstCombinerImpl::foldICmpShlConstant(ICmpInst &Cmp,
2095 BinaryOperator *Shl,
2096 const APInt &C) {
2097 const APInt *ShiftVal;
2098 if (Cmp.isEquality() && match(Shl->getOperand(0), m_APInt(ShiftVal)))
2099 return foldICmpShlConstConst(Cmp, Shl->getOperand(1), C, *ShiftVal);
2100
2101 const APInt *ShiftAmt;
2102 if (!match(Shl->getOperand(1), m_APInt(ShiftAmt)))
2103 return foldICmpShlOne(Cmp, Shl, C);
2104
2105 // Check that the shift amount is in range. If not, don't perform undefined
2106 // shifts. When the shift is visited, it will be simplified.
2107 unsigned TypeBits = C.getBitWidth();
2108 if (ShiftAmt->uge(TypeBits))
2109 return nullptr;
2110
2111 ICmpInst::Predicate Pred = Cmp.getPredicate();
2112 Value *X = Shl->getOperand(0);
2113 Type *ShType = Shl->getType();
2114
2115 // NSW guarantees that we are only shifting out sign bits from the high bits,
2116 // so we can ASHR the compare constant without needing a mask and eliminate
2117 // the shift.
2118 if (Shl->hasNoSignedWrap()) {
2119 if (Pred == ICmpInst::ICMP_SGT) {
2120 // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt)
2121 APInt ShiftedC = C.ashr(*ShiftAmt);
2122 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2123 }
2124 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2125 C.ashr(*ShiftAmt).shl(*ShiftAmt) == C) {
2126 APInt ShiftedC = C.ashr(*ShiftAmt);
2127 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2128 }
2129 if (Pred == ICmpInst::ICMP_SLT) {
2130 // SLE is the same as above, but SLE is canonicalized to SLT, so convert:
2131 // (X << S) <=s C is equiv to X <=s (C >> S) for all C
2132 // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX
2133 // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN
2134 assert(!C.isMinSignedValue() && "Unexpected icmp slt");
2135 APInt ShiftedC = (C - 1).ashr(*ShiftAmt) + 1;
2136 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2137 }
2138 // If this is a signed comparison to 0 and the shift is sign preserving,
2139 // use the shift LHS operand instead; isSignTest may change 'Pred', so only
2140 // do that if we're sure to not continue on in this function.
2141 if (isSignTest(Pred, C))
2142 return new ICmpInst(Pred, X, Constant::getNullValue(ShType));
2143 }
2144
2145 // NUW guarantees that we are only shifting out zero bits from the high bits,
2146 // so we can LSHR the compare constant without needing a mask and eliminate
2147 // the shift.
2148 if (Shl->hasNoUnsignedWrap()) {
2149 if (Pred == ICmpInst::ICMP_UGT) {
2150 // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt)
2151 APInt ShiftedC = C.lshr(*ShiftAmt);
2152 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2153 }
2154 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2155 C.lshr(*ShiftAmt).shl(*ShiftAmt) == C) {
2156 APInt ShiftedC = C.lshr(*ShiftAmt);
2157 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2158 }
2159 if (Pred == ICmpInst::ICMP_ULT) {
2160 // ULE is the same as above, but ULE is canonicalized to ULT, so convert:
2161 // (X << S) <=u C is equiv to X <=u (C >> S) for all C
2162 // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u
2163 // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0
2164 assert(C.ugt(0) && "ult 0 should have been eliminated");
2165 APInt ShiftedC = (C - 1).lshr(*ShiftAmt) + 1;
2166 return new ICmpInst(Pred, X, ConstantInt::get(ShType, ShiftedC));
2167 }
2168 }
2169
2170 if (Cmp.isEquality() && Shl->hasOneUse()) {
2171 // Strength-reduce the shift into an 'and'.
2172 Constant *Mask = ConstantInt::get(
2173 ShType,
2174 APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue()));
2175 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2176 Constant *LShrC = ConstantInt::get(ShType, C.lshr(*ShiftAmt));
2177 return new ICmpInst(Pred, And, LShrC);
2178 }
2179
2180 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
2181 bool TrueIfSigned = false;
2182 if (Shl->hasOneUse() && isSignBitCheck(Pred, C, TrueIfSigned)) {
2183 // (X << 31) <s 0 --> (X & 1) != 0
2184 Constant *Mask = ConstantInt::get(
2185 ShType,
2186 APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1));
2187 Value *And = Builder.CreateAnd(X, Mask, Shl->getName() + ".mask");
2188 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
2189 And, Constant::getNullValue(ShType));
2190 }
2191
2192 // Simplify 'shl' inequality test into 'and' equality test.
2193 if (Cmp.isUnsigned() && Shl->hasOneUse()) {
2194 // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0
2195 if ((C + 1).isPowerOf2() &&
2196 (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) {
2197 Value *And = Builder.CreateAnd(X, (~C).lshr(ShiftAmt->getZExtValue()));
2198 return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ
2199 : ICmpInst::ICMP_NE,
2200 And, Constant::getNullValue(ShType));
2201 }
2202 // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0
2203 if (C.isPowerOf2() &&
2204 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) {
2205 Value *And =
2206 Builder.CreateAnd(X, (~(C - 1)).lshr(ShiftAmt->getZExtValue()));
2207 return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ
2208 : ICmpInst::ICMP_NE,
2209 And, Constant::getNullValue(ShType));
2210 }
2211 }
2212
2213 // Transform (icmp pred iM (shl iM %v, N), C)
2214 // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N))
2215 // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N.
2216 // This enables us to get rid of the shift in favor of a trunc that may be
2217 // free on the target. It has the additional benefit of comparing to a
2218 // smaller constant that may be more target-friendly.
2219 unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
2220 if (Shl->hasOneUse() && Amt != 0 && C.countTrailingZeros() >= Amt &&
2221 DL.isLegalInteger(TypeBits - Amt)) {
2222 Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
2223 if (auto *ShVTy = dyn_cast<VectorType>(ShType))
2224 TruncTy = VectorType::get(TruncTy, ShVTy->getElementCount());
2225 Constant *NewC =
2226 ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
2227 return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
2228 }
2229
2230 return nullptr;
2231 }
2232
2233 /// Fold icmp ({al}shr X, Y), C.
foldICmpShrConstant(ICmpInst & Cmp,BinaryOperator * Shr,const APInt & C)2234 Instruction *InstCombinerImpl::foldICmpShrConstant(ICmpInst &Cmp,
2235 BinaryOperator *Shr,
2236 const APInt &C) {
2237 // An exact shr only shifts out zero bits, so:
2238 // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0
2239 Value *X = Shr->getOperand(0);
2240 CmpInst::Predicate Pred = Cmp.getPredicate();
2241 if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() && C.isZero())
2242 return new ICmpInst(Pred, X, Cmp.getOperand(1));
2243
2244 const APInt *ShiftVal;
2245 if (Cmp.isEquality() && match(Shr->getOperand(0), m_APInt(ShiftVal)))
2246 return foldICmpShrConstConst(Cmp, Shr->getOperand(1), C, *ShiftVal);
2247
2248 const APInt *ShiftAmt;
2249 if (!match(Shr->getOperand(1), m_APInt(ShiftAmt)))
2250 return nullptr;
2251
2252 // Check that the shift amount is in range. If not, don't perform undefined
2253 // shifts. When the shift is visited it will be simplified.
2254 unsigned TypeBits = C.getBitWidth();
2255 unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits);
2256 if (ShAmtVal >= TypeBits || ShAmtVal == 0)
2257 return nullptr;
2258
2259 bool IsAShr = Shr->getOpcode() == Instruction::AShr;
2260 bool IsExact = Shr->isExact();
2261 Type *ShrTy = Shr->getType();
2262 // TODO: If we could guarantee that InstSimplify would handle all of the
2263 // constant-value-based preconditions in the folds below, then we could assert
2264 // those conditions rather than checking them. This is difficult because of
2265 // undef/poison (PR34838).
2266 if (IsAShr) {
2267 if (Pred == CmpInst::ICMP_SLT || (Pred == CmpInst::ICMP_SGT && IsExact)) {
2268 // icmp slt (ashr X, ShAmtC), C --> icmp slt X, (C << ShAmtC)
2269 // icmp sgt (ashr exact X, ShAmtC), C --> icmp sgt X, (C << ShAmtC)
2270 APInt ShiftedC = C.shl(ShAmtVal);
2271 if (ShiftedC.ashr(ShAmtVal) == C)
2272 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2273 }
2274 if (Pred == CmpInst::ICMP_SGT) {
2275 // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1
2276 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2277 if (!C.isMaxSignedValue() && !(C + 1).shl(ShAmtVal).isMinSignedValue() &&
2278 (ShiftedC + 1).ashr(ShAmtVal) == (C + 1))
2279 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2280 }
2281
2282 // If the compare constant has significant bits above the lowest sign-bit,
2283 // then convert an unsigned cmp to a test of the sign-bit:
2284 // (ashr X, ShiftC) u> C --> X s< 0
2285 // (ashr X, ShiftC) u< C --> X s> -1
2286 if (C.getBitWidth() > 2 && C.getNumSignBits() <= ShAmtVal) {
2287 if (Pred == CmpInst::ICMP_UGT) {
2288 return new ICmpInst(CmpInst::ICMP_SLT, X,
2289 ConstantInt::getNullValue(ShrTy));
2290 }
2291 if (Pred == CmpInst::ICMP_ULT) {
2292 return new ICmpInst(CmpInst::ICMP_SGT, X,
2293 ConstantInt::getAllOnesValue(ShrTy));
2294 }
2295 }
2296 } else {
2297 if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) {
2298 // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC)
2299 // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC)
2300 APInt ShiftedC = C.shl(ShAmtVal);
2301 if (ShiftedC.lshr(ShAmtVal) == C)
2302 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2303 }
2304 if (Pred == CmpInst::ICMP_UGT) {
2305 // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1
2306 APInt ShiftedC = (C + 1).shl(ShAmtVal) - 1;
2307 if ((ShiftedC + 1).lshr(ShAmtVal) == (C + 1))
2308 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, ShiftedC));
2309 }
2310 }
2311
2312 if (!Cmp.isEquality())
2313 return nullptr;
2314
2315 // Handle equality comparisons of shift-by-constant.
2316
2317 // If the comparison constant changes with the shift, the comparison cannot
2318 // succeed (bits of the comparison constant cannot match the shifted value).
2319 // This should be known by InstSimplify and already be folded to true/false.
2320 assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) ||
2321 (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) &&
2322 "Expected icmp+shr simplify did not occur.");
2323
2324 // If the bits shifted out are known zero, compare the unshifted value:
2325 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
2326 if (Shr->isExact())
2327 return new ICmpInst(Pred, X, ConstantInt::get(ShrTy, C << ShAmtVal));
2328
2329 if (C.isZero()) {
2330 // == 0 is u< 1.
2331 if (Pred == CmpInst::ICMP_EQ)
2332 return new ICmpInst(CmpInst::ICMP_ULT, X,
2333 ConstantInt::get(ShrTy, (C + 1).shl(ShAmtVal)));
2334 else
2335 return new ICmpInst(CmpInst::ICMP_UGT, X,
2336 ConstantInt::get(ShrTy, (C + 1).shl(ShAmtVal) - 1));
2337 }
2338
2339 if (Shr->hasOneUse()) {
2340 // Canonicalize the shift into an 'and':
2341 // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt)
2342 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
2343 Constant *Mask = ConstantInt::get(ShrTy, Val);
2344 Value *And = Builder.CreateAnd(X, Mask, Shr->getName() + ".mask");
2345 return new ICmpInst(Pred, And, ConstantInt::get(ShrTy, C << ShAmtVal));
2346 }
2347
2348 return nullptr;
2349 }
2350
foldICmpSRemConstant(ICmpInst & Cmp,BinaryOperator * SRem,const APInt & C)2351 Instruction *InstCombinerImpl::foldICmpSRemConstant(ICmpInst &Cmp,
2352 BinaryOperator *SRem,
2353 const APInt &C) {
2354 // Match an 'is positive' or 'is negative' comparison of remainder by a
2355 // constant power-of-2 value:
2356 // (X % pow2C) sgt/slt 0
2357 const ICmpInst::Predicate Pred = Cmp.getPredicate();
2358 if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT)
2359 return nullptr;
2360
2361 // TODO: The one-use check is standard because we do not typically want to
2362 // create longer instruction sequences, but this might be a special-case
2363 // because srem is not good for analysis or codegen.
2364 if (!SRem->hasOneUse())
2365 return nullptr;
2366
2367 const APInt *DivisorC;
2368 if (!C.isZero() || !match(SRem->getOperand(1), m_Power2(DivisorC)))
2369 return nullptr;
2370
2371 // Mask off the sign bit and the modulo bits (low-bits).
2372 Type *Ty = SRem->getType();
2373 APInt SignMask = APInt::getSignMask(Ty->getScalarSizeInBits());
2374 Constant *MaskC = ConstantInt::get(Ty, SignMask | (*DivisorC - 1));
2375 Value *And = Builder.CreateAnd(SRem->getOperand(0), MaskC);
2376
2377 // For 'is positive?' check that the sign-bit is clear and at least 1 masked
2378 // bit is set. Example:
2379 // (i8 X % 32) s> 0 --> (X & 159) s> 0
2380 if (Pred == ICmpInst::ICMP_SGT)
2381 return new ICmpInst(ICmpInst::ICMP_SGT, And, ConstantInt::getNullValue(Ty));
2382
2383 // For 'is negative?' check that the sign-bit is set and at least 1 masked
2384 // bit is set. Example:
2385 // (i16 X % 4) s< 0 --> (X & 32771) u> 32768
2386 return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, SignMask));
2387 }
2388
2389 /// Fold icmp (udiv X, Y), C.
foldICmpUDivConstant(ICmpInst & Cmp,BinaryOperator * UDiv,const APInt & C)2390 Instruction *InstCombinerImpl::foldICmpUDivConstant(ICmpInst &Cmp,
2391 BinaryOperator *UDiv,
2392 const APInt &C) {
2393 const APInt *C2;
2394 if (!match(UDiv->getOperand(0), m_APInt(C2)))
2395 return nullptr;
2396
2397 assert(*C2 != 0 && "udiv 0, X should have been simplified already.");
2398
2399 // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1))
2400 Value *Y = UDiv->getOperand(1);
2401 if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) {
2402 assert(!C.isMaxValue() &&
2403 "icmp ugt X, UINT_MAX should have been simplified already.");
2404 return new ICmpInst(ICmpInst::ICMP_ULE, Y,
2405 ConstantInt::get(Y->getType(), C2->udiv(C + 1)));
2406 }
2407
2408 // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C)
2409 if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) {
2410 assert(C != 0 && "icmp ult X, 0 should have been simplified already.");
2411 return new ICmpInst(ICmpInst::ICMP_UGT, Y,
2412 ConstantInt::get(Y->getType(), C2->udiv(C)));
2413 }
2414
2415 return nullptr;
2416 }
2417
2418 /// Fold icmp ({su}div X, Y), C.
foldICmpDivConstant(ICmpInst & Cmp,BinaryOperator * Div,const APInt & C)2419 Instruction *InstCombinerImpl::foldICmpDivConstant(ICmpInst &Cmp,
2420 BinaryOperator *Div,
2421 const APInt &C) {
2422 // Fold: icmp pred ([us]div X, C2), C -> range test
2423 // Fold this div into the comparison, producing a range check.
2424 // Determine, based on the divide type, what the range is being
2425 // checked. If there is an overflow on the low or high side, remember
2426 // it, otherwise compute the range [low, hi) bounding the new value.
2427 // See: InsertRangeTest above for the kinds of replacements possible.
2428 const APInt *C2;
2429 if (!match(Div->getOperand(1), m_APInt(C2)))
2430 return nullptr;
2431
2432 // FIXME: If the operand types don't match the type of the divide
2433 // then don't attempt this transform. The code below doesn't have the
2434 // logic to deal with a signed divide and an unsigned compare (and
2435 // vice versa). This is because (x /s C2) <s C produces different
2436 // results than (x /s C2) <u C or (x /u C2) <s C or even
2437 // (x /u C2) <u C. Simply casting the operands and result won't
2438 // work. :( The if statement below tests that condition and bails
2439 // if it finds it.
2440 bool DivIsSigned = Div->getOpcode() == Instruction::SDiv;
2441 if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned())
2442 return nullptr;
2443
2444 // The ProdOV computation fails on divide by 0 and divide by -1. Cases with
2445 // INT_MIN will also fail if the divisor is 1. Although folds of all these
2446 // division-by-constant cases should be present, we can not assert that they
2447 // have happened before we reach this icmp instruction.
2448 if (C2->isZero() || C2->isOne() || (DivIsSigned && C2->isAllOnes()))
2449 return nullptr;
2450
2451 // Compute Prod = C * C2. We are essentially solving an equation of
2452 // form X / C2 = C. We solve for X by multiplying C2 and C.
2453 // By solving for X, we can turn this into a range check instead of computing
2454 // a divide.
2455 APInt Prod = C * *C2;
2456
2457 // Determine if the product overflows by seeing if the product is not equal to
2458 // the divide. Make sure we do the same kind of divide as in the LHS
2459 // instruction that we're folding.
2460 bool ProdOV = (DivIsSigned ? Prod.sdiv(*C2) : Prod.udiv(*C2)) != C;
2461
2462 ICmpInst::Predicate Pred = Cmp.getPredicate();
2463
2464 // If the division is known to be exact, then there is no remainder from the
2465 // divide, so the covered range size is unit, otherwise it is the divisor.
2466 APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2;
2467
2468 // Figure out the interval that is being checked. For example, a comparison
2469 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
2470 // Compute this interval based on the constants involved and the signedness of
2471 // the compare/divide. This computes a half-open interval, keeping track of
2472 // whether either value in the interval overflows. After analysis each
2473 // overflow variable is set to 0 if it's corresponding bound variable is valid
2474 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
2475 int LoOverflow = 0, HiOverflow = 0;
2476 APInt LoBound, HiBound;
2477
2478 if (!DivIsSigned) { // udiv
2479 // e.g. X/5 op 3 --> [15, 20)
2480 LoBound = Prod;
2481 HiOverflow = LoOverflow = ProdOV;
2482 if (!HiOverflow) {
2483 // If this is not an exact divide, then many values in the range collapse
2484 // to the same result value.
2485 HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false);
2486 }
2487 } else if (C2->isStrictlyPositive()) { // Divisor is > 0.
2488 if (C.isZero()) { // (X / pos) op 0
2489 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
2490 LoBound = -(RangeSize - 1);
2491 HiBound = RangeSize;
2492 } else if (C.isStrictlyPositive()) { // (X / pos) op pos
2493 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
2494 HiOverflow = LoOverflow = ProdOV;
2495 if (!HiOverflow)
2496 HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true);
2497 } else { // (X / pos) op neg
2498 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
2499 HiBound = Prod + 1;
2500 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
2501 if (!LoOverflow) {
2502 APInt DivNeg = -RangeSize;
2503 LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
2504 }
2505 }
2506 } else if (C2->isNegative()) { // Divisor is < 0.
2507 if (Div->isExact())
2508 RangeSize.negate();
2509 if (C.isZero()) { // (X / neg) op 0
2510 // e.g. X/-5 op 0 --> [-4, 5)
2511 LoBound = RangeSize + 1;
2512 HiBound = -RangeSize;
2513 if (HiBound == *C2) { // -INTMIN = INTMIN
2514 HiOverflow = 1; // [INTMIN+1, overflow)
2515 HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN
2516 }
2517 } else if (C.isStrictlyPositive()) { // (X / neg) op pos
2518 // e.g. X/-5 op 3 --> [-19, -14)
2519 HiBound = Prod + 1;
2520 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
2521 if (!LoOverflow)
2522 LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
2523 } else { // (X / neg) op neg
2524 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
2525 LoOverflow = HiOverflow = ProdOV;
2526 if (!HiOverflow)
2527 HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true);
2528 }
2529
2530 // Dividing by a negative swaps the condition. LT <-> GT
2531 Pred = ICmpInst::getSwappedPredicate(Pred);
2532 }
2533
2534 Value *X = Div->getOperand(0);
2535 switch (Pred) {
2536 default: llvm_unreachable("Unhandled icmp opcode!");
2537 case ICmpInst::ICMP_EQ:
2538 if (LoOverflow && HiOverflow)
2539 return replaceInstUsesWith(Cmp, Builder.getFalse());
2540 if (HiOverflow)
2541 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2542 ICmpInst::ICMP_UGE, X,
2543 ConstantInt::get(Div->getType(), LoBound));
2544 if (LoOverflow)
2545 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2546 ICmpInst::ICMP_ULT, X,
2547 ConstantInt::get(Div->getType(), HiBound));
2548 return replaceInstUsesWith(
2549 Cmp, insertRangeTest(X, LoBound, HiBound, DivIsSigned, true));
2550 case ICmpInst::ICMP_NE:
2551 if (LoOverflow && HiOverflow)
2552 return replaceInstUsesWith(Cmp, Builder.getTrue());
2553 if (HiOverflow)
2554 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
2555 ICmpInst::ICMP_ULT, X,
2556 ConstantInt::get(Div->getType(), LoBound));
2557 if (LoOverflow)
2558 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
2559 ICmpInst::ICMP_UGE, X,
2560 ConstantInt::get(Div->getType(), HiBound));
2561 return replaceInstUsesWith(Cmp,
2562 insertRangeTest(X, LoBound, HiBound,
2563 DivIsSigned, false));
2564 case ICmpInst::ICMP_ULT:
2565 case ICmpInst::ICMP_SLT:
2566 if (LoOverflow == +1) // Low bound is greater than input range.
2567 return replaceInstUsesWith(Cmp, Builder.getTrue());
2568 if (LoOverflow == -1) // Low bound is less than input range.
2569 return replaceInstUsesWith(Cmp, Builder.getFalse());
2570 return new ICmpInst(Pred, X, ConstantInt::get(Div->getType(), LoBound));
2571 case ICmpInst::ICMP_UGT:
2572 case ICmpInst::ICMP_SGT:
2573 if (HiOverflow == +1) // High bound greater than input range.
2574 return replaceInstUsesWith(Cmp, Builder.getFalse());
2575 if (HiOverflow == -1) // High bound less than input range.
2576 return replaceInstUsesWith(Cmp, Builder.getTrue());
2577 if (Pred == ICmpInst::ICMP_UGT)
2578 return new ICmpInst(ICmpInst::ICMP_UGE, X,
2579 ConstantInt::get(Div->getType(), HiBound));
2580 return new ICmpInst(ICmpInst::ICMP_SGE, X,
2581 ConstantInt::get(Div->getType(), HiBound));
2582 }
2583
2584 return nullptr;
2585 }
2586
2587 /// Fold icmp (sub X, Y), C.
foldICmpSubConstant(ICmpInst & Cmp,BinaryOperator * Sub,const APInt & C)2588 Instruction *InstCombinerImpl::foldICmpSubConstant(ICmpInst &Cmp,
2589 BinaryOperator *Sub,
2590 const APInt &C) {
2591 Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1);
2592 ICmpInst::Predicate Pred = Cmp.getPredicate();
2593 Type *Ty = Sub->getType();
2594
2595 // (SubC - Y) == C) --> Y == (SubC - C)
2596 // (SubC - Y) != C) --> Y != (SubC - C)
2597 Constant *SubC;
2598 if (Cmp.isEquality() && match(X, m_ImmConstant(SubC))) {
2599 return new ICmpInst(Pred, Y,
2600 ConstantExpr::getSub(SubC, ConstantInt::get(Ty, C)));
2601 }
2602
2603 // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C)
2604 const APInt *C2;
2605 APInt SubResult;
2606 ICmpInst::Predicate SwappedPred = Cmp.getSwappedPredicate();
2607 bool HasNSW = Sub->hasNoSignedWrap();
2608 bool HasNUW = Sub->hasNoUnsignedWrap();
2609 if (match(X, m_APInt(C2)) &&
2610 ((Cmp.isUnsigned() && HasNUW) || (Cmp.isSigned() && HasNSW)) &&
2611 !subWithOverflow(SubResult, *C2, C, Cmp.isSigned()))
2612 return new ICmpInst(SwappedPred, Y, ConstantInt::get(Ty, SubResult));
2613
2614 // The following transforms are only worth it if the only user of the subtract
2615 // is the icmp.
2616 // TODO: This is an artificial restriction for all of the transforms below
2617 // that only need a single replacement icmp.
2618 if (!Sub->hasOneUse())
2619 return nullptr;
2620
2621 // X - Y == 0 --> X == Y.
2622 // X - Y != 0 --> X != Y.
2623 if (Cmp.isEquality() && C.isZero())
2624 return new ICmpInst(Pred, X, Y);
2625
2626 if (Sub->hasNoSignedWrap()) {
2627 // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y)
2628 if (Pred == ICmpInst::ICMP_SGT && C.isAllOnes())
2629 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
2630
2631 // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y)
2632 if (Pred == ICmpInst::ICMP_SGT && C.isZero())
2633 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
2634
2635 // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y)
2636 if (Pred == ICmpInst::ICMP_SLT && C.isZero())
2637 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
2638
2639 // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y)
2640 if (Pred == ICmpInst::ICMP_SLT && C.isOne())
2641 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
2642 }
2643
2644 if (!match(X, m_APInt(C2)))
2645 return nullptr;
2646
2647 // C2 - Y <u C -> (Y | (C - 1)) == C2
2648 // iff (C2 & (C - 1)) == C - 1 and C is a power of 2
2649 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() &&
2650 (*C2 & (C - 1)) == (C - 1))
2651 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(Y, C - 1), X);
2652
2653 // C2 - Y >u C -> (Y | C) != C2
2654 // iff C2 & C == C and C + 1 is a power of 2
2655 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C)
2656 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(Y, C), X);
2657
2658 // We have handled special cases that reduce.
2659 // Canonicalize any remaining sub to add as:
2660 // (C2 - Y) > C --> (Y + ~C2) < ~C
2661 Value *Add = Builder.CreateAdd(Y, ConstantInt::get(Ty, ~(*C2)), "notsub",
2662 HasNUW, HasNSW);
2663 return new ICmpInst(SwappedPred, Add, ConstantInt::get(Ty, ~C));
2664 }
2665
2666 /// Fold icmp (add X, Y), C.
foldICmpAddConstant(ICmpInst & Cmp,BinaryOperator * Add,const APInt & C)2667 Instruction *InstCombinerImpl::foldICmpAddConstant(ICmpInst &Cmp,
2668 BinaryOperator *Add,
2669 const APInt &C) {
2670 Value *Y = Add->getOperand(1);
2671 const APInt *C2;
2672 if (Cmp.isEquality() || !match(Y, m_APInt(C2)))
2673 return nullptr;
2674
2675 // Fold icmp pred (add X, C2), C.
2676 Value *X = Add->getOperand(0);
2677 Type *Ty = Add->getType();
2678 const CmpInst::Predicate Pred = Cmp.getPredicate();
2679
2680 // If the add does not wrap, we can always adjust the compare by subtracting
2681 // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE
2682 // are canonicalized to SGT/SLT/UGT/ULT.
2683 if ((Add->hasNoSignedWrap() &&
2684 (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) ||
2685 (Add->hasNoUnsignedWrap() &&
2686 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) {
2687 bool Overflow;
2688 APInt NewC =
2689 Cmp.isSigned() ? C.ssub_ov(*C2, Overflow) : C.usub_ov(*C2, Overflow);
2690 // If there is overflow, the result must be true or false.
2691 // TODO: Can we assert there is no overflow because InstSimplify always
2692 // handles those cases?
2693 if (!Overflow)
2694 // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2)
2695 return new ICmpInst(Pred, X, ConstantInt::get(Ty, NewC));
2696 }
2697
2698 auto CR = ConstantRange::makeExactICmpRegion(Pred, C).subtract(*C2);
2699 const APInt &Upper = CR.getUpper();
2700 const APInt &Lower = CR.getLower();
2701 if (Cmp.isSigned()) {
2702 if (Lower.isSignMask())
2703 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper));
2704 if (Upper.isSignMask())
2705 return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower));
2706 } else {
2707 if (Lower.isMinValue())
2708 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper));
2709 if (Upper.isMinValue())
2710 return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower));
2711 }
2712
2713 // This set of folds is intentionally placed after folds that use no-wrapping
2714 // flags because those folds are likely better for later analysis/codegen.
2715 const APInt SMax = APInt::getSignedMaxValue(Ty->getScalarSizeInBits());
2716 const APInt SMin = APInt::getSignedMinValue(Ty->getScalarSizeInBits());
2717
2718 // Fold compare with offset to opposite sign compare if it eliminates offset:
2719 // (X + C2) >u C --> X <s -C2 (if C == C2 + SMAX)
2720 if (Pred == CmpInst::ICMP_UGT && C == *C2 + SMax)
2721 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, -(*C2)));
2722
2723 // (X + C2) <u C --> X >s ~C2 (if C == C2 + SMIN)
2724 if (Pred == CmpInst::ICMP_ULT && C == *C2 + SMin)
2725 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantInt::get(Ty, ~(*C2)));
2726
2727 // (X + C2) >s C --> X <u (SMAX - C) (if C == C2 - 1)
2728 if (Pred == CmpInst::ICMP_SGT && C == *C2 - 1)
2729 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, SMax - C));
2730
2731 // (X + C2) <s C --> X >u (C ^ SMAX) (if C == C2)
2732 if (Pred == CmpInst::ICMP_SLT && C == *C2)
2733 return new ICmpInst(ICmpInst::ICMP_UGT, X, ConstantInt::get(Ty, C ^ SMax));
2734
2735 if (!Add->hasOneUse())
2736 return nullptr;
2737
2738 // X+C <u C2 -> (X & -C2) == C
2739 // iff C & (C2-1) == 0
2740 // C2 is a power of 2
2741 if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0)
2742 return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(X, -C),
2743 ConstantExpr::getNeg(cast<Constant>(Y)));
2744
2745 // X+C >u C2 -> (X & ~C2) != C
2746 // iff C & C2 == 0
2747 // C2+1 is a power of 2
2748 if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0)
2749 return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(X, ~C),
2750 ConstantExpr::getNeg(cast<Constant>(Y)));
2751
2752 return nullptr;
2753 }
2754
matchThreeWayIntCompare(SelectInst * SI,Value * & LHS,Value * & RHS,ConstantInt * & Less,ConstantInt * & Equal,ConstantInt * & Greater)2755 bool InstCombinerImpl::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS,
2756 Value *&RHS, ConstantInt *&Less,
2757 ConstantInt *&Equal,
2758 ConstantInt *&Greater) {
2759 // TODO: Generalize this to work with other comparison idioms or ensure
2760 // they get canonicalized into this form.
2761
2762 // select i1 (a == b),
2763 // i32 Equal,
2764 // i32 (select i1 (a < b), i32 Less, i32 Greater)
2765 // where Equal, Less and Greater are placeholders for any three constants.
2766 ICmpInst::Predicate PredA;
2767 if (!match(SI->getCondition(), m_ICmp(PredA, m_Value(LHS), m_Value(RHS))) ||
2768 !ICmpInst::isEquality(PredA))
2769 return false;
2770 Value *EqualVal = SI->getTrueValue();
2771 Value *UnequalVal = SI->getFalseValue();
2772 // We still can get non-canonical predicate here, so canonicalize.
2773 if (PredA == ICmpInst::ICMP_NE)
2774 std::swap(EqualVal, UnequalVal);
2775 if (!match(EqualVal, m_ConstantInt(Equal)))
2776 return false;
2777 ICmpInst::Predicate PredB;
2778 Value *LHS2, *RHS2;
2779 if (!match(UnequalVal, m_Select(m_ICmp(PredB, m_Value(LHS2), m_Value(RHS2)),
2780 m_ConstantInt(Less), m_ConstantInt(Greater))))
2781 return false;
2782 // We can get predicate mismatch here, so canonicalize if possible:
2783 // First, ensure that 'LHS' match.
2784 if (LHS2 != LHS) {
2785 // x sgt y <--> y slt x
2786 std::swap(LHS2, RHS2);
2787 PredB = ICmpInst::getSwappedPredicate(PredB);
2788 }
2789 if (LHS2 != LHS)
2790 return false;
2791 // We also need to canonicalize 'RHS'.
2792 if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(RHS2)) {
2793 // x sgt C-1 <--> x sge C <--> not(x slt C)
2794 auto FlippedStrictness =
2795 InstCombiner::getFlippedStrictnessPredicateAndConstant(
2796 PredB, cast<Constant>(RHS2));
2797 if (!FlippedStrictness)
2798 return false;
2799 assert(FlippedStrictness->first == ICmpInst::ICMP_SGE && "Sanity check");
2800 RHS2 = FlippedStrictness->second;
2801 // And kind-of perform the result swap.
2802 std::swap(Less, Greater);
2803 PredB = ICmpInst::ICMP_SLT;
2804 }
2805 return PredB == ICmpInst::ICMP_SLT && RHS == RHS2;
2806 }
2807
foldICmpSelectConstant(ICmpInst & Cmp,SelectInst * Select,ConstantInt * C)2808 Instruction *InstCombinerImpl::foldICmpSelectConstant(ICmpInst &Cmp,
2809 SelectInst *Select,
2810 ConstantInt *C) {
2811
2812 assert(C && "Cmp RHS should be a constant int!");
2813 // If we're testing a constant value against the result of a three way
2814 // comparison, the result can be expressed directly in terms of the
2815 // original values being compared. Note: We could possibly be more
2816 // aggressive here and remove the hasOneUse test. The original select is
2817 // really likely to simplify or sink when we remove a test of the result.
2818 Value *OrigLHS, *OrigRHS;
2819 ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan;
2820 if (Cmp.hasOneUse() &&
2821 matchThreeWayIntCompare(Select, OrigLHS, OrigRHS, C1LessThan, C2Equal,
2822 C3GreaterThan)) {
2823 assert(C1LessThan && C2Equal && C3GreaterThan);
2824
2825 bool TrueWhenLessThan =
2826 ConstantExpr::getCompare(Cmp.getPredicate(), C1LessThan, C)
2827 ->isAllOnesValue();
2828 bool TrueWhenEqual =
2829 ConstantExpr::getCompare(Cmp.getPredicate(), C2Equal, C)
2830 ->isAllOnesValue();
2831 bool TrueWhenGreaterThan =
2832 ConstantExpr::getCompare(Cmp.getPredicate(), C3GreaterThan, C)
2833 ->isAllOnesValue();
2834
2835 // This generates the new instruction that will replace the original Cmp
2836 // Instruction. Instead of enumerating the various combinations when
2837 // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus
2838 // false, we rely on chaining of ORs and future passes of InstCombine to
2839 // simplify the OR further (i.e. a s< b || a == b becomes a s<= b).
2840
2841 // When none of the three constants satisfy the predicate for the RHS (C),
2842 // the entire original Cmp can be simplified to a false.
2843 Value *Cond = Builder.getFalse();
2844 if (TrueWhenLessThan)
2845 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SLT,
2846 OrigLHS, OrigRHS));
2847 if (TrueWhenEqual)
2848 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_EQ,
2849 OrigLHS, OrigRHS));
2850 if (TrueWhenGreaterThan)
2851 Cond = Builder.CreateOr(Cond, Builder.CreateICmp(ICmpInst::ICMP_SGT,
2852 OrigLHS, OrigRHS));
2853
2854 return replaceInstUsesWith(Cmp, Cond);
2855 }
2856 return nullptr;
2857 }
2858
foldICmpBitCast(ICmpInst & Cmp)2859 Instruction *InstCombinerImpl::foldICmpBitCast(ICmpInst &Cmp) {
2860 auto *Bitcast = dyn_cast<BitCastInst>(Cmp.getOperand(0));
2861 if (!Bitcast)
2862 return nullptr;
2863
2864 ICmpInst::Predicate Pred = Cmp.getPredicate();
2865 Value *Op1 = Cmp.getOperand(1);
2866 Value *BCSrcOp = Bitcast->getOperand(0);
2867
2868 // Make sure the bitcast doesn't change the number of vector elements.
2869 if (Bitcast->getSrcTy()->getScalarSizeInBits() ==
2870 Bitcast->getDestTy()->getScalarSizeInBits()) {
2871 // Zero-equality and sign-bit checks are preserved through sitofp + bitcast.
2872 Value *X;
2873 if (match(BCSrcOp, m_SIToFP(m_Value(X)))) {
2874 // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0
2875 // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0
2876 // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0
2877 // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0
2878 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT ||
2879 Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) &&
2880 match(Op1, m_Zero()))
2881 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2882
2883 // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1
2884 if (Pred == ICmpInst::ICMP_SLT && match(Op1, m_One()))
2885 return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), 1));
2886
2887 // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1
2888 if (Pred == ICmpInst::ICMP_SGT && match(Op1, m_AllOnes()))
2889 return new ICmpInst(Pred, X,
2890 ConstantInt::getAllOnesValue(X->getType()));
2891 }
2892
2893 // Zero-equality checks are preserved through unsigned floating-point casts:
2894 // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0
2895 // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0
2896 if (match(BCSrcOp, m_UIToFP(m_Value(X))))
2897 if (Cmp.isEquality() && match(Op1, m_Zero()))
2898 return new ICmpInst(Pred, X, ConstantInt::getNullValue(X->getType()));
2899
2900 // If this is a sign-bit test of a bitcast of a casted FP value, eliminate
2901 // the FP extend/truncate because that cast does not change the sign-bit.
2902 // This is true for all standard IEEE-754 types and the X86 80-bit type.
2903 // The sign-bit is always the most significant bit in those types.
2904 const APInt *C;
2905 bool TrueIfSigned;
2906 if (match(Op1, m_APInt(C)) && Bitcast->hasOneUse() &&
2907 InstCombiner::isSignBitCheck(Pred, *C, TrueIfSigned)) {
2908 if (match(BCSrcOp, m_FPExt(m_Value(X))) ||
2909 match(BCSrcOp, m_FPTrunc(m_Value(X)))) {
2910 // (bitcast (fpext/fptrunc X)) to iX) < 0 --> (bitcast X to iY) < 0
2911 // (bitcast (fpext/fptrunc X)) to iX) > -1 --> (bitcast X to iY) > -1
2912 Type *XType = X->getType();
2913
2914 // We can't currently handle Power style floating point operations here.
2915 if (!(XType->isPPC_FP128Ty() || BCSrcOp->getType()->isPPC_FP128Ty())) {
2916
2917 Type *NewType = Builder.getIntNTy(XType->getScalarSizeInBits());
2918 if (auto *XVTy = dyn_cast<VectorType>(XType))
2919 NewType = VectorType::get(NewType, XVTy->getElementCount());
2920 Value *NewBitcast = Builder.CreateBitCast(X, NewType);
2921 if (TrueIfSigned)
2922 return new ICmpInst(ICmpInst::ICMP_SLT, NewBitcast,
2923 ConstantInt::getNullValue(NewType));
2924 else
2925 return new ICmpInst(ICmpInst::ICMP_SGT, NewBitcast,
2926 ConstantInt::getAllOnesValue(NewType));
2927 }
2928 }
2929 }
2930 }
2931
2932 // Test to see if the operands of the icmp are casted versions of other
2933 // values. If the ptr->ptr cast can be stripped off both arguments, do so.
2934 if (Bitcast->getType()->isPointerTy() &&
2935 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2936 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2937 // so eliminate it as well.
2938 if (auto *BC2 = dyn_cast<BitCastInst>(Op1))
2939 Op1 = BC2->getOperand(0);
2940
2941 Op1 = Builder.CreateBitCast(Op1, BCSrcOp->getType());
2942 return new ICmpInst(Pred, BCSrcOp, Op1);
2943 }
2944
2945 const APInt *C;
2946 if (!match(Cmp.getOperand(1), m_APInt(C)) ||
2947 !Bitcast->getType()->isIntegerTy() ||
2948 !Bitcast->getSrcTy()->isIntOrIntVectorTy())
2949 return nullptr;
2950
2951 // If this is checking if all elements of a vector compare are set or not,
2952 // invert the casted vector equality compare and test if all compare
2953 // elements are clear or not. Compare against zero is generally easier for
2954 // analysis and codegen.
2955 // icmp eq/ne (bitcast (not X) to iN), -1 --> icmp eq/ne (bitcast X to iN), 0
2956 // Example: are all elements equal? --> are zero elements not equal?
2957 // TODO: Try harder to reduce compare of 2 freely invertible operands?
2958 if (Cmp.isEquality() && C->isAllOnes() && Bitcast->hasOneUse() &&
2959 isFreeToInvert(BCSrcOp, BCSrcOp->hasOneUse())) {
2960 Type *ScalarTy = Bitcast->getType();
2961 Value *Cast = Builder.CreateBitCast(Builder.CreateNot(BCSrcOp), ScalarTy);
2962 return new ICmpInst(Pred, Cast, ConstantInt::getNullValue(ScalarTy));
2963 }
2964
2965 // If this is checking if all elements of an extended vector are clear or not,
2966 // compare in a narrow type to eliminate the extend:
2967 // icmp eq/ne (bitcast (ext X) to iN), 0 --> icmp eq/ne (bitcast X to iM), 0
2968 Value *X;
2969 if (Cmp.isEquality() && C->isZero() && Bitcast->hasOneUse() &&
2970 match(BCSrcOp, m_ZExtOrSExt(m_Value(X)))) {
2971 if (auto *VecTy = dyn_cast<FixedVectorType>(X->getType())) {
2972 Type *NewType = Builder.getIntNTy(VecTy->getPrimitiveSizeInBits());
2973 Value *NewCast = Builder.CreateBitCast(X, NewType);
2974 return new ICmpInst(Pred, NewCast, ConstantInt::getNullValue(NewType));
2975 }
2976 }
2977
2978 // Folding: icmp <pred> iN X, C
2979 // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN
2980 // and C is a splat of a K-bit pattern
2981 // and SC is a constant vector = <C', C', C', ..., C'>
2982 // Into:
2983 // %E = extractelement <M x iK> %vec, i32 C'
2984 // icmp <pred> iK %E, trunc(C)
2985 Value *Vec;
2986 ArrayRef<int> Mask;
2987 if (match(BCSrcOp, m_Shuffle(m_Value(Vec), m_Undef(), m_Mask(Mask)))) {
2988 // Check whether every element of Mask is the same constant
2989 if (is_splat(Mask)) {
2990 auto *VecTy = cast<VectorType>(BCSrcOp->getType());
2991 auto *EltTy = cast<IntegerType>(VecTy->getElementType());
2992 if (C->isSplat(EltTy->getBitWidth())) {
2993 // Fold the icmp based on the value of C
2994 // If C is M copies of an iK sized bit pattern,
2995 // then:
2996 // => %E = extractelement <N x iK> %vec, i32 Elem
2997 // icmp <pred> iK %SplatVal, <pattern>
2998 Value *Elem = Builder.getInt32(Mask[0]);
2999 Value *Extract = Builder.CreateExtractElement(Vec, Elem);
3000 Value *NewC = ConstantInt::get(EltTy, C->trunc(EltTy->getBitWidth()));
3001 return new ICmpInst(Pred, Extract, NewC);
3002 }
3003 }
3004 }
3005 return nullptr;
3006 }
3007
3008 /// Try to fold integer comparisons with a constant operand: icmp Pred X, C
3009 /// where X is some kind of instruction.
foldICmpInstWithConstant(ICmpInst & Cmp)3010 Instruction *InstCombinerImpl::foldICmpInstWithConstant(ICmpInst &Cmp) {
3011 const APInt *C;
3012 if (!match(Cmp.getOperand(1), m_APInt(C)))
3013 return nullptr;
3014
3015 if (auto *BO = dyn_cast<BinaryOperator>(Cmp.getOperand(0))) {
3016 switch (BO->getOpcode()) {
3017 case Instruction::Xor:
3018 if (Instruction *I = foldICmpXorConstant(Cmp, BO, *C))
3019 return I;
3020 break;
3021 case Instruction::And:
3022 if (Instruction *I = foldICmpAndConstant(Cmp, BO, *C))
3023 return I;
3024 break;
3025 case Instruction::Or:
3026 if (Instruction *I = foldICmpOrConstant(Cmp, BO, *C))
3027 return I;
3028 break;
3029 case Instruction::Mul:
3030 if (Instruction *I = foldICmpMulConstant(Cmp, BO, *C))
3031 return I;
3032 break;
3033 case Instruction::Shl:
3034 if (Instruction *I = foldICmpShlConstant(Cmp, BO, *C))
3035 return I;
3036 break;
3037 case Instruction::LShr:
3038 case Instruction::AShr:
3039 if (Instruction *I = foldICmpShrConstant(Cmp, BO, *C))
3040 return I;
3041 break;
3042 case Instruction::SRem:
3043 if (Instruction *I = foldICmpSRemConstant(Cmp, BO, *C))
3044 return I;
3045 break;
3046 case Instruction::UDiv:
3047 if (Instruction *I = foldICmpUDivConstant(Cmp, BO, *C))
3048 return I;
3049 LLVM_FALLTHROUGH;
3050 case Instruction::SDiv:
3051 if (Instruction *I = foldICmpDivConstant(Cmp, BO, *C))
3052 return I;
3053 break;
3054 case Instruction::Sub:
3055 if (Instruction *I = foldICmpSubConstant(Cmp, BO, *C))
3056 return I;
3057 break;
3058 case Instruction::Add:
3059 if (Instruction *I = foldICmpAddConstant(Cmp, BO, *C))
3060 return I;
3061 break;
3062 default:
3063 break;
3064 }
3065 // TODO: These folds could be refactored to be part of the above calls.
3066 if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, *C))
3067 return I;
3068 }
3069
3070 // Match against CmpInst LHS being instructions other than binary operators.
3071
3072 if (auto *SI = dyn_cast<SelectInst>(Cmp.getOperand(0))) {
3073 // For now, we only support constant integers while folding the
3074 // ICMP(SELECT)) pattern. We can extend this to support vector of integers
3075 // similar to the cases handled by binary ops above.
3076 if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(Cmp.getOperand(1)))
3077 if (Instruction *I = foldICmpSelectConstant(Cmp, SI, ConstRHS))
3078 return I;
3079 }
3080
3081 if (auto *TI = dyn_cast<TruncInst>(Cmp.getOperand(0))) {
3082 if (Instruction *I = foldICmpTruncConstant(Cmp, TI, *C))
3083 return I;
3084 }
3085
3086 if (auto *II = dyn_cast<IntrinsicInst>(Cmp.getOperand(0)))
3087 if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, II, *C))
3088 return I;
3089
3090 return nullptr;
3091 }
3092
3093 /// Fold an icmp equality instruction with binary operator LHS and constant RHS:
3094 /// icmp eq/ne BO, C.
foldICmpBinOpEqualityWithConstant(ICmpInst & Cmp,BinaryOperator * BO,const APInt & C)3095 Instruction *InstCombinerImpl::foldICmpBinOpEqualityWithConstant(
3096 ICmpInst &Cmp, BinaryOperator *BO, const APInt &C) {
3097 // TODO: Some of these folds could work with arbitrary constants, but this
3098 // function is limited to scalar and vector splat constants.
3099 if (!Cmp.isEquality())
3100 return nullptr;
3101
3102 ICmpInst::Predicate Pred = Cmp.getPredicate();
3103 bool isICMP_NE = Pred == ICmpInst::ICMP_NE;
3104 Constant *RHS = cast<Constant>(Cmp.getOperand(1));
3105 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
3106
3107 switch (BO->getOpcode()) {
3108 case Instruction::SRem:
3109 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
3110 if (C.isZero() && BO->hasOneUse()) {
3111 const APInt *BOC;
3112 if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) {
3113 Value *NewRem = Builder.CreateURem(BOp0, BOp1, BO->getName());
3114 return new ICmpInst(Pred, NewRem,
3115 Constant::getNullValue(BO->getType()));
3116 }
3117 }
3118 break;
3119 case Instruction::Add: {
3120 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
3121 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3122 if (BO->hasOneUse())
3123 return new ICmpInst(Pred, BOp0, ConstantExpr::getSub(RHS, BOC));
3124 } else if (C.isZero()) {
3125 // Replace ((add A, B) != 0) with (A != -B) if A or B is
3126 // efficiently invertible, or if the add has just this one use.
3127 if (Value *NegVal = dyn_castNegVal(BOp1))
3128 return new ICmpInst(Pred, BOp0, NegVal);
3129 if (Value *NegVal = dyn_castNegVal(BOp0))
3130 return new ICmpInst(Pred, NegVal, BOp1);
3131 if (BO->hasOneUse()) {
3132 Value *Neg = Builder.CreateNeg(BOp1);
3133 Neg->takeName(BO);
3134 return new ICmpInst(Pred, BOp0, Neg);
3135 }
3136 }
3137 break;
3138 }
3139 case Instruction::Xor:
3140 if (BO->hasOneUse()) {
3141 if (Constant *BOC = dyn_cast<Constant>(BOp1)) {
3142 // For the xor case, we can xor two constants together, eliminating
3143 // the explicit xor.
3144 return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC));
3145 } else if (C.isZero()) {
3146 // Replace ((xor A, B) != 0) with (A != B)
3147 return new ICmpInst(Pred, BOp0, BOp1);
3148 }
3149 }
3150 break;
3151 case Instruction::Or: {
3152 const APInt *BOC;
3153 if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) {
3154 // Comparing if all bits outside of a constant mask are set?
3155 // Replace (X | C) == -1 with (X & ~C) == ~C.
3156 // This removes the -1 constant.
3157 Constant *NotBOC = ConstantExpr::getNot(cast<Constant>(BOp1));
3158 Value *And = Builder.CreateAnd(BOp0, NotBOC);
3159 return new ICmpInst(Pred, And, NotBOC);
3160 }
3161 break;
3162 }
3163 case Instruction::And: {
3164 const APInt *BOC;
3165 if (match(BOp1, m_APInt(BOC))) {
3166 // If we have ((X & C) == C), turn it into ((X & C) != 0).
3167 if (C == *BOC && C.isPowerOf2())
3168 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
3169 BO, Constant::getNullValue(RHS->getType()));
3170 }
3171 break;
3172 }
3173 case Instruction::UDiv:
3174 if (C.isZero()) {
3175 // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A)
3176 auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3177 return new ICmpInst(NewPred, BOp1, BOp0);
3178 }
3179 break;
3180 default:
3181 break;
3182 }
3183 return nullptr;
3184 }
3185
3186 /// Fold an equality icmp with LLVM intrinsic and constant operand.
foldICmpEqIntrinsicWithConstant(ICmpInst & Cmp,IntrinsicInst * II,const APInt & C)3187 Instruction *InstCombinerImpl::foldICmpEqIntrinsicWithConstant(
3188 ICmpInst &Cmp, IntrinsicInst *II, const APInt &C) {
3189 Type *Ty = II->getType();
3190 unsigned BitWidth = C.getBitWidth();
3191 const ICmpInst::Predicate Pred = Cmp.getPredicate();
3192
3193 switch (II->getIntrinsicID()) {
3194 case Intrinsic::abs:
3195 // abs(A) == 0 -> A == 0
3196 // abs(A) == INT_MIN -> A == INT_MIN
3197 if (C.isZero() || C.isMinSignedValue())
3198 return new ICmpInst(Pred, II->getArgOperand(0), ConstantInt::get(Ty, C));
3199 break;
3200
3201 case Intrinsic::bswap:
3202 // bswap(A) == C -> A == bswap(C)
3203 return new ICmpInst(Pred, II->getArgOperand(0),
3204 ConstantInt::get(Ty, C.byteSwap()));
3205
3206 case Intrinsic::ctlz:
3207 case Intrinsic::cttz: {
3208 // ctz(A) == bitwidth(A) -> A == 0 and likewise for !=
3209 if (C == BitWidth)
3210 return new ICmpInst(Pred, II->getArgOperand(0),
3211 ConstantInt::getNullValue(Ty));
3212
3213 // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set
3214 // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits.
3215 // Limit to one use to ensure we don't increase instruction count.
3216 unsigned Num = C.getLimitedValue(BitWidth);
3217 if (Num != BitWidth && II->hasOneUse()) {
3218 bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz;
3219 APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(BitWidth, Num + 1)
3220 : APInt::getHighBitsSet(BitWidth, Num + 1);
3221 APInt Mask2 = IsTrailing
3222 ? APInt::getOneBitSet(BitWidth, Num)
3223 : APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3224 return new ICmpInst(Pred, Builder.CreateAnd(II->getArgOperand(0), Mask1),
3225 ConstantInt::get(Ty, Mask2));
3226 }
3227 break;
3228 }
3229
3230 case Intrinsic::ctpop: {
3231 // popcount(A) == 0 -> A == 0 and likewise for !=
3232 // popcount(A) == bitwidth(A) -> A == -1 and likewise for !=
3233 bool IsZero = C.isZero();
3234 if (IsZero || C == BitWidth)
3235 return new ICmpInst(Pred, II->getArgOperand(0),
3236 IsZero ? Constant::getNullValue(Ty)
3237 : Constant::getAllOnesValue(Ty));
3238
3239 break;
3240 }
3241
3242 case Intrinsic::fshl:
3243 case Intrinsic::fshr:
3244 if (II->getArgOperand(0) == II->getArgOperand(1)) {
3245 // (rot X, ?) == 0/-1 --> X == 0/-1
3246 // TODO: This transform is safe to re-use undef elts in a vector, but
3247 // the constant value passed in by the caller doesn't allow that.
3248 if (C.isZero() || C.isAllOnes())
3249 return new ICmpInst(Pred, II->getArgOperand(0), Cmp.getOperand(1));
3250
3251 const APInt *RotAmtC;
3252 // ror(X, RotAmtC) == C --> X == rol(C, RotAmtC)
3253 // rol(X, RotAmtC) == C --> X == ror(C, RotAmtC)
3254 if (match(II->getArgOperand(2), m_APInt(RotAmtC)))
3255 return new ICmpInst(Pred, II->getArgOperand(0),
3256 II->getIntrinsicID() == Intrinsic::fshl
3257 ? ConstantInt::get(Ty, C.rotr(*RotAmtC))
3258 : ConstantInt::get(Ty, C.rotl(*RotAmtC)));
3259 }
3260 break;
3261
3262 case Intrinsic::uadd_sat: {
3263 // uadd.sat(a, b) == 0 -> (a | b) == 0
3264 if (C.isZero()) {
3265 Value *Or = Builder.CreateOr(II->getArgOperand(0), II->getArgOperand(1));
3266 return new ICmpInst(Pred, Or, Constant::getNullValue(Ty));
3267 }
3268 break;
3269 }
3270
3271 case Intrinsic::usub_sat: {
3272 // usub.sat(a, b) == 0 -> a <= b
3273 if (C.isZero()) {
3274 ICmpInst::Predicate NewPred =
3275 Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT;
3276 return new ICmpInst(NewPred, II->getArgOperand(0), II->getArgOperand(1));
3277 }
3278 break;
3279 }
3280 default:
3281 break;
3282 }
3283
3284 return nullptr;
3285 }
3286
3287 /// Fold an icmp with LLVM intrinsics
foldICmpIntrinsicWithIntrinsic(ICmpInst & Cmp)3288 static Instruction *foldICmpIntrinsicWithIntrinsic(ICmpInst &Cmp) {
3289 assert(Cmp.isEquality());
3290
3291 ICmpInst::Predicate Pred = Cmp.getPredicate();
3292 Value *Op0 = Cmp.getOperand(0);
3293 Value *Op1 = Cmp.getOperand(1);
3294 const auto *IIOp0 = dyn_cast<IntrinsicInst>(Op0);
3295 const auto *IIOp1 = dyn_cast<IntrinsicInst>(Op1);
3296 if (!IIOp0 || !IIOp1 || IIOp0->getIntrinsicID() != IIOp1->getIntrinsicID())
3297 return nullptr;
3298
3299 switch (IIOp0->getIntrinsicID()) {
3300 case Intrinsic::bswap:
3301 case Intrinsic::bitreverse:
3302 // If both operands are byte-swapped or bit-reversed, just compare the
3303 // original values.
3304 return new ICmpInst(Pred, IIOp0->getOperand(0), IIOp1->getOperand(0));
3305 case Intrinsic::fshl:
3306 case Intrinsic::fshr:
3307 // If both operands are rotated by same amount, just compare the
3308 // original values.
3309 if (IIOp0->getOperand(0) != IIOp0->getOperand(1))
3310 break;
3311 if (IIOp1->getOperand(0) != IIOp1->getOperand(1))
3312 break;
3313 if (IIOp0->getOperand(2) != IIOp1->getOperand(2))
3314 break;
3315 return new ICmpInst(Pred, IIOp0->getOperand(0), IIOp1->getOperand(0));
3316 default:
3317 break;
3318 }
3319
3320 return nullptr;
3321 }
3322
3323 /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C.
foldICmpIntrinsicWithConstant(ICmpInst & Cmp,IntrinsicInst * II,const APInt & C)3324 Instruction *InstCombinerImpl::foldICmpIntrinsicWithConstant(ICmpInst &Cmp,
3325 IntrinsicInst *II,
3326 const APInt &C) {
3327 if (Cmp.isEquality())
3328 return foldICmpEqIntrinsicWithConstant(Cmp, II, C);
3329
3330 Type *Ty = II->getType();
3331 unsigned BitWidth = C.getBitWidth();
3332 ICmpInst::Predicate Pred = Cmp.getPredicate();
3333 switch (II->getIntrinsicID()) {
3334 case Intrinsic::ctpop: {
3335 // (ctpop X > BitWidth - 1) --> X == -1
3336 Value *X = II->getArgOperand(0);
3337 if (C == BitWidth - 1 && Pred == ICmpInst::ICMP_UGT)
3338 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ, X,
3339 ConstantInt::getAllOnesValue(Ty));
3340 // (ctpop X < BitWidth) --> X != -1
3341 if (C == BitWidth && Pred == ICmpInst::ICMP_ULT)
3342 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE, X,
3343 ConstantInt::getAllOnesValue(Ty));
3344 break;
3345 }
3346 case Intrinsic::ctlz: {
3347 // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000
3348 if (Pred == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3349 unsigned Num = C.getLimitedValue();
3350 APInt Limit = APInt::getOneBitSet(BitWidth, BitWidth - Num - 1);
3351 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_ULT,
3352 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3353 }
3354
3355 // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111
3356 if (Pred == ICmpInst::ICMP_ULT && C.uge(1) && C.ule(BitWidth)) {
3357 unsigned Num = C.getLimitedValue();
3358 APInt Limit = APInt::getLowBitsSet(BitWidth, BitWidth - Num);
3359 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_UGT,
3360 II->getArgOperand(0), ConstantInt::get(Ty, Limit));
3361 }
3362 break;
3363 }
3364 case Intrinsic::cttz: {
3365 // Limit to one use to ensure we don't increase instruction count.
3366 if (!II->hasOneUse())
3367 return nullptr;
3368
3369 // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0
3370 if (Pred == ICmpInst::ICMP_UGT && C.ult(BitWidth)) {
3371 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue() + 1);
3372 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_EQ,
3373 Builder.CreateAnd(II->getArgOperand(0), Mask),
3374 ConstantInt::getNullValue(Ty));
3375 }
3376
3377 // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0
3378 if (Pred == ICmpInst::ICMP_ULT && C.uge(1) && C.ule(BitWidth)) {
3379 APInt Mask = APInt::getLowBitsSet(BitWidth, C.getLimitedValue());
3380 return CmpInst::Create(Instruction::ICmp, ICmpInst::ICMP_NE,
3381 Builder.CreateAnd(II->getArgOperand(0), Mask),
3382 ConstantInt::getNullValue(Ty));
3383 }
3384 break;
3385 }
3386 default:
3387 break;
3388 }
3389
3390 return nullptr;
3391 }
3392
3393 /// Handle icmp with constant (but not simple integer constant) RHS.
foldICmpInstWithConstantNotInt(ICmpInst & I)3394 Instruction *InstCombinerImpl::foldICmpInstWithConstantNotInt(ICmpInst &I) {
3395 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3396 Constant *RHSC = dyn_cast<Constant>(Op1);
3397 Instruction *LHSI = dyn_cast<Instruction>(Op0);
3398 if (!RHSC || !LHSI)
3399 return nullptr;
3400
3401 switch (LHSI->getOpcode()) {
3402 case Instruction::GetElementPtr:
3403 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
3404 if (RHSC->isNullValue() &&
3405 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
3406 return new ICmpInst(
3407 I.getPredicate(), LHSI->getOperand(0),
3408 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3409 break;
3410 case Instruction::PHI:
3411 // Only fold icmp into the PHI if the phi and icmp are in the same
3412 // block. If in the same block, we're encouraging jump threading. If
3413 // not, we are just pessimizing the code by making an i1 phi.
3414 if (LHSI->getParent() == I.getParent())
3415 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
3416 return NV;
3417 break;
3418 case Instruction::Select: {
3419 // If either operand of the select is a constant, we can fold the
3420 // comparison into the select arms, which will cause one to be
3421 // constant folded and the select turned into a bitwise or.
3422 Value *Op1 = nullptr, *Op2 = nullptr;
3423 ConstantInt *CI = nullptr;
3424
3425 auto SimplifyOp = [&](Value *V) {
3426 Value *Op = nullptr;
3427 if (Constant *C = dyn_cast<Constant>(V)) {
3428 Op = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
3429 } else if (RHSC->isNullValue()) {
3430 // If null is being compared, check if it can be further simplified.
3431 Op = SimplifyICmpInst(I.getPredicate(), V, RHSC, SQ);
3432 }
3433 return Op;
3434 };
3435 Op1 = SimplifyOp(LHSI->getOperand(1));
3436 if (Op1)
3437 CI = dyn_cast<ConstantInt>(Op1);
3438
3439 Op2 = SimplifyOp(LHSI->getOperand(2));
3440 if (Op2)
3441 CI = dyn_cast<ConstantInt>(Op2);
3442
3443 // We only want to perform this transformation if it will not lead to
3444 // additional code. This is true if either both sides of the select
3445 // fold to a constant (in which case the icmp is replaced with a select
3446 // which will usually simplify) or this is the only user of the
3447 // select (in which case we are trading a select+icmp for a simpler
3448 // select+icmp) or all uses of the select can be replaced based on
3449 // dominance information ("Global cases").
3450 bool Transform = false;
3451 if (Op1 && Op2)
3452 Transform = true;
3453 else if (Op1 || Op2) {
3454 // Local case
3455 if (LHSI->hasOneUse())
3456 Transform = true;
3457 // Global cases
3458 else if (CI && !CI->isZero())
3459 // When Op1 is constant try replacing select with second operand.
3460 // Otherwise Op2 is constant and try replacing select with first
3461 // operand.
3462 Transform =
3463 replacedSelectWithOperand(cast<SelectInst>(LHSI), &I, Op1 ? 2 : 1);
3464 }
3465 if (Transform) {
3466 if (!Op1)
3467 Op1 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(1), RHSC,
3468 I.getName());
3469 if (!Op2)
3470 Op2 = Builder.CreateICmp(I.getPredicate(), LHSI->getOperand(2), RHSC,
3471 I.getName());
3472 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
3473 }
3474 break;
3475 }
3476 case Instruction::IntToPtr:
3477 // icmp pred inttoptr(X), null -> icmp pred X, 0
3478 if (RHSC->isNullValue() &&
3479 DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(0)->getType())
3480 return new ICmpInst(
3481 I.getPredicate(), LHSI->getOperand(0),
3482 Constant::getNullValue(LHSI->getOperand(0)->getType()));
3483 break;
3484
3485 case Instruction::Load:
3486 // Try to optimize things like "A[i] > 4" to index computations.
3487 if (GetElementPtrInst *GEP =
3488 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
3489 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
3490 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
3491 !cast<LoadInst>(LHSI)->isVolatile())
3492 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
3493 return Res;
3494 }
3495 break;
3496 }
3497
3498 return nullptr;
3499 }
3500
3501 /// Some comparisons can be simplified.
3502 /// In this case, we are looking for comparisons that look like
3503 /// a check for a lossy truncation.
3504 /// Folds:
3505 /// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask
3506 /// Where Mask is some pattern that produces all-ones in low bits:
3507 /// (-1 >> y)
3508 /// ((-1 << y) >> y) <- non-canonical, has extra uses
3509 /// ~(-1 << y)
3510 /// ((1 << y) + (-1)) <- non-canonical, has extra uses
3511 /// The Mask can be a constant, too.
3512 /// For some predicates, the operands are commutative.
3513 /// For others, x can only be on a specific side.
foldICmpWithLowBitMaskedVal(ICmpInst & I,InstCombiner::BuilderTy & Builder)3514 static Value *foldICmpWithLowBitMaskedVal(ICmpInst &I,
3515 InstCombiner::BuilderTy &Builder) {
3516 ICmpInst::Predicate SrcPred;
3517 Value *X, *M, *Y;
3518 auto m_VariableMask = m_CombineOr(
3519 m_CombineOr(m_Not(m_Shl(m_AllOnes(), m_Value())),
3520 m_Add(m_Shl(m_One(), m_Value()), m_AllOnes())),
3521 m_CombineOr(m_LShr(m_AllOnes(), m_Value()),
3522 m_LShr(m_Shl(m_AllOnes(), m_Value(Y)), m_Deferred(Y))));
3523 auto m_Mask = m_CombineOr(m_VariableMask, m_LowBitMask());
3524 if (!match(&I, m_c_ICmp(SrcPred,
3525 m_c_And(m_CombineAnd(m_Mask, m_Value(M)), m_Value(X)),
3526 m_Deferred(X))))
3527 return nullptr;
3528
3529 ICmpInst::Predicate DstPred;
3530 switch (SrcPred) {
3531 case ICmpInst::Predicate::ICMP_EQ:
3532 // x & (-1 >> y) == x -> x u<= (-1 >> y)
3533 DstPred = ICmpInst::Predicate::ICMP_ULE;
3534 break;
3535 case ICmpInst::Predicate::ICMP_NE:
3536 // x & (-1 >> y) != x -> x u> (-1 >> y)
3537 DstPred = ICmpInst::Predicate::ICMP_UGT;
3538 break;
3539 case ICmpInst::Predicate::ICMP_ULT:
3540 // x & (-1 >> y) u< x -> x u> (-1 >> y)
3541 // x u> x & (-1 >> y) -> x u> (-1 >> y)
3542 DstPred = ICmpInst::Predicate::ICMP_UGT;
3543 break;
3544 case ICmpInst::Predicate::ICMP_UGE:
3545 // x & (-1 >> y) u>= x -> x u<= (-1 >> y)
3546 // x u<= x & (-1 >> y) -> x u<= (-1 >> y)
3547 DstPred = ICmpInst::Predicate::ICMP_ULE;
3548 break;
3549 case ICmpInst::Predicate::ICMP_SLT:
3550 // x & (-1 >> y) s< x -> x s> (-1 >> y)
3551 // x s> x & (-1 >> y) -> x s> (-1 >> y)
3552 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3553 return nullptr;
3554 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3555 return nullptr;
3556 DstPred = ICmpInst::Predicate::ICMP_SGT;
3557 break;
3558 case ICmpInst::Predicate::ICMP_SGE:
3559 // x & (-1 >> y) s>= x -> x s<= (-1 >> y)
3560 // x s<= x & (-1 >> y) -> x s<= (-1 >> y)
3561 if (!match(M, m_Constant())) // Can not do this fold with non-constant.
3562 return nullptr;
3563 if (!match(M, m_NonNegative())) // Must not have any -1 vector elements.
3564 return nullptr;
3565 DstPred = ICmpInst::Predicate::ICMP_SLE;
3566 break;
3567 case ICmpInst::Predicate::ICMP_SGT:
3568 case ICmpInst::Predicate::ICMP_SLE:
3569 return nullptr;
3570 case ICmpInst::Predicate::ICMP_UGT:
3571 case ICmpInst::Predicate::ICMP_ULE:
3572 llvm_unreachable("Instsimplify took care of commut. variant");
3573 break;
3574 default:
3575 llvm_unreachable("All possible folds are handled.");
3576 }
3577
3578 // The mask value may be a vector constant that has undefined elements. But it
3579 // may not be safe to propagate those undefs into the new compare, so replace
3580 // those elements by copying an existing, defined, and safe scalar constant.
3581 Type *OpTy = M->getType();
3582 auto *VecC = dyn_cast<Constant>(M);
3583 auto *OpVTy = dyn_cast<FixedVectorType>(OpTy);
3584 if (OpVTy && VecC && VecC->containsUndefOrPoisonElement()) {
3585 Constant *SafeReplacementConstant = nullptr;
3586 for (unsigned i = 0, e = OpVTy->getNumElements(); i != e; ++i) {
3587 if (!isa<UndefValue>(VecC->getAggregateElement(i))) {
3588 SafeReplacementConstant = VecC->getAggregateElement(i);
3589 break;
3590 }
3591 }
3592 assert(SafeReplacementConstant && "Failed to find undef replacement");
3593 M = Constant::replaceUndefsWith(VecC, SafeReplacementConstant);
3594 }
3595
3596 return Builder.CreateICmp(DstPred, X, M);
3597 }
3598
3599 /// Some comparisons can be simplified.
3600 /// In this case, we are looking for comparisons that look like
3601 /// a check for a lossy signed truncation.
3602 /// Folds: (MaskedBits is a constant.)
3603 /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x
3604 /// Into:
3605 /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits)
3606 /// Where KeptBits = bitwidth(%x) - MaskedBits
3607 static Value *
foldICmpWithTruncSignExtendedVal(ICmpInst & I,InstCombiner::BuilderTy & Builder)3608 foldICmpWithTruncSignExtendedVal(ICmpInst &I,
3609 InstCombiner::BuilderTy &Builder) {
3610 ICmpInst::Predicate SrcPred;
3611 Value *X;
3612 const APInt *C0, *C1; // FIXME: non-splats, potentially with undef.
3613 // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use.
3614 if (!match(&I, m_c_ICmp(SrcPred,
3615 m_OneUse(m_AShr(m_Shl(m_Value(X), m_APInt(C0)),
3616 m_APInt(C1))),
3617 m_Deferred(X))))
3618 return nullptr;
3619
3620 // Potential handling of non-splats: for each element:
3621 // * if both are undef, replace with constant 0.
3622 // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0.
3623 // * if both are not undef, and are different, bailout.
3624 // * else, only one is undef, then pick the non-undef one.
3625
3626 // The shift amount must be equal.
3627 if (*C0 != *C1)
3628 return nullptr;
3629 const APInt &MaskedBits = *C0;
3630 assert(MaskedBits != 0 && "shift by zero should be folded away already.");
3631
3632 ICmpInst::Predicate DstPred;
3633 switch (SrcPred) {
3634 case ICmpInst::Predicate::ICMP_EQ:
3635 // ((%x << MaskedBits) a>> MaskedBits) == %x
3636 // =>
3637 // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits)
3638 DstPred = ICmpInst::Predicate::ICMP_ULT;
3639 break;
3640 case ICmpInst::Predicate::ICMP_NE:
3641 // ((%x << MaskedBits) a>> MaskedBits) != %x
3642 // =>
3643 // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits)
3644 DstPred = ICmpInst::Predicate::ICMP_UGE;
3645 break;
3646 // FIXME: are more folds possible?
3647 default:
3648 return nullptr;
3649 }
3650
3651 auto *XType = X->getType();
3652 const unsigned XBitWidth = XType->getScalarSizeInBits();
3653 const APInt BitWidth = APInt(XBitWidth, XBitWidth);
3654 assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched");
3655
3656 // KeptBits = bitwidth(%x) - MaskedBits
3657 const APInt KeptBits = BitWidth - MaskedBits;
3658 assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable");
3659 // ICmpCst = (1 << KeptBits)
3660 const APInt ICmpCst = APInt(XBitWidth, 1).shl(KeptBits);
3661 assert(ICmpCst.isPowerOf2());
3662 // AddCst = (1 << (KeptBits-1))
3663 const APInt AddCst = ICmpCst.lshr(1);
3664 assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2());
3665
3666 // T0 = add %x, AddCst
3667 Value *T0 = Builder.CreateAdd(X, ConstantInt::get(XType, AddCst));
3668 // T1 = T0 DstPred ICmpCst
3669 Value *T1 = Builder.CreateICmp(DstPred, T0, ConstantInt::get(XType, ICmpCst));
3670
3671 return T1;
3672 }
3673
3674 // Given pattern:
3675 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3676 // we should move shifts to the same hand of 'and', i.e. rewrite as
3677 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3678 // We are only interested in opposite logical shifts here.
3679 // One of the shifts can be truncated.
3680 // If we can, we want to end up creating 'lshr' shift.
3681 static Value *
foldShiftIntoShiftInAnotherHandOfAndInICmp(ICmpInst & I,const SimplifyQuery SQ,InstCombiner::BuilderTy & Builder)3682 foldShiftIntoShiftInAnotherHandOfAndInICmp(ICmpInst &I, const SimplifyQuery SQ,
3683 InstCombiner::BuilderTy &Builder) {
3684 if (!I.isEquality() || !match(I.getOperand(1), m_Zero()) ||
3685 !I.getOperand(0)->hasOneUse())
3686 return nullptr;
3687
3688 auto m_AnyLogicalShift = m_LogicalShift(m_Value(), m_Value());
3689
3690 // Look for an 'and' of two logical shifts, one of which may be truncated.
3691 // We use m_TruncOrSelf() on the RHS to correctly handle commutative case.
3692 Instruction *XShift, *MaybeTruncation, *YShift;
3693 if (!match(
3694 I.getOperand(0),
3695 m_c_And(m_CombineAnd(m_AnyLogicalShift, m_Instruction(XShift)),
3696 m_CombineAnd(m_TruncOrSelf(m_CombineAnd(
3697 m_AnyLogicalShift, m_Instruction(YShift))),
3698 m_Instruction(MaybeTruncation)))))
3699 return nullptr;
3700
3701 // We potentially looked past 'trunc', but only when matching YShift,
3702 // therefore YShift must have the widest type.
3703 Instruction *WidestShift = YShift;
3704 // Therefore XShift must have the shallowest type.
3705 // Or they both have identical types if there was no truncation.
3706 Instruction *NarrowestShift = XShift;
3707
3708 Type *WidestTy = WidestShift->getType();
3709 Type *NarrowestTy = NarrowestShift->getType();
3710 assert(NarrowestTy == I.getOperand(0)->getType() &&
3711 "We did not look past any shifts while matching XShift though.");
3712 bool HadTrunc = WidestTy != I.getOperand(0)->getType();
3713
3714 // If YShift is a 'lshr', swap the shifts around.
3715 if (match(YShift, m_LShr(m_Value(), m_Value())))
3716 std::swap(XShift, YShift);
3717
3718 // The shifts must be in opposite directions.
3719 auto XShiftOpcode = XShift->getOpcode();
3720 if (XShiftOpcode == YShift->getOpcode())
3721 return nullptr; // Do not care about same-direction shifts here.
3722
3723 Value *X, *XShAmt, *Y, *YShAmt;
3724 match(XShift, m_BinOp(m_Value(X), m_ZExtOrSelf(m_Value(XShAmt))));
3725 match(YShift, m_BinOp(m_Value(Y), m_ZExtOrSelf(m_Value(YShAmt))));
3726
3727 // If one of the values being shifted is a constant, then we will end with
3728 // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not,
3729 // however, we will need to ensure that we won't increase instruction count.
3730 if (!isa<Constant>(X) && !isa<Constant>(Y)) {
3731 // At least one of the hands of the 'and' should be one-use shift.
3732 if (!match(I.getOperand(0),
3733 m_c_And(m_OneUse(m_AnyLogicalShift), m_Value())))
3734 return nullptr;
3735 if (HadTrunc) {
3736 // Due to the 'trunc', we will need to widen X. For that either the old
3737 // 'trunc' or the shift amt in the non-truncated shift should be one-use.
3738 if (!MaybeTruncation->hasOneUse() &&
3739 !NarrowestShift->getOperand(1)->hasOneUse())
3740 return nullptr;
3741 }
3742 }
3743
3744 // We have two shift amounts from two different shifts. The types of those
3745 // shift amounts may not match. If that's the case let's bailout now.
3746 if (XShAmt->getType() != YShAmt->getType())
3747 return nullptr;
3748
3749 // As input, we have the following pattern:
3750 // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0
3751 // We want to rewrite that as:
3752 // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x)
3753 // While we know that originally (Q+K) would not overflow
3754 // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
3755 // shift amounts. so it may now overflow in smaller bitwidth.
3756 // To ensure that does not happen, we need to ensure that the total maximal
3757 // shift amount is still representable in that smaller bit width.
3758 unsigned MaximalPossibleTotalShiftAmount =
3759 (WidestTy->getScalarSizeInBits() - 1) +
3760 (NarrowestTy->getScalarSizeInBits() - 1);
3761 APInt MaximalRepresentableShiftAmount =
3762 APInt::getAllOnes(XShAmt->getType()->getScalarSizeInBits());
3763 if (MaximalRepresentableShiftAmount.ult(MaximalPossibleTotalShiftAmount))
3764 return nullptr;
3765
3766 // Can we fold (XShAmt+YShAmt) ?
3767 auto *NewShAmt = dyn_cast_or_null<Constant>(
3768 SimplifyAddInst(XShAmt, YShAmt, /*isNSW=*/false,
3769 /*isNUW=*/false, SQ.getWithInstruction(&I)));
3770 if (!NewShAmt)
3771 return nullptr;
3772 NewShAmt = ConstantExpr::getZExtOrBitCast(NewShAmt, WidestTy);
3773 unsigned WidestBitWidth = WidestTy->getScalarSizeInBits();
3774
3775 // Is the new shift amount smaller than the bit width?
3776 // FIXME: could also rely on ConstantRange.
3777 if (!match(NewShAmt,
3778 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_ULT,
3779 APInt(WidestBitWidth, WidestBitWidth))))
3780 return nullptr;
3781
3782 // An extra legality check is needed if we had trunc-of-lshr.
3783 if (HadTrunc && match(WidestShift, m_LShr(m_Value(), m_Value()))) {
3784 auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ,
3785 WidestShift]() {
3786 // It isn't obvious whether it's worth it to analyze non-constants here.
3787 // Also, let's basically give up on non-splat cases, pessimizing vectors.
3788 // If *any* of these preconditions matches we can perform the fold.
3789 Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy()
3790 ? NewShAmt->getSplatValue()
3791 : NewShAmt;
3792 // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold.
3793 if (NewShAmtSplat &&
3794 (NewShAmtSplat->isNullValue() ||
3795 NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1))
3796 return true;
3797 // We consider *min* leading zeros so a single outlier
3798 // blocks the transform as opposed to allowing it.
3799 if (auto *C = dyn_cast<Constant>(NarrowestShift->getOperand(0))) {
3800 KnownBits Known = computeKnownBits(C, SQ.DL);
3801 unsigned MinLeadZero = Known.countMinLeadingZeros();
3802 // If the value being shifted has at most lowest bit set we can fold.
3803 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3804 if (MaxActiveBits <= 1)
3805 return true;
3806 // Precondition: NewShAmt u<= countLeadingZeros(C)
3807 if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(MinLeadZero))
3808 return true;
3809 }
3810 if (auto *C = dyn_cast<Constant>(WidestShift->getOperand(0))) {
3811 KnownBits Known = computeKnownBits(C, SQ.DL);
3812 unsigned MinLeadZero = Known.countMinLeadingZeros();
3813 // If the value being shifted has at most lowest bit set we can fold.
3814 unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero;
3815 if (MaxActiveBits <= 1)
3816 return true;
3817 // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C)
3818 if (NewShAmtSplat) {
3819 APInt AdjNewShAmt =
3820 (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger();
3821 if (AdjNewShAmt.ule(MinLeadZero))
3822 return true;
3823 }
3824 }
3825 return false; // Can't tell if it's ok.
3826 };
3827 if (!CanFold())
3828 return nullptr;
3829 }
3830
3831 // All good, we can do this fold.
3832 X = Builder.CreateZExt(X, WidestTy);
3833 Y = Builder.CreateZExt(Y, WidestTy);
3834 // The shift is the same that was for X.
3835 Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr
3836 ? Builder.CreateLShr(X, NewShAmt)
3837 : Builder.CreateShl(X, NewShAmt);
3838 Value *T1 = Builder.CreateAnd(T0, Y);
3839 return Builder.CreateICmp(I.getPredicate(), T1,
3840 Constant::getNullValue(WidestTy));
3841 }
3842
3843 /// Fold
3844 /// (-1 u/ x) u< y
3845 /// ((x * y) ?/ x) != y
3846 /// to
3847 /// @llvm.?mul.with.overflow(x, y) plus extraction of overflow bit
3848 /// Note that the comparison is commutative, while inverted (u>=, ==) predicate
3849 /// will mean that we are looking for the opposite answer.
foldMultiplicationOverflowCheck(ICmpInst & I)3850 Value *InstCombinerImpl::foldMultiplicationOverflowCheck(ICmpInst &I) {
3851 ICmpInst::Predicate Pred;
3852 Value *X, *Y;
3853 Instruction *Mul;
3854 Instruction *Div;
3855 bool NeedNegation;
3856 // Look for: (-1 u/ x) u</u>= y
3857 if (!I.isEquality() &&
3858 match(&I, m_c_ICmp(Pred,
3859 m_CombineAnd(m_OneUse(m_UDiv(m_AllOnes(), m_Value(X))),
3860 m_Instruction(Div)),
3861 m_Value(Y)))) {
3862 Mul = nullptr;
3863
3864 // Are we checking that overflow does not happen, or does happen?
3865 switch (Pred) {
3866 case ICmpInst::Predicate::ICMP_ULT:
3867 NeedNegation = false;
3868 break; // OK
3869 case ICmpInst::Predicate::ICMP_UGE:
3870 NeedNegation = true;
3871 break; // OK
3872 default:
3873 return nullptr; // Wrong predicate.
3874 }
3875 } else // Look for: ((x * y) / x) !=/== y
3876 if (I.isEquality() &&
3877 match(&I,
3878 m_c_ICmp(Pred, m_Value(Y),
3879 m_CombineAnd(
3880 m_OneUse(m_IDiv(m_CombineAnd(m_c_Mul(m_Deferred(Y),
3881 m_Value(X)),
3882 m_Instruction(Mul)),
3883 m_Deferred(X))),
3884 m_Instruction(Div))))) {
3885 NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ;
3886 } else
3887 return nullptr;
3888
3889 BuilderTy::InsertPointGuard Guard(Builder);
3890 // If the pattern included (x * y), we'll want to insert new instructions
3891 // right before that original multiplication so that we can replace it.
3892 bool MulHadOtherUses = Mul && !Mul->hasOneUse();
3893 if (MulHadOtherUses)
3894 Builder.SetInsertPoint(Mul);
3895
3896 Function *F = Intrinsic::getDeclaration(I.getModule(),
3897 Div->getOpcode() == Instruction::UDiv
3898 ? Intrinsic::umul_with_overflow
3899 : Intrinsic::smul_with_overflow,
3900 X->getType());
3901 CallInst *Call = Builder.CreateCall(F, {X, Y}, "mul");
3902
3903 // If the multiplication was used elsewhere, to ensure that we don't leave
3904 // "duplicate" instructions, replace uses of that original multiplication
3905 // with the multiplication result from the with.overflow intrinsic.
3906 if (MulHadOtherUses)
3907 replaceInstUsesWith(*Mul, Builder.CreateExtractValue(Call, 0, "mul.val"));
3908
3909 Value *Res = Builder.CreateExtractValue(Call, 1, "mul.ov");
3910 if (NeedNegation) // This technically increases instruction count.
3911 Res = Builder.CreateNot(Res, "mul.not.ov");
3912
3913 // If we replaced the mul, erase it. Do this after all uses of Builder,
3914 // as the mul is used as insertion point.
3915 if (MulHadOtherUses)
3916 eraseInstFromFunction(*Mul);
3917
3918 return Res;
3919 }
3920
foldICmpXNegX(ICmpInst & I)3921 static Instruction *foldICmpXNegX(ICmpInst &I) {
3922 CmpInst::Predicate Pred;
3923 Value *X;
3924 if (!match(&I, m_c_ICmp(Pred, m_NSWNeg(m_Value(X)), m_Deferred(X))))
3925 return nullptr;
3926
3927 if (ICmpInst::isSigned(Pred))
3928 Pred = ICmpInst::getSwappedPredicate(Pred);
3929 else if (ICmpInst::isUnsigned(Pred))
3930 Pred = ICmpInst::getSignedPredicate(Pred);
3931 // else for equality-comparisons just keep the predicate.
3932
3933 return ICmpInst::Create(Instruction::ICmp, Pred, X,
3934 Constant::getNullValue(X->getType()), I.getName());
3935 }
3936
3937 /// Try to fold icmp (binop), X or icmp X, (binop).
3938 /// TODO: A large part of this logic is duplicated in InstSimplify's
3939 /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code
3940 /// duplication.
foldICmpBinOp(ICmpInst & I,const SimplifyQuery & SQ)3941 Instruction *InstCombinerImpl::foldICmpBinOp(ICmpInst &I,
3942 const SimplifyQuery &SQ) {
3943 const SimplifyQuery Q = SQ.getWithInstruction(&I);
3944 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3945
3946 // Special logic for binary operators.
3947 BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Op0);
3948 BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Op1);
3949 if (!BO0 && !BO1)
3950 return nullptr;
3951
3952 if (Instruction *NewICmp = foldICmpXNegX(I))
3953 return NewICmp;
3954
3955 const CmpInst::Predicate Pred = I.getPredicate();
3956 Value *X;
3957
3958 // Convert add-with-unsigned-overflow comparisons into a 'not' with compare.
3959 // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X
3960 if (match(Op0, m_OneUse(m_c_Add(m_Specific(Op1), m_Value(X)))) &&
3961 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
3962 return new ICmpInst(Pred, Builder.CreateNot(Op1), X);
3963 // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0
3964 if (match(Op1, m_OneUse(m_c_Add(m_Specific(Op0), m_Value(X)))) &&
3965 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
3966 return new ICmpInst(Pred, X, Builder.CreateNot(Op0));
3967
3968 bool NoOp0WrapProblem = false, NoOp1WrapProblem = false;
3969 if (BO0 && isa<OverflowingBinaryOperator>(BO0))
3970 NoOp0WrapProblem =
3971 ICmpInst::isEquality(Pred) ||
3972 (CmpInst::isUnsigned(Pred) && BO0->hasNoUnsignedWrap()) ||
3973 (CmpInst::isSigned(Pred) && BO0->hasNoSignedWrap());
3974 if (BO1 && isa<OverflowingBinaryOperator>(BO1))
3975 NoOp1WrapProblem =
3976 ICmpInst::isEquality(Pred) ||
3977 (CmpInst::isUnsigned(Pred) && BO1->hasNoUnsignedWrap()) ||
3978 (CmpInst::isSigned(Pred) && BO1->hasNoSignedWrap());
3979
3980 // Analyze the case when either Op0 or Op1 is an add instruction.
3981 // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null).
3982 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
3983 if (BO0 && BO0->getOpcode() == Instruction::Add) {
3984 A = BO0->getOperand(0);
3985 B = BO0->getOperand(1);
3986 }
3987 if (BO1 && BO1->getOpcode() == Instruction::Add) {
3988 C = BO1->getOperand(0);
3989 D = BO1->getOperand(1);
3990 }
3991
3992 // icmp (A+B), A -> icmp B, 0 for equalities or if there is no overflow.
3993 // icmp (A+B), B -> icmp A, 0 for equalities or if there is no overflow.
3994 if ((A == Op1 || B == Op1) && NoOp0WrapProblem)
3995 return new ICmpInst(Pred, A == Op1 ? B : A,
3996 Constant::getNullValue(Op1->getType()));
3997
3998 // icmp C, (C+D) -> icmp 0, D for equalities or if there is no overflow.
3999 // icmp D, (C+D) -> icmp 0, C for equalities or if there is no overflow.
4000 if ((C == Op0 || D == Op0) && NoOp1WrapProblem)
4001 return new ICmpInst(Pred, Constant::getNullValue(Op0->getType()),
4002 C == Op0 ? D : C);
4003
4004 // icmp (A+B), (A+D) -> icmp B, D for equalities or if there is no overflow.
4005 if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem &&
4006 NoOp1WrapProblem) {
4007 // Determine Y and Z in the form icmp (X+Y), (X+Z).
4008 Value *Y, *Z;
4009 if (A == C) {
4010 // C + B == C + D -> B == D
4011 Y = B;
4012 Z = D;
4013 } else if (A == D) {
4014 // D + B == C + D -> B == C
4015 Y = B;
4016 Z = C;
4017 } else if (B == C) {
4018 // A + C == C + D -> A == D
4019 Y = A;
4020 Z = D;
4021 } else {
4022 assert(B == D);
4023 // A + D == C + D -> A == C
4024 Y = A;
4025 Z = C;
4026 }
4027 return new ICmpInst(Pred, Y, Z);
4028 }
4029
4030 // icmp slt (A + -1), Op1 -> icmp sle A, Op1
4031 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT &&
4032 match(B, m_AllOnes()))
4033 return new ICmpInst(CmpInst::ICMP_SLE, A, Op1);
4034
4035 // icmp sge (A + -1), Op1 -> icmp sgt A, Op1
4036 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE &&
4037 match(B, m_AllOnes()))
4038 return new ICmpInst(CmpInst::ICMP_SGT, A, Op1);
4039
4040 // icmp sle (A + 1), Op1 -> icmp slt A, Op1
4041 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(B, m_One()))
4042 return new ICmpInst(CmpInst::ICMP_SLT, A, Op1);
4043
4044 // icmp sgt (A + 1), Op1 -> icmp sge A, Op1
4045 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(B, m_One()))
4046 return new ICmpInst(CmpInst::ICMP_SGE, A, Op1);
4047
4048 // icmp sgt Op0, (C + -1) -> icmp sge Op0, C
4049 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT &&
4050 match(D, m_AllOnes()))
4051 return new ICmpInst(CmpInst::ICMP_SGE, Op0, C);
4052
4053 // icmp sle Op0, (C + -1) -> icmp slt Op0, C
4054 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE &&
4055 match(D, m_AllOnes()))
4056 return new ICmpInst(CmpInst::ICMP_SLT, Op0, C);
4057
4058 // icmp sge Op0, (C + 1) -> icmp sgt Op0, C
4059 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(D, m_One()))
4060 return new ICmpInst(CmpInst::ICMP_SGT, Op0, C);
4061
4062 // icmp slt Op0, (C + 1) -> icmp sle Op0, C
4063 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(D, m_One()))
4064 return new ICmpInst(CmpInst::ICMP_SLE, Op0, C);
4065
4066 // TODO: The subtraction-related identities shown below also hold, but
4067 // canonicalization from (X -nuw 1) to (X + -1) means that the combinations
4068 // wouldn't happen even if they were implemented.
4069 //
4070 // icmp ult (A - 1), Op1 -> icmp ule A, Op1
4071 // icmp uge (A - 1), Op1 -> icmp ugt A, Op1
4072 // icmp ugt Op0, (C - 1) -> icmp uge Op0, C
4073 // icmp ule Op0, (C - 1) -> icmp ult Op0, C
4074
4075 // icmp ule (A + 1), Op0 -> icmp ult A, Op1
4076 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(B, m_One()))
4077 return new ICmpInst(CmpInst::ICMP_ULT, A, Op1);
4078
4079 // icmp ugt (A + 1), Op0 -> icmp uge A, Op1
4080 if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(B, m_One()))
4081 return new ICmpInst(CmpInst::ICMP_UGE, A, Op1);
4082
4083 // icmp uge Op0, (C + 1) -> icmp ugt Op0, C
4084 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(D, m_One()))
4085 return new ICmpInst(CmpInst::ICMP_UGT, Op0, C);
4086
4087 // icmp ult Op0, (C + 1) -> icmp ule Op0, C
4088 if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(D, m_One()))
4089 return new ICmpInst(CmpInst::ICMP_ULE, Op0, C);
4090
4091 // if C1 has greater magnitude than C2:
4092 // icmp (A + C1), (C + C2) -> icmp (A + C3), C
4093 // s.t. C3 = C1 - C2
4094 //
4095 // if C2 has greater magnitude than C1:
4096 // icmp (A + C1), (C + C2) -> icmp A, (C + C3)
4097 // s.t. C3 = C2 - C1
4098 if (A && C && NoOp0WrapProblem && NoOp1WrapProblem &&
4099 (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned())
4100 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
4101 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D)) {
4102 const APInt &AP1 = C1->getValue();
4103 const APInt &AP2 = C2->getValue();
4104 if (AP1.isNegative() == AP2.isNegative()) {
4105 APInt AP1Abs = C1->getValue().abs();
4106 APInt AP2Abs = C2->getValue().abs();
4107 if (AP1Abs.uge(AP2Abs)) {
4108 ConstantInt *C3 = Builder.getInt(AP1 - AP2);
4109 bool HasNUW = BO0->hasNoUnsignedWrap() && C3->getValue().ule(AP1);
4110 bool HasNSW = BO0->hasNoSignedWrap();
4111 Value *NewAdd = Builder.CreateAdd(A, C3, "", HasNUW, HasNSW);
4112 return new ICmpInst(Pred, NewAdd, C);
4113 } else {
4114 ConstantInt *C3 = Builder.getInt(AP2 - AP1);
4115 bool HasNUW = BO1->hasNoUnsignedWrap() && C3->getValue().ule(AP2);
4116 bool HasNSW = BO1->hasNoSignedWrap();
4117 Value *NewAdd = Builder.CreateAdd(C, C3, "", HasNUW, HasNSW);
4118 return new ICmpInst(Pred, A, NewAdd);
4119 }
4120 }
4121 }
4122
4123 // Analyze the case when either Op0 or Op1 is a sub instruction.
4124 // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null).
4125 A = nullptr;
4126 B = nullptr;
4127 C = nullptr;
4128 D = nullptr;
4129 if (BO0 && BO0->getOpcode() == Instruction::Sub) {
4130 A = BO0->getOperand(0);
4131 B = BO0->getOperand(1);
4132 }
4133 if (BO1 && BO1->getOpcode() == Instruction::Sub) {
4134 C = BO1->getOperand(0);
4135 D = BO1->getOperand(1);
4136 }
4137
4138 // icmp (A-B), A -> icmp 0, B for equalities or if there is no overflow.
4139 if (A == Op1 && NoOp0WrapProblem)
4140 return new ICmpInst(Pred, Constant::getNullValue(Op1->getType()), B);
4141 // icmp C, (C-D) -> icmp D, 0 for equalities or if there is no overflow.
4142 if (C == Op0 && NoOp1WrapProblem)
4143 return new ICmpInst(Pred, D, Constant::getNullValue(Op0->getType()));
4144
4145 // Convert sub-with-unsigned-overflow comparisons into a comparison of args.
4146 // (A - B) u>/u<= A --> B u>/u<= A
4147 if (A == Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE))
4148 return new ICmpInst(Pred, B, A);
4149 // C u</u>= (C - D) --> C u</u>= D
4150 if (C == Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE))
4151 return new ICmpInst(Pred, C, D);
4152 // (A - B) u>=/u< A --> B u>/u<= A iff B != 0
4153 if (A == Op1 && (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) &&
4154 isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
4155 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), B, A);
4156 // C u<=/u> (C - D) --> C u</u>= D iff B != 0
4157 if (C == Op0 && (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) &&
4158 isKnownNonZero(D, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
4159 return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(Pred), C, D);
4160
4161 // icmp (A-B), (C-B) -> icmp A, C for equalities or if there is no overflow.
4162 if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem)
4163 return new ICmpInst(Pred, A, C);
4164
4165 // icmp (A-B), (A-D) -> icmp D, B for equalities or if there is no overflow.
4166 if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem)
4167 return new ICmpInst(Pred, D, B);
4168
4169 // icmp (0-X) < cst --> x > -cst
4170 if (NoOp0WrapProblem && ICmpInst::isSigned(Pred)) {
4171 Value *X;
4172 if (match(BO0, m_Neg(m_Value(X))))
4173 if (Constant *RHSC = dyn_cast<Constant>(Op1))
4174 if (RHSC->isNotMinSignedValue())
4175 return new ICmpInst(I.getSwappedPredicate(), X,
4176 ConstantExpr::getNeg(RHSC));
4177 }
4178
4179 {
4180 // Try to remove shared constant multiplier from equality comparison:
4181 // X * C == Y * C (with no overflowing/aliasing) --> X == Y
4182 Value *X, *Y;
4183 const APInt *C;
4184 if (match(Op0, m_Mul(m_Value(X), m_APInt(C))) && *C != 0 &&
4185 match(Op1, m_Mul(m_Value(Y), m_SpecificInt(*C))) && I.isEquality())
4186 if (!C->countTrailingZeros() ||
4187 (BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap()) ||
4188 (BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap()))
4189 return new ICmpInst(Pred, X, Y);
4190 }
4191
4192 BinaryOperator *SRem = nullptr;
4193 // icmp (srem X, Y), Y
4194 if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(1))
4195 SRem = BO0;
4196 // icmp Y, (srem X, Y)
4197 else if (BO1 && BO1->getOpcode() == Instruction::SRem &&
4198 Op0 == BO1->getOperand(1))
4199 SRem = BO1;
4200 if (SRem) {
4201 // We don't check hasOneUse to avoid increasing register pressure because
4202 // the value we use is the same value this instruction was already using.
4203 switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(Pred) : Pred) {
4204 default:
4205 break;
4206 case ICmpInst::ICMP_EQ:
4207 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
4208 case ICmpInst::ICMP_NE:
4209 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
4210 case ICmpInst::ICMP_SGT:
4211 case ICmpInst::ICMP_SGE:
4212 return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(1),
4213 Constant::getAllOnesValue(SRem->getType()));
4214 case ICmpInst::ICMP_SLT:
4215 case ICmpInst::ICMP_SLE:
4216 return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(1),
4217 Constant::getNullValue(SRem->getType()));
4218 }
4219 }
4220
4221 if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
4222 BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
4223 switch (BO0->getOpcode()) {
4224 default:
4225 break;
4226 case Instruction::Add:
4227 case Instruction::Sub:
4228 case Instruction::Xor: {
4229 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
4230 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4231
4232 const APInt *C;
4233 if (match(BO0->getOperand(1), m_APInt(C))) {
4234 // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b
4235 if (C->isSignMask()) {
4236 ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
4237 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4238 }
4239
4240 // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b
4241 if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) {
4242 ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate();
4243 NewPred = I.getSwappedPredicate(NewPred);
4244 return new ICmpInst(NewPred, BO0->getOperand(0), BO1->getOperand(0));
4245 }
4246 }
4247 break;
4248 }
4249 case Instruction::Mul: {
4250 if (!I.isEquality())
4251 break;
4252
4253 const APInt *C;
4254 if (match(BO0->getOperand(1), m_APInt(C)) && !C->isZero() &&
4255 !C->isOne()) {
4256 // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask)
4257 // Mask = -1 >> count-trailing-zeros(C).
4258 if (unsigned TZs = C->countTrailingZeros()) {
4259 Constant *Mask = ConstantInt::get(
4260 BO0->getType(),
4261 APInt::getLowBitsSet(C->getBitWidth(), C->getBitWidth() - TZs));
4262 Value *And1 = Builder.CreateAnd(BO0->getOperand(0), Mask);
4263 Value *And2 = Builder.CreateAnd(BO1->getOperand(0), Mask);
4264 return new ICmpInst(Pred, And1, And2);
4265 }
4266 }
4267 break;
4268 }
4269 case Instruction::UDiv:
4270 case Instruction::LShr:
4271 if (I.isSigned() || !BO0->isExact() || !BO1->isExact())
4272 break;
4273 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4274
4275 case Instruction::SDiv:
4276 if (!I.isEquality() || !BO0->isExact() || !BO1->isExact())
4277 break;
4278 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4279
4280 case Instruction::AShr:
4281 if (!BO0->isExact() || !BO1->isExact())
4282 break;
4283 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4284
4285 case Instruction::Shl: {
4286 bool NUW = BO0->hasNoUnsignedWrap() && BO1->hasNoUnsignedWrap();
4287 bool NSW = BO0->hasNoSignedWrap() && BO1->hasNoSignedWrap();
4288 if (!NUW && !NSW)
4289 break;
4290 if (!NSW && I.isSigned())
4291 break;
4292 return new ICmpInst(Pred, BO0->getOperand(0), BO1->getOperand(0));
4293 }
4294 }
4295 }
4296
4297 if (BO0) {
4298 // Transform A & (L - 1) `ult` L --> L != 0
4299 auto LSubOne = m_Add(m_Specific(Op1), m_AllOnes());
4300 auto BitwiseAnd = m_c_And(m_Value(), LSubOne);
4301
4302 if (match(BO0, BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) {
4303 auto *Zero = Constant::getNullValue(BO0->getType());
4304 return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero);
4305 }
4306 }
4307
4308 if (Value *V = foldMultiplicationOverflowCheck(I))
4309 return replaceInstUsesWith(I, V);
4310
4311 if (Value *V = foldICmpWithLowBitMaskedVal(I, Builder))
4312 return replaceInstUsesWith(I, V);
4313
4314 if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder))
4315 return replaceInstUsesWith(I, V);
4316
4317 if (Value *V = foldShiftIntoShiftInAnotherHandOfAndInICmp(I, SQ, Builder))
4318 return replaceInstUsesWith(I, V);
4319
4320 return nullptr;
4321 }
4322
4323 /// Fold icmp Pred min|max(X, Y), X.
foldICmpWithMinMax(ICmpInst & Cmp)4324 static Instruction *foldICmpWithMinMax(ICmpInst &Cmp) {
4325 ICmpInst::Predicate Pred = Cmp.getPredicate();
4326 Value *Op0 = Cmp.getOperand(0);
4327 Value *X = Cmp.getOperand(1);
4328
4329 // Canonicalize minimum or maximum operand to LHS of the icmp.
4330 if (match(X, m_c_SMin(m_Specific(Op0), m_Value())) ||
4331 match(X, m_c_SMax(m_Specific(Op0), m_Value())) ||
4332 match(X, m_c_UMin(m_Specific(Op0), m_Value())) ||
4333 match(X, m_c_UMax(m_Specific(Op0), m_Value()))) {
4334 std::swap(Op0, X);
4335 Pred = Cmp.getSwappedPredicate();
4336 }
4337
4338 Value *Y;
4339 if (match(Op0, m_c_SMin(m_Specific(X), m_Value(Y)))) {
4340 // smin(X, Y) == X --> X s<= Y
4341 // smin(X, Y) s>= X --> X s<= Y
4342 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SGE)
4343 return new ICmpInst(ICmpInst::ICMP_SLE, X, Y);
4344
4345 // smin(X, Y) != X --> X s> Y
4346 // smin(X, Y) s< X --> X s> Y
4347 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SLT)
4348 return new ICmpInst(ICmpInst::ICMP_SGT, X, Y);
4349
4350 // These cases should be handled in InstSimplify:
4351 // smin(X, Y) s<= X --> true
4352 // smin(X, Y) s> X --> false
4353 return nullptr;
4354 }
4355
4356 if (match(Op0, m_c_SMax(m_Specific(X), m_Value(Y)))) {
4357 // smax(X, Y) == X --> X s>= Y
4358 // smax(X, Y) s<= X --> X s>= Y
4359 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_SLE)
4360 return new ICmpInst(ICmpInst::ICMP_SGE, X, Y);
4361
4362 // smax(X, Y) != X --> X s< Y
4363 // smax(X, Y) s> X --> X s< Y
4364 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_SGT)
4365 return new ICmpInst(ICmpInst::ICMP_SLT, X, Y);
4366
4367 // These cases should be handled in InstSimplify:
4368 // smax(X, Y) s>= X --> true
4369 // smax(X, Y) s< X --> false
4370 return nullptr;
4371 }
4372
4373 if (match(Op0, m_c_UMin(m_Specific(X), m_Value(Y)))) {
4374 // umin(X, Y) == X --> X u<= Y
4375 // umin(X, Y) u>= X --> X u<= Y
4376 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_UGE)
4377 return new ICmpInst(ICmpInst::ICMP_ULE, X, Y);
4378
4379 // umin(X, Y) != X --> X u> Y
4380 // umin(X, Y) u< X --> X u> Y
4381 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT)
4382 return new ICmpInst(ICmpInst::ICMP_UGT, X, Y);
4383
4384 // These cases should be handled in InstSimplify:
4385 // umin(X, Y) u<= X --> true
4386 // umin(X, Y) u> X --> false
4387 return nullptr;
4388 }
4389
4390 if (match(Op0, m_c_UMax(m_Specific(X), m_Value(Y)))) {
4391 // umax(X, Y) == X --> X u>= Y
4392 // umax(X, Y) u<= X --> X u>= Y
4393 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_ULE)
4394 return new ICmpInst(ICmpInst::ICMP_UGE, X, Y);
4395
4396 // umax(X, Y) != X --> X u< Y
4397 // umax(X, Y) u> X --> X u< Y
4398 if (Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_UGT)
4399 return new ICmpInst(ICmpInst::ICMP_ULT, X, Y);
4400
4401 // These cases should be handled in InstSimplify:
4402 // umax(X, Y) u>= X --> true
4403 // umax(X, Y) u< X --> false
4404 return nullptr;
4405 }
4406
4407 return nullptr;
4408 }
4409
foldICmpEquality(ICmpInst & I)4410 Instruction *InstCombinerImpl::foldICmpEquality(ICmpInst &I) {
4411 if (!I.isEquality())
4412 return nullptr;
4413
4414 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4415 const CmpInst::Predicate Pred = I.getPredicate();
4416 Value *A, *B, *C, *D;
4417 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4418 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
4419 Value *OtherVal = A == Op1 ? B : A;
4420 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4421 }
4422
4423 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
4424 // A^c1 == C^c2 --> A == C^(c1^c2)
4425 ConstantInt *C1, *C2;
4426 if (match(B, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2)) &&
4427 Op1->hasOneUse()) {
4428 Constant *NC = Builder.getInt(C1->getValue() ^ C2->getValue());
4429 Value *Xor = Builder.CreateXor(C, NC);
4430 return new ICmpInst(Pred, A, Xor);
4431 }
4432
4433 // A^B == A^D -> B == D
4434 if (A == C)
4435 return new ICmpInst(Pred, B, D);
4436 if (A == D)
4437 return new ICmpInst(Pred, B, C);
4438 if (B == C)
4439 return new ICmpInst(Pred, A, D);
4440 if (B == D)
4441 return new ICmpInst(Pred, A, C);
4442 }
4443 }
4444
4445 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) && (A == Op0 || B == Op0)) {
4446 // A == (A^B) -> B == 0
4447 Value *OtherVal = A == Op0 ? B : A;
4448 return new ICmpInst(Pred, OtherVal, Constant::getNullValue(A->getType()));
4449 }
4450
4451 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
4452 if (match(Op0, m_OneUse(m_And(m_Value(A), m_Value(B)))) &&
4453 match(Op1, m_OneUse(m_And(m_Value(C), m_Value(D))))) {
4454 Value *X = nullptr, *Y = nullptr, *Z = nullptr;
4455
4456 if (A == C) {
4457 X = B;
4458 Y = D;
4459 Z = A;
4460 } else if (A == D) {
4461 X = B;
4462 Y = C;
4463 Z = A;
4464 } else if (B == C) {
4465 X = A;
4466 Y = D;
4467 Z = B;
4468 } else if (B == D) {
4469 X = A;
4470 Y = C;
4471 Z = B;
4472 }
4473
4474 if (X) { // Build (X^Y) & Z
4475 Op1 = Builder.CreateXor(X, Y);
4476 Op1 = Builder.CreateAnd(Op1, Z);
4477 return new ICmpInst(Pred, Op1, Constant::getNullValue(Op1->getType()));
4478 }
4479 }
4480
4481 {
4482 // Similar to above, but specialized for constant because invert is needed:
4483 // (X | C) == (Y | C) --> (X ^ Y) & ~C == 0
4484 Value *X, *Y;
4485 Constant *C;
4486 if (match(Op0, m_OneUse(m_Or(m_Value(X), m_Constant(C)))) &&
4487 match(Op1, m_OneUse(m_Or(m_Value(Y), m_Specific(C))))) {
4488 Value *Xor = Builder.CreateXor(X, Y);
4489 Value *And = Builder.CreateAnd(Xor, ConstantExpr::getNot(C));
4490 return new ICmpInst(Pred, And, Constant::getNullValue(And->getType()));
4491 }
4492 }
4493
4494 // Transform (zext A) == (B & (1<<X)-1) --> A == (trunc B)
4495 // and (B & (1<<X)-1) == (zext A) --> A == (trunc B)
4496 ConstantInt *Cst1;
4497 if ((Op0->hasOneUse() && match(Op0, m_ZExt(m_Value(A))) &&
4498 match(Op1, m_And(m_Value(B), m_ConstantInt(Cst1)))) ||
4499 (Op1->hasOneUse() && match(Op0, m_And(m_Value(B), m_ConstantInt(Cst1))) &&
4500 match(Op1, m_ZExt(m_Value(A))))) {
4501 APInt Pow2 = Cst1->getValue() + 1;
4502 if (Pow2.isPowerOf2() && isa<IntegerType>(A->getType()) &&
4503 Pow2.logBase2() == cast<IntegerType>(A->getType())->getBitWidth())
4504 return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
4505 }
4506
4507 // (A >> C) == (B >> C) --> (A^B) u< (1 << C)
4508 // For lshr and ashr pairs.
4509 if ((match(Op0, m_OneUse(m_LShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4510 match(Op1, m_OneUse(m_LShr(m_Value(B), m_Specific(Cst1))))) ||
4511 (match(Op0, m_OneUse(m_AShr(m_Value(A), m_ConstantInt(Cst1)))) &&
4512 match(Op1, m_OneUse(m_AShr(m_Value(B), m_Specific(Cst1)))))) {
4513 unsigned TypeBits = Cst1->getBitWidth();
4514 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4515 if (ShAmt < TypeBits && ShAmt != 0) {
4516 ICmpInst::Predicate NewPred =
4517 Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
4518 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4519 APInt CmpVal = APInt::getOneBitSet(TypeBits, ShAmt);
4520 return new ICmpInst(NewPred, Xor, Builder.getInt(CmpVal));
4521 }
4522 }
4523
4524 // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0
4525 if (match(Op0, m_OneUse(m_Shl(m_Value(A), m_ConstantInt(Cst1)))) &&
4526 match(Op1, m_OneUse(m_Shl(m_Value(B), m_Specific(Cst1))))) {
4527 unsigned TypeBits = Cst1->getBitWidth();
4528 unsigned ShAmt = (unsigned)Cst1->getLimitedValue(TypeBits);
4529 if (ShAmt < TypeBits && ShAmt != 0) {
4530 Value *Xor = Builder.CreateXor(A, B, I.getName() + ".unshifted");
4531 APInt AndVal = APInt::getLowBitsSet(TypeBits, TypeBits - ShAmt);
4532 Value *And = Builder.CreateAnd(Xor, Builder.getInt(AndVal),
4533 I.getName() + ".mask");
4534 return new ICmpInst(Pred, And, Constant::getNullValue(Cst1->getType()));
4535 }
4536 }
4537
4538 // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to
4539 // "icmp (and X, mask), cst"
4540 uint64_t ShAmt = 0;
4541 if (Op0->hasOneUse() &&
4542 match(Op0, m_Trunc(m_OneUse(m_LShr(m_Value(A), m_ConstantInt(ShAmt))))) &&
4543 match(Op1, m_ConstantInt(Cst1)) &&
4544 // Only do this when A has multiple uses. This is most important to do
4545 // when it exposes other optimizations.
4546 !A->hasOneUse()) {
4547 unsigned ASize = cast<IntegerType>(A->getType())->getPrimitiveSizeInBits();
4548
4549 if (ShAmt < ASize) {
4550 APInt MaskV =
4551 APInt::getLowBitsSet(ASize, Op0->getType()->getPrimitiveSizeInBits());
4552 MaskV <<= ShAmt;
4553
4554 APInt CmpV = Cst1->getValue().zext(ASize);
4555 CmpV <<= ShAmt;
4556
4557 Value *Mask = Builder.CreateAnd(A, Builder.getInt(MaskV));
4558 return new ICmpInst(Pred, Mask, Builder.getInt(CmpV));
4559 }
4560 }
4561
4562 if (Instruction *ICmp = foldICmpIntrinsicWithIntrinsic(I))
4563 return ICmp;
4564
4565 // Canonicalize checking for a power-of-2-or-zero value:
4566 // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
4567 // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
4568 if (!match(Op0, m_OneUse(m_c_And(m_Add(m_Value(A), m_AllOnes()),
4569 m_Deferred(A)))) ||
4570 !match(Op1, m_ZeroInt()))
4571 A = nullptr;
4572
4573 // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
4574 // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
4575 if (match(Op0, m_OneUse(m_c_And(m_Neg(m_Specific(Op1)), m_Specific(Op1)))))
4576 A = Op1;
4577 else if (match(Op1,
4578 m_OneUse(m_c_And(m_Neg(m_Specific(Op0)), m_Specific(Op0)))))
4579 A = Op0;
4580
4581 if (A) {
4582 Type *Ty = A->getType();
4583 CallInst *CtPop = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, A);
4584 return Pred == ICmpInst::ICMP_EQ
4585 ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop, ConstantInt::get(Ty, 2))
4586 : new ICmpInst(ICmpInst::ICMP_UGT, CtPop, ConstantInt::get(Ty, 1));
4587 }
4588
4589 return nullptr;
4590 }
4591
foldICmpWithZextOrSext(ICmpInst & ICmp,InstCombiner::BuilderTy & Builder)4592 static Instruction *foldICmpWithZextOrSext(ICmpInst &ICmp,
4593 InstCombiner::BuilderTy &Builder) {
4594 assert(isa<CastInst>(ICmp.getOperand(0)) && "Expected cast for operand 0");
4595 auto *CastOp0 = cast<CastInst>(ICmp.getOperand(0));
4596 Value *X;
4597 if (!match(CastOp0, m_ZExtOrSExt(m_Value(X))))
4598 return nullptr;
4599
4600 bool IsSignedExt = CastOp0->getOpcode() == Instruction::SExt;
4601 bool IsSignedCmp = ICmp.isSigned();
4602 if (auto *CastOp1 = dyn_cast<CastInst>(ICmp.getOperand(1))) {
4603 // If the signedness of the two casts doesn't agree (i.e. one is a sext
4604 // and the other is a zext), then we can't handle this.
4605 // TODO: This is too strict. We can handle some predicates (equality?).
4606 if (CastOp0->getOpcode() != CastOp1->getOpcode())
4607 return nullptr;
4608
4609 // Not an extension from the same type?
4610 Value *Y = CastOp1->getOperand(0);
4611 Type *XTy = X->getType(), *YTy = Y->getType();
4612 if (XTy != YTy) {
4613 // One of the casts must have one use because we are creating a new cast.
4614 if (!CastOp0->hasOneUse() && !CastOp1->hasOneUse())
4615 return nullptr;
4616 // Extend the narrower operand to the type of the wider operand.
4617 if (XTy->getScalarSizeInBits() < YTy->getScalarSizeInBits())
4618 X = Builder.CreateCast(CastOp0->getOpcode(), X, YTy);
4619 else if (YTy->getScalarSizeInBits() < XTy->getScalarSizeInBits())
4620 Y = Builder.CreateCast(CastOp0->getOpcode(), Y, XTy);
4621 else
4622 return nullptr;
4623 }
4624
4625 // (zext X) == (zext Y) --> X == Y
4626 // (sext X) == (sext Y) --> X == Y
4627 if (ICmp.isEquality())
4628 return new ICmpInst(ICmp.getPredicate(), X, Y);
4629
4630 // A signed comparison of sign extended values simplifies into a
4631 // signed comparison.
4632 if (IsSignedCmp && IsSignedExt)
4633 return new ICmpInst(ICmp.getPredicate(), X, Y);
4634
4635 // The other three cases all fold into an unsigned comparison.
4636 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Y);
4637 }
4638
4639 // Below here, we are only folding a compare with constant.
4640 auto *C = dyn_cast<Constant>(ICmp.getOperand(1));
4641 if (!C)
4642 return nullptr;
4643
4644 // Compute the constant that would happen if we truncated to SrcTy then
4645 // re-extended to DestTy.
4646 Type *SrcTy = CastOp0->getSrcTy();
4647 Type *DestTy = CastOp0->getDestTy();
4648 Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy);
4649 Constant *Res2 = ConstantExpr::getCast(CastOp0->getOpcode(), Res1, DestTy);
4650
4651 // If the re-extended constant didn't change...
4652 if (Res2 == C) {
4653 if (ICmp.isEquality())
4654 return new ICmpInst(ICmp.getPredicate(), X, Res1);
4655
4656 // A signed comparison of sign extended values simplifies into a
4657 // signed comparison.
4658 if (IsSignedExt && IsSignedCmp)
4659 return new ICmpInst(ICmp.getPredicate(), X, Res1);
4660
4661 // The other three cases all fold into an unsigned comparison.
4662 return new ICmpInst(ICmp.getUnsignedPredicate(), X, Res1);
4663 }
4664
4665 // The re-extended constant changed, partly changed (in the case of a vector),
4666 // or could not be determined to be equal (in the case of a constant
4667 // expression), so the constant cannot be represented in the shorter type.
4668 // All the cases that fold to true or false will have already been handled
4669 // by SimplifyICmpInst, so only deal with the tricky case.
4670 if (IsSignedCmp || !IsSignedExt || !isa<ConstantInt>(C))
4671 return nullptr;
4672
4673 // Is source op positive?
4674 // icmp ult (sext X), C --> icmp sgt X, -1
4675 if (ICmp.getPredicate() == ICmpInst::ICMP_ULT)
4676 return new ICmpInst(CmpInst::ICMP_SGT, X, Constant::getAllOnesValue(SrcTy));
4677
4678 // Is source op negative?
4679 // icmp ugt (sext X), C --> icmp slt X, 0
4680 assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
4681 return new ICmpInst(CmpInst::ICMP_SLT, X, Constant::getNullValue(SrcTy));
4682 }
4683
4684 /// Handle icmp (cast x), (cast or constant).
foldICmpWithCastOp(ICmpInst & ICmp)4685 Instruction *InstCombinerImpl::foldICmpWithCastOp(ICmpInst &ICmp) {
4686 // If any operand of ICmp is a inttoptr roundtrip cast then remove it as
4687 // icmp compares only pointer's value.
4688 // icmp (inttoptr (ptrtoint p1)), p2 --> icmp p1, p2.
4689 Value *SimplifiedOp0 = simplifyIntToPtrRoundTripCast(ICmp.getOperand(0));
4690 Value *SimplifiedOp1 = simplifyIntToPtrRoundTripCast(ICmp.getOperand(1));
4691 if (SimplifiedOp0 || SimplifiedOp1)
4692 return new ICmpInst(ICmp.getPredicate(),
4693 SimplifiedOp0 ? SimplifiedOp0 : ICmp.getOperand(0),
4694 SimplifiedOp1 ? SimplifiedOp1 : ICmp.getOperand(1));
4695
4696 auto *CastOp0 = dyn_cast<CastInst>(ICmp.getOperand(0));
4697 if (!CastOp0)
4698 return nullptr;
4699 if (!isa<Constant>(ICmp.getOperand(1)) && !isa<CastInst>(ICmp.getOperand(1)))
4700 return nullptr;
4701
4702 Value *Op0Src = CastOp0->getOperand(0);
4703 Type *SrcTy = CastOp0->getSrcTy();
4704 Type *DestTy = CastOp0->getDestTy();
4705
4706 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
4707 // integer type is the same size as the pointer type.
4708 auto CompatibleSizes = [&](Type *SrcTy, Type *DestTy) {
4709 if (isa<VectorType>(SrcTy)) {
4710 SrcTy = cast<VectorType>(SrcTy)->getElementType();
4711 DestTy = cast<VectorType>(DestTy)->getElementType();
4712 }
4713 return DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth();
4714 };
4715 if (CastOp0->getOpcode() == Instruction::PtrToInt &&
4716 CompatibleSizes(SrcTy, DestTy)) {
4717 Value *NewOp1 = nullptr;
4718 if (auto *PtrToIntOp1 = dyn_cast<PtrToIntOperator>(ICmp.getOperand(1))) {
4719 Value *PtrSrc = PtrToIntOp1->getOperand(0);
4720 if (PtrSrc->getType()->getPointerAddressSpace() ==
4721 Op0Src->getType()->getPointerAddressSpace()) {
4722 NewOp1 = PtrToIntOp1->getOperand(0);
4723 // If the pointer types don't match, insert a bitcast.
4724 if (Op0Src->getType() != NewOp1->getType())
4725 NewOp1 = Builder.CreateBitCast(NewOp1, Op0Src->getType());
4726 }
4727 } else if (auto *RHSC = dyn_cast<Constant>(ICmp.getOperand(1))) {
4728 NewOp1 = ConstantExpr::getIntToPtr(RHSC, SrcTy);
4729 }
4730
4731 if (NewOp1)
4732 return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1);
4733 }
4734
4735 return foldICmpWithZextOrSext(ICmp, Builder);
4736 }
4737
isNeutralValue(Instruction::BinaryOps BinaryOp,Value * RHS)4738 static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS) {
4739 switch (BinaryOp) {
4740 default:
4741 llvm_unreachable("Unsupported binary op");
4742 case Instruction::Add:
4743 case Instruction::Sub:
4744 return match(RHS, m_Zero());
4745 case Instruction::Mul:
4746 return match(RHS, m_One());
4747 }
4748 }
4749
4750 OverflowResult
computeOverflow(Instruction::BinaryOps BinaryOp,bool IsSigned,Value * LHS,Value * RHS,Instruction * CxtI) const4751 InstCombinerImpl::computeOverflow(Instruction::BinaryOps BinaryOp,
4752 bool IsSigned, Value *LHS, Value *RHS,
4753 Instruction *CxtI) const {
4754 switch (BinaryOp) {
4755 default:
4756 llvm_unreachable("Unsupported binary op");
4757 case Instruction::Add:
4758 if (IsSigned)
4759 return computeOverflowForSignedAdd(LHS, RHS, CxtI);
4760 else
4761 return computeOverflowForUnsignedAdd(LHS, RHS, CxtI);
4762 case Instruction::Sub:
4763 if (IsSigned)
4764 return computeOverflowForSignedSub(LHS, RHS, CxtI);
4765 else
4766 return computeOverflowForUnsignedSub(LHS, RHS, CxtI);
4767 case Instruction::Mul:
4768 if (IsSigned)
4769 return computeOverflowForSignedMul(LHS, RHS, CxtI);
4770 else
4771 return computeOverflowForUnsignedMul(LHS, RHS, CxtI);
4772 }
4773 }
4774
OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp,bool IsSigned,Value * LHS,Value * RHS,Instruction & OrigI,Value * & Result,Constant * & Overflow)4775 bool InstCombinerImpl::OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp,
4776 bool IsSigned, Value *LHS,
4777 Value *RHS, Instruction &OrigI,
4778 Value *&Result,
4779 Constant *&Overflow) {
4780 if (OrigI.isCommutative() && isa<Constant>(LHS) && !isa<Constant>(RHS))
4781 std::swap(LHS, RHS);
4782
4783 // If the overflow check was an add followed by a compare, the insertion point
4784 // may be pointing to the compare. We want to insert the new instructions
4785 // before the add in case there are uses of the add between the add and the
4786 // compare.
4787 Builder.SetInsertPoint(&OrigI);
4788
4789 Type *OverflowTy = Type::getInt1Ty(LHS->getContext());
4790 if (auto *LHSTy = dyn_cast<VectorType>(LHS->getType()))
4791 OverflowTy = VectorType::get(OverflowTy, LHSTy->getElementCount());
4792
4793 if (isNeutralValue(BinaryOp, RHS)) {
4794 Result = LHS;
4795 Overflow = ConstantInt::getFalse(OverflowTy);
4796 return true;
4797 }
4798
4799 switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, &OrigI)) {
4800 case OverflowResult::MayOverflow:
4801 return false;
4802 case OverflowResult::AlwaysOverflowsLow:
4803 case OverflowResult::AlwaysOverflowsHigh:
4804 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4805 Result->takeName(&OrigI);
4806 Overflow = ConstantInt::getTrue(OverflowTy);
4807 return true;
4808 case OverflowResult::NeverOverflows:
4809 Result = Builder.CreateBinOp(BinaryOp, LHS, RHS);
4810 Result->takeName(&OrigI);
4811 Overflow = ConstantInt::getFalse(OverflowTy);
4812 if (auto *Inst = dyn_cast<Instruction>(Result)) {
4813 if (IsSigned)
4814 Inst->setHasNoSignedWrap();
4815 else
4816 Inst->setHasNoUnsignedWrap();
4817 }
4818 return true;
4819 }
4820
4821 llvm_unreachable("Unexpected overflow result");
4822 }
4823
4824 /// Recognize and process idiom involving test for multiplication
4825 /// overflow.
4826 ///
4827 /// The caller has matched a pattern of the form:
4828 /// I = cmp u (mul(zext A, zext B), V
4829 /// The function checks if this is a test for overflow and if so replaces
4830 /// multiplication with call to 'mul.with.overflow' intrinsic.
4831 ///
4832 /// \param I Compare instruction.
4833 /// \param MulVal Result of 'mult' instruction. It is one of the arguments of
4834 /// the compare instruction. Must be of integer type.
4835 /// \param OtherVal The other argument of compare instruction.
4836 /// \returns Instruction which must replace the compare instruction, NULL if no
4837 /// replacement required.
processUMulZExtIdiom(ICmpInst & I,Value * MulVal,Value * OtherVal,InstCombinerImpl & IC)4838 static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal,
4839 Value *OtherVal,
4840 InstCombinerImpl &IC) {
4841 // Don't bother doing this transformation for pointers, don't do it for
4842 // vectors.
4843 if (!isa<IntegerType>(MulVal->getType()))
4844 return nullptr;
4845
4846 assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal);
4847 assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal);
4848 auto *MulInstr = dyn_cast<Instruction>(MulVal);
4849 if (!MulInstr)
4850 return nullptr;
4851 assert(MulInstr->getOpcode() == Instruction::Mul);
4852
4853 auto *LHS = cast<ZExtOperator>(MulInstr->getOperand(0)),
4854 *RHS = cast<ZExtOperator>(MulInstr->getOperand(1));
4855 assert(LHS->getOpcode() == Instruction::ZExt);
4856 assert(RHS->getOpcode() == Instruction::ZExt);
4857 Value *A = LHS->getOperand(0), *B = RHS->getOperand(0);
4858
4859 // Calculate type and width of the result produced by mul.with.overflow.
4860 Type *TyA = A->getType(), *TyB = B->getType();
4861 unsigned WidthA = TyA->getPrimitiveSizeInBits(),
4862 WidthB = TyB->getPrimitiveSizeInBits();
4863 unsigned MulWidth;
4864 Type *MulType;
4865 if (WidthB > WidthA) {
4866 MulWidth = WidthB;
4867 MulType = TyB;
4868 } else {
4869 MulWidth = WidthA;
4870 MulType = TyA;
4871 }
4872
4873 // In order to replace the original mul with a narrower mul.with.overflow,
4874 // all uses must ignore upper bits of the product. The number of used low
4875 // bits must be not greater than the width of mul.with.overflow.
4876 if (MulVal->hasNUsesOrMore(2))
4877 for (User *U : MulVal->users()) {
4878 if (U == &I)
4879 continue;
4880 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4881 // Check if truncation ignores bits above MulWidth.
4882 unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits();
4883 if (TruncWidth > MulWidth)
4884 return nullptr;
4885 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
4886 // Check if AND ignores bits above MulWidth.
4887 if (BO->getOpcode() != Instruction::And)
4888 return nullptr;
4889 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4890 const APInt &CVal = CI->getValue();
4891 if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth)
4892 return nullptr;
4893 } else {
4894 // In this case we could have the operand of the binary operation
4895 // being defined in another block, and performing the replacement
4896 // could break the dominance relation.
4897 return nullptr;
4898 }
4899 } else {
4900 // Other uses prohibit this transformation.
4901 return nullptr;
4902 }
4903 }
4904
4905 // Recognize patterns
4906 switch (I.getPredicate()) {
4907 case ICmpInst::ICMP_EQ:
4908 case ICmpInst::ICMP_NE:
4909 // Recognize pattern:
4910 // mulval = mul(zext A, zext B)
4911 // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits.
4912 ConstantInt *CI;
4913 Value *ValToMask;
4914 if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) {
4915 if (ValToMask != MulVal)
4916 return nullptr;
4917 const APInt &CVal = CI->getValue() + 1;
4918 if (CVal.isPowerOf2()) {
4919 unsigned MaskWidth = CVal.logBase2();
4920 if (MaskWidth == MulWidth)
4921 break; // Recognized
4922 }
4923 }
4924 return nullptr;
4925
4926 case ICmpInst::ICMP_UGT:
4927 // Recognize pattern:
4928 // mulval = mul(zext A, zext B)
4929 // cmp ugt mulval, max
4930 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4931 APInt MaxVal = APInt::getMaxValue(MulWidth);
4932 MaxVal = MaxVal.zext(CI->getBitWidth());
4933 if (MaxVal.eq(CI->getValue()))
4934 break; // Recognized
4935 }
4936 return nullptr;
4937
4938 case ICmpInst::ICMP_UGE:
4939 // Recognize pattern:
4940 // mulval = mul(zext A, zext B)
4941 // cmp uge mulval, max+1
4942 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4943 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4944 if (MaxVal.eq(CI->getValue()))
4945 break; // Recognized
4946 }
4947 return nullptr;
4948
4949 case ICmpInst::ICMP_ULE:
4950 // Recognize pattern:
4951 // mulval = mul(zext A, zext B)
4952 // cmp ule mulval, max
4953 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4954 APInt MaxVal = APInt::getMaxValue(MulWidth);
4955 MaxVal = MaxVal.zext(CI->getBitWidth());
4956 if (MaxVal.eq(CI->getValue()))
4957 break; // Recognized
4958 }
4959 return nullptr;
4960
4961 case ICmpInst::ICMP_ULT:
4962 // Recognize pattern:
4963 // mulval = mul(zext A, zext B)
4964 // cmp ule mulval, max + 1
4965 if (ConstantInt *CI = dyn_cast<ConstantInt>(OtherVal)) {
4966 APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth);
4967 if (MaxVal.eq(CI->getValue()))
4968 break; // Recognized
4969 }
4970 return nullptr;
4971
4972 default:
4973 return nullptr;
4974 }
4975
4976 InstCombiner::BuilderTy &Builder = IC.Builder;
4977 Builder.SetInsertPoint(MulInstr);
4978
4979 // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B)
4980 Value *MulA = A, *MulB = B;
4981 if (WidthA < MulWidth)
4982 MulA = Builder.CreateZExt(A, MulType);
4983 if (WidthB < MulWidth)
4984 MulB = Builder.CreateZExt(B, MulType);
4985 Function *F = Intrinsic::getDeclaration(
4986 I.getModule(), Intrinsic::umul_with_overflow, MulType);
4987 CallInst *Call = Builder.CreateCall(F, {MulA, MulB}, "umul");
4988 IC.addToWorklist(MulInstr);
4989
4990 // If there are uses of mul result other than the comparison, we know that
4991 // they are truncation or binary AND. Change them to use result of
4992 // mul.with.overflow and adjust properly mask/size.
4993 if (MulVal->hasNUsesOrMore(2)) {
4994 Value *Mul = Builder.CreateExtractValue(Call, 0, "umul.value");
4995 for (User *U : make_early_inc_range(MulVal->users())) {
4996 if (U == &I || U == OtherVal)
4997 continue;
4998 if (TruncInst *TI = dyn_cast<TruncInst>(U)) {
4999 if (TI->getType()->getPrimitiveSizeInBits() == MulWidth)
5000 IC.replaceInstUsesWith(*TI, Mul);
5001 else
5002 TI->setOperand(0, Mul);
5003 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
5004 assert(BO->getOpcode() == Instruction::And);
5005 // Replace (mul & mask) --> zext (mul.with.overflow & short_mask)
5006 ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
5007 APInt ShortMask = CI->getValue().trunc(MulWidth);
5008 Value *ShortAnd = Builder.CreateAnd(Mul, ShortMask);
5009 Value *Zext = Builder.CreateZExt(ShortAnd, BO->getType());
5010 IC.replaceInstUsesWith(*BO, Zext);
5011 } else {
5012 llvm_unreachable("Unexpected Binary operation");
5013 }
5014 IC.addToWorklist(cast<Instruction>(U));
5015 }
5016 }
5017 if (isa<Instruction>(OtherVal))
5018 IC.addToWorklist(cast<Instruction>(OtherVal));
5019
5020 // The original icmp gets replaced with the overflow value, maybe inverted
5021 // depending on predicate.
5022 bool Inverse = false;
5023 switch (I.getPredicate()) {
5024 case ICmpInst::ICMP_NE:
5025 break;
5026 case ICmpInst::ICMP_EQ:
5027 Inverse = true;
5028 break;
5029 case ICmpInst::ICMP_UGT:
5030 case ICmpInst::ICMP_UGE:
5031 if (I.getOperand(0) == MulVal)
5032 break;
5033 Inverse = true;
5034 break;
5035 case ICmpInst::ICMP_ULT:
5036 case ICmpInst::ICMP_ULE:
5037 if (I.getOperand(1) == MulVal)
5038 break;
5039 Inverse = true;
5040 break;
5041 default:
5042 llvm_unreachable("Unexpected predicate");
5043 }
5044 if (Inverse) {
5045 Value *Res = Builder.CreateExtractValue(Call, 1);
5046 return BinaryOperator::CreateNot(Res);
5047 }
5048
5049 return ExtractValueInst::Create(Call, 1);
5050 }
5051
5052 /// When performing a comparison against a constant, it is possible that not all
5053 /// the bits in the LHS are demanded. This helper method computes the mask that
5054 /// IS demanded.
getDemandedBitsLHSMask(ICmpInst & I,unsigned BitWidth)5055 static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) {
5056 const APInt *RHS;
5057 if (!match(I.getOperand(1), m_APInt(RHS)))
5058 return APInt::getAllOnes(BitWidth);
5059
5060 // If this is a normal comparison, it demands all bits. If it is a sign bit
5061 // comparison, it only demands the sign bit.
5062 bool UnusedBit;
5063 if (InstCombiner::isSignBitCheck(I.getPredicate(), *RHS, UnusedBit))
5064 return APInt::getSignMask(BitWidth);
5065
5066 switch (I.getPredicate()) {
5067 // For a UGT comparison, we don't care about any bits that
5068 // correspond to the trailing ones of the comparand. The value of these
5069 // bits doesn't impact the outcome of the comparison, because any value
5070 // greater than the RHS must differ in a bit higher than these due to carry.
5071 case ICmpInst::ICMP_UGT:
5072 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingOnes());
5073
5074 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
5075 // Any value less than the RHS must differ in a higher bit because of carries.
5076 case ICmpInst::ICMP_ULT:
5077 return APInt::getBitsSetFrom(BitWidth, RHS->countTrailingZeros());
5078
5079 default:
5080 return APInt::getAllOnes(BitWidth);
5081 }
5082 }
5083
5084 /// Check if the order of \p Op0 and \p Op1 as operands in an ICmpInst
5085 /// should be swapped.
5086 /// The decision is based on how many times these two operands are reused
5087 /// as subtract operands and their positions in those instructions.
5088 /// The rationale is that several architectures use the same instruction for
5089 /// both subtract and cmp. Thus, it is better if the order of those operands
5090 /// match.
5091 /// \return true if Op0 and Op1 should be swapped.
swapMayExposeCSEOpportunities(const Value * Op0,const Value * Op1)5092 static bool swapMayExposeCSEOpportunities(const Value *Op0, const Value *Op1) {
5093 // Filter out pointer values as those cannot appear directly in subtract.
5094 // FIXME: we may want to go through inttoptrs or bitcasts.
5095 if (Op0->getType()->isPointerTy())
5096 return false;
5097 // If a subtract already has the same operands as a compare, swapping would be
5098 // bad. If a subtract has the same operands as a compare but in reverse order,
5099 // then swapping is good.
5100 int GoodToSwap = 0;
5101 for (const User *U : Op0->users()) {
5102 if (match(U, m_Sub(m_Specific(Op1), m_Specific(Op0))))
5103 GoodToSwap++;
5104 else if (match(U, m_Sub(m_Specific(Op0), m_Specific(Op1))))
5105 GoodToSwap--;
5106 }
5107 return GoodToSwap > 0;
5108 }
5109
5110 /// Check that one use is in the same block as the definition and all
5111 /// other uses are in blocks dominated by a given block.
5112 ///
5113 /// \param DI Definition
5114 /// \param UI Use
5115 /// \param DB Block that must dominate all uses of \p DI outside
5116 /// the parent block
5117 /// \return true when \p UI is the only use of \p DI in the parent block
5118 /// and all other uses of \p DI are in blocks dominated by \p DB.
5119 ///
dominatesAllUses(const Instruction * DI,const Instruction * UI,const BasicBlock * DB) const5120 bool InstCombinerImpl::dominatesAllUses(const Instruction *DI,
5121 const Instruction *UI,
5122 const BasicBlock *DB) const {
5123 assert(DI && UI && "Instruction not defined\n");
5124 // Ignore incomplete definitions.
5125 if (!DI->getParent())
5126 return false;
5127 // DI and UI must be in the same block.
5128 if (DI->getParent() != UI->getParent())
5129 return false;
5130 // Protect from self-referencing blocks.
5131 if (DI->getParent() == DB)
5132 return false;
5133 for (const User *U : DI->users()) {
5134 auto *Usr = cast<Instruction>(U);
5135 if (Usr != UI && !DT.dominates(DB, Usr->getParent()))
5136 return false;
5137 }
5138 return true;
5139 }
5140
5141 /// Return true when the instruction sequence within a block is select-cmp-br.
isChainSelectCmpBranch(const SelectInst * SI)5142 static bool isChainSelectCmpBranch(const SelectInst *SI) {
5143 const BasicBlock *BB = SI->getParent();
5144 if (!BB)
5145 return false;
5146 auto *BI = dyn_cast_or_null<BranchInst>(BB->getTerminator());
5147 if (!BI || BI->getNumSuccessors() != 2)
5148 return false;
5149 auto *IC = dyn_cast<ICmpInst>(BI->getCondition());
5150 if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI))
5151 return false;
5152 return true;
5153 }
5154
5155 /// True when a select result is replaced by one of its operands
5156 /// in select-icmp sequence. This will eventually result in the elimination
5157 /// of the select.
5158 ///
5159 /// \param SI Select instruction
5160 /// \param Icmp Compare instruction
5161 /// \param SIOpd Operand that replaces the select
5162 ///
5163 /// Notes:
5164 /// - The replacement is global and requires dominator information
5165 /// - The caller is responsible for the actual replacement
5166 ///
5167 /// Example:
5168 ///
5169 /// entry:
5170 /// %4 = select i1 %3, %C* %0, %C* null
5171 /// %5 = icmp eq %C* %4, null
5172 /// br i1 %5, label %9, label %7
5173 /// ...
5174 /// ; <label>:7 ; preds = %entry
5175 /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0
5176 /// ...
5177 ///
5178 /// can be transformed to
5179 ///
5180 /// %5 = icmp eq %C* %0, null
5181 /// %6 = select i1 %3, i1 %5, i1 true
5182 /// br i1 %6, label %9, label %7
5183 /// ...
5184 /// ; <label>:7 ; preds = %entry
5185 /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0!
5186 ///
5187 /// Similar when the first operand of the select is a constant or/and
5188 /// the compare is for not equal rather than equal.
5189 ///
5190 /// NOTE: The function is only called when the select and compare constants
5191 /// are equal, the optimization can work only for EQ predicates. This is not a
5192 /// major restriction since a NE compare should be 'normalized' to an equal
5193 /// compare, which usually happens in the combiner and test case
5194 /// select-cmp-br.ll checks for it.
replacedSelectWithOperand(SelectInst * SI,const ICmpInst * Icmp,const unsigned SIOpd)5195 bool InstCombinerImpl::replacedSelectWithOperand(SelectInst *SI,
5196 const ICmpInst *Icmp,
5197 const unsigned SIOpd) {
5198 assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!");
5199 if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) {
5200 BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(1);
5201 // The check for the single predecessor is not the best that can be
5202 // done. But it protects efficiently against cases like when SI's
5203 // home block has two successors, Succ and Succ1, and Succ1 predecessor
5204 // of Succ. Then SI can't be replaced by SIOpd because the use that gets
5205 // replaced can be reached on either path. So the uniqueness check
5206 // guarantees that the path all uses of SI (outside SI's parent) are on
5207 // is disjoint from all other paths out of SI. But that information
5208 // is more expensive to compute, and the trade-off here is in favor
5209 // of compile-time. It should also be noticed that we check for a single
5210 // predecessor and not only uniqueness. This to handle the situation when
5211 // Succ and Succ1 points to the same basic block.
5212 if (Succ->getSinglePredecessor() && dominatesAllUses(SI, Icmp, Succ)) {
5213 NumSel++;
5214 SI->replaceUsesOutsideBlock(SI->getOperand(SIOpd), SI->getParent());
5215 return true;
5216 }
5217 }
5218 return false;
5219 }
5220
5221 /// Try to fold the comparison based on range information we can get by checking
5222 /// whether bits are known to be zero or one in the inputs.
foldICmpUsingKnownBits(ICmpInst & I)5223 Instruction *InstCombinerImpl::foldICmpUsingKnownBits(ICmpInst &I) {
5224 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5225 Type *Ty = Op0->getType();
5226 ICmpInst::Predicate Pred = I.getPredicate();
5227
5228 // Get scalar or pointer size.
5229 unsigned BitWidth = Ty->isIntOrIntVectorTy()
5230 ? Ty->getScalarSizeInBits()
5231 : DL.getPointerTypeSizeInBits(Ty->getScalarType());
5232
5233 if (!BitWidth)
5234 return nullptr;
5235
5236 KnownBits Op0Known(BitWidth);
5237 KnownBits Op1Known(BitWidth);
5238
5239 if (SimplifyDemandedBits(&I, 0,
5240 getDemandedBitsLHSMask(I, BitWidth),
5241 Op0Known, 0))
5242 return &I;
5243
5244 if (SimplifyDemandedBits(&I, 1, APInt::getAllOnes(BitWidth), Op1Known, 0))
5245 return &I;
5246
5247 // Given the known and unknown bits, compute a range that the LHS could be
5248 // in. Compute the Min, Max and RHS values based on the known bits. For the
5249 // EQ and NE we use unsigned values.
5250 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
5251 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
5252 if (I.isSigned()) {
5253 Op0Min = Op0Known.getSignedMinValue();
5254 Op0Max = Op0Known.getSignedMaxValue();
5255 Op1Min = Op1Known.getSignedMinValue();
5256 Op1Max = Op1Known.getSignedMaxValue();
5257 } else {
5258 Op0Min = Op0Known.getMinValue();
5259 Op0Max = Op0Known.getMaxValue();
5260 Op1Min = Op1Known.getMinValue();
5261 Op1Max = Op1Known.getMaxValue();
5262 }
5263
5264 // If Min and Max are known to be the same, then SimplifyDemandedBits figured
5265 // out that the LHS or RHS is a constant. Constant fold this now, so that
5266 // code below can assume that Min != Max.
5267 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
5268 return new ICmpInst(Pred, ConstantExpr::getIntegerValue(Ty, Op0Min), Op1);
5269 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
5270 return new ICmpInst(Pred, Op0, ConstantExpr::getIntegerValue(Ty, Op1Min));
5271
5272 // Don't break up a clamp pattern -- (min(max X, Y), Z) -- by replacing a
5273 // min/max canonical compare with some other compare. That could lead to
5274 // conflict with select canonicalization and infinite looping.
5275 // FIXME: This constraint may go away if min/max intrinsics are canonical.
5276 auto isMinMaxCmp = [&](Instruction &Cmp) {
5277 if (!Cmp.hasOneUse())
5278 return false;
5279 Value *A, *B;
5280 SelectPatternFlavor SPF = matchSelectPattern(Cmp.user_back(), A, B).Flavor;
5281 if (!SelectPatternResult::isMinOrMax(SPF))
5282 return false;
5283 return match(Op0, m_MaxOrMin(m_Value(), m_Value())) ||
5284 match(Op1, m_MaxOrMin(m_Value(), m_Value()));
5285 };
5286 if (!isMinMaxCmp(I)) {
5287 switch (Pred) {
5288 default:
5289 break;
5290 case ICmpInst::ICMP_ULT: {
5291 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
5292 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5293 const APInt *CmpC;
5294 if (match(Op1, m_APInt(CmpC))) {
5295 // A <u C -> A == C-1 if min(A)+1 == C
5296 if (*CmpC == Op0Min + 1)
5297 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5298 ConstantInt::get(Op1->getType(), *CmpC - 1));
5299 // X <u C --> X == 0, if the number of zero bits in the bottom of X
5300 // exceeds the log2 of C.
5301 if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2())
5302 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5303 Constant::getNullValue(Op1->getType()));
5304 }
5305 break;
5306 }
5307 case ICmpInst::ICMP_UGT: {
5308 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
5309 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5310 const APInt *CmpC;
5311 if (match(Op1, m_APInt(CmpC))) {
5312 // A >u C -> A == C+1 if max(a)-1 == C
5313 if (*CmpC == Op0Max - 1)
5314 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5315 ConstantInt::get(Op1->getType(), *CmpC + 1));
5316 // X >u C --> X != 0, if the number of zero bits in the bottom of X
5317 // exceeds the log2 of C.
5318 if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits())
5319 return new ICmpInst(ICmpInst::ICMP_NE, Op0,
5320 Constant::getNullValue(Op1->getType()));
5321 }
5322 break;
5323 }
5324 case ICmpInst::ICMP_SLT: {
5325 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
5326 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5327 const APInt *CmpC;
5328 if (match(Op1, m_APInt(CmpC))) {
5329 if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C
5330 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5331 ConstantInt::get(Op1->getType(), *CmpC - 1));
5332 }
5333 break;
5334 }
5335 case ICmpInst::ICMP_SGT: {
5336 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
5337 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5338 const APInt *CmpC;
5339 if (match(Op1, m_APInt(CmpC))) {
5340 if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C
5341 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
5342 ConstantInt::get(Op1->getType(), *CmpC + 1));
5343 }
5344 break;
5345 }
5346 }
5347 }
5348
5349 // Based on the range information we know about the LHS, see if we can
5350 // simplify this comparison. For example, (x&4) < 8 is always true.
5351 switch (Pred) {
5352 default:
5353 llvm_unreachable("Unknown icmp opcode!");
5354 case ICmpInst::ICMP_EQ:
5355 case ICmpInst::ICMP_NE: {
5356 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
5357 return replaceInstUsesWith(
5358 I, ConstantInt::getBool(I.getType(), Pred == CmpInst::ICMP_NE));
5359
5360 // If all bits are known zero except for one, then we know at most one bit
5361 // is set. If the comparison is against zero, then this is a check to see if
5362 // *that* bit is set.
5363 APInt Op0KnownZeroInverted = ~Op0Known.Zero;
5364 if (Op1Known.isZero()) {
5365 // If the LHS is an AND with the same constant, look through it.
5366 Value *LHS = nullptr;
5367 const APInt *LHSC;
5368 if (!match(Op0, m_And(m_Value(LHS), m_APInt(LHSC))) ||
5369 *LHSC != Op0KnownZeroInverted)
5370 LHS = Op0;
5371
5372 Value *X;
5373 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
5374 APInt ValToCheck = Op0KnownZeroInverted;
5375 Type *XTy = X->getType();
5376 if (ValToCheck.isPowerOf2()) {
5377 // ((1 << X) & 8) == 0 -> X != 3
5378 // ((1 << X) & 8) != 0 -> X == 3
5379 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
5380 auto NewPred = ICmpInst::getInversePredicate(Pred);
5381 return new ICmpInst(NewPred, X, CmpC);
5382 } else if ((++ValToCheck).isPowerOf2()) {
5383 // ((1 << X) & 7) == 0 -> X >= 3
5384 // ((1 << X) & 7) != 0 -> X < 3
5385 auto *CmpC = ConstantInt::get(XTy, ValToCheck.countTrailingZeros());
5386 auto NewPred =
5387 Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT;
5388 return new ICmpInst(NewPred, X, CmpC);
5389 }
5390 }
5391
5392 // Check if the LHS is 8 >>u x and the result is a power of 2 like 1.
5393 const APInt *CI;
5394 if (Op0KnownZeroInverted.isOne() &&
5395 match(LHS, m_LShr(m_Power2(CI), m_Value(X)))) {
5396 // ((8 >>u X) & 1) == 0 -> X != 3
5397 // ((8 >>u X) & 1) != 0 -> X == 3
5398 unsigned CmpVal = CI->countTrailingZeros();
5399 auto NewPred = ICmpInst::getInversePredicate(Pred);
5400 return new ICmpInst(NewPred, X, ConstantInt::get(X->getType(), CmpVal));
5401 }
5402 }
5403 break;
5404 }
5405 case ICmpInst::ICMP_ULT: {
5406 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
5407 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5408 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
5409 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5410 break;
5411 }
5412 case ICmpInst::ICMP_UGT: {
5413 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
5414 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5415 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
5416 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5417 break;
5418 }
5419 case ICmpInst::ICMP_SLT: {
5420 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
5421 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5422 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
5423 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5424 break;
5425 }
5426 case ICmpInst::ICMP_SGT: {
5427 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
5428 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5429 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
5430 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5431 break;
5432 }
5433 case ICmpInst::ICMP_SGE:
5434 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
5435 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
5436 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5437 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
5438 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5439 if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B)
5440 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5441 break;
5442 case ICmpInst::ICMP_SLE:
5443 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
5444 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
5445 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5446 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
5447 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5448 if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B)
5449 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5450 break;
5451 case ICmpInst::ICMP_UGE:
5452 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
5453 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
5454 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5455 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
5456 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5457 if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B)
5458 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5459 break;
5460 case ICmpInst::ICMP_ULE:
5461 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
5462 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
5463 return replaceInstUsesWith(I, ConstantInt::getTrue(I.getType()));
5464 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
5465 return replaceInstUsesWith(I, ConstantInt::getFalse(I.getType()));
5466 if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B)
5467 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5468 break;
5469 }
5470
5471 // Turn a signed comparison into an unsigned one if both operands are known to
5472 // have the same sign.
5473 if (I.isSigned() &&
5474 ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) ||
5475 (Op0Known.One.isNegative() && Op1Known.One.isNegative())))
5476 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
5477
5478 return nullptr;
5479 }
5480
5481 llvm::Optional<std::pair<CmpInst::Predicate, Constant *>>
getFlippedStrictnessPredicateAndConstant(CmpInst::Predicate Pred,Constant * C)5482 InstCombiner::getFlippedStrictnessPredicateAndConstant(CmpInst::Predicate Pred,
5483 Constant *C) {
5484 assert(ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate(Pred) &&
5485 "Only for relational integer predicates.");
5486
5487 Type *Type = C->getType();
5488 bool IsSigned = ICmpInst::isSigned(Pred);
5489
5490 CmpInst::Predicate UnsignedPred = ICmpInst::getUnsignedPredicate(Pred);
5491 bool WillIncrement =
5492 UnsignedPred == ICmpInst::ICMP_ULE || UnsignedPred == ICmpInst::ICMP_UGT;
5493
5494 // Check if the constant operand can be safely incremented/decremented
5495 // without overflowing/underflowing.
5496 auto ConstantIsOk = [WillIncrement, IsSigned](ConstantInt *C) {
5497 return WillIncrement ? !C->isMaxValue(IsSigned) : !C->isMinValue(IsSigned);
5498 };
5499
5500 Constant *SafeReplacementConstant = nullptr;
5501 if (auto *CI = dyn_cast<ConstantInt>(C)) {
5502 // Bail out if the constant can't be safely incremented/decremented.
5503 if (!ConstantIsOk(CI))
5504 return llvm::None;
5505 } else if (auto *FVTy = dyn_cast<FixedVectorType>(Type)) {
5506 unsigned NumElts = FVTy->getNumElements();
5507 for (unsigned i = 0; i != NumElts; ++i) {
5508 Constant *Elt = C->getAggregateElement(i);
5509 if (!Elt)
5510 return llvm::None;
5511
5512 if (isa<UndefValue>(Elt))
5513 continue;
5514
5515 // Bail out if we can't determine if this constant is min/max or if we
5516 // know that this constant is min/max.
5517 auto *CI = dyn_cast<ConstantInt>(Elt);
5518 if (!CI || !ConstantIsOk(CI))
5519 return llvm::None;
5520
5521 if (!SafeReplacementConstant)
5522 SafeReplacementConstant = CI;
5523 }
5524 } else {
5525 // ConstantExpr?
5526 return llvm::None;
5527 }
5528
5529 // It may not be safe to change a compare predicate in the presence of
5530 // undefined elements, so replace those elements with the first safe constant
5531 // that we found.
5532 // TODO: in case of poison, it is safe; let's replace undefs only.
5533 if (C->containsUndefOrPoisonElement()) {
5534 assert(SafeReplacementConstant && "Replacement constant not set");
5535 C = Constant::replaceUndefsWith(C, SafeReplacementConstant);
5536 }
5537
5538 CmpInst::Predicate NewPred = CmpInst::getFlippedStrictnessPredicate(Pred);
5539
5540 // Increment or decrement the constant.
5541 Constant *OneOrNegOne = ConstantInt::get(Type, WillIncrement ? 1 : -1, true);
5542 Constant *NewC = ConstantExpr::getAdd(C, OneOrNegOne);
5543
5544 return std::make_pair(NewPred, NewC);
5545 }
5546
5547 /// If we have an icmp le or icmp ge instruction with a constant operand, turn
5548 /// it into the appropriate icmp lt or icmp gt instruction. This transform
5549 /// allows them to be folded in visitICmpInst.
canonicalizeCmpWithConstant(ICmpInst & I)5550 static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) {
5551 ICmpInst::Predicate Pred = I.getPredicate();
5552 if (ICmpInst::isEquality(Pred) || !ICmpInst::isIntPredicate(Pred) ||
5553 InstCombiner::isCanonicalPredicate(Pred))
5554 return nullptr;
5555
5556 Value *Op0 = I.getOperand(0);
5557 Value *Op1 = I.getOperand(1);
5558 auto *Op1C = dyn_cast<Constant>(Op1);
5559 if (!Op1C)
5560 return nullptr;
5561
5562 auto FlippedStrictness =
5563 InstCombiner::getFlippedStrictnessPredicateAndConstant(Pred, Op1C);
5564 if (!FlippedStrictness)
5565 return nullptr;
5566
5567 return new ICmpInst(FlippedStrictness->first, Op0, FlippedStrictness->second);
5568 }
5569
5570 /// If we have a comparison with a non-canonical predicate, if we can update
5571 /// all the users, invert the predicate and adjust all the users.
canonicalizeICmpPredicate(CmpInst & I)5572 CmpInst *InstCombinerImpl::canonicalizeICmpPredicate(CmpInst &I) {
5573 // Is the predicate already canonical?
5574 CmpInst::Predicate Pred = I.getPredicate();
5575 if (InstCombiner::isCanonicalPredicate(Pred))
5576 return nullptr;
5577
5578 // Can all users be adjusted to predicate inversion?
5579 if (!InstCombiner::canFreelyInvertAllUsersOf(&I, /*IgnoredUser=*/nullptr))
5580 return nullptr;
5581
5582 // Ok, we can canonicalize comparison!
5583 // Let's first invert the comparison's predicate.
5584 I.setPredicate(CmpInst::getInversePredicate(Pred));
5585 I.setName(I.getName() + ".not");
5586
5587 // And, adapt users.
5588 freelyInvertAllUsersOf(&I);
5589
5590 return &I;
5591 }
5592
5593 /// Integer compare with boolean values can always be turned into bitwise ops.
canonicalizeICmpBool(ICmpInst & I,InstCombiner::BuilderTy & Builder)5594 static Instruction *canonicalizeICmpBool(ICmpInst &I,
5595 InstCombiner::BuilderTy &Builder) {
5596 Value *A = I.getOperand(0), *B = I.getOperand(1);
5597 assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only");
5598
5599 // A boolean compared to true/false can be simplified to Op0/true/false in
5600 // 14 out of the 20 (10 predicates * 2 constants) possible combinations.
5601 // Cases not handled by InstSimplify are always 'not' of Op0.
5602 if (match(B, m_Zero())) {
5603 switch (I.getPredicate()) {
5604 case CmpInst::ICMP_EQ: // A == 0 -> !A
5605 case CmpInst::ICMP_ULE: // A <=u 0 -> !A
5606 case CmpInst::ICMP_SGE: // A >=s 0 -> !A
5607 return BinaryOperator::CreateNot(A);
5608 default:
5609 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
5610 }
5611 } else if (match(B, m_One())) {
5612 switch (I.getPredicate()) {
5613 case CmpInst::ICMP_NE: // A != 1 -> !A
5614 case CmpInst::ICMP_ULT: // A <u 1 -> !A
5615 case CmpInst::ICMP_SGT: // A >s -1 -> !A
5616 return BinaryOperator::CreateNot(A);
5617 default:
5618 llvm_unreachable("ICmp i1 X, C not simplified as expected.");
5619 }
5620 }
5621
5622 switch (I.getPredicate()) {
5623 default:
5624 llvm_unreachable("Invalid icmp instruction!");
5625 case ICmpInst::ICMP_EQ:
5626 // icmp eq i1 A, B -> ~(A ^ B)
5627 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
5628
5629 case ICmpInst::ICMP_NE:
5630 // icmp ne i1 A, B -> A ^ B
5631 return BinaryOperator::CreateXor(A, B);
5632
5633 case ICmpInst::ICMP_UGT:
5634 // icmp ugt -> icmp ult
5635 std::swap(A, B);
5636 LLVM_FALLTHROUGH;
5637 case ICmpInst::ICMP_ULT:
5638 // icmp ult i1 A, B -> ~A & B
5639 return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
5640
5641 case ICmpInst::ICMP_SGT:
5642 // icmp sgt -> icmp slt
5643 std::swap(A, B);
5644 LLVM_FALLTHROUGH;
5645 case ICmpInst::ICMP_SLT:
5646 // icmp slt i1 A, B -> A & ~B
5647 return BinaryOperator::CreateAnd(Builder.CreateNot(B), A);
5648
5649 case ICmpInst::ICMP_UGE:
5650 // icmp uge -> icmp ule
5651 std::swap(A, B);
5652 LLVM_FALLTHROUGH;
5653 case ICmpInst::ICMP_ULE:
5654 // icmp ule i1 A, B -> ~A | B
5655 return BinaryOperator::CreateOr(Builder.CreateNot(A), B);
5656
5657 case ICmpInst::ICMP_SGE:
5658 // icmp sge -> icmp sle
5659 std::swap(A, B);
5660 LLVM_FALLTHROUGH;
5661 case ICmpInst::ICMP_SLE:
5662 // icmp sle i1 A, B -> A | ~B
5663 return BinaryOperator::CreateOr(Builder.CreateNot(B), A);
5664 }
5665 }
5666
5667 // Transform pattern like:
5668 // (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X
5669 // (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X
5670 // Into:
5671 // (X l>> Y) != 0
5672 // (X l>> Y) == 0
foldICmpWithHighBitMask(ICmpInst & Cmp,InstCombiner::BuilderTy & Builder)5673 static Instruction *foldICmpWithHighBitMask(ICmpInst &Cmp,
5674 InstCombiner::BuilderTy &Builder) {
5675 ICmpInst::Predicate Pred, NewPred;
5676 Value *X, *Y;
5677 if (match(&Cmp,
5678 m_c_ICmp(Pred, m_OneUse(m_Shl(m_One(), m_Value(Y))), m_Value(X)))) {
5679 switch (Pred) {
5680 case ICmpInst::ICMP_ULE:
5681 NewPred = ICmpInst::ICMP_NE;
5682 break;
5683 case ICmpInst::ICMP_UGT:
5684 NewPred = ICmpInst::ICMP_EQ;
5685 break;
5686 default:
5687 return nullptr;
5688 }
5689 } else if (match(&Cmp, m_c_ICmp(Pred,
5690 m_OneUse(m_CombineOr(
5691 m_Not(m_Shl(m_AllOnes(), m_Value(Y))),
5692 m_Add(m_Shl(m_One(), m_Value(Y)),
5693 m_AllOnes()))),
5694 m_Value(X)))) {
5695 // The variant with 'add' is not canonical, (the variant with 'not' is)
5696 // we only get it because it has extra uses, and can't be canonicalized,
5697
5698 switch (Pred) {
5699 case ICmpInst::ICMP_ULT:
5700 NewPred = ICmpInst::ICMP_NE;
5701 break;
5702 case ICmpInst::ICMP_UGE:
5703 NewPred = ICmpInst::ICMP_EQ;
5704 break;
5705 default:
5706 return nullptr;
5707 }
5708 } else
5709 return nullptr;
5710
5711 Value *NewX = Builder.CreateLShr(X, Y, X->getName() + ".highbits");
5712 Constant *Zero = Constant::getNullValue(NewX->getType());
5713 return CmpInst::Create(Instruction::ICmp, NewPred, NewX, Zero);
5714 }
5715
foldVectorCmp(CmpInst & Cmp,InstCombiner::BuilderTy & Builder)5716 static Instruction *foldVectorCmp(CmpInst &Cmp,
5717 InstCombiner::BuilderTy &Builder) {
5718 const CmpInst::Predicate Pred = Cmp.getPredicate();
5719 Value *LHS = Cmp.getOperand(0), *RHS = Cmp.getOperand(1);
5720 Value *V1, *V2;
5721 ArrayRef<int> M;
5722 if (!match(LHS, m_Shuffle(m_Value(V1), m_Undef(), m_Mask(M))))
5723 return nullptr;
5724
5725 // If both arguments of the cmp are shuffles that use the same mask and
5726 // shuffle within a single vector, move the shuffle after the cmp:
5727 // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M
5728 Type *V1Ty = V1->getType();
5729 if (match(RHS, m_Shuffle(m_Value(V2), m_Undef(), m_SpecificMask(M))) &&
5730 V1Ty == V2->getType() && (LHS->hasOneUse() || RHS->hasOneUse())) {
5731 Value *NewCmp = Builder.CreateCmp(Pred, V1, V2);
5732 return new ShuffleVectorInst(NewCmp, M);
5733 }
5734
5735 // Try to canonicalize compare with splatted operand and splat constant.
5736 // TODO: We could generalize this for more than splats. See/use the code in
5737 // InstCombiner::foldVectorBinop().
5738 Constant *C;
5739 if (!LHS->hasOneUse() || !match(RHS, m_Constant(C)))
5740 return nullptr;
5741
5742 // Length-changing splats are ok, so adjust the constants as needed:
5743 // cmp (shuffle V1, M), C --> shuffle (cmp V1, C'), M
5744 Constant *ScalarC = C->getSplatValue(/* AllowUndefs */ true);
5745 int MaskSplatIndex;
5746 if (ScalarC && match(M, m_SplatOrUndefMask(MaskSplatIndex))) {
5747 // We allow undefs in matching, but this transform removes those for safety.
5748 // Demanded elements analysis should be able to recover some/all of that.
5749 C = ConstantVector::getSplat(cast<VectorType>(V1Ty)->getElementCount(),
5750 ScalarC);
5751 SmallVector<int, 8> NewM(M.size(), MaskSplatIndex);
5752 Value *NewCmp = Builder.CreateCmp(Pred, V1, C);
5753 return new ShuffleVectorInst(NewCmp, NewM);
5754 }
5755
5756 return nullptr;
5757 }
5758
5759 // extract(uadd.with.overflow(A, B), 0) ult A
5760 // -> extract(uadd.with.overflow(A, B), 1)
foldICmpOfUAddOv(ICmpInst & I)5761 static Instruction *foldICmpOfUAddOv(ICmpInst &I) {
5762 CmpInst::Predicate Pred = I.getPredicate();
5763 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5764
5765 Value *UAddOv;
5766 Value *A, *B;
5767 auto UAddOvResultPat = m_ExtractValue<0>(
5768 m_Intrinsic<Intrinsic::uadd_with_overflow>(m_Value(A), m_Value(B)));
5769 if (match(Op0, UAddOvResultPat) &&
5770 ((Pred == ICmpInst::ICMP_ULT && (Op1 == A || Op1 == B)) ||
5771 (Pred == ICmpInst::ICMP_EQ && match(Op1, m_ZeroInt()) &&
5772 (match(A, m_One()) || match(B, m_One()))) ||
5773 (Pred == ICmpInst::ICMP_NE && match(Op1, m_AllOnes()) &&
5774 (match(A, m_AllOnes()) || match(B, m_AllOnes())))))
5775 // extract(uadd.with.overflow(A, B), 0) < A
5776 // extract(uadd.with.overflow(A, 1), 0) == 0
5777 // extract(uadd.with.overflow(A, -1), 0) != -1
5778 UAddOv = cast<ExtractValueInst>(Op0)->getAggregateOperand();
5779 else if (match(Op1, UAddOvResultPat) &&
5780 Pred == ICmpInst::ICMP_UGT && (Op0 == A || Op0 == B))
5781 // A > extract(uadd.with.overflow(A, B), 0)
5782 UAddOv = cast<ExtractValueInst>(Op1)->getAggregateOperand();
5783 else
5784 return nullptr;
5785
5786 return ExtractValueInst::Create(UAddOv, 1);
5787 }
5788
foldICmpInvariantGroup(ICmpInst & I)5789 static Instruction *foldICmpInvariantGroup(ICmpInst &I) {
5790 if (!I.getOperand(0)->getType()->isPointerTy() ||
5791 NullPointerIsDefined(
5792 I.getParent()->getParent(),
5793 I.getOperand(0)->getType()->getPointerAddressSpace())) {
5794 return nullptr;
5795 }
5796 Instruction *Op;
5797 if (match(I.getOperand(0), m_Instruction(Op)) &&
5798 match(I.getOperand(1), m_Zero()) &&
5799 Op->isLaunderOrStripInvariantGroup()) {
5800 return ICmpInst::Create(Instruction::ICmp, I.getPredicate(),
5801 Op->getOperand(0), I.getOperand(1));
5802 }
5803 return nullptr;
5804 }
5805
visitICmpInst(ICmpInst & I)5806 Instruction *InstCombinerImpl::visitICmpInst(ICmpInst &I) {
5807 bool Changed = false;
5808 const SimplifyQuery Q = SQ.getWithInstruction(&I);
5809 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5810 unsigned Op0Cplxity = getComplexity(Op0);
5811 unsigned Op1Cplxity = getComplexity(Op1);
5812
5813 /// Orders the operands of the compare so that they are listed from most
5814 /// complex to least complex. This puts constants before unary operators,
5815 /// before binary operators.
5816 if (Op0Cplxity < Op1Cplxity ||
5817 (Op0Cplxity == Op1Cplxity && swapMayExposeCSEOpportunities(Op0, Op1))) {
5818 I.swapOperands();
5819 std::swap(Op0, Op1);
5820 Changed = true;
5821 }
5822
5823 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, Q))
5824 return replaceInstUsesWith(I, V);
5825
5826 // Comparing -val or val with non-zero is the same as just comparing val
5827 // ie, abs(val) != 0 -> val != 0
5828 if (I.getPredicate() == ICmpInst::ICMP_NE && match(Op1, m_Zero())) {
5829 Value *Cond, *SelectTrue, *SelectFalse;
5830 if (match(Op0, m_Select(m_Value(Cond), m_Value(SelectTrue),
5831 m_Value(SelectFalse)))) {
5832 if (Value *V = dyn_castNegVal(SelectTrue)) {
5833 if (V == SelectFalse)
5834 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
5835 }
5836 else if (Value *V = dyn_castNegVal(SelectFalse)) {
5837 if (V == SelectTrue)
5838 return CmpInst::Create(Instruction::ICmp, I.getPredicate(), V, Op1);
5839 }
5840 }
5841 }
5842
5843 if (Op0->getType()->isIntOrIntVectorTy(1))
5844 if (Instruction *Res = canonicalizeICmpBool(I, Builder))
5845 return Res;
5846
5847 if (Instruction *Res = canonicalizeCmpWithConstant(I))
5848 return Res;
5849
5850 if (Instruction *Res = canonicalizeICmpPredicate(I))
5851 return Res;
5852
5853 if (Instruction *Res = foldICmpWithConstant(I))
5854 return Res;
5855
5856 if (Instruction *Res = foldICmpWithDominatingICmp(I))
5857 return Res;
5858
5859 if (Instruction *Res = foldICmpUsingKnownBits(I))
5860 return Res;
5861
5862 // Test if the ICmpInst instruction is used exclusively by a select as
5863 // part of a minimum or maximum operation. If so, refrain from doing
5864 // any other folding. This helps out other analyses which understand
5865 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5866 // and CodeGen. And in this case, at least one of the comparison
5867 // operands has at least one user besides the compare (the select),
5868 // which would often largely negate the benefit of folding anyway.
5869 //
5870 // Do the same for the other patterns recognized by matchSelectPattern.
5871 if (I.hasOneUse())
5872 if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
5873 Value *A, *B;
5874 SelectPatternResult SPR = matchSelectPattern(SI, A, B);
5875 if (SPR.Flavor != SPF_UNKNOWN)
5876 return nullptr;
5877 }
5878
5879 // Do this after checking for min/max to prevent infinite looping.
5880 if (Instruction *Res = foldICmpWithZero(I))
5881 return Res;
5882
5883 // FIXME: We only do this after checking for min/max to prevent infinite
5884 // looping caused by a reverse canonicalization of these patterns for min/max.
5885 // FIXME: The organization of folds is a mess. These would naturally go into
5886 // canonicalizeCmpWithConstant(), but we can't move all of the above folds
5887 // down here after the min/max restriction.
5888 ICmpInst::Predicate Pred = I.getPredicate();
5889 const APInt *C;
5890 if (match(Op1, m_APInt(C))) {
5891 // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set
5892 if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) {
5893 Constant *Zero = Constant::getNullValue(Op0->getType());
5894 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero);
5895 }
5896
5897 // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear
5898 if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) {
5899 Constant *AllOnes = Constant::getAllOnesValue(Op0->getType());
5900 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes);
5901 }
5902 }
5903
5904 // The folds in here may rely on wrapping flags and special constants, so
5905 // they can break up min/max idioms in some cases but not seemingly similar
5906 // patterns.
5907 // FIXME: It may be possible to enhance select folding to make this
5908 // unnecessary. It may also be moot if we canonicalize to min/max
5909 // intrinsics.
5910 if (Instruction *Res = foldICmpBinOp(I, Q))
5911 return Res;
5912
5913 if (Instruction *Res = foldICmpInstWithConstant(I))
5914 return Res;
5915
5916 // Try to match comparison as a sign bit test. Intentionally do this after
5917 // foldICmpInstWithConstant() to potentially let other folds to happen first.
5918 if (Instruction *New = foldSignBitTest(I))
5919 return New;
5920
5921 if (Instruction *Res = foldICmpInstWithConstantNotInt(I))
5922 return Res;
5923
5924 // Try to optimize 'icmp GEP, P' or 'icmp P, GEP'.
5925 if (auto *GEP = dyn_cast<GEPOperator>(Op0))
5926 if (Instruction *NI = foldGEPICmp(GEP, Op1, I.getPredicate(), I))
5927 return NI;
5928 if (auto *GEP = dyn_cast<GEPOperator>(Op1))
5929 if (Instruction *NI = foldGEPICmp(GEP, Op0, I.getSwappedPredicate(), I))
5930 return NI;
5931
5932 // Try to optimize equality comparisons against alloca-based pointers.
5933 if (Op0->getType()->isPointerTy() && I.isEquality()) {
5934 assert(Op1->getType()->isPointerTy() && "Comparing pointer with non-pointer?");
5935 if (auto *Alloca = dyn_cast<AllocaInst>(getUnderlyingObject(Op0)))
5936 if (Instruction *New = foldAllocaCmp(I, Alloca, Op1))
5937 return New;
5938 if (auto *Alloca = dyn_cast<AllocaInst>(getUnderlyingObject(Op1)))
5939 if (Instruction *New = foldAllocaCmp(I, Alloca, Op0))
5940 return New;
5941 }
5942
5943 if (Instruction *Res = foldICmpBitCast(I))
5944 return Res;
5945
5946 // TODO: Hoist this above the min/max bailout.
5947 if (Instruction *R = foldICmpWithCastOp(I))
5948 return R;
5949
5950 if (Instruction *Res = foldICmpWithMinMax(I))
5951 return Res;
5952
5953 {
5954 Value *A, *B;
5955 // Transform (A & ~B) == 0 --> (A & B) != 0
5956 // and (A & ~B) != 0 --> (A & B) == 0
5957 // if A is a power of 2.
5958 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
5959 match(Op1, m_Zero()) &&
5960 isKnownToBeAPowerOfTwo(A, false, 0, &I) && I.isEquality())
5961 return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(A, B),
5962 Op1);
5963
5964 // ~X < ~Y --> Y < X
5965 // ~X < C --> X > ~C
5966 if (match(Op0, m_Not(m_Value(A)))) {
5967 if (match(Op1, m_Not(m_Value(B))))
5968 return new ICmpInst(I.getPredicate(), B, A);
5969
5970 const APInt *C;
5971 if (match(Op1, m_APInt(C)))
5972 return new ICmpInst(I.getSwappedPredicate(), A,
5973 ConstantInt::get(Op1->getType(), ~(*C)));
5974 }
5975
5976 Instruction *AddI = nullptr;
5977 if (match(&I, m_UAddWithOverflow(m_Value(A), m_Value(B),
5978 m_Instruction(AddI))) &&
5979 isa<IntegerType>(A->getType())) {
5980 Value *Result;
5981 Constant *Overflow;
5982 // m_UAddWithOverflow can match patterns that do not include an explicit
5983 // "add" instruction, so check the opcode of the matched op.
5984 if (AddI->getOpcode() == Instruction::Add &&
5985 OptimizeOverflowCheck(Instruction::Add, /*Signed*/ false, A, B, *AddI,
5986 Result, Overflow)) {
5987 replaceInstUsesWith(*AddI, Result);
5988 eraseInstFromFunction(*AddI);
5989 return replaceInstUsesWith(I, Overflow);
5990 }
5991 }
5992
5993 // (zext a) * (zext b) --> llvm.umul.with.overflow.
5994 if (match(Op0, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
5995 if (Instruction *R = processUMulZExtIdiom(I, Op0, Op1, *this))
5996 return R;
5997 }
5998 if (match(Op1, m_Mul(m_ZExt(m_Value(A)), m_ZExt(m_Value(B))))) {
5999 if (Instruction *R = processUMulZExtIdiom(I, Op1, Op0, *this))
6000 return R;
6001 }
6002 }
6003
6004 if (Instruction *Res = foldICmpEquality(I))
6005 return Res;
6006
6007 if (Instruction *Res = foldICmpOfUAddOv(I))
6008 return Res;
6009
6010 // The 'cmpxchg' instruction returns an aggregate containing the old value and
6011 // an i1 which indicates whether or not we successfully did the swap.
6012 //
6013 // Replace comparisons between the old value and the expected value with the
6014 // indicator that 'cmpxchg' returns.
6015 //
6016 // N.B. This transform is only valid when the 'cmpxchg' is not permitted to
6017 // spuriously fail. In those cases, the old value may equal the expected
6018 // value but it is possible for the swap to not occur.
6019 if (I.getPredicate() == ICmpInst::ICMP_EQ)
6020 if (auto *EVI = dyn_cast<ExtractValueInst>(Op0))
6021 if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(EVI->getAggregateOperand()))
6022 if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 &&
6023 !ACXI->isWeak())
6024 return ExtractValueInst::Create(ACXI, 1);
6025
6026 {
6027 Value *X;
6028 const APInt *C;
6029 // icmp X+Cst, X
6030 if (match(Op0, m_Add(m_Value(X), m_APInt(C))) && Op1 == X)
6031 return foldICmpAddOpConst(X, *C, I.getPredicate());
6032
6033 // icmp X, X+Cst
6034 if (match(Op1, m_Add(m_Value(X), m_APInt(C))) && Op0 == X)
6035 return foldICmpAddOpConst(X, *C, I.getSwappedPredicate());
6036 }
6037
6038 if (Instruction *Res = foldICmpWithHighBitMask(I, Builder))
6039 return Res;
6040
6041 if (I.getType()->isVectorTy())
6042 if (Instruction *Res = foldVectorCmp(I, Builder))
6043 return Res;
6044
6045 if (Instruction *Res = foldICmpInvariantGroup(I))
6046 return Res;
6047
6048 return Changed ? &I : nullptr;
6049 }
6050
6051 /// Fold fcmp ([us]itofp x, cst) if possible.
foldFCmpIntToFPConst(FCmpInst & I,Instruction * LHSI,Constant * RHSC)6052 Instruction *InstCombinerImpl::foldFCmpIntToFPConst(FCmpInst &I,
6053 Instruction *LHSI,
6054 Constant *RHSC) {
6055 if (!isa<ConstantFP>(RHSC)) return nullptr;
6056 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
6057
6058 // Get the width of the mantissa. We don't want to hack on conversions that
6059 // might lose information from the integer, e.g. "i64 -> float"
6060 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
6061 if (MantissaWidth == -1) return nullptr; // Unknown.
6062
6063 IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
6064
6065 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
6066
6067 if (I.isEquality()) {
6068 FCmpInst::Predicate P = I.getPredicate();
6069 bool IsExact = false;
6070 APSInt RHSCvt(IntTy->getBitWidth(), LHSUnsigned);
6071 RHS.convertToInteger(RHSCvt, APFloat::rmNearestTiesToEven, &IsExact);
6072
6073 // If the floating point constant isn't an integer value, we know if we will
6074 // ever compare equal / not equal to it.
6075 if (!IsExact) {
6076 // TODO: Can never be -0.0 and other non-representable values
6077 APFloat RHSRoundInt(RHS);
6078 RHSRoundInt.roundToIntegral(APFloat::rmNearestTiesToEven);
6079 if (RHS != RHSRoundInt) {
6080 if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ)
6081 return replaceInstUsesWith(I, Builder.getFalse());
6082
6083 assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE);
6084 return replaceInstUsesWith(I, Builder.getTrue());
6085 }
6086 }
6087
6088 // TODO: If the constant is exactly representable, is it always OK to do
6089 // equality compares as integer?
6090 }
6091
6092 // Check to see that the input is converted from an integer type that is small
6093 // enough that preserves all bits. TODO: check here for "known" sign bits.
6094 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
6095 unsigned InputSize = IntTy->getScalarSizeInBits();
6096
6097 // Following test does NOT adjust InputSize downwards for signed inputs,
6098 // because the most negative value still requires all the mantissa bits
6099 // to distinguish it from one less than that value.
6100 if ((int)InputSize > MantissaWidth) {
6101 // Conversion would lose accuracy. Check if loss can impact comparison.
6102 int Exp = ilogb(RHS);
6103 if (Exp == APFloat::IEK_Inf) {
6104 int MaxExponent = ilogb(APFloat::getLargest(RHS.getSemantics()));
6105 if (MaxExponent < (int)InputSize - !LHSUnsigned)
6106 // Conversion could create infinity.
6107 return nullptr;
6108 } else {
6109 // Note that if RHS is zero or NaN, then Exp is negative
6110 // and first condition is trivially false.
6111 if (MantissaWidth <= Exp && Exp <= (int)InputSize - !LHSUnsigned)
6112 // Conversion could affect comparison.
6113 return nullptr;
6114 }
6115 }
6116
6117 // Otherwise, we can potentially simplify the comparison. We know that it
6118 // will always come through as an integer value and we know the constant is
6119 // not a NAN (it would have been previously simplified).
6120 assert(!RHS.isNaN() && "NaN comparison not already folded!");
6121
6122 ICmpInst::Predicate Pred;
6123 switch (I.getPredicate()) {
6124 default: llvm_unreachable("Unexpected predicate!");
6125 case FCmpInst::FCMP_UEQ:
6126 case FCmpInst::FCMP_OEQ:
6127 Pred = ICmpInst::ICMP_EQ;
6128 break;
6129 case FCmpInst::FCMP_UGT:
6130 case FCmpInst::FCMP_OGT:
6131 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
6132 break;
6133 case FCmpInst::FCMP_UGE:
6134 case FCmpInst::FCMP_OGE:
6135 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
6136 break;
6137 case FCmpInst::FCMP_ULT:
6138 case FCmpInst::FCMP_OLT:
6139 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
6140 break;
6141 case FCmpInst::FCMP_ULE:
6142 case FCmpInst::FCMP_OLE:
6143 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
6144 break;
6145 case FCmpInst::FCMP_UNE:
6146 case FCmpInst::FCMP_ONE:
6147 Pred = ICmpInst::ICMP_NE;
6148 break;
6149 case FCmpInst::FCMP_ORD:
6150 return replaceInstUsesWith(I, Builder.getTrue());
6151 case FCmpInst::FCMP_UNO:
6152 return replaceInstUsesWith(I, Builder.getFalse());
6153 }
6154
6155 // Now we know that the APFloat is a normal number, zero or inf.
6156
6157 // See if the FP constant is too large for the integer. For example,
6158 // comparing an i8 to 300.0.
6159 unsigned IntWidth = IntTy->getScalarSizeInBits();
6160
6161 if (!LHSUnsigned) {
6162 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
6163 // and large values.
6164 APFloat SMax(RHS.getSemantics());
6165 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
6166 APFloat::rmNearestTiesToEven);
6167 if (SMax < RHS) { // smax < 13123.0
6168 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
6169 Pred == ICmpInst::ICMP_SLE)
6170 return replaceInstUsesWith(I, Builder.getTrue());
6171 return replaceInstUsesWith(I, Builder.getFalse());
6172 }
6173 } else {
6174 // If the RHS value is > UnsignedMax, fold the comparison. This handles
6175 // +INF and large values.
6176 APFloat UMax(RHS.getSemantics());
6177 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
6178 APFloat::rmNearestTiesToEven);
6179 if (UMax < RHS) { // umax < 13123.0
6180 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
6181 Pred == ICmpInst::ICMP_ULE)
6182 return replaceInstUsesWith(I, Builder.getTrue());
6183 return replaceInstUsesWith(I, Builder.getFalse());
6184 }
6185 }
6186
6187 if (!LHSUnsigned) {
6188 // See if the RHS value is < SignedMin.
6189 APFloat SMin(RHS.getSemantics());
6190 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
6191 APFloat::rmNearestTiesToEven);
6192 if (SMin > RHS) { // smin > 12312.0
6193 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
6194 Pred == ICmpInst::ICMP_SGE)
6195 return replaceInstUsesWith(I, Builder.getTrue());
6196 return replaceInstUsesWith(I, Builder.getFalse());
6197 }
6198 } else {
6199 // See if the RHS value is < UnsignedMin.
6200 APFloat UMin(RHS.getSemantics());
6201 UMin.convertFromAPInt(APInt::getMinValue(IntWidth), false,
6202 APFloat::rmNearestTiesToEven);
6203 if (UMin > RHS) { // umin > 12312.0
6204 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT ||
6205 Pred == ICmpInst::ICMP_UGE)
6206 return replaceInstUsesWith(I, Builder.getTrue());
6207 return replaceInstUsesWith(I, Builder.getFalse());
6208 }
6209 }
6210
6211 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
6212 // [0, UMAX], but it may still be fractional. See if it is fractional by
6213 // casting the FP value to the integer value and back, checking for equality.
6214 // Don't do this for zero, because -0.0 is not fractional.
6215 Constant *RHSInt = LHSUnsigned
6216 ? ConstantExpr::getFPToUI(RHSC, IntTy)
6217 : ConstantExpr::getFPToSI(RHSC, IntTy);
6218 if (!RHS.isZero()) {
6219 bool Equal = LHSUnsigned
6220 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
6221 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
6222 if (!Equal) {
6223 // If we had a comparison against a fractional value, we have to adjust
6224 // the compare predicate and sometimes the value. RHSC is rounded towards
6225 // zero at this point.
6226 switch (Pred) {
6227 default: llvm_unreachable("Unexpected integer comparison!");
6228 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
6229 return replaceInstUsesWith(I, Builder.getTrue());
6230 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
6231 return replaceInstUsesWith(I, Builder.getFalse());
6232 case ICmpInst::ICMP_ULE:
6233 // (float)int <= 4.4 --> int <= 4
6234 // (float)int <= -4.4 --> false
6235 if (RHS.isNegative())
6236 return replaceInstUsesWith(I, Builder.getFalse());
6237 break;
6238 case ICmpInst::ICMP_SLE:
6239 // (float)int <= 4.4 --> int <= 4
6240 // (float)int <= -4.4 --> int < -4
6241 if (RHS.isNegative())
6242 Pred = ICmpInst::ICMP_SLT;
6243 break;
6244 case ICmpInst::ICMP_ULT:
6245 // (float)int < -4.4 --> false
6246 // (float)int < 4.4 --> int <= 4
6247 if (RHS.isNegative())
6248 return replaceInstUsesWith(I, Builder.getFalse());
6249 Pred = ICmpInst::ICMP_ULE;
6250 break;
6251 case ICmpInst::ICMP_SLT:
6252 // (float)int < -4.4 --> int < -4
6253 // (float)int < 4.4 --> int <= 4
6254 if (!RHS.isNegative())
6255 Pred = ICmpInst::ICMP_SLE;
6256 break;
6257 case ICmpInst::ICMP_UGT:
6258 // (float)int > 4.4 --> int > 4
6259 // (float)int > -4.4 --> true
6260 if (RHS.isNegative())
6261 return replaceInstUsesWith(I, Builder.getTrue());
6262 break;
6263 case ICmpInst::ICMP_SGT:
6264 // (float)int > 4.4 --> int > 4
6265 // (float)int > -4.4 --> int >= -4
6266 if (RHS.isNegative())
6267 Pred = ICmpInst::ICMP_SGE;
6268 break;
6269 case ICmpInst::ICMP_UGE:
6270 // (float)int >= -4.4 --> true
6271 // (float)int >= 4.4 --> int > 4
6272 if (RHS.isNegative())
6273 return replaceInstUsesWith(I, Builder.getTrue());
6274 Pred = ICmpInst::ICMP_UGT;
6275 break;
6276 case ICmpInst::ICMP_SGE:
6277 // (float)int >= -4.4 --> int >= -4
6278 // (float)int >= 4.4 --> int > 4
6279 if (!RHS.isNegative())
6280 Pred = ICmpInst::ICMP_SGT;
6281 break;
6282 }
6283 }
6284 }
6285
6286 // Lower this FP comparison into an appropriate integer version of the
6287 // comparison.
6288 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
6289 }
6290
6291 /// Fold (C / X) < 0.0 --> X < 0.0 if possible. Swap predicate if necessary.
foldFCmpReciprocalAndZero(FCmpInst & I,Instruction * LHSI,Constant * RHSC)6292 static Instruction *foldFCmpReciprocalAndZero(FCmpInst &I, Instruction *LHSI,
6293 Constant *RHSC) {
6294 // When C is not 0.0 and infinities are not allowed:
6295 // (C / X) < 0.0 is a sign-bit test of X
6296 // (C / X) < 0.0 --> X < 0.0 (if C is positive)
6297 // (C / X) < 0.0 --> X > 0.0 (if C is negative, swap the predicate)
6298 //
6299 // Proof:
6300 // Multiply (C / X) < 0.0 by X * X / C.
6301 // - X is non zero, if it is the flag 'ninf' is violated.
6302 // - C defines the sign of X * X * C. Thus it also defines whether to swap
6303 // the predicate. C is also non zero by definition.
6304 //
6305 // Thus X * X / C is non zero and the transformation is valid. [qed]
6306
6307 FCmpInst::Predicate Pred = I.getPredicate();
6308
6309 // Check that predicates are valid.
6310 if ((Pred != FCmpInst::FCMP_OGT) && (Pred != FCmpInst::FCMP_OLT) &&
6311 (Pred != FCmpInst::FCMP_OGE) && (Pred != FCmpInst::FCMP_OLE))
6312 return nullptr;
6313
6314 // Check that RHS operand is zero.
6315 if (!match(RHSC, m_AnyZeroFP()))
6316 return nullptr;
6317
6318 // Check fastmath flags ('ninf').
6319 if (!LHSI->hasNoInfs() || !I.hasNoInfs())
6320 return nullptr;
6321
6322 // Check the properties of the dividend. It must not be zero to avoid a
6323 // division by zero (see Proof).
6324 const APFloat *C;
6325 if (!match(LHSI->getOperand(0), m_APFloat(C)))
6326 return nullptr;
6327
6328 if (C->isZero())
6329 return nullptr;
6330
6331 // Get swapped predicate if necessary.
6332 if (C->isNegative())
6333 Pred = I.getSwappedPredicate();
6334
6335 return new FCmpInst(Pred, LHSI->getOperand(1), RHSC, "", &I);
6336 }
6337
6338 /// Optimize fabs(X) compared with zero.
foldFabsWithFcmpZero(FCmpInst & I,InstCombinerImpl & IC)6339 static Instruction *foldFabsWithFcmpZero(FCmpInst &I, InstCombinerImpl &IC) {
6340 Value *X;
6341 if (!match(I.getOperand(0), m_FAbs(m_Value(X))) ||
6342 !match(I.getOperand(1), m_PosZeroFP()))
6343 return nullptr;
6344
6345 auto replacePredAndOp0 = [&IC](FCmpInst *I, FCmpInst::Predicate P, Value *X) {
6346 I->setPredicate(P);
6347 return IC.replaceOperand(*I, 0, X);
6348 };
6349
6350 switch (I.getPredicate()) {
6351 case FCmpInst::FCMP_UGE:
6352 case FCmpInst::FCMP_OLT:
6353 // fabs(X) >= 0.0 --> true
6354 // fabs(X) < 0.0 --> false
6355 llvm_unreachable("fcmp should have simplified");
6356
6357 case FCmpInst::FCMP_OGT:
6358 // fabs(X) > 0.0 --> X != 0.0
6359 return replacePredAndOp0(&I, FCmpInst::FCMP_ONE, X);
6360
6361 case FCmpInst::FCMP_UGT:
6362 // fabs(X) u> 0.0 --> X u!= 0.0
6363 return replacePredAndOp0(&I, FCmpInst::FCMP_UNE, X);
6364
6365 case FCmpInst::FCMP_OLE:
6366 // fabs(X) <= 0.0 --> X == 0.0
6367 return replacePredAndOp0(&I, FCmpInst::FCMP_OEQ, X);
6368
6369 case FCmpInst::FCMP_ULE:
6370 // fabs(X) u<= 0.0 --> X u== 0.0
6371 return replacePredAndOp0(&I, FCmpInst::FCMP_UEQ, X);
6372
6373 case FCmpInst::FCMP_OGE:
6374 // fabs(X) >= 0.0 --> !isnan(X)
6375 assert(!I.hasNoNaNs() && "fcmp should have simplified");
6376 return replacePredAndOp0(&I, FCmpInst::FCMP_ORD, X);
6377
6378 case FCmpInst::FCMP_ULT:
6379 // fabs(X) u< 0.0 --> isnan(X)
6380 assert(!I.hasNoNaNs() && "fcmp should have simplified");
6381 return replacePredAndOp0(&I, FCmpInst::FCMP_UNO, X);
6382
6383 case FCmpInst::FCMP_OEQ:
6384 case FCmpInst::FCMP_UEQ:
6385 case FCmpInst::FCMP_ONE:
6386 case FCmpInst::FCMP_UNE:
6387 case FCmpInst::FCMP_ORD:
6388 case FCmpInst::FCMP_UNO:
6389 // Look through the fabs() because it doesn't change anything but the sign.
6390 // fabs(X) == 0.0 --> X == 0.0,
6391 // fabs(X) != 0.0 --> X != 0.0
6392 // isnan(fabs(X)) --> isnan(X)
6393 // !isnan(fabs(X) --> !isnan(X)
6394 return replacePredAndOp0(&I, I.getPredicate(), X);
6395
6396 default:
6397 return nullptr;
6398 }
6399 }
6400
visitFCmpInst(FCmpInst & I)6401 Instruction *InstCombinerImpl::visitFCmpInst(FCmpInst &I) {
6402 bool Changed = false;
6403
6404 /// Orders the operands of the compare so that they are listed from most
6405 /// complex to least complex. This puts constants before unary operators,
6406 /// before binary operators.
6407 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
6408 I.swapOperands();
6409 Changed = true;
6410 }
6411
6412 const CmpInst::Predicate Pred = I.getPredicate();
6413 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6414 if (Value *V = SimplifyFCmpInst(Pred, Op0, Op1, I.getFastMathFlags(),
6415 SQ.getWithInstruction(&I)))
6416 return replaceInstUsesWith(I, V);
6417
6418 // Simplify 'fcmp pred X, X'
6419 Type *OpType = Op0->getType();
6420 assert(OpType == Op1->getType() && "fcmp with different-typed operands?");
6421 if (Op0 == Op1) {
6422 switch (Pred) {
6423 default: break;
6424 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
6425 case FCmpInst::FCMP_ULT: // True if unordered or less than
6426 case FCmpInst::FCMP_UGT: // True if unordered or greater than
6427 case FCmpInst::FCMP_UNE: // True if unordered or not equal
6428 // Canonicalize these to be 'fcmp uno %X, 0.0'.
6429 I.setPredicate(FCmpInst::FCMP_UNO);
6430 I.setOperand(1, Constant::getNullValue(OpType));
6431 return &I;
6432
6433 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
6434 case FCmpInst::FCMP_OEQ: // True if ordered and equal
6435 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
6436 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
6437 // Canonicalize these to be 'fcmp ord %X, 0.0'.
6438 I.setPredicate(FCmpInst::FCMP_ORD);
6439 I.setOperand(1, Constant::getNullValue(OpType));
6440 return &I;
6441 }
6442 }
6443
6444 // If we're just checking for a NaN (ORD/UNO) and have a non-NaN operand,
6445 // then canonicalize the operand to 0.0.
6446 if (Pred == CmpInst::FCMP_ORD || Pred == CmpInst::FCMP_UNO) {
6447 if (!match(Op0, m_PosZeroFP()) && isKnownNeverNaN(Op0, &TLI))
6448 return replaceOperand(I, 0, ConstantFP::getNullValue(OpType));
6449
6450 if (!match(Op1, m_PosZeroFP()) && isKnownNeverNaN(Op1, &TLI))
6451 return replaceOperand(I, 1, ConstantFP::getNullValue(OpType));
6452 }
6453
6454 // fcmp pred (fneg X), (fneg Y) -> fcmp swap(pred) X, Y
6455 Value *X, *Y;
6456 if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
6457 return new FCmpInst(I.getSwappedPredicate(), X, Y, "", &I);
6458
6459 // Test if the FCmpInst instruction is used exclusively by a select as
6460 // part of a minimum or maximum operation. If so, refrain from doing
6461 // any other folding. This helps out other analyses which understand
6462 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6463 // and CodeGen. And in this case, at least one of the comparison
6464 // operands has at least one user besides the compare (the select),
6465 // which would often largely negate the benefit of folding anyway.
6466 if (I.hasOneUse())
6467 if (SelectInst *SI = dyn_cast<SelectInst>(I.user_back())) {
6468 Value *A, *B;
6469 SelectPatternResult SPR = matchSelectPattern(SI, A, B);
6470 if (SPR.Flavor != SPF_UNKNOWN)
6471 return nullptr;
6472 }
6473
6474 // The sign of 0.0 is ignored by fcmp, so canonicalize to +0.0:
6475 // fcmp Pred X, -0.0 --> fcmp Pred X, 0.0
6476 if (match(Op1, m_AnyZeroFP()) && !match(Op1, m_PosZeroFP()))
6477 return replaceOperand(I, 1, ConstantFP::getNullValue(OpType));
6478
6479 // Handle fcmp with instruction LHS and constant RHS.
6480 Instruction *LHSI;
6481 Constant *RHSC;
6482 if (match(Op0, m_Instruction(LHSI)) && match(Op1, m_Constant(RHSC))) {
6483 switch (LHSI->getOpcode()) {
6484 case Instruction::PHI:
6485 // Only fold fcmp into the PHI if the phi and fcmp are in the same
6486 // block. If in the same block, we're encouraging jump threading. If
6487 // not, we are just pessimizing the code by making an i1 phi.
6488 if (LHSI->getParent() == I.getParent())
6489 if (Instruction *NV = foldOpIntoPhi(I, cast<PHINode>(LHSI)))
6490 return NV;
6491 break;
6492 case Instruction::SIToFP:
6493 case Instruction::UIToFP:
6494 if (Instruction *NV = foldFCmpIntToFPConst(I, LHSI, RHSC))
6495 return NV;
6496 break;
6497 case Instruction::FDiv:
6498 if (Instruction *NV = foldFCmpReciprocalAndZero(I, LHSI, RHSC))
6499 return NV;
6500 break;
6501 case Instruction::Load:
6502 if (auto *GEP = dyn_cast<GetElementPtrInst>(LHSI->getOperand(0)))
6503 if (auto *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
6504 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
6505 !cast<LoadInst>(LHSI)->isVolatile())
6506 if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, I))
6507 return Res;
6508 break;
6509 }
6510 }
6511
6512 if (Instruction *R = foldFabsWithFcmpZero(I, *this))
6513 return R;
6514
6515 if (match(Op0, m_FNeg(m_Value(X)))) {
6516 // fcmp pred (fneg X), C --> fcmp swap(pred) X, -C
6517 Constant *C;
6518 if (match(Op1, m_Constant(C))) {
6519 Constant *NegC = ConstantExpr::getFNeg(C);
6520 return new FCmpInst(I.getSwappedPredicate(), X, NegC, "", &I);
6521 }
6522 }
6523
6524 if (match(Op0, m_FPExt(m_Value(X)))) {
6525 // fcmp (fpext X), (fpext Y) -> fcmp X, Y
6526 if (match(Op1, m_FPExt(m_Value(Y))) && X->getType() == Y->getType())
6527 return new FCmpInst(Pred, X, Y, "", &I);
6528
6529 // fcmp (fpext X), C -> fcmp X, (fptrunc C) if fptrunc is lossless
6530 const APFloat *C;
6531 if (match(Op1, m_APFloat(C))) {
6532 const fltSemantics &FPSem =
6533 X->getType()->getScalarType()->getFltSemantics();
6534 bool Lossy;
6535 APFloat TruncC = *C;
6536 TruncC.convert(FPSem, APFloat::rmNearestTiesToEven, &Lossy);
6537
6538 // Avoid lossy conversions and denormals.
6539 // Zero is a special case that's OK to convert.
6540 APFloat Fabs = TruncC;
6541 Fabs.clearSign();
6542 if (!Lossy &&
6543 (!(Fabs < APFloat::getSmallestNormalized(FPSem)) || Fabs.isZero())) {
6544 Constant *NewC = ConstantFP::get(X->getType(), TruncC);
6545 return new FCmpInst(Pred, X, NewC, "", &I);
6546 }
6547 }
6548 }
6549
6550 // Convert a sign-bit test of an FP value into a cast and integer compare.
6551 // TODO: Simplify if the copysign constant is 0.0 or NaN.
6552 // TODO: Handle non-zero compare constants.
6553 // TODO: Handle other predicates.
6554 const APFloat *C;
6555 if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::copysign>(m_APFloat(C),
6556 m_Value(X)))) &&
6557 match(Op1, m_AnyZeroFP()) && !C->isZero() && !C->isNaN()) {
6558 Type *IntType = Builder.getIntNTy(X->getType()->getScalarSizeInBits());
6559 if (auto *VecTy = dyn_cast<VectorType>(OpType))
6560 IntType = VectorType::get(IntType, VecTy->getElementCount());
6561
6562 // copysign(non-zero constant, X) < 0.0 --> (bitcast X) < 0
6563 if (Pred == FCmpInst::FCMP_OLT) {
6564 Value *IntX = Builder.CreateBitCast(X, IntType);
6565 return new ICmpInst(ICmpInst::ICMP_SLT, IntX,
6566 ConstantInt::getNullValue(IntType));
6567 }
6568 }
6569
6570 if (I.getType()->isVectorTy())
6571 if (Instruction *Res = foldVectorCmp(I, Builder))
6572 return Res;
6573
6574 return Changed ? &I : nullptr;
6575 }
6576