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/CaptureTracking.h" |
18 | #include "llvm/Analysis/CmpInstAnalysis.h" |
19 | #include "llvm/Analysis/ConstantFolding.h" |
20 | #include "llvm/Analysis/InstructionSimplify.h" |
21 | #include "llvm/Analysis/Utils/Local.h" |
22 | #include "llvm/Analysis/VectorUtils.h" |
23 | #include "llvm/IR/ConstantRange.h" |
24 | #include "llvm/IR/DataLayout.h" |
25 | #include "llvm/IR/InstrTypes.h" |
26 | #include "llvm/IR/IntrinsicInst.h" |
27 | #include "llvm/IR/PatternMatch.h" |
28 | #include "llvm/Support/KnownBits.h" |
29 | #include "llvm/Transforms/InstCombine/InstCombiner.h" |
30 | #include <bitset> |
31 | |
32 | using namespace llvm; |
33 | using namespace PatternMatch; |
34 | |
35 | #define DEBUG_TYPE "instcombine" |
36 | |
37 | // How many times is a select replaced by one of its operands? |
38 | STATISTIC(NumSel, "Number of select opts" ); |
39 | |
40 | /// Compute Result = In1+In2, returning true if the result overflowed for this |
41 | /// type. |
42 | static bool addWithOverflow(APInt &Result, const APInt &In1, const APInt &In2, |
43 | bool IsSigned = false) { |
44 | bool Overflow; |
45 | if (IsSigned) |
46 | Result = In1.sadd_ov(RHS: In2, Overflow); |
47 | else |
48 | Result = In1.uadd_ov(RHS: In2, Overflow); |
49 | |
50 | return Overflow; |
51 | } |
52 | |
53 | /// Compute Result = In1-In2, returning true if the result overflowed for this |
54 | /// type. |
55 | static bool subWithOverflow(APInt &Result, const APInt &In1, const APInt &In2, |
56 | bool IsSigned = false) { |
57 | bool Overflow; |
58 | if (IsSigned) |
59 | Result = In1.ssub_ov(RHS: In2, Overflow); |
60 | else |
61 | Result = In1.usub_ov(RHS: In2, Overflow); |
62 | |
63 | return Overflow; |
64 | } |
65 | |
66 | /// Given an icmp instruction, return true if any use of this comparison is a |
67 | /// branch on sign bit comparison. |
68 | static bool hasBranchUse(ICmpInst &I) { |
69 | for (auto *U : I.users()) |
70 | if (isa<BranchInst>(Val: U)) |
71 | return true; |
72 | return false; |
73 | } |
74 | |
75 | /// Returns true if the exploded icmp can be expressed as a signed comparison |
76 | /// to zero and updates the predicate accordingly. |
77 | /// The signedness of the comparison is preserved. |
78 | /// TODO: Refactor with decomposeBitTestICmp()? |
79 | static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) { |
80 | if (!ICmpInst::isSigned(predicate: Pred)) |
81 | return false; |
82 | |
83 | if (C.isZero()) |
84 | return ICmpInst::isRelational(P: Pred); |
85 | |
86 | if (C.isOne()) { |
87 | if (Pred == ICmpInst::ICMP_SLT) { |
88 | Pred = ICmpInst::ICMP_SLE; |
89 | return true; |
90 | } |
91 | } else if (C.isAllOnes()) { |
92 | if (Pred == ICmpInst::ICMP_SGT) { |
93 | Pred = ICmpInst::ICMP_SGE; |
94 | return true; |
95 | } |
96 | } |
97 | |
98 | return false; |
99 | } |
100 | |
101 | /// This is called when we see this pattern: |
102 | /// cmp pred (load (gep GV, ...)), cmpcst |
103 | /// where GV is a global variable with a constant initializer. Try to simplify |
104 | /// this into some simple computation that does not need the load. For example |
105 | /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3". |
106 | /// |
107 | /// If AndCst is non-null, then the loaded value is masked with that constant |
108 | /// before doing the comparison. This handles cases like "A[i]&4 == 0". |
109 | Instruction *InstCombinerImpl::foldCmpLoadFromIndexedGlobal( |
110 | LoadInst *LI, GetElementPtrInst *GEP, GlobalVariable *GV, CmpInst &ICI, |
111 | ConstantInt *AndCst) { |
112 | if (LI->isVolatile() || LI->getType() != GEP->getResultElementType() || |
113 | GV->getValueType() != GEP->getSourceElementType() || !GV->isConstant() || |
114 | !GV->hasDefinitiveInitializer()) |
115 | return nullptr; |
116 | |
117 | Constant *Init = GV->getInitializer(); |
118 | if (!isa<ConstantArray>(Val: Init) && !isa<ConstantDataArray>(Val: Init)) |
119 | return nullptr; |
120 | |
121 | uint64_t ArrayElementCount = Init->getType()->getArrayNumElements(); |
122 | // Don't blow up on huge arrays. |
123 | if (ArrayElementCount > MaxArraySizeForCombine) |
124 | return nullptr; |
125 | |
126 | // There are many forms of this optimization we can handle, for now, just do |
127 | // the simple index into a single-dimensional array. |
128 | // |
129 | // Require: GEP GV, 0, i {{, constant indices}} |
130 | if (GEP->getNumOperands() < 3 || !isa<ConstantInt>(Val: GEP->getOperand(i_nocapture: 1)) || |
131 | !cast<ConstantInt>(Val: GEP->getOperand(i_nocapture: 1))->isZero() || |
132 | isa<Constant>(Val: GEP->getOperand(i_nocapture: 2))) |
133 | return nullptr; |
134 | |
135 | // Check that indices after the variable are constants and in-range for the |
136 | // type they index. Collect the indices. This is typically for arrays of |
137 | // structs. |
138 | SmallVector<unsigned, 4> LaterIndices; |
139 | |
140 | Type *EltTy = Init->getType()->getArrayElementType(); |
141 | for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) { |
142 | ConstantInt *Idx = dyn_cast<ConstantInt>(Val: GEP->getOperand(i_nocapture: i)); |
143 | if (!Idx) |
144 | return nullptr; // Variable index. |
145 | |
146 | uint64_t IdxVal = Idx->getZExtValue(); |
147 | if ((unsigned)IdxVal != IdxVal) |
148 | return nullptr; // Too large array index. |
149 | |
150 | if (StructType *STy = dyn_cast<StructType>(Val: EltTy)) |
151 | EltTy = STy->getElementType(N: IdxVal); |
152 | else if (ArrayType *ATy = dyn_cast<ArrayType>(Val: EltTy)) { |
153 | if (IdxVal >= ATy->getNumElements()) |
154 | return nullptr; |
155 | EltTy = ATy->getElementType(); |
156 | } else { |
157 | return nullptr; // Unknown type. |
158 | } |
159 | |
160 | LaterIndices.push_back(Elt: IdxVal); |
161 | } |
162 | |
163 | enum { Overdefined = -3, Undefined = -2 }; |
164 | |
165 | // Variables for our state machines. |
166 | |
167 | // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form |
168 | // "i == 47 | i == 87", where 47 is the first index the condition is true for, |
169 | // and 87 is the second (and last) index. FirstTrueElement is -2 when |
170 | // undefined, otherwise set to the first true element. SecondTrueElement is |
171 | // -2 when undefined, -3 when overdefined and >= 0 when that index is true. |
172 | int FirstTrueElement = Undefined, SecondTrueElement = Undefined; |
173 | |
174 | // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the |
175 | // form "i != 47 & i != 87". Same state transitions as for true elements. |
176 | int FirstFalseElement = Undefined, SecondFalseElement = Undefined; |
177 | |
178 | /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these |
179 | /// define a state machine that triggers for ranges of values that the index |
180 | /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'. |
181 | /// This is -2 when undefined, -3 when overdefined, and otherwise the last |
182 | /// index in the range (inclusive). We use -2 for undefined here because we |
183 | /// use relative comparisons and don't want 0-1 to match -1. |
184 | int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined; |
185 | |
186 | // MagicBitvector - This is a magic bitvector where we set a bit if the |
187 | // comparison is true for element 'i'. If there are 64 elements or less in |
188 | // the array, this will fully represent all the comparison results. |
189 | uint64_t MagicBitvector = 0; |
190 | |
191 | // Scan the array and see if one of our patterns matches. |
192 | Constant *CompareRHS = cast<Constant>(Val: ICI.getOperand(i_nocapture: 1)); |
193 | for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) { |
194 | Constant *Elt = Init->getAggregateElement(Elt: i); |
195 | if (!Elt) |
196 | return nullptr; |
197 | |
198 | // If this is indexing an array of structures, get the structure element. |
199 | if (!LaterIndices.empty()) { |
200 | Elt = ConstantFoldExtractValueInstruction(Agg: Elt, Idxs: LaterIndices); |
201 | if (!Elt) |
202 | return nullptr; |
203 | } |
204 | |
205 | // If the element is masked, handle it. |
206 | if (AndCst) { |
207 | Elt = ConstantFoldBinaryOpOperands(Opcode: Instruction::And, LHS: Elt, RHS: AndCst, DL); |
208 | if (!Elt) |
209 | return nullptr; |
210 | } |
211 | |
212 | // Find out if the comparison would be true or false for the i'th element. |
213 | Constant *C = ConstantFoldCompareInstOperands(Predicate: ICI.getPredicate(), LHS: Elt, |
214 | RHS: CompareRHS, DL, TLI: &TLI); |
215 | if (!C) |
216 | return nullptr; |
217 | |
218 | // If the result is undef for this element, ignore it. |
219 | if (isa<UndefValue>(Val: C)) { |
220 | // Extend range state machines to cover this element in case there is an |
221 | // undef in the middle of the range. |
222 | if (TrueRangeEnd == (int)i - 1) |
223 | TrueRangeEnd = i; |
224 | if (FalseRangeEnd == (int)i - 1) |
225 | FalseRangeEnd = i; |
226 | continue; |
227 | } |
228 | |
229 | // If we can't compute the result for any of the elements, we have to give |
230 | // up evaluating the entire conditional. |
231 | if (!isa<ConstantInt>(Val: C)) |
232 | return nullptr; |
233 | |
234 | // Otherwise, we know if the comparison is true or false for this element, |
235 | // update our state machines. |
236 | bool IsTrueForElt = !cast<ConstantInt>(Val: C)->isZero(); |
237 | |
238 | // State machine for single/double/range index comparison. |
239 | if (IsTrueForElt) { |
240 | // Update the TrueElement state machine. |
241 | if (FirstTrueElement == Undefined) |
242 | FirstTrueElement = TrueRangeEnd = i; // First true element. |
243 | else { |
244 | // Update double-compare state machine. |
245 | if (SecondTrueElement == Undefined) |
246 | SecondTrueElement = i; |
247 | else |
248 | SecondTrueElement = Overdefined; |
249 | |
250 | // Update range state machine. |
251 | if (TrueRangeEnd == (int)i - 1) |
252 | TrueRangeEnd = i; |
253 | else |
254 | TrueRangeEnd = Overdefined; |
255 | } |
256 | } else { |
257 | // Update the FalseElement state machine. |
258 | if (FirstFalseElement == Undefined) |
259 | FirstFalseElement = FalseRangeEnd = i; // First false element. |
260 | else { |
261 | // Update double-compare state machine. |
262 | if (SecondFalseElement == Undefined) |
263 | SecondFalseElement = i; |
264 | else |
265 | SecondFalseElement = Overdefined; |
266 | |
267 | // Update range state machine. |
268 | if (FalseRangeEnd == (int)i - 1) |
269 | FalseRangeEnd = i; |
270 | else |
271 | FalseRangeEnd = Overdefined; |
272 | } |
273 | } |
274 | |
275 | // If this element is in range, update our magic bitvector. |
276 | if (i < 64 && IsTrueForElt) |
277 | MagicBitvector |= 1ULL << i; |
278 | |
279 | // If all of our states become overdefined, bail out early. Since the |
280 | // predicate is expensive, only check it every 8 elements. This is only |
281 | // really useful for really huge arrays. |
282 | if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined && |
283 | SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined && |
284 | FalseRangeEnd == Overdefined) |
285 | return nullptr; |
286 | } |
287 | |
288 | // Now that we've scanned the entire array, emit our new comparison(s). We |
289 | // order the state machines in complexity of the generated code. |
290 | Value *Idx = GEP->getOperand(i_nocapture: 2); |
291 | |
292 | // If the index is larger than the pointer offset size of the target, truncate |
293 | // the index down like the GEP would do implicitly. We don't have to do this |
294 | // for an inbounds GEP because the index can't be out of range. |
295 | if (!GEP->isInBounds()) { |
296 | Type *PtrIdxTy = DL.getIndexType(PtrTy: GEP->getType()); |
297 | unsigned OffsetSize = PtrIdxTy->getIntegerBitWidth(); |
298 | if (Idx->getType()->getPrimitiveSizeInBits().getFixedValue() > OffsetSize) |
299 | Idx = Builder.CreateTrunc(V: Idx, DestTy: PtrIdxTy); |
300 | } |
301 | |
302 | // If inbounds keyword is not present, Idx * ElementSize can overflow. |
303 | // Let's assume that ElementSize is 2 and the wanted value is at offset 0. |
304 | // Then, there are two possible values for Idx to match offset 0: |
305 | // 0x00..00, 0x80..00. |
306 | // Emitting 'icmp eq Idx, 0' isn't correct in this case because the |
307 | // comparison is false if Idx was 0x80..00. |
308 | // We need to erase the highest countTrailingZeros(ElementSize) bits of Idx. |
309 | unsigned ElementSize = |
310 | DL.getTypeAllocSize(Ty: Init->getType()->getArrayElementType()); |
311 | auto MaskIdx = [&](Value *Idx) { |
312 | if (!GEP->isInBounds() && llvm::countr_zero(Val: ElementSize) != 0) { |
313 | Value *Mask = Constant::getAllOnesValue(Ty: Idx->getType()); |
314 | Mask = Builder.CreateLShr(LHS: Mask, RHS: llvm::countr_zero(Val: ElementSize)); |
315 | Idx = Builder.CreateAnd(LHS: Idx, RHS: Mask); |
316 | } |
317 | return Idx; |
318 | }; |
319 | |
320 | // If the comparison is only true for one or two elements, emit direct |
321 | // comparisons. |
322 | if (SecondTrueElement != Overdefined) { |
323 | Idx = MaskIdx(Idx); |
324 | // None true -> false. |
325 | if (FirstTrueElement == Undefined) |
326 | return replaceInstUsesWith(I&: ICI, V: Builder.getFalse()); |
327 | |
328 | Value *FirstTrueIdx = ConstantInt::get(Ty: Idx->getType(), V: FirstTrueElement); |
329 | |
330 | // True for one element -> 'i == 47'. |
331 | if (SecondTrueElement == Undefined) |
332 | return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx); |
333 | |
334 | // True for two elements -> 'i == 47 | i == 72'. |
335 | Value *C1 = Builder.CreateICmpEQ(LHS: Idx, RHS: FirstTrueIdx); |
336 | Value *SecondTrueIdx = ConstantInt::get(Ty: Idx->getType(), V: SecondTrueElement); |
337 | Value *C2 = Builder.CreateICmpEQ(LHS: Idx, RHS: SecondTrueIdx); |
338 | return BinaryOperator::CreateOr(V1: C1, V2: C2); |
339 | } |
340 | |
341 | // If the comparison is only false for one or two elements, emit direct |
342 | // comparisons. |
343 | if (SecondFalseElement != Overdefined) { |
344 | Idx = MaskIdx(Idx); |
345 | // None false -> true. |
346 | if (FirstFalseElement == Undefined) |
347 | return replaceInstUsesWith(I&: ICI, V: Builder.getTrue()); |
348 | |
349 | Value *FirstFalseIdx = ConstantInt::get(Ty: Idx->getType(), V: FirstFalseElement); |
350 | |
351 | // False for one element -> 'i != 47'. |
352 | if (SecondFalseElement == Undefined) |
353 | return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx); |
354 | |
355 | // False for two elements -> 'i != 47 & i != 72'. |
356 | Value *C1 = Builder.CreateICmpNE(LHS: Idx, RHS: FirstFalseIdx); |
357 | Value *SecondFalseIdx = |
358 | ConstantInt::get(Ty: Idx->getType(), V: SecondFalseElement); |
359 | Value *C2 = Builder.CreateICmpNE(LHS: Idx, RHS: SecondFalseIdx); |
360 | return BinaryOperator::CreateAnd(V1: C1, V2: C2); |
361 | } |
362 | |
363 | // If the comparison can be replaced with a range comparison for the elements |
364 | // where it is true, emit the range check. |
365 | if (TrueRangeEnd != Overdefined) { |
366 | assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare" ); |
367 | Idx = MaskIdx(Idx); |
368 | |
369 | // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1). |
370 | if (FirstTrueElement) { |
371 | Value *Offs = ConstantInt::get(Ty: Idx->getType(), V: -FirstTrueElement); |
372 | Idx = Builder.CreateAdd(LHS: Idx, RHS: Offs); |
373 | } |
374 | |
375 | Value *End = |
376 | ConstantInt::get(Ty: Idx->getType(), V: TrueRangeEnd - FirstTrueElement + 1); |
377 | return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End); |
378 | } |
379 | |
380 | // False range check. |
381 | if (FalseRangeEnd != Overdefined) { |
382 | assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare" ); |
383 | Idx = MaskIdx(Idx); |
384 | // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse). |
385 | if (FirstFalseElement) { |
386 | Value *Offs = ConstantInt::get(Ty: Idx->getType(), V: -FirstFalseElement); |
387 | Idx = Builder.CreateAdd(LHS: Idx, RHS: Offs); |
388 | } |
389 | |
390 | Value *End = |
391 | ConstantInt::get(Ty: Idx->getType(), V: FalseRangeEnd - FirstFalseElement); |
392 | return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End); |
393 | } |
394 | |
395 | // If a magic bitvector captures the entire comparison state |
396 | // of this load, replace it with computation that does: |
397 | // ((magic_cst >> i) & 1) != 0 |
398 | { |
399 | Type *Ty = nullptr; |
400 | |
401 | // Look for an appropriate type: |
402 | // - The type of Idx if the magic fits |
403 | // - The smallest fitting legal type |
404 | if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth()) |
405 | Ty = Idx->getType(); |
406 | else |
407 | Ty = DL.getSmallestLegalIntType(C&: Init->getContext(), Width: ArrayElementCount); |
408 | |
409 | if (Ty) { |
410 | Idx = MaskIdx(Idx); |
411 | Value *V = Builder.CreateIntCast(V: Idx, DestTy: Ty, isSigned: false); |
412 | V = Builder.CreateLShr(LHS: ConstantInt::get(Ty, V: MagicBitvector), RHS: V); |
413 | V = Builder.CreateAnd(LHS: ConstantInt::get(Ty, V: 1), RHS: V); |
414 | return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, V: 0)); |
415 | } |
416 | } |
417 | |
418 | return nullptr; |
419 | } |
420 | |
421 | /// Returns true if we can rewrite Start as a GEP with pointer Base |
422 | /// and some integer offset. The nodes that need to be re-written |
423 | /// for this transformation will be added to Explored. |
424 | static bool canRewriteGEPAsOffset(Value *Start, Value *Base, GEPNoWrapFlags &NW, |
425 | const DataLayout &DL, |
426 | SetVector<Value *> &Explored) { |
427 | SmallVector<Value *, 16> WorkList(1, Start); |
428 | Explored.insert(X: Base); |
429 | |
430 | // The following traversal gives us an order which can be used |
431 | // when doing the final transformation. Since in the final |
432 | // transformation we create the PHI replacement instructions first, |
433 | // we don't have to get them in any particular order. |
434 | // |
435 | // However, for other instructions we will have to traverse the |
436 | // operands of an instruction first, which means that we have to |
437 | // do a post-order traversal. |
438 | while (!WorkList.empty()) { |
439 | SetVector<PHINode *> PHIs; |
440 | |
441 | while (!WorkList.empty()) { |
442 | if (Explored.size() >= 100) |
443 | return false; |
444 | |
445 | Value *V = WorkList.back(); |
446 | |
447 | if (Explored.contains(key: V)) { |
448 | WorkList.pop_back(); |
449 | continue; |
450 | } |
451 | |
452 | if (!isa<GetElementPtrInst>(Val: V) && !isa<PHINode>(Val: V)) |
453 | // We've found some value that we can't explore which is different from |
454 | // the base. Therefore we can't do this transformation. |
455 | return false; |
456 | |
457 | if (auto *GEP = dyn_cast<GEPOperator>(Val: V)) { |
458 | // Only allow inbounds GEPs with at most one variable offset. |
459 | auto IsNonConst = [](Value *V) { return !isa<ConstantInt>(Val: V); }; |
460 | if (!GEP->isInBounds() || count_if(Range: GEP->indices(), P: IsNonConst) > 1) |
461 | return false; |
462 | |
463 | NW = NW.intersectForOffsetAdd(Other: GEP->getNoWrapFlags()); |
464 | if (!Explored.contains(key: GEP->getOperand(i_nocapture: 0))) |
465 | WorkList.push_back(Elt: GEP->getOperand(i_nocapture: 0)); |
466 | } |
467 | |
468 | if (WorkList.back() == V) { |
469 | WorkList.pop_back(); |
470 | // We've finished visiting this node, mark it as such. |
471 | Explored.insert(X: V); |
472 | } |
473 | |
474 | if (auto *PN = dyn_cast<PHINode>(Val: V)) { |
475 | // We cannot transform PHIs on unsplittable basic blocks. |
476 | if (isa<CatchSwitchInst>(Val: PN->getParent()->getTerminator())) |
477 | return false; |
478 | Explored.insert(X: PN); |
479 | PHIs.insert(X: PN); |
480 | } |
481 | } |
482 | |
483 | // Explore the PHI nodes further. |
484 | for (auto *PN : PHIs) |
485 | for (Value *Op : PN->incoming_values()) |
486 | if (!Explored.contains(key: Op)) |
487 | WorkList.push_back(Elt: Op); |
488 | } |
489 | |
490 | // Make sure that we can do this. Since we can't insert GEPs in a basic |
491 | // block before a PHI node, we can't easily do this transformation if |
492 | // we have PHI node users of transformed instructions. |
493 | for (Value *Val : Explored) { |
494 | for (Value *Use : Val->uses()) { |
495 | |
496 | auto *PHI = dyn_cast<PHINode>(Val: Use); |
497 | auto *Inst = dyn_cast<Instruction>(Val); |
498 | |
499 | if (Inst == Base || Inst == PHI || !Inst || !PHI || |
500 | !Explored.contains(key: PHI)) |
501 | continue; |
502 | |
503 | if (PHI->getParent() == Inst->getParent()) |
504 | return false; |
505 | } |
506 | } |
507 | return true; |
508 | } |
509 | |
510 | // Sets the appropriate insert point on Builder where we can add |
511 | // a replacement Instruction for V (if that is possible). |
512 | static void setInsertionPoint(IRBuilder<> &Builder, Value *V, |
513 | bool Before = true) { |
514 | if (auto *PHI = dyn_cast<PHINode>(Val: V)) { |
515 | BasicBlock *Parent = PHI->getParent(); |
516 | Builder.SetInsertPoint(TheBB: Parent, IP: Parent->getFirstInsertionPt()); |
517 | return; |
518 | } |
519 | if (auto *I = dyn_cast<Instruction>(Val: V)) { |
520 | if (!Before) |
521 | I = &*std::next(x: I->getIterator()); |
522 | Builder.SetInsertPoint(I); |
523 | return; |
524 | } |
525 | if (auto *A = dyn_cast<Argument>(Val: V)) { |
526 | // Set the insertion point in the entry block. |
527 | BasicBlock &Entry = A->getParent()->getEntryBlock(); |
528 | Builder.SetInsertPoint(TheBB: &Entry, IP: Entry.getFirstInsertionPt()); |
529 | return; |
530 | } |
531 | // Otherwise, this is a constant and we don't need to set a new |
532 | // insertion point. |
533 | assert(isa<Constant>(V) && "Setting insertion point for unknown value!" ); |
534 | } |
535 | |
536 | /// Returns a re-written value of Start as an indexed GEP using Base as a |
537 | /// pointer. |
538 | static Value *rewriteGEPAsOffset(Value *Start, Value *Base, GEPNoWrapFlags NW, |
539 | const DataLayout &DL, |
540 | SetVector<Value *> &Explored, |
541 | InstCombiner &IC) { |
542 | // Perform all the substitutions. This is a bit tricky because we can |
543 | // have cycles in our use-def chains. |
544 | // 1. Create the PHI nodes without any incoming values. |
545 | // 2. Create all the other values. |
546 | // 3. Add the edges for the PHI nodes. |
547 | // 4. Emit GEPs to get the original pointers. |
548 | // 5. Remove the original instructions. |
549 | Type *IndexType = IntegerType::get( |
550 | C&: Base->getContext(), NumBits: DL.getIndexTypeSizeInBits(Ty: Start->getType())); |
551 | |
552 | DenseMap<Value *, Value *> NewInsts; |
553 | NewInsts[Base] = ConstantInt::getNullValue(Ty: IndexType); |
554 | |
555 | // Create the new PHI nodes, without adding any incoming values. |
556 | for (Value *Val : Explored) { |
557 | if (Val == Base) |
558 | continue; |
559 | // Create empty phi nodes. This avoids cyclic dependencies when creating |
560 | // the remaining instructions. |
561 | if (auto *PHI = dyn_cast<PHINode>(Val)) |
562 | NewInsts[PHI] = |
563 | PHINode::Create(Ty: IndexType, NumReservedValues: PHI->getNumIncomingValues(), |
564 | NameStr: PHI->getName() + ".idx" , InsertBefore: PHI->getIterator()); |
565 | } |
566 | IRBuilder<> Builder(Base->getContext()); |
567 | |
568 | // Create all the other instructions. |
569 | for (Value *Val : Explored) { |
570 | if (NewInsts.contains(Val)) |
571 | continue; |
572 | |
573 | if (auto *GEP = dyn_cast<GEPOperator>(Val)) { |
574 | setInsertionPoint(Builder, V: GEP); |
575 | Value *Op = NewInsts[GEP->getOperand(i_nocapture: 0)]; |
576 | Value *OffsetV = emitGEPOffset(Builder: &Builder, DL, GEP); |
577 | if (isa<ConstantInt>(Val: Op) && cast<ConstantInt>(Val: Op)->isZero()) |
578 | NewInsts[GEP] = OffsetV; |
579 | else |
580 | NewInsts[GEP] = Builder.CreateAdd( |
581 | LHS: Op, RHS: OffsetV, Name: GEP->getOperand(i_nocapture: 0)->getName() + ".add" , |
582 | /*NUW=*/HasNUW: NW.hasNoUnsignedWrap(), |
583 | /*NSW=*/HasNSW: NW.hasNoUnsignedSignedWrap()); |
584 | continue; |
585 | } |
586 | if (isa<PHINode>(Val)) |
587 | continue; |
588 | |
589 | llvm_unreachable("Unexpected instruction type" ); |
590 | } |
591 | |
592 | // Add the incoming values to the PHI nodes. |
593 | for (Value *Val : Explored) { |
594 | if (Val == Base) |
595 | continue; |
596 | // All the instructions have been created, we can now add edges to the |
597 | // phi nodes. |
598 | if (auto *PHI = dyn_cast<PHINode>(Val)) { |
599 | PHINode *NewPhi = static_cast<PHINode *>(NewInsts[PHI]); |
600 | for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) { |
601 | Value *NewIncoming = PHI->getIncomingValue(i: I); |
602 | |
603 | auto It = NewInsts.find(Val: NewIncoming); |
604 | if (It != NewInsts.end()) |
605 | NewIncoming = It->second; |
606 | |
607 | NewPhi->addIncoming(V: NewIncoming, BB: PHI->getIncomingBlock(i: I)); |
608 | } |
609 | } |
610 | } |
611 | |
612 | for (Value *Val : Explored) { |
613 | if (Val == Base) |
614 | continue; |
615 | |
616 | setInsertionPoint(Builder, V: Val, Before: false); |
617 | // Create GEP for external users. |
618 | Value *NewVal = Builder.CreateGEP(Ty: Builder.getInt8Ty(), Ptr: Base, IdxList: NewInsts[Val], |
619 | Name: Val->getName() + ".ptr" , NW); |
620 | IC.replaceInstUsesWith(I&: *cast<Instruction>(Val), V: NewVal); |
621 | // Add old instruction to worklist for DCE. We don't directly remove it |
622 | // here because the original compare is one of the users. |
623 | IC.addToWorklist(I: cast<Instruction>(Val)); |
624 | } |
625 | |
626 | return NewInsts[Start]; |
627 | } |
628 | |
629 | /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant. |
630 | /// We can look through PHIs, GEPs and casts in order to determine a common base |
631 | /// between GEPLHS and RHS. |
632 | static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS, |
633 | CmpPredicate Cond, |
634 | const DataLayout &DL, |
635 | InstCombiner &IC) { |
636 | // FIXME: Support vector of pointers. |
637 | if (GEPLHS->getType()->isVectorTy()) |
638 | return nullptr; |
639 | |
640 | if (!GEPLHS->hasAllConstantIndices()) |
641 | return nullptr; |
642 | |
643 | APInt Offset(DL.getIndexTypeSizeInBits(Ty: GEPLHS->getType()), 0); |
644 | Value *PtrBase = |
645 | GEPLHS->stripAndAccumulateConstantOffsets(DL, Offset, |
646 | /*AllowNonInbounds*/ false); |
647 | |
648 | // Bail if we looked through addrspacecast. |
649 | if (PtrBase->getType() != GEPLHS->getType()) |
650 | return nullptr; |
651 | |
652 | // The set of nodes that will take part in this transformation. |
653 | SetVector<Value *> Nodes; |
654 | GEPNoWrapFlags NW = GEPLHS->getNoWrapFlags(); |
655 | if (!canRewriteGEPAsOffset(Start: RHS, Base: PtrBase, NW, DL, Explored&: Nodes)) |
656 | return nullptr; |
657 | |
658 | // We know we can re-write this as |
659 | // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) |
660 | // Since we've only looked through inbouds GEPs we know that we |
661 | // can't have overflow on either side. We can therefore re-write |
662 | // this as: |
663 | // OFFSET1 cmp OFFSET2 |
664 | Value *NewRHS = rewriteGEPAsOffset(Start: RHS, Base: PtrBase, NW, DL, Explored&: Nodes, IC); |
665 | |
666 | // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written |
667 | // GEP having PtrBase as the pointer base, and has returned in NewRHS the |
668 | // offset. Since Index is the offset of LHS to the base pointer, we will now |
669 | // compare the offsets instead of comparing the pointers. |
670 | return new ICmpInst(ICmpInst::getSignedPredicate(Pred: Cond), |
671 | IC.Builder.getInt(AI: Offset), NewRHS); |
672 | } |
673 | |
674 | /// Fold comparisons between a GEP instruction and something else. At this point |
675 | /// we know that the GEP is on the LHS of the comparison. |
676 | Instruction *InstCombinerImpl::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS, |
677 | CmpPredicate Cond, Instruction &I) { |
678 | // Don't transform signed compares of GEPs into index compares. Even if the |
679 | // GEP is inbounds, the final add of the base pointer can have signed overflow |
680 | // and would change the result of the icmp. |
681 | // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be |
682 | // the maximum signed value for the pointer type. |
683 | if (ICmpInst::isSigned(predicate: Cond)) |
684 | return nullptr; |
685 | |
686 | // Look through bitcasts and addrspacecasts. We do not however want to remove |
687 | // 0 GEPs. |
688 | if (!isa<GetElementPtrInst>(Val: RHS)) |
689 | RHS = RHS->stripPointerCasts(); |
690 | |
691 | auto CanFold = [Cond](GEPNoWrapFlags NW) { |
692 | if (ICmpInst::isEquality(P: Cond)) |
693 | return true; |
694 | |
695 | // Unsigned predicates can be folded if the GEPs have *any* nowrap flags. |
696 | assert(ICmpInst::isUnsigned(Cond)); |
697 | return NW != GEPNoWrapFlags::none(); |
698 | }; |
699 | |
700 | auto NewICmp = [Cond](GEPNoWrapFlags NW, Value *Op1, Value *Op2) { |
701 | if (!NW.hasNoUnsignedWrap()) { |
702 | // Convert signed to unsigned comparison. |
703 | return new ICmpInst(ICmpInst::getSignedPredicate(Pred: Cond), Op1, Op2); |
704 | } |
705 | |
706 | auto *I = new ICmpInst(Cond, Op1, Op2); |
707 | I->setSameSign(NW.hasNoUnsignedSignedWrap()); |
708 | return I; |
709 | }; |
710 | |
711 | CommonPointerBase Base = CommonPointerBase::compute(LHS: GEPLHS, RHS); |
712 | if (Base.Ptr == RHS && CanFold(Base.LHSNW)) { |
713 | // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0). |
714 | Type *IdxTy = DL.getIndexType(PtrTy: GEPLHS->getType()); |
715 | Value *Offset = |
716 | EmitGEPOffsets(GEPs: Base.LHSGEPs, NW: Base.LHSNW, IdxTy, /*RewriteGEPs=*/true); |
717 | return NewICmp(Base.LHSNW, Offset, |
718 | Constant::getNullValue(Ty: Offset->getType())); |
719 | } |
720 | |
721 | if (GEPLHS->isInBounds() && ICmpInst::isEquality(P: Cond) && |
722 | isa<Constant>(Val: RHS) && cast<Constant>(Val: RHS)->isNullValue() && |
723 | !NullPointerIsDefined(F: I.getFunction(), |
724 | AS: RHS->getType()->getPointerAddressSpace())) { |
725 | // For most address spaces, an allocation can't be placed at null, but null |
726 | // itself is treated as a 0 size allocation in the in bounds rules. Thus, |
727 | // the only valid inbounds address derived from null, is null itself. |
728 | // Thus, we have four cases to consider: |
729 | // 1) Base == nullptr, Offset == 0 -> inbounds, null |
730 | // 2) Base == nullptr, Offset != 0 -> poison as the result is out of bounds |
731 | // 3) Base != nullptr, Offset == (-base) -> poison (crossing allocations) |
732 | // 4) Base != nullptr, Offset != (-base) -> nonnull (and possibly poison) |
733 | // |
734 | // (Note if we're indexing a type of size 0, that simply collapses into one |
735 | // of the buckets above.) |
736 | // |
737 | // In general, we're allowed to make values less poison (i.e. remove |
738 | // sources of full UB), so in this case, we just select between the two |
739 | // non-poison cases (1 and 4 above). |
740 | // |
741 | // For vectors, we apply the same reasoning on a per-lane basis. |
742 | auto *Base = GEPLHS->getPointerOperand(); |
743 | if (GEPLHS->getType()->isVectorTy() && Base->getType()->isPointerTy()) { |
744 | auto EC = cast<VectorType>(Val: GEPLHS->getType())->getElementCount(); |
745 | Base = Builder.CreateVectorSplat(EC, V: Base); |
746 | } |
747 | return new ICmpInst(Cond, Base, |
748 | ConstantExpr::getPointerBitCastOrAddrSpaceCast( |
749 | C: cast<Constant>(Val: RHS), Ty: Base->getType())); |
750 | } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(Val: RHS)) { |
751 | GEPNoWrapFlags NW = GEPLHS->getNoWrapFlags() & GEPRHS->getNoWrapFlags(); |
752 | |
753 | // If the base pointers are different, but the indices are the same, just |
754 | // compare the base pointer. |
755 | Value *PtrBase = GEPLHS->getOperand(i_nocapture: 0); |
756 | if (PtrBase != GEPRHS->getOperand(i_nocapture: 0)) { |
757 | bool IndicesTheSame = |
758 | GEPLHS->getNumOperands() == GEPRHS->getNumOperands() && |
759 | GEPLHS->getPointerOperand()->getType() == |
760 | GEPRHS->getPointerOperand()->getType() && |
761 | GEPLHS->getSourceElementType() == GEPRHS->getSourceElementType(); |
762 | if (IndicesTheSame) |
763 | for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) |
764 | if (GEPLHS->getOperand(i_nocapture: i) != GEPRHS->getOperand(i_nocapture: i)) { |
765 | IndicesTheSame = false; |
766 | break; |
767 | } |
768 | |
769 | // If all indices are the same, just compare the base pointers. |
770 | Type *BaseType = GEPLHS->getOperand(i_nocapture: 0)->getType(); |
771 | if (IndicesTheSame && |
772 | CmpInst::makeCmpResultType(opnd_type: BaseType) == I.getType() && CanFold(NW)) |
773 | return new ICmpInst(Cond, GEPLHS->getOperand(i_nocapture: 0), GEPRHS->getOperand(i_nocapture: 0)); |
774 | |
775 | // If we're comparing GEPs with two base pointers that only differ in type |
776 | // and both GEPs have only constant indices or just one use, then fold |
777 | // the compare with the adjusted indices. |
778 | // FIXME: Support vector of pointers. |
779 | if (GEPLHS->isInBounds() && GEPRHS->isInBounds() && |
780 | (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) && |
781 | (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) && |
782 | PtrBase->stripPointerCasts() == |
783 | GEPRHS->getOperand(i_nocapture: 0)->stripPointerCasts() && |
784 | !GEPLHS->getType()->isVectorTy()) { |
785 | Value *LOffset = EmitGEPOffset(GEP: GEPLHS); |
786 | Value *ROffset = EmitGEPOffset(GEP: GEPRHS); |
787 | |
788 | // If we looked through an addrspacecast between different sized address |
789 | // spaces, the LHS and RHS pointers are different sized |
790 | // integers. Truncate to the smaller one. |
791 | Type *LHSIndexTy = LOffset->getType(); |
792 | Type *RHSIndexTy = ROffset->getType(); |
793 | if (LHSIndexTy != RHSIndexTy) { |
794 | if (LHSIndexTy->getPrimitiveSizeInBits().getFixedValue() < |
795 | RHSIndexTy->getPrimitiveSizeInBits().getFixedValue()) { |
796 | ROffset = Builder.CreateTrunc(V: ROffset, DestTy: LHSIndexTy); |
797 | } else |
798 | LOffset = Builder.CreateTrunc(V: LOffset, DestTy: RHSIndexTy); |
799 | } |
800 | |
801 | Value *Cmp = Builder.CreateICmp(P: ICmpInst::getSignedPredicate(Pred: Cond), |
802 | LHS: LOffset, RHS: ROffset); |
803 | return replaceInstUsesWith(I, V: Cmp); |
804 | } |
805 | |
806 | // Otherwise, the base pointers are different and the indices are |
807 | // different. Try convert this to an indexed compare by looking through |
808 | // PHIs/casts. |
809 | return transformToIndexedCompare(GEPLHS, RHS, Cond, DL, IC&: *this); |
810 | } |
811 | |
812 | if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands() && |
813 | GEPLHS->getSourceElementType() == GEPRHS->getSourceElementType()) { |
814 | // If the GEPs only differ by one index, compare it. |
815 | unsigned NumDifferences = 0; // Keep track of # differences. |
816 | unsigned DiffOperand = 0; // The operand that differs. |
817 | for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) |
818 | if (GEPLHS->getOperand(i_nocapture: i) != GEPRHS->getOperand(i_nocapture: i)) { |
819 | Type *LHSType = GEPLHS->getOperand(i_nocapture: i)->getType(); |
820 | Type *RHSType = GEPRHS->getOperand(i_nocapture: i)->getType(); |
821 | // FIXME: Better support for vector of pointers. |
822 | if (LHSType->getPrimitiveSizeInBits() != |
823 | RHSType->getPrimitiveSizeInBits() || |
824 | (GEPLHS->getType()->isVectorTy() && |
825 | (!LHSType->isVectorTy() || !RHSType->isVectorTy()))) { |
826 | // Irreconcilable differences. |
827 | NumDifferences = 2; |
828 | break; |
829 | } |
830 | |
831 | if (NumDifferences++) |
832 | break; |
833 | DiffOperand = i; |
834 | } |
835 | |
836 | if (NumDifferences == 0) // SAME GEP? |
837 | return replaceInstUsesWith( |
838 | I, // No comparison is needed here. |
839 | V: ConstantInt::get(Ty: I.getType(), V: ICmpInst::isTrueWhenEqual(predicate: Cond))); |
840 | // If two GEPs only differ by an index, compare them. |
841 | // Note that nowrap flags are always needed when comparing two indices. |
842 | else if (NumDifferences == 1 && NW != GEPNoWrapFlags::none()) { |
843 | Value *LHSV = GEPLHS->getOperand(i_nocapture: DiffOperand); |
844 | Value *RHSV = GEPRHS->getOperand(i_nocapture: DiffOperand); |
845 | return NewICmp(NW, LHSV, RHSV); |
846 | } |
847 | } |
848 | |
849 | if (CanFold(NW)) { |
850 | // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2) |
851 | Value *L = EmitGEPOffset(GEP: GEPLHS, /*RewriteGEP=*/true); |
852 | Value *R = EmitGEPOffset(GEP: GEPRHS, /*RewriteGEP=*/true); |
853 | return NewICmp(NW, L, R); |
854 | } |
855 | } |
856 | |
857 | // Try convert this to an indexed compare by looking through PHIs/casts as a |
858 | // last resort. |
859 | return transformToIndexedCompare(GEPLHS, RHS, Cond, DL, IC&: *this); |
860 | } |
861 | |
862 | bool InstCombinerImpl::foldAllocaCmp(AllocaInst *Alloca) { |
863 | // It would be tempting to fold away comparisons between allocas and any |
864 | // pointer not based on that alloca (e.g. an argument). However, even |
865 | // though such pointers cannot alias, they can still compare equal. |
866 | // |
867 | // But LLVM doesn't specify where allocas get their memory, so if the alloca |
868 | // doesn't escape we can argue that it's impossible to guess its value, and we |
869 | // can therefore act as if any such guesses are wrong. |
870 | // |
871 | // However, we need to ensure that this folding is consistent: We can't fold |
872 | // one comparison to false, and then leave a different comparison against the |
873 | // same value alone (as it might evaluate to true at runtime, leading to a |
874 | // contradiction). As such, this code ensures that all comparisons are folded |
875 | // at the same time, and there are no other escapes. |
876 | |
877 | struct CmpCaptureTracker : public CaptureTracker { |
878 | AllocaInst *Alloca; |
879 | bool Captured = false; |
880 | /// The value of the map is a bit mask of which icmp operands the alloca is |
881 | /// used in. |
882 | SmallMapVector<ICmpInst *, unsigned, 4> ICmps; |
883 | |
884 | CmpCaptureTracker(AllocaInst *Alloca) : Alloca(Alloca) {} |
885 | |
886 | void tooManyUses() override { Captured = true; } |
887 | |
888 | Action captured(const Use *U, UseCaptureInfo CI) override { |
889 | // TODO(captures): Use UseCaptureInfo. |
890 | auto *ICmp = dyn_cast<ICmpInst>(Val: U->getUser()); |
891 | // We need to check that U is based *only* on the alloca, and doesn't |
892 | // have other contributions from a select/phi operand. |
893 | // TODO: We could check whether getUnderlyingObjects() reduces to one |
894 | // object, which would allow looking through phi nodes. |
895 | if (ICmp && ICmp->isEquality() && getUnderlyingObject(V: *U) == Alloca) { |
896 | // Collect equality icmps of the alloca, and don't treat them as |
897 | // captures. |
898 | ICmps[ICmp] |= 1u << U->getOperandNo(); |
899 | return Continue; |
900 | } |
901 | |
902 | Captured = true; |
903 | return Stop; |
904 | } |
905 | }; |
906 | |
907 | CmpCaptureTracker Tracker(Alloca); |
908 | PointerMayBeCaptured(V: Alloca, Tracker: &Tracker); |
909 | if (Tracker.Captured) |
910 | return false; |
911 | |
912 | bool Changed = false; |
913 | for (auto [ICmp, Operands] : Tracker.ICmps) { |
914 | switch (Operands) { |
915 | case 1: |
916 | case 2: { |
917 | // The alloca is only used in one icmp operand. Assume that the |
918 | // equality is false. |
919 | auto *Res = ConstantInt::get(Ty: ICmp->getType(), |
920 | V: ICmp->getPredicate() == ICmpInst::ICMP_NE); |
921 | replaceInstUsesWith(I&: *ICmp, V: Res); |
922 | eraseInstFromFunction(I&: *ICmp); |
923 | Changed = true; |
924 | break; |
925 | } |
926 | case 3: |
927 | // Both icmp operands are based on the alloca, so this is comparing |
928 | // pointer offsets, without leaking any information about the address |
929 | // of the alloca. Ignore such comparisons. |
930 | break; |
931 | default: |
932 | llvm_unreachable("Cannot happen" ); |
933 | } |
934 | } |
935 | |
936 | return Changed; |
937 | } |
938 | |
939 | /// Fold "icmp pred (X+C), X". |
940 | Instruction *InstCombinerImpl::foldICmpAddOpConst(Value *X, const APInt &C, |
941 | CmpPredicate Pred) { |
942 | // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0, |
943 | // so the values can never be equal. Similarly for all other "or equals" |
944 | // operators. |
945 | assert(!!C && "C should not be zero!" ); |
946 | |
947 | // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255 |
948 | // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253 |
949 | // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0 |
950 | if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { |
951 | Constant *R = |
952 | ConstantInt::get(Ty: X->getType(), V: APInt::getMaxValue(numBits: C.getBitWidth()) - C); |
953 | return new ICmpInst(ICmpInst::ICMP_UGT, X, R); |
954 | } |
955 | |
956 | // (X+1) >u X --> X <u (0-1) --> X != 255 |
957 | // (X+2) >u X --> X <u (0-2) --> X <u 254 |
958 | // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0 |
959 | if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) |
960 | return new ICmpInst(ICmpInst::ICMP_ULT, X, |
961 | ConstantInt::get(Ty: X->getType(), V: -C)); |
962 | |
963 | APInt SMax = APInt::getSignedMaxValue(numBits: C.getBitWidth()); |
964 | |
965 | // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127 |
966 | // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125 |
967 | // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0 |
968 | // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1 |
969 | // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126 |
970 | // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127 |
971 | if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) |
972 | return new ICmpInst(ICmpInst::ICMP_SGT, X, |
973 | ConstantInt::get(Ty: X->getType(), V: SMax - C)); |
974 | |
975 | // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127 |
976 | // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126 |
977 | // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1 |
978 | // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2 |
979 | // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126 |
980 | // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128 |
981 | |
982 | assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE); |
983 | return new ICmpInst(ICmpInst::ICMP_SLT, X, |
984 | ConstantInt::get(Ty: X->getType(), V: SMax - (C - 1))); |
985 | } |
986 | |
987 | /// Handle "(icmp eq/ne (ashr/lshr AP2, A), AP1)" -> |
988 | /// (icmp eq/ne A, Log2(AP2/AP1)) -> |
989 | /// (icmp eq/ne A, Log2(AP2) - Log2(AP1)). |
990 | Instruction *InstCombinerImpl::foldICmpShrConstConst(ICmpInst &I, Value *A, |
991 | const APInt &AP1, |
992 | const APInt &AP2) { |
993 | assert(I.isEquality() && "Cannot fold icmp gt/lt" ); |
994 | |
995 | auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) { |
996 | if (I.getPredicate() == I.ICMP_NE) |
997 | Pred = CmpInst::getInversePredicate(pred: Pred); |
998 | return new ICmpInst(Pred, LHS, RHS); |
999 | }; |
1000 | |
1001 | // Don't bother doing any work for cases which InstSimplify handles. |
1002 | if (AP2.isZero()) |
1003 | return nullptr; |
1004 | |
1005 | bool IsAShr = isa<AShrOperator>(Val: I.getOperand(i_nocapture: 0)); |
1006 | if (IsAShr) { |
1007 | if (AP2.isAllOnes()) |
1008 | return nullptr; |
1009 | if (AP2.isNegative() != AP1.isNegative()) |
1010 | return nullptr; |
1011 | if (AP2.sgt(RHS: AP1)) |
1012 | return nullptr; |
1013 | } |
1014 | |
1015 | if (!AP1) |
1016 | // 'A' must be large enough to shift out the highest set bit. |
1017 | return getICmp(I.ICMP_UGT, A, |
1018 | ConstantInt::get(Ty: A->getType(), V: AP2.logBase2())); |
1019 | |
1020 | if (AP1 == AP2) |
1021 | return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(Ty: A->getType())); |
1022 | |
1023 | int Shift; |
1024 | if (IsAShr && AP1.isNegative()) |
1025 | Shift = AP1.countl_one() - AP2.countl_one(); |
1026 | else |
1027 | Shift = AP1.countl_zero() - AP2.countl_zero(); |
1028 | |
1029 | if (Shift > 0) { |
1030 | if (IsAShr && AP1 == AP2.ashr(ShiftAmt: Shift)) { |
1031 | // There are multiple solutions if we are comparing against -1 and the LHS |
1032 | // of the ashr is not a power of two. |
1033 | if (AP1.isAllOnes() && !AP2.isPowerOf2()) |
1034 | return getICmp(I.ICMP_UGE, A, ConstantInt::get(Ty: A->getType(), V: Shift)); |
1035 | return getICmp(I.ICMP_EQ, A, ConstantInt::get(Ty: A->getType(), V: Shift)); |
1036 | } else if (AP1 == AP2.lshr(shiftAmt: Shift)) { |
1037 | return getICmp(I.ICMP_EQ, A, ConstantInt::get(Ty: A->getType(), V: Shift)); |
1038 | } |
1039 | } |
1040 | |
1041 | // Shifting const2 will never be equal to const1. |
1042 | // FIXME: This should always be handled by InstSimplify? |
1043 | auto *TorF = ConstantInt::get(Ty: I.getType(), V: I.getPredicate() == I.ICMP_NE); |
1044 | return replaceInstUsesWith(I, V: TorF); |
1045 | } |
1046 | |
1047 | /// Handle "(icmp eq/ne (shl AP2, A), AP1)" -> |
1048 | /// (icmp eq/ne A, TrailingZeros(AP1) - TrailingZeros(AP2)). |
1049 | Instruction *InstCombinerImpl::foldICmpShlConstConst(ICmpInst &I, Value *A, |
1050 | const APInt &AP1, |
1051 | const APInt &AP2) { |
1052 | assert(I.isEquality() && "Cannot fold icmp gt/lt" ); |
1053 | |
1054 | auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) { |
1055 | if (I.getPredicate() == I.ICMP_NE) |
1056 | Pred = CmpInst::getInversePredicate(pred: Pred); |
1057 | return new ICmpInst(Pred, LHS, RHS); |
1058 | }; |
1059 | |
1060 | // Don't bother doing any work for cases which InstSimplify handles. |
1061 | if (AP2.isZero()) |
1062 | return nullptr; |
1063 | |
1064 | unsigned AP2TrailingZeros = AP2.countr_zero(); |
1065 | |
1066 | if (!AP1 && AP2TrailingZeros != 0) |
1067 | return getICmp( |
1068 | I.ICMP_UGE, A, |
1069 | ConstantInt::get(Ty: A->getType(), V: AP2.getBitWidth() - AP2TrailingZeros)); |
1070 | |
1071 | if (AP1 == AP2) |
1072 | return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(Ty: A->getType())); |
1073 | |
1074 | // Get the distance between the lowest bits that are set. |
1075 | int Shift = AP1.countr_zero() - AP2TrailingZeros; |
1076 | |
1077 | if (Shift > 0 && AP2.shl(shiftAmt: Shift) == AP1) |
1078 | return getICmp(I.ICMP_EQ, A, ConstantInt::get(Ty: A->getType(), V: Shift)); |
1079 | |
1080 | // Shifting const2 will never be equal to const1. |
1081 | // FIXME: This should always be handled by InstSimplify? |
1082 | auto *TorF = ConstantInt::get(Ty: I.getType(), V: I.getPredicate() == I.ICMP_NE); |
1083 | return replaceInstUsesWith(I, V: TorF); |
1084 | } |
1085 | |
1086 | /// The caller has matched a pattern of the form: |
1087 | /// I = icmp ugt (add (add A, B), CI2), CI1 |
1088 | /// If this is of the form: |
1089 | /// sum = a + b |
1090 | /// if (sum+128 >u 255) |
1091 | /// Then replace it with llvm.sadd.with.overflow.i8. |
1092 | /// |
1093 | static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B, |
1094 | ConstantInt *CI2, ConstantInt *CI1, |
1095 | InstCombinerImpl &IC) { |
1096 | // The transformation we're trying to do here is to transform this into an |
1097 | // llvm.sadd.with.overflow. To do this, we have to replace the original add |
1098 | // with a narrower add, and discard the add-with-constant that is part of the |
1099 | // range check (if we can't eliminate it, this isn't profitable). |
1100 | |
1101 | // In order to eliminate the add-with-constant, the compare can be its only |
1102 | // use. |
1103 | Instruction *AddWithCst = cast<Instruction>(Val: I.getOperand(i_nocapture: 0)); |
1104 | if (!AddWithCst->hasOneUse()) |
1105 | return nullptr; |
1106 | |
1107 | // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow. |
1108 | if (!CI2->getValue().isPowerOf2()) |
1109 | return nullptr; |
1110 | unsigned NewWidth = CI2->getValue().countr_zero(); |
1111 | if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) |
1112 | return nullptr; |
1113 | |
1114 | // The width of the new add formed is 1 more than the bias. |
1115 | ++NewWidth; |
1116 | |
1117 | // Check to see that CI1 is an all-ones value with NewWidth bits. |
1118 | if (CI1->getBitWidth() == NewWidth || |
1119 | CI1->getValue() != APInt::getLowBitsSet(numBits: CI1->getBitWidth(), loBitsSet: NewWidth)) |
1120 | return nullptr; |
1121 | |
1122 | // This is only really a signed overflow check if the inputs have been |
1123 | // sign-extended; check for that condition. For example, if CI2 is 2^31 and |
1124 | // the operands of the add are 64 bits wide, we need at least 33 sign bits. |
1125 | if (IC.ComputeMaxSignificantBits(Op: A, CxtI: &I) > NewWidth || |
1126 | IC.ComputeMaxSignificantBits(Op: B, CxtI: &I) > NewWidth) |
1127 | return nullptr; |
1128 | |
1129 | // In order to replace the original add with a narrower |
1130 | // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant |
1131 | // and truncates that discard the high bits of the add. Verify that this is |
1132 | // the case. |
1133 | Instruction *OrigAdd = cast<Instruction>(Val: AddWithCst->getOperand(i: 0)); |
1134 | for (User *U : OrigAdd->users()) { |
1135 | if (U == AddWithCst) |
1136 | continue; |
1137 | |
1138 | // Only accept truncates for now. We would really like a nice recursive |
1139 | // predicate like SimplifyDemandedBits, but which goes downwards the use-def |
1140 | // chain to see which bits of a value are actually demanded. If the |
1141 | // original add had another add which was then immediately truncated, we |
1142 | // could still do the transformation. |
1143 | TruncInst *TI = dyn_cast<TruncInst>(Val: U); |
1144 | if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth) |
1145 | return nullptr; |
1146 | } |
1147 | |
1148 | // If the pattern matches, truncate the inputs to the narrower type and |
1149 | // use the sadd_with_overflow intrinsic to efficiently compute both the |
1150 | // result and the overflow bit. |
1151 | Type *NewType = IntegerType::get(C&: OrigAdd->getContext(), NumBits: NewWidth); |
1152 | Function *F = Intrinsic::getOrInsertDeclaration( |
1153 | M: I.getModule(), id: Intrinsic::sadd_with_overflow, Tys: NewType); |
1154 | |
1155 | InstCombiner::BuilderTy &Builder = IC.Builder; |
1156 | |
1157 | // Put the new code above the original add, in case there are any uses of the |
1158 | // add between the add and the compare. |
1159 | Builder.SetInsertPoint(OrigAdd); |
1160 | |
1161 | Value *TruncA = Builder.CreateTrunc(V: A, DestTy: NewType, Name: A->getName() + ".trunc" ); |
1162 | Value *TruncB = Builder.CreateTrunc(V: B, DestTy: NewType, Name: B->getName() + ".trunc" ); |
1163 | CallInst *Call = Builder.CreateCall(Callee: F, Args: {TruncA, TruncB}, Name: "sadd" ); |
1164 | Value *Add = Builder.CreateExtractValue(Agg: Call, Idxs: 0, Name: "sadd.result" ); |
1165 | Value *ZExt = Builder.CreateZExt(V: Add, DestTy: OrigAdd->getType()); |
1166 | |
1167 | // The inner add was the result of the narrow add, zero extended to the |
1168 | // wider type. Replace it with the result computed by the intrinsic. |
1169 | IC.replaceInstUsesWith(I&: *OrigAdd, V: ZExt); |
1170 | IC.eraseInstFromFunction(I&: *OrigAdd); |
1171 | |
1172 | // The original icmp gets replaced with the overflow value. |
1173 | return ExtractValueInst::Create(Agg: Call, Idxs: 1, NameStr: "sadd.overflow" ); |
1174 | } |
1175 | |
1176 | /// If we have: |
1177 | /// icmp eq/ne (urem/srem %x, %y), 0 |
1178 | /// iff %y is a power-of-two, we can replace this with a bit test: |
1179 | /// icmp eq/ne (and %x, (add %y, -1)), 0 |
1180 | Instruction *InstCombinerImpl::foldIRemByPowerOfTwoToBitTest(ICmpInst &I) { |
1181 | // This fold is only valid for equality predicates. |
1182 | if (!I.isEquality()) |
1183 | return nullptr; |
1184 | CmpPredicate Pred; |
1185 | Value *X, *Y, *Zero; |
1186 | if (!match(V: &I, P: m_ICmp(Pred, L: m_OneUse(SubPattern: m_IRem(L: m_Value(V&: X), R: m_Value(V&: Y))), |
1187 | R: m_CombineAnd(L: m_Zero(), R: m_Value(V&: Zero))))) |
1188 | return nullptr; |
1189 | if (!isKnownToBeAPowerOfTwo(V: Y, /*OrZero*/ true, CxtI: &I)) |
1190 | return nullptr; |
1191 | // This may increase instruction count, we don't enforce that Y is a constant. |
1192 | Value *Mask = Builder.CreateAdd(LHS: Y, RHS: Constant::getAllOnesValue(Ty: Y->getType())); |
1193 | Value *Masked = Builder.CreateAnd(LHS: X, RHS: Mask); |
1194 | return ICmpInst::Create(Op: Instruction::ICmp, Pred, S1: Masked, S2: Zero); |
1195 | } |
1196 | |
1197 | /// Fold equality-comparison between zero and any (maybe truncated) right-shift |
1198 | /// by one-less-than-bitwidth into a sign test on the original value. |
1199 | Instruction *InstCombinerImpl::foldSignBitTest(ICmpInst &I) { |
1200 | Instruction *Val; |
1201 | CmpPredicate Pred; |
1202 | if (!I.isEquality() || !match(V: &I, P: m_ICmp(Pred, L: m_Instruction(I&: Val), R: m_Zero()))) |
1203 | return nullptr; |
1204 | |
1205 | Value *X; |
1206 | Type *XTy; |
1207 | |
1208 | Constant *C; |
1209 | if (match(V: Val, P: m_TruncOrSelf(Op: m_Shr(L: m_Value(V&: X), R: m_Constant(C))))) { |
1210 | XTy = X->getType(); |
1211 | unsigned XBitWidth = XTy->getScalarSizeInBits(); |
1212 | if (!match(V: C, P: m_SpecificInt_ICMP(Predicate: ICmpInst::Predicate::ICMP_EQ, |
1213 | Threshold: APInt(XBitWidth, XBitWidth - 1)))) |
1214 | return nullptr; |
1215 | } else if (isa<BinaryOperator>(Val) && |
1216 | (X = reassociateShiftAmtsOfTwoSameDirectionShifts( |
1217 | Sh0: cast<BinaryOperator>(Val), SQ: SQ.getWithInstruction(I: Val), |
1218 | /*AnalyzeForSignBitExtraction=*/true))) { |
1219 | XTy = X->getType(); |
1220 | } else |
1221 | return nullptr; |
1222 | |
1223 | return ICmpInst::Create(Op: Instruction::ICmp, |
1224 | Pred: Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_SGE |
1225 | : ICmpInst::ICMP_SLT, |
1226 | S1: X, S2: ConstantInt::getNullValue(Ty: XTy)); |
1227 | } |
1228 | |
1229 | // Handle icmp pred X, 0 |
1230 | Instruction *InstCombinerImpl::foldICmpWithZero(ICmpInst &Cmp) { |
1231 | CmpInst::Predicate Pred = Cmp.getPredicate(); |
1232 | if (!match(V: Cmp.getOperand(i_nocapture: 1), P: m_Zero())) |
1233 | return nullptr; |
1234 | |
1235 | // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0) |
1236 | if (Pred == ICmpInst::ICMP_SGT) { |
1237 | Value *A, *B; |
1238 | if (match(V: Cmp.getOperand(i_nocapture: 0), P: m_SMin(L: m_Value(V&: A), R: m_Value(V&: B)))) { |
1239 | if (isKnownPositive(V: A, SQ: SQ.getWithInstruction(I: &Cmp))) |
1240 | return new ICmpInst(Pred, B, Cmp.getOperand(i_nocapture: 1)); |
1241 | if (isKnownPositive(V: B, SQ: SQ.getWithInstruction(I: &Cmp))) |
1242 | return new ICmpInst(Pred, A, Cmp.getOperand(i_nocapture: 1)); |
1243 | } |
1244 | } |
1245 | |
1246 | if (Instruction *New = foldIRemByPowerOfTwoToBitTest(I&: Cmp)) |
1247 | return New; |
1248 | |
1249 | // Given: |
1250 | // icmp eq/ne (urem %x, %y), 0 |
1251 | // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem': |
1252 | // icmp eq/ne %x, 0 |
1253 | Value *X, *Y; |
1254 | if (match(V: Cmp.getOperand(i_nocapture: 0), P: m_URem(L: m_Value(V&: X), R: m_Value(V&: Y))) && |
1255 | ICmpInst::isEquality(P: Pred)) { |
1256 | KnownBits XKnown = computeKnownBits(V: X, CxtI: &Cmp); |
1257 | KnownBits YKnown = computeKnownBits(V: Y, CxtI: &Cmp); |
1258 | if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2) |
1259 | return new ICmpInst(Pred, X, Cmp.getOperand(i_nocapture: 1)); |
1260 | } |
1261 | |
1262 | // (icmp eq/ne (mul X Y)) -> (icmp eq/ne X/Y) if we know about whether X/Y are |
1263 | // odd/non-zero/there is no overflow. |
1264 | if (match(V: Cmp.getOperand(i_nocapture: 0), P: m_Mul(L: m_Value(V&: X), R: m_Value(V&: Y))) && |
1265 | ICmpInst::isEquality(P: Pred)) { |
1266 | |
1267 | KnownBits XKnown = computeKnownBits(V: X, CxtI: &Cmp); |
1268 | // if X % 2 != 0 |
1269 | // (icmp eq/ne Y) |
1270 | if (XKnown.countMaxTrailingZeros() == 0) |
1271 | return new ICmpInst(Pred, Y, Cmp.getOperand(i_nocapture: 1)); |
1272 | |
1273 | KnownBits YKnown = computeKnownBits(V: Y, CxtI: &Cmp); |
1274 | // if Y % 2 != 0 |
1275 | // (icmp eq/ne X) |
1276 | if (YKnown.countMaxTrailingZeros() == 0) |
1277 | return new ICmpInst(Pred, X, Cmp.getOperand(i_nocapture: 1)); |
1278 | |
1279 | auto *BO0 = cast<OverflowingBinaryOperator>(Val: Cmp.getOperand(i_nocapture: 0)); |
1280 | if (BO0->hasNoUnsignedWrap() || BO0->hasNoSignedWrap()) { |
1281 | const SimplifyQuery Q = SQ.getWithInstruction(I: &Cmp); |
1282 | // `isKnownNonZero` does more analysis than just `!KnownBits.One.isZero()` |
1283 | // but to avoid unnecessary work, first just if this is an obvious case. |
1284 | |
1285 | // if X non-zero and NoOverflow(X * Y) |
1286 | // (icmp eq/ne Y) |
1287 | if (!XKnown.One.isZero() || isKnownNonZero(V: X, Q)) |
1288 | return new ICmpInst(Pred, Y, Cmp.getOperand(i_nocapture: 1)); |
1289 | |
1290 | // if Y non-zero and NoOverflow(X * Y) |
1291 | // (icmp eq/ne X) |
1292 | if (!YKnown.One.isZero() || isKnownNonZero(V: Y, Q)) |
1293 | return new ICmpInst(Pred, X, Cmp.getOperand(i_nocapture: 1)); |
1294 | } |
1295 | // Note, we are skipping cases: |
1296 | // if Y % 2 != 0 AND X % 2 != 0 |
1297 | // (false/true) |
1298 | // if X non-zero and Y non-zero and NoOverflow(X * Y) |
1299 | // (false/true) |
1300 | // Those can be simplified later as we would have already replaced the (icmp |
1301 | // eq/ne (mul X, Y)) with (icmp eq/ne X/Y) and if X/Y is known non-zero that |
1302 | // will fold to a constant elsewhere. |
1303 | } |
1304 | |
1305 | // (icmp eq/ne f(X), 0) -> (icmp eq/ne X, 0) |
1306 | // where f(X) == 0 if and only if X == 0 |
1307 | if (ICmpInst::isEquality(P: Pred)) |
1308 | if (Value *Stripped = stripNullTest(V: Cmp.getOperand(i_nocapture: 0))) |
1309 | return new ICmpInst(Pred, Stripped, |
1310 | Constant::getNullValue(Ty: Stripped->getType())); |
1311 | |
1312 | return nullptr; |
1313 | } |
1314 | |
1315 | /// Fold icmp Pred X, C. |
1316 | /// TODO: This code structure does not make sense. The saturating add fold |
1317 | /// should be moved to some other helper and extended as noted below (it is also |
1318 | /// possible that code has been made unnecessary - do we canonicalize IR to |
1319 | /// overflow/saturating intrinsics or not?). |
1320 | Instruction *InstCombinerImpl::foldICmpWithConstant(ICmpInst &Cmp) { |
1321 | // Match the following pattern, which is a common idiom when writing |
1322 | // overflow-safe integer arithmetic functions. The source performs an addition |
1323 | // in wider type and explicitly checks for overflow using comparisons against |
1324 | // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic. |
1325 | // |
1326 | // TODO: This could probably be generalized to handle other overflow-safe |
1327 | // operations if we worked out the formulas to compute the appropriate magic |
1328 | // constants. |
1329 | // |
1330 | // sum = a + b |
1331 | // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8 |
1332 | CmpInst::Predicate Pred = Cmp.getPredicate(); |
1333 | Value *Op0 = Cmp.getOperand(i_nocapture: 0), *Op1 = Cmp.getOperand(i_nocapture: 1); |
1334 | Value *A, *B; |
1335 | ConstantInt *CI, *CI2; // I = icmp ugt (add (add A, B), CI2), CI |
1336 | if (Pred == ICmpInst::ICMP_UGT && match(V: Op1, P: m_ConstantInt(CI)) && |
1337 | match(V: Op0, P: m_Add(L: m_Add(L: m_Value(V&: A), R: m_Value(V&: B)), R: m_ConstantInt(CI&: CI2)))) |
1338 | if (Instruction *Res = processUGT_ADDCST_ADD(I&: Cmp, A, B, CI2, CI1: CI, IC&: *this)) |
1339 | return Res; |
1340 | |
1341 | // icmp(phi(C1, C2, ...), C) -> phi(icmp(C1, C), icmp(C2, C), ...). |
1342 | Constant *C = dyn_cast<Constant>(Val: Op1); |
1343 | if (!C) |
1344 | return nullptr; |
1345 | |
1346 | if (auto *Phi = dyn_cast<PHINode>(Val: Op0)) |
1347 | if (all_of(Range: Phi->operands(), P: [](Value *V) { return isa<Constant>(Val: V); })) { |
1348 | SmallVector<Constant *> Ops; |
1349 | for (Value *V : Phi->incoming_values()) { |
1350 | Constant *Res = |
1351 | ConstantFoldCompareInstOperands(Predicate: Pred, LHS: cast<Constant>(Val: V), RHS: C, DL); |
1352 | if (!Res) |
1353 | return nullptr; |
1354 | Ops.push_back(Elt: Res); |
1355 | } |
1356 | Builder.SetInsertPoint(Phi); |
1357 | PHINode *NewPhi = Builder.CreatePHI(Ty: Cmp.getType(), NumReservedValues: Phi->getNumOperands()); |
1358 | for (auto [V, Pred] : zip(t&: Ops, u: Phi->blocks())) |
1359 | NewPhi->addIncoming(V, BB: Pred); |
1360 | return replaceInstUsesWith(I&: Cmp, V: NewPhi); |
1361 | } |
1362 | |
1363 | if (Instruction *R = tryFoldInstWithCtpopWithNot(I: &Cmp)) |
1364 | return R; |
1365 | |
1366 | return nullptr; |
1367 | } |
1368 | |
1369 | /// Canonicalize icmp instructions based on dominating conditions. |
1370 | Instruction *InstCombinerImpl::foldICmpWithDominatingICmp(ICmpInst &Cmp) { |
1371 | // We already checked simple implication in InstSimplify, only handle complex |
1372 | // cases here. |
1373 | Value *X = Cmp.getOperand(i_nocapture: 0), *Y = Cmp.getOperand(i_nocapture: 1); |
1374 | const APInt *C; |
1375 | if (!match(V: Y, P: m_APInt(Res&: C))) |
1376 | return nullptr; |
1377 | |
1378 | CmpInst::Predicate Pred = Cmp.getPredicate(); |
1379 | ConstantRange CR = ConstantRange::makeExactICmpRegion(Pred, Other: *C); |
1380 | |
1381 | auto handleDomCond = [&](ICmpInst::Predicate DomPred, |
1382 | const APInt *DomC) -> Instruction * { |
1383 | // We have 2 compares of a variable with constants. Calculate the constant |
1384 | // ranges of those compares to see if we can transform the 2nd compare: |
1385 | // DomBB: |
1386 | // DomCond = icmp DomPred X, DomC |
1387 | // br DomCond, CmpBB, FalseBB |
1388 | // CmpBB: |
1389 | // Cmp = icmp Pred X, C |
1390 | ConstantRange DominatingCR = |
1391 | ConstantRange::makeExactICmpRegion(Pred: DomPred, Other: *DomC); |
1392 | ConstantRange Intersection = DominatingCR.intersectWith(CR); |
1393 | ConstantRange Difference = DominatingCR.difference(CR); |
1394 | if (Intersection.isEmptySet()) |
1395 | return replaceInstUsesWith(I&: Cmp, V: Builder.getFalse()); |
1396 | if (Difference.isEmptySet()) |
1397 | return replaceInstUsesWith(I&: Cmp, V: Builder.getTrue()); |
1398 | |
1399 | // Canonicalizing a sign bit comparison that gets used in a branch, |
1400 | // pessimizes codegen by generating branch on zero instruction instead |
1401 | // of a test and branch. So we avoid canonicalizing in such situations |
1402 | // because test and branch instruction has better branch displacement |
1403 | // than compare and branch instruction. |
1404 | bool UnusedBit; |
1405 | bool IsSignBit = isSignBitCheck(Pred, RHS: *C, TrueIfSigned&: UnusedBit); |
1406 | if (Cmp.isEquality() || (IsSignBit && hasBranchUse(I&: Cmp))) |
1407 | return nullptr; |
1408 | |
1409 | // Avoid an infinite loop with min/max canonicalization. |
1410 | // TODO: This will be unnecessary if we canonicalize to min/max intrinsics. |
1411 | if (Cmp.hasOneUse() && |
1412 | match(V: Cmp.user_back(), P: m_MaxOrMin(L: m_Value(), R: m_Value()))) |
1413 | return nullptr; |
1414 | |
1415 | if (const APInt *EqC = Intersection.getSingleElement()) |
1416 | return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder.getInt(AI: *EqC)); |
1417 | if (const APInt *NeC = Difference.getSingleElement()) |
1418 | return new ICmpInst(ICmpInst::ICMP_NE, X, Builder.getInt(AI: *NeC)); |
1419 | return nullptr; |
1420 | }; |
1421 | |
1422 | for (BranchInst *BI : DC.conditionsFor(V: X)) { |
1423 | CmpPredicate DomPred; |
1424 | const APInt *DomC; |
1425 | if (!match(V: BI->getCondition(), |
1426 | P: m_ICmp(Pred&: DomPred, L: m_Specific(V: X), R: m_APInt(Res&: DomC)))) |
1427 | continue; |
1428 | |
1429 | BasicBlockEdge Edge0(BI->getParent(), BI->getSuccessor(i: 0)); |
1430 | if (DT.dominates(BBE: Edge0, BB: Cmp.getParent())) { |
1431 | if (auto *V = handleDomCond(DomPred, DomC)) |
1432 | return V; |
1433 | } else { |
1434 | BasicBlockEdge Edge1(BI->getParent(), BI->getSuccessor(i: 1)); |
1435 | if (DT.dominates(BBE: Edge1, BB: Cmp.getParent())) |
1436 | if (auto *V = |
1437 | handleDomCond(CmpInst::getInversePredicate(pred: DomPred), DomC)) |
1438 | return V; |
1439 | } |
1440 | } |
1441 | |
1442 | return nullptr; |
1443 | } |
1444 | |
1445 | /// Fold icmp (trunc X), C. |
1446 | Instruction *InstCombinerImpl::foldICmpTruncConstant(ICmpInst &Cmp, |
1447 | TruncInst *Trunc, |
1448 | const APInt &C) { |
1449 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
1450 | Value *X = Trunc->getOperand(i_nocapture: 0); |
1451 | Type *SrcTy = X->getType(); |
1452 | unsigned DstBits = Trunc->getType()->getScalarSizeInBits(), |
1453 | SrcBits = SrcTy->getScalarSizeInBits(); |
1454 | |
1455 | // Match (icmp pred (trunc nuw/nsw X), C) |
1456 | // Which we can convert to (icmp pred X, (sext/zext C)) |
1457 | if (shouldChangeType(From: Trunc->getType(), To: SrcTy)) { |
1458 | if (Trunc->hasNoSignedWrap()) |
1459 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: SrcTy, V: C.sext(width: SrcBits))); |
1460 | if (!Cmp.isSigned() && Trunc->hasNoUnsignedWrap()) |
1461 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: SrcTy, V: C.zext(width: SrcBits))); |
1462 | } |
1463 | |
1464 | if (C.isOne() && C.getBitWidth() > 1) { |
1465 | // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1 |
1466 | Value *V = nullptr; |
1467 | if (Pred == ICmpInst::ICMP_SLT && match(V: X, P: m_Signum(V: m_Value(V)))) |
1468 | return new ICmpInst(ICmpInst::ICMP_SLT, V, |
1469 | ConstantInt::get(Ty: V->getType(), V: 1)); |
1470 | } |
1471 | |
1472 | // TODO: Handle any shifted constant by subtracting trailing zeros. |
1473 | // TODO: Handle non-equality predicates. |
1474 | Value *Y; |
1475 | if (Cmp.isEquality() && match(V: X, P: m_Shl(L: m_One(), R: m_Value(V&: Y)))) { |
1476 | // (trunc (1 << Y) to iN) == 0 --> Y u>= N |
1477 | // (trunc (1 << Y) to iN) != 0 --> Y u< N |
1478 | if (C.isZero()) { |
1479 | auto NewPred = (Pred == Cmp.ICMP_EQ) ? Cmp.ICMP_UGE : Cmp.ICMP_ULT; |
1480 | return new ICmpInst(NewPred, Y, ConstantInt::get(Ty: SrcTy, V: DstBits)); |
1481 | } |
1482 | // (trunc (1 << Y) to iN) == 2**C --> Y == C |
1483 | // (trunc (1 << Y) to iN) != 2**C --> Y != C |
1484 | if (C.isPowerOf2()) |
1485 | return new ICmpInst(Pred, Y, ConstantInt::get(Ty: SrcTy, V: C.logBase2())); |
1486 | } |
1487 | |
1488 | if (Cmp.isEquality() && Trunc->hasOneUse()) { |
1489 | // Canonicalize to a mask and wider compare if the wide type is suitable: |
1490 | // (trunc X to i8) == C --> (X & 0xff) == (zext C) |
1491 | if (!SrcTy->isVectorTy() && shouldChangeType(FromBitWidth: DstBits, ToBitWidth: SrcBits)) { |
1492 | Constant *Mask = |
1493 | ConstantInt::get(Ty: SrcTy, V: APInt::getLowBitsSet(numBits: SrcBits, loBitsSet: DstBits)); |
1494 | Value *And = Builder.CreateAnd(LHS: X, RHS: Mask); |
1495 | Constant *WideC = ConstantInt::get(Ty: SrcTy, V: C.zext(width: SrcBits)); |
1496 | return new ICmpInst(Pred, And, WideC); |
1497 | } |
1498 | |
1499 | // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all |
1500 | // of the high bits truncated out of x are known. |
1501 | KnownBits Known = computeKnownBits(V: X, CxtI: &Cmp); |
1502 | |
1503 | // If all the high bits are known, we can do this xform. |
1504 | if ((Known.Zero | Known.One).countl_one() >= SrcBits - DstBits) { |
1505 | // Pull in the high bits from known-ones set. |
1506 | APInt NewRHS = C.zext(width: SrcBits); |
1507 | NewRHS |= Known.One & APInt::getHighBitsSet(numBits: SrcBits, hiBitsSet: SrcBits - DstBits); |
1508 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: SrcTy, V: NewRHS)); |
1509 | } |
1510 | } |
1511 | |
1512 | // Look through truncated right-shift of the sign-bit for a sign-bit check: |
1513 | // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] < 0 --> ShOp < 0 |
1514 | // trunc iN (ShOp >> ShAmtC) to i[N - ShAmtC] > -1 --> ShOp > -1 |
1515 | Value *ShOp; |
1516 | const APInt *ShAmtC; |
1517 | bool TrueIfSigned; |
1518 | if (isSignBitCheck(Pred, RHS: C, TrueIfSigned) && |
1519 | match(V: X, P: m_Shr(L: m_Value(V&: ShOp), R: m_APInt(Res&: ShAmtC))) && |
1520 | DstBits == SrcBits - ShAmtC->getZExtValue()) { |
1521 | return TrueIfSigned ? new ICmpInst(ICmpInst::ICMP_SLT, ShOp, |
1522 | ConstantInt::getNullValue(Ty: SrcTy)) |
1523 | : new ICmpInst(ICmpInst::ICMP_SGT, ShOp, |
1524 | ConstantInt::getAllOnesValue(Ty: SrcTy)); |
1525 | } |
1526 | |
1527 | return nullptr; |
1528 | } |
1529 | |
1530 | /// Fold icmp (trunc nuw/nsw X), (trunc nuw/nsw Y). |
1531 | /// Fold icmp (trunc nuw/nsw X), (zext/sext Y). |
1532 | Instruction * |
1533 | InstCombinerImpl::foldICmpTruncWithTruncOrExt(ICmpInst &Cmp, |
1534 | const SimplifyQuery &Q) { |
1535 | Value *X, *Y; |
1536 | CmpPredicate Pred; |
1537 | bool YIsSExt = false; |
1538 | // Try to match icmp (trunc X), (trunc Y) |
1539 | if (match(V: &Cmp, P: m_ICmp(Pred, L: m_Trunc(Op: m_Value(V&: X)), R: m_Trunc(Op: m_Value(V&: Y))))) { |
1540 | unsigned NoWrapFlags = cast<TruncInst>(Val: Cmp.getOperand(i_nocapture: 0))->getNoWrapKind() & |
1541 | cast<TruncInst>(Val: Cmp.getOperand(i_nocapture: 1))->getNoWrapKind(); |
1542 | if (Cmp.isSigned()) { |
1543 | // For signed comparisons, both truncs must be nsw. |
1544 | if (!(NoWrapFlags & TruncInst::NoSignedWrap)) |
1545 | return nullptr; |
1546 | } else { |
1547 | // For unsigned and equality comparisons, either both must be nuw or |
1548 | // both must be nsw, we don't care which. |
1549 | if (!NoWrapFlags) |
1550 | return nullptr; |
1551 | } |
1552 | |
1553 | if (X->getType() != Y->getType() && |
1554 | (!Cmp.getOperand(i_nocapture: 0)->hasOneUse() || !Cmp.getOperand(i_nocapture: 1)->hasOneUse())) |
1555 | return nullptr; |
1556 | if (!isDesirableIntType(BitWidth: X->getType()->getScalarSizeInBits()) && |
1557 | isDesirableIntType(BitWidth: Y->getType()->getScalarSizeInBits())) { |
1558 | std::swap(a&: X, b&: Y); |
1559 | Pred = Cmp.getSwappedPredicate(pred: Pred); |
1560 | } |
1561 | YIsSExt = !(NoWrapFlags & TruncInst::NoUnsignedWrap); |
1562 | } |
1563 | // Try to match icmp (trunc nuw X), (zext Y) |
1564 | else if (!Cmp.isSigned() && |
1565 | match(V: &Cmp, P: m_c_ICmp(Pred, L: m_NUWTrunc(Op: m_Value(V&: X)), |
1566 | R: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: Y)))))) { |
1567 | // Can fold trunc nuw + zext for unsigned and equality predicates. |
1568 | } |
1569 | // Try to match icmp (trunc nsw X), (sext Y) |
1570 | else if (match(V: &Cmp, P: m_c_ICmp(Pred, L: m_NSWTrunc(Op: m_Value(V&: X)), |
1571 | R: m_OneUse(SubPattern: m_ZExtOrSExt(Op: m_Value(V&: Y)))))) { |
1572 | // Can fold trunc nsw + zext/sext for all predicates. |
1573 | YIsSExt = |
1574 | isa<SExtInst>(Val: Cmp.getOperand(i_nocapture: 0)) || isa<SExtInst>(Val: Cmp.getOperand(i_nocapture: 1)); |
1575 | } else |
1576 | return nullptr; |
1577 | |
1578 | Type *TruncTy = Cmp.getOperand(i_nocapture: 0)->getType(); |
1579 | unsigned TruncBits = TruncTy->getScalarSizeInBits(); |
1580 | |
1581 | // If this transform will end up changing from desirable types -> undesirable |
1582 | // types skip it. |
1583 | if (isDesirableIntType(BitWidth: TruncBits) && |
1584 | !isDesirableIntType(BitWidth: X->getType()->getScalarSizeInBits())) |
1585 | return nullptr; |
1586 | |
1587 | Value *NewY = Builder.CreateIntCast(V: Y, DestTy: X->getType(), isSigned: YIsSExt); |
1588 | return new ICmpInst(Pred, X, NewY); |
1589 | } |
1590 | |
1591 | /// Fold icmp (xor X, Y), C. |
1592 | Instruction *InstCombinerImpl::foldICmpXorConstant(ICmpInst &Cmp, |
1593 | BinaryOperator *Xor, |
1594 | const APInt &C) { |
1595 | if (Instruction *I = foldICmpXorShiftConst(Cmp, Xor, C)) |
1596 | return I; |
1597 | |
1598 | Value *X = Xor->getOperand(i_nocapture: 0); |
1599 | Value *Y = Xor->getOperand(i_nocapture: 1); |
1600 | const APInt *XorC; |
1601 | if (!match(V: Y, P: m_APInt(Res&: XorC))) |
1602 | return nullptr; |
1603 | |
1604 | // If this is a comparison that tests the signbit (X < 0) or (x > -1), |
1605 | // fold the xor. |
1606 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
1607 | bool TrueIfSigned = false; |
1608 | if (isSignBitCheck(Pred: Cmp.getPredicate(), RHS: C, TrueIfSigned)) { |
1609 | |
1610 | // If the sign bit of the XorCst is not set, there is no change to |
1611 | // the operation, just stop using the Xor. |
1612 | if (!XorC->isNegative()) |
1613 | return replaceOperand(I&: Cmp, OpNum: 0, V: X); |
1614 | |
1615 | // Emit the opposite comparison. |
1616 | if (TrueIfSigned) |
1617 | return new ICmpInst(ICmpInst::ICMP_SGT, X, |
1618 | ConstantInt::getAllOnesValue(Ty: X->getType())); |
1619 | else |
1620 | return new ICmpInst(ICmpInst::ICMP_SLT, X, |
1621 | ConstantInt::getNullValue(Ty: X->getType())); |
1622 | } |
1623 | |
1624 | if (Xor->hasOneUse()) { |
1625 | // (icmp u/s (xor X SignMask), C) -> (icmp s/u X, (xor C SignMask)) |
1626 | if (!Cmp.isEquality() && XorC->isSignMask()) { |
1627 | Pred = Cmp.getFlippedSignednessPredicate(); |
1628 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: X->getType(), V: C ^ *XorC)); |
1629 | } |
1630 | |
1631 | // (icmp u/s (xor X ~SignMask), C) -> (icmp s/u X, (xor C ~SignMask)) |
1632 | if (!Cmp.isEquality() && XorC->isMaxSignedValue()) { |
1633 | Pred = Cmp.getFlippedSignednessPredicate(); |
1634 | Pred = Cmp.getSwappedPredicate(pred: Pred); |
1635 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: X->getType(), V: C ^ *XorC)); |
1636 | } |
1637 | } |
1638 | |
1639 | // Mask constant magic can eliminate an 'xor' with unsigned compares. |
1640 | if (Pred == ICmpInst::ICMP_UGT) { |
1641 | // (xor X, ~C) >u C --> X <u ~C (when C+1 is a power of 2) |
1642 | if (*XorC == ~C && (C + 1).isPowerOf2()) |
1643 | return new ICmpInst(ICmpInst::ICMP_ULT, X, Y); |
1644 | // (xor X, C) >u C --> X >u C (when C+1 is a power of 2) |
1645 | if (*XorC == C && (C + 1).isPowerOf2()) |
1646 | return new ICmpInst(ICmpInst::ICMP_UGT, X, Y); |
1647 | } |
1648 | if (Pred == ICmpInst::ICMP_ULT) { |
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(Ty: X->getType(), V: ~C)); |
1653 | // (xor X, C) <u C --> X >u ~C (when -C is a power of 2) |
1654 | if (*XorC == C && (-C).isPowerOf2()) |
1655 | return new ICmpInst(ICmpInst::ICMP_UGT, X, |
1656 | ConstantInt::get(Ty: X->getType(), V: ~C)); |
1657 | } |
1658 | return nullptr; |
1659 | } |
1660 | |
1661 | /// For power-of-2 C: |
1662 | /// ((X s>> ShiftC) ^ X) u< C --> (X + C) u< (C << 1) |
1663 | /// ((X s>> ShiftC) ^ X) u> (C - 1) --> (X + C) u> ((C << 1) - 1) |
1664 | Instruction *InstCombinerImpl::foldICmpXorShiftConst(ICmpInst &Cmp, |
1665 | BinaryOperator *Xor, |
1666 | const APInt &C) { |
1667 | CmpInst::Predicate Pred = Cmp.getPredicate(); |
1668 | APInt PowerOf2; |
1669 | if (Pred == ICmpInst::ICMP_ULT) |
1670 | PowerOf2 = C; |
1671 | else if (Pred == ICmpInst::ICMP_UGT && !C.isMaxValue()) |
1672 | PowerOf2 = C + 1; |
1673 | else |
1674 | return nullptr; |
1675 | if (!PowerOf2.isPowerOf2()) |
1676 | return nullptr; |
1677 | Value *X; |
1678 | const APInt *ShiftC; |
1679 | if (!match(V: Xor, P: m_OneUse(SubPattern: m_c_Xor(L: m_Value(V&: X), |
1680 | R: m_AShr(L: m_Deferred(V: X), R: m_APInt(Res&: ShiftC)))))) |
1681 | return nullptr; |
1682 | uint64_t Shift = ShiftC->getLimitedValue(); |
1683 | Type *XType = X->getType(); |
1684 | if (Shift == 0 || PowerOf2.isMinSignedValue()) |
1685 | return nullptr; |
1686 | Value *Add = Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty: XType, V: PowerOf2)); |
1687 | APInt Bound = |
1688 | Pred == ICmpInst::ICMP_ULT ? PowerOf2 << 1 : ((PowerOf2 << 1) - 1); |
1689 | return new ICmpInst(Pred, Add, ConstantInt::get(Ty: XType, V: Bound)); |
1690 | } |
1691 | |
1692 | /// Fold icmp (and (sh X, Y), C2), C1. |
1693 | Instruction *InstCombinerImpl::foldICmpAndShift(ICmpInst &Cmp, |
1694 | BinaryOperator *And, |
1695 | const APInt &C1, |
1696 | const APInt &C2) { |
1697 | BinaryOperator *Shift = dyn_cast<BinaryOperator>(Val: And->getOperand(i_nocapture: 0)); |
1698 | if (!Shift || !Shift->isShift()) |
1699 | return nullptr; |
1700 | |
1701 | // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could |
1702 | // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in |
1703 | // code produced by the clang front-end, for bitfield access. |
1704 | // This seemingly simple opportunity to fold away a shift turns out to be |
1705 | // rather complicated. See PR17827 for details. |
1706 | unsigned ShiftOpcode = Shift->getOpcode(); |
1707 | bool IsShl = ShiftOpcode == Instruction::Shl; |
1708 | const APInt *C3; |
1709 | if (match(V: Shift->getOperand(i_nocapture: 1), P: m_APInt(Res&: C3))) { |
1710 | APInt NewAndCst, NewCmpCst; |
1711 | bool AnyCmpCstBitsShiftedOut; |
1712 | if (ShiftOpcode == Instruction::Shl) { |
1713 | // For a left shift, we can fold if the comparison is not signed. We can |
1714 | // also fold a signed comparison if the mask value and comparison value |
1715 | // are not negative. These constraints may not be obvious, but we can |
1716 | // prove that they are correct using an SMT solver. |
1717 | if (Cmp.isSigned() && (C2.isNegative() || C1.isNegative())) |
1718 | return nullptr; |
1719 | |
1720 | NewCmpCst = C1.lshr(ShiftAmt: *C3); |
1721 | NewAndCst = C2.lshr(ShiftAmt: *C3); |
1722 | AnyCmpCstBitsShiftedOut = NewCmpCst.shl(ShiftAmt: *C3) != C1; |
1723 | } else if (ShiftOpcode == Instruction::LShr) { |
1724 | // For a logical right shift, we can fold if the comparison is not signed. |
1725 | // We can also fold a signed comparison if the shifted mask value and the |
1726 | // shifted comparison value are not negative. These constraints may not be |
1727 | // obvious, but we can prove that they are correct using an SMT solver. |
1728 | NewCmpCst = C1.shl(ShiftAmt: *C3); |
1729 | NewAndCst = C2.shl(ShiftAmt: *C3); |
1730 | AnyCmpCstBitsShiftedOut = NewCmpCst.lshr(ShiftAmt: *C3) != C1; |
1731 | if (Cmp.isSigned() && (NewAndCst.isNegative() || NewCmpCst.isNegative())) |
1732 | return nullptr; |
1733 | } else { |
1734 | // For an arithmetic shift, check that both constants don't use (in a |
1735 | // signed sense) the top bits being shifted out. |
1736 | assert(ShiftOpcode == Instruction::AShr && "Unknown shift opcode" ); |
1737 | NewCmpCst = C1.shl(ShiftAmt: *C3); |
1738 | NewAndCst = C2.shl(ShiftAmt: *C3); |
1739 | AnyCmpCstBitsShiftedOut = NewCmpCst.ashr(ShiftAmt: *C3) != C1; |
1740 | if (NewAndCst.ashr(ShiftAmt: *C3) != C2) |
1741 | return nullptr; |
1742 | } |
1743 | |
1744 | if (AnyCmpCstBitsShiftedOut) { |
1745 | // If we shifted bits out, the fold is not going to work out. As a |
1746 | // special case, check to see if this means that the result is always |
1747 | // true or false now. |
1748 | if (Cmp.getPredicate() == ICmpInst::ICMP_EQ) |
1749 | return replaceInstUsesWith(I&: Cmp, V: ConstantInt::getFalse(Ty: Cmp.getType())); |
1750 | if (Cmp.getPredicate() == ICmpInst::ICMP_NE) |
1751 | return replaceInstUsesWith(I&: Cmp, V: ConstantInt::getTrue(Ty: Cmp.getType())); |
1752 | } else { |
1753 | Value *NewAnd = Builder.CreateAnd( |
1754 | LHS: Shift->getOperand(i_nocapture: 0), RHS: ConstantInt::get(Ty: And->getType(), V: NewAndCst)); |
1755 | return new ICmpInst(Cmp.getPredicate(), NewAnd, |
1756 | ConstantInt::get(Ty: And->getType(), V: NewCmpCst)); |
1757 | } |
1758 | } |
1759 | |
1760 | // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is |
1761 | // preferable because it allows the C2 << Y expression to be hoisted out of a |
1762 | // loop if Y is invariant and X is not. |
1763 | if (Shift->hasOneUse() && C1.isZero() && Cmp.isEquality() && |
1764 | !Shift->isArithmeticShift() && |
1765 | ((!IsShl && C2.isOne()) || !isa<Constant>(Val: Shift->getOperand(i_nocapture: 0)))) { |
1766 | // Compute C2 << Y. |
1767 | Value *NewShift = |
1768 | IsShl ? Builder.CreateLShr(LHS: And->getOperand(i_nocapture: 1), RHS: Shift->getOperand(i_nocapture: 1)) |
1769 | : Builder.CreateShl(LHS: And->getOperand(i_nocapture: 1), RHS: Shift->getOperand(i_nocapture: 1)); |
1770 | |
1771 | // Compute X & (C2 << Y). |
1772 | Value *NewAnd = Builder.CreateAnd(LHS: Shift->getOperand(i_nocapture: 0), RHS: NewShift); |
1773 | return new ICmpInst(Cmp.getPredicate(), NewAnd, Cmp.getOperand(i_nocapture: 1)); |
1774 | } |
1775 | |
1776 | return nullptr; |
1777 | } |
1778 | |
1779 | /// Fold icmp (and X, C2), C1. |
1780 | Instruction *InstCombinerImpl::foldICmpAndConstConst(ICmpInst &Cmp, |
1781 | BinaryOperator *And, |
1782 | const APInt &C1) { |
1783 | bool isICMP_NE = Cmp.getPredicate() == ICmpInst::ICMP_NE; |
1784 | |
1785 | // For vectors: icmp ne (and X, 1), 0 --> trunc X to N x i1 |
1786 | // TODO: We canonicalize to the longer form for scalars because we have |
1787 | // better analysis/folds for icmp, and codegen may be better with icmp. |
1788 | if (isICMP_NE && Cmp.getType()->isVectorTy() && C1.isZero() && |
1789 | match(V: And->getOperand(i_nocapture: 1), P: m_One())) |
1790 | return new TruncInst(And->getOperand(i_nocapture: 0), Cmp.getType()); |
1791 | |
1792 | const APInt *C2; |
1793 | Value *X; |
1794 | if (!match(V: And, P: m_And(L: m_Value(V&: X), R: m_APInt(Res&: C2)))) |
1795 | return nullptr; |
1796 | |
1797 | // (and X, highmask) s> [0, ~highmask] --> X s> ~highmask |
1798 | if (Cmp.getPredicate() == ICmpInst::ICMP_SGT && C1.ule(RHS: ~*C2) && |
1799 | C2->isNegatedPowerOf2()) |
1800 | return new ICmpInst(ICmpInst::ICMP_SGT, X, |
1801 | ConstantInt::get(Ty: X->getType(), V: ~*C2)); |
1802 | // (and X, highmask) s< [1, -highmask] --> X s< -highmask |
1803 | if (Cmp.getPredicate() == ICmpInst::ICMP_SLT && !C1.isSignMask() && |
1804 | (C1 - 1).ule(RHS: ~*C2) && C2->isNegatedPowerOf2() && !C2->isSignMask()) |
1805 | return new ICmpInst(ICmpInst::ICMP_SLT, X, |
1806 | ConstantInt::get(Ty: X->getType(), V: -*C2)); |
1807 | |
1808 | // Don't perform the following transforms if the AND has multiple uses |
1809 | if (!And->hasOneUse()) |
1810 | return nullptr; |
1811 | |
1812 | if (Cmp.isEquality() && C1.isZero()) { |
1813 | // Restrict this fold to single-use 'and' (PR10267). |
1814 | // Replace (and X, (1 << size(X)-1) != 0) with X s< 0 |
1815 | if (C2->isSignMask()) { |
1816 | Constant *Zero = Constant::getNullValue(Ty: X->getType()); |
1817 | auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE; |
1818 | return new ICmpInst(NewPred, X, Zero); |
1819 | } |
1820 | |
1821 | APInt NewC2 = *C2; |
1822 | KnownBits Know = computeKnownBits(V: And->getOperand(i_nocapture: 0), CxtI: And); |
1823 | // Set high zeros of C2 to allow matching negated power-of-2. |
1824 | NewC2 = *C2 | APInt::getHighBitsSet(numBits: C2->getBitWidth(), |
1825 | hiBitsSet: Know.countMinLeadingZeros()); |
1826 | |
1827 | // Restrict this fold only for single-use 'and' (PR10267). |
1828 | // ((%x & C) == 0) --> %x u< (-C) iff (-C) is power of two. |
1829 | if (NewC2.isNegatedPowerOf2()) { |
1830 | Constant *NegBOC = ConstantInt::get(Ty: And->getType(), V: -NewC2); |
1831 | auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT; |
1832 | return new ICmpInst(NewPred, X, NegBOC); |
1833 | } |
1834 | } |
1835 | |
1836 | // If the LHS is an 'and' of a truncate and we can widen the and/compare to |
1837 | // the input width without changing the value produced, eliminate the cast: |
1838 | // |
1839 | // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1' |
1840 | // |
1841 | // We can do this transformation if the constants do not have their sign bits |
1842 | // set or if it is an equality comparison. Extending a relational comparison |
1843 | // when we're checking the sign bit would not work. |
1844 | Value *W; |
1845 | if (match(V: And->getOperand(i_nocapture: 0), P: m_OneUse(SubPattern: m_Trunc(Op: m_Value(V&: W)))) && |
1846 | (Cmp.isEquality() || (!C1.isNegative() && !C2->isNegative()))) { |
1847 | // TODO: Is this a good transform for vectors? Wider types may reduce |
1848 | // throughput. Should this transform be limited (even for scalars) by using |
1849 | // shouldChangeType()? |
1850 | if (!Cmp.getType()->isVectorTy()) { |
1851 | Type *WideType = W->getType(); |
1852 | unsigned WideScalarBits = WideType->getScalarSizeInBits(); |
1853 | Constant *ZextC1 = ConstantInt::get(Ty: WideType, V: C1.zext(width: WideScalarBits)); |
1854 | Constant *ZextC2 = ConstantInt::get(Ty: WideType, V: C2->zext(width: WideScalarBits)); |
1855 | Value *NewAnd = Builder.CreateAnd(LHS: W, RHS: ZextC2, Name: And->getName()); |
1856 | return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1); |
1857 | } |
1858 | } |
1859 | |
1860 | if (Instruction *I = foldICmpAndShift(Cmp, And, C1, C2: *C2)) |
1861 | return I; |
1862 | |
1863 | // (icmp pred (and (or (lshr A, B), A), 1), 0) --> |
1864 | // (icmp pred (and A, (or (shl 1, B), 1), 0)) |
1865 | // |
1866 | // iff pred isn't signed |
1867 | if (!Cmp.isSigned() && C1.isZero() && And->getOperand(i_nocapture: 0)->hasOneUse() && |
1868 | match(V: And->getOperand(i_nocapture: 1), P: m_One())) { |
1869 | Constant *One = cast<Constant>(Val: And->getOperand(i_nocapture: 1)); |
1870 | Value *Or = And->getOperand(i_nocapture: 0); |
1871 | Value *A, *B, *LShr; |
1872 | if (match(V: Or, P: m_Or(L: m_Value(V&: LShr), R: m_Value(V&: A))) && |
1873 | match(V: LShr, P: m_LShr(L: m_Specific(V: A), R: m_Value(V&: B)))) { |
1874 | unsigned UsesRemoved = 0; |
1875 | if (And->hasOneUse()) |
1876 | ++UsesRemoved; |
1877 | if (Or->hasOneUse()) |
1878 | ++UsesRemoved; |
1879 | if (LShr->hasOneUse()) |
1880 | ++UsesRemoved; |
1881 | |
1882 | // Compute A & ((1 << B) | 1) |
1883 | unsigned RequireUsesRemoved = match(V: B, P: m_ImmConstant()) ? 1 : 3; |
1884 | if (UsesRemoved >= RequireUsesRemoved) { |
1885 | Value *NewOr = |
1886 | Builder.CreateOr(LHS: Builder.CreateShl(LHS: One, RHS: B, Name: LShr->getName(), |
1887 | /*HasNUW=*/true), |
1888 | RHS: One, Name: Or->getName()); |
1889 | Value *NewAnd = Builder.CreateAnd(LHS: A, RHS: NewOr, Name: And->getName()); |
1890 | return new ICmpInst(Cmp.getPredicate(), NewAnd, Cmp.getOperand(i_nocapture: 1)); |
1891 | } |
1892 | } |
1893 | } |
1894 | |
1895 | // (icmp eq (and (bitcast X to int), ExponentMask), ExponentMask) --> |
1896 | // llvm.is.fpclass(X, fcInf|fcNan) |
1897 | // (icmp ne (and (bitcast X to int), ExponentMask), ExponentMask) --> |
1898 | // llvm.is.fpclass(X, ~(fcInf|fcNan)) |
1899 | // (icmp eq (and (bitcast X to int), ExponentMask), 0) --> |
1900 | // llvm.is.fpclass(X, fcSubnormal|fcZero) |
1901 | // (icmp ne (and (bitcast X to int), ExponentMask), 0) --> |
1902 | // llvm.is.fpclass(X, ~(fcSubnormal|fcZero)) |
1903 | Value *V; |
1904 | if (!Cmp.getParent()->getParent()->hasFnAttribute( |
1905 | Kind: Attribute::NoImplicitFloat) && |
1906 | Cmp.isEquality() && |
1907 | match(V: X, P: m_OneUse(SubPattern: m_ElementWiseBitCast(Op: m_Value(V))))) { |
1908 | Type *FPType = V->getType()->getScalarType(); |
1909 | if (FPType->isIEEELikeFPTy() && (C1.isZero() || C1 == *C2)) { |
1910 | APInt ExponentMask = |
1911 | APFloat::getInf(Sem: FPType->getFltSemantics()).bitcastToAPInt(); |
1912 | if (*C2 == ExponentMask) { |
1913 | unsigned Mask = C1.isZero() |
1914 | ? FPClassTest::fcZero | FPClassTest::fcSubnormal |
1915 | : FPClassTest::fcNan | FPClassTest::fcInf; |
1916 | if (isICMP_NE) |
1917 | Mask = ~Mask & fcAllFlags; |
1918 | return replaceInstUsesWith(I&: Cmp, V: Builder.createIsFPClass(FPNum: V, Test: Mask)); |
1919 | } |
1920 | } |
1921 | } |
1922 | |
1923 | return nullptr; |
1924 | } |
1925 | |
1926 | /// Fold icmp (and X, Y), C. |
1927 | Instruction *InstCombinerImpl::foldICmpAndConstant(ICmpInst &Cmp, |
1928 | BinaryOperator *And, |
1929 | const APInt &C) { |
1930 | if (Instruction *I = foldICmpAndConstConst(Cmp, And, C1: C)) |
1931 | return I; |
1932 | |
1933 | const ICmpInst::Predicate Pred = Cmp.getPredicate(); |
1934 | bool TrueIfNeg; |
1935 | if (isSignBitCheck(Pred, RHS: C, TrueIfSigned&: TrueIfNeg)) { |
1936 | // ((X - 1) & ~X) < 0 --> X == 0 |
1937 | // ((X - 1) & ~X) >= 0 --> X != 0 |
1938 | Value *X; |
1939 | if (match(V: And->getOperand(i_nocapture: 0), P: m_Add(L: m_Value(V&: X), R: m_AllOnes())) && |
1940 | match(V: And->getOperand(i_nocapture: 1), P: m_Not(V: m_Specific(V: X)))) { |
1941 | auto NewPred = TrueIfNeg ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE; |
1942 | return new ICmpInst(NewPred, X, ConstantInt::getNullValue(Ty: X->getType())); |
1943 | } |
1944 | // (X & -X) < 0 --> X == MinSignedC |
1945 | // (X & -X) > -1 --> X != MinSignedC |
1946 | if (match(V: And, P: m_c_And(L: m_Neg(V: m_Value(V&: X)), R: m_Deferred(V: X)))) { |
1947 | Constant *MinSignedC = ConstantInt::get( |
1948 | Ty: X->getType(), |
1949 | V: APInt::getSignedMinValue(numBits: X->getType()->getScalarSizeInBits())); |
1950 | auto NewPred = TrueIfNeg ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE; |
1951 | return new ICmpInst(NewPred, X, MinSignedC); |
1952 | } |
1953 | } |
1954 | |
1955 | // TODO: These all require that Y is constant too, so refactor with the above. |
1956 | |
1957 | // Try to optimize things like "A[i] & 42 == 0" to index computations. |
1958 | Value *X = And->getOperand(i_nocapture: 0); |
1959 | Value *Y = And->getOperand(i_nocapture: 1); |
1960 | if (auto *C2 = dyn_cast<ConstantInt>(Val: Y)) |
1961 | if (auto *LI = dyn_cast<LoadInst>(Val: X)) |
1962 | if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: LI->getOperand(i_nocapture: 0))) |
1963 | if (auto *GV = dyn_cast<GlobalVariable>(Val: GEP->getOperand(i_nocapture: 0))) |
1964 | if (Instruction *Res = |
1965 | foldCmpLoadFromIndexedGlobal(LI, GEP, GV, ICI&: Cmp, AndCst: C2)) |
1966 | return Res; |
1967 | |
1968 | if (!Cmp.isEquality()) |
1969 | return nullptr; |
1970 | |
1971 | // X & -C == -C -> X > u ~C |
1972 | // X & -C != -C -> X <= u ~C |
1973 | // iff C is a power of 2 |
1974 | if (Cmp.getOperand(i_nocapture: 1) == Y && C.isNegatedPowerOf2()) { |
1975 | auto NewPred = |
1976 | Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT : CmpInst::ICMP_ULE; |
1977 | return new ICmpInst(NewPred, X, SubOne(C: cast<Constant>(Val: Cmp.getOperand(i_nocapture: 1)))); |
1978 | } |
1979 | |
1980 | // ((zext i1 X) & Y) == 0 --> !((trunc Y) & X) |
1981 | // ((zext i1 X) & Y) != 0 --> ((trunc Y) & X) |
1982 | // ((zext i1 X) & Y) == 1 --> ((trunc Y) & X) |
1983 | // ((zext i1 X) & Y) != 1 --> !((trunc Y) & X) |
1984 | if (match(V: And, P: m_OneUse(SubPattern: m_c_And(L: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X))), R: m_Value(V&: Y)))) && |
1985 | X->getType()->isIntOrIntVectorTy(BitWidth: 1) && (C.isZero() || C.isOne())) { |
1986 | Value *TruncY = Builder.CreateTrunc(V: Y, DestTy: X->getType()); |
1987 | if (C.isZero() ^ (Pred == CmpInst::ICMP_NE)) { |
1988 | Value *And = Builder.CreateAnd(LHS: TruncY, RHS: X); |
1989 | return BinaryOperator::CreateNot(Op: And); |
1990 | } |
1991 | return BinaryOperator::CreateAnd(V1: TruncY, V2: X); |
1992 | } |
1993 | |
1994 | // (icmp eq/ne (and (shl -1, X), Y), 0) |
1995 | // -> (icmp eq/ne (lshr Y, X), 0) |
1996 | // We could technically handle any C == 0 or (C < 0 && isOdd(C)) but it seems |
1997 | // highly unlikely the non-zero case will ever show up in code. |
1998 | if (C.isZero() && |
1999 | match(V: And, P: m_OneUse(SubPattern: m_c_And(L: m_OneUse(SubPattern: m_Shl(L: m_AllOnes(), R: m_Value(V&: X))), |
2000 | R: m_Value(V&: Y))))) { |
2001 | Value *LShr = Builder.CreateLShr(LHS: Y, RHS: X); |
2002 | return new ICmpInst(Pred, LShr, Constant::getNullValue(Ty: LShr->getType())); |
2003 | } |
2004 | |
2005 | // (icmp eq/ne (and (add A, Addend), Msk), C) |
2006 | // -> (icmp eq/ne (and A, Msk), (and (sub C, Addend), Msk)) |
2007 | { |
2008 | Value *A; |
2009 | const APInt *Addend, *Msk; |
2010 | if (match(V: And, P: m_And(L: m_OneUse(SubPattern: m_Add(L: m_Value(V&: A), R: m_APInt(Res&: Addend))), |
2011 | R: m_LowBitMask(V&: Msk))) && |
2012 | C.ule(RHS: *Msk)) { |
2013 | APInt NewComperand = (C - *Addend) & *Msk; |
2014 | Value *MaskA = Builder.CreateAnd(LHS: A, RHS: ConstantInt::get(Ty: A->getType(), V: *Msk)); |
2015 | return new ICmpInst(Pred, MaskA, |
2016 | ConstantInt::get(Ty: MaskA->getType(), V: NewComperand)); |
2017 | } |
2018 | } |
2019 | |
2020 | return nullptr; |
2021 | } |
2022 | |
2023 | /// Fold icmp eq/ne (or (xor/sub (X1, X2), xor/sub (X3, X4))), 0. |
2024 | static Value *foldICmpOrXorSubChain(ICmpInst &Cmp, BinaryOperator *Or, |
2025 | InstCombiner::BuilderTy &Builder) { |
2026 | // Are we using xors or subs to bitwise check for a pair or pairs of |
2027 | // (in)equalities? Convert to a shorter form that has more potential to be |
2028 | // folded even further. |
2029 | // ((X1 ^/- X2) || (X3 ^/- X4)) == 0 --> (X1 == X2) && (X3 == X4) |
2030 | // ((X1 ^/- X2) || (X3 ^/- X4)) != 0 --> (X1 != X2) || (X3 != X4) |
2031 | // ((X1 ^/- X2) || (X3 ^/- X4) || (X5 ^/- X6)) == 0 --> |
2032 | // (X1 == X2) && (X3 == X4) && (X5 == X6) |
2033 | // ((X1 ^/- X2) || (X3 ^/- X4) || (X5 ^/- X6)) != 0 --> |
2034 | // (X1 != X2) || (X3 != X4) || (X5 != X6) |
2035 | SmallVector<std::pair<Value *, Value *>, 2> CmpValues; |
2036 | SmallVector<Value *, 16> WorkList(1, Or); |
2037 | |
2038 | while (!WorkList.empty()) { |
2039 | auto MatchOrOperatorArgument = [&](Value *OrOperatorArgument) { |
2040 | Value *Lhs, *Rhs; |
2041 | |
2042 | if (match(V: OrOperatorArgument, |
2043 | P: m_OneUse(SubPattern: m_Xor(L: m_Value(V&: Lhs), R: m_Value(V&: Rhs))))) { |
2044 | CmpValues.emplace_back(Args&: Lhs, Args&: Rhs); |
2045 | return; |
2046 | } |
2047 | |
2048 | if (match(V: OrOperatorArgument, |
2049 | P: m_OneUse(SubPattern: m_Sub(L: m_Value(V&: Lhs), R: m_Value(V&: Rhs))))) { |
2050 | CmpValues.emplace_back(Args&: Lhs, Args&: Rhs); |
2051 | return; |
2052 | } |
2053 | |
2054 | WorkList.push_back(Elt: OrOperatorArgument); |
2055 | }; |
2056 | |
2057 | Value *CurrentValue = WorkList.pop_back_val(); |
2058 | Value *OrOperatorLhs, *OrOperatorRhs; |
2059 | |
2060 | if (!match(V: CurrentValue, |
2061 | P: m_Or(L: m_Value(V&: OrOperatorLhs), R: m_Value(V&: OrOperatorRhs)))) { |
2062 | return nullptr; |
2063 | } |
2064 | |
2065 | MatchOrOperatorArgument(OrOperatorRhs); |
2066 | MatchOrOperatorArgument(OrOperatorLhs); |
2067 | } |
2068 | |
2069 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
2070 | auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; |
2071 | Value *LhsCmp = Builder.CreateICmp(P: Pred, LHS: CmpValues.rbegin()->first, |
2072 | RHS: CmpValues.rbegin()->second); |
2073 | |
2074 | for (auto It = CmpValues.rbegin() + 1; It != CmpValues.rend(); ++It) { |
2075 | Value *RhsCmp = Builder.CreateICmp(P: Pred, LHS: It->first, RHS: It->second); |
2076 | LhsCmp = Builder.CreateBinOp(Opc: BOpc, LHS: LhsCmp, RHS: RhsCmp); |
2077 | } |
2078 | |
2079 | return LhsCmp; |
2080 | } |
2081 | |
2082 | /// Fold icmp (or X, Y), C. |
2083 | Instruction *InstCombinerImpl::foldICmpOrConstant(ICmpInst &Cmp, |
2084 | BinaryOperator *Or, |
2085 | const APInt &C) { |
2086 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
2087 | if (C.isOne()) { |
2088 | // icmp slt signum(V) 1 --> icmp slt V, 1 |
2089 | Value *V = nullptr; |
2090 | if (Pred == ICmpInst::ICMP_SLT && match(V: Or, P: m_Signum(V: m_Value(V)))) |
2091 | return new ICmpInst(ICmpInst::ICMP_SLT, V, |
2092 | ConstantInt::get(Ty: V->getType(), V: 1)); |
2093 | } |
2094 | |
2095 | Value *OrOp0 = Or->getOperand(i_nocapture: 0), *OrOp1 = Or->getOperand(i_nocapture: 1); |
2096 | |
2097 | // (icmp eq/ne (or disjoint x, C0), C1) |
2098 | // -> (icmp eq/ne x, C0^C1) |
2099 | if (Cmp.isEquality() && match(V: OrOp1, P: m_ImmConstant()) && |
2100 | cast<PossiblyDisjointInst>(Val: Or)->isDisjoint()) { |
2101 | Value *NewC = |
2102 | Builder.CreateXor(LHS: OrOp1, RHS: ConstantInt::get(Ty: OrOp1->getType(), V: C)); |
2103 | return new ICmpInst(Pred, OrOp0, NewC); |
2104 | } |
2105 | |
2106 | const APInt *MaskC; |
2107 | if (match(V: OrOp1, P: m_APInt(Res&: MaskC)) && Cmp.isEquality()) { |
2108 | if (*MaskC == C && (C + 1).isPowerOf2()) { |
2109 | // X | C == C --> X <=u C |
2110 | // X | C != C --> X >u C |
2111 | // iff C+1 is a power of 2 (C is a bitmask of the low bits) |
2112 | Pred = (Pred == CmpInst::ICMP_EQ) ? CmpInst::ICMP_ULE : CmpInst::ICMP_UGT; |
2113 | return new ICmpInst(Pred, OrOp0, OrOp1); |
2114 | } |
2115 | |
2116 | // More general: canonicalize 'equality with set bits mask' to |
2117 | // 'equality with clear bits mask'. |
2118 | // (X | MaskC) == C --> (X & ~MaskC) == C ^ MaskC |
2119 | // (X | MaskC) != C --> (X & ~MaskC) != C ^ MaskC |
2120 | if (Or->hasOneUse()) { |
2121 | Value *And = Builder.CreateAnd(LHS: OrOp0, RHS: ~(*MaskC)); |
2122 | Constant *NewC = ConstantInt::get(Ty: Or->getType(), V: C ^ (*MaskC)); |
2123 | return new ICmpInst(Pred, And, NewC); |
2124 | } |
2125 | } |
2126 | |
2127 | // (X | (X-1)) s< 0 --> X s< 1 |
2128 | // (X | (X-1)) s> -1 --> X s> 0 |
2129 | Value *X; |
2130 | bool TrueIfSigned; |
2131 | if (isSignBitCheck(Pred, RHS: C, TrueIfSigned) && |
2132 | match(V: Or, P: m_c_Or(L: m_Add(L: m_Value(V&: X), R: m_AllOnes()), R: m_Deferred(V: X)))) { |
2133 | auto NewPred = TrueIfSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGT; |
2134 | Constant *NewC = ConstantInt::get(Ty: X->getType(), V: TrueIfSigned ? 1 : 0); |
2135 | return new ICmpInst(NewPred, X, NewC); |
2136 | } |
2137 | |
2138 | const APInt *OrC; |
2139 | // icmp(X | OrC, C) --> icmp(X, 0) |
2140 | if (C.isNonNegative() && match(V: Or, P: m_Or(L: m_Value(V&: X), R: m_APInt(Res&: OrC)))) { |
2141 | switch (Pred) { |
2142 | // X | OrC s< C --> X s< 0 iff OrC s>= C s>= 0 |
2143 | case ICmpInst::ICMP_SLT: |
2144 | // X | OrC s>= C --> X s>= 0 iff OrC s>= C s>= 0 |
2145 | case ICmpInst::ICMP_SGE: |
2146 | if (OrC->sge(RHS: C)) |
2147 | return new ICmpInst(Pred, X, ConstantInt::getNullValue(Ty: X->getType())); |
2148 | break; |
2149 | // X | OrC s<= C --> X s< 0 iff OrC s> C s>= 0 |
2150 | case ICmpInst::ICMP_SLE: |
2151 | // X | OrC s> C --> X s>= 0 iff OrC s> C s>= 0 |
2152 | case ICmpInst::ICMP_SGT: |
2153 | if (OrC->sgt(RHS: C)) |
2154 | return new ICmpInst(ICmpInst::getFlippedStrictnessPredicate(pred: Pred), X, |
2155 | ConstantInt::getNullValue(Ty: X->getType())); |
2156 | break; |
2157 | default: |
2158 | break; |
2159 | } |
2160 | } |
2161 | |
2162 | if (!Cmp.isEquality() || !C.isZero() || !Or->hasOneUse()) |
2163 | return nullptr; |
2164 | |
2165 | Value *P, *Q; |
2166 | if (match(V: Or, P: m_Or(L: m_PtrToInt(Op: m_Value(V&: P)), R: m_PtrToInt(Op: m_Value(V&: Q))))) { |
2167 | // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0 |
2168 | // -> and (icmp eq P, null), (icmp eq Q, null). |
2169 | Value *CmpP = |
2170 | Builder.CreateICmp(P: Pred, LHS: P, RHS: ConstantInt::getNullValue(Ty: P->getType())); |
2171 | Value *CmpQ = |
2172 | Builder.CreateICmp(P: Pred, LHS: Q, RHS: ConstantInt::getNullValue(Ty: Q->getType())); |
2173 | auto BOpc = Pred == CmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; |
2174 | return BinaryOperator::Create(Op: BOpc, S1: CmpP, S2: CmpQ); |
2175 | } |
2176 | |
2177 | if (Value *V = foldICmpOrXorSubChain(Cmp, Or, Builder)) |
2178 | return replaceInstUsesWith(I&: Cmp, V); |
2179 | |
2180 | return nullptr; |
2181 | } |
2182 | |
2183 | /// Fold icmp (mul X, Y), C. |
2184 | Instruction *InstCombinerImpl::foldICmpMulConstant(ICmpInst &Cmp, |
2185 | BinaryOperator *Mul, |
2186 | const APInt &C) { |
2187 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
2188 | Type *MulTy = Mul->getType(); |
2189 | Value *X = Mul->getOperand(i_nocapture: 0); |
2190 | |
2191 | // If there's no overflow: |
2192 | // X * X == 0 --> X == 0 |
2193 | // X * X != 0 --> X != 0 |
2194 | if (Cmp.isEquality() && C.isZero() && X == Mul->getOperand(i_nocapture: 1) && |
2195 | (Mul->hasNoUnsignedWrap() || Mul->hasNoSignedWrap())) |
2196 | return new ICmpInst(Pred, X, ConstantInt::getNullValue(Ty: MulTy)); |
2197 | |
2198 | const APInt *MulC; |
2199 | if (!match(V: Mul->getOperand(i_nocapture: 1), P: m_APInt(Res&: MulC))) |
2200 | return nullptr; |
2201 | |
2202 | // If this is a test of the sign bit and the multiply is sign-preserving with |
2203 | // a constant operand, use the multiply LHS operand instead: |
2204 | // (X * +MulC) < 0 --> X < 0 |
2205 | // (X * -MulC) < 0 --> X > 0 |
2206 | if (isSignTest(Pred, C) && Mul->hasNoSignedWrap()) { |
2207 | if (MulC->isNegative()) |
2208 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
2209 | return new ICmpInst(Pred, X, ConstantInt::getNullValue(Ty: MulTy)); |
2210 | } |
2211 | |
2212 | if (MulC->isZero()) |
2213 | return nullptr; |
2214 | |
2215 | // If the multiply does not wrap or the constant is odd, try to divide the |
2216 | // compare constant by the multiplication factor. |
2217 | if (Cmp.isEquality()) { |
2218 | // (mul nsw X, MulC) eq/ne C --> X eq/ne C /s MulC |
2219 | if (Mul->hasNoSignedWrap() && C.srem(RHS: *MulC).isZero()) { |
2220 | Constant *NewC = ConstantInt::get(Ty: MulTy, V: C.sdiv(RHS: *MulC)); |
2221 | return new ICmpInst(Pred, X, NewC); |
2222 | } |
2223 | |
2224 | // C % MulC == 0 is weaker than we could use if MulC is odd because it |
2225 | // correct to transform if MulC * N == C including overflow. I.e with i8 |
2226 | // (icmp eq (mul X, 5), 101) -> (icmp eq X, 225) but since 101 % 5 != 0, we |
2227 | // miss that case. |
2228 | if (C.urem(RHS: *MulC).isZero()) { |
2229 | // (mul nuw X, MulC) eq/ne C --> X eq/ne C /u MulC |
2230 | // (mul X, OddC) eq/ne N * C --> X eq/ne N |
2231 | if ((*MulC & 1).isOne() || Mul->hasNoUnsignedWrap()) { |
2232 | Constant *NewC = ConstantInt::get(Ty: MulTy, V: C.udiv(RHS: *MulC)); |
2233 | return new ICmpInst(Pred, X, NewC); |
2234 | } |
2235 | } |
2236 | } |
2237 | |
2238 | // With a matching no-overflow guarantee, fold the constants: |
2239 | // (X * MulC) < C --> X < (C / MulC) |
2240 | // (X * MulC) > C --> X > (C / MulC) |
2241 | // TODO: Assert that Pred is not equal to SGE, SLE, UGE, ULE? |
2242 | Constant *NewC = nullptr; |
2243 | if (Mul->hasNoSignedWrap() && ICmpInst::isSigned(predicate: Pred)) { |
2244 | // MININT / -1 --> overflow. |
2245 | if (C.isMinSignedValue() && MulC->isAllOnes()) |
2246 | return nullptr; |
2247 | if (MulC->isNegative()) |
2248 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
2249 | |
2250 | if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) { |
2251 | NewC = ConstantInt::get( |
2252 | Ty: MulTy, V: APIntOps::RoundingSDiv(A: C, B: *MulC, RM: APInt::Rounding::UP)); |
2253 | } else { |
2254 | assert((Pred == ICmpInst::ICMP_SLE || Pred == ICmpInst::ICMP_SGT) && |
2255 | "Unexpected predicate" ); |
2256 | NewC = ConstantInt::get( |
2257 | Ty: MulTy, V: APIntOps::RoundingSDiv(A: C, B: *MulC, RM: APInt::Rounding::DOWN)); |
2258 | } |
2259 | } else if (Mul->hasNoUnsignedWrap() && ICmpInst::isUnsigned(predicate: Pred)) { |
2260 | if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE) { |
2261 | NewC = ConstantInt::get( |
2262 | Ty: MulTy, V: APIntOps::RoundingUDiv(A: C, B: *MulC, RM: APInt::Rounding::UP)); |
2263 | } else { |
2264 | assert((Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) && |
2265 | "Unexpected predicate" ); |
2266 | NewC = ConstantInt::get( |
2267 | Ty: MulTy, V: APIntOps::RoundingUDiv(A: C, B: *MulC, RM: APInt::Rounding::DOWN)); |
2268 | } |
2269 | } |
2270 | |
2271 | return NewC ? new ICmpInst(Pred, X, NewC) : nullptr; |
2272 | } |
2273 | |
2274 | /// Fold icmp (shl nuw C2, Y), C. |
2275 | static Instruction *foldICmpShlLHSC(ICmpInst &Cmp, Instruction *Shl, |
2276 | const APInt &C) { |
2277 | Value *Y; |
2278 | const APInt *C2; |
2279 | if (!match(V: Shl, P: m_NUWShl(L: m_APInt(Res&: C2), R: m_Value(V&: Y)))) |
2280 | return nullptr; |
2281 | |
2282 | Type *ShiftType = Shl->getType(); |
2283 | unsigned TypeBits = C.getBitWidth(); |
2284 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
2285 | if (Cmp.isUnsigned()) { |
2286 | if (C2->isZero() || C2->ugt(RHS: C)) |
2287 | return nullptr; |
2288 | APInt Div, Rem; |
2289 | APInt::udivrem(LHS: C, RHS: *C2, Quotient&: Div, Remainder&: Rem); |
2290 | bool CIsPowerOf2 = Rem.isZero() && Div.isPowerOf2(); |
2291 | |
2292 | // (1 << Y) pred C -> Y pred Log2(C) |
2293 | if (!CIsPowerOf2) { |
2294 | // (1 << Y) < 30 -> Y <= 4 |
2295 | // (1 << Y) <= 30 -> Y <= 4 |
2296 | // (1 << Y) >= 30 -> Y > 4 |
2297 | // (1 << Y) > 30 -> Y > 4 |
2298 | if (Pred == ICmpInst::ICMP_ULT) |
2299 | Pred = ICmpInst::ICMP_ULE; |
2300 | else if (Pred == ICmpInst::ICMP_UGE) |
2301 | Pred = ICmpInst::ICMP_UGT; |
2302 | } |
2303 | |
2304 | unsigned CLog2 = Div.logBase2(); |
2305 | return new ICmpInst(Pred, Y, ConstantInt::get(Ty: ShiftType, V: CLog2)); |
2306 | } else if (Cmp.isSigned() && C2->isOne()) { |
2307 | Constant *BitWidthMinusOne = ConstantInt::get(Ty: ShiftType, V: TypeBits - 1); |
2308 | // (1 << Y) > 0 -> Y != 31 |
2309 | // (1 << Y) > C -> Y != 31 if C is negative. |
2310 | if (Pred == ICmpInst::ICMP_SGT && C.sle(RHS: 0)) |
2311 | return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne); |
2312 | |
2313 | // (1 << Y) < 0 -> Y == 31 |
2314 | // (1 << Y) < 1 -> Y == 31 |
2315 | // (1 << Y) < C -> Y == 31 if C is negative and not signed min. |
2316 | // Exclude signed min by subtracting 1 and lower the upper bound to 0. |
2317 | if (Pred == ICmpInst::ICMP_SLT && (C - 1).sle(RHS: 0)) |
2318 | return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne); |
2319 | } |
2320 | |
2321 | return nullptr; |
2322 | } |
2323 | |
2324 | /// Fold icmp (shl X, Y), C. |
2325 | Instruction *InstCombinerImpl::foldICmpShlConstant(ICmpInst &Cmp, |
2326 | BinaryOperator *Shl, |
2327 | const APInt &C) { |
2328 | const APInt *ShiftVal; |
2329 | if (Cmp.isEquality() && match(V: Shl->getOperand(i_nocapture: 0), P: m_APInt(Res&: ShiftVal))) |
2330 | return foldICmpShlConstConst(I&: Cmp, A: Shl->getOperand(i_nocapture: 1), AP1: C, AP2: *ShiftVal); |
2331 | |
2332 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
2333 | // (icmp pred (shl nuw&nsw X, Y), Csle0) |
2334 | // -> (icmp pred X, Csle0) |
2335 | // |
2336 | // The idea is the nuw/nsw essentially freeze the sign bit for the shift op |
2337 | // so X's must be what is used. |
2338 | if (C.sle(RHS: 0) && Shl->hasNoUnsignedWrap() && Shl->hasNoSignedWrap()) |
2339 | return new ICmpInst(Pred, Shl->getOperand(i_nocapture: 0), Cmp.getOperand(i_nocapture: 1)); |
2340 | |
2341 | // (icmp eq/ne (shl nuw|nsw X, Y), 0) |
2342 | // -> (icmp eq/ne X, 0) |
2343 | if (ICmpInst::isEquality(P: Pred) && C.isZero() && |
2344 | (Shl->hasNoUnsignedWrap() || Shl->hasNoSignedWrap())) |
2345 | return new ICmpInst(Pred, Shl->getOperand(i_nocapture: 0), Cmp.getOperand(i_nocapture: 1)); |
2346 | |
2347 | // (icmp slt (shl nsw X, Y), 0/1) |
2348 | // -> (icmp slt X, 0/1) |
2349 | // (icmp sgt (shl nsw X, Y), 0/-1) |
2350 | // -> (icmp sgt X, 0/-1) |
2351 | // |
2352 | // NB: sge/sle with a constant will canonicalize to sgt/slt. |
2353 | if (Shl->hasNoSignedWrap() && |
2354 | (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) |
2355 | if (C.isZero() || (Pred == ICmpInst::ICMP_SGT ? C.isAllOnes() : C.isOne())) |
2356 | return new ICmpInst(Pred, Shl->getOperand(i_nocapture: 0), Cmp.getOperand(i_nocapture: 1)); |
2357 | |
2358 | const APInt *ShiftAmt; |
2359 | if (!match(V: Shl->getOperand(i_nocapture: 1), P: m_APInt(Res&: ShiftAmt))) |
2360 | return foldICmpShlLHSC(Cmp, Shl, C); |
2361 | |
2362 | // Check that the shift amount is in range. If not, don't perform undefined |
2363 | // shifts. When the shift is visited, it will be simplified. |
2364 | unsigned TypeBits = C.getBitWidth(); |
2365 | if (ShiftAmt->uge(RHS: TypeBits)) |
2366 | return nullptr; |
2367 | |
2368 | Value *X = Shl->getOperand(i_nocapture: 0); |
2369 | Type *ShType = Shl->getType(); |
2370 | |
2371 | // NSW guarantees that we are only shifting out sign bits from the high bits, |
2372 | // so we can ASHR the compare constant without needing a mask and eliminate |
2373 | // the shift. |
2374 | if (Shl->hasNoSignedWrap()) { |
2375 | if (Pred == ICmpInst::ICMP_SGT) { |
2376 | // icmp Pred (shl nsw X, ShiftAmt), C --> icmp Pred X, (C >>s ShiftAmt) |
2377 | APInt ShiftedC = C.ashr(ShiftAmt: *ShiftAmt); |
2378 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC)); |
2379 | } |
2380 | if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) && |
2381 | C.ashr(ShiftAmt: *ShiftAmt).shl(ShiftAmt: *ShiftAmt) == C) { |
2382 | APInt ShiftedC = C.ashr(ShiftAmt: *ShiftAmt); |
2383 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC)); |
2384 | } |
2385 | if (Pred == ICmpInst::ICMP_SLT) { |
2386 | // SLE is the same as above, but SLE is canonicalized to SLT, so convert: |
2387 | // (X << S) <=s C is equiv to X <=s (C >> S) for all C |
2388 | // (X << S) <s (C + 1) is equiv to X <s (C >> S) + 1 if C <s SMAX |
2389 | // (X << S) <s C is equiv to X <s ((C - 1) >> S) + 1 if C >s SMIN |
2390 | assert(!C.isMinSignedValue() && "Unexpected icmp slt" ); |
2391 | APInt ShiftedC = (C - 1).ashr(ShiftAmt: *ShiftAmt) + 1; |
2392 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC)); |
2393 | } |
2394 | } |
2395 | |
2396 | // NUW guarantees that we are only shifting out zero bits from the high bits, |
2397 | // so we can LSHR the compare constant without needing a mask and eliminate |
2398 | // the shift. |
2399 | if (Shl->hasNoUnsignedWrap()) { |
2400 | if (Pred == ICmpInst::ICMP_UGT) { |
2401 | // icmp Pred (shl nuw X, ShiftAmt), C --> icmp Pred X, (C >>u ShiftAmt) |
2402 | APInt ShiftedC = C.lshr(ShiftAmt: *ShiftAmt); |
2403 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC)); |
2404 | } |
2405 | if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) && |
2406 | C.lshr(ShiftAmt: *ShiftAmt).shl(ShiftAmt: *ShiftAmt) == C) { |
2407 | APInt ShiftedC = C.lshr(ShiftAmt: *ShiftAmt); |
2408 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC)); |
2409 | } |
2410 | if (Pred == ICmpInst::ICMP_ULT) { |
2411 | // ULE is the same as above, but ULE is canonicalized to ULT, so convert: |
2412 | // (X << S) <=u C is equiv to X <=u (C >> S) for all C |
2413 | // (X << S) <u (C + 1) is equiv to X <u (C >> S) + 1 if C <u ~0u |
2414 | // (X << S) <u C is equiv to X <u ((C - 1) >> S) + 1 if C >u 0 |
2415 | assert(C.ugt(0) && "ult 0 should have been eliminated" ); |
2416 | APInt ShiftedC = (C - 1).lshr(ShiftAmt: *ShiftAmt) + 1; |
2417 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShType, V: ShiftedC)); |
2418 | } |
2419 | } |
2420 | |
2421 | if (Cmp.isEquality() && Shl->hasOneUse()) { |
2422 | // Strength-reduce the shift into an 'and'. |
2423 | Constant *Mask = ConstantInt::get( |
2424 | Ty: ShType, |
2425 | V: APInt::getLowBitsSet(numBits: TypeBits, loBitsSet: TypeBits - ShiftAmt->getZExtValue())); |
2426 | Value *And = Builder.CreateAnd(LHS: X, RHS: Mask, Name: Shl->getName() + ".mask" ); |
2427 | Constant *LShrC = ConstantInt::get(Ty: ShType, V: C.lshr(ShiftAmt: *ShiftAmt)); |
2428 | return new ICmpInst(Pred, And, LShrC); |
2429 | } |
2430 | |
2431 | // Otherwise, if this is a comparison of the sign bit, simplify to and/test. |
2432 | bool TrueIfSigned = false; |
2433 | if (Shl->hasOneUse() && isSignBitCheck(Pred, RHS: C, TrueIfSigned)) { |
2434 | // (X << 31) <s 0 --> (X & 1) != 0 |
2435 | Constant *Mask = ConstantInt::get( |
2436 | Ty: ShType, |
2437 | V: APInt::getOneBitSet(numBits: TypeBits, BitNo: TypeBits - ShiftAmt->getZExtValue() - 1)); |
2438 | Value *And = Builder.CreateAnd(LHS: X, RHS: Mask, Name: Shl->getName() + ".mask" ); |
2439 | return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ, |
2440 | And, Constant::getNullValue(Ty: ShType)); |
2441 | } |
2442 | |
2443 | // Simplify 'shl' inequality test into 'and' equality test. |
2444 | if (Cmp.isUnsigned() && Shl->hasOneUse()) { |
2445 | // (X l<< C2) u<=/u> C1 iff C1+1 is power of two -> X & (~C1 l>> C2) ==/!= 0 |
2446 | if ((C + 1).isPowerOf2() && |
2447 | (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT)) { |
2448 | Value *And = Builder.CreateAnd(LHS: X, RHS: (~C).lshr(shiftAmt: ShiftAmt->getZExtValue())); |
2449 | return new ICmpInst(Pred == ICmpInst::ICMP_ULE ? ICmpInst::ICMP_EQ |
2450 | : ICmpInst::ICMP_NE, |
2451 | And, Constant::getNullValue(Ty: ShType)); |
2452 | } |
2453 | // (X l<< C2) u</u>= C1 iff C1 is power of two -> X & (-C1 l>> C2) ==/!= 0 |
2454 | if (C.isPowerOf2() && |
2455 | (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) { |
2456 | Value *And = |
2457 | Builder.CreateAnd(LHS: X, RHS: (~(C - 1)).lshr(shiftAmt: ShiftAmt->getZExtValue())); |
2458 | return new ICmpInst(Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_EQ |
2459 | : ICmpInst::ICMP_NE, |
2460 | And, Constant::getNullValue(Ty: ShType)); |
2461 | } |
2462 | } |
2463 | |
2464 | // Transform (icmp pred iM (shl iM %v, N), C) |
2465 | // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N)) |
2466 | // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N. |
2467 | // This enables us to get rid of the shift in favor of a trunc that may be |
2468 | // free on the target. It has the additional benefit of comparing to a |
2469 | // smaller constant that may be more target-friendly. |
2470 | unsigned Amt = ShiftAmt->getLimitedValue(Limit: TypeBits - 1); |
2471 | if (Shl->hasOneUse() && Amt != 0 && |
2472 | shouldChangeType(FromBitWidth: ShType->getScalarSizeInBits(), ToBitWidth: TypeBits - Amt)) { |
2473 | ICmpInst::Predicate CmpPred = Pred; |
2474 | APInt RHSC = C; |
2475 | |
2476 | if (RHSC.countr_zero() < Amt && ICmpInst::isStrictPredicate(predicate: CmpPred)) { |
2477 | // Try the flipped strictness predicate. |
2478 | // e.g.: |
2479 | // icmp ult i64 (shl X, 32), 8589934593 -> |
2480 | // icmp ule i64 (shl X, 32), 8589934592 -> |
2481 | // icmp ule i32 (trunc X, i32), 2 -> |
2482 | // icmp ult i32 (trunc X, i32), 3 |
2483 | if (auto FlippedStrictness = getFlippedStrictnessPredicateAndConstant( |
2484 | Pred, C: ConstantInt::get(Context&: ShType->getContext(), V: C))) { |
2485 | CmpPred = FlippedStrictness->first; |
2486 | RHSC = cast<ConstantInt>(Val: FlippedStrictness->second)->getValue(); |
2487 | } |
2488 | } |
2489 | |
2490 | if (RHSC.countr_zero() >= Amt) { |
2491 | Type *TruncTy = ShType->getWithNewBitWidth(NewBitWidth: TypeBits - Amt); |
2492 | Constant *NewC = |
2493 | ConstantInt::get(Ty: TruncTy, V: RHSC.ashr(ShiftAmt: *ShiftAmt).trunc(width: TypeBits - Amt)); |
2494 | return new ICmpInst(CmpPred, |
2495 | Builder.CreateTrunc(V: X, DestTy: TruncTy, Name: "" , /*IsNUW=*/false, |
2496 | IsNSW: Shl->hasNoSignedWrap()), |
2497 | NewC); |
2498 | } |
2499 | } |
2500 | |
2501 | return nullptr; |
2502 | } |
2503 | |
2504 | /// Fold icmp ({al}shr X, Y), C. |
2505 | Instruction *InstCombinerImpl::foldICmpShrConstant(ICmpInst &Cmp, |
2506 | BinaryOperator *Shr, |
2507 | const APInt &C) { |
2508 | // An exact shr only shifts out zero bits, so: |
2509 | // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0 |
2510 | Value *X = Shr->getOperand(i_nocapture: 0); |
2511 | CmpInst::Predicate Pred = Cmp.getPredicate(); |
2512 | if (Cmp.isEquality() && Shr->isExact() && C.isZero()) |
2513 | return new ICmpInst(Pred, X, Cmp.getOperand(i_nocapture: 1)); |
2514 | |
2515 | bool IsAShr = Shr->getOpcode() == Instruction::AShr; |
2516 | const APInt *ShiftValC; |
2517 | if (match(V: X, P: m_APInt(Res&: ShiftValC))) { |
2518 | if (Cmp.isEquality()) |
2519 | return foldICmpShrConstConst(I&: Cmp, A: Shr->getOperand(i_nocapture: 1), AP1: C, AP2: *ShiftValC); |
2520 | |
2521 | // (ShiftValC >> Y) >s -1 --> Y != 0 with ShiftValC < 0 |
2522 | // (ShiftValC >> Y) <s 0 --> Y == 0 with ShiftValC < 0 |
2523 | bool TrueIfSigned; |
2524 | if (!IsAShr && ShiftValC->isNegative() && |
2525 | isSignBitCheck(Pred, RHS: C, TrueIfSigned)) |
2526 | return new ICmpInst(TrueIfSigned ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE, |
2527 | Shr->getOperand(i_nocapture: 1), |
2528 | ConstantInt::getNullValue(Ty: X->getType())); |
2529 | |
2530 | // If the shifted constant is a power-of-2, test the shift amount directly: |
2531 | // (ShiftValC >> Y) >u C --> X <u (LZ(C) - LZ(ShiftValC)) |
2532 | // (ShiftValC >> Y) <u C --> X >=u (LZ(C-1) - LZ(ShiftValC)) |
2533 | if (!IsAShr && ShiftValC->isPowerOf2() && |
2534 | (Pred == CmpInst::ICMP_UGT || Pred == CmpInst::ICMP_ULT)) { |
2535 | bool IsUGT = Pred == CmpInst::ICMP_UGT; |
2536 | assert(ShiftValC->uge(C) && "Expected simplify of compare" ); |
2537 | assert((IsUGT || !C.isZero()) && "Expected X u< 0 to simplify" ); |
2538 | |
2539 | unsigned CmpLZ = IsUGT ? C.countl_zero() : (C - 1).countl_zero(); |
2540 | unsigned ShiftLZ = ShiftValC->countl_zero(); |
2541 | Constant *NewC = ConstantInt::get(Ty: Shr->getType(), V: CmpLZ - ShiftLZ); |
2542 | auto NewPred = IsUGT ? CmpInst::ICMP_ULT : CmpInst::ICMP_UGE; |
2543 | return new ICmpInst(NewPred, Shr->getOperand(i_nocapture: 1), NewC); |
2544 | } |
2545 | } |
2546 | |
2547 | const APInt *ShiftAmtC; |
2548 | if (!match(V: Shr->getOperand(i_nocapture: 1), P: m_APInt(Res&: ShiftAmtC))) |
2549 | return nullptr; |
2550 | |
2551 | // Check that the shift amount is in range. If not, don't perform undefined |
2552 | // shifts. When the shift is visited it will be simplified. |
2553 | unsigned TypeBits = C.getBitWidth(); |
2554 | unsigned ShAmtVal = ShiftAmtC->getLimitedValue(Limit: TypeBits); |
2555 | if (ShAmtVal >= TypeBits || ShAmtVal == 0) |
2556 | return nullptr; |
2557 | |
2558 | bool IsExact = Shr->isExact(); |
2559 | Type *ShrTy = Shr->getType(); |
2560 | // TODO: If we could guarantee that InstSimplify would handle all of the |
2561 | // constant-value-based preconditions in the folds below, then we could assert |
2562 | // those conditions rather than checking them. This is difficult because of |
2563 | // undef/poison (PR34838). |
2564 | if (IsAShr && Shr->hasOneUse()) { |
2565 | if (IsExact && (Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_ULT) && |
2566 | (C - 1).isPowerOf2() && C.countLeadingZeros() > ShAmtVal) { |
2567 | // When C - 1 is a power of two and the transform can be legally |
2568 | // performed, prefer this form so the produced constant is close to a |
2569 | // power of two. |
2570 | // icmp slt/ult (ashr exact X, ShAmtC), C |
2571 | // --> icmp slt/ult X, (C - 1) << ShAmtC) + 1 |
2572 | APInt ShiftedC = (C - 1).shl(shiftAmt: ShAmtVal) + 1; |
2573 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC)); |
2574 | } |
2575 | if (IsExact || Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_ULT) { |
2576 | // When ShAmtC can be shifted losslessly: |
2577 | // icmp PRED (ashr exact X, ShAmtC), C --> icmp PRED X, (C << ShAmtC) |
2578 | // icmp slt/ult (ashr X, ShAmtC), C --> icmp slt/ult X, (C << ShAmtC) |
2579 | APInt ShiftedC = C.shl(shiftAmt: ShAmtVal); |
2580 | if (ShiftedC.ashr(ShiftAmt: ShAmtVal) == C) |
2581 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC)); |
2582 | } |
2583 | if (Pred == CmpInst::ICMP_SGT) { |
2584 | // icmp sgt (ashr X, ShAmtC), C --> icmp sgt X, ((C + 1) << ShAmtC) - 1 |
2585 | APInt ShiftedC = (C + 1).shl(shiftAmt: ShAmtVal) - 1; |
2586 | if (!C.isMaxSignedValue() && !(C + 1).shl(shiftAmt: ShAmtVal).isMinSignedValue() && |
2587 | (ShiftedC + 1).ashr(ShiftAmt: ShAmtVal) == (C + 1)) |
2588 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC)); |
2589 | } |
2590 | if (Pred == CmpInst::ICMP_UGT) { |
2591 | // icmp ugt (ashr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1 |
2592 | // 'C + 1 << ShAmtC' can overflow as a signed number, so the 2nd |
2593 | // clause accounts for that pattern. |
2594 | APInt ShiftedC = (C + 1).shl(shiftAmt: ShAmtVal) - 1; |
2595 | if ((ShiftedC + 1).ashr(ShiftAmt: ShAmtVal) == (C + 1) || |
2596 | (C + 1).shl(shiftAmt: ShAmtVal).isMinSignedValue()) |
2597 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC)); |
2598 | } |
2599 | |
2600 | // If the compare constant has significant bits above the lowest sign-bit, |
2601 | // then convert an unsigned cmp to a test of the sign-bit: |
2602 | // (ashr X, ShiftC) u> C --> X s< 0 |
2603 | // (ashr X, ShiftC) u< C --> X s> -1 |
2604 | if (C.getBitWidth() > 2 && C.getNumSignBits() <= ShAmtVal) { |
2605 | if (Pred == CmpInst::ICMP_UGT) { |
2606 | return new ICmpInst(CmpInst::ICMP_SLT, X, |
2607 | ConstantInt::getNullValue(Ty: ShrTy)); |
2608 | } |
2609 | if (Pred == CmpInst::ICMP_ULT) { |
2610 | return new ICmpInst(CmpInst::ICMP_SGT, X, |
2611 | ConstantInt::getAllOnesValue(Ty: ShrTy)); |
2612 | } |
2613 | } |
2614 | } else if (!IsAShr) { |
2615 | if (Pred == CmpInst::ICMP_ULT || (Pred == CmpInst::ICMP_UGT && IsExact)) { |
2616 | // icmp ult (lshr X, ShAmtC), C --> icmp ult X, (C << ShAmtC) |
2617 | // icmp ugt (lshr exact X, ShAmtC), C --> icmp ugt X, (C << ShAmtC) |
2618 | APInt ShiftedC = C.shl(shiftAmt: ShAmtVal); |
2619 | if (ShiftedC.lshr(shiftAmt: ShAmtVal) == C) |
2620 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC)); |
2621 | } |
2622 | if (Pred == CmpInst::ICMP_UGT) { |
2623 | // icmp ugt (lshr X, ShAmtC), C --> icmp ugt X, ((C + 1) << ShAmtC) - 1 |
2624 | APInt ShiftedC = (C + 1).shl(shiftAmt: ShAmtVal) - 1; |
2625 | if ((ShiftedC + 1).lshr(shiftAmt: ShAmtVal) == (C + 1)) |
2626 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: ShiftedC)); |
2627 | } |
2628 | } |
2629 | |
2630 | if (!Cmp.isEquality()) |
2631 | return nullptr; |
2632 | |
2633 | // Handle equality comparisons of shift-by-constant. |
2634 | |
2635 | // If the comparison constant changes with the shift, the comparison cannot |
2636 | // succeed (bits of the comparison constant cannot match the shifted value). |
2637 | // This should be known by InstSimplify and already be folded to true/false. |
2638 | assert(((IsAShr && C.shl(ShAmtVal).ashr(ShAmtVal) == C) || |
2639 | (!IsAShr && C.shl(ShAmtVal).lshr(ShAmtVal) == C)) && |
2640 | "Expected icmp+shr simplify did not occur." ); |
2641 | |
2642 | // If the bits shifted out are known zero, compare the unshifted value: |
2643 | // (X & 4) >> 1 == 2 --> (X & 4) == 4. |
2644 | if (Shr->isExact()) |
2645 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: ShrTy, V: C << ShAmtVal)); |
2646 | |
2647 | if (C.isZero()) { |
2648 | // == 0 is u< 1. |
2649 | if (Pred == CmpInst::ICMP_EQ) |
2650 | return new ICmpInst(CmpInst::ICMP_ULT, X, |
2651 | ConstantInt::get(Ty: ShrTy, V: (C + 1).shl(shiftAmt: ShAmtVal))); |
2652 | else |
2653 | return new ICmpInst(CmpInst::ICMP_UGT, X, |
2654 | ConstantInt::get(Ty: ShrTy, V: (C + 1).shl(shiftAmt: ShAmtVal) - 1)); |
2655 | } |
2656 | |
2657 | if (Shr->hasOneUse()) { |
2658 | // Canonicalize the shift into an 'and': |
2659 | // icmp eq/ne (shr X, ShAmt), C --> icmp eq/ne (and X, HiMask), (C << ShAmt) |
2660 | APInt Val(APInt::getHighBitsSet(numBits: TypeBits, hiBitsSet: TypeBits - ShAmtVal)); |
2661 | Constant *Mask = ConstantInt::get(Ty: ShrTy, V: Val); |
2662 | Value *And = Builder.CreateAnd(LHS: X, RHS: Mask, Name: Shr->getName() + ".mask" ); |
2663 | return new ICmpInst(Pred, And, ConstantInt::get(Ty: ShrTy, V: C << ShAmtVal)); |
2664 | } |
2665 | |
2666 | return nullptr; |
2667 | } |
2668 | |
2669 | Instruction *InstCombinerImpl::foldICmpSRemConstant(ICmpInst &Cmp, |
2670 | BinaryOperator *SRem, |
2671 | const APInt &C) { |
2672 | const ICmpInst::Predicate Pred = Cmp.getPredicate(); |
2673 | if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT) { |
2674 | // Canonicalize unsigned predicates to signed: |
2675 | // (X s% DivisorC) u> C -> (X s% DivisorC) s< 0 |
2676 | // iff (C s< 0 ? ~C : C) u>= abs(DivisorC)-1 |
2677 | // (X s% DivisorC) u< C+1 -> (X s% DivisorC) s> -1 |
2678 | // iff (C+1 s< 0 ? ~C : C) u>= abs(DivisorC)-1 |
2679 | |
2680 | const APInt *DivisorC; |
2681 | if (!match(V: SRem->getOperand(i_nocapture: 1), P: m_APInt(Res&: DivisorC))) |
2682 | return nullptr; |
2683 | |
2684 | APInt NormalizedC = C; |
2685 | if (Pred == ICmpInst::ICMP_ULT) { |
2686 | assert(!NormalizedC.isZero() && |
2687 | "ult X, 0 should have been simplified already." ); |
2688 | --NormalizedC; |
2689 | } |
2690 | if (C.isNegative()) |
2691 | NormalizedC.flipAllBits(); |
2692 | assert(!DivisorC->isZero() && |
2693 | "srem X, 0 should have been simplified already." ); |
2694 | if (!NormalizedC.uge(RHS: DivisorC->abs() - 1)) |
2695 | return nullptr; |
2696 | |
2697 | Type *Ty = SRem->getType(); |
2698 | if (Pred == ICmpInst::ICMP_UGT) |
2699 | return new ICmpInst(ICmpInst::ICMP_SLT, SRem, |
2700 | ConstantInt::getNullValue(Ty)); |
2701 | return new ICmpInst(ICmpInst::ICMP_SGT, SRem, |
2702 | ConstantInt::getAllOnesValue(Ty)); |
2703 | } |
2704 | // Match an 'is positive' or 'is negative' comparison of remainder by a |
2705 | // constant power-of-2 value: |
2706 | // (X % pow2C) sgt/slt 0 |
2707 | if (Pred != ICmpInst::ICMP_SGT && Pred != ICmpInst::ICMP_SLT && |
2708 | Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE) |
2709 | return nullptr; |
2710 | |
2711 | // TODO: The one-use check is standard because we do not typically want to |
2712 | // create longer instruction sequences, but this might be a special-case |
2713 | // because srem is not good for analysis or codegen. |
2714 | if (!SRem->hasOneUse()) |
2715 | return nullptr; |
2716 | |
2717 | const APInt *DivisorC; |
2718 | if (!match(V: SRem->getOperand(i_nocapture: 1), P: m_Power2(V&: DivisorC))) |
2719 | return nullptr; |
2720 | |
2721 | // For cmp_sgt/cmp_slt only zero valued C is handled. |
2722 | // For cmp_eq/cmp_ne only positive valued C is handled. |
2723 | if (((Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT) && |
2724 | !C.isZero()) || |
2725 | ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) && |
2726 | !C.isStrictlyPositive())) |
2727 | return nullptr; |
2728 | |
2729 | // Mask off the sign bit and the modulo bits (low-bits). |
2730 | Type *Ty = SRem->getType(); |
2731 | APInt SignMask = APInt::getSignMask(BitWidth: Ty->getScalarSizeInBits()); |
2732 | Constant *MaskC = ConstantInt::get(Ty, V: SignMask | (*DivisorC - 1)); |
2733 | Value *And = Builder.CreateAnd(LHS: SRem->getOperand(i_nocapture: 0), RHS: MaskC); |
2734 | |
2735 | if (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) |
2736 | return new ICmpInst(Pred, And, ConstantInt::get(Ty, V: C)); |
2737 | |
2738 | // For 'is positive?' check that the sign-bit is clear and at least 1 masked |
2739 | // bit is set. Example: |
2740 | // (i8 X % 32) s> 0 --> (X & 159) s> 0 |
2741 | if (Pred == ICmpInst::ICMP_SGT) |
2742 | return new ICmpInst(ICmpInst::ICMP_SGT, And, ConstantInt::getNullValue(Ty)); |
2743 | |
2744 | // For 'is negative?' check that the sign-bit is set and at least 1 masked |
2745 | // bit is set. Example: |
2746 | // (i16 X % 4) s< 0 --> (X & 32771) u> 32768 |
2747 | return new ICmpInst(ICmpInst::ICMP_UGT, And, ConstantInt::get(Ty, V: SignMask)); |
2748 | } |
2749 | |
2750 | /// Fold icmp (udiv X, Y), C. |
2751 | Instruction *InstCombinerImpl::foldICmpUDivConstant(ICmpInst &Cmp, |
2752 | BinaryOperator *UDiv, |
2753 | const APInt &C) { |
2754 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
2755 | Value *X = UDiv->getOperand(i_nocapture: 0); |
2756 | Value *Y = UDiv->getOperand(i_nocapture: 1); |
2757 | Type *Ty = UDiv->getType(); |
2758 | |
2759 | const APInt *C2; |
2760 | if (!match(V: X, P: m_APInt(Res&: C2))) |
2761 | return nullptr; |
2762 | |
2763 | assert(*C2 != 0 && "udiv 0, X should have been simplified already." ); |
2764 | |
2765 | // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1)) |
2766 | if (Pred == ICmpInst::ICMP_UGT) { |
2767 | assert(!C.isMaxValue() && |
2768 | "icmp ugt X, UINT_MAX should have been simplified already." ); |
2769 | return new ICmpInst(ICmpInst::ICMP_ULE, Y, |
2770 | ConstantInt::get(Ty, V: C2->udiv(RHS: C + 1))); |
2771 | } |
2772 | |
2773 | // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C) |
2774 | if (Pred == ICmpInst::ICMP_ULT) { |
2775 | assert(C != 0 && "icmp ult X, 0 should have been simplified already." ); |
2776 | return new ICmpInst(ICmpInst::ICMP_UGT, Y, |
2777 | ConstantInt::get(Ty, V: C2->udiv(RHS: C))); |
2778 | } |
2779 | |
2780 | return nullptr; |
2781 | } |
2782 | |
2783 | /// Fold icmp ({su}div X, Y), C. |
2784 | Instruction *InstCombinerImpl::foldICmpDivConstant(ICmpInst &Cmp, |
2785 | BinaryOperator *Div, |
2786 | const APInt &C) { |
2787 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
2788 | Value *X = Div->getOperand(i_nocapture: 0); |
2789 | Value *Y = Div->getOperand(i_nocapture: 1); |
2790 | Type *Ty = Div->getType(); |
2791 | bool DivIsSigned = Div->getOpcode() == Instruction::SDiv; |
2792 | |
2793 | // If unsigned division and the compare constant is bigger than |
2794 | // UMAX/2 (negative), there's only one pair of values that satisfies an |
2795 | // equality check, so eliminate the division: |
2796 | // (X u/ Y) == C --> (X == C) && (Y == 1) |
2797 | // (X u/ Y) != C --> (X != C) || (Y != 1) |
2798 | // Similarly, if signed division and the compare constant is exactly SMIN: |
2799 | // (X s/ Y) == SMIN --> (X == SMIN) && (Y == 1) |
2800 | // (X s/ Y) != SMIN --> (X != SMIN) || (Y != 1) |
2801 | if (Cmp.isEquality() && Div->hasOneUse() && C.isSignBitSet() && |
2802 | (!DivIsSigned || C.isMinSignedValue())) { |
2803 | Value *XBig = Builder.CreateICmp(P: Pred, LHS: X, RHS: ConstantInt::get(Ty, V: C)); |
2804 | Value *YOne = Builder.CreateICmp(P: Pred, LHS: Y, RHS: ConstantInt::get(Ty, V: 1)); |
2805 | auto Logic = Pred == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or; |
2806 | return BinaryOperator::Create(Op: Logic, S1: XBig, S2: YOne); |
2807 | } |
2808 | |
2809 | // Fold: icmp pred ([us]div X, C2), C -> range test |
2810 | // Fold this div into the comparison, producing a range check. |
2811 | // Determine, based on the divide type, what the range is being |
2812 | // checked. If there is an overflow on the low or high side, remember |
2813 | // it, otherwise compute the range [low, hi) bounding the new value. |
2814 | // See: InsertRangeTest above for the kinds of replacements possible. |
2815 | const APInt *C2; |
2816 | if (!match(V: Y, P: m_APInt(Res&: C2))) |
2817 | return nullptr; |
2818 | |
2819 | // FIXME: If the operand types don't match the type of the divide |
2820 | // then don't attempt this transform. The code below doesn't have the |
2821 | // logic to deal with a signed divide and an unsigned compare (and |
2822 | // vice versa). This is because (x /s C2) <s C produces different |
2823 | // results than (x /s C2) <u C or (x /u C2) <s C or even |
2824 | // (x /u C2) <u C. Simply casting the operands and result won't |
2825 | // work. :( The if statement below tests that condition and bails |
2826 | // if it finds it. |
2827 | if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned()) |
2828 | return nullptr; |
2829 | |
2830 | // The ProdOV computation fails on divide by 0 and divide by -1. Cases with |
2831 | // INT_MIN will also fail if the divisor is 1. Although folds of all these |
2832 | // division-by-constant cases should be present, we can not assert that they |
2833 | // have happened before we reach this icmp instruction. |
2834 | if (C2->isZero() || C2->isOne() || (DivIsSigned && C2->isAllOnes())) |
2835 | return nullptr; |
2836 | |
2837 | // Compute Prod = C * C2. We are essentially solving an equation of |
2838 | // form X / C2 = C. We solve for X by multiplying C2 and C. |
2839 | // By solving for X, we can turn this into a range check instead of computing |
2840 | // a divide. |
2841 | APInt Prod = C * *C2; |
2842 | |
2843 | // Determine if the product overflows by seeing if the product is not equal to |
2844 | // the divide. Make sure we do the same kind of divide as in the LHS |
2845 | // instruction that we're folding. |
2846 | bool ProdOV = (DivIsSigned ? Prod.sdiv(RHS: *C2) : Prod.udiv(RHS: *C2)) != C; |
2847 | |
2848 | // If the division is known to be exact, then there is no remainder from the |
2849 | // divide, so the covered range size is unit, otherwise it is the divisor. |
2850 | APInt RangeSize = Div->isExact() ? APInt(C2->getBitWidth(), 1) : *C2; |
2851 | |
2852 | // Figure out the interval that is being checked. For example, a comparison |
2853 | // like "X /u 5 == 0" is really checking that X is in the interval [0, 5). |
2854 | // Compute this interval based on the constants involved and the signedness of |
2855 | // the compare/divide. This computes a half-open interval, keeping track of |
2856 | // whether either value in the interval overflows. After analysis each |
2857 | // overflow variable is set to 0 if it's corresponding bound variable is valid |
2858 | // -1 if overflowed off the bottom end, or +1 if overflowed off the top end. |
2859 | int LoOverflow = 0, HiOverflow = 0; |
2860 | APInt LoBound, HiBound; |
2861 | |
2862 | if (!DivIsSigned) { // udiv |
2863 | // e.g. X/5 op 3 --> [15, 20) |
2864 | LoBound = Prod; |
2865 | HiOverflow = LoOverflow = ProdOV; |
2866 | if (!HiOverflow) { |
2867 | // If this is not an exact divide, then many values in the range collapse |
2868 | // to the same result value. |
2869 | HiOverflow = addWithOverflow(Result&: HiBound, In1: LoBound, In2: RangeSize, IsSigned: false); |
2870 | } |
2871 | } else if (C2->isStrictlyPositive()) { // Divisor is > 0. |
2872 | if (C.isZero()) { // (X / pos) op 0 |
2873 | // Can't overflow. e.g. X/2 op 0 --> [-1, 2) |
2874 | LoBound = -(RangeSize - 1); |
2875 | HiBound = RangeSize; |
2876 | } else if (C.isStrictlyPositive()) { // (X / pos) op pos |
2877 | LoBound = Prod; // e.g. X/5 op 3 --> [15, 20) |
2878 | HiOverflow = LoOverflow = ProdOV; |
2879 | if (!HiOverflow) |
2880 | HiOverflow = addWithOverflow(Result&: HiBound, In1: Prod, In2: RangeSize, IsSigned: true); |
2881 | } else { // (X / pos) op neg |
2882 | // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14) |
2883 | HiBound = Prod + 1; |
2884 | LoOverflow = HiOverflow = ProdOV ? -1 : 0; |
2885 | if (!LoOverflow) { |
2886 | APInt DivNeg = -RangeSize; |
2887 | LoOverflow = addWithOverflow(Result&: LoBound, In1: HiBound, In2: DivNeg, IsSigned: true) ? -1 : 0; |
2888 | } |
2889 | } |
2890 | } else if (C2->isNegative()) { // Divisor is < 0. |
2891 | if (Div->isExact()) |
2892 | RangeSize.negate(); |
2893 | if (C.isZero()) { // (X / neg) op 0 |
2894 | // e.g. X/-5 op 0 --> [-4, 5) |
2895 | LoBound = RangeSize + 1; |
2896 | HiBound = -RangeSize; |
2897 | if (HiBound == *C2) { // -INTMIN = INTMIN |
2898 | HiOverflow = 1; // [INTMIN+1, overflow) |
2899 | HiBound = APInt(); // e.g. X/INTMIN = 0 --> X > INTMIN |
2900 | } |
2901 | } else if (C.isStrictlyPositive()) { // (X / neg) op pos |
2902 | // e.g. X/-5 op 3 --> [-19, -14) |
2903 | HiBound = Prod + 1; |
2904 | HiOverflow = LoOverflow = ProdOV ? -1 : 0; |
2905 | if (!LoOverflow) |
2906 | LoOverflow = |
2907 | addWithOverflow(Result&: LoBound, In1: HiBound, In2: RangeSize, IsSigned: true) ? -1 : 0; |
2908 | } else { // (X / neg) op neg |
2909 | LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20) |
2910 | LoOverflow = HiOverflow = ProdOV; |
2911 | if (!HiOverflow) |
2912 | HiOverflow = subWithOverflow(Result&: HiBound, In1: Prod, In2: RangeSize, IsSigned: true); |
2913 | } |
2914 | |
2915 | // Dividing by a negative swaps the condition. LT <-> GT |
2916 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
2917 | } |
2918 | |
2919 | switch (Pred) { |
2920 | default: |
2921 | llvm_unreachable("Unhandled icmp predicate!" ); |
2922 | case ICmpInst::ICMP_EQ: |
2923 | if (LoOverflow && HiOverflow) |
2924 | return replaceInstUsesWith(I&: Cmp, V: Builder.getFalse()); |
2925 | if (HiOverflow) |
2926 | return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE, |
2927 | X, ConstantInt::get(Ty, V: LoBound)); |
2928 | if (LoOverflow) |
2929 | return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT, |
2930 | X, ConstantInt::get(Ty, V: HiBound)); |
2931 | return replaceInstUsesWith( |
2932 | I&: Cmp, V: insertRangeTest(V: X, Lo: LoBound, Hi: HiBound, isSigned: DivIsSigned, Inside: true)); |
2933 | case ICmpInst::ICMP_NE: |
2934 | if (LoOverflow && HiOverflow) |
2935 | return replaceInstUsesWith(I&: Cmp, V: Builder.getTrue()); |
2936 | if (HiOverflow) |
2937 | return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT, |
2938 | X, ConstantInt::get(Ty, V: LoBound)); |
2939 | if (LoOverflow) |
2940 | return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE, |
2941 | X, ConstantInt::get(Ty, V: HiBound)); |
2942 | return replaceInstUsesWith( |
2943 | I&: Cmp, V: insertRangeTest(V: X, Lo: LoBound, Hi: HiBound, isSigned: DivIsSigned, Inside: false)); |
2944 | case ICmpInst::ICMP_ULT: |
2945 | case ICmpInst::ICMP_SLT: |
2946 | if (LoOverflow == +1) // Low bound is greater than input range. |
2947 | return replaceInstUsesWith(I&: Cmp, V: Builder.getTrue()); |
2948 | if (LoOverflow == -1) // Low bound is less than input range. |
2949 | return replaceInstUsesWith(I&: Cmp, V: Builder.getFalse()); |
2950 | return new ICmpInst(Pred, X, ConstantInt::get(Ty, V: LoBound)); |
2951 | case ICmpInst::ICMP_UGT: |
2952 | case ICmpInst::ICMP_SGT: |
2953 | if (HiOverflow == +1) // High bound greater than input range. |
2954 | return replaceInstUsesWith(I&: Cmp, V: Builder.getFalse()); |
2955 | if (HiOverflow == -1) // High bound less than input range. |
2956 | return replaceInstUsesWith(I&: Cmp, V: Builder.getTrue()); |
2957 | if (Pred == ICmpInst::ICMP_UGT) |
2958 | return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, V: HiBound)); |
2959 | return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, V: HiBound)); |
2960 | } |
2961 | |
2962 | return nullptr; |
2963 | } |
2964 | |
2965 | /// Fold icmp (sub X, Y), C. |
2966 | Instruction *InstCombinerImpl::foldICmpSubConstant(ICmpInst &Cmp, |
2967 | BinaryOperator *Sub, |
2968 | const APInt &C) { |
2969 | Value *X = Sub->getOperand(i_nocapture: 0), *Y = Sub->getOperand(i_nocapture: 1); |
2970 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
2971 | Type *Ty = Sub->getType(); |
2972 | |
2973 | // (SubC - Y) == C) --> Y == (SubC - C) |
2974 | // (SubC - Y) != C) --> Y != (SubC - C) |
2975 | Constant *SubC; |
2976 | if (Cmp.isEquality() && match(V: X, P: m_ImmConstant(C&: SubC))) { |
2977 | return new ICmpInst(Pred, Y, |
2978 | ConstantExpr::getSub(C1: SubC, C2: ConstantInt::get(Ty, V: C))); |
2979 | } |
2980 | |
2981 | // (icmp P (sub nuw|nsw C2, Y), C) -> (icmp swap(P) Y, C2-C) |
2982 | const APInt *C2; |
2983 | APInt SubResult; |
2984 | ICmpInst::Predicate SwappedPred = Cmp.getSwappedPredicate(); |
2985 | bool HasNSW = Sub->hasNoSignedWrap(); |
2986 | bool HasNUW = Sub->hasNoUnsignedWrap(); |
2987 | if (match(V: X, P: m_APInt(Res&: C2)) && |
2988 | ((Cmp.isUnsigned() && HasNUW) || (Cmp.isSigned() && HasNSW)) && |
2989 | !subWithOverflow(Result&: SubResult, In1: *C2, In2: C, IsSigned: Cmp.isSigned())) |
2990 | return new ICmpInst(SwappedPred, Y, ConstantInt::get(Ty, V: SubResult)); |
2991 | |
2992 | // X - Y == 0 --> X == Y. |
2993 | // X - Y != 0 --> X != Y. |
2994 | // TODO: We allow this with multiple uses as long as the other uses are not |
2995 | // in phis. The phi use check is guarding against a codegen regression |
2996 | // for a loop test. If the backend could undo this (and possibly |
2997 | // subsequent transforms), we would not need this hack. |
2998 | if (Cmp.isEquality() && C.isZero() && |
2999 | none_of(Range: (Sub->users()), P: [](const User *U) { return isa<PHINode>(Val: U); })) |
3000 | return new ICmpInst(Pred, X, Y); |
3001 | |
3002 | // The following transforms are only worth it if the only user of the subtract |
3003 | // is the icmp. |
3004 | // TODO: This is an artificial restriction for all of the transforms below |
3005 | // that only need a single replacement icmp. Can these use the phi test |
3006 | // like the transform above here? |
3007 | if (!Sub->hasOneUse()) |
3008 | return nullptr; |
3009 | |
3010 | if (Sub->hasNoSignedWrap()) { |
3011 | // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y) |
3012 | if (Pred == ICmpInst::ICMP_SGT && C.isAllOnes()) |
3013 | return new ICmpInst(ICmpInst::ICMP_SGE, X, Y); |
3014 | |
3015 | // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y) |
3016 | if (Pred == ICmpInst::ICMP_SGT && C.isZero()) |
3017 | return new ICmpInst(ICmpInst::ICMP_SGT, X, Y); |
3018 | |
3019 | // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y) |
3020 | if (Pred == ICmpInst::ICMP_SLT && C.isZero()) |
3021 | return new ICmpInst(ICmpInst::ICMP_SLT, X, Y); |
3022 | |
3023 | // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y) |
3024 | if (Pred == ICmpInst::ICMP_SLT && C.isOne()) |
3025 | return new ICmpInst(ICmpInst::ICMP_SLE, X, Y); |
3026 | } |
3027 | |
3028 | if (!match(V: X, P: m_APInt(Res&: C2))) |
3029 | return nullptr; |
3030 | |
3031 | // C2 - Y <u C -> (Y | (C - 1)) == C2 |
3032 | // iff (C2 & (C - 1)) == C - 1 and C is a power of 2 |
3033 | if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && |
3034 | (*C2 & (C - 1)) == (C - 1)) |
3035 | return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateOr(LHS: Y, RHS: C - 1), X); |
3036 | |
3037 | // C2 - Y >u C -> (Y | C) != C2 |
3038 | // iff C2 & C == C and C + 1 is a power of 2 |
3039 | if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == C) |
3040 | return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateOr(LHS: Y, RHS: C), X); |
3041 | |
3042 | // We have handled special cases that reduce. |
3043 | // Canonicalize any remaining sub to add as: |
3044 | // (C2 - Y) > C --> (Y + ~C2) < ~C |
3045 | Value *Add = Builder.CreateAdd(LHS: Y, RHS: ConstantInt::get(Ty, V: ~(*C2)), Name: "notsub" , |
3046 | HasNUW, HasNSW); |
3047 | return new ICmpInst(SwappedPred, Add, ConstantInt::get(Ty, V: ~C)); |
3048 | } |
3049 | |
3050 | static Value *createLogicFromTable(const std::bitset<4> &Table, Value *Op0, |
3051 | Value *Op1, IRBuilderBase &Builder, |
3052 | bool HasOneUse) { |
3053 | auto FoldConstant = [&](bool Val) { |
3054 | Constant *Res = Val ? Builder.getTrue() : Builder.getFalse(); |
3055 | if (Op0->getType()->isVectorTy()) |
3056 | Res = ConstantVector::getSplat( |
3057 | EC: cast<VectorType>(Val: Op0->getType())->getElementCount(), Elt: Res); |
3058 | return Res; |
3059 | }; |
3060 | |
3061 | switch (Table.to_ulong()) { |
3062 | case 0: // 0 0 0 0 |
3063 | return FoldConstant(false); |
3064 | case 1: // 0 0 0 1 |
3065 | return HasOneUse ? Builder.CreateNot(V: Builder.CreateOr(LHS: Op0, RHS: Op1)) : nullptr; |
3066 | case 2: // 0 0 1 0 |
3067 | return HasOneUse ? Builder.CreateAnd(LHS: Builder.CreateNot(V: Op0), RHS: Op1) : nullptr; |
3068 | case 3: // 0 0 1 1 |
3069 | return Builder.CreateNot(V: Op0); |
3070 | case 4: // 0 1 0 0 |
3071 | return HasOneUse ? Builder.CreateAnd(LHS: Op0, RHS: Builder.CreateNot(V: Op1)) : nullptr; |
3072 | case 5: // 0 1 0 1 |
3073 | return Builder.CreateNot(V: Op1); |
3074 | case 6: // 0 1 1 0 |
3075 | return Builder.CreateXor(LHS: Op0, RHS: Op1); |
3076 | case 7: // 0 1 1 1 |
3077 | return HasOneUse ? Builder.CreateNot(V: Builder.CreateAnd(LHS: Op0, RHS: Op1)) : nullptr; |
3078 | case 8: // 1 0 0 0 |
3079 | return Builder.CreateAnd(LHS: Op0, RHS: Op1); |
3080 | case 9: // 1 0 0 1 |
3081 | return HasOneUse ? Builder.CreateNot(V: Builder.CreateXor(LHS: Op0, RHS: Op1)) : nullptr; |
3082 | case 10: // 1 0 1 0 |
3083 | return Op1; |
3084 | case 11: // 1 0 1 1 |
3085 | return HasOneUse ? Builder.CreateOr(LHS: Builder.CreateNot(V: Op0), RHS: Op1) : nullptr; |
3086 | case 12: // 1 1 0 0 |
3087 | return Op0; |
3088 | case 13: // 1 1 0 1 |
3089 | return HasOneUse ? Builder.CreateOr(LHS: Op0, RHS: Builder.CreateNot(V: Op1)) : nullptr; |
3090 | case 14: // 1 1 1 0 |
3091 | return Builder.CreateOr(LHS: Op0, RHS: Op1); |
3092 | case 15: // 1 1 1 1 |
3093 | return FoldConstant(true); |
3094 | default: |
3095 | llvm_unreachable("Invalid Operation" ); |
3096 | } |
3097 | return nullptr; |
3098 | } |
3099 | |
3100 | Instruction *InstCombinerImpl::foldICmpBinOpWithConstantViaTruthTable( |
3101 | ICmpInst &Cmp, BinaryOperator *BO, const APInt &C) { |
3102 | Value *A, *B; |
3103 | Constant *C1, *C2, *C3, *C4; |
3104 | if (!(match(V: BO->getOperand(i_nocapture: 0), |
3105 | P: m_Select(C: m_Value(V&: A), L: m_Constant(C&: C1), R: m_Constant(C&: C2)))) || |
3106 | !match(V: BO->getOperand(i_nocapture: 1), |
3107 | P: m_Select(C: m_Value(V&: B), L: m_Constant(C&: C3), R: m_Constant(C&: C4))) || |
3108 | Cmp.getType() != A->getType()) |
3109 | return nullptr; |
3110 | |
3111 | std::bitset<4> Table; |
3112 | auto ComputeTable = [&](bool First, bool Second) -> std::optional<bool> { |
3113 | Constant *L = First ? C1 : C2; |
3114 | Constant *R = Second ? C3 : C4; |
3115 | if (auto *Res = ConstantFoldBinaryOpOperands(Opcode: BO->getOpcode(), LHS: L, RHS: R, DL)) { |
3116 | auto *Val = Res->getType()->isVectorTy() ? Res->getSplatValue() : Res; |
3117 | if (auto *CI = dyn_cast_or_null<ConstantInt>(Val)) |
3118 | return ICmpInst::compare(LHS: CI->getValue(), RHS: C, Pred: Cmp.getPredicate()); |
3119 | } |
3120 | return std::nullopt; |
3121 | }; |
3122 | |
3123 | for (unsigned I = 0; I < 4; ++I) { |
3124 | bool First = (I >> 1) & 1; |
3125 | bool Second = I & 1; |
3126 | if (auto Res = ComputeTable(First, Second)) |
3127 | Table[I] = *Res; |
3128 | else |
3129 | return nullptr; |
3130 | } |
3131 | |
3132 | // Synthesize optimal logic. |
3133 | if (auto *Cond = createLogicFromTable(Table, Op0: A, Op1: B, Builder, HasOneUse: BO->hasOneUse())) |
3134 | return replaceInstUsesWith(I&: Cmp, V: Cond); |
3135 | return nullptr; |
3136 | } |
3137 | |
3138 | /// Fold icmp (add X, Y), C. |
3139 | Instruction *InstCombinerImpl::foldICmpAddConstant(ICmpInst &Cmp, |
3140 | BinaryOperator *Add, |
3141 | const APInt &C) { |
3142 | Value *Y = Add->getOperand(i_nocapture: 1); |
3143 | Value *X = Add->getOperand(i_nocapture: 0); |
3144 | |
3145 | Value *Op0, *Op1; |
3146 | Instruction *Ext0, *Ext1; |
3147 | const CmpInst::Predicate Pred = Cmp.getPredicate(); |
3148 | if (match(V: Add, |
3149 | P: m_Add(L: m_CombineAnd(L: m_Instruction(I&: Ext0), R: m_ZExtOrSExt(Op: m_Value(V&: Op0))), |
3150 | R: m_CombineAnd(L: m_Instruction(I&: Ext1), |
3151 | R: m_ZExtOrSExt(Op: m_Value(V&: Op1))))) && |
3152 | Op0->getType()->isIntOrIntVectorTy(BitWidth: 1) && |
3153 | Op1->getType()->isIntOrIntVectorTy(BitWidth: 1)) { |
3154 | unsigned BW = C.getBitWidth(); |
3155 | std::bitset<4> Table; |
3156 | auto ComputeTable = [&](bool Op0Val, bool Op1Val) { |
3157 | APInt Res(BW, 0); |
3158 | if (Op0Val) |
3159 | Res += APInt(BW, isa<ZExtInst>(Val: Ext0) ? 1 : -1, /*isSigned=*/true); |
3160 | if (Op1Val) |
3161 | Res += APInt(BW, isa<ZExtInst>(Val: Ext1) ? 1 : -1, /*isSigned=*/true); |
3162 | return ICmpInst::compare(LHS: Res, RHS: C, Pred); |
3163 | }; |
3164 | |
3165 | Table[0] = ComputeTable(false, false); |
3166 | Table[1] = ComputeTable(false, true); |
3167 | Table[2] = ComputeTable(true, false); |
3168 | Table[3] = ComputeTable(true, true); |
3169 | if (auto *Cond = |
3170 | createLogicFromTable(Table, Op0, Op1, Builder, HasOneUse: Add->hasOneUse())) |
3171 | return replaceInstUsesWith(I&: Cmp, V: Cond); |
3172 | } |
3173 | const APInt *C2; |
3174 | if (Cmp.isEquality() || !match(V: Y, P: m_APInt(Res&: C2))) |
3175 | return nullptr; |
3176 | |
3177 | // Fold icmp pred (add X, C2), C. |
3178 | Type *Ty = Add->getType(); |
3179 | |
3180 | // If the add does not wrap, we can always adjust the compare by subtracting |
3181 | // the constants. Equality comparisons are handled elsewhere. SGE/SLE/UGE/ULE |
3182 | // are canonicalized to SGT/SLT/UGT/ULT. |
3183 | if ((Add->hasNoSignedWrap() && |
3184 | (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLT)) || |
3185 | (Add->hasNoUnsignedWrap() && |
3186 | (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULT))) { |
3187 | bool Overflow; |
3188 | APInt NewC = |
3189 | Cmp.isSigned() ? C.ssub_ov(RHS: *C2, Overflow) : C.usub_ov(RHS: *C2, Overflow); |
3190 | // If there is overflow, the result must be true or false. |
3191 | // TODO: Can we assert there is no overflow because InstSimplify always |
3192 | // handles those cases? |
3193 | if (!Overflow) |
3194 | // icmp Pred (add nsw X, C2), C --> icmp Pred X, (C - C2) |
3195 | return new ICmpInst(Pred, X, ConstantInt::get(Ty, V: NewC)); |
3196 | } |
3197 | |
3198 | if (ICmpInst::isUnsigned(predicate: Pred) && Add->hasNoSignedWrap() && |
3199 | C.isNonNegative() && (C - *C2).isNonNegative() && |
3200 | computeConstantRange(V: X, /*ForSigned=*/true).add(Other: *C2).isAllNonNegative()) |
3201 | return new ICmpInst(ICmpInst::getSignedPredicate(Pred), X, |
3202 | ConstantInt::get(Ty, V: C - *C2)); |
3203 | |
3204 | auto CR = ConstantRange::makeExactICmpRegion(Pred, Other: C).subtract(CI: *C2); |
3205 | const APInt &Upper = CR.getUpper(); |
3206 | const APInt &Lower = CR.getLower(); |
3207 | if (Cmp.isSigned()) { |
3208 | if (Lower.isSignMask()) |
3209 | return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, V: Upper)); |
3210 | if (Upper.isSignMask()) |
3211 | return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, V: Lower)); |
3212 | } else { |
3213 | if (Lower.isMinValue()) |
3214 | return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, V: Upper)); |
3215 | if (Upper.isMinValue()) |
3216 | return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, V: Lower)); |
3217 | } |
3218 | |
3219 | // This set of folds is intentionally placed after folds that use no-wrapping |
3220 | // flags because those folds are likely better for later analysis/codegen. |
3221 | const APInt SMax = APInt::getSignedMaxValue(numBits: Ty->getScalarSizeInBits()); |
3222 | const APInt SMin = APInt::getSignedMinValue(numBits: Ty->getScalarSizeInBits()); |
3223 | |
3224 | // Fold compare with offset to opposite sign compare if it eliminates offset: |
3225 | // (X + C2) >u C --> X <s -C2 (if C == C2 + SMAX) |
3226 | if (Pred == CmpInst::ICMP_UGT && C == *C2 + SMax) |
3227 | return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, V: -(*C2))); |
3228 | |
3229 | // (X + C2) <u C --> X >s ~C2 (if C == C2 + SMIN) |
3230 | if (Pred == CmpInst::ICMP_ULT && C == *C2 + SMin) |
3231 | return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantInt::get(Ty, V: ~(*C2))); |
3232 | |
3233 | // (X + C2) >s C --> X <u (SMAX - C) (if C == C2 - 1) |
3234 | if (Pred == CmpInst::ICMP_SGT && C == *C2 - 1) |
3235 | return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, V: SMax - C)); |
3236 | |
3237 | // (X + C2) <s C --> X >u (C ^ SMAX) (if C == C2) |
3238 | if (Pred == CmpInst::ICMP_SLT && C == *C2) |
3239 | return new ICmpInst(ICmpInst::ICMP_UGT, X, ConstantInt::get(Ty, V: C ^ SMax)); |
3240 | |
3241 | // (X + -1) <u C --> X <=u C (if X is never null) |
3242 | if (Pred == CmpInst::ICMP_ULT && C2->isAllOnes()) { |
3243 | const SimplifyQuery Q = SQ.getWithInstruction(I: &Cmp); |
3244 | if (llvm::isKnownNonZero(V: X, Q)) |
3245 | return new ICmpInst(ICmpInst::ICMP_ULE, X, ConstantInt::get(Ty, V: C)); |
3246 | } |
3247 | |
3248 | if (!Add->hasOneUse()) |
3249 | return nullptr; |
3250 | |
3251 | // X+C <u C2 -> (X & -C2) == C |
3252 | // iff C & (C2-1) == 0 |
3253 | // C2 is a power of 2 |
3254 | if (Pred == ICmpInst::ICMP_ULT && C.isPowerOf2() && (*C2 & (C - 1)) == 0) |
3255 | return new ICmpInst(ICmpInst::ICMP_EQ, Builder.CreateAnd(LHS: X, RHS: -C), |
3256 | ConstantExpr::getNeg(C: cast<Constant>(Val: Y))); |
3257 | |
3258 | // X+C2 <u C -> (X & C) == 2C |
3259 | // iff C == -(C2) |
3260 | // C2 is a power of 2 |
3261 | if (Pred == ICmpInst::ICMP_ULT && C2->isPowerOf2() && C == -*C2) |
3262 | return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(LHS: X, RHS: C), |
3263 | ConstantInt::get(Ty, V: C * 2)); |
3264 | |
3265 | // X+C >u C2 -> (X & ~C2) != C |
3266 | // iff C & C2 == 0 |
3267 | // C2+1 is a power of 2 |
3268 | if (Pred == ICmpInst::ICMP_UGT && (C + 1).isPowerOf2() && (*C2 & C) == 0) |
3269 | return new ICmpInst(ICmpInst::ICMP_NE, Builder.CreateAnd(LHS: X, RHS: ~C), |
3270 | ConstantExpr::getNeg(C: cast<Constant>(Val: Y))); |
3271 | |
3272 | // The range test idiom can use either ult or ugt. Arbitrarily canonicalize |
3273 | // to the ult form. |
3274 | // X+C2 >u C -> X+(C2-C-1) <u ~C |
3275 | if (Pred == ICmpInst::ICMP_UGT) |
3276 | return new ICmpInst(ICmpInst::ICMP_ULT, |
3277 | Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty, V: *C2 - C - 1)), |
3278 | ConstantInt::get(Ty, V: ~C)); |
3279 | |
3280 | // zext(V) + C2 pred C -> V + C3 pred' C4 |
3281 | Value *V; |
3282 | if (match(V: X, P: m_ZExt(Op: m_Value(V)))) { |
3283 | Type *NewCmpTy = V->getType(); |
3284 | unsigned NewCmpBW = NewCmpTy->getScalarSizeInBits(); |
3285 | if (shouldChangeType(From: Ty, To: NewCmpTy)) { |
3286 | if (CR.getActiveBits() <= NewCmpBW) { |
3287 | ConstantRange SrcCR = CR.truncate(BitWidth: NewCmpBW); |
3288 | CmpInst::Predicate EquivPred; |
3289 | APInt EquivInt; |
3290 | APInt EquivOffset; |
3291 | |
3292 | SrcCR.getEquivalentICmp(Pred&: EquivPred, RHS&: EquivInt, Offset&: EquivOffset); |
3293 | return new ICmpInst( |
3294 | EquivPred, |
3295 | EquivOffset.isZero() |
3296 | ? V |
3297 | : Builder.CreateAdd(LHS: V, RHS: ConstantInt::get(Ty: NewCmpTy, V: EquivOffset)), |
3298 | ConstantInt::get(Ty: NewCmpTy, V: EquivInt)); |
3299 | } |
3300 | } |
3301 | } |
3302 | |
3303 | return nullptr; |
3304 | } |
3305 | |
3306 | bool InstCombinerImpl::matchThreeWayIntCompare(SelectInst *SI, Value *&LHS, |
3307 | Value *&RHS, ConstantInt *&Less, |
3308 | ConstantInt *&Equal, |
3309 | ConstantInt *&Greater) { |
3310 | // TODO: Generalize this to work with other comparison idioms or ensure |
3311 | // they get canonicalized into this form. |
3312 | |
3313 | // select i1 (a == b), |
3314 | // i32 Equal, |
3315 | // i32 (select i1 (a < b), i32 Less, i32 Greater) |
3316 | // where Equal, Less and Greater are placeholders for any three constants. |
3317 | CmpPredicate PredA; |
3318 | if (!match(V: SI->getCondition(), P: m_ICmp(Pred&: PredA, L: m_Value(V&: LHS), R: m_Value(V&: RHS))) || |
3319 | !ICmpInst::isEquality(P: PredA)) |
3320 | return false; |
3321 | Value *EqualVal = SI->getTrueValue(); |
3322 | Value *UnequalVal = SI->getFalseValue(); |
3323 | // We still can get non-canonical predicate here, so canonicalize. |
3324 | if (PredA == ICmpInst::ICMP_NE) |
3325 | std::swap(a&: EqualVal, b&: UnequalVal); |
3326 | if (!match(V: EqualVal, P: m_ConstantInt(CI&: Equal))) |
3327 | return false; |
3328 | CmpPredicate PredB; |
3329 | Value *LHS2, *RHS2; |
3330 | if (!match(V: UnequalVal, P: m_Select(C: m_ICmp(Pred&: PredB, L: m_Value(V&: LHS2), R: m_Value(V&: RHS2)), |
3331 | L: m_ConstantInt(CI&: Less), R: m_ConstantInt(CI&: Greater)))) |
3332 | return false; |
3333 | // We can get predicate mismatch here, so canonicalize if possible: |
3334 | // First, ensure that 'LHS' match. |
3335 | if (LHS2 != LHS) { |
3336 | // x sgt y <--> y slt x |
3337 | std::swap(a&: LHS2, b&: RHS2); |
3338 | PredB = ICmpInst::getSwappedPredicate(pred: PredB); |
3339 | } |
3340 | if (LHS2 != LHS) |
3341 | return false; |
3342 | // We also need to canonicalize 'RHS'. |
3343 | if (PredB == ICmpInst::ICMP_SGT && isa<Constant>(Val: RHS2)) { |
3344 | // x sgt C-1 <--> x sge C <--> not(x slt C) |
3345 | auto FlippedStrictness = |
3346 | getFlippedStrictnessPredicateAndConstant(Pred: PredB, C: cast<Constant>(Val: RHS2)); |
3347 | if (!FlippedStrictness) |
3348 | return false; |
3349 | assert(FlippedStrictness->first == ICmpInst::ICMP_SGE && |
3350 | "basic correctness failure" ); |
3351 | RHS2 = FlippedStrictness->second; |
3352 | // And kind-of perform the result swap. |
3353 | std::swap(a&: Less, b&: Greater); |
3354 | PredB = ICmpInst::ICMP_SLT; |
3355 | } |
3356 | return PredB == ICmpInst::ICMP_SLT && RHS == RHS2; |
3357 | } |
3358 | |
3359 | Instruction *InstCombinerImpl::foldICmpSelectConstant(ICmpInst &Cmp, |
3360 | SelectInst *Select, |
3361 | ConstantInt *C) { |
3362 | |
3363 | assert(C && "Cmp RHS should be a constant int!" ); |
3364 | // If we're testing a constant value against the result of a three way |
3365 | // comparison, the result can be expressed directly in terms of the |
3366 | // original values being compared. Note: We could possibly be more |
3367 | // aggressive here and remove the hasOneUse test. The original select is |
3368 | // really likely to simplify or sink when we remove a test of the result. |
3369 | Value *OrigLHS, *OrigRHS; |
3370 | ConstantInt *C1LessThan, *C2Equal, *C3GreaterThan; |
3371 | if (Cmp.hasOneUse() && |
3372 | matchThreeWayIntCompare(SI: Select, LHS&: OrigLHS, RHS&: OrigRHS, Less&: C1LessThan, Equal&: C2Equal, |
3373 | Greater&: C3GreaterThan)) { |
3374 | assert(C1LessThan && C2Equal && C3GreaterThan); |
3375 | |
3376 | bool TrueWhenLessThan = ICmpInst::compare( |
3377 | LHS: C1LessThan->getValue(), RHS: C->getValue(), Pred: Cmp.getPredicate()); |
3378 | bool TrueWhenEqual = ICmpInst::compare(LHS: C2Equal->getValue(), RHS: C->getValue(), |
3379 | Pred: Cmp.getPredicate()); |
3380 | bool TrueWhenGreaterThan = ICmpInst::compare( |
3381 | LHS: C3GreaterThan->getValue(), RHS: C->getValue(), Pred: Cmp.getPredicate()); |
3382 | |
3383 | // This generates the new instruction that will replace the original Cmp |
3384 | // Instruction. Instead of enumerating the various combinations when |
3385 | // TrueWhenLessThan, TrueWhenEqual and TrueWhenGreaterThan are true versus |
3386 | // false, we rely on chaining of ORs and future passes of InstCombine to |
3387 | // simplify the OR further (i.e. a s< b || a == b becomes a s<= b). |
3388 | |
3389 | // When none of the three constants satisfy the predicate for the RHS (C), |
3390 | // the entire original Cmp can be simplified to a false. |
3391 | Value *Cond = Builder.getFalse(); |
3392 | if (TrueWhenLessThan) |
3393 | Cond = Builder.CreateOr( |
3394 | LHS: Cond, RHS: Builder.CreateICmp(P: ICmpInst::ICMP_SLT, LHS: OrigLHS, RHS: OrigRHS)); |
3395 | if (TrueWhenEqual) |
3396 | Cond = Builder.CreateOr( |
3397 | LHS: Cond, RHS: Builder.CreateICmp(P: ICmpInst::ICMP_EQ, LHS: OrigLHS, RHS: OrigRHS)); |
3398 | if (TrueWhenGreaterThan) |
3399 | Cond = Builder.CreateOr( |
3400 | LHS: Cond, RHS: Builder.CreateICmp(P: ICmpInst::ICMP_SGT, LHS: OrigLHS, RHS: OrigRHS)); |
3401 | |
3402 | return replaceInstUsesWith(I&: Cmp, V: Cond); |
3403 | } |
3404 | return nullptr; |
3405 | } |
3406 | |
3407 | Instruction *InstCombinerImpl::foldICmpBitCast(ICmpInst &Cmp) { |
3408 | auto *Bitcast = dyn_cast<BitCastInst>(Val: Cmp.getOperand(i_nocapture: 0)); |
3409 | if (!Bitcast) |
3410 | return nullptr; |
3411 | |
3412 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
3413 | Value *Op1 = Cmp.getOperand(i_nocapture: 1); |
3414 | Value *BCSrcOp = Bitcast->getOperand(i_nocapture: 0); |
3415 | Type *SrcType = Bitcast->getSrcTy(); |
3416 | Type *DstType = Bitcast->getType(); |
3417 | |
3418 | // Make sure the bitcast doesn't change between scalar and vector and |
3419 | // doesn't change the number of vector elements. |
3420 | if (SrcType->isVectorTy() == DstType->isVectorTy() && |
3421 | SrcType->getScalarSizeInBits() == DstType->getScalarSizeInBits()) { |
3422 | // Zero-equality and sign-bit checks are preserved through sitofp + bitcast. |
3423 | Value *X; |
3424 | if (match(V: BCSrcOp, P: m_SIToFP(Op: m_Value(V&: X)))) { |
3425 | // icmp eq (bitcast (sitofp X)), 0 --> icmp eq X, 0 |
3426 | // icmp ne (bitcast (sitofp X)), 0 --> icmp ne X, 0 |
3427 | // icmp slt (bitcast (sitofp X)), 0 --> icmp slt X, 0 |
3428 | // icmp sgt (bitcast (sitofp X)), 0 --> icmp sgt X, 0 |
3429 | if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_SLT || |
3430 | Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT) && |
3431 | match(V: Op1, P: m_Zero())) |
3432 | return new ICmpInst(Pred, X, ConstantInt::getNullValue(Ty: X->getType())); |
3433 | |
3434 | // icmp slt (bitcast (sitofp X)), 1 --> icmp slt X, 1 |
3435 | if (Pred == ICmpInst::ICMP_SLT && match(V: Op1, P: m_One())) |
3436 | return new ICmpInst(Pred, X, ConstantInt::get(Ty: X->getType(), V: 1)); |
3437 | |
3438 | // icmp sgt (bitcast (sitofp X)), -1 --> icmp sgt X, -1 |
3439 | if (Pred == ICmpInst::ICMP_SGT && match(V: Op1, P: m_AllOnes())) |
3440 | return new ICmpInst(Pred, X, |
3441 | ConstantInt::getAllOnesValue(Ty: X->getType())); |
3442 | } |
3443 | |
3444 | // Zero-equality checks are preserved through unsigned floating-point casts: |
3445 | // icmp eq (bitcast (uitofp X)), 0 --> icmp eq X, 0 |
3446 | // icmp ne (bitcast (uitofp X)), 0 --> icmp ne X, 0 |
3447 | if (match(V: BCSrcOp, P: m_UIToFP(Op: m_Value(V&: X)))) |
3448 | if (Cmp.isEquality() && match(V: Op1, P: m_Zero())) |
3449 | return new ICmpInst(Pred, X, ConstantInt::getNullValue(Ty: X->getType())); |
3450 | |
3451 | const APInt *C; |
3452 | bool TrueIfSigned; |
3453 | if (match(V: Op1, P: m_APInt(Res&: C)) && Bitcast->hasOneUse()) { |
3454 | // If this is a sign-bit test of a bitcast of a casted FP value, eliminate |
3455 | // the FP extend/truncate because that cast does not change the sign-bit. |
3456 | // This is true for all standard IEEE-754 types and the X86 80-bit type. |
3457 | // The sign-bit is always the most significant bit in those types. |
3458 | if (isSignBitCheck(Pred, RHS: *C, TrueIfSigned) && |
3459 | (match(V: BCSrcOp, P: m_FPExt(Op: m_Value(V&: X))) || |
3460 | match(V: BCSrcOp, P: m_FPTrunc(Op: m_Value(V&: X))))) { |
3461 | // (bitcast (fpext/fptrunc X)) to iX) < 0 --> (bitcast X to iY) < 0 |
3462 | // (bitcast (fpext/fptrunc X)) to iX) > -1 --> (bitcast X to iY) > -1 |
3463 | Type *XType = X->getType(); |
3464 | |
3465 | // We can't currently handle Power style floating point operations here. |
3466 | if (!(XType->isPPC_FP128Ty() || SrcType->isPPC_FP128Ty())) { |
3467 | Type *NewType = Builder.getIntNTy(N: XType->getScalarSizeInBits()); |
3468 | if (auto *XVTy = dyn_cast<VectorType>(Val: XType)) |
3469 | NewType = VectorType::get(ElementType: NewType, EC: XVTy->getElementCount()); |
3470 | Value *NewBitcast = Builder.CreateBitCast(V: X, DestTy: NewType); |
3471 | if (TrueIfSigned) |
3472 | return new ICmpInst(ICmpInst::ICMP_SLT, NewBitcast, |
3473 | ConstantInt::getNullValue(Ty: NewType)); |
3474 | else |
3475 | return new ICmpInst(ICmpInst::ICMP_SGT, NewBitcast, |
3476 | ConstantInt::getAllOnesValue(Ty: NewType)); |
3477 | } |
3478 | } |
3479 | |
3480 | // icmp eq/ne (bitcast X to int), special fp -> llvm.is.fpclass(X, class) |
3481 | Type *FPType = SrcType->getScalarType(); |
3482 | if (!Cmp.getParent()->getParent()->hasFnAttribute( |
3483 | Kind: Attribute::NoImplicitFloat) && |
3484 | Cmp.isEquality() && FPType->isIEEELikeFPTy()) { |
3485 | FPClassTest Mask = APFloat(FPType->getFltSemantics(), *C).classify(); |
3486 | if (Mask & (fcInf | fcZero)) { |
3487 | if (Pred == ICmpInst::ICMP_NE) |
3488 | Mask = ~Mask; |
3489 | return replaceInstUsesWith(I&: Cmp, |
3490 | V: Builder.createIsFPClass(FPNum: BCSrcOp, Test: Mask)); |
3491 | } |
3492 | } |
3493 | } |
3494 | } |
3495 | |
3496 | const APInt *C; |
3497 | if (!match(V: Cmp.getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) || !DstType->isIntegerTy() || |
3498 | !SrcType->isIntOrIntVectorTy()) |
3499 | return nullptr; |
3500 | |
3501 | // If this is checking if all elements of a vector compare are set or not, |
3502 | // invert the casted vector equality compare and test if all compare |
3503 | // elements are clear or not. Compare against zero is generally easier for |
3504 | // analysis and codegen. |
3505 | // icmp eq/ne (bitcast (not X) to iN), -1 --> icmp eq/ne (bitcast X to iN), 0 |
3506 | // Example: are all elements equal? --> are zero elements not equal? |
3507 | // TODO: Try harder to reduce compare of 2 freely invertible operands? |
3508 | if (Cmp.isEquality() && C->isAllOnes() && Bitcast->hasOneUse()) { |
3509 | if (Value *NotBCSrcOp = |
3510 | getFreelyInverted(V: BCSrcOp, WillInvertAllUses: BCSrcOp->hasOneUse(), Builder: &Builder)) { |
3511 | Value *Cast = Builder.CreateBitCast(V: NotBCSrcOp, DestTy: DstType); |
3512 | return new ICmpInst(Pred, Cast, ConstantInt::getNullValue(Ty: DstType)); |
3513 | } |
3514 | } |
3515 | |
3516 | // If this is checking if all elements of an extended vector are clear or not, |
3517 | // compare in a narrow type to eliminate the extend: |
3518 | // icmp eq/ne (bitcast (ext X) to iN), 0 --> icmp eq/ne (bitcast X to iM), 0 |
3519 | Value *X; |
3520 | if (Cmp.isEquality() && C->isZero() && Bitcast->hasOneUse() && |
3521 | match(V: BCSrcOp, P: m_ZExtOrSExt(Op: m_Value(V&: X)))) { |
3522 | if (auto *VecTy = dyn_cast<FixedVectorType>(Val: X->getType())) { |
3523 | Type *NewType = Builder.getIntNTy(N: VecTy->getPrimitiveSizeInBits()); |
3524 | Value *NewCast = Builder.CreateBitCast(V: X, DestTy: NewType); |
3525 | return new ICmpInst(Pred, NewCast, ConstantInt::getNullValue(Ty: NewType)); |
3526 | } |
3527 | } |
3528 | |
3529 | // Folding: icmp <pred> iN X, C |
3530 | // where X = bitcast <M x iK> (shufflevector <M x iK> %vec, undef, SC)) to iN |
3531 | // and C is a splat of a K-bit pattern |
3532 | // and SC is a constant vector = <C', C', C', ..., C'> |
3533 | // Into: |
3534 | // %E = extractelement <M x iK> %vec, i32 C' |
3535 | // icmp <pred> iK %E, trunc(C) |
3536 | Value *Vec; |
3537 | ArrayRef<int> Mask; |
3538 | if (match(V: BCSrcOp, P: m_Shuffle(v1: m_Value(V&: Vec), v2: m_Undef(), mask: m_Mask(Mask)))) { |
3539 | // Check whether every element of Mask is the same constant |
3540 | if (all_equal(Range&: Mask)) { |
3541 | auto *VecTy = cast<VectorType>(Val: SrcType); |
3542 | auto *EltTy = cast<IntegerType>(Val: VecTy->getElementType()); |
3543 | if (C->isSplat(SplatSizeInBits: EltTy->getBitWidth())) { |
3544 | // Fold the icmp based on the value of C |
3545 | // If C is M copies of an iK sized bit pattern, |
3546 | // then: |
3547 | // => %E = extractelement <N x iK> %vec, i32 Elem |
3548 | // icmp <pred> iK %SplatVal, <pattern> |
3549 | Value *Elem = Builder.getInt32(C: Mask[0]); |
3550 | Value * = Builder.CreateExtractElement(Vec, Idx: Elem); |
3551 | Value *NewC = ConstantInt::get(Ty: EltTy, V: C->trunc(width: EltTy->getBitWidth())); |
3552 | return new ICmpInst(Pred, Extract, NewC); |
3553 | } |
3554 | } |
3555 | } |
3556 | return nullptr; |
3557 | } |
3558 | |
3559 | /// Try to fold integer comparisons with a constant operand: icmp Pred X, C |
3560 | /// where X is some kind of instruction. |
3561 | Instruction *InstCombinerImpl::foldICmpInstWithConstant(ICmpInst &Cmp) { |
3562 | const APInt *C; |
3563 | |
3564 | if (match(V: Cmp.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) { |
3565 | if (auto *BO = dyn_cast<BinaryOperator>(Val: Cmp.getOperand(i_nocapture: 0))) |
3566 | if (Instruction *I = foldICmpBinOpWithConstant(Cmp, BO, C: *C)) |
3567 | return I; |
3568 | |
3569 | if (auto *SI = dyn_cast<SelectInst>(Val: Cmp.getOperand(i_nocapture: 0))) |
3570 | // For now, we only support constant integers while folding the |
3571 | // ICMP(SELECT)) pattern. We can extend this to support vector of integers |
3572 | // similar to the cases handled by binary ops above. |
3573 | if (auto *ConstRHS = dyn_cast<ConstantInt>(Val: Cmp.getOperand(i_nocapture: 1))) |
3574 | if (Instruction *I = foldICmpSelectConstant(Cmp, Select: SI, C: ConstRHS)) |
3575 | return I; |
3576 | |
3577 | if (auto *TI = dyn_cast<TruncInst>(Val: Cmp.getOperand(i_nocapture: 0))) |
3578 | if (Instruction *I = foldICmpTruncConstant(Cmp, Trunc: TI, C: *C)) |
3579 | return I; |
3580 | |
3581 | if (auto *II = dyn_cast<IntrinsicInst>(Val: Cmp.getOperand(i_nocapture: 0))) |
3582 | if (Instruction *I = foldICmpIntrinsicWithConstant(ICI&: Cmp, II, C: *C)) |
3583 | return I; |
3584 | |
3585 | // (extractval ([s/u]subo X, Y), 0) == 0 --> X == Y |
3586 | // (extractval ([s/u]subo X, Y), 0) != 0 --> X != Y |
3587 | // TODO: This checks one-use, but that is not strictly necessary. |
3588 | Value *Cmp0 = Cmp.getOperand(i_nocapture: 0); |
3589 | Value *X, *Y; |
3590 | if (C->isZero() && Cmp.isEquality() && Cmp0->hasOneUse() && |
3591 | (match(V: Cmp0, |
3592 | P: m_ExtractValue<0>(V: m_Intrinsic<Intrinsic::ssub_with_overflow>( |
3593 | Op0: m_Value(V&: X), Op1: m_Value(V&: Y)))) || |
3594 | match(V: Cmp0, |
3595 | P: m_ExtractValue<0>(V: m_Intrinsic<Intrinsic::usub_with_overflow>( |
3596 | Op0: m_Value(V&: X), Op1: m_Value(V&: Y)))))) |
3597 | return new ICmpInst(Cmp.getPredicate(), X, Y); |
3598 | } |
3599 | |
3600 | if (match(V: Cmp.getOperand(i_nocapture: 1), P: m_APIntAllowPoison(Res&: C))) |
3601 | return foldICmpInstWithConstantAllowPoison(Cmp, C: *C); |
3602 | |
3603 | return nullptr; |
3604 | } |
3605 | |
3606 | /// Fold an icmp equality instruction with binary operator LHS and constant RHS: |
3607 | /// icmp eq/ne BO, C. |
3608 | Instruction *InstCombinerImpl::foldICmpBinOpEqualityWithConstant( |
3609 | ICmpInst &Cmp, BinaryOperator *BO, const APInt &C) { |
3610 | // TODO: Some of these folds could work with arbitrary constants, but this |
3611 | // function is limited to scalar and vector splat constants. |
3612 | if (!Cmp.isEquality()) |
3613 | return nullptr; |
3614 | |
3615 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
3616 | bool isICMP_NE = Pred == ICmpInst::ICMP_NE; |
3617 | Constant *RHS = cast<Constant>(Val: Cmp.getOperand(i_nocapture: 1)); |
3618 | Value *BOp0 = BO->getOperand(i_nocapture: 0), *BOp1 = BO->getOperand(i_nocapture: 1); |
3619 | |
3620 | switch (BO->getOpcode()) { |
3621 | case Instruction::SRem: |
3622 | // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one. |
3623 | if (C.isZero() && BO->hasOneUse()) { |
3624 | const APInt *BOC; |
3625 | if (match(V: BOp1, P: m_APInt(Res&: BOC)) && BOC->sgt(RHS: 1) && BOC->isPowerOf2()) { |
3626 | Value *NewRem = Builder.CreateURem(LHS: BOp0, RHS: BOp1, Name: BO->getName()); |
3627 | return new ICmpInst(Pred, NewRem, |
3628 | Constant::getNullValue(Ty: BO->getType())); |
3629 | } |
3630 | } |
3631 | break; |
3632 | case Instruction::Add: { |
3633 | // (A + C2) == C --> A == (C - C2) |
3634 | // (A + C2) != C --> A != (C - C2) |
3635 | // TODO: Remove the one-use limitation? See discussion in D58633. |
3636 | if (Constant *C2 = dyn_cast<Constant>(Val: BOp1)) { |
3637 | if (BO->hasOneUse()) |
3638 | return new ICmpInst(Pred, BOp0, ConstantExpr::getSub(C1: RHS, C2)); |
3639 | } else if (C.isZero()) { |
3640 | // Replace ((add A, B) != 0) with (A != -B) if A or B is |
3641 | // efficiently invertible, or if the add has just this one use. |
3642 | if (Value *NegVal = dyn_castNegVal(V: BOp1)) |
3643 | return new ICmpInst(Pred, BOp0, NegVal); |
3644 | if (Value *NegVal = dyn_castNegVal(V: BOp0)) |
3645 | return new ICmpInst(Pred, NegVal, BOp1); |
3646 | if (BO->hasOneUse()) { |
3647 | // (add nuw A, B) != 0 -> (or A, B) != 0 |
3648 | if (match(V: BO, P: m_NUWAdd(L: m_Value(), R: m_Value()))) { |
3649 | Value *Or = Builder.CreateOr(LHS: BOp0, RHS: BOp1); |
3650 | return new ICmpInst(Pred, Or, Constant::getNullValue(Ty: BO->getType())); |
3651 | } |
3652 | Value *Neg = Builder.CreateNeg(V: BOp1); |
3653 | Neg->takeName(V: BO); |
3654 | return new ICmpInst(Pred, BOp0, Neg); |
3655 | } |
3656 | } |
3657 | break; |
3658 | } |
3659 | case Instruction::Xor: |
3660 | if (Constant *BOC = dyn_cast<Constant>(Val: BOp1)) { |
3661 | // For the xor case, we can xor two constants together, eliminating |
3662 | // the explicit xor. |
3663 | return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(C1: RHS, C2: BOC)); |
3664 | } else if (C.isZero()) { |
3665 | // Replace ((xor A, B) != 0) with (A != B) |
3666 | return new ICmpInst(Pred, BOp0, BOp1); |
3667 | } |
3668 | break; |
3669 | case Instruction::Or: { |
3670 | const APInt *BOC; |
3671 | if (match(V: BOp1, P: m_APInt(Res&: BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) { |
3672 | // Comparing if all bits outside of a constant mask are set? |
3673 | // Replace (X | C) == -1 with (X & ~C) == ~C. |
3674 | // This removes the -1 constant. |
3675 | Constant *NotBOC = ConstantExpr::getNot(C: cast<Constant>(Val: BOp1)); |
3676 | Value *And = Builder.CreateAnd(LHS: BOp0, RHS: NotBOC); |
3677 | return new ICmpInst(Pred, And, NotBOC); |
3678 | } |
3679 | // (icmp eq (or (select cond, 0, NonZero), Other), 0) |
3680 | // -> (and cond, (icmp eq Other, 0)) |
3681 | // (icmp ne (or (select cond, NonZero, 0), Other), 0) |
3682 | // -> (or cond, (icmp ne Other, 0)) |
3683 | Value *Cond, *TV, *FV, *Other, *Sel; |
3684 | if (C.isZero() && |
3685 | match(V: BO, |
3686 | P: m_OneUse(SubPattern: m_c_Or(L: m_CombineAnd(L: m_Value(V&: Sel), |
3687 | R: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: TV), |
3688 | R: m_Value(V&: FV))), |
3689 | R: m_Value(V&: Other)))) && |
3690 | Cond->getType() == Cmp.getType()) { |
3691 | const SimplifyQuery Q = SQ.getWithInstruction(I: &Cmp); |
3692 | // Easy case is if eq/ne matches whether 0 is trueval/falseval. |
3693 | if (Pred == ICmpInst::ICMP_EQ |
3694 | ? (match(V: TV, P: m_Zero()) && isKnownNonZero(V: FV, Q)) |
3695 | : (match(V: FV, P: m_Zero()) && isKnownNonZero(V: TV, Q))) { |
3696 | Value *Cmp = Builder.CreateICmp( |
3697 | P: Pred, LHS: Other, RHS: Constant::getNullValue(Ty: Other->getType())); |
3698 | return BinaryOperator::Create( |
3699 | Op: Pred == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or, S1: Cmp, |
3700 | S2: Cond); |
3701 | } |
3702 | // Harder case is if eq/ne matches whether 0 is falseval/trueval. In this |
3703 | // case we need to invert the select condition so we need to be careful to |
3704 | // avoid creating extra instructions. |
3705 | // (icmp ne (or (select cond, 0, NonZero), Other), 0) |
3706 | // -> (or (not cond), (icmp ne Other, 0)) |
3707 | // (icmp eq (or (select cond, NonZero, 0), Other), 0) |
3708 | // -> (and (not cond), (icmp eq Other, 0)) |
3709 | // |
3710 | // Only do this if the inner select has one use, in which case we are |
3711 | // replacing `select` with `(not cond)`. Otherwise, we will create more |
3712 | // uses. NB: Trying to freely invert cond doesn't make sense here, as if |
3713 | // cond was freely invertable, the select arms would have been inverted. |
3714 | if (Sel->hasOneUse() && |
3715 | (Pred == ICmpInst::ICMP_EQ |
3716 | ? (match(V: FV, P: m_Zero()) && isKnownNonZero(V: TV, Q)) |
3717 | : (match(V: TV, P: m_Zero()) && isKnownNonZero(V: FV, Q)))) { |
3718 | Value *NotCond = Builder.CreateNot(V: Cond); |
3719 | Value *Cmp = Builder.CreateICmp( |
3720 | P: Pred, LHS: Other, RHS: Constant::getNullValue(Ty: Other->getType())); |
3721 | return BinaryOperator::Create( |
3722 | Op: Pred == ICmpInst::ICMP_EQ ? Instruction::And : Instruction::Or, S1: Cmp, |
3723 | S2: NotCond); |
3724 | } |
3725 | } |
3726 | break; |
3727 | } |
3728 | case Instruction::UDiv: |
3729 | case Instruction::SDiv: |
3730 | if (BO->isExact()) { |
3731 | // div exact X, Y eq/ne 0 -> X eq/ne 0 |
3732 | // div exact X, Y eq/ne 1 -> X eq/ne Y |
3733 | // div exact X, Y eq/ne C -> |
3734 | // if Y * C never-overflow && OneUse: |
3735 | // -> Y * C eq/ne X |
3736 | if (C.isZero()) |
3737 | return new ICmpInst(Pred, BOp0, Constant::getNullValue(Ty: BO->getType())); |
3738 | else if (C.isOne()) |
3739 | return new ICmpInst(Pred, BOp0, BOp1); |
3740 | else if (BO->hasOneUse()) { |
3741 | OverflowResult OR = computeOverflow( |
3742 | BinaryOp: Instruction::Mul, IsSigned: BO->getOpcode() == Instruction::SDiv, LHS: BOp1, |
3743 | RHS: Cmp.getOperand(i_nocapture: 1), CxtI: BO); |
3744 | if (OR == OverflowResult::NeverOverflows) { |
3745 | Value *YC = |
3746 | Builder.CreateMul(LHS: BOp1, RHS: ConstantInt::get(Ty: BO->getType(), V: C)); |
3747 | return new ICmpInst(Pred, YC, BOp0); |
3748 | } |
3749 | } |
3750 | } |
3751 | if (BO->getOpcode() == Instruction::UDiv && C.isZero()) { |
3752 | // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A) |
3753 | auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT; |
3754 | return new ICmpInst(NewPred, BOp1, BOp0); |
3755 | } |
3756 | break; |
3757 | default: |
3758 | break; |
3759 | } |
3760 | return nullptr; |
3761 | } |
3762 | |
3763 | static Instruction *foldCtpopPow2Test(ICmpInst &I, IntrinsicInst *CtpopLhs, |
3764 | const APInt &CRhs, |
3765 | InstCombiner::BuilderTy &Builder, |
3766 | const SimplifyQuery &Q) { |
3767 | assert(CtpopLhs->getIntrinsicID() == Intrinsic::ctpop && |
3768 | "Non-ctpop intrin in ctpop fold" ); |
3769 | if (!CtpopLhs->hasOneUse()) |
3770 | return nullptr; |
3771 | |
3772 | // Power of 2 test: |
3773 | // isPow2OrZero : ctpop(X) u< 2 |
3774 | // isPow2 : ctpop(X) == 1 |
3775 | // NotPow2OrZero: ctpop(X) u> 1 |
3776 | // NotPow2 : ctpop(X) != 1 |
3777 | // If we know any bit of X can be folded to: |
3778 | // IsPow2 : X & (~Bit) == 0 |
3779 | // NotPow2 : X & (~Bit) != 0 |
3780 | const ICmpInst::Predicate Pred = I.getPredicate(); |
3781 | if (((I.isEquality() || Pred == ICmpInst::ICMP_UGT) && CRhs == 1) || |
3782 | (Pred == ICmpInst::ICMP_ULT && CRhs == 2)) { |
3783 | Value *Op = CtpopLhs->getArgOperand(i: 0); |
3784 | KnownBits OpKnown = computeKnownBits(V: Op, DL: Q.DL, AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT); |
3785 | // No need to check for count > 1, that should be already constant folded. |
3786 | if (OpKnown.countMinPopulation() == 1) { |
3787 | Value *And = Builder.CreateAnd( |
3788 | LHS: Op, RHS: Constant::getIntegerValue(Ty: Op->getType(), V: ~(OpKnown.One))); |
3789 | return new ICmpInst( |
3790 | (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_ULT) |
3791 | ? ICmpInst::ICMP_EQ |
3792 | : ICmpInst::ICMP_NE, |
3793 | And, Constant::getNullValue(Ty: Op->getType())); |
3794 | } |
3795 | } |
3796 | |
3797 | return nullptr; |
3798 | } |
3799 | |
3800 | /// Fold an equality icmp with LLVM intrinsic and constant operand. |
3801 | Instruction *InstCombinerImpl::foldICmpEqIntrinsicWithConstant( |
3802 | ICmpInst &Cmp, IntrinsicInst *II, const APInt &C) { |
3803 | Type *Ty = II->getType(); |
3804 | unsigned BitWidth = C.getBitWidth(); |
3805 | const ICmpInst::Predicate Pred = Cmp.getPredicate(); |
3806 | |
3807 | switch (II->getIntrinsicID()) { |
3808 | case Intrinsic::abs: |
3809 | // abs(A) == 0 -> A == 0 |
3810 | // abs(A) == INT_MIN -> A == INT_MIN |
3811 | if (C.isZero() || C.isMinSignedValue()) |
3812 | return new ICmpInst(Pred, II->getArgOperand(i: 0), ConstantInt::get(Ty, V: C)); |
3813 | break; |
3814 | |
3815 | case Intrinsic::bswap: |
3816 | // bswap(A) == C -> A == bswap(C) |
3817 | return new ICmpInst(Pred, II->getArgOperand(i: 0), |
3818 | ConstantInt::get(Ty, V: C.byteSwap())); |
3819 | |
3820 | case Intrinsic::bitreverse: |
3821 | // bitreverse(A) == C -> A == bitreverse(C) |
3822 | return new ICmpInst(Pred, II->getArgOperand(i: 0), |
3823 | ConstantInt::get(Ty, V: C.reverseBits())); |
3824 | |
3825 | case Intrinsic::ctlz: |
3826 | case Intrinsic::cttz: { |
3827 | // ctz(A) == bitwidth(A) -> A == 0 and likewise for != |
3828 | if (C == BitWidth) |
3829 | return new ICmpInst(Pred, II->getArgOperand(i: 0), |
3830 | ConstantInt::getNullValue(Ty)); |
3831 | |
3832 | // ctz(A) == C -> A & Mask1 == Mask2, where Mask2 only has bit C set |
3833 | // and Mask1 has bits 0..C+1 set. Similar for ctl, but for high bits. |
3834 | // Limit to one use to ensure we don't increase instruction count. |
3835 | unsigned Num = C.getLimitedValue(Limit: BitWidth); |
3836 | if (Num != BitWidth && II->hasOneUse()) { |
3837 | bool IsTrailing = II->getIntrinsicID() == Intrinsic::cttz; |
3838 | APInt Mask1 = IsTrailing ? APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: Num + 1) |
3839 | : APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: Num + 1); |
3840 | APInt Mask2 = IsTrailing |
3841 | ? APInt::getOneBitSet(numBits: BitWidth, BitNo: Num) |
3842 | : APInt::getOneBitSet(numBits: BitWidth, BitNo: BitWidth - Num - 1); |
3843 | return new ICmpInst(Pred, Builder.CreateAnd(LHS: II->getArgOperand(i: 0), RHS: Mask1), |
3844 | ConstantInt::get(Ty, V: Mask2)); |
3845 | } |
3846 | break; |
3847 | } |
3848 | |
3849 | case Intrinsic::ctpop: { |
3850 | // popcount(A) == 0 -> A == 0 and likewise for != |
3851 | // popcount(A) == bitwidth(A) -> A == -1 and likewise for != |
3852 | bool IsZero = C.isZero(); |
3853 | if (IsZero || C == BitWidth) |
3854 | return new ICmpInst(Pred, II->getArgOperand(i: 0), |
3855 | IsZero ? Constant::getNullValue(Ty) |
3856 | : Constant::getAllOnesValue(Ty)); |
3857 | |
3858 | break; |
3859 | } |
3860 | |
3861 | case Intrinsic::fshl: |
3862 | case Intrinsic::fshr: |
3863 | if (II->getArgOperand(i: 0) == II->getArgOperand(i: 1)) { |
3864 | const APInt *RotAmtC; |
3865 | // ror(X, RotAmtC) == C --> X == rol(C, RotAmtC) |
3866 | // rol(X, RotAmtC) == C --> X == ror(C, RotAmtC) |
3867 | if (match(V: II->getArgOperand(i: 2), P: m_APInt(Res&: RotAmtC))) |
3868 | return new ICmpInst(Pred, II->getArgOperand(i: 0), |
3869 | II->getIntrinsicID() == Intrinsic::fshl |
3870 | ? ConstantInt::get(Ty, V: C.rotr(rotateAmt: *RotAmtC)) |
3871 | : ConstantInt::get(Ty, V: C.rotl(rotateAmt: *RotAmtC))); |
3872 | } |
3873 | break; |
3874 | |
3875 | case Intrinsic::umax: |
3876 | case Intrinsic::uadd_sat: { |
3877 | // uadd.sat(a, b) == 0 -> (a | b) == 0 |
3878 | // umax(a, b) == 0 -> (a | b) == 0 |
3879 | if (C.isZero() && II->hasOneUse()) { |
3880 | Value *Or = Builder.CreateOr(LHS: II->getArgOperand(i: 0), RHS: II->getArgOperand(i: 1)); |
3881 | return new ICmpInst(Pred, Or, Constant::getNullValue(Ty)); |
3882 | } |
3883 | break; |
3884 | } |
3885 | |
3886 | case Intrinsic::ssub_sat: |
3887 | // ssub.sat(a, b) == 0 -> a == b |
3888 | if (C.isZero()) |
3889 | return new ICmpInst(Pred, II->getArgOperand(i: 0), II->getArgOperand(i: 1)); |
3890 | break; |
3891 | case Intrinsic::usub_sat: { |
3892 | // usub.sat(a, b) == 0 -> a <= b |
3893 | if (C.isZero()) { |
3894 | ICmpInst::Predicate NewPred = |
3895 | Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT; |
3896 | return new ICmpInst(NewPred, II->getArgOperand(i: 0), II->getArgOperand(i: 1)); |
3897 | } |
3898 | break; |
3899 | } |
3900 | default: |
3901 | break; |
3902 | } |
3903 | |
3904 | return nullptr; |
3905 | } |
3906 | |
3907 | /// Fold an icmp with LLVM intrinsics |
3908 | static Instruction * |
3909 | foldICmpIntrinsicWithIntrinsic(ICmpInst &Cmp, |
3910 | InstCombiner::BuilderTy &Builder) { |
3911 | assert(Cmp.isEquality()); |
3912 | |
3913 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
3914 | Value *Op0 = Cmp.getOperand(i_nocapture: 0); |
3915 | Value *Op1 = Cmp.getOperand(i_nocapture: 1); |
3916 | const auto *IIOp0 = dyn_cast<IntrinsicInst>(Val: Op0); |
3917 | const auto *IIOp1 = dyn_cast<IntrinsicInst>(Val: Op1); |
3918 | if (!IIOp0 || !IIOp1 || IIOp0->getIntrinsicID() != IIOp1->getIntrinsicID()) |
3919 | return nullptr; |
3920 | |
3921 | switch (IIOp0->getIntrinsicID()) { |
3922 | case Intrinsic::bswap: |
3923 | case Intrinsic::bitreverse: |
3924 | // If both operands are byte-swapped or bit-reversed, just compare the |
3925 | // original values. |
3926 | return new ICmpInst(Pred, IIOp0->getOperand(i_nocapture: 0), IIOp1->getOperand(i_nocapture: 0)); |
3927 | case Intrinsic::fshl: |
3928 | case Intrinsic::fshr: { |
3929 | // If both operands are rotated by same amount, just compare the |
3930 | // original values. |
3931 | if (IIOp0->getOperand(i_nocapture: 0) != IIOp0->getOperand(i_nocapture: 1)) |
3932 | break; |
3933 | if (IIOp1->getOperand(i_nocapture: 0) != IIOp1->getOperand(i_nocapture: 1)) |
3934 | break; |
3935 | if (IIOp0->getOperand(i_nocapture: 2) == IIOp1->getOperand(i_nocapture: 2)) |
3936 | return new ICmpInst(Pred, IIOp0->getOperand(i_nocapture: 0), IIOp1->getOperand(i_nocapture: 0)); |
3937 | |
3938 | // rotate(X, AmtX) == rotate(Y, AmtY) |
3939 | // -> rotate(X, AmtX - AmtY) == Y |
3940 | // Do this if either both rotates have one use or if only one has one use |
3941 | // and AmtX/AmtY are constants. |
3942 | unsigned OneUses = IIOp0->hasOneUse() + IIOp1->hasOneUse(); |
3943 | if (OneUses == 2 || |
3944 | (OneUses == 1 && match(V: IIOp0->getOperand(i_nocapture: 2), P: m_ImmConstant()) && |
3945 | match(V: IIOp1->getOperand(i_nocapture: 2), P: m_ImmConstant()))) { |
3946 | Value *SubAmt = |
3947 | Builder.CreateSub(LHS: IIOp0->getOperand(i_nocapture: 2), RHS: IIOp1->getOperand(i_nocapture: 2)); |
3948 | Value *CombinedRotate = Builder.CreateIntrinsic( |
3949 | RetTy: Op0->getType(), ID: IIOp0->getIntrinsicID(), |
3950 | Args: {IIOp0->getOperand(i_nocapture: 0), IIOp0->getOperand(i_nocapture: 0), SubAmt}); |
3951 | return new ICmpInst(Pred, IIOp1->getOperand(i_nocapture: 0), CombinedRotate); |
3952 | } |
3953 | } break; |
3954 | default: |
3955 | break; |
3956 | } |
3957 | |
3958 | return nullptr; |
3959 | } |
3960 | |
3961 | /// Try to fold integer comparisons with a constant operand: icmp Pred X, C |
3962 | /// where X is some kind of instruction and C is AllowPoison. |
3963 | /// TODO: Move more folds which allow poison to this function. |
3964 | Instruction * |
3965 | InstCombinerImpl::foldICmpInstWithConstantAllowPoison(ICmpInst &Cmp, |
3966 | const APInt &C) { |
3967 | const ICmpInst::Predicate Pred = Cmp.getPredicate(); |
3968 | if (auto *II = dyn_cast<IntrinsicInst>(Val: Cmp.getOperand(i_nocapture: 0))) { |
3969 | switch (II->getIntrinsicID()) { |
3970 | default: |
3971 | break; |
3972 | case Intrinsic::fshl: |
3973 | case Intrinsic::fshr: |
3974 | if (Cmp.isEquality() && II->getArgOperand(i: 0) == II->getArgOperand(i: 1)) { |
3975 | // (rot X, ?) == 0/-1 --> X == 0/-1 |
3976 | if (C.isZero() || C.isAllOnes()) |
3977 | return new ICmpInst(Pred, II->getArgOperand(i: 0), Cmp.getOperand(i_nocapture: 1)); |
3978 | } |
3979 | break; |
3980 | } |
3981 | } |
3982 | |
3983 | return nullptr; |
3984 | } |
3985 | |
3986 | /// Fold an icmp with BinaryOp and constant operand: icmp Pred BO, C. |
3987 | Instruction *InstCombinerImpl::foldICmpBinOpWithConstant(ICmpInst &Cmp, |
3988 | BinaryOperator *BO, |
3989 | const APInt &C) { |
3990 | switch (BO->getOpcode()) { |
3991 | case Instruction::Xor: |
3992 | if (Instruction *I = foldICmpXorConstant(Cmp, Xor: BO, C)) |
3993 | return I; |
3994 | break; |
3995 | case Instruction::And: |
3996 | if (Instruction *I = foldICmpAndConstant(Cmp, And: BO, C)) |
3997 | return I; |
3998 | break; |
3999 | case Instruction::Or: |
4000 | if (Instruction *I = foldICmpOrConstant(Cmp, Or: BO, C)) |
4001 | return I; |
4002 | break; |
4003 | case Instruction::Mul: |
4004 | if (Instruction *I = foldICmpMulConstant(Cmp, Mul: BO, C)) |
4005 | return I; |
4006 | break; |
4007 | case Instruction::Shl: |
4008 | if (Instruction *I = foldICmpShlConstant(Cmp, Shl: BO, C)) |
4009 | return I; |
4010 | break; |
4011 | case Instruction::LShr: |
4012 | case Instruction::AShr: |
4013 | if (Instruction *I = foldICmpShrConstant(Cmp, Shr: BO, C)) |
4014 | return I; |
4015 | break; |
4016 | case Instruction::SRem: |
4017 | if (Instruction *I = foldICmpSRemConstant(Cmp, SRem: BO, C)) |
4018 | return I; |
4019 | break; |
4020 | case Instruction::UDiv: |
4021 | if (Instruction *I = foldICmpUDivConstant(Cmp, UDiv: BO, C)) |
4022 | return I; |
4023 | [[fallthrough]]; |
4024 | case Instruction::SDiv: |
4025 | if (Instruction *I = foldICmpDivConstant(Cmp, Div: BO, C)) |
4026 | return I; |
4027 | break; |
4028 | case Instruction::Sub: |
4029 | if (Instruction *I = foldICmpSubConstant(Cmp, Sub: BO, C)) |
4030 | return I; |
4031 | break; |
4032 | case Instruction::Add: |
4033 | if (Instruction *I = foldICmpAddConstant(Cmp, Add: BO, C)) |
4034 | return I; |
4035 | break; |
4036 | default: |
4037 | break; |
4038 | } |
4039 | |
4040 | // TODO: These folds could be refactored to be part of the above calls. |
4041 | if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, C)) |
4042 | return I; |
4043 | |
4044 | // Fall back to handling `icmp pred (select A ? C1 : C2) binop (select B ? C3 |
4045 | // : C4), C5` pattern, by computing a truth table of the four constant |
4046 | // variants. |
4047 | return foldICmpBinOpWithConstantViaTruthTable(Cmp, BO, C); |
4048 | } |
4049 | |
4050 | static Instruction * |
4051 | foldICmpUSubSatOrUAddSatWithConstant(CmpPredicate Pred, SaturatingInst *II, |
4052 | const APInt &C, |
4053 | InstCombiner::BuilderTy &Builder) { |
4054 | // This transform may end up producing more than one instruction for the |
4055 | // intrinsic, so limit it to one user of the intrinsic. |
4056 | if (!II->hasOneUse()) |
4057 | return nullptr; |
4058 | |
4059 | // Let Y = [add/sub]_sat(X, C) pred C2 |
4060 | // SatVal = The saturating value for the operation |
4061 | // WillWrap = Whether or not the operation will underflow / overflow |
4062 | // => Y = (WillWrap ? SatVal : (X binop C)) pred C2 |
4063 | // => Y = WillWrap ? (SatVal pred C2) : ((X binop C) pred C2) |
4064 | // |
4065 | // When (SatVal pred C2) is true, then |
4066 | // Y = WillWrap ? true : ((X binop C) pred C2) |
4067 | // => Y = WillWrap || ((X binop C) pred C2) |
4068 | // else |
4069 | // Y = WillWrap ? false : ((X binop C) pred C2) |
4070 | // => Y = !WillWrap ? ((X binop C) pred C2) : false |
4071 | // => Y = !WillWrap && ((X binop C) pred C2) |
4072 | Value *Op0 = II->getOperand(i_nocapture: 0); |
4073 | Value *Op1 = II->getOperand(i_nocapture: 1); |
4074 | |
4075 | const APInt *COp1; |
4076 | // This transform only works when the intrinsic has an integral constant or |
4077 | // splat vector as the second operand. |
4078 | if (!match(V: Op1, P: m_APInt(Res&: COp1))) |
4079 | return nullptr; |
4080 | |
4081 | APInt SatVal; |
4082 | switch (II->getIntrinsicID()) { |
4083 | default: |
4084 | llvm_unreachable( |
4085 | "This function only works with usub_sat and uadd_sat for now!" ); |
4086 | case Intrinsic::uadd_sat: |
4087 | SatVal = APInt::getAllOnes(numBits: C.getBitWidth()); |
4088 | break; |
4089 | case Intrinsic::usub_sat: |
4090 | SatVal = APInt::getZero(numBits: C.getBitWidth()); |
4091 | break; |
4092 | } |
4093 | |
4094 | // Check (SatVal pred C2) |
4095 | bool SatValCheck = ICmpInst::compare(LHS: SatVal, RHS: C, Pred); |
4096 | |
4097 | // !WillWrap. |
4098 | ConstantRange C1 = ConstantRange::makeExactNoWrapRegion( |
4099 | BinOp: II->getBinaryOp(), Other: *COp1, NoWrapKind: II->getNoWrapKind()); |
4100 | |
4101 | // WillWrap. |
4102 | if (SatValCheck) |
4103 | C1 = C1.inverse(); |
4104 | |
4105 | ConstantRange C2 = ConstantRange::makeExactICmpRegion(Pred, Other: C); |
4106 | if (II->getBinaryOp() == Instruction::Add) |
4107 | C2 = C2.sub(Other: *COp1); |
4108 | else |
4109 | C2 = C2.add(Other: *COp1); |
4110 | |
4111 | Instruction::BinaryOps CombiningOp = |
4112 | SatValCheck ? Instruction::BinaryOps::Or : Instruction::BinaryOps::And; |
4113 | |
4114 | std::optional<ConstantRange> Combination; |
4115 | if (CombiningOp == Instruction::BinaryOps::Or) |
4116 | Combination = C1.exactUnionWith(CR: C2); |
4117 | else /* CombiningOp == Instruction::BinaryOps::And */ |
4118 | Combination = C1.exactIntersectWith(CR: C2); |
4119 | |
4120 | if (!Combination) |
4121 | return nullptr; |
4122 | |
4123 | CmpInst::Predicate EquivPred; |
4124 | APInt EquivInt; |
4125 | APInt EquivOffset; |
4126 | |
4127 | Combination->getEquivalentICmp(Pred&: EquivPred, RHS&: EquivInt, Offset&: EquivOffset); |
4128 | |
4129 | return new ICmpInst( |
4130 | EquivPred, |
4131 | Builder.CreateAdd(LHS: Op0, RHS: ConstantInt::get(Ty: Op1->getType(), V: EquivOffset)), |
4132 | ConstantInt::get(Ty: Op1->getType(), V: EquivInt)); |
4133 | } |
4134 | |
4135 | static Instruction * |
4136 | foldICmpOfCmpIntrinsicWithConstant(CmpPredicate Pred, IntrinsicInst *I, |
4137 | const APInt &C, |
4138 | InstCombiner::BuilderTy &Builder) { |
4139 | std::optional<ICmpInst::Predicate> NewPredicate = std::nullopt; |
4140 | switch (Pred) { |
4141 | case ICmpInst::ICMP_EQ: |
4142 | case ICmpInst::ICMP_NE: |
4143 | if (C.isZero()) |
4144 | NewPredicate = Pred; |
4145 | else if (C.isOne()) |
4146 | NewPredicate = |
4147 | Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_ULE; |
4148 | else if (C.isAllOnes()) |
4149 | NewPredicate = |
4150 | Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE; |
4151 | break; |
4152 | |
4153 | case ICmpInst::ICMP_SGT: |
4154 | if (C.isAllOnes()) |
4155 | NewPredicate = ICmpInst::ICMP_UGE; |
4156 | else if (C.isZero()) |
4157 | NewPredicate = ICmpInst::ICMP_UGT; |
4158 | break; |
4159 | |
4160 | case ICmpInst::ICMP_SLT: |
4161 | if (C.isZero()) |
4162 | NewPredicate = ICmpInst::ICMP_ULT; |
4163 | else if (C.isOne()) |
4164 | NewPredicate = ICmpInst::ICMP_ULE; |
4165 | break; |
4166 | |
4167 | case ICmpInst::ICMP_ULT: |
4168 | if (C.ugt(RHS: 1)) |
4169 | NewPredicate = ICmpInst::ICMP_UGE; |
4170 | break; |
4171 | |
4172 | case ICmpInst::ICMP_UGT: |
4173 | if (!C.isZero() && !C.isAllOnes()) |
4174 | NewPredicate = ICmpInst::ICMP_ULT; |
4175 | break; |
4176 | |
4177 | default: |
4178 | break; |
4179 | } |
4180 | |
4181 | if (!NewPredicate) |
4182 | return nullptr; |
4183 | |
4184 | if (I->getIntrinsicID() == Intrinsic::scmp) |
4185 | NewPredicate = ICmpInst::getSignedPredicate(Pred: *NewPredicate); |
4186 | Value *LHS = I->getOperand(i_nocapture: 0); |
4187 | Value *RHS = I->getOperand(i_nocapture: 1); |
4188 | return new ICmpInst(*NewPredicate, LHS, RHS); |
4189 | } |
4190 | |
4191 | /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C. |
4192 | Instruction *InstCombinerImpl::foldICmpIntrinsicWithConstant(ICmpInst &Cmp, |
4193 | IntrinsicInst *II, |
4194 | const APInt &C) { |
4195 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
4196 | |
4197 | // Handle folds that apply for any kind of icmp. |
4198 | switch (II->getIntrinsicID()) { |
4199 | default: |
4200 | break; |
4201 | case Intrinsic::uadd_sat: |
4202 | case Intrinsic::usub_sat: |
4203 | if (auto *Folded = foldICmpUSubSatOrUAddSatWithConstant( |
4204 | Pred, II: cast<SaturatingInst>(Val: II), C, Builder)) |
4205 | return Folded; |
4206 | break; |
4207 | case Intrinsic::ctpop: { |
4208 | const SimplifyQuery Q = SQ.getWithInstruction(I: &Cmp); |
4209 | if (Instruction *R = foldCtpopPow2Test(I&: Cmp, CtpopLhs: II, CRhs: C, Builder, Q)) |
4210 | return R; |
4211 | } break; |
4212 | case Intrinsic::scmp: |
4213 | case Intrinsic::ucmp: |
4214 | if (auto *Folded = foldICmpOfCmpIntrinsicWithConstant(Pred, I: II, C, Builder)) |
4215 | return Folded; |
4216 | break; |
4217 | } |
4218 | |
4219 | if (Cmp.isEquality()) |
4220 | return foldICmpEqIntrinsicWithConstant(Cmp, II, C); |
4221 | |
4222 | Type *Ty = II->getType(); |
4223 | unsigned BitWidth = C.getBitWidth(); |
4224 | switch (II->getIntrinsicID()) { |
4225 | case Intrinsic::ctpop: { |
4226 | // (ctpop X > BitWidth - 1) --> X == -1 |
4227 | Value *X = II->getArgOperand(i: 0); |
4228 | if (C == BitWidth - 1 && Pred == ICmpInst::ICMP_UGT) |
4229 | return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_EQ, S1: X, |
4230 | S2: ConstantInt::getAllOnesValue(Ty)); |
4231 | // (ctpop X < BitWidth) --> X != -1 |
4232 | if (C == BitWidth && Pred == ICmpInst::ICMP_ULT) |
4233 | return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_NE, S1: X, |
4234 | S2: ConstantInt::getAllOnesValue(Ty)); |
4235 | break; |
4236 | } |
4237 | case Intrinsic::ctlz: { |
4238 | // ctlz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX < 0b00010000 |
4239 | if (Pred == ICmpInst::ICMP_UGT && C.ult(RHS: BitWidth)) { |
4240 | unsigned Num = C.getLimitedValue(); |
4241 | APInt Limit = APInt::getOneBitSet(numBits: BitWidth, BitNo: BitWidth - Num - 1); |
4242 | return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_ULT, |
4243 | S1: II->getArgOperand(i: 0), S2: ConstantInt::get(Ty, V: Limit)); |
4244 | } |
4245 | |
4246 | // ctlz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX > 0b00011111 |
4247 | if (Pred == ICmpInst::ICMP_ULT && C.uge(RHS: 1) && C.ule(RHS: BitWidth)) { |
4248 | unsigned Num = C.getLimitedValue(); |
4249 | APInt Limit = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: BitWidth - Num); |
4250 | return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_UGT, |
4251 | S1: II->getArgOperand(i: 0), S2: ConstantInt::get(Ty, V: Limit)); |
4252 | } |
4253 | break; |
4254 | } |
4255 | case Intrinsic::cttz: { |
4256 | // Limit to one use to ensure we don't increase instruction count. |
4257 | if (!II->hasOneUse()) |
4258 | return nullptr; |
4259 | |
4260 | // cttz(0bXXXXXXXX) > 3 -> 0bXXXXXXXX & 0b00001111 == 0 |
4261 | if (Pred == ICmpInst::ICMP_UGT && C.ult(RHS: BitWidth)) { |
4262 | APInt Mask = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: C.getLimitedValue() + 1); |
4263 | return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_EQ, |
4264 | S1: Builder.CreateAnd(LHS: II->getArgOperand(i: 0), RHS: Mask), |
4265 | S2: ConstantInt::getNullValue(Ty)); |
4266 | } |
4267 | |
4268 | // cttz(0bXXXXXXXX) < 3 -> 0bXXXXXXXX & 0b00000111 != 0 |
4269 | if (Pred == ICmpInst::ICMP_ULT && C.uge(RHS: 1) && C.ule(RHS: BitWidth)) { |
4270 | APInt Mask = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: C.getLimitedValue()); |
4271 | return CmpInst::Create(Op: Instruction::ICmp, Pred: ICmpInst::ICMP_NE, |
4272 | S1: Builder.CreateAnd(LHS: II->getArgOperand(i: 0), RHS: Mask), |
4273 | S2: ConstantInt::getNullValue(Ty)); |
4274 | } |
4275 | break; |
4276 | } |
4277 | case Intrinsic::ssub_sat: |
4278 | // ssub.sat(a, b) spred 0 -> a spred b |
4279 | if (ICmpInst::isSigned(predicate: Pred)) { |
4280 | if (C.isZero()) |
4281 | return new ICmpInst(Pred, II->getArgOperand(i: 0), II->getArgOperand(i: 1)); |
4282 | // X s<= 0 is cannonicalized to X s< 1 |
4283 | if (Pred == ICmpInst::ICMP_SLT && C.isOne()) |
4284 | return new ICmpInst(ICmpInst::ICMP_SLE, II->getArgOperand(i: 0), |
4285 | II->getArgOperand(i: 1)); |
4286 | // X s>= 0 is cannonicalized to X s> -1 |
4287 | if (Pred == ICmpInst::ICMP_SGT && C.isAllOnes()) |
4288 | return new ICmpInst(ICmpInst::ICMP_SGE, II->getArgOperand(i: 0), |
4289 | II->getArgOperand(i: 1)); |
4290 | } |
4291 | break; |
4292 | default: |
4293 | break; |
4294 | } |
4295 | |
4296 | return nullptr; |
4297 | } |
4298 | |
4299 | /// Handle icmp with constant (but not simple integer constant) RHS. |
4300 | Instruction *InstCombinerImpl::foldICmpInstWithConstantNotInt(ICmpInst &I) { |
4301 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
4302 | Constant *RHSC = dyn_cast<Constant>(Val: Op1); |
4303 | Instruction *LHSI = dyn_cast<Instruction>(Val: Op0); |
4304 | if (!RHSC || !LHSI) |
4305 | return nullptr; |
4306 | |
4307 | switch (LHSI->getOpcode()) { |
4308 | case Instruction::PHI: |
4309 | if (Instruction *NV = foldOpIntoPhi(I, PN: cast<PHINode>(Val: LHSI))) |
4310 | return NV; |
4311 | break; |
4312 | case Instruction::IntToPtr: |
4313 | // icmp pred inttoptr(X), null -> icmp pred X, 0 |
4314 | if (RHSC->isNullValue() && |
4315 | DL.getIntPtrType(RHSC->getType()) == LHSI->getOperand(i: 0)->getType()) |
4316 | return new ICmpInst( |
4317 | I.getPredicate(), LHSI->getOperand(i: 0), |
4318 | Constant::getNullValue(Ty: LHSI->getOperand(i: 0)->getType())); |
4319 | break; |
4320 | |
4321 | case Instruction::Load: |
4322 | // Try to optimize things like "A[i] > 4" to index computations. |
4323 | if (GetElementPtrInst *GEP = |
4324 | dyn_cast<GetElementPtrInst>(Val: LHSI->getOperand(i: 0))) |
4325 | if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Val: GEP->getOperand(i_nocapture: 0))) |
4326 | if (Instruction *Res = |
4327 | foldCmpLoadFromIndexedGlobal(LI: cast<LoadInst>(Val: LHSI), GEP, GV, ICI&: I)) |
4328 | return Res; |
4329 | break; |
4330 | } |
4331 | |
4332 | return nullptr; |
4333 | } |
4334 | |
4335 | Instruction *InstCombinerImpl::foldSelectICmp(CmpPredicate Pred, SelectInst *SI, |
4336 | Value *RHS, const ICmpInst &I) { |
4337 | // Try to fold the comparison into the select arms, which will cause the |
4338 | // select to be converted into a logical and/or. |
4339 | auto SimplifyOp = [&](Value *Op, bool SelectCondIsTrue) -> Value * { |
4340 | if (Value *Res = simplifyICmpInst(Pred, LHS: Op, RHS, Q: SQ)) |
4341 | return Res; |
4342 | if (std::optional<bool> Impl = isImpliedCondition( |
4343 | LHS: SI->getCondition(), RHSPred: Pred, RHSOp0: Op, RHSOp1: RHS, DL, LHSIsTrue: SelectCondIsTrue)) |
4344 | return ConstantInt::get(Ty: I.getType(), V: *Impl); |
4345 | return nullptr; |
4346 | }; |
4347 | |
4348 | ConstantInt *CI = nullptr; |
4349 | Value *Op1 = SimplifyOp(SI->getOperand(i_nocapture: 1), true); |
4350 | if (Op1) |
4351 | CI = dyn_cast<ConstantInt>(Val: Op1); |
4352 | |
4353 | Value *Op2 = SimplifyOp(SI->getOperand(i_nocapture: 2), false); |
4354 | if (Op2) |
4355 | CI = dyn_cast<ConstantInt>(Val: Op2); |
4356 | |
4357 | auto Simplifies = [&](Value *Op, unsigned Idx) { |
4358 | // A comparison of ucmp/scmp with a constant will fold into an icmp. |
4359 | const APInt *Dummy; |
4360 | return Op || |
4361 | (isa<CmpIntrinsic>(Val: SI->getOperand(i_nocapture: Idx)) && |
4362 | SI->getOperand(i_nocapture: Idx)->hasOneUse() && match(V: RHS, P: m_APInt(Res&: Dummy))); |
4363 | }; |
4364 | |
4365 | // We only want to perform this transformation if it will not lead to |
4366 | // additional code. This is true if either both sides of the select |
4367 | // fold to a constant (in which case the icmp is replaced with a select |
4368 | // which will usually simplify) or this is the only user of the |
4369 | // select (in which case we are trading a select+icmp for a simpler |
4370 | // select+icmp) or all uses of the select can be replaced based on |
4371 | // dominance information ("Global cases"). |
4372 | bool Transform = false; |
4373 | if (Op1 && Op2) |
4374 | Transform = true; |
4375 | else if (Simplifies(Op1, 1) || Simplifies(Op2, 2)) { |
4376 | // Local case |
4377 | if (SI->hasOneUse()) |
4378 | Transform = true; |
4379 | // Global cases |
4380 | else if (CI && !CI->isZero()) |
4381 | // When Op1 is constant try replacing select with second operand. |
4382 | // Otherwise Op2 is constant and try replacing select with first |
4383 | // operand. |
4384 | Transform = replacedSelectWithOperand(SI, Icmp: &I, SIOpd: Op1 ? 2 : 1); |
4385 | } |
4386 | if (Transform) { |
4387 | if (!Op1) |
4388 | Op1 = Builder.CreateICmp(P: Pred, LHS: SI->getOperand(i_nocapture: 1), RHS, Name: I.getName()); |
4389 | if (!Op2) |
4390 | Op2 = Builder.CreateICmp(P: Pred, LHS: SI->getOperand(i_nocapture: 2), RHS, Name: I.getName()); |
4391 | return SelectInst::Create(C: SI->getOperand(i_nocapture: 0), S1: Op1, S2: Op2); |
4392 | } |
4393 | |
4394 | return nullptr; |
4395 | } |
4396 | |
4397 | // Returns whether V is a Mask ((X + 1) & X == 0) or ~Mask (-Pow2OrZero) |
4398 | static bool isMaskOrZero(const Value *V, bool Not, const SimplifyQuery &Q, |
4399 | unsigned Depth = 0) { |
4400 | if (Not ? match(V, P: m_NegatedPower2OrZero()) : match(V, P: m_LowBitMaskOrZero())) |
4401 | return true; |
4402 | if (V->getType()->getScalarSizeInBits() == 1) |
4403 | return true; |
4404 | if (Depth++ >= MaxAnalysisRecursionDepth) |
4405 | return false; |
4406 | Value *X; |
4407 | const Instruction *I = dyn_cast<Instruction>(Val: V); |
4408 | if (!I) |
4409 | return false; |
4410 | switch (I->getOpcode()) { |
4411 | case Instruction::ZExt: |
4412 | // ZExt(Mask) is a Mask. |
4413 | return !Not && isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth); |
4414 | case Instruction::SExt: |
4415 | // SExt(Mask) is a Mask. |
4416 | // SExt(~Mask) is a ~Mask. |
4417 | return isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth); |
4418 | case Instruction::And: |
4419 | case Instruction::Or: |
4420 | // Mask0 | Mask1 is a Mask. |
4421 | // Mask0 & Mask1 is a Mask. |
4422 | // ~Mask0 | ~Mask1 is a ~Mask. |
4423 | // ~Mask0 & ~Mask1 is a ~Mask. |
4424 | return isMaskOrZero(V: I->getOperand(i: 1), Not, Q, Depth) && |
4425 | isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth); |
4426 | case Instruction::Xor: |
4427 | if (match(V, P: m_Not(V: m_Value(V&: X)))) |
4428 | return isMaskOrZero(V: X, Not: !Not, Q, Depth); |
4429 | |
4430 | // (X ^ -X) is a ~Mask |
4431 | if (Not) |
4432 | return match(V, P: m_c_Xor(L: m_Value(V&: X), R: m_Neg(V: m_Deferred(V: X)))); |
4433 | // (X ^ (X - 1)) is a Mask |
4434 | else |
4435 | return match(V, P: m_c_Xor(L: m_Value(V&: X), R: m_Add(L: m_Deferred(V: X), R: m_AllOnes()))); |
4436 | case Instruction::Select: |
4437 | // c ? Mask0 : Mask1 is a Mask. |
4438 | return isMaskOrZero(V: I->getOperand(i: 1), Not, Q, Depth) && |
4439 | isMaskOrZero(V: I->getOperand(i: 2), Not, Q, Depth); |
4440 | case Instruction::Shl: |
4441 | // (~Mask) << X is a ~Mask. |
4442 | return Not && isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth); |
4443 | case Instruction::LShr: |
4444 | // Mask >> X is a Mask. |
4445 | return !Not && isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth); |
4446 | case Instruction::AShr: |
4447 | // Mask s>> X is a Mask. |
4448 | // ~Mask s>> X is a ~Mask. |
4449 | return isMaskOrZero(V: I->getOperand(i: 0), Not, Q, Depth); |
4450 | case Instruction::Add: |
4451 | // Pow2 - 1 is a Mask. |
4452 | if (!Not && match(V: I->getOperand(i: 1), P: m_AllOnes())) |
4453 | return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), DL: Q.DL, /*OrZero*/ true, |
4454 | AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT, UseInstrInfo: Depth); |
4455 | break; |
4456 | case Instruction::Sub: |
4457 | // -Pow2 is a ~Mask. |
4458 | if (Not && match(V: I->getOperand(i: 0), P: m_Zero())) |
4459 | return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 1), DL: Q.DL, /*OrZero*/ true, |
4460 | AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT, UseInstrInfo: Depth); |
4461 | break; |
4462 | case Instruction::Call: { |
4463 | if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) { |
4464 | switch (II->getIntrinsicID()) { |
4465 | // min/max(Mask0, Mask1) is a Mask. |
4466 | // min/max(~Mask0, ~Mask1) is a ~Mask. |
4467 | case Intrinsic::umax: |
4468 | case Intrinsic::smax: |
4469 | case Intrinsic::umin: |
4470 | case Intrinsic::smin: |
4471 | return isMaskOrZero(V: II->getArgOperand(i: 1), Not, Q, Depth) && |
4472 | isMaskOrZero(V: II->getArgOperand(i: 0), Not, Q, Depth); |
4473 | |
4474 | // In the context of masks, bitreverse(Mask) == ~Mask |
4475 | case Intrinsic::bitreverse: |
4476 | return isMaskOrZero(V: II->getArgOperand(i: 0), Not: !Not, Q, Depth); |
4477 | default: |
4478 | break; |
4479 | } |
4480 | } |
4481 | break; |
4482 | } |
4483 | default: |
4484 | break; |
4485 | } |
4486 | return false; |
4487 | } |
4488 | |
4489 | /// Some comparisons can be simplified. |
4490 | /// In this case, we are looking for comparisons that look like |
4491 | /// a check for a lossy truncation. |
4492 | /// Folds: |
4493 | /// icmp SrcPred (x & Mask), x to icmp DstPred x, Mask |
4494 | /// icmp SrcPred (x & ~Mask), ~Mask to icmp DstPred x, ~Mask |
4495 | /// icmp eq/ne (x & ~Mask), 0 to icmp DstPred x, Mask |
4496 | /// icmp eq/ne (~x | Mask), -1 to icmp DstPred x, Mask |
4497 | /// Where Mask is some pattern that produces all-ones in low bits: |
4498 | /// (-1 >> y) |
4499 | /// ((-1 << y) >> y) <- non-canonical, has extra uses |
4500 | /// ~(-1 << y) |
4501 | /// ((1 << y) + (-1)) <- non-canonical, has extra uses |
4502 | /// The Mask can be a constant, too. |
4503 | /// For some predicates, the operands are commutative. |
4504 | /// For others, x can only be on a specific side. |
4505 | static Value *foldICmpWithLowBitMaskedVal(CmpPredicate Pred, Value *Op0, |
4506 | Value *Op1, const SimplifyQuery &Q, |
4507 | InstCombiner &IC) { |
4508 | |
4509 | ICmpInst::Predicate DstPred; |
4510 | switch (Pred) { |
4511 | case ICmpInst::Predicate::ICMP_EQ: |
4512 | // x & Mask == x |
4513 | // x & ~Mask == 0 |
4514 | // ~x | Mask == -1 |
4515 | // -> x u<= Mask |
4516 | // x & ~Mask == ~Mask |
4517 | // -> ~Mask u<= x |
4518 | DstPred = ICmpInst::Predicate::ICMP_ULE; |
4519 | break; |
4520 | case ICmpInst::Predicate::ICMP_NE: |
4521 | // x & Mask != x |
4522 | // x & ~Mask != 0 |
4523 | // ~x | Mask != -1 |
4524 | // -> x u> Mask |
4525 | // x & ~Mask != ~Mask |
4526 | // -> ~Mask u> x |
4527 | DstPred = ICmpInst::Predicate::ICMP_UGT; |
4528 | break; |
4529 | case ICmpInst::Predicate::ICMP_ULT: |
4530 | // x & Mask u< x |
4531 | // -> x u> Mask |
4532 | // x & ~Mask u< ~Mask |
4533 | // -> ~Mask u> x |
4534 | DstPred = ICmpInst::Predicate::ICMP_UGT; |
4535 | break; |
4536 | case ICmpInst::Predicate::ICMP_UGE: |
4537 | // x & Mask u>= x |
4538 | // -> x u<= Mask |
4539 | // x & ~Mask u>= ~Mask |
4540 | // -> ~Mask u<= x |
4541 | DstPred = ICmpInst::Predicate::ICMP_ULE; |
4542 | break; |
4543 | case ICmpInst::Predicate::ICMP_SLT: |
4544 | // x & Mask s< x [iff Mask s>= 0] |
4545 | // -> x s> Mask |
4546 | // x & ~Mask s< ~Mask [iff ~Mask != 0] |
4547 | // -> ~Mask s> x |
4548 | DstPred = ICmpInst::Predicate::ICMP_SGT; |
4549 | break; |
4550 | case ICmpInst::Predicate::ICMP_SGE: |
4551 | // x & Mask s>= x [iff Mask s>= 0] |
4552 | // -> x s<= Mask |
4553 | // x & ~Mask s>= ~Mask [iff ~Mask != 0] |
4554 | // -> ~Mask s<= x |
4555 | DstPred = ICmpInst::Predicate::ICMP_SLE; |
4556 | break; |
4557 | default: |
4558 | // We don't support sgt,sle |
4559 | // ult/ugt are simplified to true/false respectively. |
4560 | return nullptr; |
4561 | } |
4562 | |
4563 | Value *X, *M; |
4564 | // Put search code in lambda for early positive returns. |
4565 | auto IsLowBitMask = [&]() { |
4566 | if (match(V: Op0, P: m_c_And(L: m_Specific(V: Op1), R: m_Value(V&: M)))) { |
4567 | X = Op1; |
4568 | // Look for: x & Mask pred x |
4569 | if (isMaskOrZero(V: M, /*Not=*/false, Q)) { |
4570 | return !ICmpInst::isSigned(predicate: Pred) || |
4571 | (match(V: M, P: m_NonNegative()) || isKnownNonNegative(V: M, SQ: Q)); |
4572 | } |
4573 | |
4574 | // Look for: x & ~Mask pred ~Mask |
4575 | if (isMaskOrZero(V: X, /*Not=*/true, Q)) { |
4576 | return !ICmpInst::isSigned(predicate: Pred) || isKnownNonZero(V: X, Q); |
4577 | } |
4578 | return false; |
4579 | } |
4580 | if (ICmpInst::isEquality(P: Pred) && match(V: Op1, P: m_AllOnes()) && |
4581 | match(V: Op0, P: m_OneUse(SubPattern: m_Or(L: m_Value(V&: X), R: m_Value(V&: M))))) { |
4582 | |
4583 | auto Check = [&]() { |
4584 | // Look for: ~x | Mask == -1 |
4585 | if (isMaskOrZero(V: M, /*Not=*/false, Q)) { |
4586 | if (Value *NotX = |
4587 | IC.getFreelyInverted(V: X, WillInvertAllUses: X->hasOneUse(), Builder: &IC.Builder)) { |
4588 | X = NotX; |
4589 | return true; |
4590 | } |
4591 | } |
4592 | return false; |
4593 | }; |
4594 | if (Check()) |
4595 | return true; |
4596 | std::swap(a&: X, b&: M); |
4597 | return Check(); |
4598 | } |
4599 | if (ICmpInst::isEquality(P: Pred) && match(V: Op1, P: m_Zero()) && |
4600 | match(V: Op0, P: m_OneUse(SubPattern: m_And(L: m_Value(V&: X), R: m_Value(V&: M))))) { |
4601 | auto Check = [&]() { |
4602 | // Look for: x & ~Mask == 0 |
4603 | if (isMaskOrZero(V: M, /*Not=*/true, Q)) { |
4604 | if (Value *NotM = |
4605 | IC.getFreelyInverted(V: M, WillInvertAllUses: M->hasOneUse(), Builder: &IC.Builder)) { |
4606 | M = NotM; |
4607 | return true; |
4608 | } |
4609 | } |
4610 | return false; |
4611 | }; |
4612 | if (Check()) |
4613 | return true; |
4614 | std::swap(a&: X, b&: M); |
4615 | return Check(); |
4616 | } |
4617 | return false; |
4618 | }; |
4619 | |
4620 | if (!IsLowBitMask()) |
4621 | return nullptr; |
4622 | |
4623 | return IC.Builder.CreateICmp(P: DstPred, LHS: X, RHS: M); |
4624 | } |
4625 | |
4626 | /// Some comparisons can be simplified. |
4627 | /// In this case, we are looking for comparisons that look like |
4628 | /// a check for a lossy signed truncation. |
4629 | /// Folds: (MaskedBits is a constant.) |
4630 | /// ((%x << MaskedBits) a>> MaskedBits) SrcPred %x |
4631 | /// Into: |
4632 | /// (add %x, (1 << (KeptBits-1))) DstPred (1 << KeptBits) |
4633 | /// Where KeptBits = bitwidth(%x) - MaskedBits |
4634 | static Value * |
4635 | foldICmpWithTruncSignExtendedVal(ICmpInst &I, |
4636 | InstCombiner::BuilderTy &Builder) { |
4637 | CmpPredicate SrcPred; |
4638 | Value *X; |
4639 | const APInt *C0, *C1; // FIXME: non-splats, potentially with undef. |
4640 | // We are ok with 'shl' having multiple uses, but 'ashr' must be one-use. |
4641 | if (!match(V: &I, P: m_c_ICmp(Pred&: SrcPred, |
4642 | L: m_OneUse(SubPattern: m_AShr(L: m_Shl(L: m_Value(V&: X), R: m_APInt(Res&: C0)), |
4643 | R: m_APInt(Res&: C1))), |
4644 | R: m_Deferred(V: X)))) |
4645 | return nullptr; |
4646 | |
4647 | // Potential handling of non-splats: for each element: |
4648 | // * if both are undef, replace with constant 0. |
4649 | // Because (1<<0) is OK and is 1, and ((1<<0)>>1) is also OK and is 0. |
4650 | // * if both are not undef, and are different, bailout. |
4651 | // * else, only one is undef, then pick the non-undef one. |
4652 | |
4653 | // The shift amount must be equal. |
4654 | if (*C0 != *C1) |
4655 | return nullptr; |
4656 | const APInt &MaskedBits = *C0; |
4657 | assert(MaskedBits != 0 && "shift by zero should be folded away already." ); |
4658 | |
4659 | ICmpInst::Predicate DstPred; |
4660 | switch (SrcPred) { |
4661 | case ICmpInst::Predicate::ICMP_EQ: |
4662 | // ((%x << MaskedBits) a>> MaskedBits) == %x |
4663 | // => |
4664 | // (add %x, (1 << (KeptBits-1))) u< (1 << KeptBits) |
4665 | DstPred = ICmpInst::Predicate::ICMP_ULT; |
4666 | break; |
4667 | case ICmpInst::Predicate::ICMP_NE: |
4668 | // ((%x << MaskedBits) a>> MaskedBits) != %x |
4669 | // => |
4670 | // (add %x, (1 << (KeptBits-1))) u>= (1 << KeptBits) |
4671 | DstPred = ICmpInst::Predicate::ICMP_UGE; |
4672 | break; |
4673 | // FIXME: are more folds possible? |
4674 | default: |
4675 | return nullptr; |
4676 | } |
4677 | |
4678 | auto *XType = X->getType(); |
4679 | const unsigned XBitWidth = XType->getScalarSizeInBits(); |
4680 | const APInt BitWidth = APInt(XBitWidth, XBitWidth); |
4681 | assert(BitWidth.ugt(MaskedBits) && "shifts should leave some bits untouched" ); |
4682 | |
4683 | // KeptBits = bitwidth(%x) - MaskedBits |
4684 | const APInt KeptBits = BitWidth - MaskedBits; |
4685 | assert(KeptBits.ugt(0) && KeptBits.ult(BitWidth) && "unreachable" ); |
4686 | // ICmpCst = (1 << KeptBits) |
4687 | const APInt ICmpCst = APInt(XBitWidth, 1).shl(ShiftAmt: KeptBits); |
4688 | assert(ICmpCst.isPowerOf2()); |
4689 | // AddCst = (1 << (KeptBits-1)) |
4690 | const APInt AddCst = ICmpCst.lshr(shiftAmt: 1); |
4691 | assert(AddCst.ult(ICmpCst) && AddCst.isPowerOf2()); |
4692 | |
4693 | // T0 = add %x, AddCst |
4694 | Value *T0 = Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty: XType, V: AddCst)); |
4695 | // T1 = T0 DstPred ICmpCst |
4696 | Value *T1 = Builder.CreateICmp(P: DstPred, LHS: T0, RHS: ConstantInt::get(Ty: XType, V: ICmpCst)); |
4697 | |
4698 | return T1; |
4699 | } |
4700 | |
4701 | // Given pattern: |
4702 | // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0 |
4703 | // we should move shifts to the same hand of 'and', i.e. rewrite as |
4704 | // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x) |
4705 | // We are only interested in opposite logical shifts here. |
4706 | // One of the shifts can be truncated. |
4707 | // If we can, we want to end up creating 'lshr' shift. |
4708 | static Value * |
4709 | foldShiftIntoShiftInAnotherHandOfAndInICmp(ICmpInst &I, const SimplifyQuery SQ, |
4710 | InstCombiner::BuilderTy &Builder) { |
4711 | if (!I.isEquality() || !match(V: I.getOperand(i_nocapture: 1), P: m_Zero()) || |
4712 | !I.getOperand(i_nocapture: 0)->hasOneUse()) |
4713 | return nullptr; |
4714 | |
4715 | auto m_AnyLogicalShift = m_LogicalShift(L: m_Value(), R: m_Value()); |
4716 | |
4717 | // Look for an 'and' of two logical shifts, one of which may be truncated. |
4718 | // We use m_TruncOrSelf() on the RHS to correctly handle commutative case. |
4719 | Instruction *XShift, *MaybeTruncation, *YShift; |
4720 | if (!match( |
4721 | V: I.getOperand(i_nocapture: 0), |
4722 | P: m_c_And(L: m_CombineAnd(L: m_AnyLogicalShift, R: m_Instruction(I&: XShift)), |
4723 | R: m_CombineAnd(L: m_TruncOrSelf(Op: m_CombineAnd( |
4724 | L: m_AnyLogicalShift, R: m_Instruction(I&: YShift))), |
4725 | R: m_Instruction(I&: MaybeTruncation))))) |
4726 | return nullptr; |
4727 | |
4728 | // We potentially looked past 'trunc', but only when matching YShift, |
4729 | // therefore YShift must have the widest type. |
4730 | Instruction *WidestShift = YShift; |
4731 | // Therefore XShift must have the shallowest type. |
4732 | // Or they both have identical types if there was no truncation. |
4733 | Instruction *NarrowestShift = XShift; |
4734 | |
4735 | Type *WidestTy = WidestShift->getType(); |
4736 | Type *NarrowestTy = NarrowestShift->getType(); |
4737 | assert(NarrowestTy == I.getOperand(0)->getType() && |
4738 | "We did not look past any shifts while matching XShift though." ); |
4739 | bool HadTrunc = WidestTy != I.getOperand(i_nocapture: 0)->getType(); |
4740 | |
4741 | // If YShift is a 'lshr', swap the shifts around. |
4742 | if (match(V: YShift, P: m_LShr(L: m_Value(), R: m_Value()))) |
4743 | std::swap(a&: XShift, b&: YShift); |
4744 | |
4745 | // The shifts must be in opposite directions. |
4746 | auto XShiftOpcode = XShift->getOpcode(); |
4747 | if (XShiftOpcode == YShift->getOpcode()) |
4748 | return nullptr; // Do not care about same-direction shifts here. |
4749 | |
4750 | Value *X, *XShAmt, *Y, *YShAmt; |
4751 | match(V: XShift, P: m_BinOp(L: m_Value(V&: X), R: m_ZExtOrSelf(Op: m_Value(V&: XShAmt)))); |
4752 | match(V: YShift, P: m_BinOp(L: m_Value(V&: Y), R: m_ZExtOrSelf(Op: m_Value(V&: YShAmt)))); |
4753 | |
4754 | // If one of the values being shifted is a constant, then we will end with |
4755 | // and+icmp, and [zext+]shift instrs will be constant-folded. If they are not, |
4756 | // however, we will need to ensure that we won't increase instruction count. |
4757 | if (!isa<Constant>(Val: X) && !isa<Constant>(Val: Y)) { |
4758 | // At least one of the hands of the 'and' should be one-use shift. |
4759 | if (!match(V: I.getOperand(i_nocapture: 0), |
4760 | P: m_c_And(L: m_OneUse(SubPattern: m_AnyLogicalShift), R: m_Value()))) |
4761 | return nullptr; |
4762 | if (HadTrunc) { |
4763 | // Due to the 'trunc', we will need to widen X. For that either the old |
4764 | // 'trunc' or the shift amt in the non-truncated shift should be one-use. |
4765 | if (!MaybeTruncation->hasOneUse() && |
4766 | !NarrowestShift->getOperand(i: 1)->hasOneUse()) |
4767 | return nullptr; |
4768 | } |
4769 | } |
4770 | |
4771 | // We have two shift amounts from two different shifts. The types of those |
4772 | // shift amounts may not match. If that's the case let's bailout now. |
4773 | if (XShAmt->getType() != YShAmt->getType()) |
4774 | return nullptr; |
4775 | |
4776 | // As input, we have the following pattern: |
4777 | // icmp eq/ne (and ((x shift Q), (y oppositeshift K))), 0 |
4778 | // We want to rewrite that as: |
4779 | // icmp eq/ne (and (x shift (Q+K)), y), 0 iff (Q+K) u< bitwidth(x) |
4780 | // While we know that originally (Q+K) would not overflow |
4781 | // (because 2 * (N-1) u<= iN -1), we have looked past extensions of |
4782 | // shift amounts. so it may now overflow in smaller bitwidth. |
4783 | // To ensure that does not happen, we need to ensure that the total maximal |
4784 | // shift amount is still representable in that smaller bit width. |
4785 | unsigned MaximalPossibleTotalShiftAmount = |
4786 | (WidestTy->getScalarSizeInBits() - 1) + |
4787 | (NarrowestTy->getScalarSizeInBits() - 1); |
4788 | APInt MaximalRepresentableShiftAmount = |
4789 | APInt::getAllOnes(numBits: XShAmt->getType()->getScalarSizeInBits()); |
4790 | if (MaximalRepresentableShiftAmount.ult(RHS: MaximalPossibleTotalShiftAmount)) |
4791 | return nullptr; |
4792 | |
4793 | // Can we fold (XShAmt+YShAmt) ? |
4794 | auto *NewShAmt = dyn_cast_or_null<Constant>( |
4795 | Val: simplifyAddInst(LHS: XShAmt, RHS: YShAmt, /*isNSW=*/IsNSW: false, |
4796 | /*isNUW=*/IsNUW: false, Q: SQ.getWithInstruction(I: &I))); |
4797 | if (!NewShAmt) |
4798 | return nullptr; |
4799 | if (NewShAmt->getType() != WidestTy) { |
4800 | NewShAmt = |
4801 | ConstantFoldCastOperand(Opcode: Instruction::ZExt, C: NewShAmt, DestTy: WidestTy, DL: SQ.DL); |
4802 | if (!NewShAmt) |
4803 | return nullptr; |
4804 | } |
4805 | unsigned WidestBitWidth = WidestTy->getScalarSizeInBits(); |
4806 | |
4807 | // Is the new shift amount smaller than the bit width? |
4808 | // FIXME: could also rely on ConstantRange. |
4809 | if (!match(V: NewShAmt, |
4810 | P: m_SpecificInt_ICMP(Predicate: ICmpInst::Predicate::ICMP_ULT, |
4811 | Threshold: APInt(WidestBitWidth, WidestBitWidth)))) |
4812 | return nullptr; |
4813 | |
4814 | // An extra legality check is needed if we had trunc-of-lshr. |
4815 | if (HadTrunc && match(V: WidestShift, P: m_LShr(L: m_Value(), R: m_Value()))) { |
4816 | auto CanFold = [NewShAmt, WidestBitWidth, NarrowestShift, SQ, |
4817 | WidestShift]() { |
4818 | // It isn't obvious whether it's worth it to analyze non-constants here. |
4819 | // Also, let's basically give up on non-splat cases, pessimizing vectors. |
4820 | // If *any* of these preconditions matches we can perform the fold. |
4821 | Constant *NewShAmtSplat = NewShAmt->getType()->isVectorTy() |
4822 | ? NewShAmt->getSplatValue() |
4823 | : NewShAmt; |
4824 | // If it's edge-case shift (by 0 or by WidestBitWidth-1) we can fold. |
4825 | if (NewShAmtSplat && |
4826 | (NewShAmtSplat->isNullValue() || |
4827 | NewShAmtSplat->getUniqueInteger() == WidestBitWidth - 1)) |
4828 | return true; |
4829 | // We consider *min* leading zeros so a single outlier |
4830 | // blocks the transform as opposed to allowing it. |
4831 | if (auto *C = dyn_cast<Constant>(Val: NarrowestShift->getOperand(i: 0))) { |
4832 | KnownBits Known = computeKnownBits(V: C, DL: SQ.DL); |
4833 | unsigned MinLeadZero = Known.countMinLeadingZeros(); |
4834 | // If the value being shifted has at most lowest bit set we can fold. |
4835 | unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero; |
4836 | if (MaxActiveBits <= 1) |
4837 | return true; |
4838 | // Precondition: NewShAmt u<= countLeadingZeros(C) |
4839 | if (NewShAmtSplat && NewShAmtSplat->getUniqueInteger().ule(RHS: MinLeadZero)) |
4840 | return true; |
4841 | } |
4842 | if (auto *C = dyn_cast<Constant>(Val: WidestShift->getOperand(i: 0))) { |
4843 | KnownBits Known = computeKnownBits(V: C, DL: SQ.DL); |
4844 | unsigned MinLeadZero = Known.countMinLeadingZeros(); |
4845 | // If the value being shifted has at most lowest bit set we can fold. |
4846 | unsigned MaxActiveBits = Known.getBitWidth() - MinLeadZero; |
4847 | if (MaxActiveBits <= 1) |
4848 | return true; |
4849 | // Precondition: ((WidestBitWidth-1)-NewShAmt) u<= countLeadingZeros(C) |
4850 | if (NewShAmtSplat) { |
4851 | APInt AdjNewShAmt = |
4852 | (WidestBitWidth - 1) - NewShAmtSplat->getUniqueInteger(); |
4853 | if (AdjNewShAmt.ule(RHS: MinLeadZero)) |
4854 | return true; |
4855 | } |
4856 | } |
4857 | return false; // Can't tell if it's ok. |
4858 | }; |
4859 | if (!CanFold()) |
4860 | return nullptr; |
4861 | } |
4862 | |
4863 | // All good, we can do this fold. |
4864 | X = Builder.CreateZExt(V: X, DestTy: WidestTy); |
4865 | Y = Builder.CreateZExt(V: Y, DestTy: WidestTy); |
4866 | // The shift is the same that was for X. |
4867 | Value *T0 = XShiftOpcode == Instruction::BinaryOps::LShr |
4868 | ? Builder.CreateLShr(LHS: X, RHS: NewShAmt) |
4869 | : Builder.CreateShl(LHS: X, RHS: NewShAmt); |
4870 | Value *T1 = Builder.CreateAnd(LHS: T0, RHS: Y); |
4871 | return Builder.CreateICmp(P: I.getPredicate(), LHS: T1, |
4872 | RHS: Constant::getNullValue(Ty: WidestTy)); |
4873 | } |
4874 | |
4875 | /// Fold |
4876 | /// (-1 u/ x) u< y |
4877 | /// ((x * y) ?/ x) != y |
4878 | /// to |
4879 | /// @llvm.?mul.with.overflow(x, y) plus extraction of overflow bit |
4880 | /// Note that the comparison is commutative, while inverted (u>=, ==) predicate |
4881 | /// will mean that we are looking for the opposite answer. |
4882 | Value *InstCombinerImpl::foldMultiplicationOverflowCheck(ICmpInst &I) { |
4883 | CmpPredicate Pred; |
4884 | Value *X, *Y; |
4885 | Instruction *Mul; |
4886 | Instruction *Div; |
4887 | bool NeedNegation; |
4888 | // Look for: (-1 u/ x) u</u>= y |
4889 | if (!I.isEquality() && |
4890 | match(V: &I, P: m_c_ICmp(Pred, |
4891 | L: m_CombineAnd(L: m_OneUse(SubPattern: m_UDiv(L: m_AllOnes(), R: m_Value(V&: X))), |
4892 | R: m_Instruction(I&: Div)), |
4893 | R: m_Value(V&: Y)))) { |
4894 | Mul = nullptr; |
4895 | |
4896 | // Are we checking that overflow does not happen, or does happen? |
4897 | switch (Pred) { |
4898 | case ICmpInst::Predicate::ICMP_ULT: |
4899 | NeedNegation = false; |
4900 | break; // OK |
4901 | case ICmpInst::Predicate::ICMP_UGE: |
4902 | NeedNegation = true; |
4903 | break; // OK |
4904 | default: |
4905 | return nullptr; // Wrong predicate. |
4906 | } |
4907 | } else // Look for: ((x * y) / x) !=/== y |
4908 | if (I.isEquality() && |
4909 | match(V: &I, P: m_c_ICmp(Pred, L: m_Value(V&: Y), |
4910 | R: m_CombineAnd(L: m_OneUse(SubPattern: m_IDiv( |
4911 | L: m_CombineAnd(L: m_c_Mul(L: m_Deferred(V: Y), |
4912 | R: m_Value(V&: X)), |
4913 | R: m_Instruction(I&: Mul)), |
4914 | R: m_Deferred(V: X))), |
4915 | R: m_Instruction(I&: Div))))) { |
4916 | NeedNegation = Pred == ICmpInst::Predicate::ICMP_EQ; |
4917 | } else |
4918 | return nullptr; |
4919 | |
4920 | BuilderTy::InsertPointGuard Guard(Builder); |
4921 | // If the pattern included (x * y), we'll want to insert new instructions |
4922 | // right before that original multiplication so that we can replace it. |
4923 | bool MulHadOtherUses = Mul && !Mul->hasOneUse(); |
4924 | if (MulHadOtherUses) |
4925 | Builder.SetInsertPoint(Mul); |
4926 | |
4927 | CallInst *Call = Builder.CreateIntrinsic( |
4928 | ID: Div->getOpcode() == Instruction::UDiv ? Intrinsic::umul_with_overflow |
4929 | : Intrinsic::smul_with_overflow, |
4930 | Types: X->getType(), Args: {X, Y}, /*FMFSource=*/nullptr, Name: "mul" ); |
4931 | |
4932 | // If the multiplication was used elsewhere, to ensure that we don't leave |
4933 | // "duplicate" instructions, replace uses of that original multiplication |
4934 | // with the multiplication result from the with.overflow intrinsic. |
4935 | if (MulHadOtherUses) |
4936 | replaceInstUsesWith(I&: *Mul, V: Builder.CreateExtractValue(Agg: Call, Idxs: 0, Name: "mul.val" )); |
4937 | |
4938 | Value *Res = Builder.CreateExtractValue(Agg: Call, Idxs: 1, Name: "mul.ov" ); |
4939 | if (NeedNegation) // This technically increases instruction count. |
4940 | Res = Builder.CreateNot(V: Res, Name: "mul.not.ov" ); |
4941 | |
4942 | // If we replaced the mul, erase it. Do this after all uses of Builder, |
4943 | // as the mul is used as insertion point. |
4944 | if (MulHadOtherUses) |
4945 | eraseInstFromFunction(I&: *Mul); |
4946 | |
4947 | return Res; |
4948 | } |
4949 | |
4950 | static Instruction *foldICmpXNegX(ICmpInst &I, |
4951 | InstCombiner::BuilderTy &Builder) { |
4952 | CmpPredicate Pred; |
4953 | Value *X; |
4954 | if (match(V: &I, P: m_c_ICmp(Pred, L: m_NSWNeg(V: m_Value(V&: X)), R: m_Deferred(V: X)))) { |
4955 | |
4956 | if (ICmpInst::isSigned(predicate: Pred)) |
4957 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
4958 | else if (ICmpInst::isUnsigned(predicate: Pred)) |
4959 | Pred = ICmpInst::getSignedPredicate(Pred); |
4960 | // else for equality-comparisons just keep the predicate. |
4961 | |
4962 | return ICmpInst::Create(Op: Instruction::ICmp, Pred, S1: X, |
4963 | S2: Constant::getNullValue(Ty: X->getType()), Name: I.getName()); |
4964 | } |
4965 | |
4966 | // A value is not equal to its negation unless that value is 0 or |
4967 | // MinSignedValue, ie: a != -a --> (a & MaxSignedVal) != 0 |
4968 | if (match(V: &I, P: m_c_ICmp(Pred, L: m_OneUse(SubPattern: m_Neg(V: m_Value(V&: X))), R: m_Deferred(V: X))) && |
4969 | ICmpInst::isEquality(P: Pred)) { |
4970 | Type *Ty = X->getType(); |
4971 | uint32_t BitWidth = Ty->getScalarSizeInBits(); |
4972 | Constant *MaxSignedVal = |
4973 | ConstantInt::get(Ty, V: APInt::getSignedMaxValue(numBits: BitWidth)); |
4974 | Value *And = Builder.CreateAnd(LHS: X, RHS: MaxSignedVal); |
4975 | Constant *Zero = Constant::getNullValue(Ty); |
4976 | return CmpInst::Create(Op: Instruction::ICmp, Pred, S1: And, S2: Zero); |
4977 | } |
4978 | |
4979 | return nullptr; |
4980 | } |
4981 | |
4982 | static Instruction *foldICmpAndXX(ICmpInst &I, const SimplifyQuery &Q, |
4983 | InstCombinerImpl &IC) { |
4984 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1), *A; |
4985 | // Normalize and operand as operand 0. |
4986 | CmpInst::Predicate Pred = I.getPredicate(); |
4987 | if (match(V: Op1, P: m_c_And(L: m_Specific(V: Op0), R: m_Value()))) { |
4988 | std::swap(a&: Op0, b&: Op1); |
4989 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
4990 | } |
4991 | |
4992 | if (!match(V: Op0, P: m_c_And(L: m_Specific(V: Op1), R: m_Value(V&: A)))) |
4993 | return nullptr; |
4994 | |
4995 | // (icmp (X & Y) u< X --> (X & Y) != X |
4996 | if (Pred == ICmpInst::ICMP_ULT) |
4997 | return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); |
4998 | |
4999 | // (icmp (X & Y) u>= X --> (X & Y) == X |
5000 | if (Pred == ICmpInst::ICMP_UGE) |
5001 | return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1); |
5002 | |
5003 | if (ICmpInst::isEquality(P: Pred) && Op0->hasOneUse()) { |
5004 | // icmp (X & Y) eq/ne Y --> (X | ~Y) eq/ne -1 if Y is freely invertible and |
5005 | // Y is non-constant. If Y is constant the `X & C == C` form is preferable |
5006 | // so don't do this fold. |
5007 | if (!match(V: Op1, P: m_ImmConstant())) |
5008 | if (auto *NotOp1 = |
5009 | IC.getFreelyInverted(V: Op1, WillInvertAllUses: !Op1->hasNUsesOrMore(N: 3), Builder: &IC.Builder)) |
5010 | return new ICmpInst(Pred, IC.Builder.CreateOr(LHS: A, RHS: NotOp1), |
5011 | Constant::getAllOnesValue(Ty: Op1->getType())); |
5012 | // icmp (X & Y) eq/ne Y --> (~X & Y) eq/ne 0 if X is freely invertible. |
5013 | if (auto *NotA = IC.getFreelyInverted(V: A, WillInvertAllUses: A->hasOneUse(), Builder: &IC.Builder)) |
5014 | return new ICmpInst(Pred, IC.Builder.CreateAnd(LHS: Op1, RHS: NotA), |
5015 | Constant::getNullValue(Ty: Op1->getType())); |
5016 | } |
5017 | |
5018 | if (!ICmpInst::isSigned(predicate: Pred)) |
5019 | return nullptr; |
5020 | |
5021 | KnownBits KnownY = IC.computeKnownBits(V: A, CxtI: &I); |
5022 | // (X & NegY) spred X --> (X & NegY) upred X |
5023 | if (KnownY.isNegative()) |
5024 | return new ICmpInst(ICmpInst::getUnsignedPredicate(Pred), Op0, Op1); |
5025 | |
5026 | if (Pred != ICmpInst::ICMP_SLE && Pred != ICmpInst::ICMP_SGT) |
5027 | return nullptr; |
5028 | |
5029 | if (KnownY.isNonNegative()) |
5030 | // (X & PosY) s<= X --> X s>= 0 |
5031 | // (X & PosY) s> X --> X s< 0 |
5032 | return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Op1, |
5033 | Constant::getNullValue(Ty: Op1->getType())); |
5034 | |
5035 | if (isKnownNegative(V: Op1, SQ: IC.getSimplifyQuery().getWithInstruction(I: &I))) |
5036 | // (NegX & Y) s<= NegX --> Y s< 0 |
5037 | // (NegX & Y) s> NegX --> Y s>= 0 |
5038 | return new ICmpInst(ICmpInst::getFlippedStrictnessPredicate(pred: Pred), A, |
5039 | Constant::getNullValue(Ty: A->getType())); |
5040 | |
5041 | return nullptr; |
5042 | } |
5043 | |
5044 | static Instruction *foldICmpOrXX(ICmpInst &I, const SimplifyQuery &Q, |
5045 | InstCombinerImpl &IC) { |
5046 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1), *A; |
5047 | |
5048 | // Normalize or operand as operand 0. |
5049 | CmpInst::Predicate Pred = I.getPredicate(); |
5050 | if (match(V: Op1, P: m_c_Or(L: m_Specific(V: Op0), R: m_Value(V&: A)))) { |
5051 | std::swap(a&: Op0, b&: Op1); |
5052 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
5053 | } else if (!match(V: Op0, P: m_c_Or(L: m_Specific(V: Op1), R: m_Value(V&: A)))) { |
5054 | return nullptr; |
5055 | } |
5056 | |
5057 | // icmp (X | Y) u<= X --> (X | Y) == X |
5058 | if (Pred == ICmpInst::ICMP_ULE) |
5059 | return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1); |
5060 | |
5061 | // icmp (X | Y) u> X --> (X | Y) != X |
5062 | if (Pred == ICmpInst::ICMP_UGT) |
5063 | return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); |
5064 | |
5065 | if (ICmpInst::isEquality(P: Pred) && Op0->hasOneUse()) { |
5066 | // icmp (X | Y) eq/ne Y --> (X & ~Y) eq/ne 0 if Y is freely invertible |
5067 | if (Value *NotOp1 = IC.getFreelyInverted( |
5068 | V: Op1, WillInvertAllUses: !isa<Constant>(Val: Op1) && !Op1->hasNUsesOrMore(N: 3), Builder: &IC.Builder)) |
5069 | return new ICmpInst(Pred, IC.Builder.CreateAnd(LHS: A, RHS: NotOp1), |
5070 | Constant::getNullValue(Ty: Op1->getType())); |
5071 | // icmp (X | Y) eq/ne Y --> (~X | Y) eq/ne -1 if X is freely invertible. |
5072 | if (Value *NotA = IC.getFreelyInverted(V: A, WillInvertAllUses: A->hasOneUse(), Builder: &IC.Builder)) |
5073 | return new ICmpInst(Pred, IC.Builder.CreateOr(LHS: Op1, RHS: NotA), |
5074 | Constant::getAllOnesValue(Ty: Op1->getType())); |
5075 | } |
5076 | return nullptr; |
5077 | } |
5078 | |
5079 | static Instruction *foldICmpXorXX(ICmpInst &I, const SimplifyQuery &Q, |
5080 | InstCombinerImpl &IC) { |
5081 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1), *A; |
5082 | // Normalize xor operand as operand 0. |
5083 | CmpInst::Predicate Pred = I.getPredicate(); |
5084 | if (match(V: Op1, P: m_c_Xor(L: m_Specific(V: Op0), R: m_Value()))) { |
5085 | std::swap(a&: Op0, b&: Op1); |
5086 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
5087 | } |
5088 | if (!match(V: Op0, P: m_c_Xor(L: m_Specific(V: Op1), R: m_Value(V&: A)))) |
5089 | return nullptr; |
5090 | |
5091 | // icmp (X ^ Y_NonZero) u>= X --> icmp (X ^ Y_NonZero) u> X |
5092 | // icmp (X ^ Y_NonZero) u<= X --> icmp (X ^ Y_NonZero) u< X |
5093 | // icmp (X ^ Y_NonZero) s>= X --> icmp (X ^ Y_NonZero) s> X |
5094 | // icmp (X ^ Y_NonZero) s<= X --> icmp (X ^ Y_NonZero) s< X |
5095 | CmpInst::Predicate PredOut = CmpInst::getStrictPredicate(pred: Pred); |
5096 | if (PredOut != Pred && isKnownNonZero(V: A, Q)) |
5097 | return new ICmpInst(PredOut, Op0, Op1); |
5098 | |
5099 | // These transform work when A is negative. |
5100 | // X s< X^A, X s<= X^A, X u> X^A, X u>= X^A --> X s< 0 |
5101 | // X s> X^A, X s>= X^A, X u< X^A, X u<= X^A --> X s>= 0 |
5102 | if (match(V: A, P: m_Negative())) { |
5103 | CmpInst::Predicate NewPred; |
5104 | switch (ICmpInst::getStrictPredicate(pred: Pred)) { |
5105 | default: |
5106 | return nullptr; |
5107 | case ICmpInst::ICMP_SLT: |
5108 | case ICmpInst::ICMP_UGT: |
5109 | NewPred = ICmpInst::ICMP_SLT; |
5110 | break; |
5111 | case ICmpInst::ICMP_SGT: |
5112 | case ICmpInst::ICMP_ULT: |
5113 | NewPred = ICmpInst::ICMP_SGE; |
5114 | break; |
5115 | } |
5116 | Constant *Const = Constant::getNullValue(Ty: Op0->getType()); |
5117 | return new ICmpInst(NewPred, Op0, Const); |
5118 | } |
5119 | |
5120 | return nullptr; |
5121 | } |
5122 | |
5123 | /// Try to fold icmp (binop), X or icmp X, (binop). |
5124 | /// TODO: A large part of this logic is duplicated in InstSimplify's |
5125 | /// simplifyICmpWithBinOp(). We should be able to share that and avoid the code |
5126 | /// duplication. |
5127 | Instruction *InstCombinerImpl::foldICmpBinOp(ICmpInst &I, |
5128 | const SimplifyQuery &SQ) { |
5129 | const SimplifyQuery Q = SQ.getWithInstruction(I: &I); |
5130 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
5131 | |
5132 | // Special logic for binary operators. |
5133 | BinaryOperator *BO0 = dyn_cast<BinaryOperator>(Val: Op0); |
5134 | BinaryOperator *BO1 = dyn_cast<BinaryOperator>(Val: Op1); |
5135 | if (!BO0 && !BO1) |
5136 | return nullptr; |
5137 | |
5138 | if (Instruction *NewICmp = foldICmpXNegX(I, Builder)) |
5139 | return NewICmp; |
5140 | |
5141 | const CmpInst::Predicate Pred = I.getPredicate(); |
5142 | Value *X; |
5143 | |
5144 | // Convert add-with-unsigned-overflow comparisons into a 'not' with compare. |
5145 | // (Op1 + X) u</u>= Op1 --> ~Op1 u</u>= X |
5146 | if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Add(L: m_Specific(V: Op1), R: m_Value(V&: X)))) && |
5147 | (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) |
5148 | return new ICmpInst(Pred, Builder.CreateNot(V: Op1), X); |
5149 | // Op0 u>/u<= (Op0 + X) --> X u>/u<= ~Op0 |
5150 | if (match(V: Op1, P: m_OneUse(SubPattern: m_c_Add(L: m_Specific(V: Op0), R: m_Value(V&: X)))) && |
5151 | (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE)) |
5152 | return new ICmpInst(Pred, X, Builder.CreateNot(V: Op0)); |
5153 | |
5154 | { |
5155 | // (Op1 + X) + C u</u>= Op1 --> ~C - X u</u>= Op1 |
5156 | Constant *C; |
5157 | if (match(V: Op0, P: m_OneUse(SubPattern: m_Add(L: m_c_Add(L: m_Specific(V: Op1), R: m_Value(V&: X)), |
5158 | R: m_ImmConstant(C)))) && |
5159 | (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) { |
5160 | Constant *C2 = ConstantExpr::getNot(C); |
5161 | return new ICmpInst(Pred, Builder.CreateSub(LHS: C2, RHS: X), Op1); |
5162 | } |
5163 | // Op0 u>/u<= (Op0 + X) + C --> Op0 u>/u<= ~C - X |
5164 | if (match(V: Op1, P: m_OneUse(SubPattern: m_Add(L: m_c_Add(L: m_Specific(V: Op0), R: m_Value(V&: X)), |
5165 | R: m_ImmConstant(C)))) && |
5166 | (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE)) { |
5167 | Constant *C2 = ConstantExpr::getNot(C); |
5168 | return new ICmpInst(Pred, Op0, Builder.CreateSub(LHS: C2, RHS: X)); |
5169 | } |
5170 | } |
5171 | |
5172 | // (icmp eq/ne (X, -P2), INT_MIN) |
5173 | // -> (icmp slt/sge X, INT_MIN + P2) |
5174 | if (ICmpInst::isEquality(P: Pred) && BO0 && |
5175 | match(V: I.getOperand(i_nocapture: 1), P: m_SignMask()) && |
5176 | match(V: BO0, P: m_And(L: m_Value(), R: m_NegatedPower2OrZero()))) { |
5177 | // Will Constant fold. |
5178 | Value *NewC = Builder.CreateSub(LHS: I.getOperand(i_nocapture: 1), RHS: BO0->getOperand(i_nocapture: 1)); |
5179 | return new ICmpInst(Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_SLT |
5180 | : ICmpInst::ICMP_SGE, |
5181 | BO0->getOperand(i_nocapture: 0), NewC); |
5182 | } |
5183 | |
5184 | { |
5185 | // Similar to above: an unsigned overflow comparison may use offset + mask: |
5186 | // ((Op1 + C) & C) u< Op1 --> Op1 != 0 |
5187 | // ((Op1 + C) & C) u>= Op1 --> Op1 == 0 |
5188 | // Op0 u> ((Op0 + C) & C) --> Op0 != 0 |
5189 | // Op0 u<= ((Op0 + C) & C) --> Op0 == 0 |
5190 | BinaryOperator *BO; |
5191 | const APInt *C; |
5192 | if ((Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE) && |
5193 | match(V: Op0, P: m_And(L: m_BinOp(I&: BO), R: m_LowBitMask(V&: C))) && |
5194 | match(V: BO, P: m_Add(L: m_Specific(V: Op1), R: m_SpecificIntAllowPoison(V: *C)))) { |
5195 | CmpInst::Predicate NewPred = |
5196 | Pred == ICmpInst::ICMP_ULT ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ; |
5197 | Constant *Zero = ConstantInt::getNullValue(Ty: Op1->getType()); |
5198 | return new ICmpInst(NewPred, Op1, Zero); |
5199 | } |
5200 | |
5201 | if ((Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE) && |
5202 | match(V: Op1, P: m_And(L: m_BinOp(I&: BO), R: m_LowBitMask(V&: C))) && |
5203 | match(V: BO, P: m_Add(L: m_Specific(V: Op0), R: m_SpecificIntAllowPoison(V: *C)))) { |
5204 | CmpInst::Predicate NewPred = |
5205 | Pred == ICmpInst::ICMP_UGT ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ; |
5206 | Constant *Zero = ConstantInt::getNullValue(Ty: Op1->getType()); |
5207 | return new ICmpInst(NewPred, Op0, Zero); |
5208 | } |
5209 | } |
5210 | |
5211 | bool NoOp0WrapProblem = false, NoOp1WrapProblem = false; |
5212 | bool Op0HasNUW = false, Op1HasNUW = false; |
5213 | bool Op0HasNSW = false, Op1HasNSW = false; |
5214 | // Analyze the case when either Op0 or Op1 is an add instruction. |
5215 | // Op0 = A + B (or A and B are null); Op1 = C + D (or C and D are null). |
5216 | auto hasNoWrapProblem = [](const BinaryOperator &BO, CmpInst::Predicate Pred, |
5217 | bool &HasNSW, bool &HasNUW) -> bool { |
5218 | if (isa<OverflowingBinaryOperator>(Val: BO)) { |
5219 | HasNUW = BO.hasNoUnsignedWrap(); |
5220 | HasNSW = BO.hasNoSignedWrap(); |
5221 | return ICmpInst::isEquality(P: Pred) || |
5222 | (CmpInst::isUnsigned(predicate: Pred) && HasNUW) || |
5223 | (CmpInst::isSigned(predicate: Pred) && HasNSW); |
5224 | } else if (BO.getOpcode() == Instruction::Or) { |
5225 | HasNUW = true; |
5226 | HasNSW = true; |
5227 | return true; |
5228 | } else { |
5229 | return false; |
5230 | } |
5231 | }; |
5232 | Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; |
5233 | |
5234 | if (BO0) { |
5235 | match(V: BO0, P: m_AddLike(L: m_Value(V&: A), R: m_Value(V&: B))); |
5236 | NoOp0WrapProblem = hasNoWrapProblem(*BO0, Pred, Op0HasNSW, Op0HasNUW); |
5237 | } |
5238 | if (BO1) { |
5239 | match(V: BO1, P: m_AddLike(L: m_Value(V&: C), R: m_Value(V&: D))); |
5240 | NoOp1WrapProblem = hasNoWrapProblem(*BO1, Pred, Op1HasNSW, Op1HasNUW); |
5241 | } |
5242 | |
5243 | // icmp (A+B), A -> icmp B, 0 for equalities or if there is no overflow. |
5244 | // icmp (A+B), B -> icmp A, 0 for equalities or if there is no overflow. |
5245 | if ((A == Op1 || B == Op1) && NoOp0WrapProblem) |
5246 | return new ICmpInst(Pred, A == Op1 ? B : A, |
5247 | Constant::getNullValue(Ty: Op1->getType())); |
5248 | |
5249 | // icmp C, (C+D) -> icmp 0, D for equalities or if there is no overflow. |
5250 | // icmp D, (C+D) -> icmp 0, C for equalities or if there is no overflow. |
5251 | if ((C == Op0 || D == Op0) && NoOp1WrapProblem) |
5252 | return new ICmpInst(Pred, Constant::getNullValue(Ty: Op0->getType()), |
5253 | C == Op0 ? D : C); |
5254 | |
5255 | // icmp (A+B), (A+D) -> icmp B, D for equalities or if there is no overflow. |
5256 | if (A && C && (A == C || A == D || B == C || B == D) && NoOp0WrapProblem && |
5257 | NoOp1WrapProblem) { |
5258 | // Determine Y and Z in the form icmp (X+Y), (X+Z). |
5259 | Value *Y, *Z; |
5260 | if (A == C) { |
5261 | // C + B == C + D -> B == D |
5262 | Y = B; |
5263 | Z = D; |
5264 | } else if (A == D) { |
5265 | // D + B == C + D -> B == C |
5266 | Y = B; |
5267 | Z = C; |
5268 | } else if (B == C) { |
5269 | // A + C == C + D -> A == D |
5270 | Y = A; |
5271 | Z = D; |
5272 | } else { |
5273 | assert(B == D); |
5274 | // A + D == C + D -> A == C |
5275 | Y = A; |
5276 | Z = C; |
5277 | } |
5278 | return new ICmpInst(Pred, Y, Z); |
5279 | } |
5280 | |
5281 | // icmp slt (A + -1), Op1 -> icmp sle A, Op1 |
5282 | if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLT && |
5283 | match(V: B, P: m_AllOnes())) |
5284 | return new ICmpInst(CmpInst::ICMP_SLE, A, Op1); |
5285 | |
5286 | // icmp sge (A + -1), Op1 -> icmp sgt A, Op1 |
5287 | if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGE && |
5288 | match(V: B, P: m_AllOnes())) |
5289 | return new ICmpInst(CmpInst::ICMP_SGT, A, Op1); |
5290 | |
5291 | // icmp sle (A + 1), Op1 -> icmp slt A, Op1 |
5292 | if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SLE && match(V: B, P: m_One())) |
5293 | return new ICmpInst(CmpInst::ICMP_SLT, A, Op1); |
5294 | |
5295 | // icmp sgt (A + 1), Op1 -> icmp sge A, Op1 |
5296 | if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_SGT && match(V: B, P: m_One())) |
5297 | return new ICmpInst(CmpInst::ICMP_SGE, A, Op1); |
5298 | |
5299 | // icmp sgt Op0, (C + -1) -> icmp sge Op0, C |
5300 | if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGT && |
5301 | match(V: D, P: m_AllOnes())) |
5302 | return new ICmpInst(CmpInst::ICMP_SGE, Op0, C); |
5303 | |
5304 | // icmp sle Op0, (C + -1) -> icmp slt Op0, C |
5305 | if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLE && |
5306 | match(V: D, P: m_AllOnes())) |
5307 | return new ICmpInst(CmpInst::ICMP_SLT, Op0, C); |
5308 | |
5309 | // icmp sge Op0, (C + 1) -> icmp sgt Op0, C |
5310 | if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SGE && match(V: D, P: m_One())) |
5311 | return new ICmpInst(CmpInst::ICMP_SGT, Op0, C); |
5312 | |
5313 | // icmp slt Op0, (C + 1) -> icmp sle Op0, C |
5314 | if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_SLT && match(V: D, P: m_One())) |
5315 | return new ICmpInst(CmpInst::ICMP_SLE, Op0, C); |
5316 | |
5317 | // TODO: The subtraction-related identities shown below also hold, but |
5318 | // canonicalization from (X -nuw 1) to (X + -1) means that the combinations |
5319 | // wouldn't happen even if they were implemented. |
5320 | // |
5321 | // icmp ult (A - 1), Op1 -> icmp ule A, Op1 |
5322 | // icmp uge (A - 1), Op1 -> icmp ugt A, Op1 |
5323 | // icmp ugt Op0, (C - 1) -> icmp uge Op0, C |
5324 | // icmp ule Op0, (C - 1) -> icmp ult Op0, C |
5325 | |
5326 | // icmp ule (A + 1), Op0 -> icmp ult A, Op1 |
5327 | if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_ULE && match(V: B, P: m_One())) |
5328 | return new ICmpInst(CmpInst::ICMP_ULT, A, Op1); |
5329 | |
5330 | // icmp ugt (A + 1), Op0 -> icmp uge A, Op1 |
5331 | if (A && NoOp0WrapProblem && Pred == CmpInst::ICMP_UGT && match(V: B, P: m_One())) |
5332 | return new ICmpInst(CmpInst::ICMP_UGE, A, Op1); |
5333 | |
5334 | // icmp uge Op0, (C + 1) -> icmp ugt Op0, C |
5335 | if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_UGE && match(V: D, P: m_One())) |
5336 | return new ICmpInst(CmpInst::ICMP_UGT, Op0, C); |
5337 | |
5338 | // icmp ult Op0, (C + 1) -> icmp ule Op0, C |
5339 | if (C && NoOp1WrapProblem && Pred == CmpInst::ICMP_ULT && match(V: D, P: m_One())) |
5340 | return new ICmpInst(CmpInst::ICMP_ULE, Op0, C); |
5341 | |
5342 | // if C1 has greater magnitude than C2: |
5343 | // icmp (A + C1), (C + C2) -> icmp (A + C3), C |
5344 | // s.t. C3 = C1 - C2 |
5345 | // |
5346 | // if C2 has greater magnitude than C1: |
5347 | // icmp (A + C1), (C + C2) -> icmp A, (C + C3) |
5348 | // s.t. C3 = C2 - C1 |
5349 | if (A && C && NoOp0WrapProblem && NoOp1WrapProblem && |
5350 | (BO0->hasOneUse() || BO1->hasOneUse()) && !I.isUnsigned()) { |
5351 | const APInt *AP1, *AP2; |
5352 | // TODO: Support non-uniform vectors. |
5353 | // TODO: Allow poison passthrough if B or D's element is poison. |
5354 | if (match(V: B, P: m_APIntAllowPoison(Res&: AP1)) && |
5355 | match(V: D, P: m_APIntAllowPoison(Res&: AP2)) && |
5356 | AP1->isNegative() == AP2->isNegative()) { |
5357 | APInt AP1Abs = AP1->abs(); |
5358 | APInt AP2Abs = AP2->abs(); |
5359 | if (AP1Abs.uge(RHS: AP2Abs)) { |
5360 | APInt Diff = *AP1 - *AP2; |
5361 | Constant *C3 = Constant::getIntegerValue(Ty: BO0->getType(), V: Diff); |
5362 | Value *NewAdd = Builder.CreateAdd( |
5363 | LHS: A, RHS: C3, Name: "" , HasNUW: Op0HasNUW && Diff.ule(RHS: *AP1), HasNSW: Op0HasNSW); |
5364 | return new ICmpInst(Pred, NewAdd, C); |
5365 | } else { |
5366 | APInt Diff = *AP2 - *AP1; |
5367 | Constant *C3 = Constant::getIntegerValue(Ty: BO0->getType(), V: Diff); |
5368 | Value *NewAdd = Builder.CreateAdd( |
5369 | LHS: C, RHS: C3, Name: "" , HasNUW: Op1HasNUW && Diff.ule(RHS: *AP2), HasNSW: Op1HasNSW); |
5370 | return new ICmpInst(Pred, A, NewAdd); |
5371 | } |
5372 | } |
5373 | Constant *Cst1, *Cst2; |
5374 | if (match(V: B, P: m_ImmConstant(C&: Cst1)) && match(V: D, P: m_ImmConstant(C&: Cst2)) && |
5375 | ICmpInst::isEquality(P: Pred)) { |
5376 | Constant *Diff = ConstantExpr::getSub(C1: Cst2, C2: Cst1); |
5377 | Value *NewAdd = Builder.CreateAdd(LHS: C, RHS: Diff); |
5378 | return new ICmpInst(Pred, A, NewAdd); |
5379 | } |
5380 | } |
5381 | |
5382 | // Analyze the case when either Op0 or Op1 is a sub instruction. |
5383 | // Op0 = A - B (or A and B are null); Op1 = C - D (or C and D are null). |
5384 | A = nullptr; |
5385 | B = nullptr; |
5386 | C = nullptr; |
5387 | D = nullptr; |
5388 | if (BO0 && BO0->getOpcode() == Instruction::Sub) { |
5389 | A = BO0->getOperand(i_nocapture: 0); |
5390 | B = BO0->getOperand(i_nocapture: 1); |
5391 | } |
5392 | if (BO1 && BO1->getOpcode() == Instruction::Sub) { |
5393 | C = BO1->getOperand(i_nocapture: 0); |
5394 | D = BO1->getOperand(i_nocapture: 1); |
5395 | } |
5396 | |
5397 | // icmp (A-B), A -> icmp 0, B for equalities or if there is no overflow. |
5398 | if (A == Op1 && NoOp0WrapProblem) |
5399 | return new ICmpInst(Pred, Constant::getNullValue(Ty: Op1->getType()), B); |
5400 | // icmp C, (C-D) -> icmp D, 0 for equalities or if there is no overflow. |
5401 | if (C == Op0 && NoOp1WrapProblem) |
5402 | return new ICmpInst(Pred, D, Constant::getNullValue(Ty: Op0->getType())); |
5403 | |
5404 | // Convert sub-with-unsigned-overflow comparisons into a comparison of args. |
5405 | // (A - B) u>/u<= A --> B u>/u<= A |
5406 | if (A == Op1 && (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE)) |
5407 | return new ICmpInst(Pred, B, A); |
5408 | // C u</u>= (C - D) --> C u</u>= D |
5409 | if (C == Op0 && (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_UGE)) |
5410 | return new ICmpInst(Pred, C, D); |
5411 | // (A - B) u>=/u< A --> B u>/u<= A iff B != 0 |
5412 | if (A == Op1 && (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) && |
5413 | isKnownNonZero(V: B, Q)) |
5414 | return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(pred: Pred), B, A); |
5415 | // C u<=/u> (C - D) --> C u</u>= D iff B != 0 |
5416 | if (C == Op0 && (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_UGT) && |
5417 | isKnownNonZero(V: D, Q)) |
5418 | return new ICmpInst(CmpInst::getFlippedStrictnessPredicate(pred: Pred), C, D); |
5419 | |
5420 | // icmp (A-B), (C-B) -> icmp A, C for equalities or if there is no overflow. |
5421 | if (B && D && B == D && NoOp0WrapProblem && NoOp1WrapProblem) |
5422 | return new ICmpInst(Pred, A, C); |
5423 | |
5424 | // icmp (A-B), (A-D) -> icmp D, B for equalities or if there is no overflow. |
5425 | if (A && C && A == C && NoOp0WrapProblem && NoOp1WrapProblem) |
5426 | return new ICmpInst(Pred, D, B); |
5427 | |
5428 | // icmp (0-X) < cst --> x > -cst |
5429 | if (NoOp0WrapProblem && ICmpInst::isSigned(predicate: Pred)) { |
5430 | Value *X; |
5431 | if (match(V: BO0, P: m_Neg(V: m_Value(V&: X)))) |
5432 | if (Constant *RHSC = dyn_cast<Constant>(Val: Op1)) |
5433 | if (RHSC->isNotMinSignedValue()) |
5434 | return new ICmpInst(I.getSwappedPredicate(), X, |
5435 | ConstantExpr::getNeg(C: RHSC)); |
5436 | } |
5437 | |
5438 | if (Instruction *R = foldICmpXorXX(I, Q, IC&: *this)) |
5439 | return R; |
5440 | if (Instruction *R = foldICmpOrXX(I, Q, IC&: *this)) |
5441 | return R; |
5442 | |
5443 | { |
5444 | // Try to remove shared multiplier from comparison: |
5445 | // X * Z pred Y * Z |
5446 | Value *X, *Y, *Z; |
5447 | if ((match(V: Op0, P: m_Mul(L: m_Value(V&: X), R: m_Value(V&: Z))) && |
5448 | match(V: Op1, P: m_c_Mul(L: m_Specific(V: Z), R: m_Value(V&: Y)))) || |
5449 | (match(V: Op0, P: m_Mul(L: m_Value(V&: Z), R: m_Value(V&: X))) && |
5450 | match(V: Op1, P: m_c_Mul(L: m_Specific(V: Z), R: m_Value(V&: Y))))) { |
5451 | if (ICmpInst::isSigned(predicate: Pred)) { |
5452 | if (Op0HasNSW && Op1HasNSW) { |
5453 | KnownBits ZKnown = computeKnownBits(V: Z, CxtI: &I); |
5454 | if (ZKnown.isStrictlyPositive()) |
5455 | return new ICmpInst(Pred, X, Y); |
5456 | if (ZKnown.isNegative()) |
5457 | return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), X, Y); |
5458 | Value *LessThan = simplifyICmpInst(Pred: ICmpInst::ICMP_SLT, LHS: X, RHS: Y, |
5459 | Q: SQ.getWithInstruction(I: &I)); |
5460 | if (LessThan && match(V: LessThan, P: m_One())) |
5461 | return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Z, |
5462 | Constant::getNullValue(Ty: Z->getType())); |
5463 | Value *GreaterThan = simplifyICmpInst(Pred: ICmpInst::ICMP_SGT, LHS: X, RHS: Y, |
5464 | Q: SQ.getWithInstruction(I: &I)); |
5465 | if (GreaterThan && match(V: GreaterThan, P: m_One())) |
5466 | return new ICmpInst(Pred, Z, Constant::getNullValue(Ty: Z->getType())); |
5467 | } |
5468 | } else { |
5469 | bool NonZero; |
5470 | if (ICmpInst::isEquality(P: Pred)) { |
5471 | // If X != Y, fold (X *nw Z) eq/ne (Y *nw Z) -> Z eq/ne 0 |
5472 | if (((Op0HasNSW && Op1HasNSW) || (Op0HasNUW && Op1HasNUW)) && |
5473 | isKnownNonEqual(V1: X, V2: Y, SQ)) |
5474 | return new ICmpInst(Pred, Z, Constant::getNullValue(Ty: Z->getType())); |
5475 | |
5476 | KnownBits ZKnown = computeKnownBits(V: Z, CxtI: &I); |
5477 | // if Z % 2 != 0 |
5478 | // X * Z eq/ne Y * Z -> X eq/ne Y |
5479 | if (ZKnown.countMaxTrailingZeros() == 0) |
5480 | return new ICmpInst(Pred, X, Y); |
5481 | NonZero = !ZKnown.One.isZero() || isKnownNonZero(V: Z, Q); |
5482 | // if Z != 0 and nsw(X * Z) and nsw(Y * Z) |
5483 | // X * Z eq/ne Y * Z -> X eq/ne Y |
5484 | if (NonZero && BO0 && BO1 && Op0HasNSW && Op1HasNSW) |
5485 | return new ICmpInst(Pred, X, Y); |
5486 | } else |
5487 | NonZero = isKnownNonZero(V: Z, Q); |
5488 | |
5489 | // If Z != 0 and nuw(X * Z) and nuw(Y * Z) |
5490 | // X * Z u{lt/le/gt/ge}/eq/ne Y * Z -> X u{lt/le/gt/ge}/eq/ne Y |
5491 | if (NonZero && BO0 && BO1 && Op0HasNUW && Op1HasNUW) |
5492 | return new ICmpInst(Pred, X, Y); |
5493 | } |
5494 | } |
5495 | } |
5496 | |
5497 | BinaryOperator *SRem = nullptr; |
5498 | // icmp (srem X, Y), Y |
5499 | if (BO0 && BO0->getOpcode() == Instruction::SRem && Op1 == BO0->getOperand(i_nocapture: 1)) |
5500 | SRem = BO0; |
5501 | // icmp Y, (srem X, Y) |
5502 | else if (BO1 && BO1->getOpcode() == Instruction::SRem && |
5503 | Op0 == BO1->getOperand(i_nocapture: 1)) |
5504 | SRem = BO1; |
5505 | if (SRem) { |
5506 | // We don't check hasOneUse to avoid increasing register pressure because |
5507 | // the value we use is the same value this instruction was already using. |
5508 | switch (SRem == BO0 ? ICmpInst::getSwappedPredicate(pred: Pred) : Pred) { |
5509 | default: |
5510 | break; |
5511 | case ICmpInst::ICMP_EQ: |
5512 | return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType())); |
5513 | case ICmpInst::ICMP_NE: |
5514 | return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType())); |
5515 | case ICmpInst::ICMP_SGT: |
5516 | case ICmpInst::ICMP_SGE: |
5517 | return new ICmpInst(ICmpInst::ICMP_SGT, SRem->getOperand(i_nocapture: 1), |
5518 | Constant::getAllOnesValue(Ty: SRem->getType())); |
5519 | case ICmpInst::ICMP_SLT: |
5520 | case ICmpInst::ICMP_SLE: |
5521 | return new ICmpInst(ICmpInst::ICMP_SLT, SRem->getOperand(i_nocapture: 1), |
5522 | Constant::getNullValue(Ty: SRem->getType())); |
5523 | } |
5524 | } |
5525 | |
5526 | if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && |
5527 | (BO0->hasOneUse() || BO1->hasOneUse()) && |
5528 | BO0->getOperand(i_nocapture: 1) == BO1->getOperand(i_nocapture: 1)) { |
5529 | switch (BO0->getOpcode()) { |
5530 | default: |
5531 | break; |
5532 | case Instruction::Add: |
5533 | case Instruction::Sub: |
5534 | case Instruction::Xor: { |
5535 | if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b |
5536 | return new ICmpInst(Pred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0)); |
5537 | |
5538 | const APInt *C; |
5539 | if (match(V: BO0->getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) { |
5540 | // icmp u/s (a ^ signmask), (b ^ signmask) --> icmp s/u a, b |
5541 | if (C->isSignMask()) { |
5542 | ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate(); |
5543 | return new ICmpInst(NewPred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0)); |
5544 | } |
5545 | |
5546 | // icmp u/s (a ^ maxsignval), (b ^ maxsignval) --> icmp s/u' a, b |
5547 | if (BO0->getOpcode() == Instruction::Xor && C->isMaxSignedValue()) { |
5548 | ICmpInst::Predicate NewPred = I.getFlippedSignednessPredicate(); |
5549 | NewPred = I.getSwappedPredicate(pred: NewPred); |
5550 | return new ICmpInst(NewPred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0)); |
5551 | } |
5552 | } |
5553 | break; |
5554 | } |
5555 | case Instruction::Mul: { |
5556 | if (!I.isEquality()) |
5557 | break; |
5558 | |
5559 | const APInt *C; |
5560 | if (match(V: BO0->getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) && !C->isZero() && |
5561 | !C->isOne()) { |
5562 | // icmp eq/ne (X * C), (Y * C) --> icmp (X & Mask), (Y & Mask) |
5563 | // Mask = -1 >> count-trailing-zeros(C). |
5564 | if (unsigned TZs = C->countr_zero()) { |
5565 | Constant *Mask = ConstantInt::get( |
5566 | Ty: BO0->getType(), |
5567 | V: APInt::getLowBitsSet(numBits: C->getBitWidth(), loBitsSet: C->getBitWidth() - TZs)); |
5568 | Value *And1 = Builder.CreateAnd(LHS: BO0->getOperand(i_nocapture: 0), RHS: Mask); |
5569 | Value *And2 = Builder.CreateAnd(LHS: BO1->getOperand(i_nocapture: 0), RHS: Mask); |
5570 | return new ICmpInst(Pred, And1, And2); |
5571 | } |
5572 | } |
5573 | break; |
5574 | } |
5575 | case Instruction::UDiv: |
5576 | case Instruction::LShr: |
5577 | if (I.isSigned() || !BO0->isExact() || !BO1->isExact()) |
5578 | break; |
5579 | return new ICmpInst(Pred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0)); |
5580 | |
5581 | case Instruction::SDiv: |
5582 | if (!(I.isEquality() || match(V: BO0->getOperand(i_nocapture: 1), P: m_NonNegative())) || |
5583 | !BO0->isExact() || !BO1->isExact()) |
5584 | break; |
5585 | return new ICmpInst(Pred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0)); |
5586 | |
5587 | case Instruction::AShr: |
5588 | if (!BO0->isExact() || !BO1->isExact()) |
5589 | break; |
5590 | return new ICmpInst(Pred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0)); |
5591 | |
5592 | case Instruction::Shl: { |
5593 | bool NUW = Op0HasNUW && Op1HasNUW; |
5594 | bool NSW = Op0HasNSW && Op1HasNSW; |
5595 | if (!NUW && !NSW) |
5596 | break; |
5597 | if (!NSW && I.isSigned()) |
5598 | break; |
5599 | return new ICmpInst(Pred, BO0->getOperand(i_nocapture: 0), BO1->getOperand(i_nocapture: 0)); |
5600 | } |
5601 | } |
5602 | } |
5603 | |
5604 | if (BO0) { |
5605 | // Transform A & (L - 1) `ult` L --> L != 0 |
5606 | auto LSubOne = m_Add(L: m_Specific(V: Op1), R: m_AllOnes()); |
5607 | auto BitwiseAnd = m_c_And(L: m_Value(), R: LSubOne); |
5608 | |
5609 | if (match(V: BO0, P: BitwiseAnd) && Pred == ICmpInst::ICMP_ULT) { |
5610 | auto *Zero = Constant::getNullValue(Ty: BO0->getType()); |
5611 | return new ICmpInst(ICmpInst::ICMP_NE, Op1, Zero); |
5612 | } |
5613 | } |
5614 | |
5615 | // For unsigned predicates / eq / ne: |
5616 | // icmp pred (x << 1), x --> icmp getSignedPredicate(pred) x, 0 |
5617 | // icmp pred x, (x << 1) --> icmp getSignedPredicate(pred) 0, x |
5618 | if (!ICmpInst::isSigned(predicate: Pred)) { |
5619 | if (match(V: Op0, P: m_Shl(L: m_Specific(V: Op1), R: m_One()))) |
5620 | return new ICmpInst(ICmpInst::getSignedPredicate(Pred), Op1, |
5621 | Constant::getNullValue(Ty: Op1->getType())); |
5622 | else if (match(V: Op1, P: m_Shl(L: m_Specific(V: Op0), R: m_One()))) |
5623 | return new ICmpInst(ICmpInst::getSignedPredicate(Pred), |
5624 | Constant::getNullValue(Ty: Op0->getType()), Op0); |
5625 | } |
5626 | |
5627 | if (Value *V = foldMultiplicationOverflowCheck(I)) |
5628 | return replaceInstUsesWith(I, V); |
5629 | |
5630 | if (Instruction *R = foldICmpAndXX(I, Q, IC&: *this)) |
5631 | return R; |
5632 | |
5633 | if (Value *V = foldICmpWithTruncSignExtendedVal(I, Builder)) |
5634 | return replaceInstUsesWith(I, V); |
5635 | |
5636 | if (Value *V = foldShiftIntoShiftInAnotherHandOfAndInICmp(I, SQ, Builder)) |
5637 | return replaceInstUsesWith(I, V); |
5638 | |
5639 | return nullptr; |
5640 | } |
5641 | |
5642 | /// Fold icmp Pred min|max(X, Y), Z. |
5643 | Instruction *InstCombinerImpl::foldICmpWithMinMax(Instruction &I, |
5644 | MinMaxIntrinsic *MinMax, |
5645 | Value *Z, CmpPredicate Pred) { |
5646 | Value *X = MinMax->getLHS(); |
5647 | Value *Y = MinMax->getRHS(); |
5648 | if (ICmpInst::isSigned(predicate: Pred) && !MinMax->isSigned()) |
5649 | return nullptr; |
5650 | if (ICmpInst::isUnsigned(predicate: Pred) && MinMax->isSigned()) { |
5651 | // Revert the transform signed pred -> unsigned pred |
5652 | // TODO: We can flip the signedness of predicate if both operands of icmp |
5653 | // are negative. |
5654 | if (isKnownNonNegative(V: Z, SQ: SQ.getWithInstruction(I: &I)) && |
5655 | isKnownNonNegative(V: MinMax, SQ: SQ.getWithInstruction(I: &I))) { |
5656 | Pred = ICmpInst::getFlippedSignednessPredicate(Pred); |
5657 | } else |
5658 | return nullptr; |
5659 | } |
5660 | SimplifyQuery Q = SQ.getWithInstruction(I: &I); |
5661 | auto IsCondKnownTrue = [](Value *Val) -> std::optional<bool> { |
5662 | if (!Val) |
5663 | return std::nullopt; |
5664 | if (match(V: Val, P: m_One())) |
5665 | return true; |
5666 | if (match(V: Val, P: m_Zero())) |
5667 | return false; |
5668 | return std::nullopt; |
5669 | }; |
5670 | // Remove samesign here since it is illegal to keep it when we speculatively |
5671 | // execute comparisons. For example, `icmp samesign ult umax(X, -46), -32` |
5672 | // cannot be decomposed into `(icmp samesign ult X, -46) or (icmp samesign ult |
5673 | // -46, -32)`. `X` is allowed to be non-negative here. |
5674 | Pred = Pred.dropSameSign(); |
5675 | auto CmpXZ = IsCondKnownTrue(simplifyICmpInst(Pred, LHS: X, RHS: Z, Q)); |
5676 | auto CmpYZ = IsCondKnownTrue(simplifyICmpInst(Pred, LHS: Y, RHS: Z, Q)); |
5677 | if (!CmpXZ.has_value() && !CmpYZ.has_value()) |
5678 | return nullptr; |
5679 | if (!CmpXZ.has_value()) { |
5680 | std::swap(a&: X, b&: Y); |
5681 | std::swap(lhs&: CmpXZ, rhs&: CmpYZ); |
5682 | } |
5683 | |
5684 | auto FoldIntoCmpYZ = [&]() -> Instruction * { |
5685 | if (CmpYZ.has_value()) |
5686 | return replaceInstUsesWith(I, V: ConstantInt::getBool(Ty: I.getType(), V: *CmpYZ)); |
5687 | return ICmpInst::Create(Op: Instruction::ICmp, Pred, S1: Y, S2: Z); |
5688 | }; |
5689 | |
5690 | switch (Pred) { |
5691 | case ICmpInst::ICMP_EQ: |
5692 | case ICmpInst::ICMP_NE: { |
5693 | // If X == Z: |
5694 | // Expr Result |
5695 | // min(X, Y) == Z X <= Y |
5696 | // max(X, Y) == Z X >= Y |
5697 | // min(X, Y) != Z X > Y |
5698 | // max(X, Y) != Z X < Y |
5699 | if ((Pred == ICmpInst::ICMP_EQ) == *CmpXZ) { |
5700 | ICmpInst::Predicate NewPred = |
5701 | ICmpInst::getNonStrictPredicate(pred: MinMax->getPredicate()); |
5702 | if (Pred == ICmpInst::ICMP_NE) |
5703 | NewPred = ICmpInst::getInversePredicate(pred: NewPred); |
5704 | return ICmpInst::Create(Op: Instruction::ICmp, Pred: NewPred, S1: X, S2: Y); |
5705 | } |
5706 | // Otherwise (X != Z): |
5707 | ICmpInst::Predicate NewPred = MinMax->getPredicate(); |
5708 | auto MinMaxCmpXZ = IsCondKnownTrue(simplifyICmpInst(Pred: NewPred, LHS: X, RHS: Z, Q)); |
5709 | if (!MinMaxCmpXZ.has_value()) { |
5710 | std::swap(a&: X, b&: Y); |
5711 | std::swap(lhs&: CmpXZ, rhs&: CmpYZ); |
5712 | // Re-check pre-condition X != Z |
5713 | if (!CmpXZ.has_value() || (Pred == ICmpInst::ICMP_EQ) == *CmpXZ) |
5714 | break; |
5715 | MinMaxCmpXZ = IsCondKnownTrue(simplifyICmpInst(Pred: NewPred, LHS: X, RHS: Z, Q)); |
5716 | } |
5717 | if (!MinMaxCmpXZ.has_value()) |
5718 | break; |
5719 | if (*MinMaxCmpXZ) { |
5720 | // Expr Fact Result |
5721 | // min(X, Y) == Z X < Z false |
5722 | // max(X, Y) == Z X > Z false |
5723 | // min(X, Y) != Z X < Z true |
5724 | // max(X, Y) != Z X > Z true |
5725 | return replaceInstUsesWith( |
5726 | I, V: ConstantInt::getBool(Ty: I.getType(), V: Pred == ICmpInst::ICMP_NE)); |
5727 | } else { |
5728 | // Expr Fact Result |
5729 | // min(X, Y) == Z X > Z Y == Z |
5730 | // max(X, Y) == Z X < Z Y == Z |
5731 | // min(X, Y) != Z X > Z Y != Z |
5732 | // max(X, Y) != Z X < Z Y != Z |
5733 | return FoldIntoCmpYZ(); |
5734 | } |
5735 | break; |
5736 | } |
5737 | case ICmpInst::ICMP_SLT: |
5738 | case ICmpInst::ICMP_ULT: |
5739 | case ICmpInst::ICMP_SLE: |
5740 | case ICmpInst::ICMP_ULE: |
5741 | case ICmpInst::ICMP_SGT: |
5742 | case ICmpInst::ICMP_UGT: |
5743 | case ICmpInst::ICMP_SGE: |
5744 | case ICmpInst::ICMP_UGE: { |
5745 | bool IsSame = MinMax->getPredicate() == ICmpInst::getStrictPredicate(pred: Pred); |
5746 | if (*CmpXZ) { |
5747 | if (IsSame) { |
5748 | // Expr Fact Result |
5749 | // min(X, Y) < Z X < Z true |
5750 | // min(X, Y) <= Z X <= Z true |
5751 | // max(X, Y) > Z X > Z true |
5752 | // max(X, Y) >= Z X >= Z true |
5753 | return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType())); |
5754 | } else { |
5755 | // Expr Fact Result |
5756 | // max(X, Y) < Z X < Z Y < Z |
5757 | // max(X, Y) <= Z X <= Z Y <= Z |
5758 | // min(X, Y) > Z X > Z Y > Z |
5759 | // min(X, Y) >= Z X >= Z Y >= Z |
5760 | return FoldIntoCmpYZ(); |
5761 | } |
5762 | } else { |
5763 | if (IsSame) { |
5764 | // Expr Fact Result |
5765 | // min(X, Y) < Z X >= Z Y < Z |
5766 | // min(X, Y) <= Z X > Z Y <= Z |
5767 | // max(X, Y) > Z X <= Z Y > Z |
5768 | // max(X, Y) >= Z X < Z Y >= Z |
5769 | return FoldIntoCmpYZ(); |
5770 | } else { |
5771 | // Expr Fact Result |
5772 | // max(X, Y) < Z X >= Z false |
5773 | // max(X, Y) <= Z X > Z false |
5774 | // min(X, Y) > Z X <= Z false |
5775 | // min(X, Y) >= Z X < Z false |
5776 | return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType())); |
5777 | } |
5778 | } |
5779 | break; |
5780 | } |
5781 | default: |
5782 | break; |
5783 | } |
5784 | |
5785 | return nullptr; |
5786 | } |
5787 | |
5788 | // Canonicalize checking for a power-of-2-or-zero value: |
5789 | static Instruction *foldICmpPow2Test(ICmpInst &I, |
5790 | InstCombiner::BuilderTy &Builder) { |
5791 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
5792 | const CmpInst::Predicate Pred = I.getPredicate(); |
5793 | Value *A = nullptr; |
5794 | bool CheckIs; |
5795 | if (I.isEquality()) { |
5796 | // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants) |
5797 | // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants) |
5798 | if (!match(V: Op0, P: m_OneUse(SubPattern: m_c_And(L: m_Add(L: m_Value(V&: A), R: m_AllOnes()), |
5799 | R: m_Deferred(V: A)))) || |
5800 | !match(V: Op1, P: m_ZeroInt())) |
5801 | A = nullptr; |
5802 | |
5803 | // (A & -A) == A --> ctpop(A) < 2 (four commuted variants) |
5804 | // (-A & A) != A --> ctpop(A) > 1 (four commuted variants) |
5805 | if (match(V: Op0, P: m_OneUse(SubPattern: m_c_And(L: m_Neg(V: m_Specific(V: Op1)), R: m_Specific(V: Op1))))) |
5806 | A = Op1; |
5807 | else if (match(V: Op1, |
5808 | P: m_OneUse(SubPattern: m_c_And(L: m_Neg(V: m_Specific(V: Op0)), R: m_Specific(V: Op0))))) |
5809 | A = Op0; |
5810 | |
5811 | CheckIs = Pred == ICmpInst::ICMP_EQ; |
5812 | } else if (ICmpInst::isUnsigned(predicate: Pred)) { |
5813 | // (A ^ (A-1)) u>= A --> ctpop(A) < 2 (two commuted variants) |
5814 | // ((A-1) ^ A) u< A --> ctpop(A) > 1 (two commuted variants) |
5815 | |
5816 | if ((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) && |
5817 | match(V: Op0, P: m_OneUse(SubPattern: m_c_Xor(L: m_Add(L: m_Specific(V: Op1), R: m_AllOnes()), |
5818 | R: m_Specific(V: Op1))))) { |
5819 | A = Op1; |
5820 | CheckIs = Pred == ICmpInst::ICMP_UGE; |
5821 | } else if ((Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_ULE) && |
5822 | match(V: Op1, P: m_OneUse(SubPattern: m_c_Xor(L: m_Add(L: m_Specific(V: Op0), R: m_AllOnes()), |
5823 | R: m_Specific(V: Op0))))) { |
5824 | A = Op0; |
5825 | CheckIs = Pred == ICmpInst::ICMP_ULE; |
5826 | } |
5827 | } |
5828 | |
5829 | if (A) { |
5830 | Type *Ty = A->getType(); |
5831 | CallInst *CtPop = Builder.CreateUnaryIntrinsic(ID: Intrinsic::ctpop, V: A); |
5832 | return CheckIs ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop, |
5833 | ConstantInt::get(Ty, V: 2)) |
5834 | : new ICmpInst(ICmpInst::ICMP_UGT, CtPop, |
5835 | ConstantInt::get(Ty, V: 1)); |
5836 | } |
5837 | |
5838 | return nullptr; |
5839 | } |
5840 | |
5841 | /// Find all possible pairs (BinOp, RHS) that BinOp V, RHS can be simplified. |
5842 | using OffsetOp = std::pair<Instruction::BinaryOps, Value *>; |
5843 | static void collectOffsetOp(Value *V, SmallVectorImpl<OffsetOp> &Offsets, |
5844 | bool AllowRecursion) { |
5845 | Instruction *Inst = dyn_cast<Instruction>(Val: V); |
5846 | if (!Inst || !Inst->hasOneUse()) |
5847 | return; |
5848 | |
5849 | switch (Inst->getOpcode()) { |
5850 | case Instruction::Add: |
5851 | Offsets.emplace_back(Args: Instruction::Sub, Args: Inst->getOperand(i: 1)); |
5852 | Offsets.emplace_back(Args: Instruction::Sub, Args: Inst->getOperand(i: 0)); |
5853 | break; |
5854 | case Instruction::Sub: |
5855 | Offsets.emplace_back(Args: Instruction::Add, Args: Inst->getOperand(i: 1)); |
5856 | break; |
5857 | case Instruction::Xor: |
5858 | Offsets.emplace_back(Args: Instruction::Xor, Args: Inst->getOperand(i: 1)); |
5859 | Offsets.emplace_back(Args: Instruction::Xor, Args: Inst->getOperand(i: 0)); |
5860 | break; |
5861 | case Instruction::Select: |
5862 | if (AllowRecursion) { |
5863 | collectOffsetOp(V: Inst->getOperand(i: 1), Offsets, /*AllowRecursion=*/false); |
5864 | collectOffsetOp(V: Inst->getOperand(i: 2), Offsets, /*AllowRecursion=*/false); |
5865 | } |
5866 | break; |
5867 | default: |
5868 | break; |
5869 | } |
5870 | } |
5871 | |
5872 | enum class OffsetKind { Invalid, Value, Select }; |
5873 | |
5874 | struct OffsetResult { |
5875 | OffsetKind Kind; |
5876 | Value *V0, *V1, *V2; |
5877 | |
5878 | static OffsetResult invalid() { |
5879 | return {.Kind: OffsetKind::Invalid, .V0: nullptr, .V1: nullptr, .V2: nullptr}; |
5880 | } |
5881 | static OffsetResult value(Value *V) { |
5882 | return {.Kind: OffsetKind::Value, .V0: V, .V1: nullptr, .V2: nullptr}; |
5883 | } |
5884 | static OffsetResult select(Value *Cond, Value *TrueV, Value *FalseV) { |
5885 | return {.Kind: OffsetKind::Select, .V0: Cond, .V1: TrueV, .V2: FalseV}; |
5886 | } |
5887 | bool isValid() const { return Kind != OffsetKind::Invalid; } |
5888 | Value *materialize(InstCombiner::BuilderTy &Builder) const { |
5889 | switch (Kind) { |
5890 | case OffsetKind::Invalid: |
5891 | llvm_unreachable("Invalid offset result" ); |
5892 | case OffsetKind::Value: |
5893 | return V0; |
5894 | case OffsetKind::Select: |
5895 | return Builder.CreateSelect(C: V0, True: V1, False: V2); |
5896 | } |
5897 | llvm_unreachable("Unknown OffsetKind enum" ); |
5898 | } |
5899 | }; |
5900 | |
5901 | /// Offset both sides of an equality icmp to see if we can save some |
5902 | /// instructions: icmp eq/ne X, Y -> icmp eq/ne X op Z, Y op Z. |
5903 | /// Note: This operation should not introduce poison. |
5904 | static Instruction *foldICmpEqualityWithOffset(ICmpInst &I, |
5905 | InstCombiner::BuilderTy &Builder, |
5906 | const SimplifyQuery &SQ) { |
5907 | assert(I.isEquality() && "Expected an equality icmp" ); |
5908 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
5909 | if (!Op0->getType()->isIntOrIntVectorTy()) |
5910 | return nullptr; |
5911 | |
5912 | SmallVector<OffsetOp, 4> OffsetOps; |
5913 | collectOffsetOp(V: Op0, Offsets&: OffsetOps, /*AllowRecursion=*/true); |
5914 | collectOffsetOp(V: Op1, Offsets&: OffsetOps, /*AllowRecursion=*/true); |
5915 | |
5916 | auto ApplyOffsetImpl = [&](Value *V, unsigned BinOpc, Value *RHS) -> Value * { |
5917 | Value *Simplified = simplifyBinOp(Opcode: BinOpc, LHS: V, RHS, Q: SQ); |
5918 | // Avoid infinite loops by checking if RHS is an identity for the BinOp. |
5919 | if (!Simplified || Simplified == V) |
5920 | return nullptr; |
5921 | // Reject constant expressions as they don't simplify things. |
5922 | if (isa<Constant>(Val: Simplified) && !match(V: Simplified, P: m_ImmConstant())) |
5923 | return nullptr; |
5924 | // Check if the transformation introduces poison. |
5925 | return impliesPoison(ValAssumedPoison: RHS, V) ? Simplified : nullptr; |
5926 | }; |
5927 | |
5928 | auto ApplyOffset = [&](Value *V, unsigned BinOpc, |
5929 | Value *RHS) -> OffsetResult { |
5930 | if (auto *Sel = dyn_cast<SelectInst>(Val: V)) { |
5931 | if (!Sel->hasOneUse()) |
5932 | return OffsetResult::invalid(); |
5933 | Value *TrueVal = ApplyOffsetImpl(Sel->getTrueValue(), BinOpc, RHS); |
5934 | if (!TrueVal) |
5935 | return OffsetResult::invalid(); |
5936 | Value *FalseVal = ApplyOffsetImpl(Sel->getFalseValue(), BinOpc, RHS); |
5937 | if (!FalseVal) |
5938 | return OffsetResult::invalid(); |
5939 | return OffsetResult::select(Cond: Sel->getCondition(), TrueV: TrueVal, FalseV: FalseVal); |
5940 | } |
5941 | if (Value *Simplified = ApplyOffsetImpl(V, BinOpc, RHS)) |
5942 | return OffsetResult::value(V: Simplified); |
5943 | return OffsetResult::invalid(); |
5944 | }; |
5945 | |
5946 | for (auto [BinOp, RHS] : OffsetOps) { |
5947 | auto BinOpc = static_cast<unsigned>(BinOp); |
5948 | |
5949 | auto Op0Result = ApplyOffset(Op0, BinOpc, RHS); |
5950 | if (!Op0Result.isValid()) |
5951 | continue; |
5952 | auto Op1Result = ApplyOffset(Op1, BinOpc, RHS); |
5953 | if (!Op1Result.isValid()) |
5954 | continue; |
5955 | |
5956 | Value *NewLHS = Op0Result.materialize(Builder); |
5957 | Value *NewRHS = Op1Result.materialize(Builder); |
5958 | return new ICmpInst(I.getPredicate(), NewLHS, NewRHS); |
5959 | } |
5960 | |
5961 | return nullptr; |
5962 | } |
5963 | |
5964 | Instruction *InstCombinerImpl::foldICmpEquality(ICmpInst &I) { |
5965 | if (!I.isEquality()) |
5966 | return nullptr; |
5967 | |
5968 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
5969 | const CmpInst::Predicate Pred = I.getPredicate(); |
5970 | Value *A, *B, *C, *D; |
5971 | if (match(V: Op0, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B)))) { |
5972 | if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0 |
5973 | Value *OtherVal = A == Op1 ? B : A; |
5974 | return new ICmpInst(Pred, OtherVal, Constant::getNullValue(Ty: A->getType())); |
5975 | } |
5976 | |
5977 | if (match(V: Op1, P: m_Xor(L: m_Value(V&: C), R: m_Value(V&: D)))) { |
5978 | // A^c1 == C^c2 --> A == C^(c1^c2) |
5979 | ConstantInt *C1, *C2; |
5980 | if (match(V: B, P: m_ConstantInt(CI&: C1)) && match(V: D, P: m_ConstantInt(CI&: C2)) && |
5981 | Op1->hasOneUse()) { |
5982 | Constant *NC = Builder.getInt(AI: C1->getValue() ^ C2->getValue()); |
5983 | Value *Xor = Builder.CreateXor(LHS: C, RHS: NC); |
5984 | return new ICmpInst(Pred, A, Xor); |
5985 | } |
5986 | |
5987 | // A^B == A^D -> B == D |
5988 | if (A == C) |
5989 | return new ICmpInst(Pred, B, D); |
5990 | if (A == D) |
5991 | return new ICmpInst(Pred, B, C); |
5992 | if (B == C) |
5993 | return new ICmpInst(Pred, A, D); |
5994 | if (B == D) |
5995 | return new ICmpInst(Pred, A, C); |
5996 | } |
5997 | } |
5998 | |
5999 | if (match(V: Op1, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))) && (A == Op0 || B == Op0)) { |
6000 | // A == (A^B) -> B == 0 |
6001 | Value *OtherVal = A == Op0 ? B : A; |
6002 | return new ICmpInst(Pred, OtherVal, Constant::getNullValue(Ty: A->getType())); |
6003 | } |
6004 | |
6005 | // (X&Z) == (Y&Z) -> (X^Y) & Z == 0 |
6006 | if (match(V: Op0, P: m_And(L: m_Value(V&: A), R: m_Value(V&: B))) && |
6007 | match(V: Op1, P: m_And(L: m_Value(V&: C), R: m_Value(V&: D)))) { |
6008 | Value *X = nullptr, *Y = nullptr, *Z = nullptr; |
6009 | |
6010 | if (A == C) { |
6011 | X = B; |
6012 | Y = D; |
6013 | Z = A; |
6014 | } else if (A == D) { |
6015 | X = B; |
6016 | Y = C; |
6017 | Z = A; |
6018 | } else if (B == C) { |
6019 | X = A; |
6020 | Y = D; |
6021 | Z = B; |
6022 | } else if (B == D) { |
6023 | X = A; |
6024 | Y = C; |
6025 | Z = B; |
6026 | } |
6027 | |
6028 | if (X) { |
6029 | // If X^Y is a negative power of two, then `icmp eq/ne (Z & NegP2), 0` |
6030 | // will fold to `icmp ult/uge Z, -NegP2` incurringb no additional |
6031 | // instructions. |
6032 | const APInt *C0, *C1; |
6033 | bool XorIsNegP2 = match(V: X, P: m_APInt(Res&: C0)) && match(V: Y, P: m_APInt(Res&: C1)) && |
6034 | (*C0 ^ *C1).isNegatedPowerOf2(); |
6035 | |
6036 | // If either Op0/Op1 are both one use or X^Y will constant fold and one of |
6037 | // Op0/Op1 are one use, proceed. In those cases we are instruction neutral |
6038 | // but `icmp eq/ne A, 0` is easier to analyze than `icmp eq/ne A, B`. |
6039 | int UseCnt = |
6040 | int(Op0->hasOneUse()) + int(Op1->hasOneUse()) + |
6041 | (int(match(V: X, P: m_ImmConstant()) && match(V: Y, P: m_ImmConstant()))); |
6042 | if (XorIsNegP2 || UseCnt >= 2) { |
6043 | // Build (X^Y) & Z |
6044 | Op1 = Builder.CreateXor(LHS: X, RHS: Y); |
6045 | Op1 = Builder.CreateAnd(LHS: Op1, RHS: Z); |
6046 | return new ICmpInst(Pred, Op1, Constant::getNullValue(Ty: Op1->getType())); |
6047 | } |
6048 | } |
6049 | } |
6050 | |
6051 | { |
6052 | // Similar to above, but specialized for constant because invert is needed: |
6053 | // (X | C) == (Y | C) --> (X ^ Y) & ~C == 0 |
6054 | Value *X, *Y; |
6055 | Constant *C; |
6056 | if (match(V: Op0, P: m_OneUse(SubPattern: m_Or(L: m_Value(V&: X), R: m_Constant(C)))) && |
6057 | match(V: Op1, P: m_OneUse(SubPattern: m_Or(L: m_Value(V&: Y), R: m_Specific(V: C))))) { |
6058 | Value *Xor = Builder.CreateXor(LHS: X, RHS: Y); |
6059 | Value *And = Builder.CreateAnd(LHS: Xor, RHS: ConstantExpr::getNot(C)); |
6060 | return new ICmpInst(Pred, And, Constant::getNullValue(Ty: And->getType())); |
6061 | } |
6062 | } |
6063 | |
6064 | if (match(V: Op1, P: m_ZExt(Op: m_Value(V&: A))) && |
6065 | (Op0->hasOneUse() || Op1->hasOneUse())) { |
6066 | // (B & (Pow2C-1)) == zext A --> A == trunc B |
6067 | // (B & (Pow2C-1)) != zext A --> A != trunc B |
6068 | const APInt *MaskC; |
6069 | if (match(V: Op0, P: m_And(L: m_Value(V&: B), R: m_LowBitMask(V&: MaskC))) && |
6070 | MaskC->countr_one() == A->getType()->getScalarSizeInBits()) |
6071 | return new ICmpInst(Pred, A, Builder.CreateTrunc(V: B, DestTy: A->getType())); |
6072 | } |
6073 | |
6074 | // (A >> C) == (B >> C) --> (A^B) u< (1 << C) |
6075 | // For lshr and ashr pairs. |
6076 | const APInt *AP1, *AP2; |
6077 | if ((match(V: Op0, P: m_OneUse(SubPattern: m_LShr(L: m_Value(V&: A), R: m_APIntAllowPoison(Res&: AP1)))) && |
6078 | match(V: Op1, P: m_OneUse(SubPattern: m_LShr(L: m_Value(V&: B), R: m_APIntAllowPoison(Res&: AP2))))) || |
6079 | (match(V: Op0, P: m_OneUse(SubPattern: m_AShr(L: m_Value(V&: A), R: m_APIntAllowPoison(Res&: AP1)))) && |
6080 | match(V: Op1, P: m_OneUse(SubPattern: m_AShr(L: m_Value(V&: B), R: m_APIntAllowPoison(Res&: AP2)))))) { |
6081 | if (AP1 != AP2) |
6082 | return nullptr; |
6083 | unsigned TypeBits = AP1->getBitWidth(); |
6084 | unsigned ShAmt = AP1->getLimitedValue(Limit: TypeBits); |
6085 | if (ShAmt < TypeBits && ShAmt != 0) { |
6086 | ICmpInst::Predicate NewPred = |
6087 | Pred == ICmpInst::ICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT; |
6088 | Value *Xor = Builder.CreateXor(LHS: A, RHS: B, Name: I.getName() + ".unshifted" ); |
6089 | APInt CmpVal = APInt::getOneBitSet(numBits: TypeBits, BitNo: ShAmt); |
6090 | return new ICmpInst(NewPred, Xor, ConstantInt::get(Ty: A->getType(), V: CmpVal)); |
6091 | } |
6092 | } |
6093 | |
6094 | // (A << C) == (B << C) --> ((A^B) & (~0U >> C)) == 0 |
6095 | ConstantInt *Cst1; |
6096 | if (match(V: Op0, P: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: A), R: m_ConstantInt(CI&: Cst1)))) && |
6097 | match(V: Op1, P: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: B), R: m_Specific(V: Cst1))))) { |
6098 | unsigned TypeBits = Cst1->getBitWidth(); |
6099 | unsigned ShAmt = (unsigned)Cst1->getLimitedValue(Limit: TypeBits); |
6100 | if (ShAmt < TypeBits && ShAmt != 0) { |
6101 | Value *Xor = Builder.CreateXor(LHS: A, RHS: B, Name: I.getName() + ".unshifted" ); |
6102 | APInt AndVal = APInt::getLowBitsSet(numBits: TypeBits, loBitsSet: TypeBits - ShAmt); |
6103 | Value *And = |
6104 | Builder.CreateAnd(LHS: Xor, RHS: Builder.getInt(AI: AndVal), Name: I.getName() + ".mask" ); |
6105 | return new ICmpInst(Pred, And, Constant::getNullValue(Ty: Cst1->getType())); |
6106 | } |
6107 | } |
6108 | |
6109 | // Transform "icmp eq (trunc (lshr(X, cst1)), cst" to |
6110 | // "icmp (and X, mask), cst" |
6111 | uint64_t ShAmt = 0; |
6112 | if (Op0->hasOneUse() && |
6113 | match(V: Op0, P: m_Trunc(Op: m_OneUse(SubPattern: m_LShr(L: m_Value(V&: A), R: m_ConstantInt(V&: ShAmt))))) && |
6114 | match(V: Op1, P: m_ConstantInt(CI&: Cst1)) && |
6115 | // Only do this when A has multiple uses. This is most important to do |
6116 | // when it exposes other optimizations. |
6117 | !A->hasOneUse()) { |
6118 | unsigned ASize = cast<IntegerType>(Val: A->getType())->getPrimitiveSizeInBits(); |
6119 | |
6120 | if (ShAmt < ASize) { |
6121 | APInt MaskV = |
6122 | APInt::getLowBitsSet(numBits: ASize, loBitsSet: Op0->getType()->getPrimitiveSizeInBits()); |
6123 | MaskV <<= ShAmt; |
6124 | |
6125 | APInt CmpV = Cst1->getValue().zext(width: ASize); |
6126 | CmpV <<= ShAmt; |
6127 | |
6128 | Value *Mask = Builder.CreateAnd(LHS: A, RHS: Builder.getInt(AI: MaskV)); |
6129 | return new ICmpInst(Pred, Mask, Builder.getInt(AI: CmpV)); |
6130 | } |
6131 | } |
6132 | |
6133 | if (Instruction *ICmp = foldICmpIntrinsicWithIntrinsic(Cmp&: I, Builder)) |
6134 | return ICmp; |
6135 | |
6136 | // Match icmp eq (trunc (lshr A, BW), (ashr (trunc A), BW-1)), which checks |
6137 | // the top BW/2 + 1 bits are all the same. Create "A >=s INT_MIN && A <=s |
6138 | // INT_MAX", which we generate as "icmp ult (add A, 2^(BW-1)), 2^BW" to skip a |
6139 | // few steps of instcombine. |
6140 | unsigned BitWidth = Op0->getType()->getScalarSizeInBits(); |
6141 | if (match(V: Op0, P: m_AShr(L: m_Trunc(Op: m_Value(V&: A)), R: m_SpecificInt(V: BitWidth - 1))) && |
6142 | match(V: Op1, P: m_Trunc(Op: m_LShr(L: m_Specific(V: A), R: m_SpecificInt(V: BitWidth)))) && |
6143 | A->getType()->getScalarSizeInBits() == BitWidth * 2 && |
6144 | (I.getOperand(i_nocapture: 0)->hasOneUse() || I.getOperand(i_nocapture: 1)->hasOneUse())) { |
6145 | APInt C = APInt::getOneBitSet(numBits: BitWidth * 2, BitNo: BitWidth - 1); |
6146 | Value *Add = Builder.CreateAdd(LHS: A, RHS: ConstantInt::get(Ty: A->getType(), V: C)); |
6147 | return new ICmpInst(Pred == ICmpInst::ICMP_EQ ? ICmpInst::ICMP_ULT |
6148 | : ICmpInst::ICMP_UGE, |
6149 | Add, ConstantInt::get(Ty: A->getType(), V: C.shl(shiftAmt: 1))); |
6150 | } |
6151 | |
6152 | // Canonicalize: |
6153 | // Assume B_Pow2 != 0 |
6154 | // 1. A & B_Pow2 != B_Pow2 -> A & B_Pow2 == 0 |
6155 | // 2. A & B_Pow2 == B_Pow2 -> A & B_Pow2 != 0 |
6156 | if (match(V: Op0, P: m_c_And(L: m_Specific(V: Op1), R: m_Value())) && |
6157 | isKnownToBeAPowerOfTwo(V: Op1, /* OrZero */ false, CxtI: &I)) |
6158 | return new ICmpInst(CmpInst::getInversePredicate(pred: Pred), Op0, |
6159 | ConstantInt::getNullValue(Ty: Op0->getType())); |
6160 | |
6161 | if (match(V: Op1, P: m_c_And(L: m_Specific(V: Op0), R: m_Value())) && |
6162 | isKnownToBeAPowerOfTwo(V: Op0, /* OrZero */ false, CxtI: &I)) |
6163 | return new ICmpInst(CmpInst::getInversePredicate(pred: Pred), Op1, |
6164 | ConstantInt::getNullValue(Ty: Op1->getType())); |
6165 | |
6166 | // Canonicalize: |
6167 | // icmp eq/ne X, OneUse(rotate-right(X)) |
6168 | // -> icmp eq/ne X, rotate-left(X) |
6169 | // We generally try to convert rotate-right -> rotate-left, this just |
6170 | // canonicalizes another case. |
6171 | if (match(V: &I, P: m_c_ICmp(L: m_Value(V&: A), |
6172 | R: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::fshr>( |
6173 | Op0: m_Deferred(V: A), Op1: m_Deferred(V: A), Op2: m_Value(V&: B)))))) |
6174 | return new ICmpInst( |
6175 | Pred, A, |
6176 | Builder.CreateIntrinsic(RetTy: Op0->getType(), ID: Intrinsic::fshl, Args: {A, A, B})); |
6177 | |
6178 | // Canonicalize: |
6179 | // icmp eq/ne OneUse(A ^ Cst), B --> icmp eq/ne (A ^ B), Cst |
6180 | Constant *Cst; |
6181 | if (match(V: &I, P: m_c_ICmp(L: m_OneUse(SubPattern: m_Xor(L: m_Value(V&: A), R: m_ImmConstant(C&: Cst))), |
6182 | R: m_CombineAnd(L: m_Value(V&: B), R: m_Unless(M: m_ImmConstant()))))) |
6183 | return new ICmpInst(Pred, Builder.CreateXor(LHS: A, RHS: B), Cst); |
6184 | |
6185 | { |
6186 | // (icmp eq/ne (and (add/sub/xor X, P2), P2), P2) |
6187 | auto m_Matcher = |
6188 | m_CombineOr(L: m_CombineOr(L: m_c_Add(L: m_Value(V&: B), R: m_Deferred(V: A)), |
6189 | R: m_c_Xor(L: m_Value(V&: B), R: m_Deferred(V: A))), |
6190 | R: m_Sub(L: m_Value(V&: B), R: m_Deferred(V: A))); |
6191 | std::optional<bool> IsZero = std::nullopt; |
6192 | if (match(V: &I, P: m_c_ICmp(L: m_OneUse(SubPattern: m_c_And(L: m_Value(V&: A), R: m_Matcher)), |
6193 | R: m_Deferred(V: A)))) |
6194 | IsZero = false; |
6195 | // (icmp eq/ne (and (add/sub/xor X, P2), P2), 0) |
6196 | else if (match(V: &I, |
6197 | P: m_ICmp(L: m_OneUse(SubPattern: m_c_And(L: m_Value(V&: A), R: m_Matcher)), R: m_Zero()))) |
6198 | IsZero = true; |
6199 | |
6200 | if (IsZero && isKnownToBeAPowerOfTwo(V: A, /* OrZero */ true, CxtI: &I)) |
6201 | // (icmp eq/ne (and (add/sub/xor X, P2), P2), P2) |
6202 | // -> (icmp eq/ne (and X, P2), 0) |
6203 | // (icmp eq/ne (and (add/sub/xor X, P2), P2), 0) |
6204 | // -> (icmp eq/ne (and X, P2), P2) |
6205 | return new ICmpInst(Pred, Builder.CreateAnd(LHS: B, RHS: A), |
6206 | *IsZero ? A |
6207 | : ConstantInt::getNullValue(Ty: A->getType())); |
6208 | } |
6209 | |
6210 | if (auto *Res = foldICmpEqualityWithOffset( |
6211 | I, Builder, SQ: getSimplifyQuery().getWithInstruction(I: &I))) |
6212 | return Res; |
6213 | |
6214 | return nullptr; |
6215 | } |
6216 | |
6217 | Instruction *InstCombinerImpl::foldICmpWithTrunc(ICmpInst &ICmp) { |
6218 | ICmpInst::Predicate Pred = ICmp.getPredicate(); |
6219 | Value *Op0 = ICmp.getOperand(i_nocapture: 0), *Op1 = ICmp.getOperand(i_nocapture: 1); |
6220 | |
6221 | // Try to canonicalize trunc + compare-to-constant into a mask + cmp. |
6222 | // The trunc masks high bits while the compare may effectively mask low bits. |
6223 | Value *X; |
6224 | const APInt *C; |
6225 | if (!match(V: Op0, P: m_OneUse(SubPattern: m_Trunc(Op: m_Value(V&: X)))) || !match(V: Op1, P: m_APInt(Res&: C))) |
6226 | return nullptr; |
6227 | |
6228 | // This matches patterns corresponding to tests of the signbit as well as: |
6229 | // (trunc X) pred C2 --> (X & Mask) == C |
6230 | if (auto Res = decomposeBitTestICmp(LHS: Op0, RHS: Op1, Pred, /*WithTrunc=*/LookThroughTrunc: true, |
6231 | /*AllowNonZeroC=*/true)) { |
6232 | Value *And = Builder.CreateAnd(LHS: Res->X, RHS: Res->Mask); |
6233 | Constant *C = ConstantInt::get(Ty: Res->X->getType(), V: Res->C); |
6234 | return new ICmpInst(Res->Pred, And, C); |
6235 | } |
6236 | |
6237 | unsigned SrcBits = X->getType()->getScalarSizeInBits(); |
6238 | if (auto *II = dyn_cast<IntrinsicInst>(Val: X)) { |
6239 | if (II->getIntrinsicID() == Intrinsic::cttz || |
6240 | II->getIntrinsicID() == Intrinsic::ctlz) { |
6241 | unsigned MaxRet = SrcBits; |
6242 | // If the "is_zero_poison" argument is set, then we know at least |
6243 | // one bit is set in the input, so the result is always at least one |
6244 | // less than the full bitwidth of that input. |
6245 | if (match(V: II->getArgOperand(i: 1), P: m_One())) |
6246 | MaxRet--; |
6247 | |
6248 | // Make sure the destination is wide enough to hold the largest output of |
6249 | // the intrinsic. |
6250 | if (llvm::Log2_32(Value: MaxRet) + 1 <= Op0->getType()->getScalarSizeInBits()) |
6251 | if (Instruction *I = |
6252 | foldICmpIntrinsicWithConstant(Cmp&: ICmp, II, C: C->zext(width: SrcBits))) |
6253 | return I; |
6254 | } |
6255 | } |
6256 | |
6257 | return nullptr; |
6258 | } |
6259 | |
6260 | Instruction *InstCombinerImpl::foldICmpWithZextOrSext(ICmpInst &ICmp) { |
6261 | assert(isa<CastInst>(ICmp.getOperand(0)) && "Expected cast for operand 0" ); |
6262 | auto *CastOp0 = cast<CastInst>(Val: ICmp.getOperand(i_nocapture: 0)); |
6263 | Value *X; |
6264 | if (!match(V: CastOp0, P: m_ZExtOrSExt(Op: m_Value(V&: X)))) |
6265 | return nullptr; |
6266 | |
6267 | bool IsSignedExt = CastOp0->getOpcode() == Instruction::SExt; |
6268 | bool IsSignedCmp = ICmp.isSigned(); |
6269 | |
6270 | // icmp Pred (ext X), (ext Y) |
6271 | Value *Y; |
6272 | if (match(V: ICmp.getOperand(i_nocapture: 1), P: m_ZExtOrSExt(Op: m_Value(V&: Y)))) { |
6273 | bool IsZext0 = isa<ZExtInst>(Val: ICmp.getOperand(i_nocapture: 0)); |
6274 | bool IsZext1 = isa<ZExtInst>(Val: ICmp.getOperand(i_nocapture: 1)); |
6275 | |
6276 | if (IsZext0 != IsZext1) { |
6277 | // If X and Y and both i1 |
6278 | // (icmp eq/ne (zext X) (sext Y)) |
6279 | // eq -> (icmp eq (or X, Y), 0) |
6280 | // ne -> (icmp ne (or X, Y), 0) |
6281 | if (ICmp.isEquality() && X->getType()->isIntOrIntVectorTy(BitWidth: 1) && |
6282 | Y->getType()->isIntOrIntVectorTy(BitWidth: 1)) |
6283 | return new ICmpInst(ICmp.getPredicate(), Builder.CreateOr(LHS: X, RHS: Y), |
6284 | Constant::getNullValue(Ty: X->getType())); |
6285 | |
6286 | // If we have mismatched casts and zext has the nneg flag, we can |
6287 | // treat the "zext nneg" as "sext". Otherwise, we cannot fold and quit. |
6288 | |
6289 | auto *NonNegInst0 = dyn_cast<PossiblyNonNegInst>(Val: ICmp.getOperand(i_nocapture: 0)); |
6290 | auto *NonNegInst1 = dyn_cast<PossiblyNonNegInst>(Val: ICmp.getOperand(i_nocapture: 1)); |
6291 | |
6292 | bool IsNonNeg0 = NonNegInst0 && NonNegInst0->hasNonNeg(); |
6293 | bool IsNonNeg1 = NonNegInst1 && NonNegInst1->hasNonNeg(); |
6294 | |
6295 | if ((IsZext0 && IsNonNeg0) || (IsZext1 && IsNonNeg1)) |
6296 | IsSignedExt = true; |
6297 | else |
6298 | return nullptr; |
6299 | } |
6300 | |
6301 | // Not an extension from the same type? |
6302 | Type *XTy = X->getType(), *YTy = Y->getType(); |
6303 | if (XTy != YTy) { |
6304 | // One of the casts must have one use because we are creating a new cast. |
6305 | if (!ICmp.getOperand(i_nocapture: 0)->hasOneUse() && !ICmp.getOperand(i_nocapture: 1)->hasOneUse()) |
6306 | return nullptr; |
6307 | // Extend the narrower operand to the type of the wider operand. |
6308 | CastInst::CastOps CastOpcode = |
6309 | IsSignedExt ? Instruction::SExt : Instruction::ZExt; |
6310 | if (XTy->getScalarSizeInBits() < YTy->getScalarSizeInBits()) |
6311 | X = Builder.CreateCast(Op: CastOpcode, V: X, DestTy: YTy); |
6312 | else if (YTy->getScalarSizeInBits() < XTy->getScalarSizeInBits()) |
6313 | Y = Builder.CreateCast(Op: CastOpcode, V: Y, DestTy: XTy); |
6314 | else |
6315 | return nullptr; |
6316 | } |
6317 | |
6318 | // (zext X) == (zext Y) --> X == Y |
6319 | // (sext X) == (sext Y) --> X == Y |
6320 | if (ICmp.isEquality()) |
6321 | return new ICmpInst(ICmp.getPredicate(), X, Y); |
6322 | |
6323 | // A signed comparison of sign extended values simplifies into a |
6324 | // signed comparison. |
6325 | if (IsSignedCmp && IsSignedExt) |
6326 | return new ICmpInst(ICmp.getPredicate(), X, Y); |
6327 | |
6328 | // The other three cases all fold into an unsigned comparison. |
6329 | return new ICmpInst(ICmp.getUnsignedPredicate(), X, Y); |
6330 | } |
6331 | |
6332 | // Below here, we are only folding a compare with constant. |
6333 | auto *C = dyn_cast<Constant>(Val: ICmp.getOperand(i_nocapture: 1)); |
6334 | if (!C) |
6335 | return nullptr; |
6336 | |
6337 | // If a lossless truncate is possible... |
6338 | Type *SrcTy = CastOp0->getSrcTy(); |
6339 | Constant *Res = getLosslessTrunc(C, TruncTy: SrcTy, ExtOp: CastOp0->getOpcode()); |
6340 | if (Res) { |
6341 | if (ICmp.isEquality()) |
6342 | return new ICmpInst(ICmp.getPredicate(), X, Res); |
6343 | |
6344 | // A signed comparison of sign extended values simplifies into a |
6345 | // signed comparison. |
6346 | if (IsSignedExt && IsSignedCmp) |
6347 | return new ICmpInst(ICmp.getPredicate(), X, Res); |
6348 | |
6349 | // The other three cases all fold into an unsigned comparison. |
6350 | return new ICmpInst(ICmp.getUnsignedPredicate(), X, Res); |
6351 | } |
6352 | |
6353 | // The re-extended constant changed, partly changed (in the case of a vector), |
6354 | // or could not be determined to be equal (in the case of a constant |
6355 | // expression), so the constant cannot be represented in the shorter type. |
6356 | // All the cases that fold to true or false will have already been handled |
6357 | // by simplifyICmpInst, so only deal with the tricky case. |
6358 | if (IsSignedCmp || !IsSignedExt || !isa<ConstantInt>(Val: C)) |
6359 | return nullptr; |
6360 | |
6361 | // Is source op positive? |
6362 | // icmp ult (sext X), C --> icmp sgt X, -1 |
6363 | if (ICmp.getPredicate() == ICmpInst::ICMP_ULT) |
6364 | return new ICmpInst(CmpInst::ICMP_SGT, X, Constant::getAllOnesValue(Ty: SrcTy)); |
6365 | |
6366 | // Is source op negative? |
6367 | // icmp ugt (sext X), C --> icmp slt X, 0 |
6368 | assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!" ); |
6369 | return new ICmpInst(CmpInst::ICMP_SLT, X, Constant::getNullValue(Ty: SrcTy)); |
6370 | } |
6371 | |
6372 | /// Handle icmp (cast x), (cast or constant). |
6373 | Instruction *InstCombinerImpl::foldICmpWithCastOp(ICmpInst &ICmp) { |
6374 | // If any operand of ICmp is a inttoptr roundtrip cast then remove it as |
6375 | // icmp compares only pointer's value. |
6376 | // icmp (inttoptr (ptrtoint p1)), p2 --> icmp p1, p2. |
6377 | Value *SimplifiedOp0 = simplifyIntToPtrRoundTripCast(Val: ICmp.getOperand(i_nocapture: 0)); |
6378 | Value *SimplifiedOp1 = simplifyIntToPtrRoundTripCast(Val: ICmp.getOperand(i_nocapture: 1)); |
6379 | if (SimplifiedOp0 || SimplifiedOp1) |
6380 | return new ICmpInst(ICmp.getPredicate(), |
6381 | SimplifiedOp0 ? SimplifiedOp0 : ICmp.getOperand(i_nocapture: 0), |
6382 | SimplifiedOp1 ? SimplifiedOp1 : ICmp.getOperand(i_nocapture: 1)); |
6383 | |
6384 | auto *CastOp0 = dyn_cast<CastInst>(Val: ICmp.getOperand(i_nocapture: 0)); |
6385 | if (!CastOp0) |
6386 | return nullptr; |
6387 | if (!isa<Constant>(Val: ICmp.getOperand(i_nocapture: 1)) && !isa<CastInst>(Val: ICmp.getOperand(i_nocapture: 1))) |
6388 | return nullptr; |
6389 | |
6390 | Value *Op0Src = CastOp0->getOperand(i_nocapture: 0); |
6391 | Type *SrcTy = CastOp0->getSrcTy(); |
6392 | Type *DestTy = CastOp0->getDestTy(); |
6393 | |
6394 | // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the |
6395 | // integer type is the same size as the pointer type. |
6396 | auto CompatibleSizes = [&](Type *PtrTy, Type *IntTy) { |
6397 | if (isa<VectorType>(Val: PtrTy)) { |
6398 | PtrTy = cast<VectorType>(Val: PtrTy)->getElementType(); |
6399 | IntTy = cast<VectorType>(Val: IntTy)->getElementType(); |
6400 | } |
6401 | return DL.getPointerTypeSizeInBits(PtrTy) == IntTy->getIntegerBitWidth(); |
6402 | }; |
6403 | if (CastOp0->getOpcode() == Instruction::PtrToInt && |
6404 | CompatibleSizes(SrcTy, DestTy)) { |
6405 | Value *NewOp1 = nullptr; |
6406 | if (auto *PtrToIntOp1 = dyn_cast<PtrToIntOperator>(Val: ICmp.getOperand(i_nocapture: 1))) { |
6407 | Value *PtrSrc = PtrToIntOp1->getOperand(i_nocapture: 0); |
6408 | if (PtrSrc->getType() == Op0Src->getType()) |
6409 | NewOp1 = PtrToIntOp1->getOperand(i_nocapture: 0); |
6410 | } else if (auto *RHSC = dyn_cast<Constant>(Val: ICmp.getOperand(i_nocapture: 1))) { |
6411 | NewOp1 = ConstantExpr::getIntToPtr(C: RHSC, Ty: SrcTy); |
6412 | } |
6413 | |
6414 | if (NewOp1) |
6415 | return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1); |
6416 | } |
6417 | |
6418 | // Do the same in the other direction for icmp (inttoptr x), (inttoptr/c). |
6419 | if (CastOp0->getOpcode() == Instruction::IntToPtr && |
6420 | CompatibleSizes(DestTy, SrcTy)) { |
6421 | Value *NewOp1 = nullptr; |
6422 | if (auto *IntToPtrOp1 = dyn_cast<IntToPtrInst>(Val: ICmp.getOperand(i_nocapture: 1))) { |
6423 | Value *IntSrc = IntToPtrOp1->getOperand(i_nocapture: 0); |
6424 | if (IntSrc->getType() == Op0Src->getType()) |
6425 | NewOp1 = IntToPtrOp1->getOperand(i_nocapture: 0); |
6426 | } else if (auto *RHSC = dyn_cast<Constant>(Val: ICmp.getOperand(i_nocapture: 1))) { |
6427 | NewOp1 = ConstantFoldConstant(C: ConstantExpr::getPtrToInt(C: RHSC, Ty: SrcTy), DL); |
6428 | } |
6429 | |
6430 | if (NewOp1) |
6431 | return new ICmpInst(ICmp.getPredicate(), Op0Src, NewOp1); |
6432 | } |
6433 | |
6434 | if (Instruction *R = foldICmpWithTrunc(ICmp)) |
6435 | return R; |
6436 | |
6437 | return foldICmpWithZextOrSext(ICmp); |
6438 | } |
6439 | |
6440 | static bool isNeutralValue(Instruction::BinaryOps BinaryOp, Value *RHS, |
6441 | bool IsSigned) { |
6442 | switch (BinaryOp) { |
6443 | default: |
6444 | llvm_unreachable("Unsupported binary op" ); |
6445 | case Instruction::Add: |
6446 | case Instruction::Sub: |
6447 | return match(V: RHS, P: m_Zero()); |
6448 | case Instruction::Mul: |
6449 | return !(RHS->getType()->isIntOrIntVectorTy(BitWidth: 1) && IsSigned) && |
6450 | match(V: RHS, P: m_One()); |
6451 | } |
6452 | } |
6453 | |
6454 | OverflowResult |
6455 | InstCombinerImpl::computeOverflow(Instruction::BinaryOps BinaryOp, |
6456 | bool IsSigned, Value *LHS, Value *RHS, |
6457 | Instruction *CxtI) const { |
6458 | switch (BinaryOp) { |
6459 | default: |
6460 | llvm_unreachable("Unsupported binary op" ); |
6461 | case Instruction::Add: |
6462 | if (IsSigned) |
6463 | return computeOverflowForSignedAdd(LHS, RHS, CxtI); |
6464 | else |
6465 | return computeOverflowForUnsignedAdd(LHS, RHS, CxtI); |
6466 | case Instruction::Sub: |
6467 | if (IsSigned) |
6468 | return computeOverflowForSignedSub(LHS, RHS, CxtI); |
6469 | else |
6470 | return computeOverflowForUnsignedSub(LHS, RHS, CxtI); |
6471 | case Instruction::Mul: |
6472 | if (IsSigned) |
6473 | return computeOverflowForSignedMul(LHS, RHS, CxtI); |
6474 | else |
6475 | return computeOverflowForUnsignedMul(LHS, RHS, CxtI); |
6476 | } |
6477 | } |
6478 | |
6479 | bool InstCombinerImpl::OptimizeOverflowCheck(Instruction::BinaryOps BinaryOp, |
6480 | bool IsSigned, Value *LHS, |
6481 | Value *RHS, Instruction &OrigI, |
6482 | Value *&Result, |
6483 | Constant *&Overflow) { |
6484 | if (OrigI.isCommutative() && isa<Constant>(Val: LHS) && !isa<Constant>(Val: RHS)) |
6485 | std::swap(a&: LHS, b&: RHS); |
6486 | |
6487 | // If the overflow check was an add followed by a compare, the insertion point |
6488 | // may be pointing to the compare. We want to insert the new instructions |
6489 | // before the add in case there are uses of the add between the add and the |
6490 | // compare. |
6491 | Builder.SetInsertPoint(&OrigI); |
6492 | |
6493 | Type *OverflowTy = Type::getInt1Ty(C&: LHS->getContext()); |
6494 | if (auto *LHSTy = dyn_cast<VectorType>(Val: LHS->getType())) |
6495 | OverflowTy = VectorType::get(ElementType: OverflowTy, EC: LHSTy->getElementCount()); |
6496 | |
6497 | if (isNeutralValue(BinaryOp, RHS, IsSigned)) { |
6498 | Result = LHS; |
6499 | Overflow = ConstantInt::getFalse(Ty: OverflowTy); |
6500 | return true; |
6501 | } |
6502 | |
6503 | switch (computeOverflow(BinaryOp, IsSigned, LHS, RHS, CxtI: &OrigI)) { |
6504 | case OverflowResult::MayOverflow: |
6505 | return false; |
6506 | case OverflowResult::AlwaysOverflowsLow: |
6507 | case OverflowResult::AlwaysOverflowsHigh: |
6508 | Result = Builder.CreateBinOp(Opc: BinaryOp, LHS, RHS); |
6509 | Result->takeName(V: &OrigI); |
6510 | Overflow = ConstantInt::getTrue(Ty: OverflowTy); |
6511 | return true; |
6512 | case OverflowResult::NeverOverflows: |
6513 | Result = Builder.CreateBinOp(Opc: BinaryOp, LHS, RHS); |
6514 | Result->takeName(V: &OrigI); |
6515 | Overflow = ConstantInt::getFalse(Ty: OverflowTy); |
6516 | if (auto *Inst = dyn_cast<Instruction>(Val: Result)) { |
6517 | if (IsSigned) |
6518 | Inst->setHasNoSignedWrap(); |
6519 | else |
6520 | Inst->setHasNoUnsignedWrap(); |
6521 | } |
6522 | return true; |
6523 | } |
6524 | |
6525 | llvm_unreachable("Unexpected overflow result" ); |
6526 | } |
6527 | |
6528 | /// Recognize and process idiom involving test for multiplication |
6529 | /// overflow. |
6530 | /// |
6531 | /// The caller has matched a pattern of the form: |
6532 | /// I = cmp u (mul(zext A, zext B), V |
6533 | /// The function checks if this is a test for overflow and if so replaces |
6534 | /// multiplication with call to 'mul.with.overflow' intrinsic. |
6535 | /// |
6536 | /// \param I Compare instruction. |
6537 | /// \param MulVal Result of 'mult' instruction. It is one of the arguments of |
6538 | /// the compare instruction. Must be of integer type. |
6539 | /// \param OtherVal The other argument of compare instruction. |
6540 | /// \returns Instruction which must replace the compare instruction, NULL if no |
6541 | /// replacement required. |
6542 | static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal, |
6543 | const APInt *OtherVal, |
6544 | InstCombinerImpl &IC) { |
6545 | // Don't bother doing this transformation for pointers, don't do it for |
6546 | // vectors. |
6547 | if (!isa<IntegerType>(Val: MulVal->getType())) |
6548 | return nullptr; |
6549 | |
6550 | auto *MulInstr = dyn_cast<Instruction>(Val: MulVal); |
6551 | if (!MulInstr) |
6552 | return nullptr; |
6553 | assert(MulInstr->getOpcode() == Instruction::Mul); |
6554 | |
6555 | auto *LHS = cast<ZExtInst>(Val: MulInstr->getOperand(i: 0)), |
6556 | *RHS = cast<ZExtInst>(Val: MulInstr->getOperand(i: 1)); |
6557 | assert(LHS->getOpcode() == Instruction::ZExt); |
6558 | assert(RHS->getOpcode() == Instruction::ZExt); |
6559 | Value *A = LHS->getOperand(i_nocapture: 0), *B = RHS->getOperand(i_nocapture: 0); |
6560 | |
6561 | // Calculate type and width of the result produced by mul.with.overflow. |
6562 | Type *TyA = A->getType(), *TyB = B->getType(); |
6563 | unsigned WidthA = TyA->getPrimitiveSizeInBits(), |
6564 | WidthB = TyB->getPrimitiveSizeInBits(); |
6565 | unsigned MulWidth; |
6566 | Type *MulType; |
6567 | if (WidthB > WidthA) { |
6568 | MulWidth = WidthB; |
6569 | MulType = TyB; |
6570 | } else { |
6571 | MulWidth = WidthA; |
6572 | MulType = TyA; |
6573 | } |
6574 | |
6575 | // In order to replace the original mul with a narrower mul.with.overflow, |
6576 | // all uses must ignore upper bits of the product. The number of used low |
6577 | // bits must be not greater than the width of mul.with.overflow. |
6578 | if (MulVal->hasNUsesOrMore(N: 2)) |
6579 | for (User *U : MulVal->users()) { |
6580 | if (U == &I) |
6581 | continue; |
6582 | if (TruncInst *TI = dyn_cast<TruncInst>(Val: U)) { |
6583 | // Check if truncation ignores bits above MulWidth. |
6584 | unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits(); |
6585 | if (TruncWidth > MulWidth) |
6586 | return nullptr; |
6587 | } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: U)) { |
6588 | // Check if AND ignores bits above MulWidth. |
6589 | if (BO->getOpcode() != Instruction::And) |
6590 | return nullptr; |
6591 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: BO->getOperand(i_nocapture: 1))) { |
6592 | const APInt &CVal = CI->getValue(); |
6593 | if (CVal.getBitWidth() - CVal.countl_zero() > MulWidth) |
6594 | return nullptr; |
6595 | } else { |
6596 | // In this case we could have the operand of the binary operation |
6597 | // being defined in another block, and performing the replacement |
6598 | // could break the dominance relation. |
6599 | return nullptr; |
6600 | } |
6601 | } else { |
6602 | // Other uses prohibit this transformation. |
6603 | return nullptr; |
6604 | } |
6605 | } |
6606 | |
6607 | // Recognize patterns |
6608 | switch (I.getPredicate()) { |
6609 | case ICmpInst::ICMP_UGT: { |
6610 | // Recognize pattern: |
6611 | // mulval = mul(zext A, zext B) |
6612 | // cmp ugt mulval, max |
6613 | APInt MaxVal = APInt::getMaxValue(numBits: MulWidth); |
6614 | MaxVal = MaxVal.zext(width: OtherVal->getBitWidth()); |
6615 | if (MaxVal.eq(RHS: *OtherVal)) |
6616 | break; // Recognized |
6617 | return nullptr; |
6618 | } |
6619 | |
6620 | case ICmpInst::ICMP_ULT: { |
6621 | // Recognize pattern: |
6622 | // mulval = mul(zext A, zext B) |
6623 | // cmp ule mulval, max + 1 |
6624 | APInt MaxVal = APInt::getOneBitSet(numBits: OtherVal->getBitWidth(), BitNo: MulWidth); |
6625 | if (MaxVal.eq(RHS: *OtherVal)) |
6626 | break; // Recognized |
6627 | return nullptr; |
6628 | } |
6629 | |
6630 | default: |
6631 | return nullptr; |
6632 | } |
6633 | |
6634 | InstCombiner::BuilderTy &Builder = IC.Builder; |
6635 | Builder.SetInsertPoint(MulInstr); |
6636 | |
6637 | // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B) |
6638 | Value *MulA = A, *MulB = B; |
6639 | if (WidthA < MulWidth) |
6640 | MulA = Builder.CreateZExt(V: A, DestTy: MulType); |
6641 | if (WidthB < MulWidth) |
6642 | MulB = Builder.CreateZExt(V: B, DestTy: MulType); |
6643 | CallInst *Call = |
6644 | Builder.CreateIntrinsic(ID: Intrinsic::umul_with_overflow, Types: MulType, |
6645 | Args: {MulA, MulB}, /*FMFSource=*/nullptr, Name: "umul" ); |
6646 | IC.addToWorklist(I: MulInstr); |
6647 | |
6648 | // If there are uses of mul result other than the comparison, we know that |
6649 | // they are truncation or binary AND. Change them to use result of |
6650 | // mul.with.overflow and adjust properly mask/size. |
6651 | if (MulVal->hasNUsesOrMore(N: 2)) { |
6652 | Value *Mul = Builder.CreateExtractValue(Agg: Call, Idxs: 0, Name: "umul.value" ); |
6653 | for (User *U : make_early_inc_range(Range: MulVal->users())) { |
6654 | if (U == &I) |
6655 | continue; |
6656 | if (TruncInst *TI = dyn_cast<TruncInst>(Val: U)) { |
6657 | if (TI->getType()->getPrimitiveSizeInBits() == MulWidth) |
6658 | IC.replaceInstUsesWith(I&: *TI, V: Mul); |
6659 | else |
6660 | TI->setOperand(i_nocapture: 0, Val_nocapture: Mul); |
6661 | } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: U)) { |
6662 | assert(BO->getOpcode() == Instruction::And); |
6663 | // Replace (mul & mask) --> zext (mul.with.overflow & short_mask) |
6664 | ConstantInt *CI = cast<ConstantInt>(Val: BO->getOperand(i_nocapture: 1)); |
6665 | APInt ShortMask = CI->getValue().trunc(width: MulWidth); |
6666 | Value *ShortAnd = Builder.CreateAnd(LHS: Mul, RHS: ShortMask); |
6667 | Value *Zext = Builder.CreateZExt(V: ShortAnd, DestTy: BO->getType()); |
6668 | IC.replaceInstUsesWith(I&: *BO, V: Zext); |
6669 | } else { |
6670 | llvm_unreachable("Unexpected Binary operation" ); |
6671 | } |
6672 | IC.addToWorklist(I: cast<Instruction>(Val: U)); |
6673 | } |
6674 | } |
6675 | |
6676 | // The original icmp gets replaced with the overflow value, maybe inverted |
6677 | // depending on predicate. |
6678 | if (I.getPredicate() == ICmpInst::ICMP_ULT) { |
6679 | Value *Res = Builder.CreateExtractValue(Agg: Call, Idxs: 1); |
6680 | return BinaryOperator::CreateNot(Op: Res); |
6681 | } |
6682 | |
6683 | return ExtractValueInst::Create(Agg: Call, Idxs: 1); |
6684 | } |
6685 | |
6686 | /// When performing a comparison against a constant, it is possible that not all |
6687 | /// the bits in the LHS are demanded. This helper method computes the mask that |
6688 | /// IS demanded. |
6689 | static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth) { |
6690 | const APInt *RHS; |
6691 | if (!match(V: I.getOperand(i_nocapture: 1), P: m_APInt(Res&: RHS))) |
6692 | return APInt::getAllOnes(numBits: BitWidth); |
6693 | |
6694 | // If this is a normal comparison, it demands all bits. If it is a sign bit |
6695 | // comparison, it only demands the sign bit. |
6696 | bool UnusedBit; |
6697 | if (isSignBitCheck(Pred: I.getPredicate(), RHS: *RHS, TrueIfSigned&: UnusedBit)) |
6698 | return APInt::getSignMask(BitWidth); |
6699 | |
6700 | switch (I.getPredicate()) { |
6701 | // For a UGT comparison, we don't care about any bits that |
6702 | // correspond to the trailing ones of the comparand. The value of these |
6703 | // bits doesn't impact the outcome of the comparison, because any value |
6704 | // greater than the RHS must differ in a bit higher than these due to carry. |
6705 | case ICmpInst::ICMP_UGT: |
6706 | return APInt::getBitsSetFrom(numBits: BitWidth, loBit: RHS->countr_one()); |
6707 | |
6708 | // Similarly, for a ULT comparison, we don't care about the trailing zeros. |
6709 | // Any value less than the RHS must differ in a higher bit because of carries. |
6710 | case ICmpInst::ICMP_ULT: |
6711 | return APInt::getBitsSetFrom(numBits: BitWidth, loBit: RHS->countr_zero()); |
6712 | |
6713 | default: |
6714 | return APInt::getAllOnes(numBits: BitWidth); |
6715 | } |
6716 | } |
6717 | |
6718 | /// Check that one use is in the same block as the definition and all |
6719 | /// other uses are in blocks dominated by a given block. |
6720 | /// |
6721 | /// \param DI Definition |
6722 | /// \param UI Use |
6723 | /// \param DB Block that must dominate all uses of \p DI outside |
6724 | /// the parent block |
6725 | /// \return true when \p UI is the only use of \p DI in the parent block |
6726 | /// and all other uses of \p DI are in blocks dominated by \p DB. |
6727 | /// |
6728 | bool InstCombinerImpl::dominatesAllUses(const Instruction *DI, |
6729 | const Instruction *UI, |
6730 | const BasicBlock *DB) const { |
6731 | assert(DI && UI && "Instruction not defined\n" ); |
6732 | // Ignore incomplete definitions. |
6733 | if (!DI->getParent()) |
6734 | return false; |
6735 | // DI and UI must be in the same block. |
6736 | if (DI->getParent() != UI->getParent()) |
6737 | return false; |
6738 | // Protect from self-referencing blocks. |
6739 | if (DI->getParent() == DB) |
6740 | return false; |
6741 | for (const User *U : DI->users()) { |
6742 | auto *Usr = cast<Instruction>(Val: U); |
6743 | if (Usr != UI && !DT.dominates(A: DB, B: Usr->getParent())) |
6744 | return false; |
6745 | } |
6746 | return true; |
6747 | } |
6748 | |
6749 | /// Return true when the instruction sequence within a block is select-cmp-br. |
6750 | static bool isChainSelectCmpBranch(const SelectInst *SI) { |
6751 | const BasicBlock *BB = SI->getParent(); |
6752 | if (!BB) |
6753 | return false; |
6754 | auto *BI = dyn_cast_or_null<BranchInst>(Val: BB->getTerminator()); |
6755 | if (!BI || BI->getNumSuccessors() != 2) |
6756 | return false; |
6757 | auto *IC = dyn_cast<ICmpInst>(Val: BI->getCondition()); |
6758 | if (!IC || (IC->getOperand(i_nocapture: 0) != SI && IC->getOperand(i_nocapture: 1) != SI)) |
6759 | return false; |
6760 | return true; |
6761 | } |
6762 | |
6763 | /// True when a select result is replaced by one of its operands |
6764 | /// in select-icmp sequence. This will eventually result in the elimination |
6765 | /// of the select. |
6766 | /// |
6767 | /// \param SI Select instruction |
6768 | /// \param Icmp Compare instruction |
6769 | /// \param SIOpd Operand that replaces the select |
6770 | /// |
6771 | /// Notes: |
6772 | /// - The replacement is global and requires dominator information |
6773 | /// - The caller is responsible for the actual replacement |
6774 | /// |
6775 | /// Example: |
6776 | /// |
6777 | /// entry: |
6778 | /// %4 = select i1 %3, %C* %0, %C* null |
6779 | /// %5 = icmp eq %C* %4, null |
6780 | /// br i1 %5, label %9, label %7 |
6781 | /// ... |
6782 | /// ; <label>:7 ; preds = %entry |
6783 | /// %8 = getelementptr inbounds %C* %4, i64 0, i32 0 |
6784 | /// ... |
6785 | /// |
6786 | /// can be transformed to |
6787 | /// |
6788 | /// %5 = icmp eq %C* %0, null |
6789 | /// %6 = select i1 %3, i1 %5, i1 true |
6790 | /// br i1 %6, label %9, label %7 |
6791 | /// ... |
6792 | /// ; <label>:7 ; preds = %entry |
6793 | /// %8 = getelementptr inbounds %C* %0, i64 0, i32 0 // replace by %0! |
6794 | /// |
6795 | /// Similar when the first operand of the select is a constant or/and |
6796 | /// the compare is for not equal rather than equal. |
6797 | /// |
6798 | /// NOTE: The function is only called when the select and compare constants |
6799 | /// are equal, the optimization can work only for EQ predicates. This is not a |
6800 | /// major restriction since a NE compare should be 'normalized' to an equal |
6801 | /// compare, which usually happens in the combiner and test case |
6802 | /// select-cmp-br.ll checks for it. |
6803 | bool InstCombinerImpl::replacedSelectWithOperand(SelectInst *SI, |
6804 | const ICmpInst *Icmp, |
6805 | const unsigned SIOpd) { |
6806 | assert((SIOpd == 1 || SIOpd == 2) && "Invalid select operand!" ); |
6807 | if (isChainSelectCmpBranch(SI) && Icmp->getPredicate() == ICmpInst::ICMP_EQ) { |
6808 | BasicBlock *Succ = SI->getParent()->getTerminator()->getSuccessor(Idx: 1); |
6809 | // The check for the single predecessor is not the best that can be |
6810 | // done. But it protects efficiently against cases like when SI's |
6811 | // home block has two successors, Succ and Succ1, and Succ1 predecessor |
6812 | // of Succ. Then SI can't be replaced by SIOpd because the use that gets |
6813 | // replaced can be reached on either path. So the uniqueness check |
6814 | // guarantees that the path all uses of SI (outside SI's parent) are on |
6815 | // is disjoint from all other paths out of SI. But that information |
6816 | // is more expensive to compute, and the trade-off here is in favor |
6817 | // of compile-time. It should also be noticed that we check for a single |
6818 | // predecessor and not only uniqueness. This to handle the situation when |
6819 | // Succ and Succ1 points to the same basic block. |
6820 | if (Succ->getSinglePredecessor() && dominatesAllUses(DI: SI, UI: Icmp, DB: Succ)) { |
6821 | NumSel++; |
6822 | SI->replaceUsesOutsideBlock(V: SI->getOperand(i_nocapture: SIOpd), BB: SI->getParent()); |
6823 | return true; |
6824 | } |
6825 | } |
6826 | return false; |
6827 | } |
6828 | |
6829 | /// Try to fold the comparison based on range information we can get by checking |
6830 | /// whether bits are known to be zero or one in the inputs. |
6831 | Instruction *InstCombinerImpl::foldICmpUsingKnownBits(ICmpInst &I) { |
6832 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
6833 | Type *Ty = Op0->getType(); |
6834 | ICmpInst::Predicate Pred = I.getPredicate(); |
6835 | |
6836 | // Get scalar or pointer size. |
6837 | unsigned BitWidth = Ty->isIntOrIntVectorTy() |
6838 | ? Ty->getScalarSizeInBits() |
6839 | : DL.getPointerTypeSizeInBits(Ty->getScalarType()); |
6840 | |
6841 | if (!BitWidth) |
6842 | return nullptr; |
6843 | |
6844 | KnownBits Op0Known(BitWidth); |
6845 | KnownBits Op1Known(BitWidth); |
6846 | |
6847 | { |
6848 | // Don't use dominating conditions when folding icmp using known bits. This |
6849 | // may convert signed into unsigned predicates in ways that other passes |
6850 | // (especially IndVarSimplify) may not be able to reliably undo. |
6851 | SimplifyQuery Q = SQ.getWithoutDomCondCache().getWithInstruction(I: &I); |
6852 | if (SimplifyDemandedBits(I: &I, Op: 0, DemandedMask: getDemandedBitsLHSMask(I, BitWidth), |
6853 | Known&: Op0Known, Q)) |
6854 | return &I; |
6855 | |
6856 | if (SimplifyDemandedBits(I: &I, Op: 1, DemandedMask: APInt::getAllOnes(numBits: BitWidth), Known&: Op1Known, Q)) |
6857 | return &I; |
6858 | } |
6859 | |
6860 | if (!isa<Constant>(Val: Op0) && Op0Known.isConstant()) |
6861 | return new ICmpInst( |
6862 | Pred, ConstantExpr::getIntegerValue(Ty, V: Op0Known.getConstant()), Op1); |
6863 | if (!isa<Constant>(Val: Op1) && Op1Known.isConstant()) |
6864 | return new ICmpInst( |
6865 | Pred, Op0, ConstantExpr::getIntegerValue(Ty, V: Op1Known.getConstant())); |
6866 | |
6867 | if (std::optional<bool> Res = ICmpInst::compare(LHS: Op0Known, RHS: Op1Known, Pred)) |
6868 | return replaceInstUsesWith(I, V: ConstantInt::getBool(Ty: I.getType(), V: *Res)); |
6869 | |
6870 | // Given the known and unknown bits, compute a range that the LHS could be |
6871 | // in. Compute the Min, Max and RHS values based on the known bits. For the |
6872 | // EQ and NE we use unsigned values. |
6873 | APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0); |
6874 | APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0); |
6875 | if (I.isSigned()) { |
6876 | Op0Min = Op0Known.getSignedMinValue(); |
6877 | Op0Max = Op0Known.getSignedMaxValue(); |
6878 | Op1Min = Op1Known.getSignedMinValue(); |
6879 | Op1Max = Op1Known.getSignedMaxValue(); |
6880 | } else { |
6881 | Op0Min = Op0Known.getMinValue(); |
6882 | Op0Max = Op0Known.getMaxValue(); |
6883 | Op1Min = Op1Known.getMinValue(); |
6884 | Op1Max = Op1Known.getMaxValue(); |
6885 | } |
6886 | |
6887 | // Don't break up a clamp pattern -- (min(max X, Y), Z) -- by replacing a |
6888 | // min/max canonical compare with some other compare. That could lead to |
6889 | // conflict with select canonicalization and infinite looping. |
6890 | // FIXME: This constraint may go away if min/max intrinsics are canonical. |
6891 | auto isMinMaxCmp = [&](Instruction &Cmp) { |
6892 | if (!Cmp.hasOneUse()) |
6893 | return false; |
6894 | Value *A, *B; |
6895 | SelectPatternFlavor SPF = matchSelectPattern(V: Cmp.user_back(), LHS&: A, RHS&: B).Flavor; |
6896 | if (!SelectPatternResult::isMinOrMax(SPF)) |
6897 | return false; |
6898 | return match(V: Op0, P: m_MaxOrMin(L: m_Value(), R: m_Value())) || |
6899 | match(V: Op1, P: m_MaxOrMin(L: m_Value(), R: m_Value())); |
6900 | }; |
6901 | if (!isMinMaxCmp(I)) { |
6902 | switch (Pred) { |
6903 | default: |
6904 | break; |
6905 | case ICmpInst::ICMP_ULT: { |
6906 | if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B) |
6907 | return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); |
6908 | const APInt *CmpC; |
6909 | if (match(V: Op1, P: m_APInt(Res&: CmpC))) { |
6910 | // A <u C -> A == C-1 if min(A)+1 == C |
6911 | if (*CmpC == Op0Min + 1) |
6912 | return new ICmpInst(ICmpInst::ICMP_EQ, Op0, |
6913 | ConstantInt::get(Ty: Op1->getType(), V: *CmpC - 1)); |
6914 | // X <u C --> X == 0, if the number of zero bits in the bottom of X |
6915 | // exceeds the log2 of C. |
6916 | if (Op0Known.countMinTrailingZeros() >= CmpC->ceilLogBase2()) |
6917 | return new ICmpInst(ICmpInst::ICMP_EQ, Op0, |
6918 | Constant::getNullValue(Ty: Op1->getType())); |
6919 | } |
6920 | break; |
6921 | } |
6922 | case ICmpInst::ICMP_UGT: { |
6923 | if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B) |
6924 | return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); |
6925 | const APInt *CmpC; |
6926 | if (match(V: Op1, P: m_APInt(Res&: CmpC))) { |
6927 | // A >u C -> A == C+1 if max(a)-1 == C |
6928 | if (*CmpC == Op0Max - 1) |
6929 | return new ICmpInst(ICmpInst::ICMP_EQ, Op0, |
6930 | ConstantInt::get(Ty: Op1->getType(), V: *CmpC + 1)); |
6931 | // X >u C --> X != 0, if the number of zero bits in the bottom of X |
6932 | // exceeds the log2 of C. |
6933 | if (Op0Known.countMinTrailingZeros() >= CmpC->getActiveBits()) |
6934 | return new ICmpInst(ICmpInst::ICMP_NE, Op0, |
6935 | Constant::getNullValue(Ty: Op1->getType())); |
6936 | } |
6937 | break; |
6938 | } |
6939 | case ICmpInst::ICMP_SLT: { |
6940 | if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B) |
6941 | return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); |
6942 | const APInt *CmpC; |
6943 | if (match(V: Op1, P: m_APInt(Res&: CmpC))) { |
6944 | if (*CmpC == Op0Min + 1) // A <s C -> A == C-1 if min(A)+1 == C |
6945 | return new ICmpInst(ICmpInst::ICMP_EQ, Op0, |
6946 | ConstantInt::get(Ty: Op1->getType(), V: *CmpC - 1)); |
6947 | } |
6948 | break; |
6949 | } |
6950 | case ICmpInst::ICMP_SGT: { |
6951 | if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B) |
6952 | return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1); |
6953 | const APInt *CmpC; |
6954 | if (match(V: Op1, P: m_APInt(Res&: CmpC))) { |
6955 | if (*CmpC == Op0Max - 1) // A >s C -> A == C+1 if max(A)-1 == C |
6956 | return new ICmpInst(ICmpInst::ICMP_EQ, Op0, |
6957 | ConstantInt::get(Ty: Op1->getType(), V: *CmpC + 1)); |
6958 | } |
6959 | break; |
6960 | } |
6961 | } |
6962 | } |
6963 | |
6964 | // Based on the range information we know about the LHS, see if we can |
6965 | // simplify this comparison. For example, (x&4) < 8 is always true. |
6966 | switch (Pred) { |
6967 | default: |
6968 | break; |
6969 | case ICmpInst::ICMP_EQ: |
6970 | case ICmpInst::ICMP_NE: { |
6971 | // If all bits are known zero except for one, then we know at most one bit |
6972 | // is set. If the comparison is against zero, then this is a check to see if |
6973 | // *that* bit is set. |
6974 | APInt Op0KnownZeroInverted = ~Op0Known.Zero; |
6975 | if (Op1Known.isZero()) { |
6976 | // If the LHS is an AND with the same constant, look through it. |
6977 | Value *LHS = nullptr; |
6978 | const APInt *LHSC; |
6979 | if (!match(V: Op0, P: m_And(L: m_Value(V&: LHS), R: m_APInt(Res&: LHSC))) || |
6980 | *LHSC != Op0KnownZeroInverted) |
6981 | LHS = Op0; |
6982 | |
6983 | Value *X; |
6984 | const APInt *C1; |
6985 | if (match(V: LHS, P: m_Shl(L: m_Power2(V&: C1), R: m_Value(V&: X)))) { |
6986 | Type *XTy = X->getType(); |
6987 | unsigned Log2C1 = C1->countr_zero(); |
6988 | APInt C2 = Op0KnownZeroInverted; |
6989 | APInt C2Pow2 = (C2 & ~(*C1 - 1)) + *C1; |
6990 | if (C2Pow2.isPowerOf2()) { |
6991 | // iff (C1 is pow2) & ((C2 & ~(C1-1)) + C1) is pow2): |
6992 | // ((C1 << X) & C2) == 0 -> X >= (Log2(C2+C1) - Log2(C1)) |
6993 | // ((C1 << X) & C2) != 0 -> X < (Log2(C2+C1) - Log2(C1)) |
6994 | unsigned Log2C2 = C2Pow2.countr_zero(); |
6995 | auto *CmpC = ConstantInt::get(Ty: XTy, V: Log2C2 - Log2C1); |
6996 | auto NewPred = |
6997 | Pred == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGE : CmpInst::ICMP_ULT; |
6998 | return new ICmpInst(NewPred, X, CmpC); |
6999 | } |
7000 | } |
7001 | } |
7002 | |
7003 | // Op0 eq C_Pow2 -> Op0 ne 0 if Op0 is known to be C_Pow2 or zero. |
7004 | if (Op1Known.isConstant() && Op1Known.getConstant().isPowerOf2() && |
7005 | (Op0Known & Op1Known) == Op0Known) |
7006 | return new ICmpInst(CmpInst::getInversePredicate(pred: Pred), Op0, |
7007 | ConstantInt::getNullValue(Ty: Op1->getType())); |
7008 | break; |
7009 | } |
7010 | case ICmpInst::ICMP_SGE: |
7011 | if (Op1Min == Op0Max) // A >=s B -> A == B if max(A) == min(B) |
7012 | return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1); |
7013 | break; |
7014 | case ICmpInst::ICMP_SLE: |
7015 | if (Op1Max == Op0Min) // A <=s B -> A == B if min(A) == max(B) |
7016 | return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1); |
7017 | break; |
7018 | case ICmpInst::ICMP_UGE: |
7019 | if (Op1Min == Op0Max) // A >=u B -> A == B if max(A) == min(B) |
7020 | return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1); |
7021 | break; |
7022 | case ICmpInst::ICMP_ULE: |
7023 | if (Op1Max == Op0Min) // A <=u B -> A == B if min(A) == max(B) |
7024 | return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1); |
7025 | break; |
7026 | } |
7027 | |
7028 | // Turn a signed comparison into an unsigned one if both operands are known to |
7029 | // have the same sign. Set samesign if possible (except for equality |
7030 | // predicates). |
7031 | if ((I.isSigned() || (I.isUnsigned() && !I.hasSameSign())) && |
7032 | ((Op0Known.Zero.isNegative() && Op1Known.Zero.isNegative()) || |
7033 | (Op0Known.One.isNegative() && Op1Known.One.isNegative()))) { |
7034 | I.setPredicate(I.getUnsignedPredicate()); |
7035 | I.setSameSign(); |
7036 | return &I; |
7037 | } |
7038 | |
7039 | return nullptr; |
7040 | } |
7041 | |
7042 | /// If one operand of an icmp is effectively a bool (value range of {0,1}), |
7043 | /// then try to reduce patterns based on that limit. |
7044 | Instruction *InstCombinerImpl::foldICmpUsingBoolRange(ICmpInst &I) { |
7045 | Value *X, *Y; |
7046 | CmpPredicate Pred; |
7047 | |
7048 | // X must be 0 and bool must be true for "ULT": |
7049 | // X <u (zext i1 Y) --> (X == 0) & Y |
7050 | if (match(V: &I, P: m_c_ICmp(Pred, L: m_Value(V&: X), R: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: Y))))) && |
7051 | Y->getType()->isIntOrIntVectorTy(BitWidth: 1) && Pred == ICmpInst::ICMP_ULT) |
7052 | return BinaryOperator::CreateAnd(V1: Builder.CreateIsNull(Arg: X), V2: Y); |
7053 | |
7054 | // X must be 0 or bool must be true for "ULE": |
7055 | // X <=u (sext i1 Y) --> (X == 0) | Y |
7056 | if (match(V: &I, P: m_c_ICmp(Pred, L: m_Value(V&: X), R: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: Y))))) && |
7057 | Y->getType()->isIntOrIntVectorTy(BitWidth: 1) && Pred == ICmpInst::ICMP_ULE) |
7058 | return BinaryOperator::CreateOr(V1: Builder.CreateIsNull(Arg: X), V2: Y); |
7059 | |
7060 | // icmp eq/ne X, (zext/sext (icmp eq/ne X, C)) |
7061 | CmpPredicate Pred1, Pred2; |
7062 | const APInt *C; |
7063 | Instruction *ExtI; |
7064 | if (match(V: &I, P: m_c_ICmp(Pred&: Pred1, L: m_Value(V&: X), |
7065 | R: m_CombineAnd(L: m_Instruction(I&: ExtI), |
7066 | R: m_ZExtOrSExt(Op: m_ICmp(Pred&: Pred2, L: m_Deferred(V: X), |
7067 | R: m_APInt(Res&: C)))))) && |
7068 | ICmpInst::isEquality(P: Pred1) && ICmpInst::isEquality(P: Pred2)) { |
7069 | bool IsSExt = ExtI->getOpcode() == Instruction::SExt; |
7070 | bool HasOneUse = ExtI->hasOneUse() && ExtI->getOperand(i: 0)->hasOneUse(); |
7071 | auto CreateRangeCheck = [&] { |
7072 | Value *CmpV1 = |
7073 | Builder.CreateICmp(P: Pred1, LHS: X, RHS: Constant::getNullValue(Ty: X->getType())); |
7074 | Value *CmpV2 = Builder.CreateICmp( |
7075 | P: Pred1, LHS: X, RHS: ConstantInt::getSigned(Ty: X->getType(), V: IsSExt ? -1 : 1)); |
7076 | return BinaryOperator::Create( |
7077 | Op: Pred1 == ICmpInst::ICMP_EQ ? Instruction::Or : Instruction::And, |
7078 | S1: CmpV1, S2: CmpV2); |
7079 | }; |
7080 | if (C->isZero()) { |
7081 | if (Pred2 == ICmpInst::ICMP_EQ) { |
7082 | // icmp eq X, (zext/sext (icmp eq X, 0)) --> false |
7083 | // icmp ne X, (zext/sext (icmp eq X, 0)) --> true |
7084 | return replaceInstUsesWith( |
7085 | I, V: ConstantInt::getBool(Ty: I.getType(), V: Pred1 == ICmpInst::ICMP_NE)); |
7086 | } else if (!IsSExt || HasOneUse) { |
7087 | // icmp eq X, (zext (icmp ne X, 0)) --> X == 0 || X == 1 |
7088 | // icmp ne X, (zext (icmp ne X, 0)) --> X != 0 && X != 1 |
7089 | // icmp eq X, (sext (icmp ne X, 0)) --> X == 0 || X == -1 |
7090 | // icmp ne X, (sext (icmp ne X, 0)) --> X != 0 && X != -1 |
7091 | return CreateRangeCheck(); |
7092 | } |
7093 | } else if (IsSExt ? C->isAllOnes() : C->isOne()) { |
7094 | if (Pred2 == ICmpInst::ICMP_NE) { |
7095 | // icmp eq X, (zext (icmp ne X, 1)) --> false |
7096 | // icmp ne X, (zext (icmp ne X, 1)) --> true |
7097 | // icmp eq X, (sext (icmp ne X, -1)) --> false |
7098 | // icmp ne X, (sext (icmp ne X, -1)) --> true |
7099 | return replaceInstUsesWith( |
7100 | I, V: ConstantInt::getBool(Ty: I.getType(), V: Pred1 == ICmpInst::ICMP_NE)); |
7101 | } else if (!IsSExt || HasOneUse) { |
7102 | // icmp eq X, (zext (icmp eq X, 1)) --> X == 0 || X == 1 |
7103 | // icmp ne X, (zext (icmp eq X, 1)) --> X != 0 && X != 1 |
7104 | // icmp eq X, (sext (icmp eq X, -1)) --> X == 0 || X == -1 |
7105 | // icmp ne X, (sext (icmp eq X, -1)) --> X != 0 && X == -1 |
7106 | return CreateRangeCheck(); |
7107 | } |
7108 | } else { |
7109 | // when C != 0 && C != 1: |
7110 | // icmp eq X, (zext (icmp eq X, C)) --> icmp eq X, 0 |
7111 | // icmp eq X, (zext (icmp ne X, C)) --> icmp eq X, 1 |
7112 | // icmp ne X, (zext (icmp eq X, C)) --> icmp ne X, 0 |
7113 | // icmp ne X, (zext (icmp ne X, C)) --> icmp ne X, 1 |
7114 | // when C != 0 && C != -1: |
7115 | // icmp eq X, (sext (icmp eq X, C)) --> icmp eq X, 0 |
7116 | // icmp eq X, (sext (icmp ne X, C)) --> icmp eq X, -1 |
7117 | // icmp ne X, (sext (icmp eq X, C)) --> icmp ne X, 0 |
7118 | // icmp ne X, (sext (icmp ne X, C)) --> icmp ne X, -1 |
7119 | return ICmpInst::Create( |
7120 | Op: Instruction::ICmp, Pred: Pred1, S1: X, |
7121 | S2: ConstantInt::getSigned(Ty: X->getType(), V: Pred2 == ICmpInst::ICMP_NE |
7122 | ? (IsSExt ? -1 : 1) |
7123 | : 0)); |
7124 | } |
7125 | } |
7126 | |
7127 | return nullptr; |
7128 | } |
7129 | |
7130 | /// If we have an icmp le or icmp ge instruction with a constant operand, turn |
7131 | /// it into the appropriate icmp lt or icmp gt instruction. This transform |
7132 | /// allows them to be folded in visitICmpInst. |
7133 | static ICmpInst *canonicalizeCmpWithConstant(ICmpInst &I) { |
7134 | ICmpInst::Predicate Pred = I.getPredicate(); |
7135 | if (ICmpInst::isEquality(P: Pred) || !ICmpInst::isIntPredicate(P: Pred) || |
7136 | InstCombiner::isCanonicalPredicate(Pred)) |
7137 | return nullptr; |
7138 | |
7139 | Value *Op0 = I.getOperand(i_nocapture: 0); |
7140 | Value *Op1 = I.getOperand(i_nocapture: 1); |
7141 | auto *Op1C = dyn_cast<Constant>(Val: Op1); |
7142 | if (!Op1C) |
7143 | return nullptr; |
7144 | |
7145 | auto FlippedStrictness = getFlippedStrictnessPredicateAndConstant(Pred, C: Op1C); |
7146 | if (!FlippedStrictness) |
7147 | return nullptr; |
7148 | |
7149 | return new ICmpInst(FlippedStrictness->first, Op0, FlippedStrictness->second); |
7150 | } |
7151 | |
7152 | /// If we have a comparison with a non-canonical predicate, if we can update |
7153 | /// all the users, invert the predicate and adjust all the users. |
7154 | CmpInst *InstCombinerImpl::canonicalizeICmpPredicate(CmpInst &I) { |
7155 | // Is the predicate already canonical? |
7156 | CmpInst::Predicate Pred = I.getPredicate(); |
7157 | if (InstCombiner::isCanonicalPredicate(Pred)) |
7158 | return nullptr; |
7159 | |
7160 | // Can all users be adjusted to predicate inversion? |
7161 | if (!InstCombiner::canFreelyInvertAllUsersOf(V: &I, /*IgnoredUser=*/nullptr)) |
7162 | return nullptr; |
7163 | |
7164 | // Ok, we can canonicalize comparison! |
7165 | // Let's first invert the comparison's predicate. |
7166 | I.setPredicate(CmpInst::getInversePredicate(pred: Pred)); |
7167 | I.setName(I.getName() + ".not" ); |
7168 | |
7169 | // And, adapt users. |
7170 | freelyInvertAllUsersOf(V: &I); |
7171 | |
7172 | return &I; |
7173 | } |
7174 | |
7175 | /// Integer compare with boolean values can always be turned into bitwise ops. |
7176 | static Instruction *canonicalizeICmpBool(ICmpInst &I, |
7177 | InstCombiner::BuilderTy &Builder) { |
7178 | Value *A = I.getOperand(i_nocapture: 0), *B = I.getOperand(i_nocapture: 1); |
7179 | assert(A->getType()->isIntOrIntVectorTy(1) && "Bools only" ); |
7180 | |
7181 | // A boolean compared to true/false can be simplified to Op0/true/false in |
7182 | // 14 out of the 20 (10 predicates * 2 constants) possible combinations. |
7183 | // Cases not handled by InstSimplify are always 'not' of Op0. |
7184 | if (match(V: B, P: m_Zero())) { |
7185 | switch (I.getPredicate()) { |
7186 | case CmpInst::ICMP_EQ: // A == 0 -> !A |
7187 | case CmpInst::ICMP_ULE: // A <=u 0 -> !A |
7188 | case CmpInst::ICMP_SGE: // A >=s 0 -> !A |
7189 | return BinaryOperator::CreateNot(Op: A); |
7190 | default: |
7191 | llvm_unreachable("ICmp i1 X, C not simplified as expected." ); |
7192 | } |
7193 | } else if (match(V: B, P: m_One())) { |
7194 | switch (I.getPredicate()) { |
7195 | case CmpInst::ICMP_NE: // A != 1 -> !A |
7196 | case CmpInst::ICMP_ULT: // A <u 1 -> !A |
7197 | case CmpInst::ICMP_SGT: // A >s -1 -> !A |
7198 | return BinaryOperator::CreateNot(Op: A); |
7199 | default: |
7200 | llvm_unreachable("ICmp i1 X, C not simplified as expected." ); |
7201 | } |
7202 | } |
7203 | |
7204 | switch (I.getPredicate()) { |
7205 | default: |
7206 | llvm_unreachable("Invalid icmp instruction!" ); |
7207 | case ICmpInst::ICMP_EQ: |
7208 | // icmp eq i1 A, B -> ~(A ^ B) |
7209 | return BinaryOperator::CreateNot(Op: Builder.CreateXor(LHS: A, RHS: B)); |
7210 | |
7211 | case ICmpInst::ICMP_NE: |
7212 | // icmp ne i1 A, B -> A ^ B |
7213 | return BinaryOperator::CreateXor(V1: A, V2: B); |
7214 | |
7215 | case ICmpInst::ICMP_UGT: |
7216 | // icmp ugt -> icmp ult |
7217 | std::swap(a&: A, b&: B); |
7218 | [[fallthrough]]; |
7219 | case ICmpInst::ICMP_ULT: |
7220 | // icmp ult i1 A, B -> ~A & B |
7221 | return BinaryOperator::CreateAnd(V1: Builder.CreateNot(V: A), V2: B); |
7222 | |
7223 | case ICmpInst::ICMP_SGT: |
7224 | // icmp sgt -> icmp slt |
7225 | std::swap(a&: A, b&: B); |
7226 | [[fallthrough]]; |
7227 | case ICmpInst::ICMP_SLT: |
7228 | // icmp slt i1 A, B -> A & ~B |
7229 | return BinaryOperator::CreateAnd(V1: Builder.CreateNot(V: B), V2: A); |
7230 | |
7231 | case ICmpInst::ICMP_UGE: |
7232 | // icmp uge -> icmp ule |
7233 | std::swap(a&: A, b&: B); |
7234 | [[fallthrough]]; |
7235 | case ICmpInst::ICMP_ULE: |
7236 | // icmp ule i1 A, B -> ~A | B |
7237 | return BinaryOperator::CreateOr(V1: Builder.CreateNot(V: A), V2: B); |
7238 | |
7239 | case ICmpInst::ICMP_SGE: |
7240 | // icmp sge -> icmp sle |
7241 | std::swap(a&: A, b&: B); |
7242 | [[fallthrough]]; |
7243 | case ICmpInst::ICMP_SLE: |
7244 | // icmp sle i1 A, B -> A | ~B |
7245 | return BinaryOperator::CreateOr(V1: Builder.CreateNot(V: B), V2: A); |
7246 | } |
7247 | } |
7248 | |
7249 | // Transform pattern like: |
7250 | // (1 << Y) u<= X or ~(-1 << Y) u< X or ((1 << Y)+(-1)) u< X |
7251 | // (1 << Y) u> X or ~(-1 << Y) u>= X or ((1 << Y)+(-1)) u>= X |
7252 | // Into: |
7253 | // (X l>> Y) != 0 |
7254 | // (X l>> Y) == 0 |
7255 | static Instruction *foldICmpWithHighBitMask(ICmpInst &Cmp, |
7256 | InstCombiner::BuilderTy &Builder) { |
7257 | CmpPredicate Pred, NewPred; |
7258 | Value *X, *Y; |
7259 | if (match(V: &Cmp, |
7260 | P: m_c_ICmp(Pred, L: m_OneUse(SubPattern: m_Shl(L: m_One(), R: m_Value(V&: Y))), R: m_Value(V&: X)))) { |
7261 | switch (Pred) { |
7262 | case ICmpInst::ICMP_ULE: |
7263 | NewPred = ICmpInst::ICMP_NE; |
7264 | break; |
7265 | case ICmpInst::ICMP_UGT: |
7266 | NewPred = ICmpInst::ICMP_EQ; |
7267 | break; |
7268 | default: |
7269 | return nullptr; |
7270 | } |
7271 | } else if (match(V: &Cmp, P: m_c_ICmp(Pred, |
7272 | L: m_OneUse(SubPattern: m_CombineOr( |
7273 | L: m_Not(V: m_Shl(L: m_AllOnes(), R: m_Value(V&: Y))), |
7274 | R: m_Add(L: m_Shl(L: m_One(), R: m_Value(V&: Y)), |
7275 | R: m_AllOnes()))), |
7276 | R: m_Value(V&: X)))) { |
7277 | // The variant with 'add' is not canonical, (the variant with 'not' is) |
7278 | // we only get it because it has extra uses, and can't be canonicalized, |
7279 | |
7280 | switch (Pred) { |
7281 | case ICmpInst::ICMP_ULT: |
7282 | NewPred = ICmpInst::ICMP_NE; |
7283 | break; |
7284 | case ICmpInst::ICMP_UGE: |
7285 | NewPred = ICmpInst::ICMP_EQ; |
7286 | break; |
7287 | default: |
7288 | return nullptr; |
7289 | } |
7290 | } else |
7291 | return nullptr; |
7292 | |
7293 | Value *NewX = Builder.CreateLShr(LHS: X, RHS: Y, Name: X->getName() + ".highbits" ); |
7294 | Constant *Zero = Constant::getNullValue(Ty: NewX->getType()); |
7295 | return CmpInst::Create(Op: Instruction::ICmp, Pred: NewPred, S1: NewX, S2: Zero); |
7296 | } |
7297 | |
7298 | static Instruction *foldVectorCmp(CmpInst &Cmp, |
7299 | InstCombiner::BuilderTy &Builder) { |
7300 | const CmpInst::Predicate Pred = Cmp.getPredicate(); |
7301 | Value *LHS = Cmp.getOperand(i_nocapture: 0), *RHS = Cmp.getOperand(i_nocapture: 1); |
7302 | Value *V1, *V2; |
7303 | |
7304 | auto createCmpReverse = [&](CmpInst::Predicate Pred, Value *X, Value *Y) { |
7305 | Value *V = Builder.CreateCmp(Pred, LHS: X, RHS: Y, Name: Cmp.getName()); |
7306 | if (auto *I = dyn_cast<Instruction>(Val: V)) |
7307 | I->copyIRFlags(V: &Cmp); |
7308 | Module *M = Cmp.getModule(); |
7309 | Function *F = Intrinsic::getOrInsertDeclaration( |
7310 | M, id: Intrinsic::vector_reverse, Tys: V->getType()); |
7311 | return CallInst::Create(Func: F, Args: V); |
7312 | }; |
7313 | |
7314 | if (match(V: LHS, P: m_VecReverse(Op0: m_Value(V&: V1)))) { |
7315 | // cmp Pred, rev(V1), rev(V2) --> rev(cmp Pred, V1, V2) |
7316 | if (match(V: RHS, P: m_VecReverse(Op0: m_Value(V&: V2))) && |
7317 | (LHS->hasOneUse() || RHS->hasOneUse())) |
7318 | return createCmpReverse(Pred, V1, V2); |
7319 | |
7320 | // cmp Pred, rev(V1), RHSSplat --> rev(cmp Pred, V1, RHSSplat) |
7321 | if (LHS->hasOneUse() && isSplatValue(V: RHS)) |
7322 | return createCmpReverse(Pred, V1, RHS); |
7323 | } |
7324 | // cmp Pred, LHSSplat, rev(V2) --> rev(cmp Pred, LHSSplat, V2) |
7325 | else if (isSplatValue(V: LHS) && match(V: RHS, P: m_OneUse(SubPattern: m_VecReverse(Op0: m_Value(V&: V2))))) |
7326 | return createCmpReverse(Pred, LHS, V2); |
7327 | |
7328 | ArrayRef<int> M; |
7329 | if (!match(V: LHS, P: m_Shuffle(v1: m_Value(V&: V1), v2: m_Undef(), mask: m_Mask(M)))) |
7330 | return nullptr; |
7331 | |
7332 | // If both arguments of the cmp are shuffles that use the same mask and |
7333 | // shuffle within a single vector, move the shuffle after the cmp: |
7334 | // cmp (shuffle V1, M), (shuffle V2, M) --> shuffle (cmp V1, V2), M |
7335 | Type *V1Ty = V1->getType(); |
7336 | if (match(V: RHS, P: m_Shuffle(v1: m_Value(V&: V2), v2: m_Undef(), mask: m_SpecificMask(M))) && |
7337 | V1Ty == V2->getType() && (LHS->hasOneUse() || RHS->hasOneUse())) { |
7338 | Value *NewCmp = Builder.CreateCmp(Pred, LHS: V1, RHS: V2); |
7339 | return new ShuffleVectorInst(NewCmp, M); |
7340 | } |
7341 | |
7342 | // Try to canonicalize compare with splatted operand and splat constant. |
7343 | // TODO: We could generalize this for more than splats. See/use the code in |
7344 | // InstCombiner::foldVectorBinop(). |
7345 | Constant *C; |
7346 | if (!LHS->hasOneUse() || !match(V: RHS, P: m_Constant(C))) |
7347 | return nullptr; |
7348 | |
7349 | // Length-changing splats are ok, so adjust the constants as needed: |
7350 | // cmp (shuffle V1, M), C --> shuffle (cmp V1, C'), M |
7351 | Constant *ScalarC = C->getSplatValue(/* AllowPoison */ true); |
7352 | int MaskSplatIndex; |
7353 | if (ScalarC && match(Mask: M, P: m_SplatOrPoisonMask(MaskSplatIndex))) { |
7354 | // We allow poison in matching, but this transform removes it for safety. |
7355 | // Demanded elements analysis should be able to recover some/all of that. |
7356 | C = ConstantVector::getSplat(EC: cast<VectorType>(Val: V1Ty)->getElementCount(), |
7357 | Elt: ScalarC); |
7358 | SmallVector<int, 8> NewM(M.size(), MaskSplatIndex); |
7359 | Value *NewCmp = Builder.CreateCmp(Pred, LHS: V1, RHS: C); |
7360 | return new ShuffleVectorInst(NewCmp, NewM); |
7361 | } |
7362 | |
7363 | return nullptr; |
7364 | } |
7365 | |
7366 | // extract(uadd.with.overflow(A, B), 0) ult A |
7367 | // -> extract(uadd.with.overflow(A, B), 1) |
7368 | static Instruction *foldICmpOfUAddOv(ICmpInst &I) { |
7369 | CmpInst::Predicate Pred = I.getPredicate(); |
7370 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
7371 | |
7372 | Value *UAddOv; |
7373 | Value *A, *B; |
7374 | auto UAddOvResultPat = m_ExtractValue<0>( |
7375 | V: m_Intrinsic<Intrinsic::uadd_with_overflow>(Op0: m_Value(V&: A), Op1: m_Value(V&: B))); |
7376 | if (match(V: Op0, P: UAddOvResultPat) && |
7377 | ((Pred == ICmpInst::ICMP_ULT && (Op1 == A || Op1 == B)) || |
7378 | (Pred == ICmpInst::ICMP_EQ && match(V: Op1, P: m_ZeroInt()) && |
7379 | (match(V: A, P: m_One()) || match(V: B, P: m_One()))) || |
7380 | (Pred == ICmpInst::ICMP_NE && match(V: Op1, P: m_AllOnes()) && |
7381 | (match(V: A, P: m_AllOnes()) || match(V: B, P: m_AllOnes()))))) |
7382 | // extract(uadd.with.overflow(A, B), 0) < A |
7383 | // extract(uadd.with.overflow(A, 1), 0) == 0 |
7384 | // extract(uadd.with.overflow(A, -1), 0) != -1 |
7385 | UAddOv = cast<ExtractValueInst>(Val: Op0)->getAggregateOperand(); |
7386 | else if (match(V: Op1, P: UAddOvResultPat) && Pred == ICmpInst::ICMP_UGT && |
7387 | (Op0 == A || Op0 == B)) |
7388 | // A > extract(uadd.with.overflow(A, B), 0) |
7389 | UAddOv = cast<ExtractValueInst>(Val: Op1)->getAggregateOperand(); |
7390 | else |
7391 | return nullptr; |
7392 | |
7393 | return ExtractValueInst::Create(Agg: UAddOv, Idxs: 1); |
7394 | } |
7395 | |
7396 | static Instruction *foldICmpInvariantGroup(ICmpInst &I) { |
7397 | if (!I.getOperand(i_nocapture: 0)->getType()->isPointerTy() || |
7398 | NullPointerIsDefined( |
7399 | F: I.getParent()->getParent(), |
7400 | AS: I.getOperand(i_nocapture: 0)->getType()->getPointerAddressSpace())) { |
7401 | return nullptr; |
7402 | } |
7403 | Instruction *Op; |
7404 | if (match(V: I.getOperand(i_nocapture: 0), P: m_Instruction(I&: Op)) && |
7405 | match(V: I.getOperand(i_nocapture: 1), P: m_Zero()) && |
7406 | Op->isLaunderOrStripInvariantGroup()) { |
7407 | return ICmpInst::Create(Op: Instruction::ICmp, Pred: I.getPredicate(), |
7408 | S1: Op->getOperand(i: 0), S2: I.getOperand(i_nocapture: 1)); |
7409 | } |
7410 | return nullptr; |
7411 | } |
7412 | |
7413 | /// This function folds patterns produced by lowering of reduce idioms, such as |
7414 | /// llvm.vector.reduce.and which are lowered into instruction chains. This code |
7415 | /// attempts to generate fewer number of scalar comparisons instead of vector |
7416 | /// comparisons when possible. |
7417 | static Instruction *foldReductionIdiom(ICmpInst &I, |
7418 | InstCombiner::BuilderTy &Builder, |
7419 | const DataLayout &DL) { |
7420 | if (I.getType()->isVectorTy()) |
7421 | return nullptr; |
7422 | CmpPredicate OuterPred, InnerPred; |
7423 | Value *LHS, *RHS; |
7424 | |
7425 | // Match lowering of @llvm.vector.reduce.and. Turn |
7426 | /// %vec_ne = icmp ne <8 x i8> %lhs, %rhs |
7427 | /// %scalar_ne = bitcast <8 x i1> %vec_ne to i8 |
7428 | /// %res = icmp <pred> i8 %scalar_ne, 0 |
7429 | /// |
7430 | /// into |
7431 | /// |
7432 | /// %lhs.scalar = bitcast <8 x i8> %lhs to i64 |
7433 | /// %rhs.scalar = bitcast <8 x i8> %rhs to i64 |
7434 | /// %res = icmp <pred> i64 %lhs.scalar, %rhs.scalar |
7435 | /// |
7436 | /// for <pred> in {ne, eq}. |
7437 | if (!match(V: &I, P: m_ICmp(Pred&: OuterPred, |
7438 | L: m_OneUse(SubPattern: m_BitCast(Op: m_OneUse( |
7439 | SubPattern: m_ICmp(Pred&: InnerPred, L: m_Value(V&: LHS), R: m_Value(V&: RHS))))), |
7440 | R: m_Zero()))) |
7441 | return nullptr; |
7442 | auto *LHSTy = dyn_cast<FixedVectorType>(Val: LHS->getType()); |
7443 | if (!LHSTy || !LHSTy->getElementType()->isIntegerTy()) |
7444 | return nullptr; |
7445 | unsigned NumBits = |
7446 | LHSTy->getNumElements() * LHSTy->getElementType()->getIntegerBitWidth(); |
7447 | // TODO: Relax this to "not wider than max legal integer type"? |
7448 | if (!DL.isLegalInteger(Width: NumBits)) |
7449 | return nullptr; |
7450 | |
7451 | if (ICmpInst::isEquality(P: OuterPred) && InnerPred == ICmpInst::ICMP_NE) { |
7452 | auto *ScalarTy = Builder.getIntNTy(N: NumBits); |
7453 | LHS = Builder.CreateBitCast(V: LHS, DestTy: ScalarTy, Name: LHS->getName() + ".scalar" ); |
7454 | RHS = Builder.CreateBitCast(V: RHS, DestTy: ScalarTy, Name: RHS->getName() + ".scalar" ); |
7455 | return ICmpInst::Create(Op: Instruction::ICmp, Pred: OuterPred, S1: LHS, S2: RHS, |
7456 | Name: I.getName()); |
7457 | } |
7458 | |
7459 | return nullptr; |
7460 | } |
7461 | |
7462 | // This helper will be called with icmp operands in both orders. |
7463 | Instruction *InstCombinerImpl::foldICmpCommutative(CmpPredicate Pred, |
7464 | Value *Op0, Value *Op1, |
7465 | ICmpInst &CxtI) { |
7466 | // Try to optimize 'icmp GEP, P' or 'icmp P, GEP'. |
7467 | if (auto *GEP = dyn_cast<GEPOperator>(Val: Op0)) |
7468 | if (Instruction *NI = foldGEPICmp(GEPLHS: GEP, RHS: Op1, Cond: Pred, I&: CxtI)) |
7469 | return NI; |
7470 | |
7471 | if (auto *SI = dyn_cast<SelectInst>(Val: Op0)) |
7472 | if (Instruction *NI = foldSelectICmp(Pred, SI, RHS: Op1, I: CxtI)) |
7473 | return NI; |
7474 | |
7475 | if (auto *MinMax = dyn_cast<MinMaxIntrinsic>(Val: Op0)) |
7476 | if (Instruction *Res = foldICmpWithMinMax(I&: CxtI, MinMax, Z: Op1, Pred)) |
7477 | return Res; |
7478 | |
7479 | { |
7480 | Value *X; |
7481 | const APInt *C; |
7482 | // icmp X+Cst, X |
7483 | if (match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_APInt(Res&: C))) && Op1 == X) |
7484 | return foldICmpAddOpConst(X, C: *C, Pred); |
7485 | } |
7486 | |
7487 | // abs(X) >= X --> true |
7488 | // abs(X) u<= X --> true |
7489 | // abs(X) < X --> false |
7490 | // abs(X) u> X --> false |
7491 | // abs(X) u>= X --> IsIntMinPosion ? `X > -1`: `X u<= INTMIN` |
7492 | // abs(X) <= X --> IsIntMinPosion ? `X > -1`: `X u<= INTMIN` |
7493 | // abs(X) == X --> IsIntMinPosion ? `X > -1`: `X u<= INTMIN` |
7494 | // abs(X) u< X --> IsIntMinPosion ? `X < 0` : `X > INTMIN` |
7495 | // abs(X) > X --> IsIntMinPosion ? `X < 0` : `X > INTMIN` |
7496 | // abs(X) != X --> IsIntMinPosion ? `X < 0` : `X > INTMIN` |
7497 | { |
7498 | Value *X; |
7499 | Constant *C; |
7500 | if (match(V: Op0, P: m_Intrinsic<Intrinsic::abs>(Op0: m_Value(V&: X), Op1: m_Constant(C))) && |
7501 | match(V: Op1, P: m_Specific(V: X))) { |
7502 | Value *NullValue = Constant::getNullValue(Ty: X->getType()); |
7503 | Value *AllOnesValue = Constant::getAllOnesValue(Ty: X->getType()); |
7504 | const APInt SMin = |
7505 | APInt::getSignedMinValue(numBits: X->getType()->getScalarSizeInBits()); |
7506 | bool IsIntMinPosion = C->isAllOnesValue(); |
7507 | switch (Pred) { |
7508 | case CmpInst::ICMP_ULE: |
7509 | case CmpInst::ICMP_SGE: |
7510 | return replaceInstUsesWith(I&: CxtI, V: ConstantInt::getTrue(Ty: CxtI.getType())); |
7511 | case CmpInst::ICMP_UGT: |
7512 | case CmpInst::ICMP_SLT: |
7513 | return replaceInstUsesWith(I&: CxtI, V: ConstantInt::getFalse(Ty: CxtI.getType())); |
7514 | case CmpInst::ICMP_UGE: |
7515 | case CmpInst::ICMP_SLE: |
7516 | case CmpInst::ICMP_EQ: { |
7517 | return replaceInstUsesWith( |
7518 | I&: CxtI, V: IsIntMinPosion |
7519 | ? Builder.CreateICmpSGT(LHS: X, RHS: AllOnesValue) |
7520 | : Builder.CreateICmpULT( |
7521 | LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: SMin + 1))); |
7522 | } |
7523 | case CmpInst::ICMP_ULT: |
7524 | case CmpInst::ICMP_SGT: |
7525 | case CmpInst::ICMP_NE: { |
7526 | return replaceInstUsesWith( |
7527 | I&: CxtI, V: IsIntMinPosion |
7528 | ? Builder.CreateICmpSLT(LHS: X, RHS: NullValue) |
7529 | : Builder.CreateICmpUGT( |
7530 | LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: SMin))); |
7531 | } |
7532 | default: |
7533 | llvm_unreachable("Invalid predicate!" ); |
7534 | } |
7535 | } |
7536 | } |
7537 | |
7538 | const SimplifyQuery Q = SQ.getWithInstruction(I: &CxtI); |
7539 | if (Value *V = foldICmpWithLowBitMaskedVal(Pred, Op0, Op1, Q, IC&: *this)) |
7540 | return replaceInstUsesWith(I&: CxtI, V); |
7541 | |
7542 | // Folding (X / Y) pred X => X swap(pred) 0 for constant Y other than 0 or 1 |
7543 | auto CheckUGT1 = [](const APInt &Divisor) { return Divisor.ugt(RHS: 1); }; |
7544 | { |
7545 | if (match(V: Op0, P: m_UDiv(L: m_Specific(V: Op1), R: m_CheckedInt(CheckFn: CheckUGT1)))) { |
7546 | return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Op1, |
7547 | Constant::getNullValue(Ty: Op1->getType())); |
7548 | } |
7549 | |
7550 | if (!ICmpInst::isUnsigned(predicate: Pred) && |
7551 | match(V: Op0, P: m_SDiv(L: m_Specific(V: Op1), R: m_CheckedInt(CheckFn: CheckUGT1)))) { |
7552 | return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Op1, |
7553 | Constant::getNullValue(Ty: Op1->getType())); |
7554 | } |
7555 | } |
7556 | |
7557 | // Another case of this fold is (X >> Y) pred X => X swap(pred) 0 if Y != 0 |
7558 | auto CheckNE0 = [](const APInt &Shift) { return !Shift.isZero(); }; |
7559 | { |
7560 | if (match(V: Op0, P: m_LShr(L: m_Specific(V: Op1), R: m_CheckedInt(CheckFn: CheckNE0)))) { |
7561 | return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Op1, |
7562 | Constant::getNullValue(Ty: Op1->getType())); |
7563 | } |
7564 | |
7565 | if ((Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SGE) && |
7566 | match(V: Op0, P: m_AShr(L: m_Specific(V: Op1), R: m_CheckedInt(CheckFn: CheckNE0)))) { |
7567 | return new ICmpInst(ICmpInst::getSwappedPredicate(pred: Pred), Op1, |
7568 | Constant::getNullValue(Ty: Op1->getType())); |
7569 | } |
7570 | } |
7571 | |
7572 | return nullptr; |
7573 | } |
7574 | |
7575 | Instruction *InstCombinerImpl::visitICmpInst(ICmpInst &I) { |
7576 | bool Changed = false; |
7577 | const SimplifyQuery Q = SQ.getWithInstruction(I: &I); |
7578 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
7579 | unsigned Op0Cplxity = getComplexity(V: Op0); |
7580 | unsigned Op1Cplxity = getComplexity(V: Op1); |
7581 | |
7582 | /// Orders the operands of the compare so that they are listed from most |
7583 | /// complex to least complex. This puts constants before unary operators, |
7584 | /// before binary operators. |
7585 | if (Op0Cplxity < Op1Cplxity) { |
7586 | I.swapOperands(); |
7587 | std::swap(a&: Op0, b&: Op1); |
7588 | Changed = true; |
7589 | } |
7590 | |
7591 | if (Value *V = simplifyICmpInst(Pred: I.getCmpPredicate(), LHS: Op0, RHS: Op1, Q)) |
7592 | return replaceInstUsesWith(I, V); |
7593 | |
7594 | // Comparing -val or val with non-zero is the same as just comparing val |
7595 | // ie, abs(val) != 0 -> val != 0 |
7596 | if (I.getPredicate() == ICmpInst::ICMP_NE && match(V: Op1, P: m_Zero())) { |
7597 | Value *Cond, *SelectTrue, *SelectFalse; |
7598 | if (match(V: Op0, P: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: SelectTrue), |
7599 | R: m_Value(V&: SelectFalse)))) { |
7600 | if (Value *V = dyn_castNegVal(V: SelectTrue)) { |
7601 | if (V == SelectFalse) |
7602 | return CmpInst::Create(Op: Instruction::ICmp, Pred: I.getPredicate(), S1: V, S2: Op1); |
7603 | } else if (Value *V = dyn_castNegVal(V: SelectFalse)) { |
7604 | if (V == SelectTrue) |
7605 | return CmpInst::Create(Op: Instruction::ICmp, Pred: I.getPredicate(), S1: V, S2: Op1); |
7606 | } |
7607 | } |
7608 | } |
7609 | |
7610 | if (Instruction *Res = foldICmpTruncWithTruncOrExt(Cmp&: I, Q)) |
7611 | return Res; |
7612 | |
7613 | if (Op0->getType()->isIntOrIntVectorTy(BitWidth: 1)) |
7614 | if (Instruction *Res = canonicalizeICmpBool(I, Builder)) |
7615 | return Res; |
7616 | |
7617 | if (Instruction *Res = canonicalizeCmpWithConstant(I)) |
7618 | return Res; |
7619 | |
7620 | if (Instruction *Res = canonicalizeICmpPredicate(I)) |
7621 | return Res; |
7622 | |
7623 | if (Instruction *Res = foldICmpWithConstant(Cmp&: I)) |
7624 | return Res; |
7625 | |
7626 | if (Instruction *Res = foldICmpWithDominatingICmp(Cmp&: I)) |
7627 | return Res; |
7628 | |
7629 | if (Instruction *Res = foldICmpUsingBoolRange(I)) |
7630 | return Res; |
7631 | |
7632 | if (Instruction *Res = foldICmpUsingKnownBits(I)) |
7633 | return Res; |
7634 | |
7635 | // Test if the ICmpInst instruction is used exclusively by a select as |
7636 | // part of a minimum or maximum operation. If so, refrain from doing |
7637 | // any other folding. This helps out other analyses which understand |
7638 | // non-obfuscated minimum and maximum idioms, such as ScalarEvolution |
7639 | // and CodeGen. And in this case, at least one of the comparison |
7640 | // operands has at least one user besides the compare (the select), |
7641 | // which would often largely negate the benefit of folding anyway. |
7642 | // |
7643 | // Do the same for the other patterns recognized by matchSelectPattern. |
7644 | if (I.hasOneUse()) |
7645 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: I.user_back())) { |
7646 | Value *A, *B; |
7647 | SelectPatternResult SPR = matchSelectPattern(V: SI, LHS&: A, RHS&: B); |
7648 | if (SPR.Flavor != SPF_UNKNOWN) |
7649 | return nullptr; |
7650 | } |
7651 | |
7652 | // Do this after checking for min/max to prevent infinite looping. |
7653 | if (Instruction *Res = foldICmpWithZero(Cmp&: I)) |
7654 | return Res; |
7655 | |
7656 | // FIXME: We only do this after checking for min/max to prevent infinite |
7657 | // looping caused by a reverse canonicalization of these patterns for min/max. |
7658 | // FIXME: The organization of folds is a mess. These would naturally go into |
7659 | // canonicalizeCmpWithConstant(), but we can't move all of the above folds |
7660 | // down here after the min/max restriction. |
7661 | ICmpInst::Predicate Pred = I.getPredicate(); |
7662 | const APInt *C; |
7663 | if (match(V: Op1, P: m_APInt(Res&: C))) { |
7664 | // For i32: x >u 2147483647 -> x <s 0 -> true if sign bit set |
7665 | if (Pred == ICmpInst::ICMP_UGT && C->isMaxSignedValue()) { |
7666 | Constant *Zero = Constant::getNullValue(Ty: Op0->getType()); |
7667 | return new ICmpInst(ICmpInst::ICMP_SLT, Op0, Zero); |
7668 | } |
7669 | |
7670 | // For i32: x <u 2147483648 -> x >s -1 -> true if sign bit clear |
7671 | if (Pred == ICmpInst::ICMP_ULT && C->isMinSignedValue()) { |
7672 | Constant *AllOnes = Constant::getAllOnesValue(Ty: Op0->getType()); |
7673 | return new ICmpInst(ICmpInst::ICMP_SGT, Op0, AllOnes); |
7674 | } |
7675 | } |
7676 | |
7677 | // The folds in here may rely on wrapping flags and special constants, so |
7678 | // they can break up min/max idioms in some cases but not seemingly similar |
7679 | // patterns. |
7680 | // FIXME: It may be possible to enhance select folding to make this |
7681 | // unnecessary. It may also be moot if we canonicalize to min/max |
7682 | // intrinsics. |
7683 | if (Instruction *Res = foldICmpBinOp(I, SQ: Q)) |
7684 | return Res; |
7685 | |
7686 | if (Instruction *Res = foldICmpInstWithConstant(Cmp&: I)) |
7687 | return Res; |
7688 | |
7689 | // Try to match comparison as a sign bit test. Intentionally do this after |
7690 | // foldICmpInstWithConstant() to potentially let other folds to happen first. |
7691 | if (Instruction *New = foldSignBitTest(I)) |
7692 | return New; |
7693 | |
7694 | if (Instruction *Res = foldICmpInstWithConstantNotInt(I)) |
7695 | return Res; |
7696 | |
7697 | if (Instruction *Res = foldICmpCommutative(Pred: I.getCmpPredicate(), Op0, Op1, CxtI&: I)) |
7698 | return Res; |
7699 | if (Instruction *Res = |
7700 | foldICmpCommutative(Pred: I.getSwappedCmpPredicate(), Op0: Op1, Op1: Op0, CxtI&: I)) |
7701 | return Res; |
7702 | |
7703 | if (I.isCommutative()) { |
7704 | if (auto Pair = matchSymmetricPair(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1))) { |
7705 | replaceOperand(I, OpNum: 0, V: Pair->first); |
7706 | replaceOperand(I, OpNum: 1, V: Pair->second); |
7707 | return &I; |
7708 | } |
7709 | } |
7710 | |
7711 | // In case of a comparison with two select instructions having the same |
7712 | // condition, check whether one of the resulting branches can be simplified. |
7713 | // If so, just compare the other branch and select the appropriate result. |
7714 | // For example: |
7715 | // %tmp1 = select i1 %cmp, i32 %y, i32 %x |
7716 | // %tmp2 = select i1 %cmp, i32 %z, i32 %x |
7717 | // %cmp2 = icmp slt i32 %tmp2, %tmp1 |
7718 | // The icmp will result false for the false value of selects and the result |
7719 | // will depend upon the comparison of true values of selects if %cmp is |
7720 | // true. Thus, transform this into: |
7721 | // %cmp = icmp slt i32 %y, %z |
7722 | // %sel = select i1 %cond, i1 %cmp, i1 false |
7723 | // This handles similar cases to transform. |
7724 | { |
7725 | Value *Cond, *A, *B, *C, *D; |
7726 | if (match(V: Op0, P: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: A), R: m_Value(V&: B))) && |
7727 | match(V: Op1, P: m_Select(C: m_Specific(V: Cond), L: m_Value(V&: C), R: m_Value(V&: D))) && |
7728 | (Op0->hasOneUse() || Op1->hasOneUse())) { |
7729 | // Check whether comparison of TrueValues can be simplified |
7730 | if (Value *Res = simplifyICmpInst(Pred, LHS: A, RHS: C, Q: SQ)) { |
7731 | Value *NewICMP = Builder.CreateICmp(P: Pred, LHS: B, RHS: D); |
7732 | return SelectInst::Create(C: Cond, S1: Res, S2: NewICMP); |
7733 | } |
7734 | // Check whether comparison of FalseValues can be simplified |
7735 | if (Value *Res = simplifyICmpInst(Pred, LHS: B, RHS: D, Q: SQ)) { |
7736 | Value *NewICMP = Builder.CreateICmp(P: Pred, LHS: A, RHS: C); |
7737 | return SelectInst::Create(C: Cond, S1: NewICMP, S2: Res); |
7738 | } |
7739 | } |
7740 | } |
7741 | |
7742 | // icmp slt (sub nsw x, y), (add nsw x, y) --> icmp sgt y, 0 |
7743 | // icmp ult (sub nuw x, y), (add nuw x, y) --> icmp ugt y, 0 |
7744 | // icmp eq (sub nsw/nuw x, y), (add nsw/nuw x, y) --> icmp eq y, 0 |
7745 | { |
7746 | Value *A, *B; |
7747 | CmpPredicate CmpPred; |
7748 | if (match(V: &I, P: m_c_ICmp(Pred&: CmpPred, L: m_Sub(L: m_Value(V&: A), R: m_Value(V&: B)), |
7749 | R: m_c_Add(L: m_Deferred(V: A), R: m_Deferred(V: B))))) { |
7750 | auto *I0 = cast<OverflowingBinaryOperator>(Val: Op0); |
7751 | auto *I1 = cast<OverflowingBinaryOperator>(Val: Op1); |
7752 | bool I0NUW = I0->hasNoUnsignedWrap(); |
7753 | bool I1NUW = I1->hasNoUnsignedWrap(); |
7754 | bool I0NSW = I0->hasNoSignedWrap(); |
7755 | bool I1NSW = I1->hasNoSignedWrap(); |
7756 | if ((ICmpInst::isUnsigned(predicate: Pred) && I0NUW && I1NUW) || |
7757 | (ICmpInst::isSigned(predicate: Pred) && I0NSW && I1NSW) || |
7758 | (ICmpInst::isEquality(P: Pred) && |
7759 | ((I0NUW || I0NSW) && (I1NUW || I1NSW)))) { |
7760 | return new ICmpInst(CmpPredicate::getSwapped(P: CmpPred), B, |
7761 | ConstantInt::get(Ty: Op0->getType(), V: 0)); |
7762 | } |
7763 | } |
7764 | } |
7765 | |
7766 | // Try to optimize equality comparisons against alloca-based pointers. |
7767 | if (Op0->getType()->isPointerTy() && I.isEquality()) { |
7768 | assert(Op1->getType()->isPointerTy() && |
7769 | "Comparing pointer with non-pointer?" ); |
7770 | if (auto *Alloca = dyn_cast<AllocaInst>(Val: getUnderlyingObject(V: Op0))) |
7771 | if (foldAllocaCmp(Alloca)) |
7772 | return nullptr; |
7773 | if (auto *Alloca = dyn_cast<AllocaInst>(Val: getUnderlyingObject(V: Op1))) |
7774 | if (foldAllocaCmp(Alloca)) |
7775 | return nullptr; |
7776 | } |
7777 | |
7778 | if (Instruction *Res = foldICmpBitCast(Cmp&: I)) |
7779 | return Res; |
7780 | |
7781 | // TODO: Hoist this above the min/max bailout. |
7782 | if (Instruction *R = foldICmpWithCastOp(ICmp&: I)) |
7783 | return R; |
7784 | |
7785 | { |
7786 | Value *X, *Y; |
7787 | // Transform (X & ~Y) == 0 --> (X & Y) != 0 |
7788 | // and (X & ~Y) != 0 --> (X & Y) == 0 |
7789 | // if A is a power of 2. |
7790 | if (match(V: Op0, P: m_And(L: m_Value(V&: X), R: m_Not(V: m_Value(V&: Y)))) && |
7791 | match(V: Op1, P: m_Zero()) && isKnownToBeAPowerOfTwo(V: X, OrZero: false, CxtI: &I) && |
7792 | I.isEquality()) |
7793 | return new ICmpInst(I.getInversePredicate(), Builder.CreateAnd(LHS: X, RHS: Y), |
7794 | Op1); |
7795 | |
7796 | // Op0 pred Op1 -> ~Op1 pred ~Op0, if this allows us to drop an instruction. |
7797 | if (Op0->getType()->isIntOrIntVectorTy()) { |
7798 | bool ConsumesOp0, ConsumesOp1; |
7799 | if (isFreeToInvert(V: Op0, WillInvertAllUses: Op0->hasOneUse(), DoesConsume&: ConsumesOp0) && |
7800 | isFreeToInvert(V: Op1, WillInvertAllUses: Op1->hasOneUse(), DoesConsume&: ConsumesOp1) && |
7801 | (ConsumesOp0 || ConsumesOp1)) { |
7802 | Value *InvOp0 = getFreelyInverted(V: Op0, WillInvertAllUses: Op0->hasOneUse(), Builder: &Builder); |
7803 | Value *InvOp1 = getFreelyInverted(V: Op1, WillInvertAllUses: Op1->hasOneUse(), Builder: &Builder); |
7804 | assert(InvOp0 && InvOp1 && |
7805 | "Mismatch between isFreeToInvert and getFreelyInverted" ); |
7806 | return new ICmpInst(I.getSwappedPredicate(), InvOp0, InvOp1); |
7807 | } |
7808 | } |
7809 | |
7810 | Instruction *AddI = nullptr; |
7811 | if (match(V: &I, P: m_UAddWithOverflow(L: m_Value(V&: X), R: m_Value(V&: Y), |
7812 | S: m_Instruction(I&: AddI))) && |
7813 | isa<IntegerType>(Val: X->getType())) { |
7814 | Value *Result; |
7815 | Constant *Overflow; |
7816 | // m_UAddWithOverflow can match patterns that do not include an explicit |
7817 | // "add" instruction, so check the opcode of the matched op. |
7818 | if (AddI->getOpcode() == Instruction::Add && |
7819 | OptimizeOverflowCheck(BinaryOp: Instruction::Add, /*Signed*/ IsSigned: false, LHS: X, RHS: Y, OrigI&: *AddI, |
7820 | Result, Overflow)) { |
7821 | replaceInstUsesWith(I&: *AddI, V: Result); |
7822 | eraseInstFromFunction(I&: *AddI); |
7823 | return replaceInstUsesWith(I, V: Overflow); |
7824 | } |
7825 | } |
7826 | |
7827 | // (zext X) * (zext Y) --> llvm.umul.with.overflow. |
7828 | if (match(V: Op0, P: m_NUWMul(L: m_ZExt(Op: m_Value(V&: X)), R: m_ZExt(Op: m_Value(V&: Y)))) && |
7829 | match(V: Op1, P: m_APInt(Res&: C))) { |
7830 | if (Instruction *R = processUMulZExtIdiom(I, MulVal: Op0, OtherVal: C, IC&: *this)) |
7831 | return R; |
7832 | } |
7833 | |
7834 | // Signbit test folds |
7835 | // Fold (X u>> BitWidth - 1 Pred ZExt(i1)) --> X s< 0 Pred i1 |
7836 | // Fold (X s>> BitWidth - 1 Pred SExt(i1)) --> X s< 0 Pred i1 |
7837 | Instruction *ExtI; |
7838 | if ((I.isUnsigned() || I.isEquality()) && |
7839 | match(V: Op1, |
7840 | P: m_CombineAnd(L: m_Instruction(I&: ExtI), R: m_ZExtOrSExt(Op: m_Value(V&: Y)))) && |
7841 | Y->getType()->getScalarSizeInBits() == 1 && |
7842 | (Op0->hasOneUse() || Op1->hasOneUse())) { |
7843 | unsigned OpWidth = Op0->getType()->getScalarSizeInBits(); |
7844 | Instruction *ShiftI; |
7845 | if (match(V: Op0, P: m_CombineAnd(L: m_Instruction(I&: ShiftI), |
7846 | R: m_Shr(L: m_Value(V&: X), R: m_SpecificIntAllowPoison( |
7847 | V: OpWidth - 1))))) { |
7848 | unsigned ExtOpc = ExtI->getOpcode(); |
7849 | unsigned ShiftOpc = ShiftI->getOpcode(); |
7850 | if ((ExtOpc == Instruction::ZExt && ShiftOpc == Instruction::LShr) || |
7851 | (ExtOpc == Instruction::SExt && ShiftOpc == Instruction::AShr)) { |
7852 | Value *SLTZero = |
7853 | Builder.CreateICmpSLT(LHS: X, RHS: Constant::getNullValue(Ty: X->getType())); |
7854 | Value *Cmp = Builder.CreateICmp(P: Pred, LHS: SLTZero, RHS: Y, Name: I.getName()); |
7855 | return replaceInstUsesWith(I, V: Cmp); |
7856 | } |
7857 | } |
7858 | } |
7859 | } |
7860 | |
7861 | if (Instruction *Res = foldICmpEquality(I)) |
7862 | return Res; |
7863 | |
7864 | if (Instruction *Res = foldICmpPow2Test(I, Builder)) |
7865 | return Res; |
7866 | |
7867 | if (Instruction *Res = foldICmpOfUAddOv(I)) |
7868 | return Res; |
7869 | |
7870 | // The 'cmpxchg' instruction returns an aggregate containing the old value and |
7871 | // an i1 which indicates whether or not we successfully did the swap. |
7872 | // |
7873 | // Replace comparisons between the old value and the expected value with the |
7874 | // indicator that 'cmpxchg' returns. |
7875 | // |
7876 | // N.B. This transform is only valid when the 'cmpxchg' is not permitted to |
7877 | // spuriously fail. In those cases, the old value may equal the expected |
7878 | // value but it is possible for the swap to not occur. |
7879 | if (I.getPredicate() == ICmpInst::ICMP_EQ) |
7880 | if (auto *EVI = dyn_cast<ExtractValueInst>(Val: Op0)) |
7881 | if (auto *ACXI = dyn_cast<AtomicCmpXchgInst>(Val: EVI->getAggregateOperand())) |
7882 | if (EVI->getIndices()[0] == 0 && ACXI->getCompareOperand() == Op1 && |
7883 | !ACXI->isWeak()) |
7884 | return ExtractValueInst::Create(Agg: ACXI, Idxs: 1); |
7885 | |
7886 | if (Instruction *Res = foldICmpWithHighBitMask(Cmp&: I, Builder)) |
7887 | return Res; |
7888 | |
7889 | if (I.getType()->isVectorTy()) |
7890 | if (Instruction *Res = foldVectorCmp(Cmp&: I, Builder)) |
7891 | return Res; |
7892 | |
7893 | if (Instruction *Res = foldICmpInvariantGroup(I)) |
7894 | return Res; |
7895 | |
7896 | if (Instruction *Res = foldReductionIdiom(I, Builder, DL)) |
7897 | return Res; |
7898 | |
7899 | { |
7900 | Value *A; |
7901 | const APInt *C1, *C2; |
7902 | ICmpInst::Predicate Pred = I.getPredicate(); |
7903 | if (ICmpInst::isEquality(P: Pred)) { |
7904 | // sext(a) & c1 == c2 --> a & c3 == trunc(c2) |
7905 | // sext(a) & c1 != c2 --> a & c3 != trunc(c2) |
7906 | if (match(V: Op0, P: m_And(L: m_SExt(Op: m_Value(V&: A)), R: m_APInt(Res&: C1))) && |
7907 | match(V: Op1, P: m_APInt(Res&: C2))) { |
7908 | Type *InputTy = A->getType(); |
7909 | unsigned InputBitWidth = InputTy->getScalarSizeInBits(); |
7910 | // c2 must be non-negative at the bitwidth of a. |
7911 | if (C2->getActiveBits() < InputBitWidth) { |
7912 | APInt TruncC1 = C1->trunc(width: InputBitWidth); |
7913 | // Check if there are 1s in C1 high bits of size InputBitWidth. |
7914 | if (C1->uge(RHS: APInt::getOneBitSet(numBits: C1->getBitWidth(), BitNo: InputBitWidth))) |
7915 | TruncC1.setBit(InputBitWidth - 1); |
7916 | Value *AndInst = Builder.CreateAnd(LHS: A, RHS: TruncC1); |
7917 | return new ICmpInst( |
7918 | Pred, AndInst, |
7919 | ConstantInt::get(Ty: InputTy, V: C2->trunc(width: InputBitWidth))); |
7920 | } |
7921 | } |
7922 | } |
7923 | } |
7924 | |
7925 | return Changed ? &I : nullptr; |
7926 | } |
7927 | |
7928 | /// Fold fcmp ([us]itofp x, cst) if possible. |
7929 | Instruction *InstCombinerImpl::foldFCmpIntToFPConst(FCmpInst &I, |
7930 | Instruction *LHSI, |
7931 | Constant *RHSC) { |
7932 | const APFloat *RHS; |
7933 | if (!match(V: RHSC, P: m_APFloat(Res&: RHS))) |
7934 | return nullptr; |
7935 | |
7936 | // Get the width of the mantissa. We don't want to hack on conversions that |
7937 | // might lose information from the integer, e.g. "i64 -> float" |
7938 | int MantissaWidth = LHSI->getType()->getFPMantissaWidth(); |
7939 | if (MantissaWidth == -1) |
7940 | return nullptr; // Unknown. |
7941 | |
7942 | Type *IntTy = LHSI->getOperand(i: 0)->getType(); |
7943 | unsigned IntWidth = IntTy->getScalarSizeInBits(); |
7944 | bool LHSUnsigned = isa<UIToFPInst>(Val: LHSI); |
7945 | |
7946 | if (I.isEquality()) { |
7947 | FCmpInst::Predicate P = I.getPredicate(); |
7948 | bool IsExact = false; |
7949 | APSInt RHSCvt(IntWidth, LHSUnsigned); |
7950 | RHS->convertToInteger(Result&: RHSCvt, RM: APFloat::rmNearestTiesToEven, IsExact: &IsExact); |
7951 | |
7952 | // If the floating point constant isn't an integer value, we know if we will |
7953 | // ever compare equal / not equal to it. |
7954 | if (!IsExact) { |
7955 | // TODO: Can never be -0.0 and other non-representable values |
7956 | APFloat RHSRoundInt(*RHS); |
7957 | RHSRoundInt.roundToIntegral(RM: APFloat::rmNearestTiesToEven); |
7958 | if (*RHS != RHSRoundInt) { |
7959 | if (P == FCmpInst::FCMP_OEQ || P == FCmpInst::FCMP_UEQ) |
7960 | return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType())); |
7961 | |
7962 | assert(P == FCmpInst::FCMP_ONE || P == FCmpInst::FCMP_UNE); |
7963 | return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType())); |
7964 | } |
7965 | } |
7966 | |
7967 | // TODO: If the constant is exactly representable, is it always OK to do |
7968 | // equality compares as integer? |
7969 | } |
7970 | |
7971 | // Check to see that the input is converted from an integer type that is small |
7972 | // enough that preserves all bits. TODO: check here for "known" sign bits. |
7973 | // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e. |
7974 | |
7975 | // Following test does NOT adjust IntWidth downwards for signed inputs, |
7976 | // because the most negative value still requires all the mantissa bits |
7977 | // to distinguish it from one less than that value. |
7978 | if ((int)IntWidth > MantissaWidth) { |
7979 | // Conversion would lose accuracy. Check if loss can impact comparison. |
7980 | int Exp = ilogb(Arg: *RHS); |
7981 | if (Exp == APFloat::IEK_Inf) { |
7982 | int MaxExponent = ilogb(Arg: APFloat::getLargest(Sem: RHS->getSemantics())); |
7983 | if (MaxExponent < (int)IntWidth - !LHSUnsigned) |
7984 | // Conversion could create infinity. |
7985 | return nullptr; |
7986 | } else { |
7987 | // Note that if RHS is zero or NaN, then Exp is negative |
7988 | // and first condition is trivially false. |
7989 | if (MantissaWidth <= Exp && Exp <= (int)IntWidth - !LHSUnsigned) |
7990 | // Conversion could affect comparison. |
7991 | return nullptr; |
7992 | } |
7993 | } |
7994 | |
7995 | // Otherwise, we can potentially simplify the comparison. We know that it |
7996 | // will always come through as an integer value and we know the constant is |
7997 | // not a NAN (it would have been previously simplified). |
7998 | assert(!RHS->isNaN() && "NaN comparison not already folded!" ); |
7999 | |
8000 | ICmpInst::Predicate Pred; |
8001 | switch (I.getPredicate()) { |
8002 | default: |
8003 | llvm_unreachable("Unexpected predicate!" ); |
8004 | case FCmpInst::FCMP_UEQ: |
8005 | case FCmpInst::FCMP_OEQ: |
8006 | Pred = ICmpInst::ICMP_EQ; |
8007 | break; |
8008 | case FCmpInst::FCMP_UGT: |
8009 | case FCmpInst::FCMP_OGT: |
8010 | Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT; |
8011 | break; |
8012 | case FCmpInst::FCMP_UGE: |
8013 | case FCmpInst::FCMP_OGE: |
8014 | Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE; |
8015 | break; |
8016 | case FCmpInst::FCMP_ULT: |
8017 | case FCmpInst::FCMP_OLT: |
8018 | Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT; |
8019 | break; |
8020 | case FCmpInst::FCMP_ULE: |
8021 | case FCmpInst::FCMP_OLE: |
8022 | Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE; |
8023 | break; |
8024 | case FCmpInst::FCMP_UNE: |
8025 | case FCmpInst::FCMP_ONE: |
8026 | Pred = ICmpInst::ICMP_NE; |
8027 | break; |
8028 | case FCmpInst::FCMP_ORD: |
8029 | return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType())); |
8030 | case FCmpInst::FCMP_UNO: |
8031 | return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType())); |
8032 | } |
8033 | |
8034 | // Now we know that the APFloat is a normal number, zero or inf. |
8035 | |
8036 | // See if the FP constant is too large for the integer. For example, |
8037 | // comparing an i8 to 300.0. |
8038 | if (!LHSUnsigned) { |
8039 | // If the RHS value is > SignedMax, fold the comparison. This handles +INF |
8040 | // and large values. |
8041 | APFloat SMax(RHS->getSemantics()); |
8042 | SMax.convertFromAPInt(Input: APInt::getSignedMaxValue(numBits: IntWidth), IsSigned: true, |
8043 | RM: APFloat::rmNearestTiesToEven); |
8044 | if (SMax < *RHS) { // smax < 13123.0 |
8045 | if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT || |
8046 | Pred == ICmpInst::ICMP_SLE) |
8047 | return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType())); |
8048 | return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType())); |
8049 | } |
8050 | } else { |
8051 | // If the RHS value is > UnsignedMax, fold the comparison. This handles |
8052 | // +INF and large values. |
8053 | APFloat UMax(RHS->getSemantics()); |
8054 | UMax.convertFromAPInt(Input: APInt::getMaxValue(numBits: IntWidth), IsSigned: false, |
8055 | RM: APFloat::rmNearestTiesToEven); |
8056 | if (UMax < *RHS) { // umax < 13123.0 |
8057 | if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT || |
8058 | Pred == ICmpInst::ICMP_ULE) |
8059 | return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType())); |
8060 | return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType())); |
8061 | } |
8062 | } |
8063 | |
8064 | if (!LHSUnsigned) { |
8065 | // See if the RHS value is < SignedMin. |
8066 | APFloat SMin(RHS->getSemantics()); |
8067 | SMin.convertFromAPInt(Input: APInt::getSignedMinValue(numBits: IntWidth), IsSigned: true, |
8068 | RM: APFloat::rmNearestTiesToEven); |
8069 | if (SMin > *RHS) { // smin > 12312.0 |
8070 | if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT || |
8071 | Pred == ICmpInst::ICMP_SGE) |
8072 | return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType())); |
8073 | return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType())); |
8074 | } |
8075 | } else { |
8076 | // See if the RHS value is < UnsignedMin. |
8077 | APFloat UMin(RHS->getSemantics()); |
8078 | UMin.convertFromAPInt(Input: APInt::getMinValue(numBits: IntWidth), IsSigned: false, |
8079 | RM: APFloat::rmNearestTiesToEven); |
8080 | if (UMin > *RHS) { // umin > 12312.0 |
8081 | if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_UGT || |
8082 | Pred == ICmpInst::ICMP_UGE) |
8083 | return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType())); |
8084 | return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType())); |
8085 | } |
8086 | } |
8087 | |
8088 | // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or |
8089 | // [0, UMAX], but it may still be fractional. Check whether this is the case |
8090 | // using the IsExact flag. |
8091 | // Don't do this for zero, because -0.0 is not fractional. |
8092 | APSInt RHSInt(IntWidth, LHSUnsigned); |
8093 | bool IsExact; |
8094 | RHS->convertToInteger(Result&: RHSInt, RM: APFloat::rmTowardZero, IsExact: &IsExact); |
8095 | if (!RHS->isZero()) { |
8096 | if (!IsExact) { |
8097 | // If we had a comparison against a fractional value, we have to adjust |
8098 | // the compare predicate and sometimes the value. RHSC is rounded towards |
8099 | // zero at this point. |
8100 | switch (Pred) { |
8101 | default: |
8102 | llvm_unreachable("Unexpected integer comparison!" ); |
8103 | case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true |
8104 | return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType())); |
8105 | case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false |
8106 | return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType())); |
8107 | case ICmpInst::ICMP_ULE: |
8108 | // (float)int <= 4.4 --> int <= 4 |
8109 | // (float)int <= -4.4 --> false |
8110 | if (RHS->isNegative()) |
8111 | return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType())); |
8112 | break; |
8113 | case ICmpInst::ICMP_SLE: |
8114 | // (float)int <= 4.4 --> int <= 4 |
8115 | // (float)int <= -4.4 --> int < -4 |
8116 | if (RHS->isNegative()) |
8117 | Pred = ICmpInst::ICMP_SLT; |
8118 | break; |
8119 | case ICmpInst::ICMP_ULT: |
8120 | // (float)int < -4.4 --> false |
8121 | // (float)int < 4.4 --> int <= 4 |
8122 | if (RHS->isNegative()) |
8123 | return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType())); |
8124 | Pred = ICmpInst::ICMP_ULE; |
8125 | break; |
8126 | case ICmpInst::ICMP_SLT: |
8127 | // (float)int < -4.4 --> int < -4 |
8128 | // (float)int < 4.4 --> int <= 4 |
8129 | if (!RHS->isNegative()) |
8130 | Pred = ICmpInst::ICMP_SLE; |
8131 | break; |
8132 | case ICmpInst::ICMP_UGT: |
8133 | // (float)int > 4.4 --> int > 4 |
8134 | // (float)int > -4.4 --> true |
8135 | if (RHS->isNegative()) |
8136 | return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType())); |
8137 | break; |
8138 | case ICmpInst::ICMP_SGT: |
8139 | // (float)int > 4.4 --> int > 4 |
8140 | // (float)int > -4.4 --> int >= -4 |
8141 | if (RHS->isNegative()) |
8142 | Pred = ICmpInst::ICMP_SGE; |
8143 | break; |
8144 | case ICmpInst::ICMP_UGE: |
8145 | // (float)int >= -4.4 --> true |
8146 | // (float)int >= 4.4 --> int > 4 |
8147 | if (RHS->isNegative()) |
8148 | return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType())); |
8149 | Pred = ICmpInst::ICMP_UGT; |
8150 | break; |
8151 | case ICmpInst::ICMP_SGE: |
8152 | // (float)int >= -4.4 --> int >= -4 |
8153 | // (float)int >= 4.4 --> int > 4 |
8154 | if (!RHS->isNegative()) |
8155 | Pred = ICmpInst::ICMP_SGT; |
8156 | break; |
8157 | } |
8158 | } |
8159 | } |
8160 | |
8161 | // Lower this FP comparison into an appropriate integer version of the |
8162 | // comparison. |
8163 | return new ICmpInst(Pred, LHSI->getOperand(i: 0), |
8164 | ConstantInt::get(Ty: LHSI->getOperand(i: 0)->getType(), V: RHSInt)); |
8165 | } |
8166 | |
8167 | /// Fold (C / X) < 0.0 --> X < 0.0 if possible. Swap predicate if necessary. |
8168 | static Instruction *foldFCmpReciprocalAndZero(FCmpInst &I, Instruction *LHSI, |
8169 | Constant *RHSC) { |
8170 | // When C is not 0.0 and infinities are not allowed: |
8171 | // (C / X) < 0.0 is a sign-bit test of X |
8172 | // (C / X) < 0.0 --> X < 0.0 (if C is positive) |
8173 | // (C / X) < 0.0 --> X > 0.0 (if C is negative, swap the predicate) |
8174 | // |
8175 | // Proof: |
8176 | // Multiply (C / X) < 0.0 by X * X / C. |
8177 | // - X is non zero, if it is the flag 'ninf' is violated. |
8178 | // - C defines the sign of X * X * C. Thus it also defines whether to swap |
8179 | // the predicate. C is also non zero by definition. |
8180 | // |
8181 | // Thus X * X / C is non zero and the transformation is valid. [qed] |
8182 | |
8183 | FCmpInst::Predicate Pred = I.getPredicate(); |
8184 | |
8185 | // Check that predicates are valid. |
8186 | if ((Pred != FCmpInst::FCMP_OGT) && (Pred != FCmpInst::FCMP_OLT) && |
8187 | (Pred != FCmpInst::FCMP_OGE) && (Pred != FCmpInst::FCMP_OLE)) |
8188 | return nullptr; |
8189 | |
8190 | // Check that RHS operand is zero. |
8191 | if (!match(V: RHSC, P: m_AnyZeroFP())) |
8192 | return nullptr; |
8193 | |
8194 | // Check fastmath flags ('ninf'). |
8195 | if (!LHSI->hasNoInfs() || !I.hasNoInfs()) |
8196 | return nullptr; |
8197 | |
8198 | // Check the properties of the dividend. It must not be zero to avoid a |
8199 | // division by zero (see Proof). |
8200 | const APFloat *C; |
8201 | if (!match(V: LHSI->getOperand(i: 0), P: m_APFloat(Res&: C))) |
8202 | return nullptr; |
8203 | |
8204 | if (C->isZero()) |
8205 | return nullptr; |
8206 | |
8207 | // Get swapped predicate if necessary. |
8208 | if (C->isNegative()) |
8209 | Pred = I.getSwappedPredicate(); |
8210 | |
8211 | return new FCmpInst(Pred, LHSI->getOperand(i: 1), RHSC, "" , &I); |
8212 | } |
8213 | |
8214 | /// Optimize fabs(X) compared with zero. |
8215 | static Instruction *foldFabsWithFcmpZero(FCmpInst &I, InstCombinerImpl &IC) { |
8216 | Value *X; |
8217 | if (!match(V: I.getOperand(i_nocapture: 0), P: m_FAbs(Op0: m_Value(V&: X)))) |
8218 | return nullptr; |
8219 | |
8220 | const APFloat *C; |
8221 | if (!match(V: I.getOperand(i_nocapture: 1), P: m_APFloat(Res&: C))) |
8222 | return nullptr; |
8223 | |
8224 | if (!C->isPosZero()) { |
8225 | if (!C->isSmallestNormalized()) |
8226 | return nullptr; |
8227 | |
8228 | const Function *F = I.getFunction(); |
8229 | DenormalMode Mode = F->getDenormalMode(FPType: C->getSemantics()); |
8230 | if (Mode.Input == DenormalMode::PreserveSign || |
8231 | Mode.Input == DenormalMode::PositiveZero) { |
8232 | |
8233 | auto replaceFCmp = [](FCmpInst *I, FCmpInst::Predicate P, Value *X) { |
8234 | Constant *Zero = ConstantFP::getZero(Ty: X->getType()); |
8235 | return new FCmpInst(P, X, Zero, "" , I); |
8236 | }; |
8237 | |
8238 | switch (I.getPredicate()) { |
8239 | case FCmpInst::FCMP_OLT: |
8240 | // fcmp olt fabs(x), smallest_normalized_number -> fcmp oeq x, 0.0 |
8241 | return replaceFCmp(&I, FCmpInst::FCMP_OEQ, X); |
8242 | case FCmpInst::FCMP_UGE: |
8243 | // fcmp uge fabs(x), smallest_normalized_number -> fcmp une x, 0.0 |
8244 | return replaceFCmp(&I, FCmpInst::FCMP_UNE, X); |
8245 | case FCmpInst::FCMP_OGE: |
8246 | // fcmp oge fabs(x), smallest_normalized_number -> fcmp one x, 0.0 |
8247 | return replaceFCmp(&I, FCmpInst::FCMP_ONE, X); |
8248 | case FCmpInst::FCMP_ULT: |
8249 | // fcmp ult fabs(x), smallest_normalized_number -> fcmp ueq x, 0.0 |
8250 | return replaceFCmp(&I, FCmpInst::FCMP_UEQ, X); |
8251 | default: |
8252 | break; |
8253 | } |
8254 | } |
8255 | |
8256 | return nullptr; |
8257 | } |
8258 | |
8259 | auto replacePredAndOp0 = [&IC](FCmpInst *I, FCmpInst::Predicate P, Value *X) { |
8260 | I->setPredicate(P); |
8261 | return IC.replaceOperand(I&: *I, OpNum: 0, V: X); |
8262 | }; |
8263 | |
8264 | switch (I.getPredicate()) { |
8265 | case FCmpInst::FCMP_UGE: |
8266 | case FCmpInst::FCMP_OLT: |
8267 | // fabs(X) >= 0.0 --> true |
8268 | // fabs(X) < 0.0 --> false |
8269 | llvm_unreachable("fcmp should have simplified" ); |
8270 | |
8271 | case FCmpInst::FCMP_OGT: |
8272 | // fabs(X) > 0.0 --> X != 0.0 |
8273 | return replacePredAndOp0(&I, FCmpInst::FCMP_ONE, X); |
8274 | |
8275 | case FCmpInst::FCMP_UGT: |
8276 | // fabs(X) u> 0.0 --> X u!= 0.0 |
8277 | return replacePredAndOp0(&I, FCmpInst::FCMP_UNE, X); |
8278 | |
8279 | case FCmpInst::FCMP_OLE: |
8280 | // fabs(X) <= 0.0 --> X == 0.0 |
8281 | return replacePredAndOp0(&I, FCmpInst::FCMP_OEQ, X); |
8282 | |
8283 | case FCmpInst::FCMP_ULE: |
8284 | // fabs(X) u<= 0.0 --> X u== 0.0 |
8285 | return replacePredAndOp0(&I, FCmpInst::FCMP_UEQ, X); |
8286 | |
8287 | case FCmpInst::FCMP_OGE: |
8288 | // fabs(X) >= 0.0 --> !isnan(X) |
8289 | assert(!I.hasNoNaNs() && "fcmp should have simplified" ); |
8290 | return replacePredAndOp0(&I, FCmpInst::FCMP_ORD, X); |
8291 | |
8292 | case FCmpInst::FCMP_ULT: |
8293 | // fabs(X) u< 0.0 --> isnan(X) |
8294 | assert(!I.hasNoNaNs() && "fcmp should have simplified" ); |
8295 | return replacePredAndOp0(&I, FCmpInst::FCMP_UNO, X); |
8296 | |
8297 | case FCmpInst::FCMP_OEQ: |
8298 | case FCmpInst::FCMP_UEQ: |
8299 | case FCmpInst::FCMP_ONE: |
8300 | case FCmpInst::FCMP_UNE: |
8301 | case FCmpInst::FCMP_ORD: |
8302 | case FCmpInst::FCMP_UNO: |
8303 | // Look through the fabs() because it doesn't change anything but the sign. |
8304 | // fabs(X) == 0.0 --> X == 0.0, |
8305 | // fabs(X) != 0.0 --> X != 0.0 |
8306 | // isnan(fabs(X)) --> isnan(X) |
8307 | // !isnan(fabs(X) --> !isnan(X) |
8308 | return replacePredAndOp0(&I, I.getPredicate(), X); |
8309 | |
8310 | default: |
8311 | return nullptr; |
8312 | } |
8313 | } |
8314 | |
8315 | /// Optimize sqrt(X) compared with zero. |
8316 | static Instruction *foldSqrtWithFcmpZero(FCmpInst &I, InstCombinerImpl &IC) { |
8317 | Value *X; |
8318 | if (!match(V: I.getOperand(i_nocapture: 0), P: m_Sqrt(Op0: m_Value(V&: X)))) |
8319 | return nullptr; |
8320 | |
8321 | if (!match(V: I.getOperand(i_nocapture: 1), P: m_PosZeroFP())) |
8322 | return nullptr; |
8323 | |
8324 | auto ReplacePredAndOp0 = [&](FCmpInst::Predicate P) { |
8325 | I.setPredicate(P); |
8326 | return IC.replaceOperand(I, OpNum: 0, V: X); |
8327 | }; |
8328 | |
8329 | // Clear ninf flag if sqrt doesn't have it. |
8330 | if (!cast<Instruction>(Val: I.getOperand(i_nocapture: 0))->hasNoInfs()) |
8331 | I.setHasNoInfs(false); |
8332 | |
8333 | switch (I.getPredicate()) { |
8334 | case FCmpInst::FCMP_OLT: |
8335 | case FCmpInst::FCMP_UGE: |
8336 | // sqrt(X) < 0.0 --> false |
8337 | // sqrt(X) u>= 0.0 --> true |
8338 | llvm_unreachable("fcmp should have simplified" ); |
8339 | case FCmpInst::FCMP_ULT: |
8340 | case FCmpInst::FCMP_ULE: |
8341 | case FCmpInst::FCMP_OGT: |
8342 | case FCmpInst::FCMP_OGE: |
8343 | case FCmpInst::FCMP_OEQ: |
8344 | case FCmpInst::FCMP_UNE: |
8345 | // sqrt(X) u< 0.0 --> X u< 0.0 |
8346 | // sqrt(X) u<= 0.0 --> X u<= 0.0 |
8347 | // sqrt(X) > 0.0 --> X > 0.0 |
8348 | // sqrt(X) >= 0.0 --> X >= 0.0 |
8349 | // sqrt(X) == 0.0 --> X == 0.0 |
8350 | // sqrt(X) u!= 0.0 --> X u!= 0.0 |
8351 | return IC.replaceOperand(I, OpNum: 0, V: X); |
8352 | |
8353 | case FCmpInst::FCMP_OLE: |
8354 | // sqrt(X) <= 0.0 --> X == 0.0 |
8355 | return ReplacePredAndOp0(FCmpInst::FCMP_OEQ); |
8356 | case FCmpInst::FCMP_UGT: |
8357 | // sqrt(X) u> 0.0 --> X u!= 0.0 |
8358 | return ReplacePredAndOp0(FCmpInst::FCMP_UNE); |
8359 | case FCmpInst::FCMP_UEQ: |
8360 | // sqrt(X) u== 0.0 --> X u<= 0.0 |
8361 | return ReplacePredAndOp0(FCmpInst::FCMP_ULE); |
8362 | case FCmpInst::FCMP_ONE: |
8363 | // sqrt(X) != 0.0 --> X > 0.0 |
8364 | return ReplacePredAndOp0(FCmpInst::FCMP_OGT); |
8365 | case FCmpInst::FCMP_ORD: |
8366 | // !isnan(sqrt(X)) --> X >= 0.0 |
8367 | return ReplacePredAndOp0(FCmpInst::FCMP_OGE); |
8368 | case FCmpInst::FCMP_UNO: |
8369 | // isnan(sqrt(X)) --> X u< 0.0 |
8370 | return ReplacePredAndOp0(FCmpInst::FCMP_ULT); |
8371 | default: |
8372 | llvm_unreachable("Unexpected predicate!" ); |
8373 | } |
8374 | } |
8375 | |
8376 | static Instruction *foldFCmpFNegCommonOp(FCmpInst &I) { |
8377 | CmpInst::Predicate Pred = I.getPredicate(); |
8378 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
8379 | |
8380 | // Canonicalize fneg as Op1. |
8381 | if (match(V: Op0, P: m_FNeg(X: m_Value())) && !match(V: Op1, P: m_FNeg(X: m_Value()))) { |
8382 | std::swap(a&: Op0, b&: Op1); |
8383 | Pred = I.getSwappedPredicate(); |
8384 | } |
8385 | |
8386 | if (!match(V: Op1, P: m_FNeg(X: m_Specific(V: Op0)))) |
8387 | return nullptr; |
8388 | |
8389 | // Replace the negated operand with 0.0: |
8390 | // fcmp Pred Op0, -Op0 --> fcmp Pred Op0, 0.0 |
8391 | Constant *Zero = ConstantFP::getZero(Ty: Op0->getType()); |
8392 | return new FCmpInst(Pred, Op0, Zero, "" , &I); |
8393 | } |
8394 | |
8395 | static Instruction *foldFCmpFSubIntoFCmp(FCmpInst &I, Instruction *LHSI, |
8396 | Constant *RHSC, InstCombinerImpl &CI) { |
8397 | const CmpInst::Predicate Pred = I.getPredicate(); |
8398 | Value *X = LHSI->getOperand(i: 0); |
8399 | Value *Y = LHSI->getOperand(i: 1); |
8400 | switch (Pred) { |
8401 | default: |
8402 | break; |
8403 | case FCmpInst::FCMP_UGT: |
8404 | case FCmpInst::FCMP_ULT: |
8405 | case FCmpInst::FCMP_UNE: |
8406 | case FCmpInst::FCMP_OEQ: |
8407 | case FCmpInst::FCMP_OGE: |
8408 | case FCmpInst::FCMP_OLE: |
8409 | // The optimization is not valid if X and Y are infinities of the same |
8410 | // sign, i.e. the inf - inf = nan case. If the fsub has the ninf or nnan |
8411 | // flag then we can assume we do not have that case. Otherwise we might be |
8412 | // able to prove that either X or Y is not infinity. |
8413 | if (!LHSI->hasNoNaNs() && !LHSI->hasNoInfs() && |
8414 | !isKnownNeverInfinity(V: Y, |
8415 | SQ: CI.getSimplifyQuery().getWithInstruction(I: &I)) && |
8416 | !isKnownNeverInfinity(V: X, SQ: CI.getSimplifyQuery().getWithInstruction(I: &I))) |
8417 | break; |
8418 | |
8419 | [[fallthrough]]; |
8420 | case FCmpInst::FCMP_OGT: |
8421 | case FCmpInst::FCMP_OLT: |
8422 | case FCmpInst::FCMP_ONE: |
8423 | case FCmpInst::FCMP_UEQ: |
8424 | case FCmpInst::FCMP_UGE: |
8425 | case FCmpInst::FCMP_ULE: |
8426 | // fcmp pred (x - y), 0 --> fcmp pred x, y |
8427 | if (match(V: RHSC, P: m_AnyZeroFP()) && |
8428 | I.getFunction()->getDenormalMode( |
8429 | FPType: LHSI->getType()->getScalarType()->getFltSemantics()) == |
8430 | DenormalMode::getIEEE()) { |
8431 | CI.replaceOperand(I, OpNum: 0, V: X); |
8432 | CI.replaceOperand(I, OpNum: 1, V: Y); |
8433 | return &I; |
8434 | } |
8435 | break; |
8436 | } |
8437 | |
8438 | return nullptr; |
8439 | } |
8440 | |
8441 | static Instruction *foldFCmpWithFloorAndCeil(FCmpInst &I, |
8442 | InstCombinerImpl &IC) { |
8443 | Value *LHS = I.getOperand(i_nocapture: 0), *RHS = I.getOperand(i_nocapture: 1); |
8444 | Type *OpType = LHS->getType(); |
8445 | CmpInst::Predicate Pred = I.getPredicate(); |
8446 | |
8447 | bool FloorX = match(V: LHS, P: m_Intrinsic<Intrinsic::floor>(Op0: m_Specific(V: RHS))); |
8448 | bool CeilX = match(V: LHS, P: m_Intrinsic<Intrinsic::ceil>(Op0: m_Specific(V: RHS))); |
8449 | |
8450 | if (!FloorX && !CeilX) { |
8451 | if ((FloorX = match(V: RHS, P: m_Intrinsic<Intrinsic::floor>(Op0: m_Specific(V: LHS)))) || |
8452 | (CeilX = match(V: RHS, P: m_Intrinsic<Intrinsic::ceil>(Op0: m_Specific(V: LHS))))) { |
8453 | std::swap(a&: LHS, b&: RHS); |
8454 | Pred = I.getSwappedPredicate(); |
8455 | } |
8456 | } |
8457 | |
8458 | switch (Pred) { |
8459 | case FCmpInst::FCMP_OLE: |
8460 | // fcmp ole floor(x), x => fcmp ord x, 0 |
8461 | if (FloorX) |
8462 | return new FCmpInst(FCmpInst::FCMP_ORD, RHS, ConstantFP::getZero(Ty: OpType), |
8463 | "" , &I); |
8464 | break; |
8465 | case FCmpInst::FCMP_OGT: |
8466 | // fcmp ogt floor(x), x => false |
8467 | if (FloorX) |
8468 | return IC.replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType())); |
8469 | break; |
8470 | case FCmpInst::FCMP_OGE: |
8471 | // fcmp oge ceil(x), x => fcmp ord x, 0 |
8472 | if (CeilX) |
8473 | return new FCmpInst(FCmpInst::FCMP_ORD, RHS, ConstantFP::getZero(Ty: OpType), |
8474 | "" , &I); |
8475 | break; |
8476 | case FCmpInst::FCMP_OLT: |
8477 | // fcmp olt ceil(x), x => false |
8478 | if (CeilX) |
8479 | return IC.replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType())); |
8480 | break; |
8481 | case FCmpInst::FCMP_ULE: |
8482 | // fcmp ule floor(x), x => true |
8483 | if (FloorX) |
8484 | return IC.replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType())); |
8485 | break; |
8486 | case FCmpInst::FCMP_UGT: |
8487 | // fcmp ugt floor(x), x => fcmp uno x, 0 |
8488 | if (FloorX) |
8489 | return new FCmpInst(FCmpInst::FCMP_UNO, RHS, ConstantFP::getZero(Ty: OpType), |
8490 | "" , &I); |
8491 | break; |
8492 | case FCmpInst::FCMP_UGE: |
8493 | // fcmp uge ceil(x), x => true |
8494 | if (CeilX) |
8495 | return IC.replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType())); |
8496 | break; |
8497 | case FCmpInst::FCMP_ULT: |
8498 | // fcmp ult ceil(x), x => fcmp uno x, 0 |
8499 | if (CeilX) |
8500 | return new FCmpInst(FCmpInst::FCMP_UNO, RHS, ConstantFP::getZero(Ty: OpType), |
8501 | "" , &I); |
8502 | break; |
8503 | default: |
8504 | break; |
8505 | } |
8506 | |
8507 | return nullptr; |
8508 | } |
8509 | |
8510 | Instruction *InstCombinerImpl::visitFCmpInst(FCmpInst &I) { |
8511 | bool Changed = false; |
8512 | |
8513 | /// Orders the operands of the compare so that they are listed from most |
8514 | /// complex to least complex. This puts constants before unary operators, |
8515 | /// before binary operators. |
8516 | if (getComplexity(V: I.getOperand(i_nocapture: 0)) < getComplexity(V: I.getOperand(i_nocapture: 1))) { |
8517 | I.swapOperands(); |
8518 | Changed = true; |
8519 | } |
8520 | |
8521 | const CmpInst::Predicate Pred = I.getPredicate(); |
8522 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
8523 | if (Value *V = simplifyFCmpInst(Predicate: Pred, LHS: Op0, RHS: Op1, FMF: I.getFastMathFlags(), |
8524 | Q: SQ.getWithInstruction(I: &I))) |
8525 | return replaceInstUsesWith(I, V); |
8526 | |
8527 | // Simplify 'fcmp pred X, X' |
8528 | Type *OpType = Op0->getType(); |
8529 | assert(OpType == Op1->getType() && "fcmp with different-typed operands?" ); |
8530 | if (Op0 == Op1) { |
8531 | switch (Pred) { |
8532 | default: |
8533 | break; |
8534 | case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y) |
8535 | case FCmpInst::FCMP_ULT: // True if unordered or less than |
8536 | case FCmpInst::FCMP_UGT: // True if unordered or greater than |
8537 | case FCmpInst::FCMP_UNE: // True if unordered or not equal |
8538 | // Canonicalize these to be 'fcmp uno %X, 0.0'. |
8539 | I.setPredicate(FCmpInst::FCMP_UNO); |
8540 | I.setOperand(i_nocapture: 1, Val_nocapture: Constant::getNullValue(Ty: OpType)); |
8541 | return &I; |
8542 | |
8543 | case FCmpInst::FCMP_ORD: // True if ordered (no nans) |
8544 | case FCmpInst::FCMP_OEQ: // True if ordered and equal |
8545 | case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal |
8546 | case FCmpInst::FCMP_OLE: // True if ordered and less than or equal |
8547 | // Canonicalize these to be 'fcmp ord %X, 0.0'. |
8548 | I.setPredicate(FCmpInst::FCMP_ORD); |
8549 | I.setOperand(i_nocapture: 1, Val_nocapture: Constant::getNullValue(Ty: OpType)); |
8550 | return &I; |
8551 | } |
8552 | } |
8553 | |
8554 | if (I.isCommutative()) { |
8555 | if (auto Pair = matchSymmetricPair(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1))) { |
8556 | replaceOperand(I, OpNum: 0, V: Pair->first); |
8557 | replaceOperand(I, OpNum: 1, V: Pair->second); |
8558 | return &I; |
8559 | } |
8560 | } |
8561 | |
8562 | // If we're just checking for a NaN (ORD/UNO) and have a non-NaN operand, |
8563 | // then canonicalize the operand to 0.0. |
8564 | if (Pred == CmpInst::FCMP_ORD || Pred == CmpInst::FCMP_UNO) { |
8565 | if (!match(V: Op0, P: m_PosZeroFP()) && |
8566 | isKnownNeverNaN(V: Op0, SQ: getSimplifyQuery().getWithInstruction(I: &I))) |
8567 | return replaceOperand(I, OpNum: 0, V: ConstantFP::getZero(Ty: OpType)); |
8568 | |
8569 | if (!match(V: Op1, P: m_PosZeroFP()) && |
8570 | isKnownNeverNaN(V: Op1, SQ: getSimplifyQuery().getWithInstruction(I: &I))) |
8571 | return replaceOperand(I, OpNum: 1, V: ConstantFP::getZero(Ty: OpType)); |
8572 | } |
8573 | |
8574 | // fcmp pred (fneg X), (fneg Y) -> fcmp swap(pred) X, Y |
8575 | Value *X, *Y; |
8576 | if (match(V: Op0, P: m_FNeg(X: m_Value(V&: X))) && match(V: Op1, P: m_FNeg(X: m_Value(V&: Y)))) |
8577 | return new FCmpInst(I.getSwappedPredicate(), X, Y, "" , &I); |
8578 | |
8579 | if (Instruction *R = foldFCmpFNegCommonOp(I)) |
8580 | return R; |
8581 | |
8582 | // Test if the FCmpInst instruction is used exclusively by a select as |
8583 | // part of a minimum or maximum operation. If so, refrain from doing |
8584 | // any other folding. This helps out other analyses which understand |
8585 | // non-obfuscated minimum and maximum idioms, such as ScalarEvolution |
8586 | // and CodeGen. And in this case, at least one of the comparison |
8587 | // operands has at least one user besides the compare (the select), |
8588 | // which would often largely negate the benefit of folding anyway. |
8589 | if (I.hasOneUse()) |
8590 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: I.user_back())) { |
8591 | Value *A, *B; |
8592 | SelectPatternResult SPR = matchSelectPattern(V: SI, LHS&: A, RHS&: B); |
8593 | if (SPR.Flavor != SPF_UNKNOWN) |
8594 | return nullptr; |
8595 | } |
8596 | |
8597 | // The sign of 0.0 is ignored by fcmp, so canonicalize to +0.0: |
8598 | // fcmp Pred X, -0.0 --> fcmp Pred X, 0.0 |
8599 | if (match(V: Op1, P: m_AnyZeroFP()) && !match(V: Op1, P: m_PosZeroFP())) |
8600 | return replaceOperand(I, OpNum: 1, V: ConstantFP::getZero(Ty: OpType)); |
8601 | |
8602 | // Canonicalize: |
8603 | // fcmp olt X, +inf -> fcmp one X, +inf |
8604 | // fcmp ole X, +inf -> fcmp ord X, 0 |
8605 | // fcmp ogt X, +inf -> false |
8606 | // fcmp oge X, +inf -> fcmp oeq X, +inf |
8607 | // fcmp ult X, +inf -> fcmp une X, +inf |
8608 | // fcmp ule X, +inf -> true |
8609 | // fcmp ugt X, +inf -> fcmp uno X, 0 |
8610 | // fcmp uge X, +inf -> fcmp ueq X, +inf |
8611 | // fcmp olt X, -inf -> false |
8612 | // fcmp ole X, -inf -> fcmp oeq X, -inf |
8613 | // fcmp ogt X, -inf -> fcmp one X, -inf |
8614 | // fcmp oge X, -inf -> fcmp ord X, 0 |
8615 | // fcmp ult X, -inf -> fcmp uno X, 0 |
8616 | // fcmp ule X, -inf -> fcmp ueq X, -inf |
8617 | // fcmp ugt X, -inf -> fcmp une X, -inf |
8618 | // fcmp uge X, -inf -> true |
8619 | const APFloat *C; |
8620 | if (match(V: Op1, P: m_APFloat(Res&: C)) && C->isInfinity()) { |
8621 | switch (C->isNegative() ? FCmpInst::getSwappedPredicate(pred: Pred) : Pred) { |
8622 | default: |
8623 | break; |
8624 | case FCmpInst::FCMP_ORD: |
8625 | case FCmpInst::FCMP_UNO: |
8626 | case FCmpInst::FCMP_TRUE: |
8627 | case FCmpInst::FCMP_FALSE: |
8628 | case FCmpInst::FCMP_OGT: |
8629 | case FCmpInst::FCMP_ULE: |
8630 | llvm_unreachable("Should be simplified by InstSimplify" ); |
8631 | case FCmpInst::FCMP_OLT: |
8632 | return new FCmpInst(FCmpInst::FCMP_ONE, Op0, Op1, "" , &I); |
8633 | case FCmpInst::FCMP_OLE: |
8634 | return new FCmpInst(FCmpInst::FCMP_ORD, Op0, ConstantFP::getZero(Ty: OpType), |
8635 | "" , &I); |
8636 | case FCmpInst::FCMP_OGE: |
8637 | return new FCmpInst(FCmpInst::FCMP_OEQ, Op0, Op1, "" , &I); |
8638 | case FCmpInst::FCMP_ULT: |
8639 | return new FCmpInst(FCmpInst::FCMP_UNE, Op0, Op1, "" , &I); |
8640 | case FCmpInst::FCMP_UGT: |
8641 | return new FCmpInst(FCmpInst::FCMP_UNO, Op0, ConstantFP::getZero(Ty: OpType), |
8642 | "" , &I); |
8643 | case FCmpInst::FCMP_UGE: |
8644 | return new FCmpInst(FCmpInst::FCMP_UEQ, Op0, Op1, "" , &I); |
8645 | } |
8646 | } |
8647 | |
8648 | // Ignore signbit of bitcasted int when comparing equality to FP 0.0: |
8649 | // fcmp oeq/une (bitcast X), 0.0 --> (and X, SignMaskC) ==/!= 0 |
8650 | if (match(V: Op1, P: m_PosZeroFP()) && |
8651 | match(V: Op0, P: m_OneUse(SubPattern: m_ElementWiseBitCast(Op: m_Value(V&: X))))) { |
8652 | ICmpInst::Predicate IntPred = ICmpInst::BAD_ICMP_PREDICATE; |
8653 | if (Pred == FCmpInst::FCMP_OEQ) |
8654 | IntPred = ICmpInst::ICMP_EQ; |
8655 | else if (Pred == FCmpInst::FCMP_UNE) |
8656 | IntPred = ICmpInst::ICMP_NE; |
8657 | |
8658 | if (IntPred != ICmpInst::BAD_ICMP_PREDICATE) { |
8659 | Type *IntTy = X->getType(); |
8660 | const APInt &SignMask = ~APInt::getSignMask(BitWidth: IntTy->getScalarSizeInBits()); |
8661 | Value *MaskX = Builder.CreateAnd(LHS: X, RHS: ConstantInt::get(Ty: IntTy, V: SignMask)); |
8662 | return new ICmpInst(IntPred, MaskX, ConstantInt::getNullValue(Ty: IntTy)); |
8663 | } |
8664 | } |
8665 | |
8666 | // Handle fcmp with instruction LHS and constant RHS. |
8667 | Instruction *LHSI; |
8668 | Constant *RHSC; |
8669 | if (match(V: Op0, P: m_Instruction(I&: LHSI)) && match(V: Op1, P: m_Constant(C&: RHSC))) { |
8670 | switch (LHSI->getOpcode()) { |
8671 | case Instruction::Select: |
8672 | // fcmp eq (cond ? x : -x), 0 --> fcmp eq x, 0 |
8673 | if (FCmpInst::isEquality(Pred) && match(V: RHSC, P: m_AnyZeroFP()) && |
8674 | match(V: LHSI, P: m_c_Select(L: m_FNeg(X: m_Value(V&: X)), R: m_Deferred(V: X)))) |
8675 | return replaceOperand(I, OpNum: 0, V: X); |
8676 | if (Instruction *NV = FoldOpIntoSelect(Op&: I, SI: cast<SelectInst>(Val: LHSI))) |
8677 | return NV; |
8678 | break; |
8679 | case Instruction::FSub: |
8680 | if (LHSI->hasOneUse()) |
8681 | if (Instruction *NV = foldFCmpFSubIntoFCmp(I, LHSI, RHSC, CI&: *this)) |
8682 | return NV; |
8683 | break; |
8684 | case Instruction::PHI: |
8685 | if (Instruction *NV = foldOpIntoPhi(I, PN: cast<PHINode>(Val: LHSI))) |
8686 | return NV; |
8687 | break; |
8688 | case Instruction::SIToFP: |
8689 | case Instruction::UIToFP: |
8690 | if (Instruction *NV = foldFCmpIntToFPConst(I, LHSI, RHSC)) |
8691 | return NV; |
8692 | break; |
8693 | case Instruction::FDiv: |
8694 | if (Instruction *NV = foldFCmpReciprocalAndZero(I, LHSI, RHSC)) |
8695 | return NV; |
8696 | break; |
8697 | case Instruction::Load: |
8698 | if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: LHSI->getOperand(i: 0))) |
8699 | if (auto *GV = dyn_cast<GlobalVariable>(Val: GEP->getOperand(i_nocapture: 0))) |
8700 | if (Instruction *Res = foldCmpLoadFromIndexedGlobal( |
8701 | LI: cast<LoadInst>(Val: LHSI), GEP, GV, ICI&: I)) |
8702 | return Res; |
8703 | break; |
8704 | } |
8705 | } |
8706 | |
8707 | if (Instruction *R = foldFabsWithFcmpZero(I, IC&: *this)) |
8708 | return R; |
8709 | |
8710 | if (Instruction *R = foldSqrtWithFcmpZero(I, IC&: *this)) |
8711 | return R; |
8712 | |
8713 | if (Instruction *R = foldFCmpWithFloorAndCeil(I, IC&: *this)) |
8714 | return R; |
8715 | |
8716 | if (match(V: Op0, P: m_FNeg(X: m_Value(V&: X)))) { |
8717 | // fcmp pred (fneg X), C --> fcmp swap(pred) X, -C |
8718 | Constant *C; |
8719 | if (match(V: Op1, P: m_Constant(C))) |
8720 | if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL)) |
8721 | return new FCmpInst(I.getSwappedPredicate(), X, NegC, "" , &I); |
8722 | } |
8723 | |
8724 | // fcmp (fadd X, 0.0), Y --> fcmp X, Y |
8725 | if (match(V: Op0, P: m_FAdd(L: m_Value(V&: X), R: m_AnyZeroFP()))) |
8726 | return new FCmpInst(Pred, X, Op1, "" , &I); |
8727 | |
8728 | // fcmp X, (fadd Y, 0.0) --> fcmp X, Y |
8729 | if (match(V: Op1, P: m_FAdd(L: m_Value(V&: Y), R: m_AnyZeroFP()))) |
8730 | return new FCmpInst(Pred, Op0, Y, "" , &I); |
8731 | |
8732 | if (match(V: Op0, P: m_FPExt(Op: m_Value(V&: X)))) { |
8733 | // fcmp (fpext X), (fpext Y) -> fcmp X, Y |
8734 | if (match(V: Op1, P: m_FPExt(Op: m_Value(V&: Y))) && X->getType() == Y->getType()) |
8735 | return new FCmpInst(Pred, X, Y, "" , &I); |
8736 | |
8737 | const APFloat *C; |
8738 | if (match(V: Op1, P: m_APFloat(Res&: C))) { |
8739 | const fltSemantics &FPSem = |
8740 | X->getType()->getScalarType()->getFltSemantics(); |
8741 | bool Lossy; |
8742 | APFloat TruncC = *C; |
8743 | TruncC.convert(ToSemantics: FPSem, RM: APFloat::rmNearestTiesToEven, losesInfo: &Lossy); |
8744 | |
8745 | if (Lossy) { |
8746 | // X can't possibly equal the higher-precision constant, so reduce any |
8747 | // equality comparison. |
8748 | // TODO: Other predicates can be handled via getFCmpCode(). |
8749 | switch (Pred) { |
8750 | case FCmpInst::FCMP_OEQ: |
8751 | // X is ordered and equal to an impossible constant --> false |
8752 | return replaceInstUsesWith(I, V: ConstantInt::getFalse(Ty: I.getType())); |
8753 | case FCmpInst::FCMP_ONE: |
8754 | // X is ordered and not equal to an impossible constant --> ordered |
8755 | return new FCmpInst(FCmpInst::FCMP_ORD, X, |
8756 | ConstantFP::getZero(Ty: X->getType())); |
8757 | case FCmpInst::FCMP_UEQ: |
8758 | // X is unordered or equal to an impossible constant --> unordered |
8759 | return new FCmpInst(FCmpInst::FCMP_UNO, X, |
8760 | ConstantFP::getZero(Ty: X->getType())); |
8761 | case FCmpInst::FCMP_UNE: |
8762 | // X is unordered or not equal to an impossible constant --> true |
8763 | return replaceInstUsesWith(I, V: ConstantInt::getTrue(Ty: I.getType())); |
8764 | default: |
8765 | break; |
8766 | } |
8767 | } |
8768 | |
8769 | // fcmp (fpext X), C -> fcmp X, (fptrunc C) if fptrunc is lossless |
8770 | // Avoid lossy conversions and denormals. |
8771 | // Zero is a special case that's OK to convert. |
8772 | APFloat Fabs = TruncC; |
8773 | Fabs.clearSign(); |
8774 | if (!Lossy && |
8775 | (Fabs.isZero() || !(Fabs < APFloat::getSmallestNormalized(Sem: FPSem)))) { |
8776 | Constant *NewC = ConstantFP::get(Ty: X->getType(), V: TruncC); |
8777 | return new FCmpInst(Pred, X, NewC, "" , &I); |
8778 | } |
8779 | } |
8780 | } |
8781 | |
8782 | // Convert a sign-bit test of an FP value into a cast and integer compare. |
8783 | // TODO: Simplify if the copysign constant is 0.0 or NaN. |
8784 | // TODO: Handle non-zero compare constants. |
8785 | // TODO: Handle other predicates. |
8786 | if (match(V: Op0, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::copysign>(Op0: m_APFloat(Res&: C), |
8787 | Op1: m_Value(V&: X)))) && |
8788 | match(V: Op1, P: m_AnyZeroFP()) && !C->isZero() && !C->isNaN()) { |
8789 | Type *IntType = Builder.getIntNTy(N: X->getType()->getScalarSizeInBits()); |
8790 | if (auto *VecTy = dyn_cast<VectorType>(Val: OpType)) |
8791 | IntType = VectorType::get(ElementType: IntType, EC: VecTy->getElementCount()); |
8792 | |
8793 | // copysign(non-zero constant, X) < 0.0 --> (bitcast X) < 0 |
8794 | if (Pred == FCmpInst::FCMP_OLT) { |
8795 | Value *IntX = Builder.CreateBitCast(V: X, DestTy: IntType); |
8796 | return new ICmpInst(ICmpInst::ICMP_SLT, IntX, |
8797 | ConstantInt::getNullValue(Ty: IntType)); |
8798 | } |
8799 | } |
8800 | |
8801 | { |
8802 | Value *CanonLHS = nullptr, *CanonRHS = nullptr; |
8803 | match(V: Op0, P: m_Intrinsic<Intrinsic::canonicalize>(Op0: m_Value(V&: CanonLHS))); |
8804 | match(V: Op1, P: m_Intrinsic<Intrinsic::canonicalize>(Op0: m_Value(V&: CanonRHS))); |
8805 | |
8806 | // (canonicalize(x) == x) => (x == x) |
8807 | if (CanonLHS == Op1) |
8808 | return new FCmpInst(Pred, Op1, Op1, "" , &I); |
8809 | |
8810 | // (x == canonicalize(x)) => (x == x) |
8811 | if (CanonRHS == Op0) |
8812 | return new FCmpInst(Pred, Op0, Op0, "" , &I); |
8813 | |
8814 | // (canonicalize(x) == canonicalize(y)) => (x == y) |
8815 | if (CanonLHS && CanonRHS) |
8816 | return new FCmpInst(Pred, CanonLHS, CanonRHS, "" , &I); |
8817 | } |
8818 | |
8819 | if (I.getType()->isVectorTy()) |
8820 | if (Instruction *Res = foldVectorCmp(Cmp&: I, Builder)) |
8821 | return Res; |
8822 | |
8823 | return Changed ? &I : nullptr; |
8824 | } |
8825 | |