1 //===- LazyValueInfo.cpp - Value constraint analysis ------------*- C++ -*-===//
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 defines the interface for lazy computation of value constraint
10 // information.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/Analysis/LazyValueInfo.h"
15 #include "llvm/ADT/DenseSet.h"
16 #include "llvm/ADT/Optional.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/Analysis/AssumptionCache.h"
19 #include "llvm/Analysis/ConstantFolding.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/ValueLattice.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/IR/AssemblyAnnotationWriter.h"
25 #include "llvm/IR/CFG.h"
26 #include "llvm/IR/ConstantRange.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/Dominators.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/Intrinsics.h"
33 #include "llvm/IR/LLVMContext.h"
34 #include "llvm/IR/PatternMatch.h"
35 #include "llvm/IR/ValueHandle.h"
36 #include "llvm/InitializePasses.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/FormattedStream.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include <map>
41 using namespace llvm;
42 using namespace PatternMatch;
43
44 #define DEBUG_TYPE "lazy-value-info"
45
46 // This is the number of worklist items we will process to try to discover an
47 // answer for a given value.
48 static const unsigned MaxProcessedPerValue = 500;
49
50 char LazyValueInfoWrapperPass::ID = 0;
LazyValueInfoWrapperPass()51 LazyValueInfoWrapperPass::LazyValueInfoWrapperPass() : FunctionPass(ID) {
52 initializeLazyValueInfoWrapperPassPass(*PassRegistry::getPassRegistry());
53 }
54 INITIALIZE_PASS_BEGIN(LazyValueInfoWrapperPass, "lazy-value-info",
55 "Lazy Value Information Analysis", false, true)
56 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
57 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
58 INITIALIZE_PASS_END(LazyValueInfoWrapperPass, "lazy-value-info",
59 "Lazy Value Information Analysis", false, true)
60
61 namespace llvm {
createLazyValueInfoPass()62 FunctionPass *createLazyValueInfoPass() { return new LazyValueInfoWrapperPass(); }
63 }
64
65 AnalysisKey LazyValueAnalysis::Key;
66
67 /// Returns true if this lattice value represents at most one possible value.
68 /// This is as precise as any lattice value can get while still representing
69 /// reachable code.
hasSingleValue(const ValueLatticeElement & Val)70 static bool hasSingleValue(const ValueLatticeElement &Val) {
71 if (Val.isConstantRange() &&
72 Val.getConstantRange().isSingleElement())
73 // Integer constants are single element ranges
74 return true;
75 if (Val.isConstant())
76 // Non integer constants
77 return true;
78 return false;
79 }
80
81 /// Combine two sets of facts about the same value into a single set of
82 /// facts. Note that this method is not suitable for merging facts along
83 /// different paths in a CFG; that's what the mergeIn function is for. This
84 /// is for merging facts gathered about the same value at the same location
85 /// through two independent means.
86 /// Notes:
87 /// * This method does not promise to return the most precise possible lattice
88 /// value implied by A and B. It is allowed to return any lattice element
89 /// which is at least as strong as *either* A or B (unless our facts
90 /// conflict, see below).
91 /// * Due to unreachable code, the intersection of two lattice values could be
92 /// contradictory. If this happens, we return some valid lattice value so as
93 /// not confuse the rest of LVI. Ideally, we'd always return Undefined, but
94 /// we do not make this guarantee. TODO: This would be a useful enhancement.
intersect(const ValueLatticeElement & A,const ValueLatticeElement & B)95 static ValueLatticeElement intersect(const ValueLatticeElement &A,
96 const ValueLatticeElement &B) {
97 // Undefined is the strongest state. It means the value is known to be along
98 // an unreachable path.
99 if (A.isUndefined())
100 return A;
101 if (B.isUndefined())
102 return B;
103
104 // If we gave up for one, but got a useable fact from the other, use it.
105 if (A.isOverdefined())
106 return B;
107 if (B.isOverdefined())
108 return A;
109
110 // Can't get any more precise than constants.
111 if (hasSingleValue(A))
112 return A;
113 if (hasSingleValue(B))
114 return B;
115
116 // Could be either constant range or not constant here.
117 if (!A.isConstantRange() || !B.isConstantRange()) {
118 // TODO: Arbitrary choice, could be improved
119 return A;
120 }
121
122 // Intersect two constant ranges
123 ConstantRange Range =
124 A.getConstantRange().intersectWith(B.getConstantRange());
125 // Note: An empty range is implicitly converted to overdefined internally.
126 // TODO: We could instead use Undefined here since we've proven a conflict
127 // and thus know this path must be unreachable.
128 return ValueLatticeElement::getRange(std::move(Range));
129 }
130
131 //===----------------------------------------------------------------------===//
132 // LazyValueInfoCache Decl
133 //===----------------------------------------------------------------------===//
134
135 namespace {
136 /// A callback value handle updates the cache when values are erased.
137 class LazyValueInfoCache;
138 struct LVIValueHandle final : public CallbackVH {
139 // Needs to access getValPtr(), which is protected.
140 friend struct DenseMapInfo<LVIValueHandle>;
141
142 LazyValueInfoCache *Parent;
143
LVIValueHandle__anonebac50b60111::LVIValueHandle144 LVIValueHandle(Value *V, LazyValueInfoCache *P)
145 : CallbackVH(V), Parent(P) { }
146
147 void deleted() override;
allUsesReplacedWith__anonebac50b60111::LVIValueHandle148 void allUsesReplacedWith(Value *V) override {
149 deleted();
150 }
151 };
152 } // end anonymous namespace
153
154 namespace {
155 /// This is the cache kept by LazyValueInfo which
156 /// maintains information about queries across the clients' queries.
157 class LazyValueInfoCache {
158 /// This is all of the cached block information for exactly one Value*.
159 /// The entries are sorted by the BasicBlock* of the
160 /// entries, allowing us to do a lookup with a binary search.
161 /// Over-defined lattice values are recorded in OverDefinedCache to reduce
162 /// memory overhead.
163 struct ValueCacheEntryTy {
ValueCacheEntryTy__anonebac50b60211::LazyValueInfoCache::ValueCacheEntryTy164 ValueCacheEntryTy(Value *V, LazyValueInfoCache *P) : Handle(V, P) {}
165 LVIValueHandle Handle;
166 SmallDenseMap<PoisoningVH<BasicBlock>, ValueLatticeElement, 4> BlockVals;
167 };
168
169 /// This tracks, on a per-block basis, the set of values that are
170 /// over-defined at the end of that block.
171 typedef DenseMap<PoisoningVH<BasicBlock>, SmallPtrSet<Value *, 4>>
172 OverDefinedCacheTy;
173 /// Keep track of all blocks that we have ever seen, so we
174 /// don't spend time removing unused blocks from our caches.
175 DenseSet<PoisoningVH<BasicBlock> > SeenBlocks;
176
177 /// This is all of the cached information for all values,
178 /// mapped from Value* to key information.
179 DenseMap<Value *, std::unique_ptr<ValueCacheEntryTy>> ValueCache;
180 OverDefinedCacheTy OverDefinedCache;
181
182
183 public:
insertResult(Value * Val,BasicBlock * BB,const ValueLatticeElement & Result)184 void insertResult(Value *Val, BasicBlock *BB,
185 const ValueLatticeElement &Result) {
186 SeenBlocks.insert(BB);
187
188 // Insert over-defined values into their own cache to reduce memory
189 // overhead.
190 if (Result.isOverdefined())
191 OverDefinedCache[BB].insert(Val);
192 else {
193 auto It = ValueCache.find_as(Val);
194 if (It == ValueCache.end()) {
195 ValueCache[Val] = std::make_unique<ValueCacheEntryTy>(Val, this);
196 It = ValueCache.find_as(Val);
197 assert(It != ValueCache.end() && "Val was just added to the map!");
198 }
199 It->second->BlockVals[BB] = Result;
200 }
201 }
202
isOverdefined(Value * V,BasicBlock * BB) const203 bool isOverdefined(Value *V, BasicBlock *BB) const {
204 auto ODI = OverDefinedCache.find(BB);
205
206 if (ODI == OverDefinedCache.end())
207 return false;
208
209 return ODI->second.count(V);
210 }
211
hasCachedValueInfo(Value * V,BasicBlock * BB) const212 bool hasCachedValueInfo(Value *V, BasicBlock *BB) const {
213 if (isOverdefined(V, BB))
214 return true;
215
216 auto I = ValueCache.find_as(V);
217 if (I == ValueCache.end())
218 return false;
219
220 return I->second->BlockVals.count(BB);
221 }
222
getCachedValueInfo(Value * V,BasicBlock * BB) const223 ValueLatticeElement getCachedValueInfo(Value *V, BasicBlock *BB) const {
224 if (isOverdefined(V, BB))
225 return ValueLatticeElement::getOverdefined();
226
227 auto I = ValueCache.find_as(V);
228 if (I == ValueCache.end())
229 return ValueLatticeElement();
230 auto BBI = I->second->BlockVals.find(BB);
231 if (BBI == I->second->BlockVals.end())
232 return ValueLatticeElement();
233 return BBI->second;
234 }
235
236 /// clear - Empty the cache.
clear()237 void clear() {
238 SeenBlocks.clear();
239 ValueCache.clear();
240 OverDefinedCache.clear();
241 }
242
243 /// Inform the cache that a given value has been deleted.
244 void eraseValue(Value *V);
245
246 /// This is part of the update interface to inform the cache
247 /// that a block has been deleted.
248 void eraseBlock(BasicBlock *BB);
249
250 /// Updates the cache to remove any influence an overdefined value in
251 /// OldSucc might have (unless also overdefined in NewSucc). This just
252 /// flushes elements from the cache and does not add any.
253 void threadEdgeImpl(BasicBlock *OldSucc,BasicBlock *NewSucc);
254
255 friend struct LVIValueHandle;
256 };
257 }
258
eraseValue(Value * V)259 void LazyValueInfoCache::eraseValue(Value *V) {
260 for (auto I = OverDefinedCache.begin(), E = OverDefinedCache.end(); I != E;) {
261 // Copy and increment the iterator immediately so we can erase behind
262 // ourselves.
263 auto Iter = I++;
264 SmallPtrSetImpl<Value *> &ValueSet = Iter->second;
265 ValueSet.erase(V);
266 if (ValueSet.empty())
267 OverDefinedCache.erase(Iter);
268 }
269
270 ValueCache.erase(V);
271 }
272
deleted()273 void LVIValueHandle::deleted() {
274 // This erasure deallocates *this, so it MUST happen after we're done
275 // using any and all members of *this.
276 Parent->eraseValue(*this);
277 }
278
eraseBlock(BasicBlock * BB)279 void LazyValueInfoCache::eraseBlock(BasicBlock *BB) {
280 // Shortcut if we have never seen this block.
281 DenseSet<PoisoningVH<BasicBlock> >::iterator I = SeenBlocks.find(BB);
282 if (I == SeenBlocks.end())
283 return;
284 SeenBlocks.erase(I);
285
286 auto ODI = OverDefinedCache.find(BB);
287 if (ODI != OverDefinedCache.end())
288 OverDefinedCache.erase(ODI);
289
290 for (auto &I : ValueCache)
291 I.second->BlockVals.erase(BB);
292 }
293
threadEdgeImpl(BasicBlock * OldSucc,BasicBlock * NewSucc)294 void LazyValueInfoCache::threadEdgeImpl(BasicBlock *OldSucc,
295 BasicBlock *NewSucc) {
296 // When an edge in the graph has been threaded, values that we could not
297 // determine a value for before (i.e. were marked overdefined) may be
298 // possible to solve now. We do NOT try to proactively update these values.
299 // Instead, we clear their entries from the cache, and allow lazy updating to
300 // recompute them when needed.
301
302 // The updating process is fairly simple: we need to drop cached info
303 // for all values that were marked overdefined in OldSucc, and for those same
304 // values in any successor of OldSucc (except NewSucc) in which they were
305 // also marked overdefined.
306 std::vector<BasicBlock*> worklist;
307 worklist.push_back(OldSucc);
308
309 auto I = OverDefinedCache.find(OldSucc);
310 if (I == OverDefinedCache.end())
311 return; // Nothing to process here.
312 SmallVector<Value *, 4> ValsToClear(I->second.begin(), I->second.end());
313
314 // Use a worklist to perform a depth-first search of OldSucc's successors.
315 // NOTE: We do not need a visited list since any blocks we have already
316 // visited will have had their overdefined markers cleared already, and we
317 // thus won't loop to their successors.
318 while (!worklist.empty()) {
319 BasicBlock *ToUpdate = worklist.back();
320 worklist.pop_back();
321
322 // Skip blocks only accessible through NewSucc.
323 if (ToUpdate == NewSucc) continue;
324
325 // If a value was marked overdefined in OldSucc, and is here too...
326 auto OI = OverDefinedCache.find(ToUpdate);
327 if (OI == OverDefinedCache.end())
328 continue;
329 SmallPtrSetImpl<Value *> &ValueSet = OI->second;
330
331 bool changed = false;
332 for (Value *V : ValsToClear) {
333 if (!ValueSet.erase(V))
334 continue;
335
336 // If we removed anything, then we potentially need to update
337 // blocks successors too.
338 changed = true;
339
340 if (ValueSet.empty()) {
341 OverDefinedCache.erase(OI);
342 break;
343 }
344 }
345
346 if (!changed) continue;
347
348 worklist.insert(worklist.end(), succ_begin(ToUpdate), succ_end(ToUpdate));
349 }
350 }
351
352
353 namespace {
354 /// An assembly annotator class to print LazyValueCache information in
355 /// comments.
356 class LazyValueInfoImpl;
357 class LazyValueInfoAnnotatedWriter : public AssemblyAnnotationWriter {
358 LazyValueInfoImpl *LVIImpl;
359 // While analyzing which blocks we can solve values for, we need the dominator
360 // information. Since this is an optional parameter in LVI, we require this
361 // DomTreeAnalysis pass in the printer pass, and pass the dominator
362 // tree to the LazyValueInfoAnnotatedWriter.
363 DominatorTree &DT;
364
365 public:
LazyValueInfoAnnotatedWriter(LazyValueInfoImpl * L,DominatorTree & DTree)366 LazyValueInfoAnnotatedWriter(LazyValueInfoImpl *L, DominatorTree &DTree)
367 : LVIImpl(L), DT(DTree) {}
368
369 virtual void emitBasicBlockStartAnnot(const BasicBlock *BB,
370 formatted_raw_ostream &OS);
371
372 virtual void emitInstructionAnnot(const Instruction *I,
373 formatted_raw_ostream &OS);
374 };
375 }
376 namespace {
377 // The actual implementation of the lazy analysis and update. Note that the
378 // inheritance from LazyValueInfoCache is intended to be temporary while
379 // splitting the code and then transitioning to a has-a relationship.
380 class LazyValueInfoImpl {
381
382 /// Cached results from previous queries
383 LazyValueInfoCache TheCache;
384
385 /// This stack holds the state of the value solver during a query.
386 /// It basically emulates the callstack of the naive
387 /// recursive value lookup process.
388 SmallVector<std::pair<BasicBlock*, Value*>, 8> BlockValueStack;
389
390 /// Keeps track of which block-value pairs are in BlockValueStack.
391 DenseSet<std::pair<BasicBlock*, Value*> > BlockValueSet;
392
393 /// Push BV onto BlockValueStack unless it's already in there.
394 /// Returns true on success.
pushBlockValue(const std::pair<BasicBlock *,Value * > & BV)395 bool pushBlockValue(const std::pair<BasicBlock *, Value *> &BV) {
396 if (!BlockValueSet.insert(BV).second)
397 return false; // It's already in the stack.
398
399 LLVM_DEBUG(dbgs() << "PUSH: " << *BV.second << " in "
400 << BV.first->getName() << "\n");
401 BlockValueStack.push_back(BV);
402 return true;
403 }
404
405 AssumptionCache *AC; ///< A pointer to the cache of @llvm.assume calls.
406 const DataLayout &DL; ///< A mandatory DataLayout
407 DominatorTree *DT; ///< An optional DT pointer.
408 DominatorTree *DisabledDT; ///< Stores DT if it's disabled.
409
410 ValueLatticeElement getBlockValue(Value *Val, BasicBlock *BB);
411 bool getEdgeValue(Value *V, BasicBlock *F, BasicBlock *T,
412 ValueLatticeElement &Result, Instruction *CxtI = nullptr);
413 bool hasBlockValue(Value *Val, BasicBlock *BB);
414
415 // These methods process one work item and may add more. A false value
416 // returned means that the work item was not completely processed and must
417 // be revisited after going through the new items.
418 bool solveBlockValue(Value *Val, BasicBlock *BB);
419 bool solveBlockValueImpl(ValueLatticeElement &Res, Value *Val,
420 BasicBlock *BB);
421 bool solveBlockValueNonLocal(ValueLatticeElement &BBLV, Value *Val,
422 BasicBlock *BB);
423 bool solveBlockValuePHINode(ValueLatticeElement &BBLV, PHINode *PN,
424 BasicBlock *BB);
425 bool solveBlockValueSelect(ValueLatticeElement &BBLV, SelectInst *S,
426 BasicBlock *BB);
427 Optional<ConstantRange> getRangeForOperand(unsigned Op, Instruction *I,
428 BasicBlock *BB);
429 bool solveBlockValueBinaryOpImpl(
430 ValueLatticeElement &BBLV, Instruction *I, BasicBlock *BB,
431 std::function<ConstantRange(const ConstantRange &,
432 const ConstantRange &)> OpFn);
433 bool solveBlockValueBinaryOp(ValueLatticeElement &BBLV, BinaryOperator *BBI,
434 BasicBlock *BB);
435 bool solveBlockValueCast(ValueLatticeElement &BBLV, CastInst *CI,
436 BasicBlock *BB);
437 bool solveBlockValueOverflowIntrinsic(
438 ValueLatticeElement &BBLV, WithOverflowInst *WO, BasicBlock *BB);
439 bool solveBlockValueSaturatingIntrinsic(ValueLatticeElement &BBLV,
440 SaturatingInst *SI, BasicBlock *BB);
441 bool solveBlockValueIntrinsic(ValueLatticeElement &BBLV, IntrinsicInst *II,
442 BasicBlock *BB);
443 bool solveBlockValueExtractValue(ValueLatticeElement &BBLV,
444 ExtractValueInst *EVI, BasicBlock *BB);
445 void intersectAssumeOrGuardBlockValueConstantRange(Value *Val,
446 ValueLatticeElement &BBLV,
447 Instruction *BBI);
448
449 void solve();
450
451 public:
452 /// This is the query interface to determine the lattice
453 /// value for the specified Value* at the end of the specified block.
454 ValueLatticeElement getValueInBlock(Value *V, BasicBlock *BB,
455 Instruction *CxtI = nullptr);
456
457 /// This is the query interface to determine the lattice
458 /// value for the specified Value* at the specified instruction (generally
459 /// from an assume intrinsic).
460 ValueLatticeElement getValueAt(Value *V, Instruction *CxtI);
461
462 /// This is the query interface to determine the lattice
463 /// value for the specified Value* that is true on the specified edge.
464 ValueLatticeElement getValueOnEdge(Value *V, BasicBlock *FromBB,
465 BasicBlock *ToBB,
466 Instruction *CxtI = nullptr);
467
468 /// Complete flush all previously computed values
clear()469 void clear() {
470 TheCache.clear();
471 }
472
473 /// Printing the LazyValueInfo Analysis.
printLVI(Function & F,DominatorTree & DTree,raw_ostream & OS)474 void printLVI(Function &F, DominatorTree &DTree, raw_ostream &OS) {
475 LazyValueInfoAnnotatedWriter Writer(this, DTree);
476 F.print(OS, &Writer);
477 }
478
479 /// This is part of the update interface to inform the cache
480 /// that a block has been deleted.
eraseBlock(BasicBlock * BB)481 void eraseBlock(BasicBlock *BB) {
482 TheCache.eraseBlock(BB);
483 }
484
485 /// Disables use of the DominatorTree within LVI.
disableDT()486 void disableDT() {
487 if (DT) {
488 assert(!DisabledDT && "Both DT and DisabledDT are not nullptr!");
489 std::swap(DT, DisabledDT);
490 }
491 }
492
493 /// Enables use of the DominatorTree within LVI. Does nothing if the class
494 /// instance was initialized without a DT pointer.
enableDT()495 void enableDT() {
496 if (DisabledDT) {
497 assert(!DT && "Both DT and DisabledDT are not nullptr!");
498 std::swap(DT, DisabledDT);
499 }
500 }
501
502 /// This is the update interface to inform the cache that an edge from
503 /// PredBB to OldSucc has been threaded to be from PredBB to NewSucc.
504 void threadEdge(BasicBlock *PredBB,BasicBlock *OldSucc,BasicBlock *NewSucc);
505
LazyValueInfoImpl(AssumptionCache * AC,const DataLayout & DL,DominatorTree * DT=nullptr)506 LazyValueInfoImpl(AssumptionCache *AC, const DataLayout &DL,
507 DominatorTree *DT = nullptr)
508 : AC(AC), DL(DL), DT(DT), DisabledDT(nullptr) {}
509 };
510 } // end anonymous namespace
511
512
solve()513 void LazyValueInfoImpl::solve() {
514 SmallVector<std::pair<BasicBlock *, Value *>, 8> StartingStack(
515 BlockValueStack.begin(), BlockValueStack.end());
516
517 unsigned processedCount = 0;
518 while (!BlockValueStack.empty()) {
519 processedCount++;
520 // Abort if we have to process too many values to get a result for this one.
521 // Because of the design of the overdefined cache currently being per-block
522 // to avoid naming-related issues (IE it wants to try to give different
523 // results for the same name in different blocks), overdefined results don't
524 // get cached globally, which in turn means we will often try to rediscover
525 // the same overdefined result again and again. Once something like
526 // PredicateInfo is used in LVI or CVP, we should be able to make the
527 // overdefined cache global, and remove this throttle.
528 if (processedCount > MaxProcessedPerValue) {
529 LLVM_DEBUG(
530 dbgs() << "Giving up on stack because we are getting too deep\n");
531 // Fill in the original values
532 while (!StartingStack.empty()) {
533 std::pair<BasicBlock *, Value *> &e = StartingStack.back();
534 TheCache.insertResult(e.second, e.first,
535 ValueLatticeElement::getOverdefined());
536 StartingStack.pop_back();
537 }
538 BlockValueSet.clear();
539 BlockValueStack.clear();
540 return;
541 }
542 std::pair<BasicBlock *, Value *> e = BlockValueStack.back();
543 assert(BlockValueSet.count(e) && "Stack value should be in BlockValueSet!");
544
545 if (solveBlockValue(e.second, e.first)) {
546 // The work item was completely processed.
547 assert(BlockValueStack.back() == e && "Nothing should have been pushed!");
548 assert(TheCache.hasCachedValueInfo(e.second, e.first) &&
549 "Result should be in cache!");
550
551 LLVM_DEBUG(
552 dbgs() << "POP " << *e.second << " in " << e.first->getName() << " = "
553 << TheCache.getCachedValueInfo(e.second, e.first) << "\n");
554
555 BlockValueStack.pop_back();
556 BlockValueSet.erase(e);
557 } else {
558 // More work needs to be done before revisiting.
559 assert(BlockValueStack.back() != e && "Stack should have been pushed!");
560 }
561 }
562 }
563
hasBlockValue(Value * Val,BasicBlock * BB)564 bool LazyValueInfoImpl::hasBlockValue(Value *Val, BasicBlock *BB) {
565 // If already a constant, there is nothing to compute.
566 if (isa<Constant>(Val))
567 return true;
568
569 return TheCache.hasCachedValueInfo(Val, BB);
570 }
571
getBlockValue(Value * Val,BasicBlock * BB)572 ValueLatticeElement LazyValueInfoImpl::getBlockValue(Value *Val,
573 BasicBlock *BB) {
574 // If already a constant, there is nothing to compute.
575 if (Constant *VC = dyn_cast<Constant>(Val))
576 return ValueLatticeElement::get(VC);
577
578 return TheCache.getCachedValueInfo(Val, BB);
579 }
580
getFromRangeMetadata(Instruction * BBI)581 static ValueLatticeElement getFromRangeMetadata(Instruction *BBI) {
582 switch (BBI->getOpcode()) {
583 default: break;
584 case Instruction::Load:
585 case Instruction::Call:
586 case Instruction::Invoke:
587 if (MDNode *Ranges = BBI->getMetadata(LLVMContext::MD_range))
588 if (isa<IntegerType>(BBI->getType())) {
589 return ValueLatticeElement::getRange(
590 getConstantRangeFromMetadata(*Ranges));
591 }
592 break;
593 };
594 // Nothing known - will be intersected with other facts
595 return ValueLatticeElement::getOverdefined();
596 }
597
solveBlockValue(Value * Val,BasicBlock * BB)598 bool LazyValueInfoImpl::solveBlockValue(Value *Val, BasicBlock *BB) {
599 if (isa<Constant>(Val))
600 return true;
601
602 if (TheCache.hasCachedValueInfo(Val, BB)) {
603 // If we have a cached value, use that.
604 LLVM_DEBUG(dbgs() << " reuse BB '" << BB->getName() << "' val="
605 << TheCache.getCachedValueInfo(Val, BB) << '\n');
606
607 // Since we're reusing a cached value, we don't need to update the
608 // OverDefinedCache. The cache will have been properly updated whenever the
609 // cached value was inserted.
610 return true;
611 }
612
613 // Hold off inserting this value into the Cache in case we have to return
614 // false and come back later.
615 ValueLatticeElement Res;
616 if (!solveBlockValueImpl(Res, Val, BB))
617 // Work pushed, will revisit
618 return false;
619
620 TheCache.insertResult(Val, BB, Res);
621 return true;
622 }
623
solveBlockValueImpl(ValueLatticeElement & Res,Value * Val,BasicBlock * BB)624 bool LazyValueInfoImpl::solveBlockValueImpl(ValueLatticeElement &Res,
625 Value *Val, BasicBlock *BB) {
626
627 Instruction *BBI = dyn_cast<Instruction>(Val);
628 if (!BBI || BBI->getParent() != BB)
629 return solveBlockValueNonLocal(Res, Val, BB);
630
631 if (PHINode *PN = dyn_cast<PHINode>(BBI))
632 return solveBlockValuePHINode(Res, PN, BB);
633
634 if (auto *SI = dyn_cast<SelectInst>(BBI))
635 return solveBlockValueSelect(Res, SI, BB);
636
637 // If this value is a nonnull pointer, record it's range and bailout. Note
638 // that for all other pointer typed values, we terminate the search at the
639 // definition. We could easily extend this to look through geps, bitcasts,
640 // and the like to prove non-nullness, but it's not clear that's worth it
641 // compile time wise. The context-insensitive value walk done inside
642 // isKnownNonZero gets most of the profitable cases at much less expense.
643 // This does mean that we have a sensitivity to where the defining
644 // instruction is placed, even if it could legally be hoisted much higher.
645 // That is unfortunate.
646 PointerType *PT = dyn_cast<PointerType>(BBI->getType());
647 if (PT && isKnownNonZero(BBI, DL)) {
648 Res = ValueLatticeElement::getNot(ConstantPointerNull::get(PT));
649 return true;
650 }
651 if (BBI->getType()->isIntegerTy()) {
652 if (auto *CI = dyn_cast<CastInst>(BBI))
653 return solveBlockValueCast(Res, CI, BB);
654
655 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(BBI))
656 return solveBlockValueBinaryOp(Res, BO, BB);
657
658 if (auto *EVI = dyn_cast<ExtractValueInst>(BBI))
659 return solveBlockValueExtractValue(Res, EVI, BB);
660
661 if (auto *II = dyn_cast<IntrinsicInst>(BBI))
662 return solveBlockValueIntrinsic(Res, II, BB);
663 }
664
665 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()
666 << "' - unknown inst def found.\n");
667 Res = getFromRangeMetadata(BBI);
668 return true;
669 }
670
InstructionDereferencesPointer(Instruction * I,Value * Ptr)671 static bool InstructionDereferencesPointer(Instruction *I, Value *Ptr) {
672 if (LoadInst *L = dyn_cast<LoadInst>(I)) {
673 return L->getPointerAddressSpace() == 0 &&
674 GetUnderlyingObject(L->getPointerOperand(),
675 L->getModule()->getDataLayout()) == Ptr;
676 }
677 if (StoreInst *S = dyn_cast<StoreInst>(I)) {
678 return S->getPointerAddressSpace() == 0 &&
679 GetUnderlyingObject(S->getPointerOperand(),
680 S->getModule()->getDataLayout()) == Ptr;
681 }
682 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) {
683 if (MI->isVolatile()) return false;
684
685 // FIXME: check whether it has a valuerange that excludes zero?
686 ConstantInt *Len = dyn_cast<ConstantInt>(MI->getLength());
687 if (!Len || Len->isZero()) return false;
688
689 if (MI->getDestAddressSpace() == 0)
690 if (GetUnderlyingObject(MI->getRawDest(),
691 MI->getModule()->getDataLayout()) == Ptr)
692 return true;
693 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
694 if (MTI->getSourceAddressSpace() == 0)
695 if (GetUnderlyingObject(MTI->getRawSource(),
696 MTI->getModule()->getDataLayout()) == Ptr)
697 return true;
698 }
699 return false;
700 }
701
702 /// Return true if the allocation associated with Val is ever dereferenced
703 /// within the given basic block. This establishes the fact Val is not null,
704 /// but does not imply that the memory at Val is dereferenceable. (Val may
705 /// point off the end of the dereferenceable part of the object.)
isObjectDereferencedInBlock(Value * Val,BasicBlock * BB)706 static bool isObjectDereferencedInBlock(Value *Val, BasicBlock *BB) {
707 assert(Val->getType()->isPointerTy());
708
709 const DataLayout &DL = BB->getModule()->getDataLayout();
710 Value *UnderlyingVal = GetUnderlyingObject(Val, DL);
711 // If 'GetUnderlyingObject' didn't converge, skip it. It won't converge
712 // inside InstructionDereferencesPointer either.
713 if (UnderlyingVal == GetUnderlyingObject(UnderlyingVal, DL, 1))
714 for (Instruction &I : *BB)
715 if (InstructionDereferencesPointer(&I, UnderlyingVal))
716 return true;
717 return false;
718 }
719
solveBlockValueNonLocal(ValueLatticeElement & BBLV,Value * Val,BasicBlock * BB)720 bool LazyValueInfoImpl::solveBlockValueNonLocal(ValueLatticeElement &BBLV,
721 Value *Val, BasicBlock *BB) {
722 ValueLatticeElement Result; // Start Undefined.
723
724 // If this is the entry block, we must be asking about an argument. The
725 // value is overdefined.
726 if (BB == &BB->getParent()->getEntryBlock()) {
727 assert(isa<Argument>(Val) && "Unknown live-in to the entry block");
728 // Before giving up, see if we can prove the pointer non-null local to
729 // this particular block.
730 PointerType *PTy = dyn_cast<PointerType>(Val->getType());
731 if (PTy &&
732 (isKnownNonZero(Val, DL) ||
733 (isObjectDereferencedInBlock(Val, BB) &&
734 !NullPointerIsDefined(BB->getParent(), PTy->getAddressSpace())))) {
735 Result = ValueLatticeElement::getNot(ConstantPointerNull::get(PTy));
736 } else {
737 Result = ValueLatticeElement::getOverdefined();
738 }
739 BBLV = Result;
740 return true;
741 }
742
743 // Loop over all of our predecessors, merging what we know from them into
744 // result. If we encounter an unexplored predecessor, we eagerly explore it
745 // in a depth first manner. In practice, this has the effect of discovering
746 // paths we can't analyze eagerly without spending compile times analyzing
747 // other paths. This heuristic benefits from the fact that predecessors are
748 // frequently arranged such that dominating ones come first and we quickly
749 // find a path to function entry. TODO: We should consider explicitly
750 // canonicalizing to make this true rather than relying on this happy
751 // accident.
752 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
753 ValueLatticeElement EdgeResult;
754 if (!getEdgeValue(Val, *PI, BB, EdgeResult))
755 // Explore that input, then return here
756 return false;
757
758 Result.mergeIn(EdgeResult, DL);
759
760 // If we hit overdefined, exit early. The BlockVals entry is already set
761 // to overdefined.
762 if (Result.isOverdefined()) {
763 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()
764 << "' - overdefined because of pred (non local).\n");
765 // Before giving up, see if we can prove the pointer non-null local to
766 // this particular block.
767 PointerType *PTy = dyn_cast<PointerType>(Val->getType());
768 if (PTy && isObjectDereferencedInBlock(Val, BB) &&
769 !NullPointerIsDefined(BB->getParent(), PTy->getAddressSpace())) {
770 Result = ValueLatticeElement::getNot(ConstantPointerNull::get(PTy));
771 }
772
773 BBLV = Result;
774 return true;
775 }
776 }
777
778 // Return the merged value, which is more precise than 'overdefined'.
779 assert(!Result.isOverdefined());
780 BBLV = Result;
781 return true;
782 }
783
solveBlockValuePHINode(ValueLatticeElement & BBLV,PHINode * PN,BasicBlock * BB)784 bool LazyValueInfoImpl::solveBlockValuePHINode(ValueLatticeElement &BBLV,
785 PHINode *PN, BasicBlock *BB) {
786 ValueLatticeElement Result; // Start Undefined.
787
788 // Loop over all of our predecessors, merging what we know from them into
789 // result. See the comment about the chosen traversal order in
790 // solveBlockValueNonLocal; the same reasoning applies here.
791 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
792 BasicBlock *PhiBB = PN->getIncomingBlock(i);
793 Value *PhiVal = PN->getIncomingValue(i);
794 ValueLatticeElement EdgeResult;
795 // Note that we can provide PN as the context value to getEdgeValue, even
796 // though the results will be cached, because PN is the value being used as
797 // the cache key in the caller.
798 if (!getEdgeValue(PhiVal, PhiBB, BB, EdgeResult, PN))
799 // Explore that input, then return here
800 return false;
801
802 Result.mergeIn(EdgeResult, DL);
803
804 // If we hit overdefined, exit early. The BlockVals entry is already set
805 // to overdefined.
806 if (Result.isOverdefined()) {
807 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()
808 << "' - overdefined because of pred (local).\n");
809
810 BBLV = Result;
811 return true;
812 }
813 }
814
815 // Return the merged value, which is more precise than 'overdefined'.
816 assert(!Result.isOverdefined() && "Possible PHI in entry block?");
817 BBLV = Result;
818 return true;
819 }
820
821 static ValueLatticeElement getValueFromCondition(Value *Val, Value *Cond,
822 bool isTrueDest = true);
823
824 // If we can determine a constraint on the value given conditions assumed by
825 // the program, intersect those constraints with BBLV
intersectAssumeOrGuardBlockValueConstantRange(Value * Val,ValueLatticeElement & BBLV,Instruction * BBI)826 void LazyValueInfoImpl::intersectAssumeOrGuardBlockValueConstantRange(
827 Value *Val, ValueLatticeElement &BBLV, Instruction *BBI) {
828 BBI = BBI ? BBI : dyn_cast<Instruction>(Val);
829 if (!BBI)
830 return;
831
832 for (auto &AssumeVH : AC->assumptionsFor(Val)) {
833 if (!AssumeVH)
834 continue;
835 auto *I = cast<CallInst>(AssumeVH);
836 if (!isValidAssumeForContext(I, BBI, DT))
837 continue;
838
839 BBLV = intersect(BBLV, getValueFromCondition(Val, I->getArgOperand(0)));
840 }
841
842 // If guards are not used in the module, don't spend time looking for them
843 auto *GuardDecl = BBI->getModule()->getFunction(
844 Intrinsic::getName(Intrinsic::experimental_guard));
845 if (!GuardDecl || GuardDecl->use_empty())
846 return;
847
848 if (BBI->getIterator() == BBI->getParent()->begin())
849 return;
850 for (Instruction &I : make_range(std::next(BBI->getIterator().getReverse()),
851 BBI->getParent()->rend())) {
852 Value *Cond = nullptr;
853 if (match(&I, m_Intrinsic<Intrinsic::experimental_guard>(m_Value(Cond))))
854 BBLV = intersect(BBLV, getValueFromCondition(Val, Cond));
855 }
856 }
857
solveBlockValueSelect(ValueLatticeElement & BBLV,SelectInst * SI,BasicBlock * BB)858 bool LazyValueInfoImpl::solveBlockValueSelect(ValueLatticeElement &BBLV,
859 SelectInst *SI, BasicBlock *BB) {
860
861 // Recurse on our inputs if needed
862 if (!hasBlockValue(SI->getTrueValue(), BB)) {
863 if (pushBlockValue(std::make_pair(BB, SI->getTrueValue())))
864 return false;
865 BBLV = ValueLatticeElement::getOverdefined();
866 return true;
867 }
868 ValueLatticeElement TrueVal = getBlockValue(SI->getTrueValue(), BB);
869 // If we hit overdefined, don't ask more queries. We want to avoid poisoning
870 // extra slots in the table if we can.
871 if (TrueVal.isOverdefined()) {
872 BBLV = ValueLatticeElement::getOverdefined();
873 return true;
874 }
875
876 if (!hasBlockValue(SI->getFalseValue(), BB)) {
877 if (pushBlockValue(std::make_pair(BB, SI->getFalseValue())))
878 return false;
879 BBLV = ValueLatticeElement::getOverdefined();
880 return true;
881 }
882 ValueLatticeElement FalseVal = getBlockValue(SI->getFalseValue(), BB);
883 // If we hit overdefined, don't ask more queries. We want to avoid poisoning
884 // extra slots in the table if we can.
885 if (FalseVal.isOverdefined()) {
886 BBLV = ValueLatticeElement::getOverdefined();
887 return true;
888 }
889
890 if (TrueVal.isConstantRange() && FalseVal.isConstantRange()) {
891 const ConstantRange &TrueCR = TrueVal.getConstantRange();
892 const ConstantRange &FalseCR = FalseVal.getConstantRange();
893 Value *LHS = nullptr;
894 Value *RHS = nullptr;
895 SelectPatternResult SPR = matchSelectPattern(SI, LHS, RHS);
896 // Is this a min specifically of our two inputs? (Avoid the risk of
897 // ValueTracking getting smarter looking back past our immediate inputs.)
898 if (SelectPatternResult::isMinOrMax(SPR.Flavor) &&
899 LHS == SI->getTrueValue() && RHS == SI->getFalseValue()) {
900 ConstantRange ResultCR = [&]() {
901 switch (SPR.Flavor) {
902 default:
903 llvm_unreachable("unexpected minmax type!");
904 case SPF_SMIN: /// Signed minimum
905 return TrueCR.smin(FalseCR);
906 case SPF_UMIN: /// Unsigned minimum
907 return TrueCR.umin(FalseCR);
908 case SPF_SMAX: /// Signed maximum
909 return TrueCR.smax(FalseCR);
910 case SPF_UMAX: /// Unsigned maximum
911 return TrueCR.umax(FalseCR);
912 };
913 }();
914 BBLV = ValueLatticeElement::getRange(ResultCR);
915 return true;
916 }
917
918 if (SPR.Flavor == SPF_ABS) {
919 if (LHS == SI->getTrueValue()) {
920 BBLV = ValueLatticeElement::getRange(TrueCR.abs());
921 return true;
922 }
923 if (LHS == SI->getFalseValue()) {
924 BBLV = ValueLatticeElement::getRange(FalseCR.abs());
925 return true;
926 }
927 }
928
929 if (SPR.Flavor == SPF_NABS) {
930 ConstantRange Zero(APInt::getNullValue(TrueCR.getBitWidth()));
931 if (LHS == SI->getTrueValue()) {
932 BBLV = ValueLatticeElement::getRange(Zero.sub(TrueCR.abs()));
933 return true;
934 }
935 if (LHS == SI->getFalseValue()) {
936 BBLV = ValueLatticeElement::getRange(Zero.sub(FalseCR.abs()));
937 return true;
938 }
939 }
940 }
941
942 // Can we constrain the facts about the true and false values by using the
943 // condition itself? This shows up with idioms like e.g. select(a > 5, a, 5).
944 // TODO: We could potentially refine an overdefined true value above.
945 Value *Cond = SI->getCondition();
946 TrueVal = intersect(TrueVal,
947 getValueFromCondition(SI->getTrueValue(), Cond, true));
948 FalseVal = intersect(FalseVal,
949 getValueFromCondition(SI->getFalseValue(), Cond, false));
950
951 // Handle clamp idioms such as:
952 // %24 = constantrange<0, 17>
953 // %39 = icmp eq i32 %24, 0
954 // %40 = add i32 %24, -1
955 // %siv.next = select i1 %39, i32 16, i32 %40
956 // %siv.next = constantrange<0, 17> not <-1, 17>
957 // In general, this can handle any clamp idiom which tests the edge
958 // condition via an equality or inequality.
959 if (auto *ICI = dyn_cast<ICmpInst>(Cond)) {
960 ICmpInst::Predicate Pred = ICI->getPredicate();
961 Value *A = ICI->getOperand(0);
962 if (ConstantInt *CIBase = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
963 auto addConstants = [](ConstantInt *A, ConstantInt *B) {
964 assert(A->getType() == B->getType());
965 return ConstantInt::get(A->getType(), A->getValue() + B->getValue());
966 };
967 // See if either input is A + C2, subject to the constraint from the
968 // condition that A != C when that input is used. We can assume that
969 // that input doesn't include C + C2.
970 ConstantInt *CIAdded;
971 switch (Pred) {
972 default: break;
973 case ICmpInst::ICMP_EQ:
974 if (match(SI->getFalseValue(), m_Add(m_Specific(A),
975 m_ConstantInt(CIAdded)))) {
976 auto ResNot = addConstants(CIBase, CIAdded);
977 FalseVal = intersect(FalseVal,
978 ValueLatticeElement::getNot(ResNot));
979 }
980 break;
981 case ICmpInst::ICMP_NE:
982 if (match(SI->getTrueValue(), m_Add(m_Specific(A),
983 m_ConstantInt(CIAdded)))) {
984 auto ResNot = addConstants(CIBase, CIAdded);
985 TrueVal = intersect(TrueVal,
986 ValueLatticeElement::getNot(ResNot));
987 }
988 break;
989 };
990 }
991 }
992
993 ValueLatticeElement Result; // Start Undefined.
994 Result.mergeIn(TrueVal, DL);
995 Result.mergeIn(FalseVal, DL);
996 BBLV = Result;
997 return true;
998 }
999
getRangeForOperand(unsigned Op,Instruction * I,BasicBlock * BB)1000 Optional<ConstantRange> LazyValueInfoImpl::getRangeForOperand(unsigned Op,
1001 Instruction *I,
1002 BasicBlock *BB) {
1003 if (!hasBlockValue(I->getOperand(Op), BB))
1004 if (pushBlockValue(std::make_pair(BB, I->getOperand(Op))))
1005 return None;
1006
1007 const unsigned OperandBitWidth =
1008 DL.getTypeSizeInBits(I->getOperand(Op)->getType());
1009 ConstantRange Range = ConstantRange::getFull(OperandBitWidth);
1010 if (hasBlockValue(I->getOperand(Op), BB)) {
1011 ValueLatticeElement Val = getBlockValue(I->getOperand(Op), BB);
1012 intersectAssumeOrGuardBlockValueConstantRange(I->getOperand(Op), Val, I);
1013 if (Val.isConstantRange())
1014 Range = Val.getConstantRange();
1015 }
1016 return Range;
1017 }
1018
solveBlockValueCast(ValueLatticeElement & BBLV,CastInst * CI,BasicBlock * BB)1019 bool LazyValueInfoImpl::solveBlockValueCast(ValueLatticeElement &BBLV,
1020 CastInst *CI,
1021 BasicBlock *BB) {
1022 if (!CI->getOperand(0)->getType()->isSized()) {
1023 // Without knowing how wide the input is, we can't analyze it in any useful
1024 // way.
1025 BBLV = ValueLatticeElement::getOverdefined();
1026 return true;
1027 }
1028
1029 // Filter out casts we don't know how to reason about before attempting to
1030 // recurse on our operand. This can cut a long search short if we know we're
1031 // not going to be able to get any useful information anways.
1032 switch (CI->getOpcode()) {
1033 case Instruction::Trunc:
1034 case Instruction::SExt:
1035 case Instruction::ZExt:
1036 case Instruction::BitCast:
1037 break;
1038 default:
1039 // Unhandled instructions are overdefined.
1040 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()
1041 << "' - overdefined (unknown cast).\n");
1042 BBLV = ValueLatticeElement::getOverdefined();
1043 return true;
1044 }
1045
1046 // Figure out the range of the LHS. If that fails, we still apply the
1047 // transfer rule on the full set since we may be able to locally infer
1048 // interesting facts.
1049 Optional<ConstantRange> LHSRes = getRangeForOperand(0, CI, BB);
1050 if (!LHSRes.hasValue())
1051 // More work to do before applying this transfer rule.
1052 return false;
1053 ConstantRange LHSRange = LHSRes.getValue();
1054
1055 const unsigned ResultBitWidth = CI->getType()->getIntegerBitWidth();
1056
1057 // NOTE: We're currently limited by the set of operations that ConstantRange
1058 // can evaluate symbolically. Enhancing that set will allows us to analyze
1059 // more definitions.
1060 BBLV = ValueLatticeElement::getRange(LHSRange.castOp(CI->getOpcode(),
1061 ResultBitWidth));
1062 return true;
1063 }
1064
solveBlockValueBinaryOpImpl(ValueLatticeElement & BBLV,Instruction * I,BasicBlock * BB,std::function<ConstantRange (const ConstantRange &,const ConstantRange &)> OpFn)1065 bool LazyValueInfoImpl::solveBlockValueBinaryOpImpl(
1066 ValueLatticeElement &BBLV, Instruction *I, BasicBlock *BB,
1067 std::function<ConstantRange(const ConstantRange &,
1068 const ConstantRange &)> OpFn) {
1069 // Figure out the ranges of the operands. If that fails, use a
1070 // conservative range, but apply the transfer rule anyways. This
1071 // lets us pick up facts from expressions like "and i32 (call i32
1072 // @foo()), 32"
1073 Optional<ConstantRange> LHSRes = getRangeForOperand(0, I, BB);
1074 Optional<ConstantRange> RHSRes = getRangeForOperand(1, I, BB);
1075 if (!LHSRes.hasValue() || !RHSRes.hasValue())
1076 // More work to do before applying this transfer rule.
1077 return false;
1078
1079 ConstantRange LHSRange = LHSRes.getValue();
1080 ConstantRange RHSRange = RHSRes.getValue();
1081 BBLV = ValueLatticeElement::getRange(OpFn(LHSRange, RHSRange));
1082 return true;
1083 }
1084
solveBlockValueBinaryOp(ValueLatticeElement & BBLV,BinaryOperator * BO,BasicBlock * BB)1085 bool LazyValueInfoImpl::solveBlockValueBinaryOp(ValueLatticeElement &BBLV,
1086 BinaryOperator *BO,
1087 BasicBlock *BB) {
1088
1089 assert(BO->getOperand(0)->getType()->isSized() &&
1090 "all operands to binary operators are sized");
1091 if (BO->getOpcode() == Instruction::Xor) {
1092 // Xor is the only operation not supported by ConstantRange::binaryOp().
1093 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()
1094 << "' - overdefined (unknown binary operator).\n");
1095 BBLV = ValueLatticeElement::getOverdefined();
1096 return true;
1097 }
1098
1099 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(BO)) {
1100 unsigned NoWrapKind = 0;
1101 if (OBO->hasNoUnsignedWrap())
1102 NoWrapKind |= OverflowingBinaryOperator::NoUnsignedWrap;
1103 if (OBO->hasNoSignedWrap())
1104 NoWrapKind |= OverflowingBinaryOperator::NoSignedWrap;
1105
1106 return solveBlockValueBinaryOpImpl(
1107 BBLV, BO, BB,
1108 [BO, NoWrapKind](const ConstantRange &CR1, const ConstantRange &CR2) {
1109 return CR1.overflowingBinaryOp(BO->getOpcode(), CR2, NoWrapKind);
1110 });
1111 }
1112
1113 return solveBlockValueBinaryOpImpl(
1114 BBLV, BO, BB, [BO](const ConstantRange &CR1, const ConstantRange &CR2) {
1115 return CR1.binaryOp(BO->getOpcode(), CR2);
1116 });
1117 }
1118
solveBlockValueOverflowIntrinsic(ValueLatticeElement & BBLV,WithOverflowInst * WO,BasicBlock * BB)1119 bool LazyValueInfoImpl::solveBlockValueOverflowIntrinsic(
1120 ValueLatticeElement &BBLV, WithOverflowInst *WO, BasicBlock *BB) {
1121 return solveBlockValueBinaryOpImpl(BBLV, WO, BB,
1122 [WO](const ConstantRange &CR1, const ConstantRange &CR2) {
1123 return CR1.binaryOp(WO->getBinaryOp(), CR2);
1124 });
1125 }
1126
solveBlockValueSaturatingIntrinsic(ValueLatticeElement & BBLV,SaturatingInst * SI,BasicBlock * BB)1127 bool LazyValueInfoImpl::solveBlockValueSaturatingIntrinsic(
1128 ValueLatticeElement &BBLV, SaturatingInst *SI, BasicBlock *BB) {
1129 switch (SI->getIntrinsicID()) {
1130 case Intrinsic::uadd_sat:
1131 return solveBlockValueBinaryOpImpl(
1132 BBLV, SI, BB, [](const ConstantRange &CR1, const ConstantRange &CR2) {
1133 return CR1.uadd_sat(CR2);
1134 });
1135 case Intrinsic::usub_sat:
1136 return solveBlockValueBinaryOpImpl(
1137 BBLV, SI, BB, [](const ConstantRange &CR1, const ConstantRange &CR2) {
1138 return CR1.usub_sat(CR2);
1139 });
1140 case Intrinsic::sadd_sat:
1141 return solveBlockValueBinaryOpImpl(
1142 BBLV, SI, BB, [](const ConstantRange &CR1, const ConstantRange &CR2) {
1143 return CR1.sadd_sat(CR2);
1144 });
1145 case Intrinsic::ssub_sat:
1146 return solveBlockValueBinaryOpImpl(
1147 BBLV, SI, BB, [](const ConstantRange &CR1, const ConstantRange &CR2) {
1148 return CR1.ssub_sat(CR2);
1149 });
1150 default:
1151 llvm_unreachable("All llvm.sat intrinsic are handled.");
1152 }
1153 }
1154
solveBlockValueIntrinsic(ValueLatticeElement & BBLV,IntrinsicInst * II,BasicBlock * BB)1155 bool LazyValueInfoImpl::solveBlockValueIntrinsic(ValueLatticeElement &BBLV,
1156 IntrinsicInst *II,
1157 BasicBlock *BB) {
1158 if (auto *SI = dyn_cast<SaturatingInst>(II))
1159 return solveBlockValueSaturatingIntrinsic(BBLV, SI, BB);
1160
1161 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()
1162 << "' - overdefined (unknown intrinsic).\n");
1163 BBLV = ValueLatticeElement::getOverdefined();
1164 return true;
1165 }
1166
solveBlockValueExtractValue(ValueLatticeElement & BBLV,ExtractValueInst * EVI,BasicBlock * BB)1167 bool LazyValueInfoImpl::solveBlockValueExtractValue(
1168 ValueLatticeElement &BBLV, ExtractValueInst *EVI, BasicBlock *BB) {
1169 if (auto *WO = dyn_cast<WithOverflowInst>(EVI->getAggregateOperand()))
1170 if (EVI->getNumIndices() == 1 && *EVI->idx_begin() == 0)
1171 return solveBlockValueOverflowIntrinsic(BBLV, WO, BB);
1172
1173 // Handle extractvalue of insertvalue to allow further simplification
1174 // based on replaced with.overflow intrinsics.
1175 if (Value *V = SimplifyExtractValueInst(
1176 EVI->getAggregateOperand(), EVI->getIndices(),
1177 EVI->getModule()->getDataLayout())) {
1178 if (!hasBlockValue(V, BB)) {
1179 if (pushBlockValue({ BB, V }))
1180 return false;
1181 BBLV = ValueLatticeElement::getOverdefined();
1182 return true;
1183 }
1184 BBLV = getBlockValue(V, BB);
1185 return true;
1186 }
1187
1188 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()
1189 << "' - overdefined (unknown extractvalue).\n");
1190 BBLV = ValueLatticeElement::getOverdefined();
1191 return true;
1192 }
1193
getValueFromICmpCondition(Value * Val,ICmpInst * ICI,bool isTrueDest)1194 static ValueLatticeElement getValueFromICmpCondition(Value *Val, ICmpInst *ICI,
1195 bool isTrueDest) {
1196 Value *LHS = ICI->getOperand(0);
1197 Value *RHS = ICI->getOperand(1);
1198 CmpInst::Predicate Predicate = ICI->getPredicate();
1199
1200 if (isa<Constant>(RHS)) {
1201 if (ICI->isEquality() && LHS == Val) {
1202 // We know that V has the RHS constant if this is a true SETEQ or
1203 // false SETNE.
1204 if (isTrueDest == (Predicate == ICmpInst::ICMP_EQ))
1205 return ValueLatticeElement::get(cast<Constant>(RHS));
1206 else
1207 return ValueLatticeElement::getNot(cast<Constant>(RHS));
1208 }
1209 }
1210
1211 if (!Val->getType()->isIntegerTy())
1212 return ValueLatticeElement::getOverdefined();
1213
1214 // Use ConstantRange::makeAllowedICmpRegion in order to determine the possible
1215 // range of Val guaranteed by the condition. Recognize comparisons in the from
1216 // of:
1217 // icmp <pred> Val, ...
1218 // icmp <pred> (add Val, Offset), ...
1219 // The latter is the range checking idiom that InstCombine produces. Subtract
1220 // the offset from the allowed range for RHS in this case.
1221
1222 // Val or (add Val, Offset) can be on either hand of the comparison
1223 if (LHS != Val && !match(LHS, m_Add(m_Specific(Val), m_ConstantInt()))) {
1224 std::swap(LHS, RHS);
1225 Predicate = CmpInst::getSwappedPredicate(Predicate);
1226 }
1227
1228 ConstantInt *Offset = nullptr;
1229 if (LHS != Val)
1230 match(LHS, m_Add(m_Specific(Val), m_ConstantInt(Offset)));
1231
1232 if (LHS == Val || Offset) {
1233 // Calculate the range of values that are allowed by the comparison
1234 ConstantRange RHSRange(RHS->getType()->getIntegerBitWidth(),
1235 /*isFullSet=*/true);
1236 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS))
1237 RHSRange = ConstantRange(CI->getValue());
1238 else if (Instruction *I = dyn_cast<Instruction>(RHS))
1239 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
1240 RHSRange = getConstantRangeFromMetadata(*Ranges);
1241
1242 // If we're interested in the false dest, invert the condition
1243 CmpInst::Predicate Pred =
1244 isTrueDest ? Predicate : CmpInst::getInversePredicate(Predicate);
1245 ConstantRange TrueValues =
1246 ConstantRange::makeAllowedICmpRegion(Pred, RHSRange);
1247
1248 if (Offset) // Apply the offset from above.
1249 TrueValues = TrueValues.subtract(Offset->getValue());
1250
1251 return ValueLatticeElement::getRange(std::move(TrueValues));
1252 }
1253
1254 return ValueLatticeElement::getOverdefined();
1255 }
1256
1257 // Handle conditions of the form
1258 // extractvalue(op.with.overflow(%x, C), 1).
getValueFromOverflowCondition(Value * Val,WithOverflowInst * WO,bool IsTrueDest)1259 static ValueLatticeElement getValueFromOverflowCondition(
1260 Value *Val, WithOverflowInst *WO, bool IsTrueDest) {
1261 // TODO: This only works with a constant RHS for now. We could also compute
1262 // the range of the RHS, but this doesn't fit into the current structure of
1263 // the edge value calculation.
1264 const APInt *C;
1265 if (WO->getLHS() != Val || !match(WO->getRHS(), m_APInt(C)))
1266 return ValueLatticeElement::getOverdefined();
1267
1268 // Calculate the possible values of %x for which no overflow occurs.
1269 ConstantRange NWR = ConstantRange::makeExactNoWrapRegion(
1270 WO->getBinaryOp(), *C, WO->getNoWrapKind());
1271
1272 // If overflow is false, %x is constrained to NWR. If overflow is true, %x is
1273 // constrained to it's inverse (all values that might cause overflow).
1274 if (IsTrueDest)
1275 NWR = NWR.inverse();
1276 return ValueLatticeElement::getRange(NWR);
1277 }
1278
1279 static ValueLatticeElement
1280 getValueFromCondition(Value *Val, Value *Cond, bool isTrueDest,
1281 DenseMap<Value*, ValueLatticeElement> &Visited);
1282
1283 static ValueLatticeElement
getValueFromConditionImpl(Value * Val,Value * Cond,bool isTrueDest,DenseMap<Value *,ValueLatticeElement> & Visited)1284 getValueFromConditionImpl(Value *Val, Value *Cond, bool isTrueDest,
1285 DenseMap<Value*, ValueLatticeElement> &Visited) {
1286 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Cond))
1287 return getValueFromICmpCondition(Val, ICI, isTrueDest);
1288
1289 if (auto *EVI = dyn_cast<ExtractValueInst>(Cond))
1290 if (auto *WO = dyn_cast<WithOverflowInst>(EVI->getAggregateOperand()))
1291 if (EVI->getNumIndices() == 1 && *EVI->idx_begin() == 1)
1292 return getValueFromOverflowCondition(Val, WO, isTrueDest);
1293
1294 // Handle conditions in the form of (cond1 && cond2), we know that on the
1295 // true dest path both of the conditions hold. Similarly for conditions of
1296 // the form (cond1 || cond2), we know that on the false dest path neither
1297 // condition holds.
1298 BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond);
1299 if (!BO || (isTrueDest && BO->getOpcode() != BinaryOperator::And) ||
1300 (!isTrueDest && BO->getOpcode() != BinaryOperator::Or))
1301 return ValueLatticeElement::getOverdefined();
1302
1303 // Prevent infinite recursion if Cond references itself as in this example:
1304 // Cond: "%tmp4 = and i1 %tmp4, undef"
1305 // BL: "%tmp4 = and i1 %tmp4, undef"
1306 // BR: "i1 undef"
1307 Value *BL = BO->getOperand(0);
1308 Value *BR = BO->getOperand(1);
1309 if (BL == Cond || BR == Cond)
1310 return ValueLatticeElement::getOverdefined();
1311
1312 return intersect(getValueFromCondition(Val, BL, isTrueDest, Visited),
1313 getValueFromCondition(Val, BR, isTrueDest, Visited));
1314 }
1315
1316 static ValueLatticeElement
getValueFromCondition(Value * Val,Value * Cond,bool isTrueDest,DenseMap<Value *,ValueLatticeElement> & Visited)1317 getValueFromCondition(Value *Val, Value *Cond, bool isTrueDest,
1318 DenseMap<Value*, ValueLatticeElement> &Visited) {
1319 auto I = Visited.find(Cond);
1320 if (I != Visited.end())
1321 return I->second;
1322
1323 auto Result = getValueFromConditionImpl(Val, Cond, isTrueDest, Visited);
1324 Visited[Cond] = Result;
1325 return Result;
1326 }
1327
getValueFromCondition(Value * Val,Value * Cond,bool isTrueDest)1328 ValueLatticeElement getValueFromCondition(Value *Val, Value *Cond,
1329 bool isTrueDest) {
1330 assert(Cond && "precondition");
1331 DenseMap<Value*, ValueLatticeElement> Visited;
1332 return getValueFromCondition(Val, Cond, isTrueDest, Visited);
1333 }
1334
1335 // Return true if Usr has Op as an operand, otherwise false.
usesOperand(User * Usr,Value * Op)1336 static bool usesOperand(User *Usr, Value *Op) {
1337 return find(Usr->operands(), Op) != Usr->op_end();
1338 }
1339
1340 // Return true if the instruction type of Val is supported by
1341 // constantFoldUser(). Currently CastInst and BinaryOperator only. Call this
1342 // before calling constantFoldUser() to find out if it's even worth attempting
1343 // to call it.
isOperationFoldable(User * Usr)1344 static bool isOperationFoldable(User *Usr) {
1345 return isa<CastInst>(Usr) || isa<BinaryOperator>(Usr);
1346 }
1347
1348 // Check if Usr can be simplified to an integer constant when the value of one
1349 // of its operands Op is an integer constant OpConstVal. If so, return it as an
1350 // lattice value range with a single element or otherwise return an overdefined
1351 // lattice value.
constantFoldUser(User * Usr,Value * Op,const APInt & OpConstVal,const DataLayout & DL)1352 static ValueLatticeElement constantFoldUser(User *Usr, Value *Op,
1353 const APInt &OpConstVal,
1354 const DataLayout &DL) {
1355 assert(isOperationFoldable(Usr) && "Precondition");
1356 Constant* OpConst = Constant::getIntegerValue(Op->getType(), OpConstVal);
1357 // Check if Usr can be simplified to a constant.
1358 if (auto *CI = dyn_cast<CastInst>(Usr)) {
1359 assert(CI->getOperand(0) == Op && "Operand 0 isn't Op");
1360 if (auto *C = dyn_cast_or_null<ConstantInt>(
1361 SimplifyCastInst(CI->getOpcode(), OpConst,
1362 CI->getDestTy(), DL))) {
1363 return ValueLatticeElement::getRange(ConstantRange(C->getValue()));
1364 }
1365 } else if (auto *BO = dyn_cast<BinaryOperator>(Usr)) {
1366 bool Op0Match = BO->getOperand(0) == Op;
1367 bool Op1Match = BO->getOperand(1) == Op;
1368 assert((Op0Match || Op1Match) &&
1369 "Operand 0 nor Operand 1 isn't a match");
1370 Value *LHS = Op0Match ? OpConst : BO->getOperand(0);
1371 Value *RHS = Op1Match ? OpConst : BO->getOperand(1);
1372 if (auto *C = dyn_cast_or_null<ConstantInt>(
1373 SimplifyBinOp(BO->getOpcode(), LHS, RHS, DL))) {
1374 return ValueLatticeElement::getRange(ConstantRange(C->getValue()));
1375 }
1376 }
1377 return ValueLatticeElement::getOverdefined();
1378 }
1379
1380 /// Compute the value of Val on the edge BBFrom -> BBTo. Returns false if
1381 /// Val is not constrained on the edge. Result is unspecified if return value
1382 /// is false.
getEdgeValueLocal(Value * Val,BasicBlock * BBFrom,BasicBlock * BBTo,ValueLatticeElement & Result)1383 static bool getEdgeValueLocal(Value *Val, BasicBlock *BBFrom,
1384 BasicBlock *BBTo, ValueLatticeElement &Result) {
1385 // TODO: Handle more complex conditionals. If (v == 0 || v2 < 1) is false, we
1386 // know that v != 0.
1387 if (BranchInst *BI = dyn_cast<BranchInst>(BBFrom->getTerminator())) {
1388 // If this is a conditional branch and only one successor goes to BBTo, then
1389 // we may be able to infer something from the condition.
1390 if (BI->isConditional() &&
1391 BI->getSuccessor(0) != BI->getSuccessor(1)) {
1392 bool isTrueDest = BI->getSuccessor(0) == BBTo;
1393 assert(BI->getSuccessor(!isTrueDest) == BBTo &&
1394 "BBTo isn't a successor of BBFrom");
1395 Value *Condition = BI->getCondition();
1396
1397 // If V is the condition of the branch itself, then we know exactly what
1398 // it is.
1399 if (Condition == Val) {
1400 Result = ValueLatticeElement::get(ConstantInt::get(
1401 Type::getInt1Ty(Val->getContext()), isTrueDest));
1402 return true;
1403 }
1404
1405 // If the condition of the branch is an equality comparison, we may be
1406 // able to infer the value.
1407 Result = getValueFromCondition(Val, Condition, isTrueDest);
1408 if (!Result.isOverdefined())
1409 return true;
1410
1411 if (User *Usr = dyn_cast<User>(Val)) {
1412 assert(Result.isOverdefined() && "Result isn't overdefined");
1413 // Check with isOperationFoldable() first to avoid linearly iterating
1414 // over the operands unnecessarily which can be expensive for
1415 // instructions with many operands.
1416 if (isa<IntegerType>(Usr->getType()) && isOperationFoldable(Usr)) {
1417 const DataLayout &DL = BBTo->getModule()->getDataLayout();
1418 if (usesOperand(Usr, Condition)) {
1419 // If Val has Condition as an operand and Val can be folded into a
1420 // constant with either Condition == true or Condition == false,
1421 // propagate the constant.
1422 // eg.
1423 // ; %Val is true on the edge to %then.
1424 // %Val = and i1 %Condition, true.
1425 // br %Condition, label %then, label %else
1426 APInt ConditionVal(1, isTrueDest ? 1 : 0);
1427 Result = constantFoldUser(Usr, Condition, ConditionVal, DL);
1428 } else {
1429 // If one of Val's operand has an inferred value, we may be able to
1430 // infer the value of Val.
1431 // eg.
1432 // ; %Val is 94 on the edge to %then.
1433 // %Val = add i8 %Op, 1
1434 // %Condition = icmp eq i8 %Op, 93
1435 // br i1 %Condition, label %then, label %else
1436 for (unsigned i = 0; i < Usr->getNumOperands(); ++i) {
1437 Value *Op = Usr->getOperand(i);
1438 ValueLatticeElement OpLatticeVal =
1439 getValueFromCondition(Op, Condition, isTrueDest);
1440 if (Optional<APInt> OpConst = OpLatticeVal.asConstantInteger()) {
1441 Result = constantFoldUser(Usr, Op, OpConst.getValue(), DL);
1442 break;
1443 }
1444 }
1445 }
1446 }
1447 }
1448 if (!Result.isOverdefined())
1449 return true;
1450 }
1451 }
1452
1453 // If the edge was formed by a switch on the value, then we may know exactly
1454 // what it is.
1455 if (SwitchInst *SI = dyn_cast<SwitchInst>(BBFrom->getTerminator())) {
1456 Value *Condition = SI->getCondition();
1457 if (!isa<IntegerType>(Val->getType()))
1458 return false;
1459 bool ValUsesConditionAndMayBeFoldable = false;
1460 if (Condition != Val) {
1461 // Check if Val has Condition as an operand.
1462 if (User *Usr = dyn_cast<User>(Val))
1463 ValUsesConditionAndMayBeFoldable = isOperationFoldable(Usr) &&
1464 usesOperand(Usr, Condition);
1465 if (!ValUsesConditionAndMayBeFoldable)
1466 return false;
1467 }
1468 assert((Condition == Val || ValUsesConditionAndMayBeFoldable) &&
1469 "Condition != Val nor Val doesn't use Condition");
1470
1471 bool DefaultCase = SI->getDefaultDest() == BBTo;
1472 unsigned BitWidth = Val->getType()->getIntegerBitWidth();
1473 ConstantRange EdgesVals(BitWidth, DefaultCase/*isFullSet*/);
1474
1475 for (auto Case : SI->cases()) {
1476 APInt CaseValue = Case.getCaseValue()->getValue();
1477 ConstantRange EdgeVal(CaseValue);
1478 if (ValUsesConditionAndMayBeFoldable) {
1479 User *Usr = cast<User>(Val);
1480 const DataLayout &DL = BBTo->getModule()->getDataLayout();
1481 ValueLatticeElement EdgeLatticeVal =
1482 constantFoldUser(Usr, Condition, CaseValue, DL);
1483 if (EdgeLatticeVal.isOverdefined())
1484 return false;
1485 EdgeVal = EdgeLatticeVal.getConstantRange();
1486 }
1487 if (DefaultCase) {
1488 // It is possible that the default destination is the destination of
1489 // some cases. We cannot perform difference for those cases.
1490 // We know Condition != CaseValue in BBTo. In some cases we can use
1491 // this to infer Val == f(Condition) is != f(CaseValue). For now, we
1492 // only do this when f is identity (i.e. Val == Condition), but we
1493 // should be able to do this for any injective f.
1494 if (Case.getCaseSuccessor() != BBTo && Condition == Val)
1495 EdgesVals = EdgesVals.difference(EdgeVal);
1496 } else if (Case.getCaseSuccessor() == BBTo)
1497 EdgesVals = EdgesVals.unionWith(EdgeVal);
1498 }
1499 Result = ValueLatticeElement::getRange(std::move(EdgesVals));
1500 return true;
1501 }
1502 return false;
1503 }
1504
1505 /// Compute the value of Val on the edge BBFrom -> BBTo or the value at
1506 /// the basic block if the edge does not constrain Val.
getEdgeValue(Value * Val,BasicBlock * BBFrom,BasicBlock * BBTo,ValueLatticeElement & Result,Instruction * CxtI)1507 bool LazyValueInfoImpl::getEdgeValue(Value *Val, BasicBlock *BBFrom,
1508 BasicBlock *BBTo,
1509 ValueLatticeElement &Result,
1510 Instruction *CxtI) {
1511 // If already a constant, there is nothing to compute.
1512 if (Constant *VC = dyn_cast<Constant>(Val)) {
1513 Result = ValueLatticeElement::get(VC);
1514 return true;
1515 }
1516
1517 ValueLatticeElement LocalResult;
1518 if (!getEdgeValueLocal(Val, BBFrom, BBTo, LocalResult))
1519 // If we couldn't constrain the value on the edge, LocalResult doesn't
1520 // provide any information.
1521 LocalResult = ValueLatticeElement::getOverdefined();
1522
1523 if (hasSingleValue(LocalResult)) {
1524 // Can't get any more precise here
1525 Result = LocalResult;
1526 return true;
1527 }
1528
1529 if (!hasBlockValue(Val, BBFrom)) {
1530 if (pushBlockValue(std::make_pair(BBFrom, Val)))
1531 return false;
1532 // No new information.
1533 Result = LocalResult;
1534 return true;
1535 }
1536
1537 // Try to intersect ranges of the BB and the constraint on the edge.
1538 ValueLatticeElement InBlock = getBlockValue(Val, BBFrom);
1539 intersectAssumeOrGuardBlockValueConstantRange(Val, InBlock,
1540 BBFrom->getTerminator());
1541 // We can use the context instruction (generically the ultimate instruction
1542 // the calling pass is trying to simplify) here, even though the result of
1543 // this function is generally cached when called from the solve* functions
1544 // (and that cached result might be used with queries using a different
1545 // context instruction), because when this function is called from the solve*
1546 // functions, the context instruction is not provided. When called from
1547 // LazyValueInfoImpl::getValueOnEdge, the context instruction is provided,
1548 // but then the result is not cached.
1549 intersectAssumeOrGuardBlockValueConstantRange(Val, InBlock, CxtI);
1550
1551 Result = intersect(LocalResult, InBlock);
1552 return true;
1553 }
1554
getValueInBlock(Value * V,BasicBlock * BB,Instruction * CxtI)1555 ValueLatticeElement LazyValueInfoImpl::getValueInBlock(Value *V, BasicBlock *BB,
1556 Instruction *CxtI) {
1557 LLVM_DEBUG(dbgs() << "LVI Getting block end value " << *V << " at '"
1558 << BB->getName() << "'\n");
1559
1560 assert(BlockValueStack.empty() && BlockValueSet.empty());
1561 if (!hasBlockValue(V, BB)) {
1562 pushBlockValue(std::make_pair(BB, V));
1563 solve();
1564 }
1565 ValueLatticeElement Result = getBlockValue(V, BB);
1566 intersectAssumeOrGuardBlockValueConstantRange(V, Result, CxtI);
1567
1568 LLVM_DEBUG(dbgs() << " Result = " << Result << "\n");
1569 return Result;
1570 }
1571
getValueAt(Value * V,Instruction * CxtI)1572 ValueLatticeElement LazyValueInfoImpl::getValueAt(Value *V, Instruction *CxtI) {
1573 LLVM_DEBUG(dbgs() << "LVI Getting value " << *V << " at '" << CxtI->getName()
1574 << "'\n");
1575
1576 if (auto *C = dyn_cast<Constant>(V))
1577 return ValueLatticeElement::get(C);
1578
1579 ValueLatticeElement Result = ValueLatticeElement::getOverdefined();
1580 if (auto *I = dyn_cast<Instruction>(V))
1581 Result = getFromRangeMetadata(I);
1582 intersectAssumeOrGuardBlockValueConstantRange(V, Result, CxtI);
1583
1584 LLVM_DEBUG(dbgs() << " Result = " << Result << "\n");
1585 return Result;
1586 }
1587
1588 ValueLatticeElement LazyValueInfoImpl::
getValueOnEdge(Value * V,BasicBlock * FromBB,BasicBlock * ToBB,Instruction * CxtI)1589 getValueOnEdge(Value *V, BasicBlock *FromBB, BasicBlock *ToBB,
1590 Instruction *CxtI) {
1591 LLVM_DEBUG(dbgs() << "LVI Getting edge value " << *V << " from '"
1592 << FromBB->getName() << "' to '" << ToBB->getName()
1593 << "'\n");
1594
1595 ValueLatticeElement Result;
1596 if (!getEdgeValue(V, FromBB, ToBB, Result, CxtI)) {
1597 solve();
1598 bool WasFastQuery = getEdgeValue(V, FromBB, ToBB, Result, CxtI);
1599 (void)WasFastQuery;
1600 assert(WasFastQuery && "More work to do after problem solved?");
1601 }
1602
1603 LLVM_DEBUG(dbgs() << " Result = " << Result << "\n");
1604 return Result;
1605 }
1606
threadEdge(BasicBlock * PredBB,BasicBlock * OldSucc,BasicBlock * NewSucc)1607 void LazyValueInfoImpl::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc,
1608 BasicBlock *NewSucc) {
1609 TheCache.threadEdgeImpl(OldSucc, NewSucc);
1610 }
1611
1612 //===----------------------------------------------------------------------===//
1613 // LazyValueInfo Impl
1614 //===----------------------------------------------------------------------===//
1615
1616 /// This lazily constructs the LazyValueInfoImpl.
getImpl(void * & PImpl,AssumptionCache * AC,const DataLayout * DL,DominatorTree * DT=nullptr)1617 static LazyValueInfoImpl &getImpl(void *&PImpl, AssumptionCache *AC,
1618 const DataLayout *DL,
1619 DominatorTree *DT = nullptr) {
1620 if (!PImpl) {
1621 assert(DL && "getCache() called with a null DataLayout");
1622 PImpl = new LazyValueInfoImpl(AC, *DL, DT);
1623 }
1624 return *static_cast<LazyValueInfoImpl*>(PImpl);
1625 }
1626
runOnFunction(Function & F)1627 bool LazyValueInfoWrapperPass::runOnFunction(Function &F) {
1628 Info.AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1629 const DataLayout &DL = F.getParent()->getDataLayout();
1630
1631 DominatorTreeWrapperPass *DTWP =
1632 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
1633 Info.DT = DTWP ? &DTWP->getDomTree() : nullptr;
1634 Info.TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
1635
1636 if (Info.PImpl)
1637 getImpl(Info.PImpl, Info.AC, &DL, Info.DT).clear();
1638
1639 // Fully lazy.
1640 return false;
1641 }
1642
getAnalysisUsage(AnalysisUsage & AU) const1643 void LazyValueInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1644 AU.setPreservesAll();
1645 AU.addRequired<AssumptionCacheTracker>();
1646 AU.addRequired<TargetLibraryInfoWrapperPass>();
1647 }
1648
getLVI()1649 LazyValueInfo &LazyValueInfoWrapperPass::getLVI() { return Info; }
1650
~LazyValueInfo()1651 LazyValueInfo::~LazyValueInfo() { releaseMemory(); }
1652
releaseMemory()1653 void LazyValueInfo::releaseMemory() {
1654 // If the cache was allocated, free it.
1655 if (PImpl) {
1656 delete &getImpl(PImpl, AC, nullptr);
1657 PImpl = nullptr;
1658 }
1659 }
1660
invalidate(Function & F,const PreservedAnalyses & PA,FunctionAnalysisManager::Invalidator & Inv)1661 bool LazyValueInfo::invalidate(Function &F, const PreservedAnalyses &PA,
1662 FunctionAnalysisManager::Invalidator &Inv) {
1663 // We need to invalidate if we have either failed to preserve this analyses
1664 // result directly or if any of its dependencies have been invalidated.
1665 auto PAC = PA.getChecker<LazyValueAnalysis>();
1666 if (!(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>()) ||
1667 (DT && Inv.invalidate<DominatorTreeAnalysis>(F, PA)))
1668 return true;
1669
1670 return false;
1671 }
1672
releaseMemory()1673 void LazyValueInfoWrapperPass::releaseMemory() { Info.releaseMemory(); }
1674
run(Function & F,FunctionAnalysisManager & FAM)1675 LazyValueInfo LazyValueAnalysis::run(Function &F,
1676 FunctionAnalysisManager &FAM) {
1677 auto &AC = FAM.getResult<AssumptionAnalysis>(F);
1678 auto &TLI = FAM.getResult<TargetLibraryAnalysis>(F);
1679 auto *DT = FAM.getCachedResult<DominatorTreeAnalysis>(F);
1680
1681 return LazyValueInfo(&AC, &F.getParent()->getDataLayout(), &TLI, DT);
1682 }
1683
1684 /// Returns true if we can statically tell that this value will never be a
1685 /// "useful" constant. In practice, this means we've got something like an
1686 /// alloca or a malloc call for which a comparison against a constant can
1687 /// only be guarding dead code. Note that we are potentially giving up some
1688 /// precision in dead code (a constant result) in favour of avoiding a
1689 /// expensive search for a easily answered common query.
isKnownNonConstant(Value * V)1690 static bool isKnownNonConstant(Value *V) {
1691 V = V->stripPointerCasts();
1692 // The return val of alloc cannot be a Constant.
1693 if (isa<AllocaInst>(V))
1694 return true;
1695 return false;
1696 }
1697
getConstant(Value * V,BasicBlock * BB,Instruction * CxtI)1698 Constant *LazyValueInfo::getConstant(Value *V, BasicBlock *BB,
1699 Instruction *CxtI) {
1700 // Bail out early if V is known not to be a Constant.
1701 if (isKnownNonConstant(V))
1702 return nullptr;
1703
1704 const DataLayout &DL = BB->getModule()->getDataLayout();
1705 ValueLatticeElement Result =
1706 getImpl(PImpl, AC, &DL, DT).getValueInBlock(V, BB, CxtI);
1707
1708 if (Result.isConstant())
1709 return Result.getConstant();
1710 if (Result.isConstantRange()) {
1711 const ConstantRange &CR = Result.getConstantRange();
1712 if (const APInt *SingleVal = CR.getSingleElement())
1713 return ConstantInt::get(V->getContext(), *SingleVal);
1714 }
1715 return nullptr;
1716 }
1717
getConstantRange(Value * V,BasicBlock * BB,Instruction * CxtI)1718 ConstantRange LazyValueInfo::getConstantRange(Value *V, BasicBlock *BB,
1719 Instruction *CxtI) {
1720 assert(V->getType()->isIntegerTy());
1721 unsigned Width = V->getType()->getIntegerBitWidth();
1722 const DataLayout &DL = BB->getModule()->getDataLayout();
1723 ValueLatticeElement Result =
1724 getImpl(PImpl, AC, &DL, DT).getValueInBlock(V, BB, CxtI);
1725 if (Result.isUndefined())
1726 return ConstantRange::getEmpty(Width);
1727 if (Result.isConstantRange())
1728 return Result.getConstantRange();
1729 // We represent ConstantInt constants as constant ranges but other kinds
1730 // of integer constants, i.e. ConstantExpr will be tagged as constants
1731 assert(!(Result.isConstant() && isa<ConstantInt>(Result.getConstant())) &&
1732 "ConstantInt value must be represented as constantrange");
1733 return ConstantRange::getFull(Width);
1734 }
1735
1736 /// Determine whether the specified value is known to be a
1737 /// constant on the specified edge. Return null if not.
getConstantOnEdge(Value * V,BasicBlock * FromBB,BasicBlock * ToBB,Instruction * CxtI)1738 Constant *LazyValueInfo::getConstantOnEdge(Value *V, BasicBlock *FromBB,
1739 BasicBlock *ToBB,
1740 Instruction *CxtI) {
1741 const DataLayout &DL = FromBB->getModule()->getDataLayout();
1742 ValueLatticeElement Result =
1743 getImpl(PImpl, AC, &DL, DT).getValueOnEdge(V, FromBB, ToBB, CxtI);
1744
1745 if (Result.isConstant())
1746 return Result.getConstant();
1747 if (Result.isConstantRange()) {
1748 const ConstantRange &CR = Result.getConstantRange();
1749 if (const APInt *SingleVal = CR.getSingleElement())
1750 return ConstantInt::get(V->getContext(), *SingleVal);
1751 }
1752 return nullptr;
1753 }
1754
getConstantRangeOnEdge(Value * V,BasicBlock * FromBB,BasicBlock * ToBB,Instruction * CxtI)1755 ConstantRange LazyValueInfo::getConstantRangeOnEdge(Value *V,
1756 BasicBlock *FromBB,
1757 BasicBlock *ToBB,
1758 Instruction *CxtI) {
1759 unsigned Width = V->getType()->getIntegerBitWidth();
1760 const DataLayout &DL = FromBB->getModule()->getDataLayout();
1761 ValueLatticeElement Result =
1762 getImpl(PImpl, AC, &DL, DT).getValueOnEdge(V, FromBB, ToBB, CxtI);
1763
1764 if (Result.isUndefined())
1765 return ConstantRange::getEmpty(Width);
1766 if (Result.isConstantRange())
1767 return Result.getConstantRange();
1768 // We represent ConstantInt constants as constant ranges but other kinds
1769 // of integer constants, i.e. ConstantExpr will be tagged as constants
1770 assert(!(Result.isConstant() && isa<ConstantInt>(Result.getConstant())) &&
1771 "ConstantInt value must be represented as constantrange");
1772 return ConstantRange::getFull(Width);
1773 }
1774
1775 static LazyValueInfo::Tristate
getPredicateResult(unsigned Pred,Constant * C,const ValueLatticeElement & Val,const DataLayout & DL,TargetLibraryInfo * TLI)1776 getPredicateResult(unsigned Pred, Constant *C, const ValueLatticeElement &Val,
1777 const DataLayout &DL, TargetLibraryInfo *TLI) {
1778 // If we know the value is a constant, evaluate the conditional.
1779 Constant *Res = nullptr;
1780 if (Val.isConstant()) {
1781 Res = ConstantFoldCompareInstOperands(Pred, Val.getConstant(), C, DL, TLI);
1782 if (ConstantInt *ResCI = dyn_cast<ConstantInt>(Res))
1783 return ResCI->isZero() ? LazyValueInfo::False : LazyValueInfo::True;
1784 return LazyValueInfo::Unknown;
1785 }
1786
1787 if (Val.isConstantRange()) {
1788 ConstantInt *CI = dyn_cast<ConstantInt>(C);
1789 if (!CI) return LazyValueInfo::Unknown;
1790
1791 const ConstantRange &CR = Val.getConstantRange();
1792 if (Pred == ICmpInst::ICMP_EQ) {
1793 if (!CR.contains(CI->getValue()))
1794 return LazyValueInfo::False;
1795
1796 if (CR.isSingleElement())
1797 return LazyValueInfo::True;
1798 } else if (Pred == ICmpInst::ICMP_NE) {
1799 if (!CR.contains(CI->getValue()))
1800 return LazyValueInfo::True;
1801
1802 if (CR.isSingleElement())
1803 return LazyValueInfo::False;
1804 } else {
1805 // Handle more complex predicates.
1806 ConstantRange TrueValues = ConstantRange::makeExactICmpRegion(
1807 (ICmpInst::Predicate)Pred, CI->getValue());
1808 if (TrueValues.contains(CR))
1809 return LazyValueInfo::True;
1810 if (TrueValues.inverse().contains(CR))
1811 return LazyValueInfo::False;
1812 }
1813 return LazyValueInfo::Unknown;
1814 }
1815
1816 if (Val.isNotConstant()) {
1817 // If this is an equality comparison, we can try to fold it knowing that
1818 // "V != C1".
1819 if (Pred == ICmpInst::ICMP_EQ) {
1820 // !C1 == C -> false iff C1 == C.
1821 Res = ConstantFoldCompareInstOperands(ICmpInst::ICMP_NE,
1822 Val.getNotConstant(), C, DL,
1823 TLI);
1824 if (Res->isNullValue())
1825 return LazyValueInfo::False;
1826 } else if (Pred == ICmpInst::ICMP_NE) {
1827 // !C1 != C -> true iff C1 == C.
1828 Res = ConstantFoldCompareInstOperands(ICmpInst::ICMP_NE,
1829 Val.getNotConstant(), C, DL,
1830 TLI);
1831 if (Res->isNullValue())
1832 return LazyValueInfo::True;
1833 }
1834 return LazyValueInfo::Unknown;
1835 }
1836
1837 return LazyValueInfo::Unknown;
1838 }
1839
1840 /// Determine whether the specified value comparison with a constant is known to
1841 /// be true or false on the specified CFG edge. Pred is a CmpInst predicate.
1842 LazyValueInfo::Tristate
getPredicateOnEdge(unsigned Pred,Value * V,Constant * C,BasicBlock * FromBB,BasicBlock * ToBB,Instruction * CxtI)1843 LazyValueInfo::getPredicateOnEdge(unsigned Pred, Value *V, Constant *C,
1844 BasicBlock *FromBB, BasicBlock *ToBB,
1845 Instruction *CxtI) {
1846 const DataLayout &DL = FromBB->getModule()->getDataLayout();
1847 ValueLatticeElement Result =
1848 getImpl(PImpl, AC, &DL, DT).getValueOnEdge(V, FromBB, ToBB, CxtI);
1849
1850 return getPredicateResult(Pred, C, Result, DL, TLI);
1851 }
1852
1853 LazyValueInfo::Tristate
getPredicateAt(unsigned Pred,Value * V,Constant * C,Instruction * CxtI)1854 LazyValueInfo::getPredicateAt(unsigned Pred, Value *V, Constant *C,
1855 Instruction *CxtI) {
1856 // Is or is not NonNull are common predicates being queried. If
1857 // isKnownNonZero can tell us the result of the predicate, we can
1858 // return it quickly. But this is only a fastpath, and falling
1859 // through would still be correct.
1860 const DataLayout &DL = CxtI->getModule()->getDataLayout();
1861 if (V->getType()->isPointerTy() && C->isNullValue() &&
1862 isKnownNonZero(V->stripPointerCastsSameRepresentation(), DL)) {
1863 if (Pred == ICmpInst::ICMP_EQ)
1864 return LazyValueInfo::False;
1865 else if (Pred == ICmpInst::ICMP_NE)
1866 return LazyValueInfo::True;
1867 }
1868 ValueLatticeElement Result = getImpl(PImpl, AC, &DL, DT).getValueAt(V, CxtI);
1869 Tristate Ret = getPredicateResult(Pred, C, Result, DL, TLI);
1870 if (Ret != Unknown)
1871 return Ret;
1872
1873 // Note: The following bit of code is somewhat distinct from the rest of LVI;
1874 // LVI as a whole tries to compute a lattice value which is conservatively
1875 // correct at a given location. In this case, we have a predicate which we
1876 // weren't able to prove about the merged result, and we're pushing that
1877 // predicate back along each incoming edge to see if we can prove it
1878 // separately for each input. As a motivating example, consider:
1879 // bb1:
1880 // %v1 = ... ; constantrange<1, 5>
1881 // br label %merge
1882 // bb2:
1883 // %v2 = ... ; constantrange<10, 20>
1884 // br label %merge
1885 // merge:
1886 // %phi = phi [%v1, %v2] ; constantrange<1,20>
1887 // %pred = icmp eq i32 %phi, 8
1888 // We can't tell from the lattice value for '%phi' that '%pred' is false
1889 // along each path, but by checking the predicate over each input separately,
1890 // we can.
1891 // We limit the search to one step backwards from the current BB and value.
1892 // We could consider extending this to search further backwards through the
1893 // CFG and/or value graph, but there are non-obvious compile time vs quality
1894 // tradeoffs.
1895 if (CxtI) {
1896 BasicBlock *BB = CxtI->getParent();
1897
1898 // Function entry or an unreachable block. Bail to avoid confusing
1899 // analysis below.
1900 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
1901 if (PI == PE)
1902 return Unknown;
1903
1904 // If V is a PHI node in the same block as the context, we need to ask
1905 // questions about the predicate as applied to the incoming value along
1906 // each edge. This is useful for eliminating cases where the predicate is
1907 // known along all incoming edges.
1908 if (auto *PHI = dyn_cast<PHINode>(V))
1909 if (PHI->getParent() == BB) {
1910 Tristate Baseline = Unknown;
1911 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i < e; i++) {
1912 Value *Incoming = PHI->getIncomingValue(i);
1913 BasicBlock *PredBB = PHI->getIncomingBlock(i);
1914 // Note that PredBB may be BB itself.
1915 Tristate Result = getPredicateOnEdge(Pred, Incoming, C, PredBB, BB,
1916 CxtI);
1917
1918 // Keep going as long as we've seen a consistent known result for
1919 // all inputs.
1920 Baseline = (i == 0) ? Result /* First iteration */
1921 : (Baseline == Result ? Baseline : Unknown); /* All others */
1922 if (Baseline == Unknown)
1923 break;
1924 }
1925 if (Baseline != Unknown)
1926 return Baseline;
1927 }
1928
1929 // For a comparison where the V is outside this block, it's possible
1930 // that we've branched on it before. Look to see if the value is known
1931 // on all incoming edges.
1932 if (!isa<Instruction>(V) ||
1933 cast<Instruction>(V)->getParent() != BB) {
1934 // For predecessor edge, determine if the comparison is true or false
1935 // on that edge. If they're all true or all false, we can conclude
1936 // the value of the comparison in this block.
1937 Tristate Baseline = getPredicateOnEdge(Pred, V, C, *PI, BB, CxtI);
1938 if (Baseline != Unknown) {
1939 // Check that all remaining incoming values match the first one.
1940 while (++PI != PE) {
1941 Tristate Ret = getPredicateOnEdge(Pred, V, C, *PI, BB, CxtI);
1942 if (Ret != Baseline) break;
1943 }
1944 // If we terminated early, then one of the values didn't match.
1945 if (PI == PE) {
1946 return Baseline;
1947 }
1948 }
1949 }
1950 }
1951 return Unknown;
1952 }
1953
threadEdge(BasicBlock * PredBB,BasicBlock * OldSucc,BasicBlock * NewSucc)1954 void LazyValueInfo::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc,
1955 BasicBlock *NewSucc) {
1956 if (PImpl) {
1957 const DataLayout &DL = PredBB->getModule()->getDataLayout();
1958 getImpl(PImpl, AC, &DL, DT).threadEdge(PredBB, OldSucc, NewSucc);
1959 }
1960 }
1961
eraseBlock(BasicBlock * BB)1962 void LazyValueInfo::eraseBlock(BasicBlock *BB) {
1963 if (PImpl) {
1964 const DataLayout &DL = BB->getModule()->getDataLayout();
1965 getImpl(PImpl, AC, &DL, DT).eraseBlock(BB);
1966 }
1967 }
1968
1969
printLVI(Function & F,DominatorTree & DTree,raw_ostream & OS)1970 void LazyValueInfo::printLVI(Function &F, DominatorTree &DTree, raw_ostream &OS) {
1971 if (PImpl) {
1972 getImpl(PImpl, AC, DL, DT).printLVI(F, DTree, OS);
1973 }
1974 }
1975
disableDT()1976 void LazyValueInfo::disableDT() {
1977 if (PImpl)
1978 getImpl(PImpl, AC, DL, DT).disableDT();
1979 }
1980
enableDT()1981 void LazyValueInfo::enableDT() {
1982 if (PImpl)
1983 getImpl(PImpl, AC, DL, DT).enableDT();
1984 }
1985
1986 // Print the LVI for the function arguments at the start of each basic block.
emitBasicBlockStartAnnot(const BasicBlock * BB,formatted_raw_ostream & OS)1987 void LazyValueInfoAnnotatedWriter::emitBasicBlockStartAnnot(
1988 const BasicBlock *BB, formatted_raw_ostream &OS) {
1989 // Find if there are latticevalues defined for arguments of the function.
1990 auto *F = BB->getParent();
1991 for (auto &Arg : F->args()) {
1992 ValueLatticeElement Result = LVIImpl->getValueInBlock(
1993 const_cast<Argument *>(&Arg), const_cast<BasicBlock *>(BB));
1994 if (Result.isUndefined())
1995 continue;
1996 OS << "; LatticeVal for: '" << Arg << "' is: " << Result << "\n";
1997 }
1998 }
1999
2000 // This function prints the LVI analysis for the instruction I at the beginning
2001 // of various basic blocks. It relies on calculated values that are stored in
2002 // the LazyValueInfoCache, and in the absence of cached values, recalculate the
2003 // LazyValueInfo for `I`, and print that info.
emitInstructionAnnot(const Instruction * I,formatted_raw_ostream & OS)2004 void LazyValueInfoAnnotatedWriter::emitInstructionAnnot(
2005 const Instruction *I, formatted_raw_ostream &OS) {
2006
2007 auto *ParentBB = I->getParent();
2008 SmallPtrSet<const BasicBlock*, 16> BlocksContainingLVI;
2009 // We can generate (solve) LVI values only for blocks that are dominated by
2010 // the I's parent. However, to avoid generating LVI for all dominating blocks,
2011 // that contain redundant/uninteresting information, we print LVI for
2012 // blocks that may use this LVI information (such as immediate successor
2013 // blocks, and blocks that contain uses of `I`).
2014 auto printResult = [&](const BasicBlock *BB) {
2015 if (!BlocksContainingLVI.insert(BB).second)
2016 return;
2017 ValueLatticeElement Result = LVIImpl->getValueInBlock(
2018 const_cast<Instruction *>(I), const_cast<BasicBlock *>(BB));
2019 OS << "; LatticeVal for: '" << *I << "' in BB: '";
2020 BB->printAsOperand(OS, false);
2021 OS << "' is: " << Result << "\n";
2022 };
2023
2024 printResult(ParentBB);
2025 // Print the LVI analysis results for the immediate successor blocks, that
2026 // are dominated by `ParentBB`.
2027 for (auto *BBSucc : successors(ParentBB))
2028 if (DT.dominates(ParentBB, BBSucc))
2029 printResult(BBSucc);
2030
2031 // Print LVI in blocks where `I` is used.
2032 for (auto *U : I->users())
2033 if (auto *UseI = dyn_cast<Instruction>(U))
2034 if (!isa<PHINode>(UseI) || DT.dominates(ParentBB, UseI->getParent()))
2035 printResult(UseI->getParent());
2036
2037 }
2038
2039 namespace {
2040 // Printer class for LazyValueInfo results.
2041 class LazyValueInfoPrinter : public FunctionPass {
2042 public:
2043 static char ID; // Pass identification, replacement for typeid
LazyValueInfoPrinter()2044 LazyValueInfoPrinter() : FunctionPass(ID) {
2045 initializeLazyValueInfoPrinterPass(*PassRegistry::getPassRegistry());
2046 }
2047
getAnalysisUsage(AnalysisUsage & AU) const2048 void getAnalysisUsage(AnalysisUsage &AU) const override {
2049 AU.setPreservesAll();
2050 AU.addRequired<LazyValueInfoWrapperPass>();
2051 AU.addRequired<DominatorTreeWrapperPass>();
2052 }
2053
2054 // Get the mandatory dominator tree analysis and pass this in to the
2055 // LVIPrinter. We cannot rely on the LVI's DT, since it's optional.
runOnFunction(Function & F)2056 bool runOnFunction(Function &F) override {
2057 dbgs() << "LVI for function '" << F.getName() << "':\n";
2058 auto &LVI = getAnalysis<LazyValueInfoWrapperPass>().getLVI();
2059 auto &DTree = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2060 LVI.printLVI(F, DTree, dbgs());
2061 return false;
2062 }
2063 };
2064 }
2065
2066 char LazyValueInfoPrinter::ID = 0;
2067 INITIALIZE_PASS_BEGIN(LazyValueInfoPrinter, "print-lazy-value-info",
2068 "Lazy Value Info Printer Pass", false, false)
2069 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)
2070 INITIALIZE_PASS_END(LazyValueInfoPrinter, "print-lazy-value-info",
2071 "Lazy Value Info Printer Pass", false, false)
2072