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/KnownBits.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include <map>
42 using namespace llvm;
43 using namespace PatternMatch;
44 
45 #define DEBUG_TYPE "lazy-value-info"
46 
47 // This is the number of worklist items we will process to try to discover an
48 // answer for a given value.
49 static const unsigned MaxProcessedPerValue = 500;
50 
51 char LazyValueInfoWrapperPass::ID = 0;
LazyValueInfoWrapperPass()52 LazyValueInfoWrapperPass::LazyValueInfoWrapperPass() : FunctionPass(ID) {
53   initializeLazyValueInfoWrapperPassPass(*PassRegistry::getPassRegistry());
54 }
55 INITIALIZE_PASS_BEGIN(LazyValueInfoWrapperPass, "lazy-value-info",
56                 "Lazy Value Information Analysis", false, true)
57 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
58 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
59 INITIALIZE_PASS_END(LazyValueInfoWrapperPass, "lazy-value-info",
60                 "Lazy Value Information Analysis", false, true)
61 
62 namespace llvm {
createLazyValueInfoPass()63   FunctionPass *createLazyValueInfoPass() { return new LazyValueInfoWrapperPass(); }
64 }
65 
66 AnalysisKey LazyValueAnalysis::Key;
67 
68 /// Returns true if this lattice value represents at most one possible value.
69 /// This is as precise as any lattice value can get while still representing
70 /// reachable code.
hasSingleValue(const ValueLatticeElement & Val)71 static bool hasSingleValue(const ValueLatticeElement &Val) {
72   if (Val.isConstantRange() &&
73       Val.getConstantRange().isSingleElement())
74     // Integer constants are single element ranges
75     return true;
76   if (Val.isConstant())
77     // Non integer constants
78     return true;
79   return false;
80 }
81 
82 /// Combine two sets of facts about the same value into a single set of
83 /// facts.  Note that this method is not suitable for merging facts along
84 /// different paths in a CFG; that's what the mergeIn function is for.  This
85 /// is for merging facts gathered about the same value at the same location
86 /// through two independent means.
87 /// Notes:
88 /// * This method does not promise to return the most precise possible lattice
89 ///   value implied by A and B.  It is allowed to return any lattice element
90 ///   which is at least as strong as *either* A or B (unless our facts
91 ///   conflict, see below).
92 /// * Due to unreachable code, the intersection of two lattice values could be
93 ///   contradictory.  If this happens, we return some valid lattice value so as
94 ///   not confuse the rest of LVI.  Ideally, we'd always return Undefined, but
95 ///   we do not make this guarantee.  TODO: This would be a useful enhancement.
intersect(const ValueLatticeElement & A,const ValueLatticeElement & B)96 static ValueLatticeElement intersect(const ValueLatticeElement &A,
97                                      const ValueLatticeElement &B) {
98   // Undefined is the strongest state.  It means the value is known to be along
99   // an unreachable path.
100   if (A.isUnknown())
101     return A;
102   if (B.isUnknown())
103     return B;
104 
105   // If we gave up for one, but got a useable fact from the other, use it.
106   if (A.isOverdefined())
107     return B;
108   if (B.isOverdefined())
109     return A;
110 
111   // Can't get any more precise than constants.
112   if (hasSingleValue(A))
113     return A;
114   if (hasSingleValue(B))
115     return B;
116 
117   // Could be either constant range or not constant here.
118   if (!A.isConstantRange() || !B.isConstantRange()) {
119     // TODO: Arbitrary choice, could be improved
120     return A;
121   }
122 
123   // Intersect two constant ranges
124   ConstantRange Range =
125       A.getConstantRange().intersectWith(B.getConstantRange());
126   // Note: An empty range is implicitly converted to unknown or undef depending
127   // on MayIncludeUndef internally.
128   return ValueLatticeElement::getRange(
129       std::move(Range), /*MayIncludeUndef=*/A.isConstantRangeIncludingUndef() |
130                             B.isConstantRangeIncludingUndef());
131 }
132 
133 //===----------------------------------------------------------------------===//
134 //                          LazyValueInfoCache Decl
135 //===----------------------------------------------------------------------===//
136 
137 namespace {
138   /// A callback value handle updates the cache when values are erased.
139   class LazyValueInfoCache;
140   struct LVIValueHandle final : public CallbackVH {
141     LazyValueInfoCache *Parent;
142 
LVIValueHandle__anonb19ab8520111::LVIValueHandle143     LVIValueHandle(Value *V, LazyValueInfoCache *P = nullptr)
144       : CallbackVH(V), Parent(P) { }
145 
146     void deleted() override;
allUsesReplacedWith__anonb19ab8520111::LVIValueHandle147     void allUsesReplacedWith(Value *V) override {
148       deleted();
149     }
150   };
151 } // end anonymous namespace
152 
153 namespace {
154   using NonNullPointerSet = SmallDenseSet<AssertingVH<Value>, 2>;
155 
156   /// This is the cache kept by LazyValueInfo which
157   /// maintains information about queries across the clients' queries.
158   class LazyValueInfoCache {
159     /// This is all of the cached information for one basic block. It contains
160     /// the per-value lattice elements, as well as a separate set for
161     /// overdefined values to reduce memory usage. Additionally pointers
162     /// dereferenced in the block are cached for nullability queries.
163     struct BlockCacheEntry {
164       SmallDenseMap<AssertingVH<Value>, ValueLatticeElement, 4> LatticeElements;
165       SmallDenseSet<AssertingVH<Value>, 4> OverDefined;
166       // None indicates that the nonnull pointers for this basic block
167       // block have not been computed yet.
168       Optional<NonNullPointerSet> NonNullPointers;
169     };
170 
171     /// Cached information per basic block.
172     DenseMap<PoisoningVH<BasicBlock>, std::unique_ptr<BlockCacheEntry>>
173         BlockCache;
174     /// Set of value handles used to erase values from the cache on deletion.
175     DenseSet<LVIValueHandle, DenseMapInfo<Value *>> ValueHandles;
176 
getBlockEntry(BasicBlock * BB) const177     const BlockCacheEntry *getBlockEntry(BasicBlock *BB) const {
178       auto It = BlockCache.find_as(BB);
179       if (It == BlockCache.end())
180         return nullptr;
181       return It->second.get();
182     }
183 
getOrCreateBlockEntry(BasicBlock * BB)184     BlockCacheEntry *getOrCreateBlockEntry(BasicBlock *BB) {
185       auto It = BlockCache.find_as(BB);
186       if (It == BlockCache.end())
187         It = BlockCache.insert({ BB, std::make_unique<BlockCacheEntry>() })
188                        .first;
189 
190       return It->second.get();
191     }
192 
addValueHandle(Value * Val)193     void addValueHandle(Value *Val) {
194       auto HandleIt = ValueHandles.find_as(Val);
195       if (HandleIt == ValueHandles.end())
196         ValueHandles.insert({ Val, this });
197     }
198 
199   public:
insertResult(Value * Val,BasicBlock * BB,const ValueLatticeElement & Result)200     void insertResult(Value *Val, BasicBlock *BB,
201                       const ValueLatticeElement &Result) {
202       BlockCacheEntry *Entry = getOrCreateBlockEntry(BB);
203 
204       // Insert over-defined values into their own cache to reduce memory
205       // overhead.
206       if (Result.isOverdefined())
207         Entry->OverDefined.insert(Val);
208       else
209         Entry->LatticeElements.insert({ Val, Result });
210 
211       addValueHandle(Val);
212     }
213 
getCachedValueInfo(Value * V,BasicBlock * BB) const214     Optional<ValueLatticeElement> getCachedValueInfo(Value *V,
215                                                      BasicBlock *BB) const {
216       const BlockCacheEntry *Entry = getBlockEntry(BB);
217       if (!Entry)
218         return None;
219 
220       if (Entry->OverDefined.count(V))
221         return ValueLatticeElement::getOverdefined();
222 
223       auto LatticeIt = Entry->LatticeElements.find_as(V);
224       if (LatticeIt == Entry->LatticeElements.end())
225         return None;
226 
227       return LatticeIt->second;
228     }
229 
isNonNullAtEndOfBlock(Value * V,BasicBlock * BB,function_ref<NonNullPointerSet (BasicBlock *)> InitFn)230     bool isNonNullAtEndOfBlock(
231         Value *V, BasicBlock *BB,
232         function_ref<NonNullPointerSet(BasicBlock *)> InitFn) {
233       BlockCacheEntry *Entry = getOrCreateBlockEntry(BB);
234       if (!Entry->NonNullPointers) {
235         Entry->NonNullPointers = InitFn(BB);
236         for (Value *V : *Entry->NonNullPointers)
237           addValueHandle(V);
238       }
239 
240       return Entry->NonNullPointers->count(V);
241     }
242 
243     /// clear - Empty the cache.
clear()244     void clear() {
245       BlockCache.clear();
246       ValueHandles.clear();
247     }
248 
249     /// Inform the cache that a given value has been deleted.
250     void eraseValue(Value *V);
251 
252     /// This is part of the update interface to inform the cache
253     /// that a block has been deleted.
254     void eraseBlock(BasicBlock *BB);
255 
256     /// Updates the cache to remove any influence an overdefined value in
257     /// OldSucc might have (unless also overdefined in NewSucc).  This just
258     /// flushes elements from the cache and does not add any.
259     void threadEdgeImpl(BasicBlock *OldSucc,BasicBlock *NewSucc);
260   };
261 }
262 
eraseValue(Value * V)263 void LazyValueInfoCache::eraseValue(Value *V) {
264   for (auto &Pair : BlockCache) {
265     Pair.second->LatticeElements.erase(V);
266     Pair.second->OverDefined.erase(V);
267     if (Pair.second->NonNullPointers)
268       Pair.second->NonNullPointers->erase(V);
269   }
270 
271   auto HandleIt = ValueHandles.find_as(V);
272   if (HandleIt != ValueHandles.end())
273     ValueHandles.erase(HandleIt);
274 }
275 
deleted()276 void LVIValueHandle::deleted() {
277   // This erasure deallocates *this, so it MUST happen after we're done
278   // using any and all members of *this.
279   Parent->eraseValue(*this);
280 }
281 
eraseBlock(BasicBlock * BB)282 void LazyValueInfoCache::eraseBlock(BasicBlock *BB) {
283   BlockCache.erase(BB);
284 }
285 
threadEdgeImpl(BasicBlock * OldSucc,BasicBlock * NewSucc)286 void LazyValueInfoCache::threadEdgeImpl(BasicBlock *OldSucc,
287                                         BasicBlock *NewSucc) {
288   // When an edge in the graph has been threaded, values that we could not
289   // determine a value for before (i.e. were marked overdefined) may be
290   // possible to solve now. We do NOT try to proactively update these values.
291   // Instead, we clear their entries from the cache, and allow lazy updating to
292   // recompute them when needed.
293 
294   // The updating process is fairly simple: we need to drop cached info
295   // for all values that were marked overdefined in OldSucc, and for those same
296   // values in any successor of OldSucc (except NewSucc) in which they were
297   // also marked overdefined.
298   std::vector<BasicBlock*> worklist;
299   worklist.push_back(OldSucc);
300 
301   const BlockCacheEntry *Entry = getBlockEntry(OldSucc);
302   if (!Entry || Entry->OverDefined.empty())
303     return; // Nothing to process here.
304   SmallVector<Value *, 4> ValsToClear(Entry->OverDefined.begin(),
305                                       Entry->OverDefined.end());
306 
307   // Use a worklist to perform a depth-first search of OldSucc's successors.
308   // NOTE: We do not need a visited list since any blocks we have already
309   // visited will have had their overdefined markers cleared already, and we
310   // thus won't loop to their successors.
311   while (!worklist.empty()) {
312     BasicBlock *ToUpdate = worklist.back();
313     worklist.pop_back();
314 
315     // Skip blocks only accessible through NewSucc.
316     if (ToUpdate == NewSucc) continue;
317 
318     // If a value was marked overdefined in OldSucc, and is here too...
319     auto OI = BlockCache.find_as(ToUpdate);
320     if (OI == BlockCache.end() || OI->second->OverDefined.empty())
321       continue;
322     auto &ValueSet = OI->second->OverDefined;
323 
324     bool changed = false;
325     for (Value *V : ValsToClear) {
326       if (!ValueSet.erase(V))
327         continue;
328 
329       // If we removed anything, then we potentially need to update
330       // blocks successors too.
331       changed = true;
332     }
333 
334     if (!changed) continue;
335 
336     llvm::append_range(worklist, successors(ToUpdate));
337   }
338 }
339 
340 
341 namespace {
342 /// An assembly annotator class to print LazyValueCache information in
343 /// comments.
344 class LazyValueInfoImpl;
345 class LazyValueInfoAnnotatedWriter : public AssemblyAnnotationWriter {
346   LazyValueInfoImpl *LVIImpl;
347   // While analyzing which blocks we can solve values for, we need the dominator
348   // information.
349   DominatorTree &DT;
350 
351 public:
LazyValueInfoAnnotatedWriter(LazyValueInfoImpl * L,DominatorTree & DTree)352   LazyValueInfoAnnotatedWriter(LazyValueInfoImpl *L, DominatorTree &DTree)
353       : LVIImpl(L), DT(DTree) {}
354 
355   void emitBasicBlockStartAnnot(const BasicBlock *BB,
356                                 formatted_raw_ostream &OS) override;
357 
358   void emitInstructionAnnot(const Instruction *I,
359                             formatted_raw_ostream &OS) override;
360 };
361 }
362 namespace {
363 // The actual implementation of the lazy analysis and update.  Note that the
364 // inheritance from LazyValueInfoCache is intended to be temporary while
365 // splitting the code and then transitioning to a has-a relationship.
366 class LazyValueInfoImpl {
367 
368   /// Cached results from previous queries
369   LazyValueInfoCache TheCache;
370 
371   /// This stack holds the state of the value solver during a query.
372   /// It basically emulates the callstack of the naive
373   /// recursive value lookup process.
374   SmallVector<std::pair<BasicBlock*, Value*>, 8> BlockValueStack;
375 
376   /// Keeps track of which block-value pairs are in BlockValueStack.
377   DenseSet<std::pair<BasicBlock*, Value*> > BlockValueSet;
378 
379   /// Push BV onto BlockValueStack unless it's already in there.
380   /// Returns true on success.
pushBlockValue(const std::pair<BasicBlock *,Value * > & BV)381   bool pushBlockValue(const std::pair<BasicBlock *, Value *> &BV) {
382     if (!BlockValueSet.insert(BV).second)
383       return false;  // It's already in the stack.
384 
385     LLVM_DEBUG(dbgs() << "PUSH: " << *BV.second << " in "
386                       << BV.first->getName() << "\n");
387     BlockValueStack.push_back(BV);
388     return true;
389   }
390 
391   AssumptionCache *AC;  ///< A pointer to the cache of @llvm.assume calls.
392   const DataLayout &DL; ///< A mandatory DataLayout
393 
394   /// Declaration of the llvm.experimental.guard() intrinsic,
395   /// if it exists in the module.
396   Function *GuardDecl;
397 
398   Optional<ValueLatticeElement> getBlockValue(Value *Val, BasicBlock *BB);
399   Optional<ValueLatticeElement> getEdgeValue(Value *V, BasicBlock *F,
400                                 BasicBlock *T, Instruction *CxtI = nullptr);
401 
402   // These methods process one work item and may add more. A false value
403   // returned means that the work item was not completely processed and must
404   // be revisited after going through the new items.
405   bool solveBlockValue(Value *Val, BasicBlock *BB);
406   Optional<ValueLatticeElement> solveBlockValueImpl(Value *Val, BasicBlock *BB);
407   Optional<ValueLatticeElement> solveBlockValueNonLocal(Value *Val,
408                                                         BasicBlock *BB);
409   Optional<ValueLatticeElement> solveBlockValuePHINode(PHINode *PN,
410                                                        BasicBlock *BB);
411   Optional<ValueLatticeElement> solveBlockValueSelect(SelectInst *S,
412                                                       BasicBlock *BB);
413   Optional<ConstantRange> getRangeFor(Value *V, Instruction *CxtI,
414                                       BasicBlock *BB);
415   Optional<ValueLatticeElement> solveBlockValueBinaryOpImpl(
416       Instruction *I, BasicBlock *BB,
417       std::function<ConstantRange(const ConstantRange &,
418                                   const ConstantRange &)> OpFn);
419   Optional<ValueLatticeElement> solveBlockValueBinaryOp(BinaryOperator *BBI,
420                                                         BasicBlock *BB);
421   Optional<ValueLatticeElement> solveBlockValueCast(CastInst *CI,
422                                                     BasicBlock *BB);
423   Optional<ValueLatticeElement> solveBlockValueOverflowIntrinsic(
424       WithOverflowInst *WO, BasicBlock *BB);
425   Optional<ValueLatticeElement> solveBlockValueIntrinsic(IntrinsicInst *II,
426                                                          BasicBlock *BB);
427   Optional<ValueLatticeElement> solveBlockValueExtractValue(
428       ExtractValueInst *EVI, BasicBlock *BB);
429   bool isNonNullAtEndOfBlock(Value *Val, BasicBlock *BB);
430   void intersectAssumeOrGuardBlockValueConstantRange(Value *Val,
431                                                      ValueLatticeElement &BBLV,
432                                                      Instruction *BBI);
433 
434   void solve();
435 
436 public:
437   /// This is the query interface to determine the lattice value for the
438   /// specified Value* at the context instruction (if specified) or at the
439   /// start of the block.
440   ValueLatticeElement getValueInBlock(Value *V, BasicBlock *BB,
441                                       Instruction *CxtI = nullptr);
442 
443   /// This is the query interface to determine the lattice value for the
444   /// specified Value* at the specified instruction using only information
445   /// from assumes/guards and range metadata. Unlike getValueInBlock(), no
446   /// recursive query is performed.
447   ValueLatticeElement getValueAt(Value *V, Instruction *CxtI);
448 
449   /// This is the query interface to determine the lattice
450   /// value for the specified Value* that is true on the specified edge.
451   ValueLatticeElement getValueOnEdge(Value *V, BasicBlock *FromBB,
452                                      BasicBlock *ToBB,
453                                      Instruction *CxtI = nullptr);
454 
455   /// Complete flush all previously computed values
clear()456   void clear() {
457     TheCache.clear();
458   }
459 
460   /// Printing the LazyValueInfo Analysis.
printLVI(Function & F,DominatorTree & DTree,raw_ostream & OS)461   void printLVI(Function &F, DominatorTree &DTree, raw_ostream &OS) {
462     LazyValueInfoAnnotatedWriter Writer(this, DTree);
463     F.print(OS, &Writer);
464   }
465 
466   /// This is part of the update interface to inform the cache
467   /// that a block has been deleted.
eraseBlock(BasicBlock * BB)468   void eraseBlock(BasicBlock *BB) {
469     TheCache.eraseBlock(BB);
470   }
471 
472   /// This is the update interface to inform the cache that an edge from
473   /// PredBB to OldSucc has been threaded to be from PredBB to NewSucc.
474   void threadEdge(BasicBlock *PredBB,BasicBlock *OldSucc,BasicBlock *NewSucc);
475 
LazyValueInfoImpl(AssumptionCache * AC,const DataLayout & DL,Function * GuardDecl)476   LazyValueInfoImpl(AssumptionCache *AC, const DataLayout &DL,
477                     Function *GuardDecl)
478       : AC(AC), DL(DL), GuardDecl(GuardDecl) {}
479 };
480 } // end anonymous namespace
481 
482 
solve()483 void LazyValueInfoImpl::solve() {
484   SmallVector<std::pair<BasicBlock *, Value *>, 8> StartingStack(
485       BlockValueStack.begin(), BlockValueStack.end());
486 
487   unsigned processedCount = 0;
488   while (!BlockValueStack.empty()) {
489     processedCount++;
490     // Abort if we have to process too many values to get a result for this one.
491     // Because of the design of the overdefined cache currently being per-block
492     // to avoid naming-related issues (IE it wants to try to give different
493     // results for the same name in different blocks), overdefined results don't
494     // get cached globally, which in turn means we will often try to rediscover
495     // the same overdefined result again and again.  Once something like
496     // PredicateInfo is used in LVI or CVP, we should be able to make the
497     // overdefined cache global, and remove this throttle.
498     if (processedCount > MaxProcessedPerValue) {
499       LLVM_DEBUG(
500           dbgs() << "Giving up on stack because we are getting too deep\n");
501       // Fill in the original values
502       while (!StartingStack.empty()) {
503         std::pair<BasicBlock *, Value *> &e = StartingStack.back();
504         TheCache.insertResult(e.second, e.first,
505                               ValueLatticeElement::getOverdefined());
506         StartingStack.pop_back();
507       }
508       BlockValueSet.clear();
509       BlockValueStack.clear();
510       return;
511     }
512     std::pair<BasicBlock *, Value *> e = BlockValueStack.back();
513     assert(BlockValueSet.count(e) && "Stack value should be in BlockValueSet!");
514 
515     if (solveBlockValue(e.second, e.first)) {
516       // The work item was completely processed.
517       assert(BlockValueStack.back() == e && "Nothing should have been pushed!");
518 #ifndef NDEBUG
519       Optional<ValueLatticeElement> BBLV =
520           TheCache.getCachedValueInfo(e.second, e.first);
521       assert(BBLV && "Result should be in cache!");
522       LLVM_DEBUG(
523           dbgs() << "POP " << *e.second << " in " << e.first->getName() << " = "
524                  << *BBLV << "\n");
525 #endif
526 
527       BlockValueStack.pop_back();
528       BlockValueSet.erase(e);
529     } else {
530       // More work needs to be done before revisiting.
531       assert(BlockValueStack.back() != e && "Stack should have been pushed!");
532     }
533   }
534 }
535 
getBlockValue(Value * Val,BasicBlock * BB)536 Optional<ValueLatticeElement> LazyValueInfoImpl::getBlockValue(Value *Val,
537                                                                BasicBlock *BB) {
538   // If already a constant, there is nothing to compute.
539   if (Constant *VC = dyn_cast<Constant>(Val))
540     return ValueLatticeElement::get(VC);
541 
542   if (Optional<ValueLatticeElement> OptLatticeVal =
543           TheCache.getCachedValueInfo(Val, BB))
544     return OptLatticeVal;
545 
546   // We have hit a cycle, assume overdefined.
547   if (!pushBlockValue({ BB, Val }))
548     return ValueLatticeElement::getOverdefined();
549 
550   // Yet to be resolved.
551   return None;
552 }
553 
getFromRangeMetadata(Instruction * BBI)554 static ValueLatticeElement getFromRangeMetadata(Instruction *BBI) {
555   switch (BBI->getOpcode()) {
556   default: break;
557   case Instruction::Load:
558   case Instruction::Call:
559   case Instruction::Invoke:
560     if (MDNode *Ranges = BBI->getMetadata(LLVMContext::MD_range))
561       if (isa<IntegerType>(BBI->getType())) {
562         return ValueLatticeElement::getRange(
563             getConstantRangeFromMetadata(*Ranges));
564       }
565     break;
566   };
567   // Nothing known - will be intersected with other facts
568   return ValueLatticeElement::getOverdefined();
569 }
570 
solveBlockValue(Value * Val,BasicBlock * BB)571 bool LazyValueInfoImpl::solveBlockValue(Value *Val, BasicBlock *BB) {
572   assert(!isa<Constant>(Val) && "Value should not be constant");
573   assert(!TheCache.getCachedValueInfo(Val, BB) &&
574          "Value should not be in cache");
575 
576   // Hold off inserting this value into the Cache in case we have to return
577   // false and come back later.
578   Optional<ValueLatticeElement> Res = solveBlockValueImpl(Val, BB);
579   if (!Res)
580     // Work pushed, will revisit
581     return false;
582 
583   TheCache.insertResult(Val, BB, *Res);
584   return true;
585 }
586 
solveBlockValueImpl(Value * Val,BasicBlock * BB)587 Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueImpl(
588     Value *Val, BasicBlock *BB) {
589   Instruction *BBI = dyn_cast<Instruction>(Val);
590   if (!BBI || BBI->getParent() != BB)
591     return solveBlockValueNonLocal(Val, BB);
592 
593   if (PHINode *PN = dyn_cast<PHINode>(BBI))
594     return solveBlockValuePHINode(PN, BB);
595 
596   if (auto *SI = dyn_cast<SelectInst>(BBI))
597     return solveBlockValueSelect(SI, BB);
598 
599   // If this value is a nonnull pointer, record it's range and bailout.  Note
600   // that for all other pointer typed values, we terminate the search at the
601   // definition.  We could easily extend this to look through geps, bitcasts,
602   // and the like to prove non-nullness, but it's not clear that's worth it
603   // compile time wise.  The context-insensitive value walk done inside
604   // isKnownNonZero gets most of the profitable cases at much less expense.
605   // This does mean that we have a sensitivity to where the defining
606   // instruction is placed, even if it could legally be hoisted much higher.
607   // That is unfortunate.
608   PointerType *PT = dyn_cast<PointerType>(BBI->getType());
609   if (PT && isKnownNonZero(BBI, DL))
610     return ValueLatticeElement::getNot(ConstantPointerNull::get(PT));
611 
612   if (BBI->getType()->isIntegerTy()) {
613     if (auto *CI = dyn_cast<CastInst>(BBI))
614       return solveBlockValueCast(CI, BB);
615 
616     if (BinaryOperator *BO = dyn_cast<BinaryOperator>(BBI))
617       return solveBlockValueBinaryOp(BO, BB);
618 
619     if (auto *EVI = dyn_cast<ExtractValueInst>(BBI))
620       return solveBlockValueExtractValue(EVI, BB);
621 
622     if (auto *II = dyn_cast<IntrinsicInst>(BBI))
623       return solveBlockValueIntrinsic(II, BB);
624   }
625 
626   LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()
627                     << "' - unknown inst def found.\n");
628   return getFromRangeMetadata(BBI);
629 }
630 
AddNonNullPointer(Value * Ptr,NonNullPointerSet & PtrSet)631 static void AddNonNullPointer(Value *Ptr, NonNullPointerSet &PtrSet) {
632   // TODO: Use NullPointerIsDefined instead.
633   if (Ptr->getType()->getPointerAddressSpace() == 0)
634     PtrSet.insert(getUnderlyingObject(Ptr));
635 }
636 
AddNonNullPointersByInstruction(Instruction * I,NonNullPointerSet & PtrSet)637 static void AddNonNullPointersByInstruction(
638     Instruction *I, NonNullPointerSet &PtrSet) {
639   if (LoadInst *L = dyn_cast<LoadInst>(I)) {
640     AddNonNullPointer(L->getPointerOperand(), PtrSet);
641   } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
642     AddNonNullPointer(S->getPointerOperand(), PtrSet);
643   } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) {
644     if (MI->isVolatile()) return;
645 
646     // FIXME: check whether it has a valuerange that excludes zero?
647     ConstantInt *Len = dyn_cast<ConstantInt>(MI->getLength());
648     if (!Len || Len->isZero()) return;
649 
650     AddNonNullPointer(MI->getRawDest(), PtrSet);
651     if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
652       AddNonNullPointer(MTI->getRawSource(), PtrSet);
653   }
654 }
655 
isNonNullAtEndOfBlock(Value * Val,BasicBlock * BB)656 bool LazyValueInfoImpl::isNonNullAtEndOfBlock(Value *Val, BasicBlock *BB) {
657   if (NullPointerIsDefined(BB->getParent(),
658                            Val->getType()->getPointerAddressSpace()))
659     return false;
660 
661   Val = Val->stripInBoundsOffsets();
662   return TheCache.isNonNullAtEndOfBlock(Val, BB, [](BasicBlock *BB) {
663     NonNullPointerSet NonNullPointers;
664     for (Instruction &I : *BB)
665       AddNonNullPointersByInstruction(&I, NonNullPointers);
666     return NonNullPointers;
667   });
668 }
669 
solveBlockValueNonLocal(Value * Val,BasicBlock * BB)670 Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueNonLocal(
671     Value *Val, BasicBlock *BB) {
672   ValueLatticeElement Result;  // Start Undefined.
673 
674   // If this is the entry block, we must be asking about an argument.  The
675   // value is overdefined.
676   if (BB->isEntryBlock()) {
677     assert(isa<Argument>(Val) && "Unknown live-in to the entry block");
678     return ValueLatticeElement::getOverdefined();
679   }
680 
681   // Loop over all of our predecessors, merging what we know from them into
682   // result.  If we encounter an unexplored predecessor, we eagerly explore it
683   // in a depth first manner.  In practice, this has the effect of discovering
684   // paths we can't analyze eagerly without spending compile times analyzing
685   // other paths.  This heuristic benefits from the fact that predecessors are
686   // frequently arranged such that dominating ones come first and we quickly
687   // find a path to function entry.  TODO: We should consider explicitly
688   // canonicalizing to make this true rather than relying on this happy
689   // accident.
690   for (BasicBlock *Pred : predecessors(BB)) {
691     Optional<ValueLatticeElement> EdgeResult = getEdgeValue(Val, Pred, BB);
692     if (!EdgeResult)
693       // Explore that input, then return here
694       return None;
695 
696     Result.mergeIn(*EdgeResult);
697 
698     // If we hit overdefined, exit early.  The BlockVals entry is already set
699     // to overdefined.
700     if (Result.isOverdefined()) {
701       LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()
702                         << "' - overdefined because of pred (non local).\n");
703       return Result;
704     }
705   }
706 
707   // Return the merged value, which is more precise than 'overdefined'.
708   assert(!Result.isOverdefined());
709   return Result;
710 }
711 
solveBlockValuePHINode(PHINode * PN,BasicBlock * BB)712 Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValuePHINode(
713     PHINode *PN, BasicBlock *BB) {
714   ValueLatticeElement Result;  // Start Undefined.
715 
716   // Loop over all of our predecessors, merging what we know from them into
717   // result.  See the comment about the chosen traversal order in
718   // solveBlockValueNonLocal; the same reasoning applies here.
719   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
720     BasicBlock *PhiBB = PN->getIncomingBlock(i);
721     Value *PhiVal = PN->getIncomingValue(i);
722     // Note that we can provide PN as the context value to getEdgeValue, even
723     // though the results will be cached, because PN is the value being used as
724     // the cache key in the caller.
725     Optional<ValueLatticeElement> EdgeResult =
726         getEdgeValue(PhiVal, PhiBB, BB, PN);
727     if (!EdgeResult)
728       // Explore that input, then return here
729       return None;
730 
731     Result.mergeIn(*EdgeResult);
732 
733     // If we hit overdefined, exit early.  The BlockVals entry is already set
734     // to overdefined.
735     if (Result.isOverdefined()) {
736       LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()
737                         << "' - overdefined because of pred (local).\n");
738 
739       return Result;
740     }
741   }
742 
743   // Return the merged value, which is more precise than 'overdefined'.
744   assert(!Result.isOverdefined() && "Possible PHI in entry block?");
745   return Result;
746 }
747 
748 static ValueLatticeElement getValueFromCondition(Value *Val, Value *Cond,
749                                                  bool isTrueDest = true);
750 
751 // If we can determine a constraint on the value given conditions assumed by
752 // the program, intersect those constraints with BBLV
intersectAssumeOrGuardBlockValueConstantRange(Value * Val,ValueLatticeElement & BBLV,Instruction * BBI)753 void LazyValueInfoImpl::intersectAssumeOrGuardBlockValueConstantRange(
754         Value *Val, ValueLatticeElement &BBLV, Instruction *BBI) {
755   BBI = BBI ? BBI : dyn_cast<Instruction>(Val);
756   if (!BBI)
757     return;
758 
759   BasicBlock *BB = BBI->getParent();
760   for (auto &AssumeVH : AC->assumptionsFor(Val)) {
761     if (!AssumeVH)
762       continue;
763 
764     // Only check assumes in the block of the context instruction. Other
765     // assumes will have already been taken into account when the value was
766     // propagated from predecessor blocks.
767     auto *I = cast<CallInst>(AssumeVH);
768     if (I->getParent() != BB || !isValidAssumeForContext(I, BBI))
769       continue;
770 
771     BBLV = intersect(BBLV, getValueFromCondition(Val, I->getArgOperand(0)));
772   }
773 
774   // If guards are not used in the module, don't spend time looking for them
775   if (GuardDecl && !GuardDecl->use_empty() &&
776       BBI->getIterator() != BB->begin()) {
777     for (Instruction &I : make_range(std::next(BBI->getIterator().getReverse()),
778                                      BB->rend())) {
779       Value *Cond = nullptr;
780       if (match(&I, m_Intrinsic<Intrinsic::experimental_guard>(m_Value(Cond))))
781         BBLV = intersect(BBLV, getValueFromCondition(Val, Cond));
782     }
783   }
784 
785   if (BBLV.isOverdefined()) {
786     // Check whether we're checking at the terminator, and the pointer has
787     // been dereferenced in this block.
788     PointerType *PTy = dyn_cast<PointerType>(Val->getType());
789     if (PTy && BB->getTerminator() == BBI &&
790         isNonNullAtEndOfBlock(Val, BB))
791       BBLV = ValueLatticeElement::getNot(ConstantPointerNull::get(PTy));
792   }
793 }
794 
solveBlockValueSelect(SelectInst * SI,BasicBlock * BB)795 Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueSelect(
796     SelectInst *SI, BasicBlock *BB) {
797   // Recurse on our inputs if needed
798   Optional<ValueLatticeElement> OptTrueVal =
799       getBlockValue(SI->getTrueValue(), BB);
800   if (!OptTrueVal)
801     return None;
802   ValueLatticeElement &TrueVal = *OptTrueVal;
803 
804   Optional<ValueLatticeElement> OptFalseVal =
805       getBlockValue(SI->getFalseValue(), BB);
806   if (!OptFalseVal)
807     return None;
808   ValueLatticeElement &FalseVal = *OptFalseVal;
809 
810   if (TrueVal.isConstantRange() && FalseVal.isConstantRange()) {
811     const ConstantRange &TrueCR = TrueVal.getConstantRange();
812     const ConstantRange &FalseCR = FalseVal.getConstantRange();
813     Value *LHS = nullptr;
814     Value *RHS = nullptr;
815     SelectPatternResult SPR = matchSelectPattern(SI, LHS, RHS);
816     // Is this a min specifically of our two inputs?  (Avoid the risk of
817     // ValueTracking getting smarter looking back past our immediate inputs.)
818     if (SelectPatternResult::isMinOrMax(SPR.Flavor) &&
819         LHS == SI->getTrueValue() && RHS == SI->getFalseValue()) {
820       ConstantRange ResultCR = [&]() {
821         switch (SPR.Flavor) {
822         default:
823           llvm_unreachable("unexpected minmax type!");
824         case SPF_SMIN:                   /// Signed minimum
825           return TrueCR.smin(FalseCR);
826         case SPF_UMIN:                   /// Unsigned minimum
827           return TrueCR.umin(FalseCR);
828         case SPF_SMAX:                   /// Signed maximum
829           return TrueCR.smax(FalseCR);
830         case SPF_UMAX:                   /// Unsigned maximum
831           return TrueCR.umax(FalseCR);
832         };
833       }();
834       return ValueLatticeElement::getRange(
835           ResultCR, TrueVal.isConstantRangeIncludingUndef() |
836                         FalseVal.isConstantRangeIncludingUndef());
837     }
838 
839     if (SPR.Flavor == SPF_ABS) {
840       if (LHS == SI->getTrueValue())
841         return ValueLatticeElement::getRange(
842             TrueCR.abs(), TrueVal.isConstantRangeIncludingUndef());
843       if (LHS == SI->getFalseValue())
844         return ValueLatticeElement::getRange(
845             FalseCR.abs(), FalseVal.isConstantRangeIncludingUndef());
846     }
847 
848     if (SPR.Flavor == SPF_NABS) {
849       ConstantRange Zero(APInt::getNullValue(TrueCR.getBitWidth()));
850       if (LHS == SI->getTrueValue())
851         return ValueLatticeElement::getRange(
852             Zero.sub(TrueCR.abs()), FalseVal.isConstantRangeIncludingUndef());
853       if (LHS == SI->getFalseValue())
854         return ValueLatticeElement::getRange(
855             Zero.sub(FalseCR.abs()), FalseVal.isConstantRangeIncludingUndef());
856     }
857   }
858 
859   // Can we constrain the facts about the true and false values by using the
860   // condition itself?  This shows up with idioms like e.g. select(a > 5, a, 5).
861   // TODO: We could potentially refine an overdefined true value above.
862   Value *Cond = SI->getCondition();
863   TrueVal = intersect(TrueVal,
864                       getValueFromCondition(SI->getTrueValue(), Cond, true));
865   FalseVal = intersect(FalseVal,
866                        getValueFromCondition(SI->getFalseValue(), Cond, false));
867 
868   ValueLatticeElement Result = TrueVal;
869   Result.mergeIn(FalseVal);
870   return Result;
871 }
872 
getRangeFor(Value * V,Instruction * CxtI,BasicBlock * BB)873 Optional<ConstantRange> LazyValueInfoImpl::getRangeFor(Value *V,
874                                                        Instruction *CxtI,
875                                                        BasicBlock *BB) {
876   Optional<ValueLatticeElement> OptVal = getBlockValue(V, BB);
877   if (!OptVal)
878     return None;
879 
880   ValueLatticeElement &Val = *OptVal;
881   intersectAssumeOrGuardBlockValueConstantRange(V, Val, CxtI);
882   if (Val.isConstantRange())
883     return Val.getConstantRange();
884 
885   const unsigned OperandBitWidth = DL.getTypeSizeInBits(V->getType());
886   return ConstantRange::getFull(OperandBitWidth);
887 }
888 
solveBlockValueCast(CastInst * CI,BasicBlock * BB)889 Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueCast(
890     CastInst *CI, BasicBlock *BB) {
891   // Without knowing how wide the input is, we can't analyze it in any useful
892   // way.
893   if (!CI->getOperand(0)->getType()->isSized())
894     return ValueLatticeElement::getOverdefined();
895 
896   // Filter out casts we don't know how to reason about before attempting to
897   // recurse on our operand.  This can cut a long search short if we know we're
898   // not going to be able to get any useful information anways.
899   switch (CI->getOpcode()) {
900   case Instruction::Trunc:
901   case Instruction::SExt:
902   case Instruction::ZExt:
903   case Instruction::BitCast:
904     break;
905   default:
906     // Unhandled instructions are overdefined.
907     LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()
908                       << "' - overdefined (unknown cast).\n");
909     return ValueLatticeElement::getOverdefined();
910   }
911 
912   // Figure out the range of the LHS.  If that fails, we still apply the
913   // transfer rule on the full set since we may be able to locally infer
914   // interesting facts.
915   Optional<ConstantRange> LHSRes = getRangeFor(CI->getOperand(0), CI, BB);
916   if (!LHSRes.hasValue())
917     // More work to do before applying this transfer rule.
918     return None;
919   const ConstantRange &LHSRange = LHSRes.getValue();
920 
921   const unsigned ResultBitWidth = CI->getType()->getIntegerBitWidth();
922 
923   // NOTE: We're currently limited by the set of operations that ConstantRange
924   // can evaluate symbolically.  Enhancing that set will allows us to analyze
925   // more definitions.
926   return ValueLatticeElement::getRange(LHSRange.castOp(CI->getOpcode(),
927                                                        ResultBitWidth));
928 }
929 
solveBlockValueBinaryOpImpl(Instruction * I,BasicBlock * BB,std::function<ConstantRange (const ConstantRange &,const ConstantRange &)> OpFn)930 Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueBinaryOpImpl(
931     Instruction *I, BasicBlock *BB,
932     std::function<ConstantRange(const ConstantRange &,
933                                 const ConstantRange &)> OpFn) {
934   // Figure out the ranges of the operands.  If that fails, use a
935   // conservative range, but apply the transfer rule anyways.  This
936   // lets us pick up facts from expressions like "and i32 (call i32
937   // @foo()), 32"
938   Optional<ConstantRange> LHSRes = getRangeFor(I->getOperand(0), I, BB);
939   Optional<ConstantRange> RHSRes = getRangeFor(I->getOperand(1), I, BB);
940   if (!LHSRes.hasValue() || !RHSRes.hasValue())
941     // More work to do before applying this transfer rule.
942     return None;
943 
944   const ConstantRange &LHSRange = LHSRes.getValue();
945   const ConstantRange &RHSRange = RHSRes.getValue();
946   return ValueLatticeElement::getRange(OpFn(LHSRange, RHSRange));
947 }
948 
solveBlockValueBinaryOp(BinaryOperator * BO,BasicBlock * BB)949 Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueBinaryOp(
950     BinaryOperator *BO, BasicBlock *BB) {
951   assert(BO->getOperand(0)->getType()->isSized() &&
952          "all operands to binary operators are sized");
953   if (BO->getOpcode() == Instruction::Xor) {
954     // Xor is the only operation not supported by ConstantRange::binaryOp().
955     LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()
956                       << "' - overdefined (unknown binary operator).\n");
957     return ValueLatticeElement::getOverdefined();
958   }
959 
960   if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(BO)) {
961     unsigned NoWrapKind = 0;
962     if (OBO->hasNoUnsignedWrap())
963       NoWrapKind |= OverflowingBinaryOperator::NoUnsignedWrap;
964     if (OBO->hasNoSignedWrap())
965       NoWrapKind |= OverflowingBinaryOperator::NoSignedWrap;
966 
967     return solveBlockValueBinaryOpImpl(
968         BO, BB,
969         [BO, NoWrapKind](const ConstantRange &CR1, const ConstantRange &CR2) {
970           return CR1.overflowingBinaryOp(BO->getOpcode(), CR2, NoWrapKind);
971         });
972   }
973 
974   return solveBlockValueBinaryOpImpl(
975       BO, BB, [BO](const ConstantRange &CR1, const ConstantRange &CR2) {
976         return CR1.binaryOp(BO->getOpcode(), CR2);
977       });
978 }
979 
980 Optional<ValueLatticeElement>
solveBlockValueOverflowIntrinsic(WithOverflowInst * WO,BasicBlock * BB)981 LazyValueInfoImpl::solveBlockValueOverflowIntrinsic(WithOverflowInst *WO,
982                                                     BasicBlock *BB) {
983   return solveBlockValueBinaryOpImpl(
984       WO, BB, [WO](const ConstantRange &CR1, const ConstantRange &CR2) {
985         return CR1.binaryOp(WO->getBinaryOp(), CR2);
986       });
987 }
988 
solveBlockValueIntrinsic(IntrinsicInst * II,BasicBlock * BB)989 Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueIntrinsic(
990     IntrinsicInst *II, BasicBlock *BB) {
991   if (!ConstantRange::isIntrinsicSupported(II->getIntrinsicID())) {
992     LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()
993                       << "' - unknown intrinsic.\n");
994     return getFromRangeMetadata(II);
995   }
996 
997   SmallVector<ConstantRange, 2> OpRanges;
998   for (Value *Op : II->args()) {
999     Optional<ConstantRange> Range = getRangeFor(Op, II, BB);
1000     if (!Range)
1001       return None;
1002     OpRanges.push_back(*Range);
1003   }
1004 
1005   return ValueLatticeElement::getRange(
1006       ConstantRange::intrinsic(II->getIntrinsicID(), OpRanges));
1007 }
1008 
solveBlockValueExtractValue(ExtractValueInst * EVI,BasicBlock * BB)1009 Optional<ValueLatticeElement> LazyValueInfoImpl::solveBlockValueExtractValue(
1010     ExtractValueInst *EVI, BasicBlock *BB) {
1011   if (auto *WO = dyn_cast<WithOverflowInst>(EVI->getAggregateOperand()))
1012     if (EVI->getNumIndices() == 1 && *EVI->idx_begin() == 0)
1013       return solveBlockValueOverflowIntrinsic(WO, BB);
1014 
1015   // Handle extractvalue of insertvalue to allow further simplification
1016   // based on replaced with.overflow intrinsics.
1017   if (Value *V = SimplifyExtractValueInst(
1018           EVI->getAggregateOperand(), EVI->getIndices(),
1019           EVI->getModule()->getDataLayout()))
1020     return getBlockValue(V, BB);
1021 
1022   LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()
1023                     << "' - overdefined (unknown extractvalue).\n");
1024   return ValueLatticeElement::getOverdefined();
1025 }
1026 
matchICmpOperand(APInt & Offset,Value * LHS,Value * Val,ICmpInst::Predicate Pred)1027 static bool matchICmpOperand(APInt &Offset, Value *LHS, Value *Val,
1028                              ICmpInst::Predicate Pred) {
1029   if (LHS == Val)
1030     return true;
1031 
1032   // Handle range checking idiom produced by InstCombine. We will subtract the
1033   // offset from the allowed range for RHS in this case.
1034   const APInt *C;
1035   if (match(LHS, m_Add(m_Specific(Val), m_APInt(C)))) {
1036     Offset = *C;
1037     return true;
1038   }
1039 
1040   // Handle the symmetric case. This appears in saturation patterns like
1041   // (x == 16) ? 16 : (x + 1).
1042   if (match(Val, m_Add(m_Specific(LHS), m_APInt(C)))) {
1043     Offset = -*C;
1044     return true;
1045   }
1046 
1047   // If (x | y) < C, then (x < C) && (y < C).
1048   if (match(LHS, m_c_Or(m_Specific(Val), m_Value())) &&
1049       (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE))
1050     return true;
1051 
1052   // If (x & y) > C, then (x > C) && (y > C).
1053   if (match(LHS, m_c_And(m_Specific(Val), m_Value())) &&
1054       (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE))
1055     return true;
1056 
1057   return false;
1058 }
1059 
1060 /// Get value range for a "(Val + Offset) Pred RHS" condition.
getValueFromSimpleICmpCondition(CmpInst::Predicate Pred,Value * RHS,const APInt & Offset)1061 static ValueLatticeElement getValueFromSimpleICmpCondition(
1062     CmpInst::Predicate Pred, Value *RHS, const APInt &Offset) {
1063   ConstantRange RHSRange(RHS->getType()->getIntegerBitWidth(),
1064                          /*isFullSet=*/true);
1065   if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS))
1066     RHSRange = ConstantRange(CI->getValue());
1067   else if (Instruction *I = dyn_cast<Instruction>(RHS))
1068     if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
1069       RHSRange = getConstantRangeFromMetadata(*Ranges);
1070 
1071   ConstantRange TrueValues =
1072       ConstantRange::makeAllowedICmpRegion(Pred, RHSRange);
1073   return ValueLatticeElement::getRange(TrueValues.subtract(Offset));
1074 }
1075 
getValueFromICmpCondition(Value * Val,ICmpInst * ICI,bool isTrueDest)1076 static ValueLatticeElement getValueFromICmpCondition(Value *Val, ICmpInst *ICI,
1077                                                      bool isTrueDest) {
1078   Value *LHS = ICI->getOperand(0);
1079   Value *RHS = ICI->getOperand(1);
1080 
1081   // Get the predicate that must hold along the considered edge.
1082   CmpInst::Predicate EdgePred =
1083       isTrueDest ? ICI->getPredicate() : ICI->getInversePredicate();
1084 
1085   if (isa<Constant>(RHS)) {
1086     if (ICI->isEquality() && LHS == Val) {
1087       if (EdgePred == ICmpInst::ICMP_EQ)
1088         return ValueLatticeElement::get(cast<Constant>(RHS));
1089       else if (!isa<UndefValue>(RHS))
1090         return ValueLatticeElement::getNot(cast<Constant>(RHS));
1091     }
1092   }
1093 
1094   Type *Ty = Val->getType();
1095   if (!Ty->isIntegerTy())
1096     return ValueLatticeElement::getOverdefined();
1097 
1098   APInt Offset(Ty->getScalarSizeInBits(), 0);
1099   if (matchICmpOperand(Offset, LHS, Val, EdgePred))
1100     return getValueFromSimpleICmpCondition(EdgePred, RHS, Offset);
1101 
1102   CmpInst::Predicate SwappedPred = CmpInst::getSwappedPredicate(EdgePred);
1103   if (matchICmpOperand(Offset, RHS, Val, SwappedPred))
1104     return getValueFromSimpleICmpCondition(SwappedPred, LHS, Offset);
1105 
1106   const APInt *Mask, *C;
1107   if (match(LHS, m_And(m_Specific(Val), m_APInt(Mask))) &&
1108       match(RHS, m_APInt(C))) {
1109     // If (Val & Mask) == C then all the masked bits are known and we can
1110     // compute a value range based on that.
1111     if (EdgePred == ICmpInst::ICMP_EQ) {
1112       KnownBits Known;
1113       Known.Zero = ~*C & *Mask;
1114       Known.One = *C & *Mask;
1115       return ValueLatticeElement::getRange(
1116           ConstantRange::fromKnownBits(Known, /*IsSigned*/ false));
1117     }
1118     // If (Val & Mask) != 0 then the value must be larger than the lowest set
1119     // bit of Mask.
1120     if (EdgePred == ICmpInst::ICMP_NE && !Mask->isNullValue() &&
1121         C->isNullValue()) {
1122       unsigned BitWidth = Ty->getIntegerBitWidth();
1123       return ValueLatticeElement::getRange(ConstantRange::getNonEmpty(
1124           APInt::getOneBitSet(BitWidth, Mask->countTrailingZeros()),
1125           APInt::getNullValue(BitWidth)));
1126     }
1127   }
1128 
1129   return ValueLatticeElement::getOverdefined();
1130 }
1131 
1132 // Handle conditions of the form
1133 // extractvalue(op.with.overflow(%x, C), 1).
getValueFromOverflowCondition(Value * Val,WithOverflowInst * WO,bool IsTrueDest)1134 static ValueLatticeElement getValueFromOverflowCondition(
1135     Value *Val, WithOverflowInst *WO, bool IsTrueDest) {
1136   // TODO: This only works with a constant RHS for now. We could also compute
1137   // the range of the RHS, but this doesn't fit into the current structure of
1138   // the edge value calculation.
1139   const APInt *C;
1140   if (WO->getLHS() != Val || !match(WO->getRHS(), m_APInt(C)))
1141     return ValueLatticeElement::getOverdefined();
1142 
1143   // Calculate the possible values of %x for which no overflow occurs.
1144   ConstantRange NWR = ConstantRange::makeExactNoWrapRegion(
1145       WO->getBinaryOp(), *C, WO->getNoWrapKind());
1146 
1147   // If overflow is false, %x is constrained to NWR. If overflow is true, %x is
1148   // constrained to it's inverse (all values that might cause overflow).
1149   if (IsTrueDest)
1150     NWR = NWR.inverse();
1151   return ValueLatticeElement::getRange(NWR);
1152 }
1153 
1154 static Optional<ValueLatticeElement>
getValueFromConditionImpl(Value * Val,Value * Cond,bool isTrueDest,bool isRevisit,SmallDenseMap<Value *,ValueLatticeElement> & Visited,SmallVectorImpl<Value * > & Worklist)1155 getValueFromConditionImpl(Value *Val, Value *Cond, bool isTrueDest,
1156                           bool isRevisit,
1157                           SmallDenseMap<Value *, ValueLatticeElement> &Visited,
1158                           SmallVectorImpl<Value *> &Worklist) {
1159   if (!isRevisit) {
1160     if (ICmpInst *ICI = dyn_cast<ICmpInst>(Cond))
1161       return getValueFromICmpCondition(Val, ICI, isTrueDest);
1162 
1163     if (auto *EVI = dyn_cast<ExtractValueInst>(Cond))
1164       if (auto *WO = dyn_cast<WithOverflowInst>(EVI->getAggregateOperand()))
1165         if (EVI->getNumIndices() == 1 && *EVI->idx_begin() == 1)
1166           return getValueFromOverflowCondition(Val, WO, isTrueDest);
1167   }
1168 
1169   Value *L, *R;
1170   bool IsAnd;
1171   if (match(Cond, m_LogicalAnd(m_Value(L), m_Value(R))))
1172     IsAnd = true;
1173   else if (match(Cond, m_LogicalOr(m_Value(L), m_Value(R))))
1174     IsAnd = false;
1175   else
1176     return ValueLatticeElement::getOverdefined();
1177 
1178   auto LV = Visited.find(L);
1179   auto RV = Visited.find(R);
1180 
1181   // if (L && R) -> intersect L and R
1182   // if (!(L || R)) -> intersect L and R
1183   // if (L || R) -> union L and R
1184   // if (!(L && R)) -> union L and R
1185   if ((isTrueDest ^ IsAnd) && (LV != Visited.end())) {
1186     ValueLatticeElement V = LV->second;
1187     if (V.isOverdefined())
1188       return V;
1189     if (RV != Visited.end()) {
1190       V.mergeIn(RV->second);
1191       return V;
1192     }
1193   }
1194 
1195   if (LV == Visited.end() || RV == Visited.end()) {
1196     assert(!isRevisit);
1197     if (LV == Visited.end())
1198       Worklist.push_back(L);
1199     if (RV == Visited.end())
1200       Worklist.push_back(R);
1201     return None;
1202   }
1203 
1204   return intersect(LV->second, RV->second);
1205 }
1206 
getValueFromCondition(Value * Val,Value * Cond,bool isTrueDest)1207 ValueLatticeElement getValueFromCondition(Value *Val, Value *Cond,
1208                                           bool isTrueDest) {
1209   assert(Cond && "precondition");
1210   SmallDenseMap<Value*, ValueLatticeElement> Visited;
1211   SmallVector<Value *> Worklist;
1212 
1213   Worklist.push_back(Cond);
1214   do {
1215     Value *CurrentCond = Worklist.back();
1216     // Insert an Overdefined placeholder into the set to prevent
1217     // infinite recursion if there exists IRs that use not
1218     // dominated by its def as in this example:
1219     //   "%tmp3 = or i1 undef, %tmp4"
1220     //   "%tmp4 = or i1 undef, %tmp3"
1221     auto Iter =
1222         Visited.try_emplace(CurrentCond, ValueLatticeElement::getOverdefined());
1223     bool isRevisit = !Iter.second;
1224     Optional<ValueLatticeElement> Result = getValueFromConditionImpl(
1225         Val, CurrentCond, isTrueDest, isRevisit, Visited, Worklist);
1226     if (Result) {
1227       Visited[CurrentCond] = *Result;
1228       Worklist.pop_back();
1229     }
1230   } while (!Worklist.empty());
1231 
1232   auto Result = Visited.find(Cond);
1233   assert(Result != Visited.end());
1234   return Result->second;
1235 }
1236 
1237 // Return true if Usr has Op as an operand, otherwise false.
usesOperand(User * Usr,Value * Op)1238 static bool usesOperand(User *Usr, Value *Op) {
1239   return is_contained(Usr->operands(), Op);
1240 }
1241 
1242 // Return true if the instruction type of Val is supported by
1243 // constantFoldUser(). Currently CastInst, BinaryOperator and FreezeInst only.
1244 // Call this before calling constantFoldUser() to find out if it's even worth
1245 // attempting to call it.
isOperationFoldable(User * Usr)1246 static bool isOperationFoldable(User *Usr) {
1247   return isa<CastInst>(Usr) || isa<BinaryOperator>(Usr) || isa<FreezeInst>(Usr);
1248 }
1249 
1250 // Check if Usr can be simplified to an integer constant when the value of one
1251 // of its operands Op is an integer constant OpConstVal. If so, return it as an
1252 // lattice value range with a single element or otherwise return an overdefined
1253 // lattice value.
constantFoldUser(User * Usr,Value * Op,const APInt & OpConstVal,const DataLayout & DL)1254 static ValueLatticeElement constantFoldUser(User *Usr, Value *Op,
1255                                             const APInt &OpConstVal,
1256                                             const DataLayout &DL) {
1257   assert(isOperationFoldable(Usr) && "Precondition");
1258   Constant* OpConst = Constant::getIntegerValue(Op->getType(), OpConstVal);
1259   // Check if Usr can be simplified to a constant.
1260   if (auto *CI = dyn_cast<CastInst>(Usr)) {
1261     assert(CI->getOperand(0) == Op && "Operand 0 isn't Op");
1262     if (auto *C = dyn_cast_or_null<ConstantInt>(
1263             SimplifyCastInst(CI->getOpcode(), OpConst,
1264                              CI->getDestTy(), DL))) {
1265       return ValueLatticeElement::getRange(ConstantRange(C->getValue()));
1266     }
1267   } else if (auto *BO = dyn_cast<BinaryOperator>(Usr)) {
1268     bool Op0Match = BO->getOperand(0) == Op;
1269     bool Op1Match = BO->getOperand(1) == Op;
1270     assert((Op0Match || Op1Match) &&
1271            "Operand 0 nor Operand 1 isn't a match");
1272     Value *LHS = Op0Match ? OpConst : BO->getOperand(0);
1273     Value *RHS = Op1Match ? OpConst : BO->getOperand(1);
1274     if (auto *C = dyn_cast_or_null<ConstantInt>(
1275             SimplifyBinOp(BO->getOpcode(), LHS, RHS, DL))) {
1276       return ValueLatticeElement::getRange(ConstantRange(C->getValue()));
1277     }
1278   } else if (isa<FreezeInst>(Usr)) {
1279     assert(cast<FreezeInst>(Usr)->getOperand(0) == Op && "Operand 0 isn't Op");
1280     return ValueLatticeElement::getRange(ConstantRange(OpConstVal));
1281   }
1282   return ValueLatticeElement::getOverdefined();
1283 }
1284 
1285 /// Compute the value of Val on the edge BBFrom -> BBTo. Returns false if
1286 /// Val is not constrained on the edge.  Result is unspecified if return value
1287 /// is false.
getEdgeValueLocal(Value * Val,BasicBlock * BBFrom,BasicBlock * BBTo)1288 static Optional<ValueLatticeElement> getEdgeValueLocal(Value *Val,
1289                                                        BasicBlock *BBFrom,
1290                                                        BasicBlock *BBTo) {
1291   // TODO: Handle more complex conditionals. If (v == 0 || v2 < 1) is false, we
1292   // know that v != 0.
1293   if (BranchInst *BI = dyn_cast<BranchInst>(BBFrom->getTerminator())) {
1294     // If this is a conditional branch and only one successor goes to BBTo, then
1295     // we may be able to infer something from the condition.
1296     if (BI->isConditional() &&
1297         BI->getSuccessor(0) != BI->getSuccessor(1)) {
1298       bool isTrueDest = BI->getSuccessor(0) == BBTo;
1299       assert(BI->getSuccessor(!isTrueDest) == BBTo &&
1300              "BBTo isn't a successor of BBFrom");
1301       Value *Condition = BI->getCondition();
1302 
1303       // If V is the condition of the branch itself, then we know exactly what
1304       // it is.
1305       if (Condition == Val)
1306         return ValueLatticeElement::get(ConstantInt::get(
1307                               Type::getInt1Ty(Val->getContext()), isTrueDest));
1308 
1309       // If the condition of the branch is an equality comparison, we may be
1310       // able to infer the value.
1311       ValueLatticeElement Result = getValueFromCondition(Val, Condition,
1312                                                          isTrueDest);
1313       if (!Result.isOverdefined())
1314         return Result;
1315 
1316       if (User *Usr = dyn_cast<User>(Val)) {
1317         assert(Result.isOverdefined() && "Result isn't overdefined");
1318         // Check with isOperationFoldable() first to avoid linearly iterating
1319         // over the operands unnecessarily which can be expensive for
1320         // instructions with many operands.
1321         if (isa<IntegerType>(Usr->getType()) && isOperationFoldable(Usr)) {
1322           const DataLayout &DL = BBTo->getModule()->getDataLayout();
1323           if (usesOperand(Usr, Condition)) {
1324             // If Val has Condition as an operand and Val can be folded into a
1325             // constant with either Condition == true or Condition == false,
1326             // propagate the constant.
1327             // eg.
1328             //   ; %Val is true on the edge to %then.
1329             //   %Val = and i1 %Condition, true.
1330             //   br %Condition, label %then, label %else
1331             APInt ConditionVal(1, isTrueDest ? 1 : 0);
1332             Result = constantFoldUser(Usr, Condition, ConditionVal, DL);
1333           } else {
1334             // If one of Val's operand has an inferred value, we may be able to
1335             // infer the value of Val.
1336             // eg.
1337             //    ; %Val is 94 on the edge to %then.
1338             //    %Val = add i8 %Op, 1
1339             //    %Condition = icmp eq i8 %Op, 93
1340             //    br i1 %Condition, label %then, label %else
1341             for (unsigned i = 0; i < Usr->getNumOperands(); ++i) {
1342               Value *Op = Usr->getOperand(i);
1343               ValueLatticeElement OpLatticeVal =
1344                   getValueFromCondition(Op, Condition, isTrueDest);
1345               if (Optional<APInt> OpConst = OpLatticeVal.asConstantInteger()) {
1346                 Result = constantFoldUser(Usr, Op, OpConst.getValue(), DL);
1347                 break;
1348               }
1349             }
1350           }
1351         }
1352       }
1353       if (!Result.isOverdefined())
1354         return Result;
1355     }
1356   }
1357 
1358   // If the edge was formed by a switch on the value, then we may know exactly
1359   // what it is.
1360   if (SwitchInst *SI = dyn_cast<SwitchInst>(BBFrom->getTerminator())) {
1361     Value *Condition = SI->getCondition();
1362     if (!isa<IntegerType>(Val->getType()))
1363       return None;
1364     bool ValUsesConditionAndMayBeFoldable = false;
1365     if (Condition != Val) {
1366       // Check if Val has Condition as an operand.
1367       if (User *Usr = dyn_cast<User>(Val))
1368         ValUsesConditionAndMayBeFoldable = isOperationFoldable(Usr) &&
1369             usesOperand(Usr, Condition);
1370       if (!ValUsesConditionAndMayBeFoldable)
1371         return None;
1372     }
1373     assert((Condition == Val || ValUsesConditionAndMayBeFoldable) &&
1374            "Condition != Val nor Val doesn't use Condition");
1375 
1376     bool DefaultCase = SI->getDefaultDest() == BBTo;
1377     unsigned BitWidth = Val->getType()->getIntegerBitWidth();
1378     ConstantRange EdgesVals(BitWidth, DefaultCase/*isFullSet*/);
1379 
1380     for (auto Case : SI->cases()) {
1381       APInt CaseValue = Case.getCaseValue()->getValue();
1382       ConstantRange EdgeVal(CaseValue);
1383       if (ValUsesConditionAndMayBeFoldable) {
1384         User *Usr = cast<User>(Val);
1385         const DataLayout &DL = BBTo->getModule()->getDataLayout();
1386         ValueLatticeElement EdgeLatticeVal =
1387             constantFoldUser(Usr, Condition, CaseValue, DL);
1388         if (EdgeLatticeVal.isOverdefined())
1389           return None;
1390         EdgeVal = EdgeLatticeVal.getConstantRange();
1391       }
1392       if (DefaultCase) {
1393         // It is possible that the default destination is the destination of
1394         // some cases. We cannot perform difference for those cases.
1395         // We know Condition != CaseValue in BBTo.  In some cases we can use
1396         // this to infer Val == f(Condition) is != f(CaseValue).  For now, we
1397         // only do this when f is identity (i.e. Val == Condition), but we
1398         // should be able to do this for any injective f.
1399         if (Case.getCaseSuccessor() != BBTo && Condition == Val)
1400           EdgesVals = EdgesVals.difference(EdgeVal);
1401       } else if (Case.getCaseSuccessor() == BBTo)
1402         EdgesVals = EdgesVals.unionWith(EdgeVal);
1403     }
1404     return ValueLatticeElement::getRange(std::move(EdgesVals));
1405   }
1406   return None;
1407 }
1408 
1409 /// Compute the value of Val on the edge BBFrom -> BBTo or the value at
1410 /// the basic block if the edge does not constrain Val.
getEdgeValue(Value * Val,BasicBlock * BBFrom,BasicBlock * BBTo,Instruction * CxtI)1411 Optional<ValueLatticeElement> LazyValueInfoImpl::getEdgeValue(
1412     Value *Val, BasicBlock *BBFrom, BasicBlock *BBTo, Instruction *CxtI) {
1413   // If already a constant, there is nothing to compute.
1414   if (Constant *VC = dyn_cast<Constant>(Val))
1415     return ValueLatticeElement::get(VC);
1416 
1417   ValueLatticeElement LocalResult = getEdgeValueLocal(Val, BBFrom, BBTo)
1418       .getValueOr(ValueLatticeElement::getOverdefined());
1419   if (hasSingleValue(LocalResult))
1420     // Can't get any more precise here
1421     return LocalResult;
1422 
1423   Optional<ValueLatticeElement> OptInBlock = getBlockValue(Val, BBFrom);
1424   if (!OptInBlock)
1425     return None;
1426   ValueLatticeElement &InBlock = *OptInBlock;
1427 
1428   // Try to intersect ranges of the BB and the constraint on the edge.
1429   intersectAssumeOrGuardBlockValueConstantRange(Val, InBlock,
1430                                                 BBFrom->getTerminator());
1431   // We can use the context instruction (generically the ultimate instruction
1432   // the calling pass is trying to simplify) here, even though the result of
1433   // this function is generally cached when called from the solve* functions
1434   // (and that cached result might be used with queries using a different
1435   // context instruction), because when this function is called from the solve*
1436   // functions, the context instruction is not provided. When called from
1437   // LazyValueInfoImpl::getValueOnEdge, the context instruction is provided,
1438   // but then the result is not cached.
1439   intersectAssumeOrGuardBlockValueConstantRange(Val, InBlock, CxtI);
1440 
1441   return intersect(LocalResult, InBlock);
1442 }
1443 
getValueInBlock(Value * V,BasicBlock * BB,Instruction * CxtI)1444 ValueLatticeElement LazyValueInfoImpl::getValueInBlock(Value *V, BasicBlock *BB,
1445                                                        Instruction *CxtI) {
1446   LLVM_DEBUG(dbgs() << "LVI Getting block end value " << *V << " at '"
1447                     << BB->getName() << "'\n");
1448 
1449   assert(BlockValueStack.empty() && BlockValueSet.empty());
1450   Optional<ValueLatticeElement> OptResult = getBlockValue(V, BB);
1451   if (!OptResult) {
1452     solve();
1453     OptResult = getBlockValue(V, BB);
1454     assert(OptResult && "Value not available after solving");
1455   }
1456   ValueLatticeElement Result = *OptResult;
1457   intersectAssumeOrGuardBlockValueConstantRange(V, Result, CxtI);
1458 
1459   LLVM_DEBUG(dbgs() << "  Result = " << Result << "\n");
1460   return Result;
1461 }
1462 
getValueAt(Value * V,Instruction * CxtI)1463 ValueLatticeElement LazyValueInfoImpl::getValueAt(Value *V, Instruction *CxtI) {
1464   LLVM_DEBUG(dbgs() << "LVI Getting value " << *V << " at '" << CxtI->getName()
1465                     << "'\n");
1466 
1467   if (auto *C = dyn_cast<Constant>(V))
1468     return ValueLatticeElement::get(C);
1469 
1470   ValueLatticeElement Result = ValueLatticeElement::getOverdefined();
1471   if (auto *I = dyn_cast<Instruction>(V))
1472     Result = getFromRangeMetadata(I);
1473   intersectAssumeOrGuardBlockValueConstantRange(V, Result, CxtI);
1474 
1475   LLVM_DEBUG(dbgs() << "  Result = " << Result << "\n");
1476   return Result;
1477 }
1478 
1479 ValueLatticeElement LazyValueInfoImpl::
getValueOnEdge(Value * V,BasicBlock * FromBB,BasicBlock * ToBB,Instruction * CxtI)1480 getValueOnEdge(Value *V, BasicBlock *FromBB, BasicBlock *ToBB,
1481                Instruction *CxtI) {
1482   LLVM_DEBUG(dbgs() << "LVI Getting edge value " << *V << " from '"
1483                     << FromBB->getName() << "' to '" << ToBB->getName()
1484                     << "'\n");
1485 
1486   Optional<ValueLatticeElement> Result = getEdgeValue(V, FromBB, ToBB, CxtI);
1487   if (!Result) {
1488     solve();
1489     Result = getEdgeValue(V, FromBB, ToBB, CxtI);
1490     assert(Result && "More work to do after problem solved?");
1491   }
1492 
1493   LLVM_DEBUG(dbgs() << "  Result = " << *Result << "\n");
1494   return *Result;
1495 }
1496 
threadEdge(BasicBlock * PredBB,BasicBlock * OldSucc,BasicBlock * NewSucc)1497 void LazyValueInfoImpl::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc,
1498                                    BasicBlock *NewSucc) {
1499   TheCache.threadEdgeImpl(OldSucc, NewSucc);
1500 }
1501 
1502 //===----------------------------------------------------------------------===//
1503 //                            LazyValueInfo Impl
1504 //===----------------------------------------------------------------------===//
1505 
1506 /// This lazily constructs the LazyValueInfoImpl.
getImpl(void * & PImpl,AssumptionCache * AC,const Module * M)1507 static LazyValueInfoImpl &getImpl(void *&PImpl, AssumptionCache *AC,
1508                                   const Module *M) {
1509   if (!PImpl) {
1510     assert(M && "getCache() called with a null Module");
1511     const DataLayout &DL = M->getDataLayout();
1512     Function *GuardDecl = M->getFunction(
1513         Intrinsic::getName(Intrinsic::experimental_guard));
1514     PImpl = new LazyValueInfoImpl(AC, DL, GuardDecl);
1515   }
1516   return *static_cast<LazyValueInfoImpl*>(PImpl);
1517 }
1518 
runOnFunction(Function & F)1519 bool LazyValueInfoWrapperPass::runOnFunction(Function &F) {
1520   Info.AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1521   Info.TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
1522 
1523   if (Info.PImpl)
1524     getImpl(Info.PImpl, Info.AC, F.getParent()).clear();
1525 
1526   // Fully lazy.
1527   return false;
1528 }
1529 
getAnalysisUsage(AnalysisUsage & AU) const1530 void LazyValueInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1531   AU.setPreservesAll();
1532   AU.addRequired<AssumptionCacheTracker>();
1533   AU.addRequired<TargetLibraryInfoWrapperPass>();
1534 }
1535 
getLVI()1536 LazyValueInfo &LazyValueInfoWrapperPass::getLVI() { return Info; }
1537 
~LazyValueInfo()1538 LazyValueInfo::~LazyValueInfo() { releaseMemory(); }
1539 
releaseMemory()1540 void LazyValueInfo::releaseMemory() {
1541   // If the cache was allocated, free it.
1542   if (PImpl) {
1543     delete &getImpl(PImpl, AC, nullptr);
1544     PImpl = nullptr;
1545   }
1546 }
1547 
invalidate(Function & F,const PreservedAnalyses & PA,FunctionAnalysisManager::Invalidator & Inv)1548 bool LazyValueInfo::invalidate(Function &F, const PreservedAnalyses &PA,
1549                                FunctionAnalysisManager::Invalidator &Inv) {
1550   // We need to invalidate if we have either failed to preserve this analyses
1551   // result directly or if any of its dependencies have been invalidated.
1552   auto PAC = PA.getChecker<LazyValueAnalysis>();
1553   if (!(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>()))
1554     return true;
1555 
1556   return false;
1557 }
1558 
releaseMemory()1559 void LazyValueInfoWrapperPass::releaseMemory() { Info.releaseMemory(); }
1560 
run(Function & F,FunctionAnalysisManager & FAM)1561 LazyValueInfo LazyValueAnalysis::run(Function &F,
1562                                      FunctionAnalysisManager &FAM) {
1563   auto &AC = FAM.getResult<AssumptionAnalysis>(F);
1564   auto &TLI = FAM.getResult<TargetLibraryAnalysis>(F);
1565 
1566   return LazyValueInfo(&AC, &F.getParent()->getDataLayout(), &TLI);
1567 }
1568 
1569 /// Returns true if we can statically tell that this value will never be a
1570 /// "useful" constant.  In practice, this means we've got something like an
1571 /// alloca or a malloc call for which a comparison against a constant can
1572 /// only be guarding dead code.  Note that we are potentially giving up some
1573 /// precision in dead code (a constant result) in favour of avoiding a
1574 /// expensive search for a easily answered common query.
isKnownNonConstant(Value * V)1575 static bool isKnownNonConstant(Value *V) {
1576   V = V->stripPointerCasts();
1577   // The return val of alloc cannot be a Constant.
1578   if (isa<AllocaInst>(V))
1579     return true;
1580   return false;
1581 }
1582 
getConstant(Value * V,Instruction * CxtI)1583 Constant *LazyValueInfo::getConstant(Value *V, Instruction *CxtI) {
1584   // Bail out early if V is known not to be a Constant.
1585   if (isKnownNonConstant(V))
1586     return nullptr;
1587 
1588   BasicBlock *BB = CxtI->getParent();
1589   ValueLatticeElement Result =
1590       getImpl(PImpl, AC, BB->getModule()).getValueInBlock(V, BB, CxtI);
1591 
1592   if (Result.isConstant())
1593     return Result.getConstant();
1594   if (Result.isConstantRange()) {
1595     const ConstantRange &CR = Result.getConstantRange();
1596     if (const APInt *SingleVal = CR.getSingleElement())
1597       return ConstantInt::get(V->getContext(), *SingleVal);
1598   }
1599   return nullptr;
1600 }
1601 
getConstantRange(Value * V,Instruction * CxtI,bool UndefAllowed)1602 ConstantRange LazyValueInfo::getConstantRange(Value *V, Instruction *CxtI,
1603                                               bool UndefAllowed) {
1604   assert(V->getType()->isIntegerTy());
1605   unsigned Width = V->getType()->getIntegerBitWidth();
1606   BasicBlock *BB = CxtI->getParent();
1607   ValueLatticeElement Result =
1608       getImpl(PImpl, AC, BB->getModule()).getValueInBlock(V, BB, CxtI);
1609   if (Result.isUnknown())
1610     return ConstantRange::getEmpty(Width);
1611   if (Result.isConstantRange(UndefAllowed))
1612     return Result.getConstantRange(UndefAllowed);
1613   // We represent ConstantInt constants as constant ranges but other kinds
1614   // of integer constants, i.e. ConstantExpr will be tagged as constants
1615   assert(!(Result.isConstant() && isa<ConstantInt>(Result.getConstant())) &&
1616          "ConstantInt value must be represented as constantrange");
1617   return ConstantRange::getFull(Width);
1618 }
1619 
1620 /// Determine whether the specified value is known to be a
1621 /// constant on the specified edge. Return null if not.
getConstantOnEdge(Value * V,BasicBlock * FromBB,BasicBlock * ToBB,Instruction * CxtI)1622 Constant *LazyValueInfo::getConstantOnEdge(Value *V, BasicBlock *FromBB,
1623                                            BasicBlock *ToBB,
1624                                            Instruction *CxtI) {
1625   Module *M = FromBB->getModule();
1626   ValueLatticeElement Result =
1627       getImpl(PImpl, AC, M).getValueOnEdge(V, FromBB, ToBB, CxtI);
1628 
1629   if (Result.isConstant())
1630     return Result.getConstant();
1631   if (Result.isConstantRange()) {
1632     const ConstantRange &CR = Result.getConstantRange();
1633     if (const APInt *SingleVal = CR.getSingleElement())
1634       return ConstantInt::get(V->getContext(), *SingleVal);
1635   }
1636   return nullptr;
1637 }
1638 
getConstantRangeOnEdge(Value * V,BasicBlock * FromBB,BasicBlock * ToBB,Instruction * CxtI)1639 ConstantRange LazyValueInfo::getConstantRangeOnEdge(Value *V,
1640                                                     BasicBlock *FromBB,
1641                                                     BasicBlock *ToBB,
1642                                                     Instruction *CxtI) {
1643   unsigned Width = V->getType()->getIntegerBitWidth();
1644   Module *M = FromBB->getModule();
1645   ValueLatticeElement Result =
1646       getImpl(PImpl, AC, M).getValueOnEdge(V, FromBB, ToBB, CxtI);
1647 
1648   if (Result.isUnknown())
1649     return ConstantRange::getEmpty(Width);
1650   if (Result.isConstantRange())
1651     return Result.getConstantRange();
1652   // We represent ConstantInt constants as constant ranges but other kinds
1653   // of integer constants, i.e. ConstantExpr will be tagged as constants
1654   assert(!(Result.isConstant() && isa<ConstantInt>(Result.getConstant())) &&
1655          "ConstantInt value must be represented as constantrange");
1656   return ConstantRange::getFull(Width);
1657 }
1658 
1659 static LazyValueInfo::Tristate
getPredicateResult(unsigned Pred,Constant * C,const ValueLatticeElement & Val,const DataLayout & DL,TargetLibraryInfo * TLI)1660 getPredicateResult(unsigned Pred, Constant *C, const ValueLatticeElement &Val,
1661                    const DataLayout &DL, TargetLibraryInfo *TLI) {
1662   // If we know the value is a constant, evaluate the conditional.
1663   Constant *Res = nullptr;
1664   if (Val.isConstant()) {
1665     Res = ConstantFoldCompareInstOperands(Pred, Val.getConstant(), C, DL, TLI);
1666     if (ConstantInt *ResCI = dyn_cast<ConstantInt>(Res))
1667       return ResCI->isZero() ? LazyValueInfo::False : LazyValueInfo::True;
1668     return LazyValueInfo::Unknown;
1669   }
1670 
1671   if (Val.isConstantRange()) {
1672     ConstantInt *CI = dyn_cast<ConstantInt>(C);
1673     if (!CI) return LazyValueInfo::Unknown;
1674 
1675     const ConstantRange &CR = Val.getConstantRange();
1676     if (Pred == ICmpInst::ICMP_EQ) {
1677       if (!CR.contains(CI->getValue()))
1678         return LazyValueInfo::False;
1679 
1680       if (CR.isSingleElement())
1681         return LazyValueInfo::True;
1682     } else if (Pred == ICmpInst::ICMP_NE) {
1683       if (!CR.contains(CI->getValue()))
1684         return LazyValueInfo::True;
1685 
1686       if (CR.isSingleElement())
1687         return LazyValueInfo::False;
1688     } else {
1689       // Handle more complex predicates.
1690       ConstantRange TrueValues = ConstantRange::makeExactICmpRegion(
1691           (ICmpInst::Predicate)Pred, CI->getValue());
1692       if (TrueValues.contains(CR))
1693         return LazyValueInfo::True;
1694       if (TrueValues.inverse().contains(CR))
1695         return LazyValueInfo::False;
1696     }
1697     return LazyValueInfo::Unknown;
1698   }
1699 
1700   if (Val.isNotConstant()) {
1701     // If this is an equality comparison, we can try to fold it knowing that
1702     // "V != C1".
1703     if (Pred == ICmpInst::ICMP_EQ) {
1704       // !C1 == C -> false iff C1 == C.
1705       Res = ConstantFoldCompareInstOperands(ICmpInst::ICMP_NE,
1706                                             Val.getNotConstant(), C, DL,
1707                                             TLI);
1708       if (Res->isNullValue())
1709         return LazyValueInfo::False;
1710     } else if (Pred == ICmpInst::ICMP_NE) {
1711       // !C1 != C -> true iff C1 == C.
1712       Res = ConstantFoldCompareInstOperands(ICmpInst::ICMP_NE,
1713                                             Val.getNotConstant(), C, DL,
1714                                             TLI);
1715       if (Res->isNullValue())
1716         return LazyValueInfo::True;
1717     }
1718     return LazyValueInfo::Unknown;
1719   }
1720 
1721   return LazyValueInfo::Unknown;
1722 }
1723 
1724 /// Determine whether the specified value comparison with a constant is known to
1725 /// be true or false on the specified CFG edge. Pred is a CmpInst predicate.
1726 LazyValueInfo::Tristate
getPredicateOnEdge(unsigned Pred,Value * V,Constant * C,BasicBlock * FromBB,BasicBlock * ToBB,Instruction * CxtI)1727 LazyValueInfo::getPredicateOnEdge(unsigned Pred, Value *V, Constant *C,
1728                                   BasicBlock *FromBB, BasicBlock *ToBB,
1729                                   Instruction *CxtI) {
1730   Module *M = FromBB->getModule();
1731   ValueLatticeElement Result =
1732       getImpl(PImpl, AC, M).getValueOnEdge(V, FromBB, ToBB, CxtI);
1733 
1734   return getPredicateResult(Pred, C, Result, M->getDataLayout(), TLI);
1735 }
1736 
1737 LazyValueInfo::Tristate
getPredicateAt(unsigned Pred,Value * V,Constant * C,Instruction * CxtI,bool UseBlockValue)1738 LazyValueInfo::getPredicateAt(unsigned Pred, Value *V, Constant *C,
1739                               Instruction *CxtI, bool UseBlockValue) {
1740   // Is or is not NonNull are common predicates being queried. If
1741   // isKnownNonZero can tell us the result of the predicate, we can
1742   // return it quickly. But this is only a fastpath, and falling
1743   // through would still be correct.
1744   Module *M = CxtI->getModule();
1745   const DataLayout &DL = M->getDataLayout();
1746   if (V->getType()->isPointerTy() && C->isNullValue() &&
1747       isKnownNonZero(V->stripPointerCastsSameRepresentation(), DL)) {
1748     if (Pred == ICmpInst::ICMP_EQ)
1749       return LazyValueInfo::False;
1750     else if (Pred == ICmpInst::ICMP_NE)
1751       return LazyValueInfo::True;
1752   }
1753 
1754   ValueLatticeElement Result = UseBlockValue
1755       ? getImpl(PImpl, AC, M).getValueInBlock(V, CxtI->getParent(), CxtI)
1756       : getImpl(PImpl, AC, M).getValueAt(V, CxtI);
1757   Tristate Ret = getPredicateResult(Pred, C, Result, DL, TLI);
1758   if (Ret != Unknown)
1759     return Ret;
1760 
1761   // Note: The following bit of code is somewhat distinct from the rest of LVI;
1762   // LVI as a whole tries to compute a lattice value which is conservatively
1763   // correct at a given location.  In this case, we have a predicate which we
1764   // weren't able to prove about the merged result, and we're pushing that
1765   // predicate back along each incoming edge to see if we can prove it
1766   // separately for each input.  As a motivating example, consider:
1767   // bb1:
1768   //   %v1 = ... ; constantrange<1, 5>
1769   //   br label %merge
1770   // bb2:
1771   //   %v2 = ... ; constantrange<10, 20>
1772   //   br label %merge
1773   // merge:
1774   //   %phi = phi [%v1, %v2] ; constantrange<1,20>
1775   //   %pred = icmp eq i32 %phi, 8
1776   // We can't tell from the lattice value for '%phi' that '%pred' is false
1777   // along each path, but by checking the predicate over each input separately,
1778   // we can.
1779   // We limit the search to one step backwards from the current BB and value.
1780   // We could consider extending this to search further backwards through the
1781   // CFG and/or value graph, but there are non-obvious compile time vs quality
1782   // tradeoffs.
1783   if (CxtI) {
1784     BasicBlock *BB = CxtI->getParent();
1785 
1786     // Function entry or an unreachable block.  Bail to avoid confusing
1787     // analysis below.
1788     pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
1789     if (PI == PE)
1790       return Unknown;
1791 
1792     // If V is a PHI node in the same block as the context, we need to ask
1793     // questions about the predicate as applied to the incoming value along
1794     // each edge. This is useful for eliminating cases where the predicate is
1795     // known along all incoming edges.
1796     if (auto *PHI = dyn_cast<PHINode>(V))
1797       if (PHI->getParent() == BB) {
1798         Tristate Baseline = Unknown;
1799         for (unsigned i = 0, e = PHI->getNumIncomingValues(); i < e; i++) {
1800           Value *Incoming = PHI->getIncomingValue(i);
1801           BasicBlock *PredBB = PHI->getIncomingBlock(i);
1802           // Note that PredBB may be BB itself.
1803           Tristate Result = getPredicateOnEdge(Pred, Incoming, C, PredBB, BB,
1804                                                CxtI);
1805 
1806           // Keep going as long as we've seen a consistent known result for
1807           // all inputs.
1808           Baseline = (i == 0) ? Result /* First iteration */
1809             : (Baseline == Result ? Baseline : Unknown); /* All others */
1810           if (Baseline == Unknown)
1811             break;
1812         }
1813         if (Baseline != Unknown)
1814           return Baseline;
1815       }
1816 
1817     // For a comparison where the V is outside this block, it's possible
1818     // that we've branched on it before. Look to see if the value is known
1819     // on all incoming edges.
1820     if (!isa<Instruction>(V) ||
1821         cast<Instruction>(V)->getParent() != BB) {
1822       // For predecessor edge, determine if the comparison is true or false
1823       // on that edge. If they're all true or all false, we can conclude
1824       // the value of the comparison in this block.
1825       Tristate Baseline = getPredicateOnEdge(Pred, V, C, *PI, BB, CxtI);
1826       if (Baseline != Unknown) {
1827         // Check that all remaining incoming values match the first one.
1828         while (++PI != PE) {
1829           Tristate Ret = getPredicateOnEdge(Pred, V, C, *PI, BB, CxtI);
1830           if (Ret != Baseline) break;
1831         }
1832         // If we terminated early, then one of the values didn't match.
1833         if (PI == PE) {
1834           return Baseline;
1835         }
1836       }
1837     }
1838   }
1839   return Unknown;
1840 }
1841 
getPredicateAt(unsigned P,Value * LHS,Value * RHS,Instruction * CxtI,bool UseBlockValue)1842 LazyValueInfo::Tristate LazyValueInfo::getPredicateAt(unsigned P, Value *LHS,
1843                                                       Value *RHS,
1844                                                       Instruction *CxtI,
1845                                                       bool UseBlockValue) {
1846   CmpInst::Predicate Pred = (CmpInst::Predicate)P;
1847 
1848   if (auto *C = dyn_cast<Constant>(RHS))
1849     return getPredicateAt(P, LHS, C, CxtI, UseBlockValue);
1850   if (auto *C = dyn_cast<Constant>(LHS))
1851     return getPredicateAt(CmpInst::getSwappedPredicate(Pred), RHS, C, CxtI,
1852                           UseBlockValue);
1853 
1854   // Got two non-Constant values. While we could handle them somewhat,
1855   // by getting their constant ranges, and applying ConstantRange::icmp(),
1856   // so far it did not appear to be profitable.
1857   return LazyValueInfo::Unknown;
1858 }
1859 
threadEdge(BasicBlock * PredBB,BasicBlock * OldSucc,BasicBlock * NewSucc)1860 void LazyValueInfo::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc,
1861                                BasicBlock *NewSucc) {
1862   if (PImpl) {
1863     getImpl(PImpl, AC, PredBB->getModule())
1864         .threadEdge(PredBB, OldSucc, NewSucc);
1865   }
1866 }
1867 
eraseBlock(BasicBlock * BB)1868 void LazyValueInfo::eraseBlock(BasicBlock *BB) {
1869   if (PImpl) {
1870     getImpl(PImpl, AC, BB->getModule()).eraseBlock(BB);
1871   }
1872 }
1873 
1874 
printLVI(Function & F,DominatorTree & DTree,raw_ostream & OS)1875 void LazyValueInfo::printLVI(Function &F, DominatorTree &DTree, raw_ostream &OS) {
1876   if (PImpl) {
1877     getImpl(PImpl, AC, F.getParent()).printLVI(F, DTree, OS);
1878   }
1879 }
1880 
1881 // Print the LVI for the function arguments at the start of each basic block.
emitBasicBlockStartAnnot(const BasicBlock * BB,formatted_raw_ostream & OS)1882 void LazyValueInfoAnnotatedWriter::emitBasicBlockStartAnnot(
1883     const BasicBlock *BB, formatted_raw_ostream &OS) {
1884   // Find if there are latticevalues defined for arguments of the function.
1885   auto *F = BB->getParent();
1886   for (auto &Arg : F->args()) {
1887     ValueLatticeElement Result = LVIImpl->getValueInBlock(
1888         const_cast<Argument *>(&Arg), const_cast<BasicBlock *>(BB));
1889     if (Result.isUnknown())
1890       continue;
1891     OS << "; LatticeVal for: '" << Arg << "' is: " << Result << "\n";
1892   }
1893 }
1894 
1895 // This function prints the LVI analysis for the instruction I at the beginning
1896 // of various basic blocks. It relies on calculated values that are stored in
1897 // the LazyValueInfoCache, and in the absence of cached values, recalculate the
1898 // LazyValueInfo for `I`, and print that info.
emitInstructionAnnot(const Instruction * I,formatted_raw_ostream & OS)1899 void LazyValueInfoAnnotatedWriter::emitInstructionAnnot(
1900     const Instruction *I, formatted_raw_ostream &OS) {
1901 
1902   auto *ParentBB = I->getParent();
1903   SmallPtrSet<const BasicBlock*, 16> BlocksContainingLVI;
1904   // We can generate (solve) LVI values only for blocks that are dominated by
1905   // the I's parent. However, to avoid generating LVI for all dominating blocks,
1906   // that contain redundant/uninteresting information, we print LVI for
1907   // blocks that may use this LVI information (such as immediate successor
1908   // blocks, and blocks that contain uses of `I`).
1909   auto printResult = [&](const BasicBlock *BB) {
1910     if (!BlocksContainingLVI.insert(BB).second)
1911       return;
1912     ValueLatticeElement Result = LVIImpl->getValueInBlock(
1913         const_cast<Instruction *>(I), const_cast<BasicBlock *>(BB));
1914       OS << "; LatticeVal for: '" << *I << "' in BB: '";
1915       BB->printAsOperand(OS, false);
1916       OS << "' is: " << Result << "\n";
1917   };
1918 
1919   printResult(ParentBB);
1920   // Print the LVI analysis results for the immediate successor blocks, that
1921   // are dominated by `ParentBB`.
1922   for (auto *BBSucc : successors(ParentBB))
1923     if (DT.dominates(ParentBB, BBSucc))
1924       printResult(BBSucc);
1925 
1926   // Print LVI in blocks where `I` is used.
1927   for (auto *U : I->users())
1928     if (auto *UseI = dyn_cast<Instruction>(U))
1929       if (!isa<PHINode>(UseI) || DT.dominates(ParentBB, UseI->getParent()))
1930         printResult(UseI->getParent());
1931 
1932 }
1933 
1934 namespace {
1935 // Printer class for LazyValueInfo results.
1936 class LazyValueInfoPrinter : public FunctionPass {
1937 public:
1938   static char ID; // Pass identification, replacement for typeid
LazyValueInfoPrinter()1939   LazyValueInfoPrinter() : FunctionPass(ID) {
1940     initializeLazyValueInfoPrinterPass(*PassRegistry::getPassRegistry());
1941   }
1942 
getAnalysisUsage(AnalysisUsage & AU) const1943   void getAnalysisUsage(AnalysisUsage &AU) const override {
1944     AU.setPreservesAll();
1945     AU.addRequired<LazyValueInfoWrapperPass>();
1946     AU.addRequired<DominatorTreeWrapperPass>();
1947   }
1948 
1949   // Get the mandatory dominator tree analysis and pass this in to the
1950   // LVIPrinter. We cannot rely on the LVI's DT, since it's optional.
runOnFunction(Function & F)1951   bool runOnFunction(Function &F) override {
1952     dbgs() << "LVI for function '" << F.getName() << "':\n";
1953     auto &LVI = getAnalysis<LazyValueInfoWrapperPass>().getLVI();
1954     auto &DTree = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1955     LVI.printLVI(F, DTree, dbgs());
1956     return false;
1957   }
1958 };
1959 }
1960 
1961 char LazyValueInfoPrinter::ID = 0;
1962 INITIALIZE_PASS_BEGIN(LazyValueInfoPrinter, "print-lazy-value-info",
1963                 "Lazy Value Info Printer Pass", false, false)
1964 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)
1965 INITIALIZE_PASS_END(LazyValueInfoPrinter, "print-lazy-value-info",
1966                 "Lazy Value Info Printer Pass", false, false)
1967