xref: /freebsd/contrib/llvm-project/llvm/lib/Analysis/Loads.cpp (revision 2e3507c25e42292b45a5482e116d278f5515d04d)
1 //===- Loads.cpp - Local load analysis ------------------------------------===//
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 simple local analyses for load instructions.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/Analysis/Loads.h"
14 #include "llvm/Analysis/AliasAnalysis.h"
15 #include "llvm/Analysis/AssumeBundleQueries.h"
16 #include "llvm/Analysis/LoopInfo.h"
17 #include "llvm/Analysis/MemoryBuiltins.h"
18 #include "llvm/Analysis/MemoryLocation.h"
19 #include "llvm/Analysis/ScalarEvolution.h"
20 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
21 #include "llvm/Analysis/ValueTracking.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/IntrinsicInst.h"
24 #include "llvm/IR/Module.h"
25 #include "llvm/IR/Operator.h"
26 
27 using namespace llvm;
28 
29 static bool isAligned(const Value *Base, const APInt &Offset, Align Alignment,
30                       const DataLayout &DL) {
31   Align BA = Base->getPointerAlignment(DL);
32   return BA >= Alignment && Offset.isAligned(BA);
33 }
34 
35 /// Test if V is always a pointer to allocated and suitably aligned memory for
36 /// a simple load or store.
37 static bool isDereferenceableAndAlignedPointer(
38     const Value *V, Align Alignment, const APInt &Size, const DataLayout &DL,
39     const Instruction *CtxI, AssumptionCache *AC, const DominatorTree *DT,
40     const TargetLibraryInfo *TLI, SmallPtrSetImpl<const Value *> &Visited,
41     unsigned MaxDepth) {
42   assert(V->getType()->isPointerTy() && "Base must be pointer");
43 
44   // Recursion limit.
45   if (MaxDepth-- == 0)
46     return false;
47 
48   // Already visited?  Bail out, we've likely hit unreachable code.
49   if (!Visited.insert(V).second)
50     return false;
51 
52   // Note that it is not safe to speculate into a malloc'd region because
53   // malloc may return null.
54 
55   // For GEPs, determine if the indexing lands within the allocated object.
56   if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
57     const Value *Base = GEP->getPointerOperand();
58 
59     APInt Offset(DL.getIndexTypeSizeInBits(GEP->getType()), 0);
60     if (!GEP->accumulateConstantOffset(DL, Offset) || Offset.isNegative() ||
61         !Offset.urem(APInt(Offset.getBitWidth(), Alignment.value()))
62              .isMinValue())
63       return false;
64 
65     // If the base pointer is dereferenceable for Offset+Size bytes, then the
66     // GEP (== Base + Offset) is dereferenceable for Size bytes.  If the base
67     // pointer is aligned to Align bytes, and the Offset is divisible by Align
68     // then the GEP (== Base + Offset == k_0 * Align + k_1 * Align) is also
69     // aligned to Align bytes.
70 
71     // Offset and Size may have different bit widths if we have visited an
72     // addrspacecast, so we can't do arithmetic directly on the APInt values.
73     return isDereferenceableAndAlignedPointer(
74         Base, Alignment, Offset + Size.sextOrTrunc(Offset.getBitWidth()), DL,
75         CtxI, AC, DT, TLI, Visited, MaxDepth);
76   }
77 
78   // bitcast instructions are no-ops as far as dereferenceability is concerned.
79   if (const BitCastOperator *BC = dyn_cast<BitCastOperator>(V)) {
80     if (BC->getSrcTy()->isPointerTy())
81       return isDereferenceableAndAlignedPointer(
82         BC->getOperand(0), Alignment, Size, DL, CtxI, AC, DT, TLI,
83           Visited, MaxDepth);
84   }
85 
86   // Recurse into both hands of select.
87   if (const SelectInst *Sel = dyn_cast<SelectInst>(V)) {
88     return isDereferenceableAndAlignedPointer(Sel->getTrueValue(), Alignment,
89                                               Size, DL, CtxI, AC, DT, TLI,
90                                               Visited, MaxDepth) &&
91            isDereferenceableAndAlignedPointer(Sel->getFalseValue(), Alignment,
92                                               Size, DL, CtxI, AC, DT, TLI,
93                                               Visited, MaxDepth);
94   }
95 
96   bool CheckForNonNull, CheckForFreed;
97   APInt KnownDerefBytes(Size.getBitWidth(),
98                         V->getPointerDereferenceableBytes(DL, CheckForNonNull,
99                                                           CheckForFreed));
100   if (KnownDerefBytes.getBoolValue() && KnownDerefBytes.uge(Size) &&
101       !CheckForFreed)
102     if (!CheckForNonNull || isKnownNonZero(V, DL, 0, AC, CtxI, DT)) {
103       // As we recursed through GEPs to get here, we've incrementally checked
104       // that each step advanced by a multiple of the alignment. If our base is
105       // properly aligned, then the original offset accessed must also be.
106       APInt Offset(DL.getTypeStoreSizeInBits(V->getType()), 0);
107       return isAligned(V, Offset, Alignment, DL);
108     }
109 
110   /// TODO refactor this function to be able to search independently for
111   /// Dereferencability and Alignment requirements.
112 
113 
114   if (const auto *Call = dyn_cast<CallBase>(V)) {
115     if (auto *RP = getArgumentAliasingToReturnedPointer(Call, true))
116       return isDereferenceableAndAlignedPointer(RP, Alignment, Size, DL, CtxI,
117                                                 AC, DT, TLI, Visited, MaxDepth);
118 
119     // If we have a call we can't recurse through, check to see if this is an
120     // allocation function for which we can establish an minimum object size.
121     // Such a minimum object size is analogous to a deref_or_null attribute in
122     // that we still need to prove the result non-null at point of use.
123     // NOTE: We can only use the object size as a base fact as we a) need to
124     // prove alignment too, and b) don't want the compile time impact of a
125     // separate recursive walk.
126     ObjectSizeOpts Opts;
127     // TODO: It may be okay to round to align, but that would imply that
128     // accessing slightly out of bounds was legal, and we're currently
129     // inconsistent about that.  For the moment, be conservative.
130     Opts.RoundToAlign = false;
131     Opts.NullIsUnknownSize = true;
132     uint64_t ObjSize;
133     if (getObjectSize(V, ObjSize, DL, TLI, Opts)) {
134       APInt KnownDerefBytes(Size.getBitWidth(), ObjSize);
135       if (KnownDerefBytes.getBoolValue() && KnownDerefBytes.uge(Size) &&
136           isKnownNonZero(V, DL, 0, AC, CtxI, DT) && !V->canBeFreed()) {
137         // As we recursed through GEPs to get here, we've incrementally
138         // checked that each step advanced by a multiple of the alignment. If
139         // our base is properly aligned, then the original offset accessed
140         // must also be.
141         APInt Offset(DL.getTypeStoreSizeInBits(V->getType()), 0);
142         return isAligned(V, Offset, Alignment, DL);
143       }
144     }
145   }
146 
147   // For gc.relocate, look through relocations
148   if (const GCRelocateInst *RelocateInst = dyn_cast<GCRelocateInst>(V))
149     return isDereferenceableAndAlignedPointer(RelocateInst->getDerivedPtr(),
150                                               Alignment, Size, DL, CtxI, AC, DT,
151                                               TLI, Visited, MaxDepth);
152 
153   if (const AddrSpaceCastOperator *ASC = dyn_cast<AddrSpaceCastOperator>(V))
154     return isDereferenceableAndAlignedPointer(ASC->getOperand(0), Alignment,
155                                               Size, DL, CtxI, AC, DT, TLI,
156                                               Visited, MaxDepth);
157 
158   if (CtxI) {
159     /// Look through assumes to see if both dereferencability and alignment can
160     /// be provent by an assume
161     RetainedKnowledge AlignRK;
162     RetainedKnowledge DerefRK;
163     if (getKnowledgeForValue(
164             V, {Attribute::Dereferenceable, Attribute::Alignment}, AC,
165             [&](RetainedKnowledge RK, Instruction *Assume, auto) {
166               if (!isValidAssumeForContext(Assume, CtxI))
167                 return false;
168               if (RK.AttrKind == Attribute::Alignment)
169                 AlignRK = std::max(AlignRK, RK);
170               if (RK.AttrKind == Attribute::Dereferenceable)
171                 DerefRK = std::max(DerefRK, RK);
172               if (AlignRK && DerefRK && AlignRK.ArgValue >= Alignment.value() &&
173                   DerefRK.ArgValue >= Size.getZExtValue())
174                 return true; // We have found what we needed so we stop looking
175               return false;  // Other assumes may have better information. so
176                              // keep looking
177             }))
178       return true;
179   }
180 
181   // If we don't know, assume the worst.
182   return false;
183 }
184 
185 bool llvm::isDereferenceableAndAlignedPointer(
186     const Value *V, Align Alignment, const APInt &Size, const DataLayout &DL,
187     const Instruction *CtxI, AssumptionCache *AC, const DominatorTree *DT,
188     const TargetLibraryInfo *TLI) {
189   // Note: At the moment, Size can be zero.  This ends up being interpreted as
190   // a query of whether [Base, V] is dereferenceable and V is aligned (since
191   // that's what the implementation happened to do).  It's unclear if this is
192   // the desired semantic, but at least SelectionDAG does exercise this case.
193 
194   SmallPtrSet<const Value *, 32> Visited;
195   return ::isDereferenceableAndAlignedPointer(V, Alignment, Size, DL, CtxI, AC,
196                                               DT, TLI, Visited, 16);
197 }
198 
199 bool llvm::isDereferenceableAndAlignedPointer(
200     const Value *V, Type *Ty, Align Alignment, const DataLayout &DL,
201     const Instruction *CtxI, AssumptionCache *AC, const DominatorTree *DT,
202     const TargetLibraryInfo *TLI) {
203   // For unsized types or scalable vectors we don't know exactly how many bytes
204   // are dereferenced, so bail out.
205   if (!Ty->isSized() || Ty->isScalableTy())
206     return false;
207 
208   // When dereferenceability information is provided by a dereferenceable
209   // attribute, we know exactly how many bytes are dereferenceable. If we can
210   // determine the exact offset to the attributed variable, we can use that
211   // information here.
212 
213   APInt AccessSize(DL.getPointerTypeSizeInBits(V->getType()),
214                    DL.getTypeStoreSize(Ty));
215   return isDereferenceableAndAlignedPointer(V, Alignment, AccessSize, DL, CtxI,
216                                             AC, DT, TLI);
217 }
218 
219 bool llvm::isDereferenceablePointer(const Value *V, Type *Ty,
220                                     const DataLayout &DL,
221                                     const Instruction *CtxI,
222                                     AssumptionCache *AC,
223                                     const DominatorTree *DT,
224                                     const TargetLibraryInfo *TLI) {
225   return isDereferenceableAndAlignedPointer(V, Ty, Align(1), DL, CtxI, AC, DT,
226                                             TLI);
227 }
228 
229 /// Test if A and B will obviously have the same value.
230 ///
231 /// This includes recognizing that %t0 and %t1 will have the same
232 /// value in code like this:
233 /// \code
234 ///   %t0 = getelementptr \@a, 0, 3
235 ///   store i32 0, i32* %t0
236 ///   %t1 = getelementptr \@a, 0, 3
237 ///   %t2 = load i32* %t1
238 /// \endcode
239 ///
240 static bool AreEquivalentAddressValues(const Value *A, const Value *B) {
241   // Test if the values are trivially equivalent.
242   if (A == B)
243     return true;
244 
245   // Test if the values come from identical arithmetic instructions.
246   // Use isIdenticalToWhenDefined instead of isIdenticalTo because
247   // this function is only used when one address use dominates the
248   // other, which means that they'll always either have the same
249   // value or one of them will have an undefined value.
250   if (isa<BinaryOperator>(A) || isa<CastInst>(A) || isa<PHINode>(A) ||
251       isa<GetElementPtrInst>(A))
252     if (const Instruction *BI = dyn_cast<Instruction>(B))
253       if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
254         return true;
255 
256   // Otherwise they may not be equivalent.
257   return false;
258 }
259 
260 bool llvm::isDereferenceableAndAlignedInLoop(LoadInst *LI, Loop *L,
261                                              ScalarEvolution &SE,
262                                              DominatorTree &DT,
263                                              AssumptionCache *AC) {
264   auto &DL = LI->getModule()->getDataLayout();
265   Value *Ptr = LI->getPointerOperand();
266 
267   APInt EltSize(DL.getIndexTypeSizeInBits(Ptr->getType()),
268                 DL.getTypeStoreSize(LI->getType()).getFixedValue());
269   const Align Alignment = LI->getAlign();
270 
271   Instruction *HeaderFirstNonPHI = L->getHeader()->getFirstNonPHI();
272 
273   // If given a uniform (i.e. non-varying) address, see if we can prove the
274   // access is safe within the loop w/o needing predication.
275   if (L->isLoopInvariant(Ptr))
276     return isDereferenceableAndAlignedPointer(Ptr, Alignment, EltSize, DL,
277                                               HeaderFirstNonPHI, AC, &DT);
278 
279   // Otherwise, check to see if we have a repeating access pattern where we can
280   // prove that all accesses are well aligned and dereferenceable.
281   auto *AddRec = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Ptr));
282   if (!AddRec || AddRec->getLoop() != L || !AddRec->isAffine())
283     return false;
284   auto* Step = dyn_cast<SCEVConstant>(AddRec->getStepRecurrence(SE));
285   if (!Step)
286     return false;
287 
288   auto TC = SE.getSmallConstantMaxTripCount(L);
289   if (!TC)
290     return false;
291 
292   // TODO: Handle overlapping accesses.
293   // We should be computing AccessSize as (TC - 1) * Step + EltSize.
294   if (EltSize.sgt(Step->getAPInt()))
295     return false;
296 
297   // Compute the total access size for access patterns with unit stride and
298   // patterns with gaps. For patterns with unit stride, Step and EltSize are the
299   // same.
300   // For patterns with gaps (i.e. non unit stride), we are
301   // accessing EltSize bytes at every Step.
302   APInt AccessSize = TC * Step->getAPInt();
303 
304   assert(SE.isLoopInvariant(AddRec->getStart(), L) &&
305          "implied by addrec definition");
306   Value *Base = nullptr;
307   if (auto *StartS = dyn_cast<SCEVUnknown>(AddRec->getStart())) {
308     Base = StartS->getValue();
309   } else if (auto *StartS = dyn_cast<SCEVAddExpr>(AddRec->getStart())) {
310     // Handle (NewBase + offset) as start value.
311     const auto *Offset = dyn_cast<SCEVConstant>(StartS->getOperand(0));
312     const auto *NewBase = dyn_cast<SCEVUnknown>(StartS->getOperand(1));
313     if (StartS->getNumOperands() == 2 && Offset && NewBase) {
314       // For the moment, restrict ourselves to the case where the offset is a
315       // multiple of the requested alignment and the base is aligned.
316       // TODO: generalize if a case found which warrants
317       if (Offset->getAPInt().urem(Alignment.value()) != 0)
318         return false;
319       Base = NewBase->getValue();
320       bool Overflow = false;
321       AccessSize = AccessSize.uadd_ov(Offset->getAPInt(), Overflow);
322       if (Overflow)
323         return false;
324     }
325   }
326 
327   if (!Base)
328     return false;
329 
330   // For the moment, restrict ourselves to the case where the access size is a
331   // multiple of the requested alignment and the base is aligned.
332   // TODO: generalize if a case found which warrants
333   if (EltSize.urem(Alignment.value()) != 0)
334     return false;
335   return isDereferenceableAndAlignedPointer(Base, Alignment, AccessSize, DL,
336                                             HeaderFirstNonPHI, AC, &DT);
337 }
338 
339 /// Check if executing a load of this pointer value cannot trap.
340 ///
341 /// If DT and ScanFrom are specified this method performs context-sensitive
342 /// analysis and returns true if it is safe to load immediately before ScanFrom.
343 ///
344 /// If it is not obviously safe to load from the specified pointer, we do
345 /// a quick local scan of the basic block containing \c ScanFrom, to determine
346 /// if the address is already accessed.
347 ///
348 /// This uses the pointee type to determine how many bytes need to be safe to
349 /// load from the pointer.
350 bool llvm::isSafeToLoadUnconditionally(Value *V, Align Alignment, APInt &Size,
351                                        const DataLayout &DL,
352                                        Instruction *ScanFrom,
353                                        AssumptionCache *AC,
354                                        const DominatorTree *DT,
355                                        const TargetLibraryInfo *TLI) {
356   // If DT is not specified we can't make context-sensitive query
357   const Instruction* CtxI = DT ? ScanFrom : nullptr;
358   if (isDereferenceableAndAlignedPointer(V, Alignment, Size, DL, CtxI, AC, DT,
359                                          TLI))
360     return true;
361 
362   if (!ScanFrom)
363     return false;
364 
365   if (Size.getBitWidth() > 64)
366     return false;
367   const uint64_t LoadSize = Size.getZExtValue();
368 
369   // Otherwise, be a little bit aggressive by scanning the local block where we
370   // want to check to see if the pointer is already being loaded or stored
371   // from/to.  If so, the previous load or store would have already trapped,
372   // so there is no harm doing an extra load (also, CSE will later eliminate
373   // the load entirely).
374   BasicBlock::iterator BBI = ScanFrom->getIterator(),
375                        E = ScanFrom->getParent()->begin();
376 
377   // We can at least always strip pointer casts even though we can't use the
378   // base here.
379   V = V->stripPointerCasts();
380 
381   while (BBI != E) {
382     --BBI;
383 
384     // If we see a free or a call which may write to memory (i.e. which might do
385     // a free) the pointer could be marked invalid.
386     if (isa<CallInst>(BBI) && BBI->mayWriteToMemory() &&
387         !isa<LifetimeIntrinsic>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
388       return false;
389 
390     Value *AccessedPtr;
391     Type *AccessedTy;
392     Align AccessedAlign;
393     if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
394       // Ignore volatile loads. The execution of a volatile load cannot
395       // be used to prove an address is backed by regular memory; it can,
396       // for example, point to an MMIO register.
397       if (LI->isVolatile())
398         continue;
399       AccessedPtr = LI->getPointerOperand();
400       AccessedTy = LI->getType();
401       AccessedAlign = LI->getAlign();
402     } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
403       // Ignore volatile stores (see comment for loads).
404       if (SI->isVolatile())
405         continue;
406       AccessedPtr = SI->getPointerOperand();
407       AccessedTy = SI->getValueOperand()->getType();
408       AccessedAlign = SI->getAlign();
409     } else
410       continue;
411 
412     if (AccessedAlign < Alignment)
413       continue;
414 
415     // Handle trivial cases.
416     if (AccessedPtr == V &&
417         LoadSize <= DL.getTypeStoreSize(AccessedTy))
418       return true;
419 
420     if (AreEquivalentAddressValues(AccessedPtr->stripPointerCasts(), V) &&
421         LoadSize <= DL.getTypeStoreSize(AccessedTy))
422       return true;
423   }
424   return false;
425 }
426 
427 bool llvm::isSafeToLoadUnconditionally(Value *V, Type *Ty, Align Alignment,
428                                        const DataLayout &DL,
429                                        Instruction *ScanFrom,
430                                        AssumptionCache *AC,
431                                        const DominatorTree *DT,
432                                        const TargetLibraryInfo *TLI) {
433   TypeSize TySize = DL.getTypeStoreSize(Ty);
434   if (TySize.isScalable())
435     return false;
436   APInt Size(DL.getIndexTypeSizeInBits(V->getType()), TySize.getFixedValue());
437   return isSafeToLoadUnconditionally(V, Alignment, Size, DL, ScanFrom, AC, DT,
438                                      TLI);
439 }
440 
441 /// DefMaxInstsToScan - the default number of maximum instructions
442 /// to scan in the block, used by FindAvailableLoadedValue().
443 /// FindAvailableLoadedValue() was introduced in r60148, to improve jump
444 /// threading in part by eliminating partially redundant loads.
445 /// At that point, the value of MaxInstsToScan was already set to '6'
446 /// without documented explanation.
447 cl::opt<unsigned>
448 llvm::DefMaxInstsToScan("available-load-scan-limit", cl::init(6), cl::Hidden,
449   cl::desc("Use this to specify the default maximum number of instructions "
450            "to scan backward from a given instruction, when searching for "
451            "available loaded value"));
452 
453 Value *llvm::FindAvailableLoadedValue(LoadInst *Load,
454                                       BasicBlock *ScanBB,
455                                       BasicBlock::iterator &ScanFrom,
456                                       unsigned MaxInstsToScan,
457                                       AAResults *AA, bool *IsLoad,
458                                       unsigned *NumScanedInst) {
459   // Don't CSE load that is volatile or anything stronger than unordered.
460   if (!Load->isUnordered())
461     return nullptr;
462 
463   MemoryLocation Loc = MemoryLocation::get(Load);
464   return findAvailablePtrLoadStore(Loc, Load->getType(), Load->isAtomic(),
465                                    ScanBB, ScanFrom, MaxInstsToScan, AA, IsLoad,
466                                    NumScanedInst);
467 }
468 
469 // Check if the load and the store have the same base, constant offsets and
470 // non-overlapping access ranges.
471 static bool areNonOverlapSameBaseLoadAndStore(const Value *LoadPtr,
472                                               Type *LoadTy,
473                                               const Value *StorePtr,
474                                               Type *StoreTy,
475                                               const DataLayout &DL) {
476   APInt LoadOffset(DL.getIndexTypeSizeInBits(LoadPtr->getType()), 0);
477   APInt StoreOffset(DL.getIndexTypeSizeInBits(StorePtr->getType()), 0);
478   const Value *LoadBase = LoadPtr->stripAndAccumulateConstantOffsets(
479       DL, LoadOffset, /* AllowNonInbounds */ false);
480   const Value *StoreBase = StorePtr->stripAndAccumulateConstantOffsets(
481       DL, StoreOffset, /* AllowNonInbounds */ false);
482   if (LoadBase != StoreBase)
483     return false;
484   auto LoadAccessSize = LocationSize::precise(DL.getTypeStoreSize(LoadTy));
485   auto StoreAccessSize = LocationSize::precise(DL.getTypeStoreSize(StoreTy));
486   ConstantRange LoadRange(LoadOffset,
487                           LoadOffset + LoadAccessSize.toRaw());
488   ConstantRange StoreRange(StoreOffset,
489                            StoreOffset + StoreAccessSize.toRaw());
490   return LoadRange.intersectWith(StoreRange).isEmptySet();
491 }
492 
493 static Value *getAvailableLoadStore(Instruction *Inst, const Value *Ptr,
494                                     Type *AccessTy, bool AtLeastAtomic,
495                                     const DataLayout &DL, bool *IsLoadCSE) {
496   // If this is a load of Ptr, the loaded value is available.
497   // (This is true even if the load is volatile or atomic, although
498   // those cases are unlikely.)
499   if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
500     // We can value forward from an atomic to a non-atomic, but not the
501     // other way around.
502     if (LI->isAtomic() < AtLeastAtomic)
503       return nullptr;
504 
505     Value *LoadPtr = LI->getPointerOperand()->stripPointerCasts();
506     if (!AreEquivalentAddressValues(LoadPtr, Ptr))
507       return nullptr;
508 
509     if (CastInst::isBitOrNoopPointerCastable(LI->getType(), AccessTy, DL)) {
510       if (IsLoadCSE)
511         *IsLoadCSE = true;
512       return LI;
513     }
514   }
515 
516   // If this is a store through Ptr, the value is available!
517   // (This is true even if the store is volatile or atomic, although
518   // those cases are unlikely.)
519   if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
520     // We can value forward from an atomic to a non-atomic, but not the
521     // other way around.
522     if (SI->isAtomic() < AtLeastAtomic)
523       return nullptr;
524 
525     Value *StorePtr = SI->getPointerOperand()->stripPointerCasts();
526     if (!AreEquivalentAddressValues(StorePtr, Ptr))
527       return nullptr;
528 
529     if (IsLoadCSE)
530       *IsLoadCSE = false;
531 
532     Value *Val = SI->getValueOperand();
533     if (CastInst::isBitOrNoopPointerCastable(Val->getType(), AccessTy, DL))
534       return Val;
535 
536     TypeSize StoreSize = DL.getTypeSizeInBits(Val->getType());
537     TypeSize LoadSize = DL.getTypeSizeInBits(AccessTy);
538     if (TypeSize::isKnownLE(LoadSize, StoreSize))
539       if (auto *C = dyn_cast<Constant>(Val))
540         return ConstantFoldLoadFromConst(C, AccessTy, DL);
541   }
542 
543   if (auto *MSI = dyn_cast<MemSetInst>(Inst)) {
544     // Don't forward from (non-atomic) memset to atomic load.
545     if (AtLeastAtomic)
546       return nullptr;
547 
548     // Only handle constant memsets.
549     auto *Val = dyn_cast<ConstantInt>(MSI->getValue());
550     auto *Len = dyn_cast<ConstantInt>(MSI->getLength());
551     if (!Val || !Len)
552       return nullptr;
553 
554     // TODO: Handle offsets.
555     Value *Dst = MSI->getDest();
556     if (!AreEquivalentAddressValues(Dst, Ptr))
557       return nullptr;
558 
559     if (IsLoadCSE)
560       *IsLoadCSE = false;
561 
562     TypeSize LoadTypeSize = DL.getTypeSizeInBits(AccessTy);
563     if (LoadTypeSize.isScalable())
564       return nullptr;
565 
566     // Make sure the read bytes are contained in the memset.
567     uint64_t LoadSize = LoadTypeSize.getFixedValue();
568     if ((Len->getValue() * 8).ult(LoadSize))
569       return nullptr;
570 
571     APInt Splat = LoadSize >= 8 ? APInt::getSplat(LoadSize, Val->getValue())
572                                 : Val->getValue().trunc(LoadSize);
573     ConstantInt *SplatC = ConstantInt::get(MSI->getContext(), Splat);
574     if (CastInst::isBitOrNoopPointerCastable(SplatC->getType(), AccessTy, DL))
575       return SplatC;
576 
577     return nullptr;
578   }
579 
580   return nullptr;
581 }
582 
583 Value *llvm::findAvailablePtrLoadStore(
584     const MemoryLocation &Loc, Type *AccessTy, bool AtLeastAtomic,
585     BasicBlock *ScanBB, BasicBlock::iterator &ScanFrom, unsigned MaxInstsToScan,
586     AAResults *AA, bool *IsLoadCSE, unsigned *NumScanedInst) {
587   if (MaxInstsToScan == 0)
588     MaxInstsToScan = ~0U;
589 
590   const DataLayout &DL = ScanBB->getModule()->getDataLayout();
591   const Value *StrippedPtr = Loc.Ptr->stripPointerCasts();
592 
593   while (ScanFrom != ScanBB->begin()) {
594     // We must ignore debug info directives when counting (otherwise they
595     // would affect codegen).
596     Instruction *Inst = &*--ScanFrom;
597     if (Inst->isDebugOrPseudoInst())
598       continue;
599 
600     // Restore ScanFrom to expected value in case next test succeeds
601     ScanFrom++;
602 
603     if (NumScanedInst)
604       ++(*NumScanedInst);
605 
606     // Don't scan huge blocks.
607     if (MaxInstsToScan-- == 0)
608       return nullptr;
609 
610     --ScanFrom;
611 
612     if (Value *Available = getAvailableLoadStore(Inst, StrippedPtr, AccessTy,
613                                                  AtLeastAtomic, DL, IsLoadCSE))
614       return Available;
615 
616     // Try to get the store size for the type.
617     if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
618       Value *StorePtr = SI->getPointerOperand()->stripPointerCasts();
619 
620       // If both StrippedPtr and StorePtr reach all the way to an alloca or
621       // global and they are different, ignore the store. This is a trivial form
622       // of alias analysis that is important for reg2mem'd code.
623       if ((isa<AllocaInst>(StrippedPtr) || isa<GlobalVariable>(StrippedPtr)) &&
624           (isa<AllocaInst>(StorePtr) || isa<GlobalVariable>(StorePtr)) &&
625           StrippedPtr != StorePtr)
626         continue;
627 
628       if (!AA) {
629         // When AA isn't available, but if the load and the store have the same
630         // base, constant offsets and non-overlapping access ranges, ignore the
631         // store. This is a simple form of alias analysis that is used by the
632         // inliner. FIXME: use BasicAA if possible.
633         if (areNonOverlapSameBaseLoadAndStore(
634                 Loc.Ptr, AccessTy, SI->getPointerOperand(),
635                 SI->getValueOperand()->getType(), DL))
636           continue;
637       } else {
638         // If we have alias analysis and it says the store won't modify the
639         // loaded value, ignore the store.
640         if (!isModSet(AA->getModRefInfo(SI, Loc)))
641           continue;
642       }
643 
644       // Otherwise the store that may or may not alias the pointer, bail out.
645       ++ScanFrom;
646       return nullptr;
647     }
648 
649     // If this is some other instruction that may clobber Ptr, bail out.
650     if (Inst->mayWriteToMemory()) {
651       // If alias analysis claims that it really won't modify the load,
652       // ignore it.
653       if (AA && !isModSet(AA->getModRefInfo(Inst, Loc)))
654         continue;
655 
656       // May modify the pointer, bail out.
657       ++ScanFrom;
658       return nullptr;
659     }
660   }
661 
662   // Got to the start of the block, we didn't find it, but are done for this
663   // block.
664   return nullptr;
665 }
666 
667 Value *llvm::FindAvailableLoadedValue(LoadInst *Load, AAResults &AA,
668                                       bool *IsLoadCSE,
669                                       unsigned MaxInstsToScan) {
670   const DataLayout &DL = Load->getModule()->getDataLayout();
671   Value *StrippedPtr = Load->getPointerOperand()->stripPointerCasts();
672   BasicBlock *ScanBB = Load->getParent();
673   Type *AccessTy = Load->getType();
674   bool AtLeastAtomic = Load->isAtomic();
675 
676   if (!Load->isUnordered())
677     return nullptr;
678 
679   // Try to find an available value first, and delay expensive alias analysis
680   // queries until later.
681   Value *Available = nullptr;
682   SmallVector<Instruction *> MustNotAliasInsts;
683   for (Instruction &Inst : make_range(++Load->getReverseIterator(),
684                                       ScanBB->rend())) {
685     if (Inst.isDebugOrPseudoInst())
686       continue;
687 
688     if (MaxInstsToScan-- == 0)
689       return nullptr;
690 
691     Available = getAvailableLoadStore(&Inst, StrippedPtr, AccessTy,
692                                       AtLeastAtomic, DL, IsLoadCSE);
693     if (Available)
694       break;
695 
696     if (Inst.mayWriteToMemory())
697       MustNotAliasInsts.push_back(&Inst);
698   }
699 
700   // If we found an available value, ensure that the instructions in between
701   // did not modify the memory location.
702   if (Available) {
703     MemoryLocation Loc = MemoryLocation::get(Load);
704     for (Instruction *Inst : MustNotAliasInsts)
705       if (isModSet(AA.getModRefInfo(Inst, Loc)))
706         return nullptr;
707   }
708 
709   return Available;
710 }
711 
712 bool llvm::canReplacePointersIfEqual(Value *A, Value *B, const DataLayout &DL,
713                                      Instruction *CtxI) {
714   Type *Ty = A->getType();
715   assert(Ty == B->getType() && Ty->isPointerTy() &&
716          "values must have matching pointer types");
717 
718   // NOTE: The checks in the function are incomplete and currently miss illegal
719   // cases! The current implementation is a starting point and the
720   // implementation should be made stricter over time.
721   if (auto *C = dyn_cast<Constant>(B)) {
722     // Do not allow replacing a pointer with a constant pointer, unless it is
723     // either null or at least one byte is dereferenceable.
724     APInt OneByte(DL.getPointerTypeSizeInBits(Ty), 1);
725     return C->isNullValue() ||
726            isDereferenceableAndAlignedPointer(B, Align(1), OneByte, DL, CtxI);
727   }
728 
729   return true;
730 }
731