1 //===- TypeBasedAliasAnalysis.cpp - Type-Based Alias 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 the TypeBasedAliasAnalysis pass, which implements
10 // metadata-based TBAA.
11 //
12 // In LLVM IR, memory does not have types, so LLVM's own type system is not
13 // suitable for doing TBAA. Instead, metadata is added to the IR to describe
14 // a type system of a higher level language. This can be used to implement
15 // typical C/C++ TBAA, but it can also be used to implement custom alias
16 // analysis behavior for other languages.
17 //
18 // We now support two types of metadata format: scalar TBAA and struct-path
19 // aware TBAA. After all testing cases are upgraded to use struct-path aware
20 // TBAA and we can auto-upgrade existing bc files, the support for scalar TBAA
21 // can be dropped.
22 //
23 // The scalar TBAA metadata format is very simple. TBAA MDNodes have up to
24 // three fields, e.g.:
25 //   !0 = !{ !"an example type tree" }
26 //   !1 = !{ !"int", !0 }
27 //   !2 = !{ !"float", !0 }
28 //   !3 = !{ !"const float", !2, i64 1 }
29 //
30 // The first field is an identity field. It can be any value, usually
31 // an MDString, which uniquely identifies the type. The most important
32 // name in the tree is the name of the root node. Two trees with
33 // different root node names are entirely disjoint, even if they
34 // have leaves with common names.
35 //
36 // The second field identifies the type's parent node in the tree, or
37 // is null or omitted for a root node. A type is considered to alias
38 // all of its descendants and all of its ancestors in the tree. Also,
39 // a type is considered to alias all types in other trees, so that
40 // bitcode produced from multiple front-ends is handled conservatively.
41 //
42 // If the third field is present, it's an integer which if equal to 1
43 // indicates that the type is "constant" (meaning pointsToConstantMemory
44 // should return true; see
45 // http://llvm.org/docs/AliasAnalysis.html#OtherItfs).
46 //
47 // With struct-path aware TBAA, the MDNodes attached to an instruction using
48 // "!tbaa" are called path tag nodes.
49 //
50 // The path tag node has 4 fields with the last field being optional.
51 //
52 // The first field is the base type node, it can be a struct type node
53 // or a scalar type node. The second field is the access type node, it
54 // must be a scalar type node. The third field is the offset into the base type.
55 // The last field has the same meaning as the last field of our scalar TBAA:
56 // it's an integer which if equal to 1 indicates that the access is "constant".
57 //
58 // The struct type node has a name and a list of pairs, one pair for each member
59 // of the struct. The first element of each pair is a type node (a struct type
60 // node or a scalar type node), specifying the type of the member, the second
61 // element of each pair is the offset of the member.
62 //
63 // Given an example
64 // typedef struct {
65 //   short s;
66 // } A;
67 // typedef struct {
68 //   uint16_t s;
69 //   A a;
70 // } B;
71 //
72 // For an access to B.a.s, we attach !5 (a path tag node) to the load/store
73 // instruction. The base type is !4 (struct B), the access type is !2 (scalar
74 // type short) and the offset is 4.
75 //
76 // !0 = !{!"Simple C/C++ TBAA"}
77 // !1 = !{!"omnipotent char", !0} // Scalar type node
78 // !2 = !{!"short", !1}           // Scalar type node
79 // !3 = !{!"A", !2, i64 0}        // Struct type node
80 // !4 = !{!"B", !2, i64 0, !3, i64 4}
81 //                                                           // Struct type node
82 // !5 = !{!4, !2, i64 4}          // Path tag node
83 //
84 // The struct type nodes and the scalar type nodes form a type DAG.
85 //         Root (!0)
86 //         char (!1)  -- edge to Root
87 //         short (!2) -- edge to char
88 //         A (!3) -- edge with offset 0 to short
89 //         B (!4) -- edge with offset 0 to short and edge with offset 4 to A
90 //
91 // To check if two tags (tagX and tagY) can alias, we start from the base type
92 // of tagX, follow the edge with the correct offset in the type DAG and adjust
93 // the offset until we reach the base type of tagY or until we reach the Root
94 // node.
95 // If we reach the base type of tagY, compare the adjusted offset with
96 // offset of tagY, return Alias if the offsets are the same, return NoAlias
97 // otherwise.
98 // If we reach the Root node, perform the above starting from base type of tagY
99 // to see if we reach base type of tagX.
100 //
101 // If they have different roots, they're part of different potentially
102 // unrelated type systems, so we return Alias to be conservative.
103 // If neither node is an ancestor of the other and they have the same root,
104 // then we say NoAlias.
105 //
106 //===----------------------------------------------------------------------===//
107 
108 #include "llvm/Analysis/TypeBasedAliasAnalysis.h"
109 #include "llvm/ADT/SetVector.h"
110 #include "llvm/Analysis/AliasAnalysis.h"
111 #include "llvm/Analysis/MemoryLocation.h"
112 #include "llvm/IR/Constants.h"
113 #include "llvm/IR/DerivedTypes.h"
114 #include "llvm/IR/InstrTypes.h"
115 #include "llvm/IR/Instruction.h"
116 #include "llvm/IR/LLVMContext.h"
117 #include "llvm/IR/Metadata.h"
118 #include "llvm/InitializePasses.h"
119 #include "llvm/Pass.h"
120 #include "llvm/Support/Casting.h"
121 #include "llvm/Support/CommandLine.h"
122 #include "llvm/Support/ErrorHandling.h"
123 #include <cassert>
124 #include <cstdint>
125 
126 using namespace llvm;
127 
128 // A handy option for disabling TBAA functionality. The same effect can also be
129 // achieved by stripping the !tbaa tags from IR, but this option is sometimes
130 // more convenient.
131 static cl::opt<bool> EnableTBAA("enable-tbaa", cl::init(true), cl::Hidden);
132 
133 namespace {
134 
135 /// isNewFormatTypeNode - Return true iff the given type node is in the new
136 /// size-aware format.
137 static bool isNewFormatTypeNode(const MDNode *N) {
138   if (N->getNumOperands() < 3)
139     return false;
140   // In the old format the first operand is a string.
141   if (!isa<MDNode>(N->getOperand(0)))
142     return false;
143   return true;
144 }
145 
146 /// This is a simple wrapper around an MDNode which provides a higher-level
147 /// interface by hiding the details of how alias analysis information is encoded
148 /// in its operands.
149 template<typename MDNodeTy>
150 class TBAANodeImpl {
151   MDNodeTy *Node = nullptr;
152 
153 public:
154   TBAANodeImpl() = default;
155   explicit TBAANodeImpl(MDNodeTy *N) : Node(N) {}
156 
157   /// getNode - Get the MDNode for this TBAANode.
158   MDNodeTy *getNode() const { return Node; }
159 
160   /// isNewFormat - Return true iff the wrapped type node is in the new
161   /// size-aware format.
162   bool isNewFormat() const { return isNewFormatTypeNode(Node); }
163 
164   /// getParent - Get this TBAANode's Alias tree parent.
165   TBAANodeImpl<MDNodeTy> getParent() const {
166     if (isNewFormat())
167       return TBAANodeImpl(cast<MDNodeTy>(Node->getOperand(0)));
168 
169     if (Node->getNumOperands() < 2)
170       return TBAANodeImpl<MDNodeTy>();
171     MDNodeTy *P = dyn_cast_or_null<MDNodeTy>(Node->getOperand(1));
172     if (!P)
173       return TBAANodeImpl<MDNodeTy>();
174     // Ok, this node has a valid parent. Return it.
175     return TBAANodeImpl<MDNodeTy>(P);
176   }
177 
178   /// Test if this TBAANode represents a type for objects which are
179   /// not modified (by any means) in the context where this
180   /// AliasAnalysis is relevant.
181   bool isTypeImmutable() const {
182     if (Node->getNumOperands() < 3)
183       return false;
184     ConstantInt *CI = mdconst::dyn_extract<ConstantInt>(Node->getOperand(2));
185     if (!CI)
186       return false;
187     return CI->getValue()[0];
188   }
189 };
190 
191 /// \name Specializations of \c TBAANodeImpl for const and non const qualified
192 /// \c MDNode.
193 /// @{
194 using TBAANode = TBAANodeImpl<const MDNode>;
195 using MutableTBAANode = TBAANodeImpl<MDNode>;
196 /// @}
197 
198 /// This is a simple wrapper around an MDNode which provides a
199 /// higher-level interface by hiding the details of how alias analysis
200 /// information is encoded in its operands.
201 template<typename MDNodeTy>
202 class TBAAStructTagNodeImpl {
203   /// This node should be created with createTBAAAccessTag().
204   MDNodeTy *Node;
205 
206 public:
207   explicit TBAAStructTagNodeImpl(MDNodeTy *N) : Node(N) {}
208 
209   /// Get the MDNode for this TBAAStructTagNode.
210   MDNodeTy *getNode() const { return Node; }
211 
212   /// isNewFormat - Return true iff the wrapped access tag is in the new
213   /// size-aware format.
214   bool isNewFormat() const {
215     if (Node->getNumOperands() < 4)
216       return false;
217     if (MDNodeTy *AccessType = getAccessType())
218       if (!TBAANodeImpl<MDNodeTy>(AccessType).isNewFormat())
219         return false;
220     return true;
221   }
222 
223   MDNodeTy *getBaseType() const {
224     return dyn_cast_or_null<MDNode>(Node->getOperand(0));
225   }
226 
227   MDNodeTy *getAccessType() const {
228     return dyn_cast_or_null<MDNode>(Node->getOperand(1));
229   }
230 
231   uint64_t getOffset() const {
232     return mdconst::extract<ConstantInt>(Node->getOperand(2))->getZExtValue();
233   }
234 
235   uint64_t getSize() const {
236     if (!isNewFormat())
237       return UINT64_MAX;
238     return mdconst::extract<ConstantInt>(Node->getOperand(3))->getZExtValue();
239   }
240 
241   /// Test if this TBAAStructTagNode represents a type for objects
242   /// which are not modified (by any means) in the context where this
243   /// AliasAnalysis is relevant.
244   bool isTypeImmutable() const {
245     unsigned OpNo = isNewFormat() ? 4 : 3;
246     if (Node->getNumOperands() < OpNo + 1)
247       return false;
248     ConstantInt *CI = mdconst::dyn_extract<ConstantInt>(Node->getOperand(OpNo));
249     if (!CI)
250       return false;
251     return CI->getValue()[0];
252   }
253 };
254 
255 /// \name Specializations of \c TBAAStructTagNodeImpl for const and non const
256 /// qualified \c MDNods.
257 /// @{
258 using TBAAStructTagNode = TBAAStructTagNodeImpl<const MDNode>;
259 using MutableTBAAStructTagNode = TBAAStructTagNodeImpl<MDNode>;
260 /// @}
261 
262 /// This is a simple wrapper around an MDNode which provides a
263 /// higher-level interface by hiding the details of how alias analysis
264 /// information is encoded in its operands.
265 class TBAAStructTypeNode {
266   /// This node should be created with createTBAATypeNode().
267   const MDNode *Node = nullptr;
268 
269 public:
270   TBAAStructTypeNode() = default;
271   explicit TBAAStructTypeNode(const MDNode *N) : Node(N) {}
272 
273   /// Get the MDNode for this TBAAStructTypeNode.
274   const MDNode *getNode() const { return Node; }
275 
276   /// isNewFormat - Return true iff the wrapped type node is in the new
277   /// size-aware format.
278   bool isNewFormat() const { return isNewFormatTypeNode(Node); }
279 
280   bool operator==(const TBAAStructTypeNode &Other) const {
281     return getNode() == Other.getNode();
282   }
283 
284   /// getId - Return type identifier.
285   Metadata *getId() const {
286     return Node->getOperand(isNewFormat() ? 2 : 0);
287   }
288 
289   unsigned getNumFields() const {
290     unsigned FirstFieldOpNo = isNewFormat() ? 3 : 1;
291     unsigned NumOpsPerField = isNewFormat() ? 3 : 2;
292     return (getNode()->getNumOperands() - FirstFieldOpNo) / NumOpsPerField;
293   }
294 
295   TBAAStructTypeNode getFieldType(unsigned FieldIndex) const {
296     unsigned FirstFieldOpNo = isNewFormat() ? 3 : 1;
297     unsigned NumOpsPerField = isNewFormat() ? 3 : 2;
298     unsigned OpIndex = FirstFieldOpNo + FieldIndex * NumOpsPerField;
299     auto *TypeNode = cast<MDNode>(getNode()->getOperand(OpIndex));
300     return TBAAStructTypeNode(TypeNode);
301   }
302 
303   /// Get this TBAAStructTypeNode's field in the type DAG with
304   /// given offset. Update the offset to be relative to the field type.
305   TBAAStructTypeNode getField(uint64_t &Offset) const {
306     bool NewFormat = isNewFormat();
307     if (NewFormat) {
308       // New-format root and scalar type nodes have no fields.
309       if (Node->getNumOperands() < 6)
310         return TBAAStructTypeNode();
311     } else {
312       // Parent can be omitted for the root node.
313       if (Node->getNumOperands() < 2)
314         return TBAAStructTypeNode();
315 
316       // Fast path for a scalar type node and a struct type node with a single
317       // field.
318       if (Node->getNumOperands() <= 3) {
319         uint64_t Cur = Node->getNumOperands() == 2
320                            ? 0
321                            : mdconst::extract<ConstantInt>(Node->getOperand(2))
322                                  ->getZExtValue();
323         Offset -= Cur;
324         MDNode *P = dyn_cast_or_null<MDNode>(Node->getOperand(1));
325         if (!P)
326           return TBAAStructTypeNode();
327         return TBAAStructTypeNode(P);
328       }
329     }
330 
331     // Assume the offsets are in order. We return the previous field if
332     // the current offset is bigger than the given offset.
333     unsigned FirstFieldOpNo = NewFormat ? 3 : 1;
334     unsigned NumOpsPerField = NewFormat ? 3 : 2;
335     unsigned TheIdx = 0;
336     for (unsigned Idx = FirstFieldOpNo; Idx < Node->getNumOperands();
337          Idx += NumOpsPerField) {
338       uint64_t Cur = mdconst::extract<ConstantInt>(Node->getOperand(Idx + 1))
339                          ->getZExtValue();
340       if (Cur > Offset) {
341         assert(Idx >= FirstFieldOpNo + NumOpsPerField &&
342                "TBAAStructTypeNode::getField should have an offset match!");
343         TheIdx = Idx - NumOpsPerField;
344         break;
345       }
346     }
347     // Move along the last field.
348     if (TheIdx == 0)
349       TheIdx = Node->getNumOperands() - NumOpsPerField;
350     uint64_t Cur = mdconst::extract<ConstantInt>(Node->getOperand(TheIdx + 1))
351                        ->getZExtValue();
352     Offset -= Cur;
353     MDNode *P = dyn_cast_or_null<MDNode>(Node->getOperand(TheIdx));
354     if (!P)
355       return TBAAStructTypeNode();
356     return TBAAStructTypeNode(P);
357   }
358 };
359 
360 } // end anonymous namespace
361 
362 /// Check the first operand of the tbaa tag node, if it is a MDNode, we treat
363 /// it as struct-path aware TBAA format, otherwise, we treat it as scalar TBAA
364 /// format.
365 static bool isStructPathTBAA(const MDNode *MD) {
366   // Anonymous TBAA root starts with a MDNode and dragonegg uses it as
367   // a TBAA tag.
368   return isa<MDNode>(MD->getOperand(0)) && MD->getNumOperands() >= 3;
369 }
370 
371 AliasResult TypeBasedAAResult::alias(const MemoryLocation &LocA,
372                                      const MemoryLocation &LocB,
373                                      AAQueryInfo &AAQI) {
374   if (!EnableTBAA)
375     return AAResultBase::alias(LocA, LocB, AAQI);
376 
377   // If accesses may alias, chain to the next AliasAnalysis.
378   if (Aliases(LocA.AATags.TBAA, LocB.AATags.TBAA))
379     return AAResultBase::alias(LocA, LocB, AAQI);
380 
381   // Otherwise return a definitive result.
382   return AliasResult::NoAlias;
383 }
384 
385 bool TypeBasedAAResult::pointsToConstantMemory(const MemoryLocation &Loc,
386                                                AAQueryInfo &AAQI,
387                                                bool OrLocal) {
388   if (!EnableTBAA)
389     return AAResultBase::pointsToConstantMemory(Loc, AAQI, OrLocal);
390 
391   const MDNode *M = Loc.AATags.TBAA;
392   if (!M)
393     return AAResultBase::pointsToConstantMemory(Loc, AAQI, OrLocal);
394 
395   // If this is an "immutable" type, we can assume the pointer is pointing
396   // to constant memory.
397   if ((!isStructPathTBAA(M) && TBAANode(M).isTypeImmutable()) ||
398       (isStructPathTBAA(M) && TBAAStructTagNode(M).isTypeImmutable()))
399     return true;
400 
401   return AAResultBase::pointsToConstantMemory(Loc, AAQI, OrLocal);
402 }
403 
404 FunctionModRefBehavior
405 TypeBasedAAResult::getModRefBehavior(const CallBase *Call) {
406   if (!EnableTBAA)
407     return AAResultBase::getModRefBehavior(Call);
408 
409   FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior;
410 
411   // If this is an "immutable" type, we can assume the call doesn't write
412   // to memory.
413   if (const MDNode *M = Call->getMetadata(LLVMContext::MD_tbaa))
414     if ((!isStructPathTBAA(M) && TBAANode(M).isTypeImmutable()) ||
415         (isStructPathTBAA(M) && TBAAStructTagNode(M).isTypeImmutable()))
416       Min = FMRB_OnlyReadsMemory;
417 
418   return FunctionModRefBehavior(AAResultBase::getModRefBehavior(Call) & Min);
419 }
420 
421 FunctionModRefBehavior TypeBasedAAResult::getModRefBehavior(const Function *F) {
422   // Functions don't have metadata. Just chain to the next implementation.
423   return AAResultBase::getModRefBehavior(F);
424 }
425 
426 ModRefInfo TypeBasedAAResult::getModRefInfo(const CallBase *Call,
427                                             const MemoryLocation &Loc,
428                                             AAQueryInfo &AAQI) {
429   if (!EnableTBAA)
430     return AAResultBase::getModRefInfo(Call, Loc, AAQI);
431 
432   if (const MDNode *L = Loc.AATags.TBAA)
433     if (const MDNode *M = Call->getMetadata(LLVMContext::MD_tbaa))
434       if (!Aliases(L, M))
435         return ModRefInfo::NoModRef;
436 
437   return AAResultBase::getModRefInfo(Call, Loc, AAQI);
438 }
439 
440 ModRefInfo TypeBasedAAResult::getModRefInfo(const CallBase *Call1,
441                                             const CallBase *Call2,
442                                             AAQueryInfo &AAQI) {
443   if (!EnableTBAA)
444     return AAResultBase::getModRefInfo(Call1, Call2, AAQI);
445 
446   if (const MDNode *M1 = Call1->getMetadata(LLVMContext::MD_tbaa))
447     if (const MDNode *M2 = Call2->getMetadata(LLVMContext::MD_tbaa))
448       if (!Aliases(M1, M2))
449         return ModRefInfo::NoModRef;
450 
451   return AAResultBase::getModRefInfo(Call1, Call2, AAQI);
452 }
453 
454 bool MDNode::isTBAAVtableAccess() const {
455   if (!isStructPathTBAA(this)) {
456     if (getNumOperands() < 1)
457       return false;
458     if (MDString *Tag1 = dyn_cast<MDString>(getOperand(0))) {
459       if (Tag1->getString() == "vtable pointer")
460         return true;
461     }
462     return false;
463   }
464 
465   // For struct-path aware TBAA, we use the access type of the tag.
466   TBAAStructTagNode Tag(this);
467   TBAAStructTypeNode AccessType(Tag.getAccessType());
468   if(auto *Id = dyn_cast<MDString>(AccessType.getId()))
469     if (Id->getString() == "vtable pointer")
470       return true;
471   return false;
472 }
473 
474 static bool matchAccessTags(const MDNode *A, const MDNode *B,
475                             const MDNode **GenericTag = nullptr);
476 
477 MDNode *MDNode::getMostGenericTBAA(MDNode *A, MDNode *B) {
478   const MDNode *GenericTag;
479   matchAccessTags(A, B, &GenericTag);
480   return const_cast<MDNode*>(GenericTag);
481 }
482 
483 static const MDNode *getLeastCommonType(const MDNode *A, const MDNode *B) {
484   if (!A || !B)
485     return nullptr;
486 
487   if (A == B)
488     return A;
489 
490   SmallSetVector<const MDNode *, 4> PathA;
491   TBAANode TA(A);
492   while (TA.getNode()) {
493     if (PathA.count(TA.getNode()))
494       report_fatal_error("Cycle found in TBAA metadata.");
495     PathA.insert(TA.getNode());
496     TA = TA.getParent();
497   }
498 
499   SmallSetVector<const MDNode *, 4> PathB;
500   TBAANode TB(B);
501   while (TB.getNode()) {
502     if (PathB.count(TB.getNode()))
503       report_fatal_error("Cycle found in TBAA metadata.");
504     PathB.insert(TB.getNode());
505     TB = TB.getParent();
506   }
507 
508   int IA = PathA.size() - 1;
509   int IB = PathB.size() - 1;
510 
511   const MDNode *Ret = nullptr;
512   while (IA >= 0 && IB >= 0) {
513     if (PathA[IA] == PathB[IB])
514       Ret = PathA[IA];
515     else
516       break;
517     --IA;
518     --IB;
519   }
520 
521   return Ret;
522 }
523 
524 AAMDNodes AAMDNodes::merge(const AAMDNodes &Other) const {
525   AAMDNodes Result;
526   Result.TBAA = MDNode::getMostGenericTBAA(TBAA, Other.TBAA);
527   Result.TBAAStruct = nullptr;
528   Result.Scope = MDNode::getMostGenericAliasScope(Scope, Other.Scope);
529   Result.NoAlias = MDNode::intersect(NoAlias, Other.NoAlias);
530   return Result;
531 }
532 
533 AAMDNodes Instruction::getAAMetadata() const {
534   AAMDNodes Result;
535   Result.TBAA = getMetadata(LLVMContext::MD_tbaa);
536   Result.TBAAStruct = getMetadata(LLVMContext::MD_tbaa_struct);
537   Result.Scope = getMetadata(LLVMContext::MD_alias_scope);
538   Result.NoAlias = getMetadata(LLVMContext::MD_noalias);
539   return Result;
540 }
541 
542 static const MDNode *createAccessTag(const MDNode *AccessType) {
543   // If there is no access type or the access type is the root node, then
544   // we don't have any useful access tag to return.
545   if (!AccessType || AccessType->getNumOperands() < 2)
546     return nullptr;
547 
548   Type *Int64 = IntegerType::get(AccessType->getContext(), 64);
549   auto *OffsetNode = ConstantAsMetadata::get(ConstantInt::get(Int64, 0));
550 
551   if (TBAAStructTypeNode(AccessType).isNewFormat()) {
552     // TODO: Take access ranges into account when matching access tags and
553     // fix this code to generate actual access sizes for generic tags.
554     uint64_t AccessSize = UINT64_MAX;
555     auto *SizeNode =
556         ConstantAsMetadata::get(ConstantInt::get(Int64, AccessSize));
557     Metadata *Ops[] = {const_cast<MDNode*>(AccessType),
558                        const_cast<MDNode*>(AccessType),
559                        OffsetNode, SizeNode};
560     return MDNode::get(AccessType->getContext(), Ops);
561   }
562 
563   Metadata *Ops[] = {const_cast<MDNode*>(AccessType),
564                      const_cast<MDNode*>(AccessType),
565                      OffsetNode};
566   return MDNode::get(AccessType->getContext(), Ops);
567 }
568 
569 static bool hasField(TBAAStructTypeNode BaseType,
570                      TBAAStructTypeNode FieldType) {
571   for (unsigned I = 0, E = BaseType.getNumFields(); I != E; ++I) {
572     TBAAStructTypeNode T = BaseType.getFieldType(I);
573     if (T == FieldType || hasField(T, FieldType))
574       return true;
575   }
576   return false;
577 }
578 
579 /// Return true if for two given accesses, one of the accessed objects may be a
580 /// subobject of the other. The \p BaseTag and \p SubobjectTag parameters
581 /// describe the accesses to the base object and the subobject respectively.
582 /// \p CommonType must be the metadata node describing the common type of the
583 /// accessed objects. On return, \p MayAlias is set to true iff these accesses
584 /// may alias and \p Generic, if not null, points to the most generic access
585 /// tag for the given two.
586 static bool mayBeAccessToSubobjectOf(TBAAStructTagNode BaseTag,
587                                      TBAAStructTagNode SubobjectTag,
588                                      const MDNode *CommonType,
589                                      const MDNode **GenericTag,
590                                      bool &MayAlias) {
591   // If the base object is of the least common type, then this may be an access
592   // to its subobject.
593   if (BaseTag.getAccessType() == BaseTag.getBaseType() &&
594       BaseTag.getAccessType() == CommonType) {
595     if (GenericTag)
596       *GenericTag = createAccessTag(CommonType);
597     MayAlias = true;
598     return true;
599   }
600 
601   // If the access to the base object is through a field of the subobject's
602   // type, then this may be an access to that field. To check for that we start
603   // from the base type, follow the edge with the correct offset in the type DAG
604   // and adjust the offset until we reach the field type or until we reach the
605   // access type.
606   bool NewFormat = BaseTag.isNewFormat();
607   TBAAStructTypeNode BaseType(BaseTag.getBaseType());
608   uint64_t OffsetInBase = BaseTag.getOffset();
609 
610   for (;;) {
611     // In the old format there is no distinction between fields and parent
612     // types, so in this case we consider all nodes up to the root.
613     if (!BaseType.getNode()) {
614       assert(!NewFormat && "Did not see access type in access path!");
615       break;
616     }
617 
618     if (BaseType.getNode() == SubobjectTag.getBaseType()) {
619       bool SameMemberAccess = OffsetInBase == SubobjectTag.getOffset();
620       if (GenericTag) {
621         *GenericTag = SameMemberAccess ? SubobjectTag.getNode() :
622                                          createAccessTag(CommonType);
623       }
624       MayAlias = SameMemberAccess;
625       return true;
626     }
627 
628     // With new-format nodes we stop at the access type.
629     if (NewFormat && BaseType.getNode() == BaseTag.getAccessType())
630       break;
631 
632     // Follow the edge with the correct offset. Offset will be adjusted to
633     // be relative to the field type.
634     BaseType = BaseType.getField(OffsetInBase);
635   }
636 
637   // If the base object has a direct or indirect field of the subobject's type,
638   // then this may be an access to that field. We need this to check now that
639   // we support aggregates as access types.
640   if (NewFormat) {
641     // TBAAStructTypeNode BaseAccessType(BaseTag.getAccessType());
642     TBAAStructTypeNode FieldType(SubobjectTag.getBaseType());
643     if (hasField(BaseType, FieldType)) {
644       if (GenericTag)
645         *GenericTag = createAccessTag(CommonType);
646       MayAlias = true;
647       return true;
648     }
649   }
650 
651   return false;
652 }
653 
654 /// matchTags - Return true if the given couple of accesses are allowed to
655 /// overlap. If \arg GenericTag is not null, then on return it points to the
656 /// most generic access descriptor for the given two.
657 static bool matchAccessTags(const MDNode *A, const MDNode *B,
658                             const MDNode **GenericTag) {
659   if (A == B) {
660     if (GenericTag)
661       *GenericTag = A;
662     return true;
663   }
664 
665   // Accesses with no TBAA information may alias with any other accesses.
666   if (!A || !B) {
667     if (GenericTag)
668       *GenericTag = nullptr;
669     return true;
670   }
671 
672   // Verify that both input nodes are struct-path aware.  Auto-upgrade should
673   // have taken care of this.
674   assert(isStructPathTBAA(A) && "Access A is not struct-path aware!");
675   assert(isStructPathTBAA(B) && "Access B is not struct-path aware!");
676 
677   TBAAStructTagNode TagA(A), TagB(B);
678   const MDNode *CommonType = getLeastCommonType(TagA.getAccessType(),
679                                                 TagB.getAccessType());
680 
681   // If the final access types have different roots, they're part of different
682   // potentially unrelated type systems, so we must be conservative.
683   if (!CommonType) {
684     if (GenericTag)
685       *GenericTag = nullptr;
686     return true;
687   }
688 
689   // If one of the accessed objects may be a subobject of the other, then such
690   // accesses may alias.
691   bool MayAlias;
692   if (mayBeAccessToSubobjectOf(/* BaseTag= */ TagA, /* SubobjectTag= */ TagB,
693                                CommonType, GenericTag, MayAlias) ||
694       mayBeAccessToSubobjectOf(/* BaseTag= */ TagB, /* SubobjectTag= */ TagA,
695                                CommonType, GenericTag, MayAlias))
696     return MayAlias;
697 
698   // Otherwise, we've proved there's no alias.
699   if (GenericTag)
700     *GenericTag = createAccessTag(CommonType);
701   return false;
702 }
703 
704 /// Aliases - Test whether the access represented by tag A may alias the
705 /// access represented by tag B.
706 bool TypeBasedAAResult::Aliases(const MDNode *A, const MDNode *B) const {
707   return matchAccessTags(A, B);
708 }
709 
710 AnalysisKey TypeBasedAA::Key;
711 
712 TypeBasedAAResult TypeBasedAA::run(Function &F, FunctionAnalysisManager &AM) {
713   return TypeBasedAAResult();
714 }
715 
716 char TypeBasedAAWrapperPass::ID = 0;
717 INITIALIZE_PASS(TypeBasedAAWrapperPass, "tbaa", "Type-Based Alias Analysis",
718                 false, true)
719 
720 ImmutablePass *llvm::createTypeBasedAAWrapperPass() {
721   return new TypeBasedAAWrapperPass();
722 }
723 
724 TypeBasedAAWrapperPass::TypeBasedAAWrapperPass() : ImmutablePass(ID) {
725   initializeTypeBasedAAWrapperPassPass(*PassRegistry::getPassRegistry());
726 }
727 
728 bool TypeBasedAAWrapperPass::doInitialization(Module &M) {
729   Result.reset(new TypeBasedAAResult());
730   return false;
731 }
732 
733 bool TypeBasedAAWrapperPass::doFinalization(Module &M) {
734   Result.reset();
735   return false;
736 }
737 
738 void TypeBasedAAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
739   AU.setPreservesAll();
740 }
741 
742 MDNode *AAMDNodes::shiftTBAA(MDNode *MD, size_t Offset) {
743   // Fast path if there's no offset
744   if (Offset == 0)
745     return MD;
746   // Fast path if there's no path tbaa node (and thus scalar)
747   if (!isStructPathTBAA(MD))
748     return MD;
749 
750   // The correct behavior here is to add the offset into the TBAA
751   // struct node offset. The base type, however may not have defined
752   // a type at this additional offset, resulting in errors. Since
753   // this method is only used within a given load/store access
754   // the offset provided is only used to subdivide the previous load
755   // maintaining the validity of the previous TBAA.
756   //
757   // This, however, should be revisited in the future.
758   return MD;
759 }
760 
761 MDNode *AAMDNodes::shiftTBAAStruct(MDNode *MD, size_t Offset) {
762   // Fast path if there's no offset
763   if (Offset == 0)
764     return MD;
765   SmallVector<Metadata *, 3> Sub;
766   for (size_t i = 0, size = MD->getNumOperands(); i < size; i += 3) {
767     ConstantInt *InnerOffset = mdconst::extract<ConstantInt>(MD->getOperand(i));
768     ConstantInt *InnerSize =
769         mdconst::extract<ConstantInt>(MD->getOperand(i + 1));
770     // Don't include any triples that aren't in bounds
771     if (InnerOffset->getZExtValue() + InnerSize->getZExtValue() <= Offset)
772       continue;
773 
774     uint64_t NewSize = InnerSize->getZExtValue();
775     uint64_t NewOffset = InnerOffset->getZExtValue() - Offset;
776     if (InnerOffset->getZExtValue() < Offset) {
777       NewOffset = 0;
778       NewSize -= Offset - InnerOffset->getZExtValue();
779     }
780 
781     // Shift the offset of the triple
782     Sub.push_back(ConstantAsMetadata::get(
783         ConstantInt::get(InnerOffset->getType(), NewOffset)));
784     Sub.push_back(ConstantAsMetadata::get(
785         ConstantInt::get(InnerSize->getType(), NewSize)));
786     Sub.push_back(MD->getOperand(i + 2));
787   }
788   return MDNode::get(MD->getContext(), Sub);
789 }
790