1 //===- TypeBasedAliasAnalysis.cpp - Type-Based Alias Analysis -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the TypeBasedAliasAnalysis pass, which implements
11 // metadata-based TBAA.
13 // In LLVM IR, memory does not have types, so LLVM's own type system is not
14 // suitable for doing TBAA. Instead, metadata is added to the IR to describe
15 // a type system of a higher level language. This can be used to implement
16 // typical C/C++ TBAA, but it can also be used to implement custom alias
17 // analysis behavior for other languages.
19 // We now support two types of metadata format: scalar TBAA and struct-path
20 // aware TBAA. After all testing cases are upgraded to use struct-path aware
21 // TBAA and we can auto-upgrade existing bc files, the support for scalar TBAA
24 // The scalar TBAA metadata format is very simple. TBAA MDNodes have up to
25 // three fields, e.g.:
26 // !0 = metadata !{ metadata !"an example type tree" }
27 // !1 = metadata !{ metadata !"int", metadata !0 }
28 // !2 = metadata !{ metadata !"float", metadata !0 }
29 // !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
31 // The first field is an identity field. It can be any value, usually
32 // an MDString, which uniquely identifies the type. The most important
33 // name in the tree is the name of the root node. Two trees with
34 // different root node names are entirely disjoint, even if they
35 // have leaves with common names.
37 // The second field identifies the type's parent node in the tree, or
38 // is null or omitted for a root node. A type is considered to alias
39 // all of its descendants and all of its ancestors in the tree. Also,
40 // a type is considered to alias all types in other trees, so that
41 // bitcode produced from multiple front-ends is handled conservatively.
43 // If the third field is present, it's an integer which if equal to 1
44 // indicates that the type is "constant" (meaning pointsToConstantMemory
45 // should return true; see
46 // http://llvm.org/docs/AliasAnalysis.html#OtherItfs).
48 // With struct-path aware TBAA, the MDNodes attached to an instruction using
49 // "!tbaa" are called path tag nodes.
51 // The path tag node has 4 fields with the last field being optional.
53 // The first field is the base type node, it can be a struct type node
54 // or a scalar type node. The second field is the access type node, it
55 // must be a scalar type node. The third field is the offset into the base type.
56 // The last field has the same meaning as the last field of our scalar TBAA:
57 // it's an integer which if equal to 1 indicates that the access is "constant".
59 // The struct type node has a name and a list of pairs, one pair for each member
60 // of the struct. The first element of each pair is a type node (a struct type
61 // node or a sclar type node), specifying the type of the member, the second
62 // element of each pair is the offset of the member.
73 // For an acess to B.a.s, we attach !5 (a path tag node) to the load/store
74 // instruction. The base type is !4 (struct B), the access type is !2 (scalar
75 // type short) and the offset is 4.
77 // !0 = metadata !{metadata !"Simple C/C++ TBAA"}
78 // !1 = metadata !{metadata !"omnipotent char", metadata !0} // Scalar type node
79 // !2 = metadata !{metadata !"short", metadata !1} // Scalar type node
80 // !3 = metadata !{metadata !"A", metadata !2, i64 0} // Struct type node
81 // !4 = metadata !{metadata !"B", metadata !2, i64 0, metadata !3, i64 4}
82 // // Struct type node
83 // !5 = metadata !{metadata !4, metadata !2, i64 4} // Path tag node
85 // The struct type nodes and the scalar type nodes form a type DAG.
87 // char (!1) -- edge to Root
88 // short (!2) -- edge to char
89 // A (!3) -- edge with offset 0 to short
90 // B (!4) -- edge with offset 0 to short and edge with offset 4 to A
92 // To check if two tags (tagX and tagY) can alias, we start from the base type
93 // of tagX, follow the edge with the correct offset in the type DAG and adjust
94 // the offset until we reach the base type of tagY or until we reach the Root
96 // If we reach the base type of tagY, compare the adjusted offset with
97 // offset of tagY, return Alias if the offsets are the same, return NoAlias
99 // If we reach the Root node, perform the above starting from base type of tagY
100 // to see if we reach base type of tagX.
102 // If they have different roots, they're part of different potentially
103 // unrelated type systems, so we return Alias to be conservative.
104 // If neither node is an ancestor of the other and they have the same root,
105 // then we say NoAlias.
107 // TODO: The current metadata format doesn't support struct
108 // fields. For example:
113 // void foo(struct X *x, struct X *y, double *p) {
117 // Struct X has a double member, so the store to *x can alias the store to *p.
118 // Currently it's not possible to precisely describe all the things struct X
119 // aliases, so struct assignments must use conservative TBAA nodes. There's
120 // no scheme for attaching metadata to @llvm.memcpy yet either.
122 //===----------------------------------------------------------------------===//
124 #include "llvm/Analysis/Passes.h"
125 #include "llvm/Analysis/AliasAnalysis.h"
126 #include "llvm/IR/Constants.h"
127 #include "llvm/IR/LLVMContext.h"
128 #include "llvm/IR/Metadata.h"
129 #include "llvm/IR/Module.h"
130 #include "llvm/Pass.h"
131 #include "llvm/Support/CommandLine.h"
132 using namespace llvm;
134 // A handy option for disabling TBAA functionality. The same effect can also be
135 // achieved by stripping the !tbaa tags from IR, but this option is sometimes
137 static cl::opt<bool> EnableTBAA("enable-tbaa", cl::init(true));
140 /// TBAANode - This is a simple wrapper around an MDNode which provides a
141 /// higher-level interface by hiding the details of how alias analysis
142 /// information is encoded in its operands.
147 TBAANode() : Node(nullptr) {}
148 explicit TBAANode(const MDNode *N) : Node(N) {}
150 /// getNode - Get the MDNode for this TBAANode.
151 const MDNode *getNode() const { return Node; }
153 /// getParent - Get this TBAANode's Alias tree parent.
154 TBAANode getParent() const {
155 if (Node->getNumOperands() < 2)
157 MDNode *P = dyn_cast_or_null<MDNode>(Node->getOperand(1));
160 // Ok, this node has a valid parent. Return it.
164 /// TypeIsImmutable - Test if this TBAANode represents a type for objects
165 /// which are not modified (by any means) in the context where this
166 /// AliasAnalysis is relevant.
167 bool TypeIsImmutable() const {
168 if (Node->getNumOperands() < 3)
170 ConstantInt *CI = dyn_cast<ConstantInt>(Node->getOperand(2));
173 return CI->getValue()[0];
177 /// This is a simple wrapper around an MDNode which provides a
178 /// higher-level interface by hiding the details of how alias analysis
179 /// information is encoded in its operands.
180 class TBAAStructTagNode {
181 /// This node should be created with createTBAAStructTagNode.
185 explicit TBAAStructTagNode(const MDNode *N) : Node(N) {}
187 /// Get the MDNode for this TBAAStructTagNode.
188 const MDNode *getNode() const { return Node; }
190 const MDNode *getBaseType() const {
191 return dyn_cast_or_null<MDNode>(Node->getOperand(0));
193 const MDNode *getAccessType() const {
194 return dyn_cast_or_null<MDNode>(Node->getOperand(1));
196 uint64_t getOffset() const {
197 return cast<ConstantInt>(Node->getOperand(2))->getZExtValue();
199 /// TypeIsImmutable - Test if this TBAAStructTagNode represents a type for
200 /// objects which are not modified (by any means) in the context where this
201 /// AliasAnalysis is relevant.
202 bool TypeIsImmutable() const {
203 if (Node->getNumOperands() < 4)
205 ConstantInt *CI = dyn_cast<ConstantInt>(Node->getOperand(3));
208 return CI->getValue()[0];
212 /// This is a simple wrapper around an MDNode which provides a
213 /// higher-level interface by hiding the details of how alias analysis
214 /// information is encoded in its operands.
215 class TBAAStructTypeNode {
216 /// This node should be created with createTBAAStructTypeNode.
220 TBAAStructTypeNode() : Node(nullptr) {}
221 explicit TBAAStructTypeNode(const MDNode *N) : Node(N) {}
223 /// Get the MDNode for this TBAAStructTypeNode.
224 const MDNode *getNode() const { return Node; }
226 /// Get this TBAAStructTypeNode's field in the type DAG with
227 /// given offset. Update the offset to be relative to the field type.
228 TBAAStructTypeNode getParent(uint64_t &Offset) const {
229 // Parent can be omitted for the root node.
230 if (Node->getNumOperands() < 2)
231 return TBAAStructTypeNode();
233 // Fast path for a scalar type node and a struct type node with a single
235 if (Node->getNumOperands() <= 3) {
236 uint64_t Cur = Node->getNumOperands() == 2 ? 0 :
237 cast<ConstantInt>(Node->getOperand(2))->getZExtValue();
239 MDNode *P = dyn_cast_or_null<MDNode>(Node->getOperand(1));
241 return TBAAStructTypeNode();
242 return TBAAStructTypeNode(P);
245 // Assume the offsets are in order. We return the previous field if
246 // the current offset is bigger than the given offset.
248 for (unsigned Idx = 1; Idx < Node->getNumOperands(); Idx += 2) {
249 uint64_t Cur = cast<ConstantInt>(Node->getOperand(Idx + 1))->
253 "TBAAStructTypeNode::getParent should have an offset match!");
258 // Move along the last field.
260 TheIdx = Node->getNumOperands() - 2;
261 uint64_t Cur = cast<ConstantInt>(Node->getOperand(TheIdx + 1))->
264 MDNode *P = dyn_cast_or_null<MDNode>(Node->getOperand(TheIdx));
266 return TBAAStructTypeNode();
267 return TBAAStructTypeNode(P);
273 /// TypeBasedAliasAnalysis - This is a simple alias analysis
274 /// implementation that uses TypeBased to answer queries.
275 class TypeBasedAliasAnalysis : public ImmutablePass,
276 public AliasAnalysis {
278 static char ID; // Class identification, replacement for typeinfo
279 TypeBasedAliasAnalysis() : ImmutablePass(ID) {
280 initializeTypeBasedAliasAnalysisPass(*PassRegistry::getPassRegistry());
283 void initializePass() override {
284 InitializeAliasAnalysis(this);
287 /// getAdjustedAnalysisPointer - This method is used when a pass implements
288 /// an analysis interface through multiple inheritance. If needed, it
289 /// should override this to adjust the this pointer as needed for the
290 /// specified pass info.
291 void *getAdjustedAnalysisPointer(const void *PI) override {
292 if (PI == &AliasAnalysis::ID)
293 return (AliasAnalysis*)this;
297 bool Aliases(const MDNode *A, const MDNode *B) const;
298 bool PathAliases(const MDNode *A, const MDNode *B) const;
301 void getAnalysisUsage(AnalysisUsage &AU) const override;
302 AliasResult alias(const Location &LocA, const Location &LocB) override;
303 bool pointsToConstantMemory(const Location &Loc, bool OrLocal) override;
304 ModRefBehavior getModRefBehavior(ImmutableCallSite CS) override;
305 ModRefBehavior getModRefBehavior(const Function *F) override;
306 ModRefResult getModRefInfo(ImmutableCallSite CS,
307 const Location &Loc) override;
308 ModRefResult getModRefInfo(ImmutableCallSite CS1,
309 ImmutableCallSite CS2) override;
311 } // End of anonymous namespace
313 // Register this pass...
314 char TypeBasedAliasAnalysis::ID = 0;
315 INITIALIZE_AG_PASS(TypeBasedAliasAnalysis, AliasAnalysis, "tbaa",
316 "Type-Based Alias Analysis", false, true, false)
318 ImmutablePass *llvm::createTypeBasedAliasAnalysisPass() {
319 return new TypeBasedAliasAnalysis();
323 TypeBasedAliasAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
324 AU.setPreservesAll();
325 AliasAnalysis::getAnalysisUsage(AU);
328 /// Check the first operand of the tbaa tag node, if it is a MDNode, we treat
329 /// it as struct-path aware TBAA format, otherwise, we treat it as scalar TBAA
331 static bool isStructPathTBAA(const MDNode *MD) {
332 // Anonymous TBAA root starts with a MDNode and dragonegg uses it as
334 return isa<MDNode>(MD->getOperand(0)) && MD->getNumOperands() >= 3;
337 /// Aliases - Test whether the type represented by A may alias the
338 /// type represented by B.
340 TypeBasedAliasAnalysis::Aliases(const MDNode *A,
341 const MDNode *B) const {
342 if (isStructPathTBAA(A))
343 return PathAliases(A, B);
345 // Keep track of the root node for A and B.
346 TBAANode RootA, RootB;
348 // Climb the tree from A to see if we reach B.
349 for (TBAANode T(A); ; ) {
350 if (T.getNode() == B)
351 // B is an ancestor of A.
360 // Climb the tree from B to see if we reach A.
361 for (TBAANode T(B); ; ) {
362 if (T.getNode() == A)
363 // A is an ancestor of B.
372 // Neither node is an ancestor of the other.
374 // If they have different roots, they're part of different potentially
375 // unrelated type systems, so we must be conservative.
376 if (RootA.getNode() != RootB.getNode())
379 // If they have the same root, then we've proved there's no alias.
383 /// Test whether the struct-path tag represented by A may alias the
384 /// struct-path tag represented by B.
386 TypeBasedAliasAnalysis::PathAliases(const MDNode *A,
387 const MDNode *B) const {
388 // Keep track of the root node for A and B.
389 TBAAStructTypeNode RootA, RootB;
390 TBAAStructTagNode TagA(A), TagB(B);
392 // TODO: We need to check if AccessType of TagA encloses AccessType of
393 // TagB to support aggregate AccessType. If yes, return true.
395 // Start from the base type of A, follow the edge with the correct offset in
396 // the type DAG and adjust the offset until we reach the base type of B or
397 // until we reach the Root node.
398 // Compare the adjusted offset once we have the same base.
400 // Climb the type DAG from base type of A to see if we reach base type of B.
401 const MDNode *BaseA = TagA.getBaseType();
402 const MDNode *BaseB = TagB.getBaseType();
403 uint64_t OffsetA = TagA.getOffset(), OffsetB = TagB.getOffset();
404 for (TBAAStructTypeNode T(BaseA); ; ) {
405 if (T.getNode() == BaseB)
406 // Base type of A encloses base type of B, check if the offsets match.
407 return OffsetA == OffsetB;
410 // Follow the edge with the correct offset, OffsetA will be adjusted to
411 // be relative to the field type.
412 T = T.getParent(OffsetA);
417 // Reset OffsetA and climb the type DAG from base type of B to see if we reach
419 OffsetA = TagA.getOffset();
420 for (TBAAStructTypeNode T(BaseB); ; ) {
421 if (T.getNode() == BaseA)
422 // Base type of B encloses base type of A, check if the offsets match.
423 return OffsetA == OffsetB;
426 // Follow the edge with the correct offset, OffsetB will be adjusted to
427 // be relative to the field type.
428 T = T.getParent(OffsetB);
433 // Neither node is an ancestor of the other.
435 // If they have different roots, they're part of different potentially
436 // unrelated type systems, so we must be conservative.
437 if (RootA.getNode() != RootB.getNode())
440 // If they have the same root, then we've proved there's no alias.
444 AliasAnalysis::AliasResult
445 TypeBasedAliasAnalysis::alias(const Location &LocA,
446 const Location &LocB) {
448 return AliasAnalysis::alias(LocA, LocB);
450 // Get the attached MDNodes. If either value lacks a tbaa MDNode, we must
452 const MDNode *AM = LocA.TBAATag;
453 if (!AM) return AliasAnalysis::alias(LocA, LocB);
454 const MDNode *BM = LocB.TBAATag;
455 if (!BM) return AliasAnalysis::alias(LocA, LocB);
457 // If they may alias, chain to the next AliasAnalysis.
459 return AliasAnalysis::alias(LocA, LocB);
461 // Otherwise return a definitive result.
465 bool TypeBasedAliasAnalysis::pointsToConstantMemory(const Location &Loc,
468 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
470 const MDNode *M = Loc.TBAATag;
471 if (!M) return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
473 // If this is an "immutable" type, we can assume the pointer is pointing
474 // to constant memory.
475 if ((!isStructPathTBAA(M) && TBAANode(M).TypeIsImmutable()) ||
476 (isStructPathTBAA(M) && TBAAStructTagNode(M).TypeIsImmutable()))
479 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal);
482 AliasAnalysis::ModRefBehavior
483 TypeBasedAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) {
485 return AliasAnalysis::getModRefBehavior(CS);
487 ModRefBehavior Min = UnknownModRefBehavior;
489 // If this is an "immutable" type, we can assume the call doesn't write
491 if (const MDNode *M = CS.getInstruction()->getMetadata(LLVMContext::MD_tbaa))
492 if ((!isStructPathTBAA(M) && TBAANode(M).TypeIsImmutable()) ||
493 (isStructPathTBAA(M) && TBAAStructTagNode(M).TypeIsImmutable()))
494 Min = OnlyReadsMemory;
496 return ModRefBehavior(AliasAnalysis::getModRefBehavior(CS) & Min);
499 AliasAnalysis::ModRefBehavior
500 TypeBasedAliasAnalysis::getModRefBehavior(const Function *F) {
501 // Functions don't have metadata. Just chain to the next implementation.
502 return AliasAnalysis::getModRefBehavior(F);
505 AliasAnalysis::ModRefResult
506 TypeBasedAliasAnalysis::getModRefInfo(ImmutableCallSite CS,
507 const Location &Loc) {
509 return AliasAnalysis::getModRefInfo(CS, Loc);
511 if (const MDNode *L = Loc.TBAATag)
512 if (const MDNode *M =
513 CS.getInstruction()->getMetadata(LLVMContext::MD_tbaa))
517 return AliasAnalysis::getModRefInfo(CS, Loc);
520 AliasAnalysis::ModRefResult
521 TypeBasedAliasAnalysis::getModRefInfo(ImmutableCallSite CS1,
522 ImmutableCallSite CS2) {
524 return AliasAnalysis::getModRefInfo(CS1, CS2);
526 if (const MDNode *M1 =
527 CS1.getInstruction()->getMetadata(LLVMContext::MD_tbaa))
528 if (const MDNode *M2 =
529 CS2.getInstruction()->getMetadata(LLVMContext::MD_tbaa))
530 if (!Aliases(M1, M2))
533 return AliasAnalysis::getModRefInfo(CS1, CS2);
536 bool MDNode::isTBAAVtableAccess() const {
537 if (!isStructPathTBAA(this)) {
538 if (getNumOperands() < 1) return false;
539 if (MDString *Tag1 = dyn_cast<MDString>(getOperand(0))) {
540 if (Tag1->getString() == "vtable pointer") return true;
545 // For struct-path aware TBAA, we use the access type of the tag.
546 if (getNumOperands() < 2) return false;
547 MDNode *Tag = cast_or_null<MDNode>(getOperand(1));
548 if (!Tag) return false;
549 if (MDString *Tag1 = dyn_cast<MDString>(Tag->getOperand(0))) {
550 if (Tag1->getString() == "vtable pointer") return true;
555 MDNode *MDNode::getMostGenericTBAA(MDNode *A, MDNode *B) {
562 // For struct-path aware TBAA, we use the access type of the tag.
563 bool StructPath = isStructPathTBAA(A);
565 A = cast_or_null<MDNode>(A->getOperand(1));
566 if (!A) return nullptr;
567 B = cast_or_null<MDNode>(B->getOperand(1));
568 if (!B) return nullptr;
571 SmallVector<MDNode *, 4> PathA;
575 T = T->getNumOperands() >= 2 ? cast_or_null<MDNode>(T->getOperand(1))
579 SmallVector<MDNode *, 4> PathB;
583 T = T->getNumOperands() >= 2 ? cast_or_null<MDNode>(T->getOperand(1))
587 int IA = PathA.size() - 1;
588 int IB = PathB.size() - 1;
590 MDNode *Ret = nullptr;
591 while (IA >= 0 && IB >=0) {
592 if (PathA[IA] == PathB[IB])
604 // We need to convert from a type node to a tag node.
605 Type *Int64 = IntegerType::get(A->getContext(), 64);
606 Value *Ops[3] = { Ret, Ret, ConstantInt::get(Int64, 0) };
607 return MDNode::get(A->getContext(), Ops);