RecursiveASTVisitor: VisitTypeOfExprType doesn't get called

[llvmbot]

Bugzilla Link	13618
Version	trunk
OS	Windows NT
Reporter	LLVM Bugzilla Contributor
CC	@zygoloid

Extended Description

I wrote this simple recursive AST visitor:

class MyVisitor : public clang::RecursiveASTVisitor { public: MyVisitor() {}

bool VisitTypeOfExprType(clang::TypeOfExprType* t)
{
    printf("Visiting typeof expr type!\n");
    return true;
}

};

And ran it on AST of the following code (found in one open-source project):

zbx.c:

void foo() { int status = 0;

status = ((((extension (((union { typeof(status) in; int i; }) { .in = (status) }).__i))) & 0xff00) >> 8);

status = ({ union { typeof(status) in; int i; } a; a.in = status; (a.__i & 0xff00) >> 8; }); }

I got only one "Visiting typeof expr type!" as a result, for the 2nd anonymous union. I cannot find a way how to catch __typeof of the first anonymous union with RAV.

[ec04fc15-fa35-46f2-80e1-5d271f2ef708]

The CompoundLiteralExpr change would be straight forward to add to RAV.

I agree, fixed in r162133.

I'm not sure about the RecordType change though.

Right, that change is not correct (for instance, a named type can be defined in a compound literal in C). I think we want two changes here:

• Avoid repeatedly visiting the decl-specifier-seq when visiting a DeclStmt, and
• Visit the TagDecl when we visit a TagTypeLoc where isDefinition() returns true.

[llvmbot]

Thanks, this workaround seems to work, all tests from different projects pass now!

Sergejs

[llvmbot]

Here's a version of your visitor that can traverse the anonymous union:

class MyVisitor : public clang::RecursiveASTVisitor { typedef RecursiveASTVisitor VisitorBase; public: MyVisitor() {}

bool VisitTypeOfExprType(clang::TypeOfExprType* t)
{
    printf("Visiting typeof expr type!\n");
    return true;
}

bool TraverseCompoundLiteralExpr(CompoundLiteralExpr *E) {
    VisitorBase::TraverseCompoundLiteralExpr(E);
    TraverseTypeLoc(E->getTypeSourceInfo()->getTypeLoc());
    return true;
}

bool VisitRecordType(RecordType *T) {
    RecordDecl *D = T->getDecl();
    if (!D->getIdentifier())
      TraverseDecl(D);
    return true;
}

};

The CompoundLiteralExpr change would be straight forward to add to RAV.

I'm not sure about the RecordType change though. There's two problems.

First, I don't know whether getIdentifier() is the correct test. I expected isAnonymousStructOrUnion() to be a better test but it didn't work.

Second, this changes results in the 2nd anonymous union being traversed twice. This is because the declaration of 'a' is parsed into a DeclStmt containing both a RecordDecl and a VarDecl (and there is a RecordType as part of the VarDecl), whereas I would have expected only the VarDecl.