[cfe-dev] Cleaning up the representation of Decls in the AST

[Ted Kremenek]

Hi Argiris,

I think part of this dialogue stems from our different perspective of  
the ASTs.  In my mind, the ASTs should be designed to (1) faithfully  
represent the program and (2) make it easy for clients of the ASTs to  
understand their structure and purpose.  I'm personally not married to  
creating a one-to-one mapping between the terminology used in the C99  
standard to describe grammar and syntax and the ASTs themselves.  I  
*do* think, however, that faithfully following this terminology is  
important in the parser and its actions, but not the ASTs.

On Sep 12, 2008, at 4:01 AM, Argiris Kirtzidis wrote:

> To sum up my point here, it is more syntactically accurate to  
> consider RecordDecl a part of type-specifier.
> In the syntax it's the type-specifier of a syntactic construct that  
> tells us whether it's "struct s" or "struct s { int d; }".

The DeclGroup proposal (which I want to add that I'm not married to)  
handles this case.  RecordDecls indicate with a bit whether or not  
they are a forward declaration or a definition.

Note that the DeclGroup proposal can directly represent the following  
beast with complete fidelity:

struct { int x; } a, *b, **c;

The DeclGroup would point to the RecordDecl for the anonymous  
structure, as well as point the the three VarDecls for a, b, and c  
respectively.  The VarDecls would still have QualTypes that refer to  
the appropriate types.

>
> If we are not getting this kind of information, the type-specifier  
> is not represented correctly in the AST.
>
> -Argiris



One of the other design points of the ASTs is that we represent  
directly represent the information that most clients need, for the  
remaining 10% of the information that we don't directly represent we  
rely on being able to lazily recreate them, either by relexing or  
other means.  The problem with the current representation of  
declarations in the ASTs is that we lose critical source information  
as well as a clear idea of where types are defined, etc.  That's the  
motivation of DeclGroup (as well as your type-specifier proposal).

Ted