[cfe-dev] Where to put exception spec source locations?

[Douglas Gregor]

On May 31, 2009, at 11:42 AM, Sebastian Redl wrote:
> Where in the AST would be a good home for exception specification  
> source
> locations? I store the thrown types in the Type hierarchy, but I
> obviously can't do that for source locations.
>
> If it was only functions, I could put them into the FunctionDecl:
>
> void fn() throw(int);
>
> However, what about this function pointer?
>
> void (*fn)() throw(int);
>
> It only has some kind of VarDecl. It would be very weird to extend
> VarDecl with source locations for an exception spec, and it wouldn't
> work anyway:
>
> void (*fn)(void (*)() throw (int)) throw(int);
>
> What we have here is a pointer to a function that could throw int, and
> takes as its only parameter a pointer to a function that could throw
> int. One VarDecl for the whole construct, and of course I could extend
> it to as many parameters as Clang supports (2^20-1).

Given these examples, I think that the only reasonable answer is to  
put some SourceLocations into the type system.

> I need such source locations to accurately point to the types that are
> mismatches. I also need them if I want to point to anything useful  
> when
> instantiating a type that has a template parameter in its throw list:
>
> template <typename T> struct X {
>  void fn() throw(T); // Obviously, this could just as well be a
> function pointer.
> };
> X<void> x; // Oops, cannot have throw(void) in function type. But  
> where
> to point?
>
> Any suggestions?


I've actually been planning to add some new sugary Type nodes into the  
AST that contain SourceLocations, precisely for use in template  
instantiation. Right now, we provide sub-par location information in  
diagnostics relating to template instantiation, because we have no  
SourceLocations in the types. Caveat #1: canonical types should never  
have source locations. Caveat #2: we don't want to bloat the Type  
nodes in the common case, so we should only add SourceLocations where  
we need them to produce good diagnostics, e.g., in dependent types  
(for template instantiation) and in function types with exception  
specifications (for the reasons Sebastian mentioned above).

I don't expect to get to adding the new, SourceLocation-improved  
dependent type nodes for a few weeks, though.

	- Doug