Re: DWARF question/inlining/local classes

[Geoff Keating <geoffk at geoffk dot org>]

Mark Mitchell <mark@codesourcery.com> writes:

> PR 8391 contains this code:
> 
>    inline void h(const Outer &o)
>    {
>      struct Local : public Outer::Inner {
> 	 virtual bool f() {};
>      };
>      Local l;
>      o.g(l);
>    }
> 
> When we try to emit the body of Local::f (the function in the struct
> inside h), we don't actually emit assembly code because of this code
> in rest_of_compilation:
> 
>       /* If this is nested inside an inlined external function, pretend
> 	 it was only declared.  Since we cannot inline such functions,
> 	 generating code for this one is not only not necessary but will
> 	 confuse some debugging output writers.  */
>       for (parent = DECL_CONTEXT (current_function_decl);
> 	   parent != NULL_TREE;
> 	   parent = get_containing_scope (parent))
> 	if (TREE_CODE (parent) == FUNCTION_DECL
> 	    && DECL_INLINE (parent) && DECL_EXTERNAL (parent))
> 	  {
> 	    DECL_INITIAL (decl) = 0;
> 	    goto exit_rest_of_compilation;
> 	  }
>  
> This logic is wrong for C++; we need to emit Local::f.

I believe the rationale behind this code is that Local::f can't be
used from this translation unit, since the function won't be inlined.
It will be emitted when the out-of-line copy of the function is emitted.

Will this function really not be inlined?  That may be the real problem.

-- 
- Geoffrey Keating <geoffk@geoffk.org>