[tree-ssa] Reorganize expansion of nested functions

[Jan Hubicka]

> > On Sun, Feb 01, 2004 at 03:25:53AM +0100, Jan Hubicka wrote:
> > > + /* PARAM is function parameter referenced by nested function, prepare to
> > > +    produce correct code in that case.  */
> > > + void
> > > + notice_parameter_referenced_by_nested_function (tree param ATTRIBUTE_UNUSED)
> > > + {
> > > + #ifdef NEED_SEPARATE_AP
> > > +   get_arg_pointer_save_area (cfun);
> > > + #endif
> > > + }
> > 
> > Why is this needed?  Indeed, I would expect basically all of this:
> 
> It is needed to make it possible to access the argument by nested
> functions.  There is corresponding code in function.c
> 
>       /* If reached via arg pointer, get the arg pointer value
> 	 out of that function's stack frame.
> 
> 	 There are two cases:  If a separate ap is needed, allocate a
> 	 slot in the outer function for it and dereference it that way.
> 	 This is correct even if the real ap is actually a pseudo.
> 	 Otherwise, just adjust the offset from the frame pointer to
> 	 compensate.  */
> 
> #ifdef NEED_SEPARATE_AP
>       rtx addr;
> 
>       addr = get_arg_pointer_save_area (fp);
>       addr = fix_lexical_addr (XEXP (addr, 0), var);
>       addr = memory_address (Pmode, addr);
> 
>       base = gen_rtx_MEM (Pmode, addr);
>       set_mem_alias_set (base, get_frame_alias_set ());
>       base = copy_to_reg (base);
> #else
>       displacement += (FIRST_PARM_OFFSET (context) - STARTING_FRAME_OFFSET);
>       base = lookup_static_chain (var);
> #endif

To be more specific (now when I am not in hurry), the problems I am
fighting with is that during expansion of nested functions, the need for
trampolines/nonlocal goto receivers and arg_pointer_save_area is
discovered and since the outer function is expanded later, they can be
taken into account.

Because I now expand it before, i have to discover these facts in
advance so the later expansion of nested functions won't mess things up.

Honza
> > 
> > > ! examine_nested_functions_1 (tree * tp, int *walk_subtrees, void *data)
> > 
> > to be handled by rtl expansion of the parent function.
> 
> It is done by rtl expansion of nested function that hooks it into parent
> function if I understand it right.  But I can move the walk around into
> the expansion somewhere.
> > 
> > >     expand_function_end ();
> > > + 
> > > +   expand_nested_functions ();
> > 
> > As far as I can see, moving expansion of nested functions here
> > from expand_used_vars is the only substantive change.
> 
> Yes, the other stuff is needed to make it work :)
> 
> Honza
> > 
> > 
> > r~