Re: [PATCH/RFC] Per-architecture DWARF CFI register state initialization hooks

[Andrew Cagney <cagney@gnu.org>]

>    Date: Sat, 07 Feb 2004 18:48:47 -0500
>    From: Andrew Cagney <cagney@gnu.org>
> 
>    > Here's my proposal for the per-architecture DWARF CFI register state
>    > initialization hooks needed for S/390, and others.  This is a RFC,
>    > since I'm not entirely confident whether my approach is acceptable.  I
>    > chose to implement this using per-architecture data instead of adding
>    > a function to the architecture vector.  I think it is cleaner since it
>    > keeps things localized and modular, and the architecture vector is big
>    > enough as it stands.
> 
>    Yes.  Technical nit though - I think it is still better to have a local 
>    data struct and store the value in there.
> 
> I'm not sure what your idea is here.  Is it that you want me to use a
> data structure that would be allocated by the dwarf2-frame.c module
> such that I'd only need a single per-arch data key for the entire
> dwarf2-frame.c module?

Yes, just this:

> static struct gdbarch_data *frame_base_data;
> 
> struct frame_base_table
> {
>   frame_base_sniffer_ftype **sniffer;
....
> };
> 
> static struct frame_base_table *
> frame_base_table (struct gdbarch *gdbarch)
> {
>   struct frame_base_table *table = gdbarch_data (gdbarch, frame_base_data);
>   ...
>   return table;
> }

i.e, put the function inside a struct, instead of storing the function 
directly in that per-arch data pointer.

It's more consistent with:
http://sources.redhat.com/gdb/current/onlinedocs/gdbint_13.html#SEC114
but hmm, that's out of date, sigh.  Instead of free, memory is obtained 
using gdbarch_obstack_zalloc.

 > Or do you want the architecture to allocate
> and initialize the structure?  The latter would mean more work for the
> architecture; if you want to override a single member of the structure
> you'd have to fill in all the details.  I don't really like that.
> 
>    From: Daniel Jacobowitz <drow@mvista.com>
>    Date: Sat, 7 Feb 2004 18:03:30 -0500
> 
>    Hmm, I do.  You're adding a per-architecture data item which is a
>    function pointer, and what amounts to the rest of what gdbarch.sh would
>    generate (wrapper functions, default initialization.  I'd rather you
>    just used gdbarch.sh.
> 
> 
> What about Daniels objections that I'm hand-coding much what
> gdbarch.sh already does?  I'm feeling that the modularity is worth it,
> but how do you feel about that?

No. Yes.  Using gdbarch, and loosing that modularity, is far too high a 
price to pay.

Andrew