[rfc] Introduce "per-objfile" gdbarch

[Daniel Jacobowitz]

On Sun, Dec 09, 2007 at 08:40:28PM +0100, Ulrich Weigand wrote:
> The only instances that remain are related to:
> 
> - Register numbering (gdbarch_XXX_reg_to_regnum)
> 
>   This is somewhat of a fundamental problem: if we allow GDB register numbers
>   to depend on per-target information, having symbol classes like LOC_REGISTER
>   simply makes no sense, as during symbol reading the GDB register numbers 
>   are not really well-defined yet.
> 
>   This is not an issue with DWARF-2, as it always uses LOC_COMPUTED instead.
>   Maybe it would be good to change the other symbol readers in a similar way?

That would work.

Another possibility would be to make the mappings independent of the
target; I've demonstrated how you can do this by making unavailable
registers have the int0 type and no name, but still have reserved
register numbers.  It does simplify quite a lot.

> - dwarf2_tracepoint_var_ref calls gdbarch_virtual_frame_pointer.  I guess that
>   gdbarch callback really needs to be eliminated.

Jim has been, I think, working on this.

> Note that the patch makes no effort to *enforce* that the per-objfile 
> gdbarch is used only for the subset of functions described above.  I'm
> not sure whether we should do so (and how?), or whether this is simply
> a matter of proper documentation ...
> 
> What do you think about this approach?

I like it.

-- 
Daniel Jacobowitz
CodeSourcery