Anyone knows about a testcase which was fixed by Wed Dec 24 18:05:13 1997 SPARC change?

[Jakub Jelinek <jakub at redhat dot com>]

Hi!

Can anyone think of a testcase which was fixed by

Wed Dec 24 18:05:13 1997  Richard Kenner  <kenner@vlsi1.ultra.nyu.edu>

        * sparc.c: Add prototypes for static functions.
        (check_pic): Check for form of pic_pc_rtx, not it itself.
        (pic_setup_code): New function, from finalize_pic.
        (finalize_pic): Call pic_setup_code and insert after nonlocal_receiver.
        * sparc.md (nonlocal_goto_receiver): New pattern.

(well, particularly the nonlocal_goto_receiver and reloading of %l7 (pic
base register) in all nonlocal_goto_receivers)? I cannot think of any, nor
can DaveM. If noone comes with explanation why %l7 might get clobbered, I'll
submit a patch to remove this shortly.
Thanks.

On Thu, Oct 05, 2000 at 03:17:40PM -0700, David S. Miller wrote:
>    Date: Thu, 5 Oct 2000 17:12:19 -0400
>    From: Jakub Jelinek <jakub@redhat.com>
> 
>    Could you please have a look at my question why is reloading of %l7
>    necessary in sparc nonlocal_goto_receiver?  It makes no sense to
>    me, because anyone who want to jump back to that label (be it
>    longjmp, goto from nested function or whatever) has to end up in
>    function's original register window (how otherwise would be e.g. %i7
>    restored?) and thus %l7 will be reloaded without any problems.
>    I cannot work on a better fix for the libm miscompilations with -mvis.
> 
> I think it is unneeded.  We never allow %l7 to be clobbered locally
> by code generated by GCC and I can't think of any register window
> situation which would not preserve %l7.
> 
> I think Kenner just doesn't understand how register windows work and
> what they guarentee.
> 
> But, just to be sure, ask around if anyone knows of any real bug
> report which this nonlocal_goto_receiver thing actually fixed.
> 
> This all smells bogus to me, IMHO.
> 
> Later,
> David S. Miller
> davem@redhat.com

	Jakub