This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: RFA: MIPS paired single support
Richard Sandiford <rsandifo@redhat.com> writes:
> James E Wilson <wilson@specifixinc.com> writes:
>> I do think that reducing the size of the problem would make things
>> easier for both of us. If there is nothing really wrong with the patch,
>> then why not accept it in its current form, and require me to fix
>> problems after check-in?
>
> OK, go ahead.
Jim seems to be away from email at the moment, so I went ahead and
applied it. Hope that's OK. Thanks to both Jim and MIPS for the work
(and for their patience).
I committed the patch exactly as posted except that I added a copyright
banner to the new mips-ps-3d.md file. I also tweaked the changelog
(very) slightly. I didn't apply the testcases because there seemed to
be some disagreement about where they should go. Bootstrapped &
regression tested on mips-sgi-irix6.5.
FWIW, the reason for committing it now (rather than waiting a bit longer)
is that I've been doing subsequent MIPS work relative to this patch.
Most of the work is only really suitable for stage 2 and that deadline
is fast approaching.
Richard
PS. For avoidance of doubt, the rest of my mail still stands...
> WRT your TODO list, I've already volunteered to do these two items:
>
>> - reduce number of patterns, maybe encode comparison code into UNSPEC
>> parallel instead of into UNSPEC number
>> - use mode macros for v2sf patterns as appropriate
>
> and I get the impression that the following items are the ones you're
> least happy with, so I'll volunteer to do them as well:
>
>> - break dependency between mips.c bdesc table and mips.h BUILTINS_* enum table
>> - in mips_expand_* functions, use loops to reduce code duplication
>> - rewrite mips_expand_* function to avoid huge switch tables
>> - try to reduce or eliminate the MIPS_BUILTIN_* enums
>> - in general, reduce cut and paste more
>
> I'll also deal with:
>
>> - split the 4-wide (8 length) instructions after reload
>
> since it will be easier to do after (or at the same time as) macroisation.
>
> I'll leave you or MIPS to deal with:
>
>> - improve docs for new options
>> - movdf/movsf/movv2sf patterns have bad type attribute of fpstore for m/G alt
>> - add __mips3d predefined macro
>> - add new attributes for recip/sqrt step 1/2 patterns
>
> Richard