[PATCH] Add MIPS r6 support

YunQiang Su syq at debian.org
Wed Jun 19 00:33:49 UTC 2019

Yes. I am working on it. And will send new version of patch soon.

Torbjörn Granlund <tg at gmplib.org> 于2019年6月18日周二 下午5:19写道:
> I went ahead and comitted my minimal changes.
> When you have time, please give feedback on my changes.  Do they work
> for you?  I actually expect slightly better performance with just my
> changes than with your asm code; my longlong.h change allows the
> compiler to generate decent code also for functions not included in your
> patch.
>   Your mips32r6 asm code will presumably not get used as no configure.ac
>   mips32 changes analogous to the mips64 changes have been made.  Any
>   particular reason for that?  (Are there any mips32r6 CPUs out there?)
> Are there?
> If there are, I'll consider making analogous changes for mips32.  (But
> these might be hard to test.)
>   As you pointed out before, mipsr6 is in a way a new architecture, as r5
>   code does not run on an r6, or vice versa.  For that reason, I think we
>   need to separate them more cleanly in GMP.
>   My suggestion os to move any code using the LO or HI registers to a
>   subdirectory mips64/hilo.  That way, things will be less fragile, and we
>   do not need to hide HI/LO GMP files with a corresponding r6 file.
> I comitted code doing that for mips64.
>   After this change, we can add r6 asm code for functions not handled well
>   enough by the compiler.
> I need your feedback here with benchmark results from actual hardware.
> Specifically, does your suggested assembly code improve speed?  I
> understand that your main motication for providing precisely the
> functions you did provide was to "hide" the hi/lo code in the main
> mips64 asm directory.  With my code reorg, that is no longer necessary.
> --
> Torbjörn
> Please encrypt, key id 0xC8601622

More information about the gmp-devel mailing list