[FFmpeg-devel] [RFC] clobbers for XMM registers

[Måns Rullgård]

Michael Niedermayer <michaelni at gmx.at> writes:

> On Thu, Sep 30, 2010 at 08:49:04PM -0400, Alexander Strange wrote:
>> 
>> On Sep 30, 2010, at 8:31 PM, Michael Niedermayer wrote:
>> 
>> > On Fri, Oct 01, 2010 at 12:42:33AM +0100, M?ns Rullg?rd wrote:
>> >> Michael Niedermayer <michaelni at gmx.at> writes:
>> >> 
>> >>> On Thu, Sep 30, 2010 at 04:02:00PM -0300, Ramiro Polla wrote:
>> >>>> 2010/9/30 M?ns Rullg?rd <mans at mansr.com>:
>> >>>>> Ramiro Polla <ramiro.polla at gmail.com> writes:
>> >>>>>> 2010/9/30 M?ns Rullg?rd <mans at mansr.com>:
>> >>>>>>> Ramiro Polla <ramiro.polla at gmail.com> writes:
>> >>>>>>>> What about
>> >>>>>>>> #if HAVE_XMM_CLOBBERS
>> >>>>>>>> #    define XMM_CLOBBERS(a, ...) __VA_ARGS__
>> >>>>>>>> #else
>> >>>>>>>> #    define XMM_CLOBBERS(a, ...) a
>> >>>>>>>> #endif
>> >>>>>>>> 
>> >>>>>>>> to be used as in lavc/x86/fft_sse.c:
>> >>>>>>>>         :"+r"(j), "+r"(k)
>> >>>>>>>>         :"r"(output+n4), "r"(output+n4*3),
>> >>>>>>>>          "m"(*m1m1m1m1)
>> >>>>>>>>         XMM_CLOBBERS(, : "%xmm0", "%xmm1", "%xmm7")
>> >>>>>>>>     );
>> >>>>>>> 
>> >>>>>>> That falls over if any other clobbers are needed.
>> >>>>>> 
>> >>>>>> If any other clobbers are needed they could be written before the macro.
>> >>>>> 
>> >>>>> That won't work.
>> >>>> 
>> >>>> : "eax" XMM_CLOBBERS(,, "%xmm0", "%xmm1", "%xmm7")
>> >>>> 
>> >>>> But I agree having "cc" as dummy is simpler.
>> >>> 
>> >>> it may be simpler but your variant is more correct
>> >> 
>> >> If "cc" is a no-op neither variant is more correct.  Mine is a hell of
>> >> a lot prettier though.
>> > 
>> > unless "cc" is documented to be a no-op on x86 we should not assume it is
>> 
>> svn blame on gcc says it hasn't changed since x86 was rewritten in 1999.
>> (the code in question being ix86_md_asm_clobbers(), in a file too big to display in viewcvs)
>> 
>> I hardly see how it could change now since it would break all users of asm.
>
> i agree with you still i dont think the risk with cc byting us is 0

Yet you are willing to assume that gcc will never, ever start using
registers in a way incompatible with your under-specified inline asm.
The mind boggles.

> and theres a perfectly working solution, namely ramiros macro.

It may work, but perfect it is not.  The syntax is horrible and non-obvious.
Mistakes are inevitable.

> and it wont cost us anything to use ramiros macro now.

It will cost us beauty.

-- 
M?ns Rullg?rd
mans at mansr.com