[FFmpeg-devel] [PATCH] split-radix FFT

[Måns Rullgård]

Michael Niedermayer <michaelni at gmx.at> writes:

> On Tue, Jul 29, 2008 at 05:20:15PM +0100, M?ns Rullg?rd wrote:
>> 
>> Michael Niedermayer wrote:
>> > On Tue, Jul 29, 2008 at 06:26:49PM +0300, Uoti Urpala wrote:
>> >> On Tue, 2008-07-29 at 17:10 +0200, Michael Niedermayer wrote:
>> >> > And just to clarify, yes what i considered a good argument was the
>> >> sentance
>> >> > above where my reply is. That is to use MANGLE in speed critical code.
>> >> > That way most textrels are avoided while minimizing the speed impact.
>> >> >
>> >> > I do not think you ever argued for that.
>> >>
>> >> IIRC I did mention the possibility of omitting -fPIC for a subset of
>> >> files.
>> >>
>> >> >  I remember you strongly arguing
>> >> > toward replacing all MANGLE by "m" knowing that it would break gcc 2.95
>> >> > and not really caring that it would slow down code compiled with -fPIC.
>> >>
>> >> Of course the code would be slower on x86. If you want it to be as fast
>> >> as possible then compile it with -fPIC on x86. I don't think it's
>> >> worthwhile to pick only the globals used inside asm for such special
>> >> treatment.
>> >
>> > x86-64 shared libs require -fPIC, unless that has been fixed.
>> 
>> The x86-64 instruction set hasn't been "fixed", and I doubt it ever
>> will be.  You simply can't fit a 64-bit offset in a 32-bit immediate
>> operand.
>
> Thats not what i meant
>
>> 
>> > so the user does not always have the option to omit -fPIC
>> 
>> But in these cases, forcing a textrel will break the build.
>
> MANGLE forces rip relative addressing on x86-64 and thus avoids the
> occasional GOT indirection gcc adds.
>
> Heres a example:
> long globivar;
>
> void func(){
>     asm(
>         "mov globivar(%rip), %rax\n\t"
>     );
>     asm(
>         "mov %0, %%rax\n\t"
>         :: "m"(globivar)
>     );
> }
>
> results in:
> 0000000000000554 <func>:
>  554:	55                   	push   %rbp
>  555:	48 89 e5             	mov    %rsp,%rbp
>  558:	48 8b 05 d1 02 20 00 	mov    0x2002d1(%rip),%rax        # 200830 <globivar>
>  55f:	48 8b 05 8a 02 20 00 	mov    0x20028a(%rip),%rax        # 2007f0 <_DYNAMIC+0x1b8>
>  566:	48 8b 00             	mov    (%rax),%rax
>  569:	c9                   	leaveq 
>  56a:	c3                   	retq   
>
> you can see the second needs 2 instructions, the first just 1.

There is no guarantee that &globivar is reachable with a 32-bit offset
from %rip (or any other register).

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