[FFmpeg-devel] r9017 breaks WMA decoding on Intel Macs

[Guillaume Poirier]

Hi,

On May 27, 2007, at 3:08 , Guillaume Poirier wrote:

> Hi,
>
> Le 27 mai 07 ? 14:52, Guillaume POIRIER a ?crit :
>
>> On 5/27/07, Guillaume POIRIER <poirierg at gmail.com> wrote:
>>> Any vorbis should do the trick. I'll try to narrow down the
>>> problem to
>>> see which part of the patch broke it.
>>
>> This hunk is what causes the regression:
>
> Off course this should read: "applying this hunk fixes the  
> regression".
>
>
>> Index: fft_sse.c
>> ===================================================================
>> --- fft_sse.c	(revision 9017)
>> +++ fft_sse.c	(revision 6577)
>> @@ -100,33 +100,20 @@
>>              i = nloops*8;
>>              asm volatile(
>>                  "1: \n\t"
>> -                "sub $32, %0 \n\t"
>> +                "sub $16, %0 \n\t"
>>                  "movaps    (%2,%0), %%xmm1 \n\t"
>>                  "movaps    (%1,%0), %%xmm0 \n\t"
>> -                "movaps  16(%2,%0), %%xmm5 \n\t"
>> -                "movaps  16(%1,%0), %%xmm4 \n\t"
>>                  "movaps     %%xmm1, %%xmm2 \n\t"
>> -                "movaps     %%xmm5, %%xmm6 \n\t"
>>                  "shufps      $0xA0, %%xmm1, %%xmm1 \n\t"
>>                  "shufps      $0xF5, %%xmm2, %%xmm2 \n\t"
>> -                "shufps      $0xA0, %%xmm5, %%xmm5 \n\t"
>> -                "shufps      $0xF5, %%xmm6, %%xmm6 \n\t"
>>                  "mulps   (%3,%0,2), %%xmm1 \n\t" //  cre*re cim*re
>>                  "mulps 16(%3,%0,2), %%xmm2 \n\t" // -cim*im cre*im
>> -                "mulps 32(%3,%0,2), %%xmm5 \n\t" //  cre*re cim*re
>> -                "mulps 48(%3,%0,2), %%xmm6 \n\t" // -cim*im cre*im
>>                  "addps      %%xmm2, %%xmm1 \n\t"
>> -                "addps      %%xmm6, %%xmm5 \n\t"
>>                  "movaps     %%xmm0, %%xmm3 \n\t"
>> -                "movaps     %%xmm4, %%xmm7 \n\t"
>>                  "addps      %%xmm1, %%xmm0 \n\t"
>>                  "subps      %%xmm1, %%xmm3 \n\t"
>> -                "addps      %%xmm5, %%xmm4 \n\t"
>> -                "subps      %%xmm5, %%xmm7 \n\t"
>>                  "movaps     %%xmm0, (%1,%0) \n\t"
>>                  "movaps     %%xmm3, (%2,%0) \n\t"
>> -                "movaps     %%xmm4, 16(%1,%0) \n\t"
>> -                "movaps     %%xmm7, 16(%2,%0) \n\t"
>>                  "jg 1b \n\t"
>>                  :"+r"(i)
>>                  :"r"(p), "r"(p + nloops), "r"(cptr)
>>
>>
>> We're quite lucky, it's the shortest of the 2 hunks.
>>
>> Now I need to figure out what's wrong in that hunk.
>
> There's nothing wrong to this hunk!

Off course there's nothing wrong with this hunk. It's the other one  
that causes the regression. I wish I had turned my brain on this  
morning when I woke up.


> It just duplicates the original code and uses "original register
> number" + 4.
> Why on earth would it break on OSX and not on Linux?
>
> Is there's some qualified guru out there who could could enlighten me
> here?

I guess I'm off to reading the other hunk to figure out what may be  
wrong with it. If I can't figure this out, then I'll have to compare  
the assembler emitted by GCC.

Guillaume