[FFmpeg-devel] [PATCH] h264: assembly version of get_cabac for x86_64 with PIC

[Roland Scheidegger]

Am 17.04.2012 11:35, schrieb Michael Niedermayer:
> On Mon, Apr 16, 2012 at 03:50:09PM +0200, Roland Scheidegger wrote:
>> Am 16.04.2012 14:06, schrieb Michael Niedermayer:
>>> On Mon, Apr 16, 2012 at 12:55:20PM +0200, Roland Scheidegger wrote:
>>>> Am 15.04.2012 10:19, schrieb Michael Niedermayer:
>>>>> On Sat, Apr 14, 2012 at 01:50:21AM +0200, Roland Scheidegger wrote:
>>>>> [...]
>>>>>
>>>>>> +#ifdef BROKEN_RELOCATIONS
>>>>>> +        BRANCHLESS_GET_CABAC("%4", "%q4", "(%1)", "%3", "%w3",
>>>>>> +                             "%5", "%q5", "%k0", "%b0",
>>>>>> +                             "%a11(%6)", "%a12(%6)", "%13")
>>>>>> +#else
>>>>>>          BRANCHLESS_GET_CABAC("%4", "(%1)", "%3", "%w3",
>>>>>>                               "%5", "%k0", "%b0",
>>>>>>                               "%a11(%6)", "%a12(%6)")
>>>>>> +#endif
>>>>>>  
>>>>>>          "test $1, %4                            \n\t"
>>>>>>          " jz 4f                                 \n\t"
>>>>>>          "add  %10, %1                           \n\t"
>>>>>>  
>>>>>> +#ifdef BROKEN_RELOCATIONS
>>>>>> +        BRANCHLESS_GET_CABAC("%4", "%q4", "(%1)", "%3", "%w3",
>>>>>> +                             "%5", "%q5", "%k0", "%b0",
>>>>>> +                             "%a11(%6)", "%a12(%6)", "%13")
>>>>>> +#else
>>>>>>          BRANCHLESS_GET_CABAC("%4", "(%1)", "%3", "%w3",
>>>>>>                               "%5", "%k0", "%b0",
>>>>>>                               "%a11(%6)", "%a12(%6)")
>>>>>> +#endif
>>>>>>  
>>>>>>          "sub  %10, %1                           \n\t"
>>>>>>          "mov  %2, %0                            \n\t"
>>>>>
>>>>> these 2 ifdefs can be avoided by making the macro arguments between
>>>>> the 2 variants match
>>>>
>>>> Of course but then you have unnecessary arguments for the non-pic case.
>>>> In particular also one reg more. I don't think that's a good idea -
>>>> didn't test if it even compiles on x86 due to register pressure.
>>>
>>> iam talking about the macro arguments, not the asm arguments.
>>> The code would be identical after the C preprocessor so no way there
>>> could be any issue with anything, you have same C source to the
>>> C compiler and same asm to the assembler and binary identical object
>>> files except line numbers / debug symbols
>>>
>> Maybe I'm missing some macro-foo (which I'm bad at) but I still don't
>> see how that should look like, since the macro arguments directly come
>> from the operand specification of the inline assembly. Or do you suggest
>> passing in some fake immediate (which will be unused) there or something
>> similar instead of a reg if we don't need the pic version?
> 
> not sure if fake immediate is matching what i meant, but let me provide
> an example, i should have done that previously already
> 
> instead of
> 
> #if this
> #define THAT(a,b,c,d)
> ...
> #else
> #define THAT(a,c)
> ...
> #endif
> 
> asm(
>     ...
> #if this
> THAT(a,b,c,d)
> #else
> THAT(a,c)
> #endif
>     ...
> )
> 
> one can do:
> #if this
> #define THAT(a,b,c,d)
> ...
> #else
> #define THAT(a,b,c,d)
> ... //no code here uses b or d
> #endif
> 
> asm(
>     ...
> THAT(a,b,c,d)
>     ...
> ) // the b and d arguments dont exist in the second case
> 
> 
> 
> also keep in mind that code like this:
> #define this(a,b,c,d) return a
> main(){
>     this(0, not exist, "realy not", [{;:/ haha);
> }
> 
> is perfectly valid, no need for the arguments here to exist of make
> any sense as long as they arent used
> 

Ahh finally got it thanks. Somehow was completely missing that the
arguments will never be seen as they get replaced by the macro
preprocessor first.

Roland