[XviD-devel] Decoder performance

[Edouard Gomez]

Christoph Lampert (chl at math.uni-bonn.de) wrote:
> please be careful when replacing 
> 
>    if (quarterpel) 
> 
> [... usually sthg like this ...]
> (current->vol_flags & XVID_VOL_QUARTERPEL)

Oh,  i had  a  nicer  solution than  the  ? operator  for  that (on  the
paper). Completely branchless.

#define IS_SET_FLAG(a, flag) \
 (!!((a)&(flag)))

It 'returns':
 - 1 for (a & flag) != 0
 - 0 for (a & flag) == 0

It's translated to that sort of assembly in ia32:
-O2 int test(a, flag);
080482f4 <test>:
function prolog
 80482f7:       8b 45 0c                mov    eax, a
 80482fa:       85 45 08                test   flag, eax
 80482fd:       0f 95 c0                setne  al
 8048300:       0f b6 c0                movzbl eax, al
function termination

The -O0 int  test(a, flag); adds a few instructions more  , doing a real
bitwise and operation, not the 'test' fake one.

Now the point  is that the ?  operator is implemented  the same way when
doing a simple bitwise and returning 1 or 0 :-) The best thing we should
do, is define  two macros, one for boolean  returned value required, and
one just  to do the 'and'. This  would explicit what is  intended in the
code.

Something like:
FLAG_IS_SET_BOOL(a, flag) returns 1 if 'flag' is set in 'a' or 0 otherwise.
FLAG_IS_SET(a, flag) returns != 0 if 'flag' is set oin 'a' or 0 otherwise.

-- 
Edouard Gomez