Eliminate another redundant instruction in vp56/8 arithcoder

Necessary because of this GCC bug:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44474
To do this, convert some, but not all (!) of the variables in VP56RangeCoder
into local variables.
If we convert c->high into a local variable, gcc gets the stupids and refuses
to use a conditional move for the unpredictable main branch.

TODO: dispense with this bullshit and write an asm version.

Originally committed as revision 23924 to svn://svn.ffmpeg.org/ffmpeg/trunk
This commit is contained in:
Jason Garrett-Glaser 2010-06-30 23:59:27 +00:00
parent 2e79db0141
commit 4148855ee4
1 changed files with 17 additions and 8 deletions

View File

@ -194,23 +194,32 @@ static inline void vp56_init_range_decoder(VP56RangeCoder *c,
static inline int vp56_rac_get_prob(VP56RangeCoder *c, uint8_t prob)
{
/* Don't put c->high in a local variable; if we do that, gcc gets
* the stupids and turns the code below into a branch again. */
int bits = c->bits;
unsigned long code_word = c->code_word;
unsigned int low = 1 + (((c->high - 1) * prob) >> 8);
unsigned int low_shift = low << 8;
int bit = c->code_word >= low_shift;
int bit = code_word >= low_shift;
int shift;
/* Incantation to convince GCC to turn these into conditional moves
* instead of branches -- faster, as this branch is basically
* unpredictable. */
c->high = bit ? c->high - low : low;
c->code_word = bit ? c->code_word - low_shift : c->code_word;
code_word = bit ? code_word - low_shift : code_word;
/* normalize */
shift = ff_h264_norm_shift[c->high] - 1;
c->high <<= shift;
c->code_word <<= shift;
c->bits += shift;
if(c->bits >= 0 && c->buffer < c->end) {
c->code_word |= *c->buffer++ << c->bits;
c->bits -= 8;
c->high <<= shift;
code_word <<= shift;
bits += shift;
if(bits >= 0 && c->buffer < c->end) {
code_word |= *c->buffer++ << bits;
bits -= 8;
}
c->bits = bits;
c->code_word = code_word;
return bit;
}