Hi!
This little patch turns bitwise.c's mask[]
lookup table into a bit-twiddling macro (which avoids
the <<32 case without branching). This
spares the size of the 132-byte lookup table (and
then a bit more, curiously).
While I expected this to generally be faster
(because it uses a few more instructions in place of
a table lookup), it's actually slightly slower
than the LUT, in a tight benchmark (perhaps because
of the large chance of the LUT being in the dcache?),
on an i32-athlon-classic at least. But other
architectures may see some small benefit -- or
this could be of academic interest to someone. :)
It's been _V_BIT_TEST'd on i32, and I expect it to
work on 64-bit archs but haven't been able to test.
--Adam
-------------- next part --------------
Index: .libs/libvorbisidec.a
==================================================================Cannot
display: file marked as a binary type.
svn:mime-type = application/octet-stream
Index: bitwise.c
==================================================================--- bitwise.c
(revision 3086)
+++ bitwise.c (working copy)
@@ -23,14 +23,8 @@
#include "misc.h"
#include "ogg.h"
-static unsigned long
mask[]-{0x00000000,0x00000001,0x00000003,0x00000007,0x0000000f,
- 0x0000001f,0x0000003f,0x0000007f,0x000000ff,0x000001ff,
- 0x000003ff,0x000007ff,0x00000fff,0x00001fff,0x00003fff,
- 0x00007fff,0x0000ffff,0x0001ffff,0x0003ffff,0x0007ffff,
- 0x000fffff,0x001fffff,0x003fffff,0x007fffff,0x00ffffff,
- 0x01ffffff,0x03ffffff,0x07ffffff,0x0fffffff,0x1fffffff,
- 0x3fffffff,0x7fffffff,0xffffffff };
+#define MASK(b) ( (unsigned long)( (( (1<<((b)&31)) ^
(((b)&32)>>5) ) ) -1 ) \
+ & 0xffffffff )
/* spans forward, skipping as many bytes as headend is negative; if
headend is zero, simply finds next byte. If we're up to the end
@@ -80,7 +74,7 @@
/* Read in bits without advancing the bitptr; bits <= 32 */
long oggpack_look(oggpack_buffer *b,int bits){
- unsigned long m=mask[bits];
+ unsigned long m=MASK(bits);
unsigned long ret;
bits+=b->headbit;
@@ -229,15 +223,15 @@
int tbit=bits?bits:ilog(b[i]);
if((test=oggpack_look(&r,tbit))==0xffffffff)
report("out of data!\n");
- if(test!=(b[i]&mask[tbit])){
- fprintf(stderr,"%ld) %lx %lx\n",i,(b[i]&mask[tbit]),test);
+ if(test!=(b[i]&MASK(tbit))){
+ fprintf(stderr,"%ld) %lx %lx\n",i,(b[i]&MASK(tbit)),test);
report("looked at incorrect value!\n");
}
if((test=oggpack_read(&r,tbit))==0xffffffff){
report("premature end of data when reading!\n");
}
- if(test!=(b[i]&mask[tbit])){
- fprintf(stderr,"%ld) %lx %lx\n",i,(b[i]&mask[tbit]),test);
+ if(test!=(b[i]&MASK(tbit))){
+ fprintf(stderr,"%ld) %lx %lx\n",i,(b[i]&MASK(tbit)),test);
report("read incorrect value!\n");
}
bitcount+=tbit;
@@ -607,10 +601,10 @@
j-begin,ilen);
exit(1);
}
- if(temp!=(values[j]&mask[len[j]])){
+ if(temp!=(values[j]&MASK(len[j]))){
fprintf(stderr,"\nERROR: Incorrect read %lx != %lx, word %ld,
len %d\n"
,
- values[j]&mask[len[j]],temp,j-begin,len[j]);
+ values[j]&MASK(len[j]),temp,j-begin,len[j]);
exit(1);
}
bitcount+=len[j];
@@ -642,10 +636,10 @@
j-begin);
exit(1);
}
- if(temp!=(values[j]&mask[len[j]])){
+ if(temp!=(values[j]&MASK(len[j]))){
fprintf(stderr,"\nERROR: Incorrect look %lx != %lx, word %ld,
len %d\n"
,
- values[j]&mask[len[j]],temp,j-begin,len[j]);
+ values[j]&MASK(len[j]),temp,j-begin,len[j]);
exit(1);
}
oggpack_adv(&o,len[j]);
