prefetching does not help length of match
l_lx_elf64amd.S amd_d_nrv2b.S amd_d_nrv2e.S committer: jreiser <jreiser> 1132290120 +0000
This commit is contained in:
@@ -102,6 +102,17 @@ ra_setup:
|
||||
*/
|
||||
/* jump on next bit {0,1} with prediction {y==>likely, n==>unlikely} */
|
||||
/* Prediction omitted for now. */
|
||||
/* On refill: prefetch next byte, for latency reduction on literals and offsets. */
|
||||
#define jnextb0np jnextb0yp
|
||||
#define jnextb0yp GETBITp; jnc
|
||||
#define jnextb1np jnextb1yp
|
||||
#define jnextb1yp GETBITp; jc
|
||||
#define GETBITp \
|
||||
addl bits,bits; jnz 0f; \
|
||||
movl (%rsi),bits; subq $-4,%rsi; \
|
||||
adcl bits,bits; movb (%rsi),%dl; \
|
||||
0:
|
||||
/* Same, but without prefetch (not useful for length of match.) */
|
||||
#define jnextb0n jnextb0y
|
||||
#define jnextb0y GETBIT; jnc
|
||||
#define jnextb1n jnextb1y
|
||||
@@ -109,11 +120,12 @@ ra_setup:
|
||||
#define GETBIT \
|
||||
addl bits,bits; jnz 0f; \
|
||||
movl (%rsi),bits; subq $-4,%rsi; \
|
||||
adcl bits,bits; movb (%rsi),%dl; \
|
||||
adcl bits,bits; \
|
||||
0:
|
||||
|
||||
/* rotate next bit into bottom bit of reg */
|
||||
#define getnextb(reg) call *%r11; adcl reg,reg
|
||||
#define getnextbp(reg) call *%r11; adcl reg,reg
|
||||
#define getnextb(reg) getnextbp(reg)
|
||||
|
||||
ALIGN(1<<3)
|
||||
getbit:
|
||||
|
||||
Reference in New Issue
Block a user