AuroraMiddleware/glibc
1
0
Fork 0
mirror of https://sourceware.org/git/glibc.git synced 2024-11-25 22:40:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

AArch64: Optimize memchr

Browse Source 
Optimize the main loop - large strings are 40% faster on modern CPUs.

Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
This commit is contained in:
Wilco Dijkstra 2023-01-11 13:50:59 +00:00
parent 569cfcc6bf
commit ce758d4f06
1 changed files with 14 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
sysdeps/aarch64/memchr.S
View File

@ -30,7 +30,6 @@
# define MEMCHR __memchr # define MEMCHR __memchr
#endif #endif
/* Arguments and results. */
#define srcin x0 #define srcin x0
#define chrin w1 #define chrin w1
#define cntin x2 #define cntin x2
@ -73,42 +72,44 @@ ENTRY (MEMCHR)
rbit synd, synd rbit synd, synd
clz synd, synd clz synd, synd
add result, srcin, synd, lsr 2
cmp cntin, synd, lsr 2 cmp cntin, synd, lsr 2
add result, srcin, synd, lsr 2
csel result, result, xzr, hi csel result, result, xzr, hi
ret ret
.p2align 3
L(start_loop): L(start_loop):
sub tmp, src, srcin sub tmp, src, srcin
add tmp, tmp, 16 add tmp, tmp, 17
subs cntrem, cntin, tmp subs cntrem, cntin, tmp
b.ls L(nomatch) b.lo L(nomatch)
/* Make sure that it won't overread by a 16-byte chunk */ /* Make sure that it won't overread by a 16-byte chunk */
add tmp, cntrem, 15 tbz cntrem, 4, L(loop32_2)
tbnz tmp, 4, L(loop32_2) sub src, src, 16
.p2align 4 .p2align 4
L(loop32): L(loop32):
ldr qdata, [src, 16]! ldr qdata, [src, 32]!
cmeq vhas_chr.16b, vdata.16b, vrepchr.16b cmeq vhas_chr.16b, vdata.16b, vrepchr.16b
umaxp vend.16b, vhas_chr.16b, vhas_chr.16b /* 128->64 */ umaxp vend.16b, vhas_chr.16b, vhas_chr.16b /* 128->64 */
fmov synd, dend fmov synd, dend
cbnz synd, L(end) cbnz synd, L(end)
L(loop32_2): L(loop32_2):
ldr qdata, [src, 16]! ldr qdata, [src, 16]
subs cntrem, cntrem, 32
cmeq vhas_chr.16b, vdata.16b, vrepchr.16b cmeq vhas_chr.16b, vdata.16b, vrepchr.16b
b.ls L(end) subs cntrem, cntrem, 32
b.lo L(end_2)
umaxp vend.16b, vhas_chr.16b, vhas_chr.16b /* 128->64 */ umaxp vend.16b, vhas_chr.16b, vhas_chr.16b /* 128->64 */
fmov synd, dend fmov synd, dend
cbz synd, L(loop32) cbz synd, L(loop32)
L(end_2):
add src, src, 16
L(end): L(end):
shrn vend.8b, vhas_chr.8h, 4 /* 128->64 */ shrn vend.8b, vhas_chr.8h, 4 /* 128->64 */
sub cntrem, src, srcin
fmov synd, dend fmov synd, dend
add tmp, srcin, cntin sub cntrem, cntin, cntrem
sub cntrem, tmp, src
#ifndef __AARCH64EB__ #ifndef __AARCH64EB__
rbit synd, synd rbit synd, synd
#endif #endif