diff options
author | Wilco Dijkstra <wilco.dijkstra@arm.com> | 2022-08-22 13:21:53 +0100 |
---|---|---|
committer | Szabolcs Nagy <szabolcs.nagy@arm.com> | 2022-08-23 10:09:05 +0100 |
commit | a1547d148400deffeaab1cd484638ec03a519682 (patch) | |
tree | 9225fbe07f9603207547b78c4bb5c1f516e85a98 /string | |
parent | 864fc0254172a47c9ea36e12cd27511108902813 (diff) | |
download | arm-optimized-routines-a1547d148400deffeaab1cd484638ec03a519682.tar.gz |
string: Optimize memrchr
Optimize the main loop - large strings are 43% faster.
Diffstat (limited to 'string')
-rw-r--r-- | string/aarch64/memrchr.S | 21 |
1 files changed, 12 insertions, 9 deletions
diff --git a/string/aarch64/memrchr.S b/string/aarch64/memrchr.S index 47fbce2..4726618 100644 --- a/string/aarch64/memrchr.S +++ b/string/aarch64/memrchr.S @@ -51,7 +51,7 @@ ENTRY (__memrchr_aarch64) dup vrepchr.16b, chrin cmeq vhas_chr.16b, vdata.16b, vrepchr.16b neg shift, end, lsl 2 - shrn vend.8b, vhas_chr.8h, 4 /* 128->64 */ + shrn vend.8b, vhas_chr.8h, 4 /* 128->64 */ fmov synd, dend lsl synd, synd, shift cbz synd, L(start_loop) @@ -62,31 +62,34 @@ ENTRY (__memrchr_aarch64) csel result, result, xzr, hi ret + nop L(start_loop): - sub tmp, end, src - subs cntrem, cntin, tmp + subs cntrem, src, srcin b.ls L(nomatch) /* Make sure that it won't overread by a 16-byte chunk */ - add tmp, cntrem, 15 - tbnz tmp, 4, L(loop32_2) + sub cntrem, cntrem, 1 + tbz cntrem, 4, L(loop32_2) + add src, src, 16 - .p2align 4 + .p2align 5 L(loop32): - ldr qdata, [src, -16]! + ldr qdata, [src, -32]! cmeq vhas_chr.16b, vdata.16b, vrepchr.16b umaxp vend.16b, vhas_chr.16b, vhas_chr.16b /* 128->64 */ fmov synd, dend cbnz synd, L(end) L(loop32_2): - ldr qdata, [src, -16]! + ldr qdata, [src, -16] subs cntrem, cntrem, 32 cmeq vhas_chr.16b, vdata.16b, vrepchr.16b - b.ls L(end) + b.lo L(end_2) umaxp vend.16b, vhas_chr.16b, vhas_chr.16b /* 128->64 */ fmov synd, dend cbz synd, L(loop32) +L(end_2): + sub src, src, 16 L(end): shrn vend.8b, vhas_chr.8h, 4 /* 128->64 */ fmov synd, dend |