[glibc] AArch64: Optimize memchr

[Wilco Dijkstra <wilco@sourceware.org>]

https://sourceware.org/git/gitweb.cgi?p=glibc.git;h=ce758d4f063820c2bc743e12797d7454c66be718

commit ce758d4f063820c2bc743e12797d7454c66be718
Author: Wilco Dijkstra <wilco.dijkstra@arm.com>
Date:   Wed Jan 11 13:50:59 2023 +0000

    AArch64: Optimize memchr
    
    Optimize the main loop - large strings are 40% faster on modern CPUs.
    
    Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>

Diff:
---
 sysdeps/aarch64/memchr.S | 27 ++++++++++++++-------------
 1 file changed, 14 insertions(+), 13 deletions(-)

diff --git a/sysdeps/aarch64/memchr.S b/sysdeps/aarch64/memchr.S
index 1cd32bf8cf..1c99d45fbb 100644
--- a/sysdeps/aarch64/memchr.S
+++ b/sysdeps/aarch64/memchr.S
@@ -30,7 +30,6 @@
 # define MEMCHR __memchr
 #endif
 
-/* Arguments and results.  */
 #define srcin		x0
 #define chrin		w1
 #define cntin		x2
@@ -73,42 +72,44 @@ ENTRY (MEMCHR)
 
 	rbit	synd, synd
 	clz	synd, synd
-	add	result, srcin, synd, lsr 2
 	cmp	cntin, synd, lsr 2
+	add	result, srcin, synd, lsr 2
 	csel	result, result, xzr, hi
 	ret
 
+	.p2align 3
 L(start_loop):
 	sub	tmp, src, srcin
-	add	tmp, tmp, 16
+	add	tmp, tmp, 17
 	subs	cntrem, cntin, tmp
-	b.ls	L(nomatch)
+	b.lo	L(nomatch)
 
 	/* Make sure that it won't overread by a 16-byte chunk */
-	add	tmp, cntrem, 15
-	tbnz	tmp, 4, L(loop32_2)
-
+	tbz	cntrem, 4, L(loop32_2)
+	sub	src, src, 16
 	.p2align 4
 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]!
-	subs	cntrem, cntrem, 32
+	ldr	qdata, [src, 16]
 	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 */
 	fmov	synd, dend
 	cbz	synd, L(loop32)
+L(end_2):
+	add	src, src, 16
 L(end):
 	shrn	vend.8b, vhas_chr.8h, 4		/* 128->64 */
+	sub	cntrem, src, srcin
 	fmov	synd, dend
-	add	tmp, srcin, cntin
-	sub	cntrem, tmp, src
+	sub	cntrem, cntin, cntrem
 #ifndef __AARCH64EB__
 	rbit	synd, synd
 #endif