Re: [PATCH][AArch64] Adjust writeback in non-zero memset

[Wilco Dijkstra <Wilco dot Dijkstra at arm dot com>]

v2: Also bias dst in the zva_other code to avoid issues with zva sizes >= 256.

This fixes an ineffiency in the non-zero memset.  Delaying the writeback
until the end of the loop is slightly faster on some cores - this shows
~5% performance gain on Cortex-A53 when doing large non-zero memsets.

Tested against the GLIBC testsuite, OK for commit?

ChangeLog:
2018-11-14  Wilco Dijkstra  <wdijkstr@arm.com>

        * sysdeps/aarch64/memset.S (MEMSET): Improve non-zero memset loop.

---

diff --git a/sysdeps/aarch64/memset.S b/sysdeps/aarch64/memset.S
index 4a454593618f78e22c55520d56737fab5d8f63a4..9738cf5fd55a1d937fb3392cec46f37b4d5fb51d 100644
--- a/sysdeps/aarch64/memset.S
+++ b/sysdeps/aarch64/memset.S
@@ -89,10 +89,10 @@ L(set_long):
     b.eq    L(try_zva)
 L(no_zva):
     sub    count, dstend, dst    /* Count is 16 too large.  */
-    add    dst, dst, 16
+    sub    dst, dst, 16        /* Dst is biased by -32.  */
     sub    count, count, 64 + 16    /* Adjust count and bias for loop.  */
-1:    stp    q0, q0, [dst], 64
-    stp    q0, q0, [dst, -32]
+1:    stp    q0, q0, [dst, 32]
+    stp    q0, q0, [dst, 64]!
 L(tail64):
     subs    count, count, 64
     b.hi    1b
@@ -183,6 +183,7 @@ L(zva_other):
     subs    count, count, zva_len
     b.hs    3b
 4:    add    count, count, zva_len
+    sub    dst, dst, 32        /* Bias dst for tail loop.  */
     b    L(tail64)
 #endif