From 099c6f6d45c0b4cd50d768428d6c1cf0ca93c624 Mon Sep 17 00:00:00 2001
From: Mateusz Guzik <mjg@FreeBSD.org>
Date: Mon, 22 Oct 2018 06:44:20 +0000
Subject: amd64: finish the tail in memset with an overlapping store

Instead of finding the exact size to fit in we can just shift the target
by -8 + tail. Doing a blind write to a previously rep stosq'ed area comes
with a penalty so do it conditionally.

Sample win on EPYC when zeroing a 257 sized buffer (tail = 1) aligned to
16 bytes:
before: 44782846 ops/s
after:  46118614 ops/s

Idea stolen from NetBSD.

Sponsored by:	The FreeBSD Foundation
---
 sys/amd64/amd64/support.S | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/sys/amd64/amd64/support.S b/sys/amd64/amd64/support.S
index d3e5a829b67d..bbcbd54937f9 100644
--- a/sys/amd64/amd64/support.S
+++ b/sys/amd64/amd64/support.S
@@ -524,9 +524,12 @@ END(memcpy_erms)
 	rep
 	stosq
 	movq	%r9,%rax
-	movq	%rdx,%rcx
-	andb	$7,%cl
-	jne	1004b
+	andl	$7,%edx
+	jnz	1f
+	POP_FRAME_POINTER
+	ret
+1:
+	movq	%r10,-8(%rdi,%rdx)
 .endif
 	POP_FRAME_POINTER
 	ret
-- 
cgit v1.3