From 20ca271fdd651f12793f6e2580b8d19f627d80f7 Mon Sep 17 00:00:00 2001
From: Mateusz Guzik <mjg@FreeBSD.org>
Date: Wed, 9 May 2018 15:16:25 +0000
Subject: [PATCH] amd64: depessimize bcmp for small buffers

Adapt assembly generated by clang for memcmp and use it for <= 64 sized
compares (which are the vast majority).

Sample result of doing stats on Broadwell (% of samples):
before: 4.0 kernel     bcmp                 cache_lookup
after : 0.7 kernel     bcmp                 cache_lookup

The routine is most definitely still not optimal. Anyone interested in
spending time improving it is welcome to take over.

Reviewed by:	kib
---
 sys/amd64/amd64/support.S | 27 +++++++++++++++++++++++++--
 1 file changed, 25 insertions(+), 2 deletions(-)

diff --git a/sys/amd64/amd64/support.S b/sys/amd64/amd64/support.S
index 79b9db3d6283..7c1b52cdd0de 100644
--- a/sys/amd64/amd64/support.S
+++ b/sys/amd64/amd64/support.S
@@ -98,17 +98,40 @@ END(sse2_pagezero)
 
 ENTRY(bcmp)
 	PUSH_FRAME_POINTER
+	test	%rdx,%rdx
+	je	1f
+	cmpq	$64,%rdx
+	jg	4f
+
+	xor	%ecx,%ecx
+2:
+	movzbl	(%rdi,%rcx,1),%eax
+	movzbl	(%rsi,%rcx,1),%r8d
+	cmp	%r8b,%al
+	jne	3f
+	add	$0x1,%rcx
+	cmp	%rcx,%rdx
+	jne	2b
+1:
+	xor	%eax,%eax
+	POP_FRAME_POINTER
+	retq
+3:
+	mov	$1,%eax
+	POP_FRAME_POINTER
+	retq
+4:
 	movq	%rdx,%rcx
 	shrq	$3,%rcx
 	repe
 	cmpsq
-	jne	1f
+	jne	5f
 
 	movq	%rdx,%rcx
 	andq	$7,%rcx
 	repe
 	cmpsb
-1:
+5:
 	setne	%al
 	movsbl	%al,%eax
 	POP_FRAME_POINTER