[src/netbsd-1-5]: src/sys/arch/i386/i386 Pull up revisions 1.10-1.12 (request...

[he <he%NetBSD.org@localhost>]

details:   https://anonhg.NetBSD.org/src/rev/dc51628e141c
branches:  netbsd-1-5
changeset: 491023:dc51628e141c
user:      he <he%NetBSD.org@localhost>
date:      Fri Mar 30 22:29:06 2001 +0000

description:
Pull up revisions 1.10-1.12 (requested by fvdl):
  Use optimized i386 assembly version for in4_cksum as well as the
  older in_cksum.  Gets back the 10-15% throughput loss compared
  to the NetBSD 1.4 branch.

diffstat:

 sys/arch/i386/i386/in_cksum.s |  47 +++++++++++++++++++++++++++++++++++++++++-
 1 files changed, 45 insertions(+), 2 deletions(-)

diffs (70 lines):

diff -r b83bb9ab8e8f -r dc51628e141c sys/arch/i386/i386/in_cksum.s
--- a/sys/arch/i386/i386/in_cksum.s     Fri Mar 30 22:22:52 2001 +0000
+++ b/sys/arch/i386/i386/in_cksum.s     Fri Mar 30 22:29:06 2001 +0000
@@ -1,7 +1,7 @@
-/*     $NetBSD: in_cksum.s,v 1.9 1998/12/01 04:31:00 thorpej Exp $     */
+/*     $NetBSD: in_cksum.s,v 1.9.22.1 2001/03/30 22:29:06 he Exp $     */
 
 /*-
- * Copyright (c) 1998 The NetBSD Foundation, Inc.
+ * Copyright (c) 1998, 2001 The NetBSD Foundation, Inc.
  * All rights reserved.
  *
  * This code is derived from software contributed to The NetBSD Foundation
@@ -119,6 +119,48 @@
        addw    %dx, %ax        ; \
        adcw    $0, %ax
 
+ENTRY(in4_cksum)
+       pushl   %ebp
+       pushl   %ebx
+       pushl   %esi
+
+       movl    16(%esp), %ebp
+       movzbl  20(%esp), %eax          /* sum = nxt */
+       movl    28(%esp), %esi
+       movl    M_DATA(%ebp), %ebx
+       addl    %esi, %eax              /* sum += len */
+       movl    24(%esp), %edx          /* %edx = off */
+       shll    $8, %eax                /* sum = htons(sum) */
+
+       ADD(IP_SRC)                     /* sum += ip->ip_src */
+       ADC(IP_DST)                     /* sum += ip->ip_dst */
+       MOP
+
+mbuf_loop_0:
+       testl   %ebp, %ebp
+       jz      out_of_mbufs
+
+       movl    M_DATA(%ebp), %ebx      /* %ebx = m_data */
+       movl    M_LEN(%ebp), %ecx       /* %ecx = m_len */
+       movl    M_NEXT(%ebp), %ebp
+
+       subl    %ecx, %edx              /* %edx = off - m_len */
+       jnb     mbuf_loop_0
+
+       addl    %edx, %ebx              /* %ebx = m_data + off - m_len */
+       negl    %edx                    /* %edx = m_len - off */
+       addl    %ecx, %ebx              /* %ebx = m_data + off */
+       xorb    %cl, %cl
+
+       /*
+        * The len == 0 case is handled really inefficiently, by going through
+        * the whole short_mbuf path once to get back to mbuf_loop_1 -- but
+        * this case never happens in practice, so it's sufficient that it
+        * doesn't explode.
+        */
+       jmp     in4_entry
+
+
 ENTRY(in_cksum)
        pushl   %ebp
        pushl   %ebx
@@ -141,6 +183,7 @@
        movl    M_LEN(%ebp), %edx
        movl    M_NEXT(%ebp), %ebp
 
+in4_entry:
        cmpl    %esi, %edx
        jbe     1f
        movl    %esi, %edx