From owner-freebsd-hackers  Thu Jun 15 07:35:01 1995
Return-Path: hackers-owner
Received: (from majordom@localhost)
          by freefall.cdrom.com (8.6.10/8.6.6) id HAA21402
          for hackers-outgoing; Thu, 15 Jun 1995 07:35:01 -0700
Received: from insanus.matematik.su.se (insanus.matematik.su.se [130.237.198.12])
          by freefall.cdrom.com (8.6.10/8.6.6) with ESMTP id HAA21389
          for <hackers@freebsd.org>; Thu, 15 Jun 1995 07:34:58 -0700
Received: from localhost (jocosus.matematik.su.se [130.237.198.7]) by insanus.matematik.su.se (8.6.10/8.6.9) with ESMTP id OAA04557; Thu, 15 Jun 1995 14:34:06 GMT
Message-Id: <199506151434.OAA04557@insanus.matematik.su.se>
X-Address: Department of Mathematics, Stockholm University 
	      S-106 91  Stockholm
	      SWEDEN
X-Phone: int+8162000
X-Fax:   int+86126717
X-Url:   http://www.matematik.su.se
To: hackers@freebsd.org
cc: sef@cygnus.com
Subject: Poor store performance with Pentium
Date: Thu, 15 Jun 1995 16:34:01 +0200
From: Torbjorn Granlund <tege@matematik.su.se>
Sender: hackers-owner@freebsd.org
Precedence: bulk

I am exploring the Pentium pipeline, in preparation for optimizing GCC for
Pentium.  I just ran into something very weird with the performance of loads
and stores.

The loop in below takes 28 cycles/iteration if READ is undefined, and the
expected 4 cycles if READ is defined.  (The nops are irrelevant, but are
needed in my accurate timing setup.)

The only explanation I can think of to the poor write performance is that
the cache is in write-through mode, not write-back mode.

Does FreeBSD 2.x have support for setting the Pentium data cache in
write-back mode?  If it doesn't, could somebody with a good understanding of
the relevant components in the kernel outline the issues involved, and what
I would have to modify in order to enable write-back.

Torbjorn


.text
        .align 2
.globl _f
        .type    _f,@function
_f:
        movl 4(%esp),%edx
        movl $99999999,%eax
        .align 2,0x90
L2:
        nop
        nop

#if READ
        movl    (%edx),%ecx
        movl    4(%edx),%ecx
        movl    8(%edx),%ecx
        movl    12(%edx),%ecx
#else
        movl    %ecx,(%edx)
        movl    %ecx,4(%edx)
        movl    %ecx,8(%edx)
        movl    %ecx,12(%edx)
#endif

        decl %eax
        jg L2
        ret
Lfe1:
        .size    _f,Lfe1-_f