Date: Sun, 30 Mar 2003 00:27:29 +1100 (EST) From: Bruce Evans <bde@zeta.org.au> To: Dag-Erling =?iso-8859-1?q?Sm=F8rgrav?= <des@ofug.org> Cc: cvs-all@freebsd.org Subject: Re: Checksum/copy Message-ID: <20030329235643.H11074@gamplex.bde.org> In-Reply-To: <xzp3cl61hxs.fsf@flood.ping.uio.no> References: <Pine.BSF.4.21.0303260956250.27748-100000@root.org> <20030327232742.GA80113@wantadilla.lemis.com> <20030328073513.GA20464@cirb503493.alcatel.com.au> <xzp3cl61hxs.fsf@flood.ping.uio.no>
next in thread | previous in thread | raw e-mail | index | archive | help
On Sat, 29 Mar 2003, Dag-Erling [iso-8859-1] Smørgrav wrote:
> I've attached an untested patch that sets up the infrastructure for
> processor-specific pagezero() and pagecopy() on i386. At the very
> least, it helps avoid some of the #ifdef spaghetti in pmap.c, and it
> should also result in a slight speedup for page zeroing and copying.
I'm sorry I started this. My original reply was about why even larger
micro-optimizations are dubious.
> (the patch builds, but I haven't booted it)
>
> Note that this patch makes pmap_zero_page_area() identical to
> pmap_zero_page() on i386 - this was already the case for i686-class
> CPUs since they use i686_pagezero() for pmap_zero_page_area(). This
Er, they can't be identical. pmap_zero_page_area() handles partial
pages. i686_pagezero() was only used for the special case where the
partial page is actually the whole page.
> On a different note, support.s is a bloody mess. Once the dust has
> settled, I'd like to go through it and reorder its contents a little.
There is very little wrong with its order.
% Index: sys/sys/systm.h
% ===================================================================
% RCS file: /home/ncvs/src/sys/sys/systm.h,v
% retrieving revision 1.191
% diff -u -r1.191 systm.h
% --- sys/sys/systm.h 24 Mar 2003 21:15:35 -0000 1.191
% +++ sys/sys/systm.h 29 Mar 2003 12:10:28 -0000
% @@ -171,12 +171,9 @@
%
% void bcopy(const void *from, void *to, size_t len);
% void ovbcopy(const void *from, void *to, size_t len);
% -
% -#ifdef __i386__
% -extern void (*bzero)(void *buf, size_t len);
% -#else
% void bzero(void *buf, size_t len);
% -#endif
The microoptimization of making bzero a function pointer wasn't such a
good idea. The main problem with undoing it is that this breaks binary
compatibility.
% +void pagecopy(const void *from, void *to);
% +void pagezero(void *buf);
I'd prefer not to have more interfaces to combinatorially explode.
The versions of these in the patch are only generic, so using these
interfaces instead of bcopy() and bzero() bypasses non-generic
versions of the latter.
%
% void *memcpy(void *to, const void *from, size_t len);
The declarations of page*() are disordered ('p' > 'm').
% Index: sys/i386/i386/pmap.c
% ===================================================================
% RCS file: /home/ncvs/src/sys/i386/i386/pmap.c,v
% retrieving revision 1.398
% diff -u -r1.398 pmap.c
% --- sys/i386/i386/pmap.c 25 Mar 2003 00:07:02 -0000 1.398
% +++ sys/i386/i386/pmap.c 29 Mar 2003 11:59:36 -0000
% ...
% @@ -2789,12 +2784,7 @@
% #endif
% invlpg((u_int)CADDR2);
% #endif
% -#if defined(I686_CPU)
% - if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
% - i686_pagezero(CADDR2);
% - else
% -#endif
% - bzero((char *)CADDR2 + off, size);
% + pagezero(CADDR2);
See above about this bug.
% Index: sys/i386/i386/support.s
% ===================================================================
% RCS file: /home/ncvs/src/sys/i386/i386/support.s,v
% retrieving revision 1.93
% diff -u -r1.93 support.s
% --- sys/i386/i386/support.s 22 Sep 2002 04:45:20 -0000 1.93
% +++ sys/i386/i386/support.s 29 Mar 2003 12:08:54 -0000
% @@ -48,9 +48,15 @@
% .globl bcopy_vector
% bcopy_vector:
% .long generic_bcopy
% - .globl bzero
% -bzero:
% + .globl bzero_vector
% +bzero_vector:
% .long generic_bzero
% + .globl pagecopy_vector
% +pagecopy_vector:
% + .long generic_pagecopy
% + .globl pagezero_vector
% +pagezero_vector:
% + .long generic_pagezero
% .globl copyin_vector
% copyin_vector:
% .long generic_copyin
The definitions of page*() are disordered ('p' > 'c').
% Index: sys/i386/include/md_var.h
% ===================================================================
% RCS file: /home/ncvs/src/sys/i386/include/md_var.h,v
% retrieving revision 1.60
% diff -u -r1.60 md_var.h
% --- sys/i386/include/md_var.h 25 Mar 2003 00:07:03 -0000 1.60
% +++ sys/i386/include/md_var.h 29 Mar 2003 12:11:08 -0000
% @@ -36,12 +36,17 @@
% * Miscellaneous machine-dependent declarations.
% */
%
% +extern void (*bcopy_vector)(const void *from, void *to, size_t len);
% +extern void (*ovbcopy_vector)(const void *from, void *to, size_t len);
% +extern void (*bzero_vector)(void *buf, size_t len);
% +extern void (*pagecopy_vector)(const void *from, void *to);
% +extern void (*pagezero_vector)(void *buf);
% +extern int (*copyin_vector)(const void *udaddr, void *kaddr, size_t len);
% +extern int (*copyout_vector)(const void *kaddr, void *udaddr, size_t len);
This file was mostly sorted and consistently formatted.
Bruce
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20030329235643.H11074>