Date: Mon, 07 Jan 2002 13:14:02 -0800 (PST) From: John Baldwin <jhb@FreeBSD.org> To: Stephen Montgomery-Smith <stephen@math.missouri.edu> Cc: freebsd-hackers@FreeBSD.org, "Matthew D. Fuller" <fullermd@over-yonder.net> Subject: Re: Tell gcc I have a i686 Message-ID: <XFMail.020107131402.jhb@FreeBSD.org> In-Reply-To: <3C3A0674.5BD4A62C@math.missouri.edu>
next in thread | previous in thread | raw e-mail | index | archive | help
On 07-Jan-02 Stephen Montgomery-Smith wrote:
> John Baldwin wrote:
>>
>> On 07-Jan-02 Stephen Montgomery-Smith wrote:
>> > John Baldwin wrote:
>> >>
>> >> > You know, I have no idea. It is someone elses code. These are the
>> >> > instructions. Can anyone tell me?
>> >> >
>> >> > "movl 32(%0),%1\n"
>> >> > "adcl %1,32(%0)\n"
>> >> >
>> >> > Also, from this discussion, what I have decided to do is provide it as
>> >> > an option for the user to add by editing the Makefile - not to do it
>> >> > automatically.
>> >>
>> >> These instructions are 386 instructions. What we need to see are the
>> >> contraints (the stuff after the actual instructions with colons in them)
>> >> to
>> >> see
>> >> if it is somehow using Pentium Pro+ specific registers. And actually,
>> >> just
>> >> for
>> >> the record, a PPro is a 686. :)
>> >>
>> >
>> > OK, this is it in context:
>> >
>> > register Word32 *_x = x;
>> > register int _a = 0;
>> >
>> > asm("xorl %1,%1\n" /* clear C */
>> > "movl 124(%0),%1\n"
>> > "adcl %1,124(%0)\n"
>> > : : "r" (_x), "r" (_a)
>> > );
>>
>> Looks like rather silly code to double the value at x + 124. I say silly
>> casue
>> it clears carry and then does a addcl. However, since CF is zero, this is
>> the
>> same as doing an addl. Since it is just doubling the value, a shl would
>> make
>> more sense (and only be 1 instruction.)
>>
>
> More likely I am butchering the code by trying to only give you part of
> it. The whole code is designed to take N 32 bit words and do a left
> shift on it:
Ok. This isn't 686 specific at all. However, one optimization might be to
get rid of the xorl, and use 'addl' for the first instruction instead of adcl.
Anyways, If I were you, I would do it via a series of rcl (rotate through carry
left, it shits in the carry flag instead of 0), thus I would do:
"shl (%0),1\n"
#if (N >= 2)
"rcl 4(%0),1\n"
#endif
etc. using a single rcl for the rest of the shifts.
Using multiple instructions might be useful if you did multiple loads to
different destination registers by interleaving movl's and adcl's, and
alternating temporary registers to try and use the multiple pipelines on
Pentiums and better. Then again, these instructions might not work well for
that. *shrug* Or perhaps if it uses suitable temporary registers, this might
be doing all the loads in one pipeline and the adcl's in the others to get a
jumpstart on each mov. I dunno. :-P
--
John Baldwin <jhb@FreeBSD.org> <>< http://www.FreeBSD.org/~jhb/
"Power Users Use the Power to Serve!" - http://www.FreeBSD.org/
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-hackers" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?XFMail.020107131402.jhb>