X-pop3-spooler:	POP3MAIL 2.1.0 b 4 980420 -bs-
Message-ID:	<19980708045815.20996@cerebro.laendle>
Date:	Wed, 8 Jul 1998 04:58:15 +0200
From:	Marc Lehmann <pcg AT goof DOT com>
To:	beastium <beastium-list AT Desk DOT nl>
Subject:	please benchmark / MMX #2
Mail-Followup-To:	beastium <beastium-list AT desk DOT nl>
Mime-Version:	1.0
X-Operating-System:	Linux version 2.1.108 (root AT cerebro) (gcc version pgcc-2.91.43 19980628 (gcc2 ss-980502 experimental))
Status:	RO
Lines:	37

Ok, I revised the patch a bit.. due to my dumbness, gcc actually generated a
memory access for every mmx->reg transfer *sigh*

Could people benchmark again? I'm interested in P-II, Pentium, AMD and Cyrix
results. At ftp.goof.com:/pub/pcg/test/ you can find three binaries this
time:

binary		speedup on pii	cflags used

bzip2-nommx	-		-static -O6 -funroll-all-loops -mno-mmx
bzip2-mmx	5%		-static -O6 -funroll-all-loops -mmmx
bzip2-mmxonly	7.5%		-static -O6 -funroll-all-loops -mmmx-only

(please note that mmx-only is not only the smallest, it is also slightly
broken, since mmx-only doesn't care for the fpu in any way. Also, the sole
reason why I used bzip2 is that gzip doesn't compile with mmx yet, and its a
single file)

I think there is plenty of room for tuning in it, there is no handling of HImode
&c...

Maybe its time to put this into pgcc and announce the first mmx-capable
compiler... bah.

The -mmx switch could profit even more from the same optimization that would
enable register parameters.

Thanks!

      -----==-                                              |
      ----==-- _                                            |
      ---==---(_)__  __ ____  __       Marc Lehmann       +--
      --==---/ / _ \/ // /\ \/ /       pcg AT goof DOT com       |e|
      -=====/_/_//_/\_,_/ /_/\_\                          --+
    The choice of a GNU generation                        |
                                                          |

- Raw text -