X-pop3-spooler:	POP3MAIL 2.1.0 b 4 980420 -bs-
From:	Yann Dupont <dupont AT unantes DOT univ-nantes DOT fr>
Date:	Wed, 2 Sep 1998 16:44:12 +0200 (CEST)
To:	strasbur AT chkw386 DOT ch DOT pwr DOT wroc DOT pl (Krzysztof Strasburger)
cc:	beastium-list AT Desk DOT nl
Subject:	Re:pgcc-1.1a - first impression
In-Reply-To:	<m0zEBKp-0000FVC@chkw386.ch.pwr.wroc.pl>
References:	<m0zEBKp-0000FVC AT chkw386 DOT ch DOT pwr DOT wroc DOT pl>
X-Mailer:	VM 6.43 under 20.4 "Emerald" XEmacs Lucid
Message-ID:	<13805.15183.120737.920512@amalthee.unantes.sciences.univ-nantes.fr>
Sender:	Marc Lehmann <pcg AT goof DOT com>
Status:	RO
X-Status:	A
Lines:	97

Krzysztof Strasburger writes:
 > Hi!
 
 > Here are execution times (only -mcpu=pentiumpro -march=pentiumpro different
 > than in previous example) averaged over 9 runs:
 > pgcc-1.0.3a:  -O2     -O3
 >               28.24s  28.95s
 > pgcc-1.1a     -O2     -O3     -Os
 >               28.45s  27.97s  28.12s
 > The performance win is not very big, but higher optimization options 
 > improve the performance for newer pgcc instead of degrading it.
 > The code optimized for size performs surprisingly well.
 > Krzysztof

Also tried pgcc1.1.a ...

The program I want to speed up is the interpreted language python. I'm 
using it on a pentium2 333. I use the built-in sys.pystone to measure
speed. (Not sure it's a good test, it's probably very specific... But, 
well... this is A test :-)

After lot of test I ended with a total test on various computers with
various optimisations :


-O3 -m486 -fomit-frame-pointer  (486)
-O3 -mpentiumpro -march=pentiumpro -fomit-frame-pointer (Ppro)
-O3 -mpentiumpro -march=pentiumpro -malign-double -mstack-align-double
-fomit-frame-pointer (Ppro2)
-O3 -mpentium -march=pentium -fomit-frame-pointer (Pentium)
-O3 -mamk6 -march=amdk6 -fomit-frame-pointer (K6)
-O3 -m386 -fomit-frame-pointer (386)

Sizes :
[root AT amalthee src]# ls -l --sort=size py*
-rwxr-xr-x   1 root     root       730546 Sep  2 15:00 pythonppro2
-rwxr-xr-x   1 root     root       725290 Sep  2 14:58 pythonppro
-rwxr-xr-x   1 root     root       722987 Sep  2 15:05 python486
-rwxr-xr-x   1 root     root       672971 Sep  2 15:02 pythonpentium
-rwxr-xr-x   1 root     root       669131 Sep  2 15:07 python386
-rwxr-xr-x   1 root     root       667723 Sep  2 14:56 pythonK6

I also tested the original python wich comes with redhat-5.1 
(compiled with gcc 2.7.2.3, unknown optimisation)

Results : (Bigger are better, Average of 3 runs are given)

(all machine are linux redhat-5.1 standard)

On a P2-333:

386	 : 4545
486	 : 4505
K6	 : 4385
ppro2	 : 4250
ppro	 : 4200
original : 4030

On a Ppro-200:

386	 : 2740
486	 : 2730
K6	 : 2650
ppro2	 : 2600
ppro	 : 2545
original : 2485

On a amdK6-233:

386	: 3425
pentium : 3360
486	: 3230
K6	: 3200
ppro	: 3190
ppro2	: 3000

On a Pentium100:

486	 : 1185
pentium	 : 1170
386	 : 1090
K6	 : 1075
ppro2	 : 1070
ppro	 : 1065


The results are *REALLY* surprising !!!! I'll try with another program 
(gzip ?) to see if they are consistents....

Any comments ???


Yann.
\|/ ____ \|/ Fac. des sciences de Nantes-Linux-Python-IPv6-ATM-CORBA-& +
"@'/ ,. \@"  Tel :(+33) [0]251125865(AM)[0]251125857(PM)[0]251127866(Fax)
/_| \__/ |_\ Yann DOT Dupont AT sciences DOT univ-nantes DOT fr
   \__U_/    http://www.unantes.univ-nantes.fr/~dupont

- Raw text -