Date: Tue, 23 Sep 1997 19:31:29 +0800 (GMT)
From: Orlando Andico <orly AT dilnet DOT upd DOT edu DOT ph>
To: firewind <firewind AT metroid DOT dyn DOT ml DOT org>
cc: djgpp AT delorie DOT com
Subject: DJGPP and optimized code
In-Reply-To: <604rks$fj2@sjx-ixn9.ix.netcom.com>
Message-ID: <Pine.SGI.3.96.970923192505.21675A-100000@gibson.eee.upd.edu.ph>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
Precedence: bulk

On 22 Sep 1997, firewind wrote:

..
> DJGPP does not and will -never- add things not already present in the stock
> gcc.
..
> How about we fart in their general direction anyway? If you think MSVC is
> great, by all means, switch. Personally, I'll stick with gcc.

I agree that in general, sticking with gcc is smarter. But will DJGPP
_ever_ adopt anything from egcs? I know it's experimental, but the
performance improvements are quite impressive (I've attached a simple
benchmark that I've made).

There is a DJGPP binary on www.goof.com; is this ever going to be part of
the standard distribution? anyway, below are the results of the
CPU-intensive test I did.


---

Test procedure:

The program used was the publicly available ISO MPEG layer 1, 2, and 3
decoder (CD 11172-5 of April 1994) with modifications by Fraunhofer-IIS
(version 4.1).

A 32-kbps, 32kHz single-channel MPEG-1 layer II file was decoded, and the
wave output dumped to /dev/null. Time consumed by this procedure was
measured with the `time' command.

All tests were performed on an unloaded Red Hat Linux 4.2 system (kernel
2.0.30, Pentium-optimized) with a 100-MHz Pentium processor and 24MB of
core. The test is not I/O intensive because the MPEG source file is small
(<1MB) and output was directed to the null device.

Both 486- and Pentium-optimized binaries were dynamic ELF executables, and
used the system's 486-optimized libc and libm. So it is possible to get
even better improvements by recompiling the libraries with Pentium
optimizations.

In summary, egcs had the following speed improvements over gcc:

 user     +37.7%
 system   +11.9%
 elapsed  +37.3%


---

gcc version 2.7.2.1
CFLAGS="-m486 -ffast-math -fomit-frame-pointer -O2"
time ./decode486 -A /opt/share/archive/mpeg2/pieces-of-you.mp2 /dev/null

> Avg slots/frame = 144.000; b/smp = 1.00; br = 32.000 kbps
> Decoding of "/opt/share/archive/mpeg2/pieces-of-you.mp2" is finished
> The decoded PCM output file name is "/dev/null"
> "/dev/null" has been written with AIFF header information
> 129.00user 0.89system 2:10.00elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
> 0inputs+0outputs (93major+31minor)pagefaults 0swaps


gcc version egcs-2.90.06p 970907 (gcc2-970802 experimental)
CFLAGS="-mpentium -ffast-math -fomit-frame-pointer -O6"
time ./decode586 -A /opt/share/archive/mpeg2/pieces-of-you.mp2 /dev/null

> Avg slots/frame = 144.000; b/smp = 1.00; br = 32.000 kbps
> Decoding of "/opt/share/archive/mpeg2/pieces-of-you.mp2" is finished
> The decoded PCM output file name is "/dev/null"
> "/dev/null" has been written with AIFF header information
> 93.70user 0.75system 1:34.67elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
> 0inputs+0outputs (102major+32minor)pagefaults 0swaps