Message-ID: <329CFFAC.4E03@gbrmpa.gov.au>
Date: Thu, 28 Nov 1996 10:57:51 +0800
From: Leath Muller <leathm AT gbrmpa DOT gov DOT au>
Reply-To: leathm AT gbrmpa DOT gov DOT au
Organization: Great Barrier Reef Marine Park Authority
MIME-Version: 1.0
To: Benjamin D Chambers <chambersb AT juno DOT com>
CC: djgpp AT delorie DOT com
Subject: Re: Optimization
References: <Pine DOT SUN DOT 3 DOT 90 DOT 961127125213 DOT 12832A-100000 AT coop10> <19961128 DOT 160329 DOT 4455 DOT 4 DOT chambersb AT juno DOT com>
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit

> ***** Actually, after several tests I ran, I found the best perfermance
> came from C code with algorithmic optimizations, -O3 used, and NO
> ASSEMBLY OPTIMIZATIONS.  At least with my code, GCC was able to figure
> out better ways of shuffling registers than I was.  I'll admit, this
> won't work with everyone though - you'll have to either profile with
> different compilations or run some other benchmarks on them to know for
> sure.

Generally speaking, DJGPP with -O3 is pretty good at optimization, 
except I found its not so good at mixing (naturally) FPU and integer
code (although the pentium optimizations patch would probably fix
this).
 
> >types etc. For example: if I don't really need 32bits worth of int,
> >will
> >things be faster if I declare my variables as short ints?
 
> ***** No!  Benchmarks show that a 486 is slowest with 8-bit data, about
> twice as fast with 16-bit data, and even faster with 32-bit data!  On the
> Pentium, the difference between 16-bit and 32-bit is even greater.  And a
> Pentium Pro actually runs 16-bit SLOWER than a Pentium, with the 32-bit
> code much faster than a Pentium.  AVOID 16-bit DATA!
> ***** Warning, though, if you get a 32-bit DWord aligned wrong, you COULD
> actually end up with code that's less than half the speed it should be.
> I *believe* (I'm not sure) the way to make sure it's aligned properly is
> to use
> _PACKED_.
> ^^^^^^^^ Somebody check me here.  I could be wrong...

I though DJGPP automatically aligned structures on 32 bit boundaries,
and you used the __attribute__ ((packed)) when you wanted otherwise...

Leathal.