From: leathm AT solwarra DOT gbrmpa DOT gov DOT au (Leath Muller)
Message-Id: <199703030209.MAA08035@solwarra.gbrmpa.gov.au>
Subject: Re: Netlib code [was Re: flops...]
To: jbennett AT ti DOT com
Date: Mon, 3 Mar 1997 12:09:05 +1000 (EST)
Cc: djgpp AT delorie DOT com
In-Reply-To: <5fd1a8$ag6$2@superb.csc.ti.com> from "Jesse Bennett" at Mar 2, 97 11:08:56 pm
Content-Type: text

> This is a very simple function (but also very important in numerical
> applications).  Understanding how to coerce GCC into producing near
> optimal code (without obfuscating the source) for the matrix
> multiplication problem would be very beneficial to my work, since the
> required "tricks" should be widely applicable to my code.  I would
> like to hear any further thoughts or ideas on this subject.

The thing I have found about the FPU code generated by gcc is nothing 
short of weird... :)  I am coding for the pentium, and as a result have
started converting my entire texture mapping routines to asm and doing the
optimizations myself because DJGPP does a lot of repetitive things. Look
at the code generated everytime you want to store an integer.

Question: Is there any way to let DJGPP know I am running in single
precision  mode, and stop doing all the crap it does everytime I wan't
to store an integer? It would save a lot of hassle. I don't have the code
on me, so this is from memory, but to do a fistp myself running in single
precision, I normally have something like:

	flds	_a;
	fmuls	_b;
	fadds	_c;
	fistpl	_d;

Whereas DJGPP goes:

	flds	_a;
	fmuls	_b;
	fadds	_c;
	fstcw	-4(%ebp);
	fnldcw	-8(%ebp);
	fistp	_d;
	fldcw	-4(%ebp);

etc etc, only with a _lot_ more crap in the middle. Basically, I want
to inform DJGPP _not_ to load and store the FPU control word information
as there is no need! Especially when I am doing a lot of fistp's, as this
just kills performance... 

Leathal.