From:	dahms AT ifk20 DOT mach DOT uni-karlsruhe DOT de
Subject:	Re: Maths ruined when optimizing in gcc
26 Nov 1996 14:45:12 -0800 :
Sender:	daemon AT cygnus DOT com
Approved:	cygnus DOT gnu-win32 AT cygnus DOT com
Distribution:	cygnus
Message-ID:	<009ABF70.EA7639E0.5267.cygnus.gnu-win32@ifk20.mach.uni-karlsruhe.de>
Original-To:	khan AT xraylith DOT wisc DOT edu
Original-CC:	gnu-win32 AT cygnus DOT com, dahms AT ifk20 DOT mach DOT uni-karlsruhe DOT de
Original-Sender:	owner-gnu-win32 AT cygnus DOT com

Hi Mumit, you wrote:

: Another data point here. I'm in the process of testing g77 port to
: cygwin32 (it was quite simple once I figured out the right flags to
: libfI77) and the big gotcha is the every time I use -O, some of the
: programs produce NaNs all over the place! Haven't been able to pin
: point where it happens yet, but hopefully soon.
:
: Anybody have a short program that screws up the numerics with -O? If
: so, please email it me (if fortran, then either original or f2c'd will
: do fine).

Yes, here it is:

--- snip here ---
#include <stdio.h>
int main(argc, argv)
int argc;
char *argv[];
{
        volatile char c = 10;
        volatile int i = c;
        float x = c / 5.0, y = i / 5.0;
        printf("%.2f %.2f\n", x, y);
        return 0;
}
--- snip here ---

with the following results:

--- snip here ---
bash$ gcc floatbug.c -o floatbug.exe
bash$  floatbug
2.00 2.00
bash$ gcc -O floatbug.c -o floatbug.exe
bash$ floatbug
Inf Inf
bash$
--- snip here ---

This is stripped down from my big program which also produced NaNs.
The original doesn't use volatile, but that prevents compile time evaluation.
Astonishingly this shows Inf (using b16), which is not any better!

BTW, I just joined the mailinglist, so I missed the start of the thread.


Bye, Heribert (dahms AT ifk20 DOT mach DOT uni-karlsruhe DOT de)
-
For help on using this list, send a message to
"gnu-win32-request AT cygnus DOT com" with one line of text: "help".

- Raw text -