[parisc-linux] hppa-linux gcc-3.0.3: -O2 optimization does not handle fp args correctly

[parisc-linux] hppa-linux gcc-3.0.3: -O2 optimization does not handle fp args correctly

[Randolph Chung]

I was going to file this bug against gcc, but it looks like it's
just another instance of the fp relocation bug that has been
reported earlier..... Dave, Alan, any suggestions on how we should fix
this? 

The following program gives inconsistent output with -O1 vs -O2:

legolas[22:24] ~% gcc -O1 -o logtest logtest.c -lm
legolas[22:24] ~% ./logtest
float: nan
legolas[22:24] ~% gcc -O2 -o logtest logtest.c -lm
legolas[22:24] ~% ./logtest
float: -1.002099

The bug was first observed in the testsuite for 'mawk'. A simpler
testcase that illustrates the problem is:

#include <stdio.h>
#include <math.h>

void print_data(double d)
{
	fprintf(stderr, "float: %f\n", d);
}

int main(int argc, char **argv)
{
	double d = log(-8.0);
	print_data(d);
	return 0;
}

randolph
-- 
Debian Developer <tausq@debian.org>
http://www.TauSq.org/

Re: [parisc-linux] hppa-linux gcc-3.0.3: -O2 optimization does not handle fp args correctly

[John David Anglin]

> legolas[22:24] ~% gcc -O2 -o logtest logtest.c -lm
> legolas[22:24] ~% ./logtest
> float: -1.002099

This looks ok with my new build of 3.1:

dave@gsyprf11:~/gcc_test$ gcc -O2 -o logtest logtest.c -lm
dave@gsyprf11:~/gcc_test$ ./logtest
float: nan
dave@gsyprf11:~/gcc_test$ gcc -O1 -o logtest logtest.c -lm
dave@gsyprf11:~/gcc_test$ ./logtest
float: nan

The problem that needs fixing in the linker is passing floating
arguments to indirect functions.  The above program doesn't involve
an indirect call.

Dave
-- 
J. David Anglin                                  dave.anglin@nrc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6605)

Re: [parisc-linux] hppa-linux gcc-3.0.3: -O2 optimization does not handle fp args correctly

[Randolph Chung]

> The problem that needs fixing in the linker is passing floating
> arguments to indirect functions.  The above program doesn't involve
> an indirect call.

hrm... ok. any ideas when this was fixed? we can't ship 3.1 with the
upcoming debian release, so we'll need to backport the fix.

randolph
-- 
   @..@                                         http://www.TauSq.org/
  (----)
 ( >__< )
 ^^ ~~ ^^