GCC bug when -ffast-math is set on armel

[Martin Guy <martinwguy@yahoo.it>]

Hi
  I thought I should mention this here too in case anyone else gets
bitten by this armel-specific bug.

   A GCC bug has just turned up that affects the arm-*-gnueabi
architecture in gcc-4.[123]. While debugging libvorbisenc (which
produces silent output files on armel), it turns out that when -O
-ffast-math are set, GCC can produce incorrect code for the max(x,y)
macro applied to floating point values.

  Thanks to Erik de Castro Lopo for some incisive debugging and for
producing a minimal example (should print "0 0" but doesn't)

/*
**	This file is in the Public Domain.
**
**	This program demonstrates a bug in the -ffast-math option of the gcc
**	armel compiler : gcc version 4.3.2 (Debian 4.3.2-1.1)
**
**	This works as expected:
**
**	> gcc -Wall -O3 gcc-test.c -o gcc-test && ./gcc-test
**	min :       0.0000    max :       0.0000
**
**	Compile with -ffast-math and things goes screwy.
**
**	> gcc -Wall -O3 -ffast-math gcc-test.c -o gcc-test && ./gcc-test
**	min :   99999.0000    max :       0.0000
*/

#include <stdio.h>

#define	COUNT 	10

#define test_max(x,y)   ((x) <  (y) ? (y) : (x))
#define test_min(x,y)   ((x) >  (y) ? (y) : (x))

int
main (void)
{	/* C Standard says static data gets initialized to zero. */
	static float data [COUNT] ;
	float max = -99999.0, min = 99999.0 ;
	int k ;

	for (k = 0 ; k < COUNT ; k++)
	{	max = test_max (max, data [k]) ;
		min = test_min (min, data [k]) ;
		} ;

	printf ("min : %12.4f    max : %12.4f\n", min, max) ;

	return 0 ;
}

Full details at http://bugs.debian.org/515949

     M