Bug#5367: gcc -O2 fails for INT_MIN < long long < INT_MAX

[Herbert Thielen <Herbert.Thielen@lpr.e-technik.tu-muenchen.de>]

Package: gcc
Version: 2.7.2.1-1

A colleague of mine found that the gcc optimizer for ix86 fails for

	long long a;
	(a < MIN_INT || a > MAX_INT) : foo ? bar

You can reproduce the bug with the program at the end of the mail. This
bug was exploited while trying to use a cross-gdb for a 64-bit-processor
on Debian/Linux.

gcc 2.7.2-8 showed this bug too.
The bug was _not_ reproducable on a gcc 2.7.2.1 for
mips-mips-riscos4sysv, so it seem's to be limited on the ix86 optimizer.


Regards
	Herbert.

Program to reproduce the bug (o.k. with -O0, buggy with -O1, -O2, -O3):

-snip-snap-snip-snap-

#include <stdio.h>
#include <limits.h>

#define	LARGE	   2000000000

#define CHECK_FALSE(expr)	\
		printf("%s: " #expr "\n", \
		(expr) ? "BUG, assumed true" : \
			 "o.k., it is false")

void foo(long long val)
{
	printf( "val=%Ld (%#Lx)\n"
		"INT_MIN=%d (%#x)\n"
		"INT_MAX=%d (%#x)\n"
		"LARGE=%d (%#x)\n\n",
		val, val, INT_MIN, INT_MIN, INT_MAX, INT_MAX, LARGE, LARGE);

	CHECK_FALSE(val < INT_MIN);
	CHECK_FALSE(val > INT_MAX);
	CHECK_FALSE(val < INT_MIN || val > INT_MAX);	/* Buggy with -O2 !! */
	CHECK_FALSE(val < -LARGE || val > LARGE);
	CHECK_FALSE(!(val > INT_MIN && val < INT_MAX));
	return;
}

main()
{
	long long	a;

	a = 5LL;
	foo(a);
	exit(0);
}

-snip-snap-snip-snap-

--
TO UNSUBSCRIBE FROM THIS MAILING LIST: e-mail the word "unsubscribe" to
debian-devel-REQUEST@lists.debian.org . Trouble? e-mail to Bruce@Pixar.com