------- Comment #6 from rguenther at suse dot de 2010-01-13 15:08 -------
Subject: Re: possible integer wrong code bug
On Wed, 13 Jan 2010, jakub at gcc dot gnu dot org wrote:
>
>
> ------- Comment #5 from jakub at gcc dot gnu dot org 2010-01-13 15:05 -------
> The bug is in add_double_with_sign it seems.
> 319int
> 320add_double_with_sign (unsigned HOST_WIDE_INT l1, HOST_WIDE_INT h1,
> 321 unsigned HOST_WIDE_INT l2, HOST_WIDE_INT h2,
> 322 unsigned HOST_WIDE_INT *lv, HOST_WIDE_INT *hv,
> 323 bool unsigned_p)
> 324{
> 325 unsigned HOST_WIDE_INT l;
> 326 HOST_WIDE_INT h;
> 327
> 328 l = l1 + l2;
> 329 h = h1 + h2 + (l < l1);
> 330
> 331 *lv = l;
> 332 *hv = h;
> 333
> 334 if (unsigned_p)
> 335 return (unsigned HOST_WIDE_INT) h < (unsigned HOST_WIDE_INT) h1;
> 336 else
> 337 return OVERFLOW_SUM_SIGN (h1, h2, h);
> 338}
> When called with say -1U, -1, -2U, -1, &x, &y, true it pretends no overflow
> happened. That is because we don't add just 2 numbers together, but 3 (also
> (l
> < l1)), and so -1U + -1U + 1 gives -1U, so h == h1. I believe it is wrong not
> just for unsigned_p, but also for !unsigned_p. The question is how to fix
> this
> as efficiently as possible.
No-undefined-overflow branch has
{
unsigned HOST_WIDE_INT l;
HOST_WIDE_INT h;
l = l1 + l2;
h = (HOST_WIDE_INT)((unsigned HOST_WIDE_INT)h1
+ (unsigned HOST_WIDE_INT)h2
+ (l < l1));
*lv = l;
*hv = h;
if (unsigned_p)
return ((unsigned HOST_WIDE_INT) h < (unsigned HOST_WIDE_INT) h1
|| (h == h1
&& l < l1));
else
return OVERFLOW_SUM_SIGN (h1, h2, h);
}
but I think that caused fallout somewhere else... (extra TREE_OVERFLOW
stuff, etc)
Richard.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42721
