Re: [4.1 Committed] PR middle-end/26729: Fold vs. TREE_OVERFLOW

["Richard Guenther" <richard dot guenther at gmail dot com>]

On 5/14/06, Roger Sayle <roger@eyesopen.com> wrote:
> Hi Richard,
>
> On Sun, 14 May 2006, Richard Guenther wrote:
> > I don't see how this avoids the problem in general.  In fact,
> > integer_nonzerop will return false for an integer != 0 with
> > TREE_CONSTANT_OVERFLOW
> > set - so it will fail the same way, just for different overflowed constants.
> >
> > Or am I missing something?
> >
> > At least it looks like you would only get bogus warnings in this case and
> > not wrong-code.
>
> It's the other way around.  The bogus warning and the invalid code are
> caused by the same hunk.  Previously, the TREE_OVERFLOW caused fold-const
> to incorrectly determine that the condition was always a constant,
> which would issue a diagnostic and then misoptimize the expression.
> Now that the sense of TREE_OVERFLOW has been reversed in this test,
> we no longer enter the problematic clause.
>
> You'll notice that the new testcase, has two passes on 4.1 now, but had
> two unexpected failures previously (one for the unexpected warning during
> compilation, and the other for the abort() at run-time).
>
> Hopefully this clears up your confusion.

Doh, yes.  I looked up more context from the sources and it's all clear
now.  Now the only thing left is if you did look at the 14 other places
!integer_zerop is used in fold_const.c?  At least a bunch of them look
suspicious from a quick look (I'll have a closer look tomorrow and try to
produce some testcases that fail).

Thanks,
Richard.