Re: longjmp clobber warnings are utterly broken in modern gcc

Andres Freund <andres(at)2ndquadrant(dot)com> writes:
> On 2015-01-25 14:02:47 -0500, Tom Lane wrote:
>> I've been looking for other instances of the problem Mark Wilding
>> pointed out, about missing "volatile" markers on variables that
>> are modified in PG_TRY blocks and then used in the PG_CATCH stanzas.
>> There definitely are some. Current gcc versions do not warn about that.

> I think it's actually not a recent regression - in the past a lot of
> spurious instances of these warnings have been fixed by simply tacking
> on volatile on variables that didn't actually need it.

Yeah, it's not. For years and years I just automatically stuck a "volatile"
on anything gcc 2.95.3 complained about, so that's why there's so many
volatiles there now. But I've not done that lately, and comparing what
2.95.3 warns about now with what a modern version says with -Wclobbered,
it's clear that it's pretty much the same broken (and perhaps slightly
machine-dependent) algorithm :-(

>> This is scary as hell. I intend to go around and manually audit
>> every single PG_TRY in the current source code, but that is obviously
>> not a long-term solution. Anybody have an idea about how we might
>> get trustworthy mechanical detection of this type of situation?

> Not really, except convincing gcc to fix the inaccurate detection. Given
> that there've been bugs open about this (IIRC one from you even) for
> years I'm not holding my breath.

I've completed the audit, and there were a total of only five places
that need fixes (including the two I already patched over the weekend).
It's mostly pretty new code too, which probably explains why we don't
already have field reports of problems.

Interestingly, plpython seems heavily *over* volatilized. Not sure
whether to take some out there for consistency, or just leave it alone.

regards, tom lane