Re: [Patch, Fortran, OOP] PR 59654: [4.8/4.9 Regression] Broken function table with complex OO use case

[Janus Weil <janus at gcc dot gnu dot org>]

Hi Mikael,

>> I have just committed as obvious a one-line-removal patch which fixes
>> a wrong-code OOP regression:
>>
>> http://gcc.gnu.org/viewcvs/gcc?view=revision&revision=206281
>>
> Sometimes computer programming is more witchcraft than (computer) science.
> You add some code, it fixes a bug.
> Now you remove it, it also fixes this one.
> Huh?

well, if you look at the behavior of the test cases in the PR, there
is definitely some sort of black magic involved here ;)

Anyway, removing unneeded code is a pretty neat way to fix a bug. (I
checked that removing the code does not re-introduce the original
problem, for which it was inserted back then.)

To be honest, the vtab generation was never fully systematic, but
mostly followed the principle that we just generate one whenever we
need one (of course always checking that it isn't present already).


>> Since the regression also affects 4.8, I would like to backport the
>> patch (after waiting a few days and letting the reporter try the fix
>> on his original code). Ok if no problems appear on trunk?
>>
> OK with me. I would be delighted to understand why it works.

AFAICS, the problem was somehow related to the fact that we were
apparently building two instances of the vtab symbol for the derived
type 'TestResult'. But I have to admit that I did not investigate in
detail how that could lead to the observed behavior.

Possibly it was combined with some other problem, or at least
obfuscated by the delicacies of module loading and/or resolution order
etc.

In any case, we apparently end up with two versions of the vtab, and
the one that is being used probably has not been properly initialized
or something like that.

Cheers,
Janus