Re: GCC 3.0.1 Status

[scott snyder <snyder at fnal dot gov>]

Nathan Sidwell <nathan@codesourcery.com> writes:
> within a virtual base. This did not appear to be a regression from
> 2.95 (though 2.95 might not have seg faulted, it would not have
> produced correct layout). I did not investigate further.
>
> Scott, if you've any particular test case which does not fall into
> the above mentioned category, can you file it please? If you're not sure
> whether your test case is one or the other, just file it. Although
> 3145 is great at stressing the algorithm, it is pants when it comes
> to diagnosing a failure.

hi -

Well, the situation is that there is valid C++ code which executed
correctly when compiled with older versions of the compiler but fails
when compiled with the current version.  Now, it may have only been
fortuitous that it worked previously (i.e., the compiler isn't really
doing the right thing, but for this particular code, it didn't make a
difference in the execution), but from a user's point of view, this
looks like a regression --- stuff which used to work doesn't work with
the new compiler.

So, even though the underlying bugs are longstanding, for some reason
they didn't really cause problems before, but now they do
(At least for the particular applications i'm trying to build.)
[Something changed in the couple months before the release that made
this difference.  I had orbacus built and working happily with the 3.0
branch as of a few months ago (i don't remember now exactly when i built
it).  Then, a couple weeks before the 3.0 release, there were some
ABI changes, in libstdc++ i think, that forced me to recompile everything.
That's when i started running into these virtual derivation problems.]

I can go back to the orbacus code and distill another test case from that,
that i'm actually running into.  Every time i've done that in the
past, though, that specific case gets fixed, but when i go back to the
full program, i run into some other case that causes problems.  Since
isolating these test cases takes a while (it's a big hairy piece of code
that i didn't write and don't understand very well), i really liked
the approach used in c++/3145.

But i can certainly extract another testcase, if that would be helpful.
I would like to see this get fixed, since this is currently blocking
me from continuing with the tests i was doing with building our code
with gcc.  (It would also be useful for me to know if this is _not_
likely to be fixed soon --- in which case i'd either have a go at it
myself or try to come up with some other workaround.)

thanks,
sss