Alvaro Herrera <alvherre@2ndquadrant.com> writes:
> Tom Lane wrote:
>> Looking again at the spoonbill failure, it's striking that it gets as
>> far as plpython before choking. The symptoms are consistent with the
>> theory that USE_INLINE is getting #define'd by some Python header or
>> other, after having *not* been defined by pg_config.h. Maybe we should
>> rename that to PG_USE_INLINE?
> Ahh, that makes a lot of sense, yes; and I see that the Python headers
> in my machine do mention USE_INLINE (though this is not defined, even on
> pyconfig.h), so it is indeed possible that on some other machines it is
> defined somewhere in those headers. I will go rename ours. Thanks for
> the research.
I'd not tried a build with plpython on my old HPUX box recently, but
looking into its Python headers I find
$ grep -r USE_INLINE /usr/local/include/python2.5
/usr/local/include/python2.5/pyport.h:#undef USE_INLINE /* XXX - set via configure? */
/usr/local/include/python2.5/pyport.h:#elif defined(USE_INLINE)
which means on this box things would have failed in the opposite
direction (undefined references to non-provided external functions)
without the rename. So that's the right fix for sure.
>> It's (probably) not contributing to this particular breakage, but I
>> can't say that I think undef'ing things like _POSIX_C_SOURCE midstream
>> is a bright idea at all. Seems to me that could easily result in
>> inconsistent results from reading different system header files.
> Ugh.
The more I think about that, the more worried I get. It's not just
that _POSIX_C_SOURCE might be defined or not, it's that the *value*
it has affects what definitions you get. At least that's how it
looks on my box. So our existing code potentially allows Python.h
to redefine it with a different value than what we'd defined it as,
which could result in inconsistencies between the system headers
included by postgres.h and those included later. I think this needs
reconsideration. Not sure what to do about it exactly, though :-(
regards, tom lane
