Re: Solaris getopt_long and PostgreSQL

Dne 31.03.09 18:21, Tom Lane napsal(a):
> Zdenek Kotala <Zdenek.Kotala@Sun.COM> writes:
>> Yeah correct, I overlooked that optreset is also defined as a extern. 
>> There is updated patch.
> 
> On looking at this I still can't see how it's not broken.  You are
> effectively assuming that getopt_long.c must define those variables.

No I assuming that when we use getopt_long from libc then libc already 
has these variables defined. It must.

> But surely getopt_long.c should be assuming that getopt.c defines them.
> Aren't we likely to end up with the situation that everyone is
> extern'ing them?

Yeah, getopt_long assumes that optind, opterr and so on are already 
defined by getopt. If you look in current implementation in BDS you can 
see that there is getopt_internal() and getopt(), getopt_long are only 
wrapper.

http://www.openbsd.org/cgi-bin/cvsweb/src/lib/libc/stdlib/getopt_long.c?rev=1.23;content-type=text%2Fplain

The main problem what I see here is that getopt and getopt_long works 
together. Use one from system and one ported is not good idea. I think 
best solution is to port new BSD version into postgreSQL and use both 
function from libc version or ported versin. Mixing then is risky.

See also e.g. initdb.c, there is declaration of opterr...

> What appears to me to be happening is that Solaris' linker is failing to
> merge global variable definitions when it should (must) do so.  We need
> to find out why rather than solve it with a patch that will certainly
> break other platforms.

Linker cannot do it in general. Libc contains these variables and 
compiled code points to them. When you declare them again into your 
application you override them but only for your application, not for libc.

> If you can't come up with a real solution, we might have to do this
> with "#ifndef SOLARIS" or something similar.

I prefer another solution then this, but it is also possible.

Personally prefer to port new getopt_long. It is benefit also for win users.

    Zdenek