... BTW, a bit of grepping shows that there are still significant
violations of this policy with respect to header-driven inclusion of
postgres_fe.h and c.h. Also, plpgsql is doing it in the unapproved
fashion. Cleaning these up will take a bit more work, since we'll have
to actually add #includes to some .c files that are currently relying on
those headers to get the core header. However, if I believe my own
argument, then that's a good thing and we'd better go do it.
Also, ecpglib.h seems like a complete mess: it's relying on symbols
like ENABLE_NLS but there's no certainty as to whether pg_config.h
has been included first. It won't have been in the case where this
header is read from an ecpg-generated .c file. So this header will
in fact be interpreted differently in ecpg-generated programs than
in ecpglib and ecpg itself. Maybe this is okay but it sure smells
like trouble waiting to happen. I have no desire to try to fix it
myself though.
regards, tom lane