On 18 October 2016 at 04:19, Andres Freund <andres@anarazel.de> wrote: > On 2016-10-17 16:16:37 -0400, Robert Haas wrote: >> I wouldn't think that cross-file references would be especially >> common. Functions that take PG_FUNCTION_ARGS and return Datum aren't >> a lot of fun to call from C. But maybe I'm wrong. > > There's a fair number of DirectFunctionCall$Ns over the backend.
It's only an issue if one contrib calls another contrib (or the core backend code calls into a contrib, but that's unlikely) via DirectFunctionCall .
If changed to use regular OidFunctionCall they'll go via the catalogs and be fine, albeit at a small performance penalty. In many cases that can be eliminated by caching the fmgr info and using FunctionCall directly, but it requires taking a lock on the function or registering for invalidations so it's often not worth it.
Personally I think it'd be cleaner to avoid one contrib referencing another's headers directly and we have the fmgr infrastructure to make it unnecessary. But I don't really think it's a problem that especially needs solving either.
Not all called functions has V1 interface. I understand so plpgsql_check is not usual extension and it is a exception, but there is lot of cross calls. I can use a technique used by Tom in last changes in python's extension, but I am not able do check in linux gcc.
