model-writtenrun 5370 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- The thread began with Andres Freund's announcement that extension libraries would hide symbols by default and explicitly export their public interfaces, improving loading, avoiding symbol conflicts, and simplifying Windows builds #1.
- Early build failures exposed a missing configure executable bit and an omitted commit during a local merge; Andres explained the mistakes, and Tom Lane explained why check-world had not caught the configure problem #3 #4.
- Pavel Stehule reported that hidden visibility broke plpgsql_check and identified seven required functions #6; Álvaro Herrera suggested an exports.txt file, but Tom Lane recommended PGDLLEXPORT annotations and raised the possibility of pldebugger needing additional exports #7 #8.
- Álvaro submitted the seven export annotations, reported passing plpgsql_check tests, and proposed separating public and internal declarations into different headers; he also discovered that pldebugger failed during startup because it requested shared memory outside shmem_request_hook #9.
- Pavel confirmed that the patch fixed plpgsql_check #11, while Andres favored handling header restructuring separately, reiterated that PGDLLEXPORT was sufficient, and checked other dynamic symbol consumers without finding further missing exports #12 #14 #16.
- Álvaro committed the export patch from #9 with only whitespace adjustments as 1caf915ff3, then confirmed that splitting plpgsql.h remained unplanned #17.
- After the commit, Tom suggested documenting the export annotations instead of splitting the header #18; Robert Haas questioned the Windows-sounding macro name, but Tom and Andres considered retaining it acceptable for dynamically loaded libraries generally #19 #21 #22 #23.
- Álvaro questioned whether plpgsql_plugin_ptr also needed exporting #20, but Tom's experiments established that pldebugger accesses it through find_rendezvous_variable and needs no explicit linkage; Tom instead attached an independent shared-memory-hook fix for pldebugger #26.
- Dave Page reported pushing Tom's pldebugger fix with compatibility conditionals for older PostgreSQL versions #27, leaving both reported extension failures addressed and the optional plpgsql header restructuring deferred.