model-writtenrun 5738 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Justin Pryzby opened the thread with a survey of buildfarm warnings, identified a likely dangling pointer in pg_basebackup, and proposed filtering known warnings to expose new ones #1.
- Tom Lane described his existing warning-filtering script and the difficulty of automating useful diagnoses, while Andres Freund explained that skink's unusual warnings reflected its use of -O1 for Valgrind runs #2 #3.
- Tom's renewed audit uncovered duplicate xml_is_well_formed symbols on AIX and noisy unresolved-symbol warnings on Solaris; he initially proposed renaming the core XML function and invited a Solaris linker fix #4.
- Tom found that xml2's extension script referenced the core function by its C name, making renaming problematic, and instead proposed deleting xml2's obsolete pre-9.1 compatibility implementation; Andrew Dunstan endorsed that deletion #5 #7.
- Thomas Munro investigated the Solaris warnings on OpenIndiana and identified an extern mapfile as a possible solution, but supplied no patch in this thread #8.
- Tom committed the XML compatibility-function deletion directly from the discussion as fd2abeb7c2, implementing the proposal in #5.
- Andrew pointed out that the buildfarm already offered a little-publicized check_warnings setting and command-line option, providing an existing mechanism for detecting warnings #9.
- Tom confirmed that pg_basebackup retained a filename pointer beyond its stack lifetime and proposed copying it; Robert Haas explained the archive-versus-member filename behavior and recommended copying every assignment while freeing old allocations #10 #11.
- Tom settled on centralizing filename ownership in progress_update_filename(), guarded by showprogress && verbose, and committed that discussion-derived fix as 62cb7427d1 #12.
- After the fix landed, Robert questioned how to test truncated progress filenames reliably, and Andres suggested simply enabling progress in existing tests to exercise the code for crashes #13 #14 #15.
- Robert's final attachment added progress options to several pg_basebackup tests, but his experiments showed that these additions did not distinguish the dangling-pointer bug from the fixed behavior; both candidate fixes had landed, while this test proposal and the Solaris linker issue remained unresolved in the supplied thread #16 #8.