Stephen R. van den Berg wrote:
Stephen R. van den Berg wrote:
Agreed. The trouble is finding them. The problems are triggered by destruct-races, and therefore hard to reproduce under controlled circumstances.
I might have found something, in the Shuffler. The error (if it actually is what we're looking for) was already present in Pike 8.0. But since the Shuffler code in 8.0 and older is hopelessly broken in other ways as well, it's non-critical to backport only this (unless you'd backport *all* Shuffler fixes).
I test-applied a commit here locally, and will let it run tonight. If it still hasn't coredumped in the morning, I'll push it to master tomorrow.
Well, this experiment failed. I took out my reverts locally, and now changed the diagnostic Pike_error() in case of the already-destructed call into a Pike_fatal(), in hopes of finding out when it is being triggered from the next coredump.