Cyrus Harmon has been working on a port of the mostly-copying generational garbage collector in SBCL, based on previous work by Raymond Toy for CMUCL. At the seventh merge candidate, I decided I'd made him suffer enough with my tyrannical requests for whitespace changes and function renamings, so I merged it, having tested on multiple platforms, and went to sleep the sleep of the just.

Sadly, the testing wasn't quite exhaustive enough, and I discovered this morning that we'd managed to kill the overnight automatic build-and-benchmark host run by Andreas Fuchs. Fortunately Cyrus was awake (whether again or still I don't know; for him it was 0300 local time) and diagnosed the problem pretty quickly: we had managed to subvert the garbage collector almost, but not quite, completely on threaded platforms. Interestingly, the system was stable enough to build itself, but benchmarking it (or indeed running the regression tests) exercised different-enough allocation patterns that the existence of the error was pretty obvious.

A simple one-line fix later, and things started working on threaded platforms too.