[maleadt]

YieldingSpinLock (used by Completion) defined Base.assert_havelock and islocked methods. On Julia 1.13+ that invalidates Base.assert_havelock(::GenericCondition), which Base compiled for conditions with an unknown lock type, plus everything depending on it through wait/notify. One visible consequence: every process loading GPUToolbox (and so GPUArrays, CUDA.jl, oneAPI.jl, ...) recompiles Base.Filesystem.temp_cleanup_atexit at exit.

Julia	invalidations before	after	julia -e 'using GPUToolbox' before	after
1.10	11	11	123 ms	122 ms
1.11	30	30	100 ms	101 ms
1.12	10	10	130 ms	131 ms
1.13	269	0	227 ms	121 ms
nightly	488	0	142 ms	143 ms

(Invalidations from SnoopCompile's @snoop_invalidations using GPUToolbox; timings are medians of 30 interleaved runs on a loaded machine. On 1.13, --trace-compile-timing shows temp_cleanup_atexit taking about 100 ms to recompile at exit.)

The fix: on 1.13+, YieldingSpinLock is a Threads.AbstractSpinLock with the owned field that Base's spin lock methods expect, and only overrides lock. That way it reuses Base's trylock, unlock, islocked and assert_havelock, and no new methods reach the poorly-inferred condition code. Earlier versions keep the current wrapper around SpinLock: there is no AbstractSpinLock, and the remaining invalidations are minor and don't cause the exit-time recompilation. The tests now also check that assert_havelock still throws when the lock isn't held.