FEniCS / ffcx / 35657512554

86%

main: 88%

push

github

jit: locate cached modules by path, not by scanning the cache directory (#887)

get_cached_module built an importlib FileFinder over the cache directory
and invalidated its cache before every lookup, so each cache hit listed
the whole directory. The cache holds four files per form ever compiled,
so the cost grows without bound as a user accumulates compiled forms:
with 24,196 files it is 9.8 ms per fem.form call against 0.13 ms on a
fresh cache, and the gap is a constant, independent of problem size.

The file name is fully determined by the module name, so probe the
candidate extension suffixes and load with spec_from_file_location
instead. This also removes the need for invalidate_caches(), since the
wait loop re-stats on each iteration and still picks up a module written
by a concurrent process.

Measured against the same 24,196-file cache: 9.81 ms -> 0.13 ms.

Closes #886.

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

12 of 15 new or added lines in 1 file covered. (80.0%)

4713 of 5450 relevant lines covered (86.48%)

0.86 hits per line