mirror of
https://github.com/JuliaLang/julia.git
synced 2026-05-28 03:10:33 +08:00
d8d384278d
Fixes https://github.com/JuliaLang/julia/issues/53859#issuecomment-2027352004, which was actually fixed before in https://github.com/JuliaLang/julia/pull/52346, but https://github.com/JuliaLang/julia/pull/52750 undid that fix. This PR does not directly address #53859, because I can not reproduce it atm. --- The `@depot` resolution logic for `include()` files is adjusted as follows: 1. (new behavior) If the cache is not relocatable because of an absolute path, we ignore that path for the depot search. Recompilation will be triggered by `stale_cachefile()` if that absolute path does not exist. Previously this caused any `@depot` tags to be not resolved and so trigger recompilation. 2. (new behavior) If we can't find a depot for a relocatable path, we still replace it with the depot we found from other files. Recompilation will be triggered by `stale_cachefile()` because the resolved path does not exist. 3. (this behavior is kept) We require that relocatable paths all resolve to the same depot. 4. (new behavior) We no longer use the first matching depot for replacement, but instead we explicitly check that all resolve to the same depot. This has two reasons: - We want to scan all source files anyways in order to provide logs for 1. and 2. above, so the check is free. - It is possible that a depot might be missing source files. Assume that we have two depots on `DEPOT_PATH`, `depot_complete` and `depot_incomplete`. If `DEPOT_PATH=["depot_complete","depot_incomplete"]` then no recompilation shall happen, because `depot_complete` will be picked. If `DEPOT_PATH=["depot_incomplete","depot_complete"]` we trigger recompilation and hopefully a meaningful error about missing files is thrown. If we were to just select the first depot we find, then whether recompilation happens would depend on whether the first relocatable file resolves to `depot_complete` or `depot_incomplete`.