Unify TLS destructor list implementations #116850
Conversation
|
r? @m-ou-se (rustbot has picked a reviewer for you, use r? to override) |
rustbot
added
O-hermit
This comment has been minimized.
This comment has been minimized.
|
The Miri subtree was changed cc @rust-lang/miri |
|
☔ The latest upstream changes (presumably #116402) made this pull request unmergeable. Please resolve the merge conflicts. |
joboet
force-pushed
the
unify_tls_dtor_lists
branch
from
9e13c63 to
abd6645
Compare
|
☔ The latest upstream changes (presumably #118126) made this pull request unmergeable. Please resolve the merge conflicts. |
joboet
force-pushed
the
unify_tls_dtor_lists
branch
from
abd6645 to
232a3c4
Compare
|
☔ The latest upstream changes (presumably #117873) made this pull request unmergeable. Please resolve the merge conflicts. |
joboet
force-pushed
the
unify_tls_dtor_lists
branch
from
232a3c4 to
4708bb2
Compare
|
☔ The latest upstream changes (presumably #117285) made this pull request unmergeable. Please resolve the merge conflicts. |
joboet
force-pushed
the
unify_tls_dtor_lists
branch
from
4708bb2 to
79f310a
Compare
|
☔ The latest upstream changes (presumably #120486) made this pull request unmergeable. Please resolve the merge conflicts. |
Now that the fallback code has been removed, there are no users of `StaticKey` left for targets with native TLS. Therefore, move the at-exit hack to the thread-local guard module, where it can be shared by both implementations, and cfg-out the key-based TLS when it's not needed.
joboet
force-pushed
the
unify_tls_dtor_lists
branch
from
79f310a to
5e1727c
Compare
|
☔ The latest upstream changes (presumably #121569) made this pull request unmergeable. Please resolve the merge conflicts. |
rustbot
added
S-waiting-on-author
|
Closing in favour of #126523. |
6 participants
Part of #110897.
Currently, nearly every platform implements its own TLS destructor list. This adds unnecessary code and means that issues like #116390 are harder to find and fix. Therefore, this PR unifies all of these implementations into a shared one inside
sys::common::fast_local, leaving behind only the platform calls needed to ensure the list is emptied.This changes behaviour on some Linux-like platforms, because we now keep our own list. This should however not impact performance too much, as the platform calls would also have needed to allocate.
Because it is only used on UNIX and is now much easier, I've removed the fallback implementation in
sys_common. BecauseStaticKeymay now be unused, this resulted in errors for Windows, which is why I have refactored that code to move the at-exit hack to the newthread_local_guardmodule. If it causes too much review pain, I can split this part out.Tested on aarch64 macOS, checked on all other platforms. Best reviewed per-commit.
@rustbot label +T-libs +A-thread-locals