[compiler] Always use full version of RecordWrite builtin
When we know that the value in a write barrier is a map, we know that we are not going to have an old-to-new reference (maps are always in old generation). Therefore we also don't really need the generational barrier in RecordWrite. While this is technically correct, we don't gain much from this optimization. The inline and out-of-line generated code for the barrier is still the same as in all other cases. Which means that outside marking we don't even reach the RecordWrite builtin. Most write barrier executions happen outside incremental marking, hence performance of the incremental marking barrier isn't critical. This CL always uses the full RecordWrite builtin using a flag in order to allow for an easy revert. This CL is motivated by the shared heap work, which needs an additional always-on barrier in the future (similar to OLD_TO_NEW) to keep a OLD_TO_SHARED remembered set up-to-date. While maps are always in the old generation, they maybe by located in the shared heap. Bug: v8:11708 Change-Id: I71a6ded2547a0b2bbb9bbbd796dbcae0987b2232 Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3471854Reviewed-by: Michael Lippautz <mlippautz@chromium.org> Reviewed-by: Tobias Tebbi <tebbi@chromium.org> Commit-Queue: Dominik Inführ <dinfuehr@chromium.org> Cr-Commit-Position: refs/heads/main@{#79160}
Showing
Please
register
or
sign in
to comment