Different threads check this flag multiple times per function, and
currently all of them synchronize on a single mutex. It's not even a
reader-writer-lock, hence they might block each other just for checking
whether an error has been set.
Threads don't rely on precise information here, this is just a check to
abort early if compilation failed anyway. Also in the current
implementation, no ordering is guaranteed on this error field.

We can avoid taking the mutex by turning the field into an atomic
pointer. It will be updated at most once, from nullptr to the first
error detected. To check whether an error is set, we can even use
relaxed memory order, since we won't look into the object behind the
pointer.

R=titzer@chromium.org

Bug: v8:8423
Change-Id: I71354c8d463a57c219eb21e53136556ae787ebd4
Cq-Include-Trybots: luci.v8.try:v8_linux64_tsan_rel
Reviewed-on: https://chromium-review.googlesource.com/c/1375661
Commit-Queue: Clemens Hammacher <clemensh@chromium.org>
Reviewed-by: Ben Titzer <titzer@chromium.org>
Cr-Commit-Position: refs/heads/master@{#58220}