We can probably turn this off at any time; we just have drastic performance issues when we do. The real work is in deciding which parts of the stdlib should be resilient and which should not, and then verifying that we don't leak any of the resilient details into the public SIL.

I agree. We should not turn this off until the performance is OK. Is there a ticket to track this work?

Well, we need to be able to start testing things before we can actually measure the performance, so we will still need a flag for this. But maybe that's covered by SR-268.

Removing -sil-serialize-all itself is a trivial change. Just remove it the relevant option in cmake/modules/AddSwift.cmake.
So testing can be done easily.
All I'm saying is that we should not remove the option per default in ToT until the performance issues are resolved.

rdar://24057969

I just talked to Slava. He will take this.

This is mostly done, only a few tests fail, so I'm going to close this.