[concurrency] Change #isolated to mask out the TBI bits of the witness pointer of the implicit isolated any Actor pointer so we can do optimizations on TBI supporting platforms in the future.

[gottesmm]

Specifically, this will let us perform an optimization in the future where a caller can set bits in the TBI byte to signal information dynamically to a nonisolated(nonsending) callee. The intent is that this will let us signal to a nonisolated(nonsending) callee that the caller isn't performing any isolated work before it hops (or potentially a caller of the caller) hops meaning that the nonisolated(nonsending) function can avoid a tail hop that restores isolation after calling some sort of actor isolated call.

E.x.:

nonisolated(nonsending) func callsActor(_ a: MyActor) async {
  await a.doSomething()
}

@concurrent func runOnBackground() async {}

actor A {
  func doSomething() async { ... }

  func caller() async {
    await callsActor(self)
    await runOnBackgroud()
  }
}

In the above case, the hop back in callsActor is not needed dynamically in caller since we are going to immediately await ontot he background. But we do not know that. An optimization can be written that can recognize this case and dynamically tell via the tbi bits of the protocol witness table of the any Actor call that the hop to executor call can avoid the hop.

Since this only applies to the implicit isolated any Actor parameter which is only used for the purpose of hopping, it is safe to do this since when we hop we just call the Actor unowned executor dispatch thunk which directly accesses the witness table by loading from that pointer... meaning that anything set in the TBI bit is irrelevent.

The only other way to grab this implicit bit is to use #isolation and that is where this commit comes into play. Since we can access the actor via #isolation, we need to ensure that we mask out the TBI bits so that the rest of the runtime/compiled code never see the TBI bits. This is because we do not want the fact that we are using these TBI bits to ever escape into other code (e.x.: consider things that may cast).

So we have basically created a hermetically sealed world where we can pass information from callee to caller using the protocol witness TBI bits.

rdar://156525771

[gottesmm]

@swift-ci smoke test

[DougGregor]

So we never see unowned here, or is it handled is part of this call?

[gottesmm]

It will never be unowned. We always know that it will be @guaranteed. I checked.

[DougGregor]

I would have expected this in the IR gen options

[gottesmm]

The reason I didn't put it into IRGen options is that I am using it in SILGen.

[gottesmm]

@DougGregor I was thinking about this. Really the whole feature should be in IRGen and just be a flag on an apply. That being said, I want to do that later on main.

[gottesmm]

@swift-ci smoke test

[ktoso]

AFAICS looking good 👍

[gottesmm]

@swift-ci smoke test

[DougGregor]

This part shouldn't be behind the feature flag, right? It's the bit stripping we will always need to do.

[DougGregor]

Why are we checking for TBI here? It seems like we effectively need to distinguish between places where TBI will strip the bits for us (e.g., we're dereferencing the pointer) vs those where we treat the pointer as a value and we need to strip the bits ourselves (e.g., when building a runtime data structure involving this witness table).

[jamieQ]

@gottesmm just out of curiosity, how did you perform all these changes to update the test expectations? were they automated in some manner?

[gottesmm]

I did it by hand partially and partially just using query-replace. I have a quick way of munging failing lit test lines into things that I can use to quickly iterate on tests.