[Distributed] Improve getting return type metadata for distributed invocations

[ktoso]

Sometimes we may get incomplete type metadata when using the swift_func_getReturnTypeInfo function, which then could result in a crash when trying to return a result of such type from a distributed func. This happens very rarely, only if the type metadata has never been fully loaded in the process, but can happen and then cause a crash during returning from a remote call distributed method.

I was not able to reproduce the problem in the test since we always end up "touching" the type, even though I tried to avoid that by doing the remote call directly 🤔

Resolves: rdar://141313340

[ktoso]

@swift-ci please test

[ktoso]

Unrelated failure: FAIL: Swift-validation(linux-x86_64) :: SILOptimizer/large_nested_array.swift.gyb (18915 of 18915)

[ktoso]

@swift-ci please test linux platform

[al45tair]

I was not able to reproduce the problem in the test since we always end up "touching" the type, even though I tried to avoid that by doing the remote call directly 🤔

Could we maybe split the test into two separate executables and actually do a remote call, say using a pipe between them? Would that help? It'd be better if we could reproduce the problem and prove the change fixes it, if that's possible.

[mikeash]

The issue is actually that the original metadata request always comes back complete, but requests from other threads may return early, so it needs a multithreaded test to work. We can avoid the need for separate processes by calling swift_func_getReturnTypeInfo directly. I think I've got a test that hammers it enough to reliably catch this problem.

[ktoso]

We managed to form a test which reproduces reliably, thank you @mikeash !

I also corrected the impl in one more way thanks to this test -- we cannot do non-blocking metadata requests as that's how we can get those zero sizes as well.

[ktoso]

@swift-ci please test