Open Radar

 
Summary:
Closely related to rdar://22996431, if you use @autoclosure then you get nondeterministic runtime failures instead of deterministic runtime failures.

Steps to Reproduce:
This code:

struct A {} // not Equatable!

func failsNondeterministicallyAtRuntime<A>(@autoclosure a: () -> A, _ test: (A, A) -> Bool, @autoclosure _ b: () -> A) -> Bool {
	return test(a(), b())
}

func failsDeterministicallyAtRuntime<A>(a: A, _ test: (A, A) -> Bool, _ b: A) -> Bool {
	return test(a, b)
}

for i in 0..<1000000 {
	assert(failsNondeterministicallyAtRuntime([ A(), A() ], ==, [ A(), A() ]), "failed at comparison \(i)")
}



Expected Results:
should not compile (rdar://22996431) but since it does, it should fail deterministically at runtime


Actual Results:
but it’s totally nondeterministic.


Regression:
N/A


Notes:
I could be persuaded that this is undefined behaviour—after all, it _is_ selecting UnsafePointer equality. In which case rdar://22996431 is a bug but this is not.

However, both of these together are _much_ worse than rdar://22996431 alone, since it appears to succeed _most_ of the time. I don’t have the slightest idea how that could be.