Infer a missing method as returning Union{}

[martinholters]

Ref #20033 (comment) and 0b8f0b7#commitcomment-21033206.

[martinholters]

Derp. I really should get into the habit of enabling assertions.

[martinholters]

Huh, looks like [Union{}] triggers the assertion, while Any[Union{}] works.

[martinholters]

Looks like I found a reason for better inferring Any myself: #20869.
In the long run, I still think this PR is the way to go, but we need to fix #20869 first.

[StefanKarpinski]

@vtjnash: Jeff wants your opinion on this, but is "generally in favor".

[vtjnash]

As the author of this PR noted above, we can't do this until type-intersection is usually correct.

[martinholters]

OTOH, as long as type-intersection may give wrongly empty answers, things can easily break anyway; this PR only makes it even a little easier. Consider

function foo(...)
# ...
end

function bar(...)
    # ...
    foo(...)
end

Now if the inferred types for the arguments to foo in bar and the signature of foo have a wrongly empty intersection computed (i.e. the actual argument types determined at runtime match the signature), then on master, things are fine, while with this PR, calling bar will likely crash as foo unexpectedly returns a result instead of throwing. But now add

foo(...) = error("arguments must ... whatever")

with a signature more general than the first foo method to add a more descriptive error message instead of a MethodError. Now if only this second method has a non-empty intersection, then master will be equally prone to a crash as this PR.

I guess what I'm trying to say is: We really need to make type-intersection reliable; this PR only increases the pressure to do so.

[StefanKarpinski]

In a way, making the process more likely to crash in cases where inference is wrong is helpful for finding cases where it is wrong.

[martinholters]

Superseded by #21892.