Cut the Gordian Knot: Don't widen unions to transparent

[odersky]

The idea is that some unions usually make more sense than others. For instance,
if Apply and Ident are case classes that extend Tree, it makes sense to
widen Apply | Ident to Tree. But it makes less sense to widen String | Int
to Matchable.

Making sense means: (1) Matches our intuitive understanding, and (2) choosing not to
widen would usually not cause errors.

To explain (2): In the Tree case it might well be that we define a given for Inv[Tree] for
invariant class Inv, and then we would not find that given for Inv[Apply | Ident].
But it's much less likely that we are looking for a given of type Inv[Any].

This commit does two things:

• add logic not to widen a union if the result is a product of only transparent traits or classes.
• treat Any, AnyVal, Object, and Matchable as transparent.

[odersky]

test performance please

[dottybot]

performance test scheduled: 1 job(s) in queue, 0 running.

[dottybot]

Performance test finished successfully:

Visit https://dotty-bench.epfl.ch/15642/ to see the changes.

Benchmarks is based on merging with main (79d9a6f)

[sjrd]

• treat Any, AnyVal, Object, and Matchable as transparent.

If that makes sense, then it probably makes sense to also treat js.Any and js.Object as transparent, when they are there:
https://github.com/lampepfl/dotty/blob/1fb83bda632c0ed059844273e0789daeac24af8e/compiler/src/dotty/tools/backend/sjs/JSDefinitions.scala#L45-L48

---

Unrelated to the above: I expect this change to have a significant impact on source compatibility. Although we never promise source compatibility, perhaps this one is strong enough that it would warrant a minor release?

[odersky]

@sjrd I agree it would need a minor release. It might need a SIP as well, even though so far we have not talked about type inference in the SIP committee, despite many changes and improvements that happened in Scala 2 and 3.

That said, I am positively surprised how little code broke. All the tests that failed were simply testing for type inference predictions that are now improved. Nothing else in our code base or in the community build broke. I would not have expected that when I set out to try this change.

[dwijnand]

LGTM (commetting to avoid it being merged)

[dwijnand]

@odersky happy to merge this now?

[odersky]

I am OK with merging. I'll give this another 24 hours in case someone wants to object.

[smarter]

I'm a bit wary of trying to solve our union inference issues by adding more special cases, in particular it might have the unintended consequence of encouraging people to add transparent in their class hierarchies to avoid widening. In any case, this change of behavior requires an update of the documentation since both https://github.com/lampepfl/dotty/blob/main/docs/_docs/reference/new-types/union-types.md an https://github.com/lampepfl/dotty/blob/main/docs/_docs/reference/new-types/union-types-spec.md refer to the current widening logic.

[odersky]

@smarter I updated the docs on union types. Do you want to take a look?

[som-snytt]

add transparent in their class hierarchies to avoid widening.

add transparent until it compiles. Is that not the FAQ?