Allow name-based extractors to be irrefutable

[martijnhoekstra]

Extractors (SLS 2.13 8.1.9) are objects with an unapply and/or unapplySeq methods that can be used as patterns in match statements.

Name-based extractors are extractors where the unapply method doesn't return Option but some other type, with a contract that it has (1) a method get returning the extracted thing and (2) a method isEmpty returning false if the extractor matches and true if it doesn't.

Prior to this change, having the static return type Some instead of Option for unapply indicated that the extractor is "irrefutable", meaning the compiler knows the pattern always matches, which it uses for reachability and exhaustiveness analysis. This PR adds two more forms of irrefutable extractors:

1. "Boolean Match" extractors, where unapply returns the singleton type true instead of Boolean
2. Name-based extractors, where isEmpty returns the singleton type false instead of Boolean

With these changes, using such extractors removes what would otherwise be spurious false positive exhaustivity warnings.

---

PR implementation notes:

Irrefutable name-based extractors on isEmpty: false

• Check/update spec wontfix in this PR. Neither name-based extractors nor irrefutable extractors are specced.
• Align with dotty
• more tests, both neg and pos
• check if this removes the necessity for List special casing
• add tests from Vector.unapplySeq is refutable bug#12240, which is the motivating bug

Dotty: https://github.com/lampepfl/dotty/blob/cda0a9d496cc8fd6c10cb4c90c0a521a2c6820df/compiler/src/dotty/tools/dotc/transform/patmat/Space.scala#L292-L320 : constant false isEmpty is irrefutable

List special casing: Will still be needed as is, probably won't be needed anymore if we can check if all arities of unapplySeq have cases.

spun out the added testcases in 12240 to their own bug in scala/bug#12252

Fixes scala/bug#12240
Fixes scala/bug#12254

[martijnhoekstra]

Supporting exhaustiveness/unused checking over multiple arities of unapplySeq extractors is quite a bit more complicated: if element arity exhaustiveness is checked at all yet, I would appreciate a pointer to it. Otherwise, the ambition level of covering understanding that case Vector(), case Vector(_) and case Vector(_, _, _*) together cover all of it is probably too high.

[martijnhoekstra]

now also fixes scala/bug#12254:

allow true as an extractor type, and make it irrefutable

[martijnhoekstra]

Oh, I didn't see that the : true change broke a neg test output -- throw new Exception now infers as a valid extractor type as subtype of Boolean

I -update-check'ed that, and for pushed the fixup.

Happy for review in either state. Whether the error message should complain about an extractor of Nothing not having a get method as a name-based extractor or not supporting 2 extracted members as a Boolean extractor probably doesn't really matter in the end.

[martijnhoekstra]

what's so weird is that obviously the isEmpty of UnapplySeqWrapper was constant folded, but I wasn't able to construct any other source where isEmpty was getting constant folded.

[dwijnand]

See my run/patmat-no-inline-isEmpty.scala test: it only occurs under separate compilation, that's why I compile twice. I didn't get into why that was true or if it should be true.

[martijnhoekstra]

Ah, I see! That's the darndest thing -- and scary too. Maybe it doesn't just apply to this situation.

[dwijnand]

Lukas wondered if the unapply could be constant folded too, so I added a similar test for that and it's not constant folded - so all is good.

[lrytz]

Just some minor suggestions.

[SethTisue]

🎉 👏