Attempt to fix completion crash

[odersky]

Fixes #16228 (hopefully)

Note: I could not test this. Maybe somebody else take can this over.

Looking at the stack trace in #16228:

at dotty.tools.dotc.core.TypeErasure.dotty$tools$dotc$core$TypeErasure$$sigName(TypeErasure.scala:815)
	at dotty.tools.dotc.core.TypeErasure$.sigName(TypeErasure.scala:204)
	at dotty.tools.dotc.core.Signature.$anonfun$2(Signature.scala:111)
	at scala.collection.immutable.List.map(List.scala:246)
	at dotty.tools.dotc.core.Signature.prependTermParams(Signature.scala:111)
	at dotty.tools.dotc.core.Types$MethodOrPoly.computeSignature$2(Types.scala:3565)
	at dotty.tools.dotc.core.Types$MethodOrPoly.signature(Types.scala:3582)
	at dotty.tools.dotc.core.Types$MethodOrPoly.computeSignature$2(Types.scala:3557)
	at dotty.tools.dotc.core.Types$MethodOrPoly.signature(Types.scala:3582)
	at dotty.tools.dotc.core.Denotations$SingleDenotation.signature(Denotations.scala:615)
	at dotty.tools.dotc.core.Denotations$SingleDenotation.signature(Denotations.scala:605)
	at dotty.tools.dotc.core.Denotations$SingleDenotation.atSignature(Denotations.scala:647)
	at dotty.tools.dotc.core.Denotations$SingleDenotation.atSignature(Denotations.scala:645)
	at dotty.tools.dotc.core.tasty.TreeUnpickler$TreeReader.readLengthTerm$1(TreeUnpickler.scala:1299)

I think what happens is that we need to disambiguate an overloaded reference by looking at the signature of each overloaded alternative. But one of the signatures refers to a type that's not on the classpath, which causes the MissingType exception. The fix is to catch the exception and declare the alternative that caused it to be non-matching.

About the regression: I have no idea what previous fix could have caused this.

[prolativ]

This change seems to prevent the crash but at the same time this introduces a further regression - up until now the completion worked if a user added some prefix before triggering code completion given that this excluded the missing class from the completion suggestions. I'll see if I manage to somehow bisect the regression

[odersky]

This change seems to prevent the crash but at the same time this introduces a further regression - up until now the completion worked if a user added some prefix before triggering code completion given that this excluded the missing class from the completion suggestions. I'll see if I manage to somehow bisect the regression

That's weird. I am not sure how such behavior could be explained by the fix in this PR. The fix just filters out alternatives that are referring to missing classes.

[prolativ]

Interestingly, when I tried this now, I could not reproduce the bug that this fix was supposed to introduce anymore, so I think we're good to go