Fix #4720 and #4721: account for type inference being too smart

[Blaisorblade]

Fix #4720 and #4721: account for @smarter type inference. Some type arguments can be solved right away by looking at their bounds alone, so handle them gracefully.
#4720 was introduced by eeb5965, #4721 was earlier.

[smarter]

Is this comment still correct?

[Blaisorblade]

I can't make sense of "@pre param is not part of the constraint domain." if that's a precondition. What's the "constraint domain"?

Otherwise, I suspect yes.

Poly params that are known to be smaller or greater than param are not contained in the return bounds.

IIUC, this suggests that in [Y<:String, Z>:Int, W>:Z<:Y], nonParamBounds(Y) and nonParamBounds(Z) don't include W. But nonParamBounds(W) should mention Z and Y, so it might never have been accurate.

EDIT: my patch doesn't affect it, and neither does yours; this comment means, concretely, that we only account for boundsMap but not lowerMap or upperMap. How information is partitioned between those map is another matter entirely (and might be affected by your patch).

[smarter]

72d7bc4

Not sure this is necessary, but it's not obvious that having another error
downstream is guaranteed, so giving one right away sounds safer.

This isn't great, if there really are cases where this message is the only one, it's likely to be very confusing to users.

[Blaisorblade]

This isn't great, if there really are cases where this message is the only one, it's likely to be very confusing to users.

That's what I hoped to learn from you or @odersky. But I think that can't happen and I should drop the last commit. This is just inference, and we must check bounds also for type arguments written by users, or we have bigger issues, and hopefully that's checked by the same code. This is done by calling Checking.checkBounds in PostTyper, for both AppliedTypeTree and TypeApply.

So I'll drop last commit.

The resulting error is better, but it's designed for types written by the user, not synthesized by inference:

-- [E057] Type Mismatch Error: tests/neg/i4721a.scala:2:39 ---------------------
2 |  def main(args: Array[String]):Unit = m("1") // error
  |                                       ^
  |                 Type argument Int does not conform to upper bound String

Otherwise, I can propagate more info out of addToConstraint, or into it, but that might be invasive. And I'd have to assume it's safe to pretty-print types in Typer.

Maybe I can just detect if the type argument tree was inferred? With tree.pos.isSynthetic on the type argument, or otherwise?

[smarter]

Maybe I can just detect if the type argument tree was inferred? With tree.pos.isSynthetic on the type argument, or otherwise?

A synthetic tree can contain a user written tree. Use pos.isZeroExtent instead

[Blaisorblade]

@smarter neither isSynthetic nor isZeroExtent works just so — I pushed my current attempts to #4970. After asking Martin, I'd merge this PR to fix the crashes (which actual users did report!) and postpone improved reporting to a separate PR.