|
My first attempt at SI-6992 was about having whitebox expansions first
typecheck against outerPt and only then verify that the result is compatible
with innerPt.
That was a nice try, but soon after it went live in 2.11.0-M8, we've got
multiple reports with problems - both shapeless and then in a week specs2
started having issues with their whitebox macros.
In shapeless, typecheck against outerPt screwed up type inference, which
was more or less fixable by explicit type annotations, so I decided to
wait a bit before jumping to conclusions.
However, in specs2 the problem was more insidious. After being typechecked
against outerPt, expansions were being implicitly converted to a type
that became incompatible with innerPt. This revealed a fatal flaw of the
implemented approach - if allowed to typecheck against outerPt first,
whitebox macros could never be robust.
Now realizing that "outerPt > innerPt" doesn't work, I nevertheless wasn't
looking forward to rolling that back to "innerPt > outerPt", because that
would revive SI-6992 and SI-8048 that are highly unintuitive, especially
the latter one.
Therefore, this commit combines the permissiveness of "... > innerPt"
approaches with the robustness of "innerPt > outerPt", introducing
"WildcardType > innerPt > outerPt".
|