Boolean literal types based filtering

[alexarchambault]

Currently, the filter methods on HLists and Coproducts filter elements of a given type only.
Wouldn't it make more sense to have a filtering based on value types - Boolean literals - returned by a Poly? This way it would

• look closer to the filter method of standard collections (like map is),
• allow more complex filtering (filtering more than one type, or filtering based on value types)
• be closer to a filterKeys method that could be added for records and unions - as keys have value types, it doesn't make really sense to filter them only on their types (if it was, only two types of filtering would be allowed then: all-or-nothing, or filtering one or all but one key).

For example, it allows to write things like

object intOrString extends Poly1 {
  implicit def caseBoolean = at[Boolean](_ => False)
  implicit def caseInt = at[Int](_ => True)
  implicit def caseString = at[String](_ => True)
}

val l = 1 :: true :: "foo" :: 2 :: HNil
val filtered = l.filterNot(intOrString)

The problem with this new filter is that it would use boolean literal types, and in particular some macro to manipulate these value types. But these macro cannot be called during their own build (the one of the shapeless-core project). So they have to be built separately, in a bootstrap project say. So the new filtering would require a bootstrap project for some macro and definitions that they use.

It would also require renaming the former filter methods to something else, like filterType, so that would break compatibility with former shapeless versions.

I gave a try to all of this here:
https://github.com/alexarchambault/shapeless/tree/refactor/bootstrap (bootstrap project),
https://github.com/alexarchambault/shapeless/tree/refactor/filtertype (renaming the former filter to filterType), and
https://github.com/alexarchambault/shapeless/tree/feature/filter (new filter method).

So all of this is feasible, but requires big changes. I let you decide whether you think they are worth it. (I think they are as far as I'm concerned :-)

[milessabin]

I'm going to punt this to 3.0 ... or has it been subsumed by Refined?