Support case-insensitive comparisons

[chris-martin]

Previously brought up by @frasertweedale here, where the request was to add a FoldCase instance (I presume from the case-insensitive package).

I'm inclined to discourage use of the CI type; it bothers me a lot that CI has a field that == doesn't consider in its equality test, I think it's an abuse of Eq. As the docs say:

(==) is customarily expected to implement an equivalence relationship where two values comparing equal are indistinguishable by "public" functions

And I think that going against this expectation can lead us to unhappy places.

I was considering adding another package called ascii-subset, whose purpose would be to provide types like "case-insensitive character". (I can think of a handful of ways to write this: to define a new type for each ASCII subset we care to consider, to define a 128-constructor GADT with some phantom type parameters that allow you to narrow the type, or to define various smaller types like ControlCharacter, Digit, CapitalLetter, SmallLetter etc. and then build up various ASCII subsets out of those. But, setting aside these details for the moment...) What do you think, would having "case-insensitive char" and "case-insensitive string" types be helpful, or is there a particular need for the CI approach that preserves the original case?

[frasertweedale]

In my use case - email and in particular header and parameter names - case-insensitive comparison is essential, and case preservation of values to be compared case-insensitively is desirable. The case-insensitive library meets these requirements.

I take your point about (==) and observational equivalence, but what CI does it pretty clear and unsurprising. And useful. Perhaps implement FoldCase with a prominent admonition about CI's violation of the Eq substitutivity law and suggest to write (==) `on` foldedCase instead?

[frasertweedale]

And y'know... I'll just leave this here :)

λ> import Data.ByteString.Lazy as L
λ> l1 = L.fromChunks ["ab", "c"]
λ> l2 = L.fromChunks ["a", "bc"]
λ> f = L.toChunks
λ> (l1 == l2, f l1 == f l2)
(True,False)

[chris-martin]

Heh, fair point about lazy ByteString, though I don't have to like that either!

I suppose it's no big deal to add an instance. That dependency is pretty stable, in Stackage, no major transitive dependencies, and the FoldCase class has remained unchanged for nine years, so the bound would be case-insensitive >= 0.1 && < 1.3.

I do also want to add some Equivalences, but we can do both.