[Haskell-cafe] Possible extension to Haskell overloading behavior

[Chris Smith]

So I've been thinking about something, and I'm curious whether anyone
(in particular, people involved with GHC) think this is a worthwhile
idea.

I'd like to implement an extension to GHC to offer a different
behavior for literals with polymorphic types.  The current behavior is
something like:

1. Give the literal a polymorphic type, like (Integral a => a)
2. Type check the whole program, possibly giving the term a more
constrained type.
3. If the type is still ambiguous, apply defaulting rules.

I'd like to add the option to do this instead.

1. Take the polymorphic type, and immediately apply defaulting rules
to get a monomorphic type.
2. Type check the program with the monomorphic type.

Mostly, this would reduce the set of valid programs, since the type is
chosen before considering whether it meets all the relevant
constraints.  So what's the purpose?  To simplify type errors for
programmers who don't understand type classes.  What I have in mind is
domain-specific dialects of Haskell that replace the Prelude and are
aimed at less technical audiences - in my case, children around 10 to
13 years old; but I think the ideas apply elsewhere, too.  Type
classes are (debatably) the one feature of Haskell that tends to be
tricky for non-technical audiences, and yet pops up in very simple
programs (and more importantly, their error messages) even when the
programmer wasn't aware of it's existence, because of its role in
overloaded literals.

In some cases, I think it's a good trade to remove overloaded
literals, in exchange for simpler error messages.  This leaves new
programmers learning a very small, simple language, and not staring so
much at cryptic error messages.  At the same time, it's not really
changing the language, except for the need to explicitly use type
classes (via conversion functions like fromInteger) rather than get
them thrown in implicitly.  With GHC's extended defaulting rules that
apply for OverloadedStrings, this could also be used to treat all
string literals as Text, too, which might make some people happy, too.

Of course, the disadvantage is that for numeric types, you would lose
the convenience of overloaded operators, since this is only a sensible
thing to do if you're replacing the Prelude with one that doesn't use
type classes.  But in at least my intended use, I prefer to have a
single Number type anyway (and a single Text type that's not sometimes
called [Char]).  In the past, explaining these things has eaten up far
too much time that I'd rather have spent on more general skills and
creative activities.