MonoLocalBinds and hoopl

[Simon Peyton-Jones]

That is an interesting thought.  As it happens, each binding records what its free variables are, so it would not be hard to check whether all the free variables were top-level-bound.

Of course, it would make the rule a bit more complicated.  Rather than
	only top level bindings are generalised
it would be
	only binding groups whose free variables are top-level are generalised

Mind you, the rule is complicated already; for example bang-patterns are not generalised.  So maybe this would be no worse.

I must say I'm inclined to adopt this idea.  Any comments from others?

Simon

| -----Original Message-----
| From: Edward Z. Yang [mailto:ezyang at MIT.EDU]
| Sent: 14 June 2011 14:04
| To: glasgow-haskell-users; Simon Peyton-Jones
| Subject: Re: MonoLocalBinds and hoopl
| 
| I ran into some more code like this, and I realized there was something
| pretty important: the majority of let-bindings do not have any free varaibles.
| They could very well be floated to the top level without having to make any
| source level changes.
| 
| So maybe let should be generalized, if no free variables are captured.
| Some food for thought.
| 
| Cheers,
| Edward
| 
| Excerpts from Edward Z. Yang's message of Thu Dec 09 10:28:20 -0500 2010:
| > Hello all,
| >
| > Here's an experience report for porting hoopl to manage MonoLocalBinds.  The
| > Compiler.Hoop.XUtil module has a rather interesting (but probably common) style of
| code
| > writing, along the lines of this:
| >
| > fbnf3 (ff, fm, fl) block = unFF3 $ scottFoldBlock (ScottBlock f m l cat) block
| >     where f n = FF3 $ ff n
| >           m n = FF3 $ fm n
| >           l n = FF3 $ fl n
| >           FF3 f `cat` FF3 f' = FF3 $ f' . f
| >
| > f, m, l and cat are polymorphic functions that are only used once in the
| > main expression, and are floated outside to improve readability.  However, when
| > MonoLocalBinds is turned on, these all become monomorphic and the definitions
| > fail.  In contrast, this (uglier) version typechecks:
| >
| > fbnf3 (ff, fm, fl) block = unFF3 $ scottFoldBlock (ScottBlock (FF3 . ff) (FF3 . fm)
| (FF3 . fl) (\(FF3 f) (FF3 f') -> FF3 $ f' . f)) block
| >
| > One suggestion that I had was that we should generalize local bindings that
| > are only used once, but Marlow pointed out that this would make the typechecker
| > more complex and I probably would agree.
| >
| > As a userspace developer, I have two options:
| >
| >     1. Bite the bullet and put in the polymorphic type signatures (which
| >        can be quite hefty)
| >     2. Inline the definitions
| >     3. Move the polymorphic functions into the global namespace
| >
| > (3) and (2) are not so nice because it breaks the nice symmetry between these
| > definitions, which always define f, m, l for the many, many definitions in
| > Hoopl of this style.
| >
| > Cheers,
| > Edward
| >