m (Changed genericLenght to genericLength)
Revision as of 17:30, 15 July 2007
The monomorphism restriction is probably the most annoying and controversial feature of Haskell's type system. All seem to agree that it is evil, but whether or not it is considered a necessary evil depends on who you ask.
The definition of the restriction is fairly technical, but to a first approximation it means that you often cannot overload a function unless you provide an explicit type signature. In summary:
-- This is allowed f1 x = show x -- This is not allowed f2 = \x -> show x -- ...but this is allowed f3 :: (Show a) => a -> String f3 = \x -> show x -- This is not allowed f4 = show -- ...but this is allowed f5 :: (Show a) => a -> String f5 = show
Call me dense but why exactly is the first one OK and the second one not? -- anonymous
- The difference is that the first version binds x via a "simple pattern binding" (see section 188.8.131.52 of the Haskell 98 report), and is therefore unrestricted, but the second version does not. The reason why one is allowed and the other is not is that it's considered clear that sharing f1 will not share any computation, and less clear that sharing f2 will have the same effect. If this seems arbitrary, that's because it is. It is difficult to design an objective rule which disallows subjective unexpected behaviour. Some people are going to fall foul of the rule even though they're doing quite reasonable things. -- Andrew Bromage
Arguably, these should all be equivalent, but thanks to the monomorphism restriction, they are not.
So why is the restriction imposed? The reasoning behind it is fairly subtle, and is fully explained in the Haskell 98 report. Basically, it solves one practical problem (without the restriction, there would be some ambiguous types) and one semantic problem (without the restriction, there would be some repeated evaluation where a programmer might expect the evaluation to be shared). Those who are for the restriction argue that these cases should be dealt with correctly. Those who are against the restriction argue that these cases are so rare that it's not worth sacrificing the type-independence of eta reduction.
- An example, from A History of Haskell: Consider the
genericLength :: Num a => [b] -> a
- And consider the function:
f xs = (len,len) where len = genericLength xs
Num a => aand, without the monomorphism restriction, it could be computed twice. --ARG
Oversimplifying the debate somewhat: Those in favour tend to be those who have written Haskell Implementations and those against tend to be those who have written complex combinator libraries (and hence have hit their collective heads against the restriction all too often). It often boils down to the fact that programmers want to avoid legalese, and language implementors want to avoid cruft.
In almost all cases, you can get around the restriction by including explicit type declarations. Those who are for the restriction are usually quick to point out that including explicit type declarations is good programming practice anyway. In a few very rare cases, however, you may need to supply a type signature which is not valid Haskell. (Such type signatures require a type system extension such as ScopedTypeVariables.) Unless you're writing some weird combinator libraries, or are in the habit of not you're unlikely to come across it. Even so, most Haskell Implementations provide a way to turn the restriction off.
See also: Section 4.5.5, Haskell 98 report.
Some question or suggestion: As I understand the problem arises from the situation that two different forms of assignment are described by the same notation. There are two forms of assignment, namely the inspection of data structures ("unpacking", "pattern binding") and the definition of functions ("function binding"). Unique examples are:
let f x = y -- function definition let F x = y -- data structure decomposition
In the first case we have the identifier f starting with lower case. This means this is a function binding. The second assignment starts with F, which must be a constructor. That's why this is a pattern binding. The monomorphism restriction applies only to the pattern binding. I think this was not defined in order to please compiler writers, but has shown to be useful in practice, or am I wrong? But the different handling of these binding types leads to a problem since both types have a common case.
let x = y -- function or pattern binding?
So, what speaks against differentiating the assignments notationally, say
let f x = y -- function definition let F x <= y -- data structure decomposition
and keep the monomorphism restriction as it is?
The problem isn't just pattern bindings, it's that pattern bindings which are typeclass polymorphic are actually function bindings in disguise, since the usual implementation of typeclasses adds parameters to such definitions, to allow the definition to take the typeclass dictionaries involved. Thus, such pattern bindings have different properties with respect to sharing (they're generally less shared than you want). In especially bad cases, without the MR, it is possible to have programs which run exponentially slower without type signatures than when signatures are added. Just distinguishing pattern bindings with a new notation doesn't solve the problem, since they'll have to be converted into function bindings in that case anyway. If you intend to keep the MR, then you don't need to change anything. The issue with the MR is just the fact that it's annoying to have eta-reduction fail in the absence of explicit type signatures, and the fact that it makes otherwise perfectly valid programs fail to compile on speculation that there might be loss of sharing (when there usually isn't, or at least the impact isn't large enough to worry about).
John Hughes recently advocated the removal of the MR on the Haskell Prime mailing list, and suggested replacing it with two forms of pattern binding: one for call-by-name (polymorphic, not shared), and one for call-by-need (monomorphic, guaranteed shared). This might be similar to what you're suggesting. If you look at it too closely, it seems like a good solution, but the overall impact on Haskell code seems too large to me, to resolve a distinction which it ought to be statically possible to determine.
I'm of the opinion that it would be better to find a way to restore sharing lost through the typeclass transformation in some way, or else implement typeclasses in an altogether different way which doesn't run into this problem. Additional runtime machinery seems like a likely candidate for this -- the interactions with garbage collection are somewhat subtle, but I think it should be doable. It's also possible to restore the sharing via whole-program analysis, but advocates of separate compilation will probably complain, unless we were to find a mechanism to fix the problem from the object code (and potentially temporaries) at link time.