FlexibleContexts and FlexibleInstances

[Niklas Broberg]

Dear all,

This post is partly a gripe about how poor the formal documentation
for various GHC extensions is, partly a gripe about how GHC blurs the
lines between syntactic and type-level issues as well as between
various extensions, and partly a gripe about how the Haskell 98 report
is sometimes similarly blurred where syntax is concerned (or not). All
these things make the life of a poor parser implementor quite
miserable at times. All in good jest of course, but with an edge of
truth, especially regarding (lack of) formal documentation.

The issue at hand which has caused my frustration is the
FlexibleContexts [1] and FlexibleInstances [2] extensions, which lift
restrictions imposed by Haskell 98 on the forms of contexts and
instances that may be defined. Great extensions both of them - but
what do they do, really really?

The following toy program requires MultiParamTypeClasses OR
FlexibleContexts in order to be accepted by GHC(i):

> f :: (T a b) => a -> Int
> f _ = 0

This of course assumes that we import the definition of T, we *must*
have MultiParamTypeClasses enabled if we want to declare T. Both
extensions thus enable classes with more than one argument to appear
in contexts.

Changing the program to

> f :: (T a ()) => a -> Int
> f _ = 0

i.e. changing the second argument to T to () instead, means we now
*must* have FlexibleInstances, in order to allow the non-tyvar
argument. This is nothing surprising, this is what FlexibleInstances
are supposed to do. But the question is, is this a syntactic issue or
a typing issue? In GHC proper this doesn't really matter much, as long
as it is caught *somewhere* then all is dandy. GHC's parser lets
everything pass, and it's the type checker that balks at this program.
But for someone like me with *only* a parser, this is a question that
needs a clear answer. Looking at the online report, the productions
regarding contexts are

context 	-> 	class
	| 	( class1 , ... , classn ) 	(n>=0)
class 	-> 	qtycls tyvar
	| 	qtycls ( tyvar atype1 ... atypen ) 	(n>=1)
qtycls 	-> 	[ modid . ] tycls
tycls 	-> 	conid
tyvar 	-> 	varid

Ok, so clearly the () is a syntactic extension enabled by
FlexibleContexts, as it is not a tyvar nor a tyvar applied to a
sequence of types. So this is something that a parser should handle.
FlexibleContexts also enables similar parses of contexts in other
places, for instance in class declarations, for which the Haskell 98
report says

topdecl 	-> 	class [scontext =>] tycls tyvar [where cdecls]
scontext 	-> 	simpleclass
	| 	( simpleclass1 , ... , simpleclassn ) 	(n>=0)
simpleclass 	-> 	qtycls tyvar

The difference here is that the simpleclass doesn't allow the tyvar
applied to a sequence of types bit. FlexibleContexts lifts that
restriction too, so there should be no difference between the two
kinds of contexts. So the new formal productions for flexible contexts
should be something like

fcontext	-> 	fclass
	| 	( fclass1 , ... , fclassn ) 	(n>=0)
fclass 	-> 	qtycls type1 ... typen  	(n>=1)

topdecl 	-> 	data [fcontext =>] simpletype = constrs [deriving]
	| 	newtype [fcontext =>] simpletype = newconstr [deriving]
	| 	class [fcontext =>] tycls tyvar [where cdecls]
	| 	instance [fcontext =>] qtycls inst [where idecls]

gendecl 	-> 	vars :: [fcontext =>] type

Does this seem correct?

Now let's turn to FlexibleInstances, which similarly lifts
restrictions, only to instance declarations instead of contexts. The
Haskell 98 report says on instance declarations:

topdecl 	-> 	instance [scontext =>] qtycls inst [where idecls]
inst 	-> 	gtycon
	| 	( gtycon tyvar1 ... tyvark ) 	(k>=0, tyvars distinct)
	| 	( tyvar1 , ... , tyvark ) 	(k>=2, tyvars distinct)
	| 	[ tyvar ]
	| 	( tyvar1 -> tyvar2 ) 	(tyvar1 and tyvar2 distinct)

Note the re-appearance of scontext, which is the same as above. The
instance head must be a type constructor, possibly applied to a number
of type variables, or one of three built-in syntactic cases. This is
where I consider the Haskell 98 report blurry - the fact that the
tyvars must be distinct, is that truly a syntactic issue? It might be,
it's certainly something that could be checked syntactically. But when
you take into account that with the proper extensions, they no longer
need to be distinct, at what level would we expect such a check to
happen? My gut feeling is that this check for distinctness is
something that a type checker might do better than a parser, though
it's not clear cut by any means. But since I don't do any other kind
of name resolution or checking in my parser even if it would be
possible (e.g. multiple declarations of the same symbol), I would find
it a bit anomalous to check this too.

Turning on FlexibleInstances, we shouldn't need to follow any of the
above restrictions on inst. In other words, the flexible production
should simply be something like

finst 	-> 	type

Right?

Now, FlexibleInstances *also* lifts the restriction on contexts, just
like FlexibleContexts - but *only* for the contexts of instance
declarations. This may seem like a reasonable thing, but it certainly
gives me some headaches. It means I could not treat the contexts
uniformly, but would need to have separate syntactic categories (or
rather post-parse checks) that look different between instance
contexts and other contexts (including class). So with
FlexibleInstances on, there are *three* different kinds of contexts
allowed: scontext for class declaration, fcontext for instance
declarations, and context for all other uses of contexts (type
signatures, data/newtype declarations). Just a small headache, since I
already apparently needed two categories from Haskell 98, but still.

I'm not sure I find it reasonable, that flexible instances are enabled
just for instance declarations but not elsewhere, but I'm sure a lot
of thought was given to that. If it was up to *me* though, I would
leave the flexible contexts with FlexibleContexts entirely, which
means you would have to use both flags if you wanted both in your
instance declarations. Would that be a bad thing? Separation of
concern seems desirable to me (and no, I'm not saying that just
because it would be easier to implement in the parser)...

At any rate, to make a long rant short:

* Are my interpretations of the lifted restrictions by
FlexibleContexts and FlexibleInstances correct?
* Is it reasonable that the issue of checking that tyvars are distinct
should not be considered syntactic?
* Would it also be reasonable to make the separation of concern
between FlexibleContexts and FlexibleInstances more clean?

Thanks for reading, and please give me input!

Cheers,

/Niklas

[1] http://www.haskell.org/ghc/docs/latest/html/users_guide/other-type-extensions.html#flexible-contexts
[2] http://www.haskell.org/ghc/docs/latest/html/users_guide/type-class-extensions.html#instance-rules