infered type does not typecheck

[Duncan Coutts]

Hi,

This was posted to the Haskell list earlier,
http://www.haskell.org//pipermail/haskell/2004-June/014286.html
but just to record what appears to be a typechecker bug.

I have a function with this type:
buildUArray' :: (Ix i, MArray (array s) a (ST s)) =>
		(i,i) -> (i -> a) -> ST s (array s i a)

Then I ask ghci (6.2.1 and 6.3 recent cvs) the type of the following
function:

buildUArray bounds f = do arr <- buildUArray' bounds f
                          unsafeFreeze arr

ghci says:
buildUArray :: forall a i b s array.
	       (MArray (array s) a (ST s), Ix i, IArray b a) =>
	       (i, i) -> (i -> a) -> ST s (b i a)

However if I add this annotation back in the script, ghc complains:

Could not deduce (MArray (array1 s) a (ST s))
	from the context (MArray (array s) a (ST s), Ix i, IArray b a)
      arising from use of `unsafeFreeze' at BuildArray.hs:25
    Probable fix:
	Add (MArray (array1 s) a (ST s))
	to the type signature(s) for `buildUArray'
	Or add an instance declaration for (MArray (array1 s) a (ST s))
    In the result of a 'do' expression: unsafeFreeze arr
    In the definition of `buildUArray':
	buildUArray bounds f
		      = do
			  arr <- buildUArray' bounds f
			  unsafeFreeze arr


We can get something that checks ok if we make the type of buildUArray'
a bit more specific, (specialise 'array' tvar to 'STUArray s'):

buildUArray' :: (Ix i, MArray (STUArray s) a (ST s)) =>
		(i,i) -> (i -> a) -> ST s (STUArray s i a)

Then ghc infers and happily accepts the following type for buildUArray:
buildUArray :: forall a i b s.
	       (MArray (STUArray s) a (ST s), Ix i, IArray b a) =>
	       (i, i) -> (i -> a) -> ST s (b i a)

I originally got this error message when using runST on buildUArray:
foo bounds f = runST (buildUArray bounds f)

The error message says:
    Probable fix:
	Add (MArray (STUArray s) a (ST s))
	to the expected type of an expression

However it is not clear where to add an annotation that will properly
express the class constraint on the 's' variable (which is of course
bound by a universal quantifier not at the top level).

Duncan