On Wed, Jul 13, 2016 at 5:42 AM, C Maeder
seeing
aexp -> qvar (variable) | gcon (general constructor) ... | qcon { fbind1 … fbindn } (labeled construction) | aexp { fbind1 … fbindn } (labelled update)
and https://www.haskell.org/onlinereport/haskell2010/haskellch3.html#x8-220003
I realise that the update requires at least one field binding whereas for a construction "C {}" (n = 0) could be used. ("C {}" makes sense for patterns!)
And due to the meta-rule a labelled update is not possible for a lambda abstraction, let expression, or conditional (as aexp), but it is for case (and do if the record type happens to be a monad). So a further less obvious example is:
case e of p -> r { f = v }
that will be parsed as: (case e of p -> r) { f = v }
(I'm sure the grammar could be fully disambiguated, but this would not improve readability. Preferring shift over reduce is common and fine for such cases.)
Upon reading this example, I believed this to be simply a matter of the layout rule. case e of p -> r { f = v } would become case e of { p -> r } { f = v } This, on the other hand case e of p -> r { f = v } would be equivalent to case e of { p -> (r { f = v }) } I just tested this after writing the preceding as I was confused about what you found confusing, and I am surprised that the example you showed does indeed yield a parse error. I very much expected this to be valid Haskell: data X = X { x :: Bool } someX = X True foo = case () of _ -> someX { x = False } Am I alone in my surprise?