InvalidInstanceHead error can be subverted via vacuous fundep

[Quelklef]

Description

An instance head containing the empty row () normally causes an InvalidInstanceHead. As the compiler instructs, by adding appropriate fundeps to the class one can make the instance valid. Surprisingly, one can also make the instance valid by adding a fundep of the form a -> a where a is the type variable being instantiated to the empty row ().

To Reproduce

Run the following program in Purescript 0.15.8 (online playground):

module Main where

-- Fails, as expected
class A a b
-- instance A Int ()

-- Works, as expected
class B a b | a -> b
instance B Int () 

-- Works, surprisingly
class C a b | b -> b
instance C Int ()

Expected behavior

The instance with the vacuous fundep a -> a should (?) also cause an InvalidInstanceHead error

Additional context

The docs for InvalidInstanceHead say "when you see this error, it's because you're trying to write an instance of a given class that pattern matches on some row type [...] there isn't really a sensible way to match on row."

I don't know if this is worth anything, but it's kinda unclear to me what's unsensible about matching on row types, and I can't find an explanation online. In fact, by abusing this "vacuous fundeps circumvents InvalidInstanceHead" bug I was able to write a split function which can split a record into subrecords decided at compile time. eg,

splitMe :: { a :: Int, b :: String } -> { a :: Int } /\ { b :: String }
splitMe = split

This function is quite handy in a project I'm working on. Of course it could be implemented with RowToList, but I think that would require unsafeCoerce whereas the current implementation does not.

PureScript version

0.15.8

[Quelklef]

FWIW version 0.15.4 does not seem to exhibit this behaviour -- the example code fails with InvalidInsteadHead as expected

[rhendric]

AFAIK the only thing likely to have caused this in that timeframe would be #4195.