I'm trying, and failing, to understand why ForeignPtr is defined the way it is. It all seems rather complicated and redundant. I was thinking we might want to switch to something simpler and more explicit, like this:

data ForeignPtr a = ForeignPtr {

  fptr :: !(Ptr a) -- What we point to

  froot :: MutableByteArray# RealWorld

  fwk :: MutVar# RealWorld Any -- The finalizers

  }

froot represents the object the ForeignPtr points into. When created by one of the malloc variants, this will actually be the allocated byte array. Otherwise, it will be a 0-length array.

fwk should be seen as having type

  fwk :: MutVar# RealWorld (Weak# (MutableByteArray# RealWorld))

but we can't express that directly right now. fwk is keyed on froot, and in the malloc case also points to it, to prevent it from being freed prematurely.

It seems that this representation requires a couple extra primitive features from Weak#. I think that's okay: Weak# already has some primitive features designed to support ForeignPtr.

1. An operation to add a Haskell finalizer to a Weak#, similar to the one that adds a C finalizer.

2. A way to deal with mixed finalizers: either make the finalizer-adding primops report distinguishable failure on mixing or arrange to run the C finalizers after any Haskell finalizers (since C finalizers are much more likely to actually invalidate the pointer).

3. A documented guarantee about the order in which the finalizers attached to a particular Weak# run, compatible with the ForeignPtr documentation.