[Haskell-cafe] manage effects in a DSL

[Lindsey Kuper]

On Tue, Jan 28, 2014 at 3:03 AM, Corentin Dupont
<corentin.dupont at gmail.com> wrote:
> Hi Haskell-Caféists!
> I have a small DSL for a game. Some instructions have effects (change the
> game state), some not.
> -> In short, my question is: how can I semantically separate instructions
> with effect from the others? i.e. how can I mark down and track those
> effects?

Hi, Corentin,

This is very much like what we did in our work on adding fine-grained
effect specification to the Par monad in the LVish library
(http://hackage.haskell.org/package/lvish).  Just as Jacques suggested
up-thread, we did it by adding an extra phantom type parameter to the
monad which we call the "determinism level".  At its simplest, there
are only two of these levels, Det and QuasiDet, for deterministic and
quasi-deterministic:

  data Determinism = Det | QuasiDet

and the first parameter passed to the Par type constructor is of kind
Determinism (we have to turn on the DataKinds extension for this).

Doing so allows the static type of a Par computation to reflect its
determinism or quasi-determinism guarantee.  (The effects you can
perform in deterministic computations are a subset of the ones you can
do in quasi-deterministic computations.)  The beginning of section 6
of our POPL paper on LVish
(http://www.cs.indiana.edu/~lkuper/papers/lvish-popl14.pdf) discusses
this feature.  In more recent work, we've extended this to allow a
more fine-grained menu of effects to choose from, to the point where
we've begun calling it an "effect level" rather than merely
"determinism level".

Good luck with it!  Lately I have gotten a lot of mileage out of
thinking of monads as embedded DSLs.  Projects like yours really put
that analogy to use. :)

Lindsey