[Template-haskell] RE: Partial evaluation using TH

[Simon Peyton-Jones]

Duncan

Interesting.  I'm ccing the TH mailing list.  What do TH-afficionados
think?


Nested brackets
~~~~~~~~~~

| > evalProg' = [| evalProg |]		--same evalProg as before
| 
| > compileProg :: ExpQ -> ExpQ
| > compileProg = $( cogen [| evalProg' |] )

I'm very keen that TH should obey the beta rule.  That is, given
	x = e
you can always replace x by e.  So in this case you ought to be able to
write

	compileProg = $(cogen [| [| evalProg |] |])

I'm not saying anything about efficiency yet -- but it ought to be
correct. 

===> Conclusion 1: TH should allow nested brackets (and presumably
nested $'s).


Quoting
~~~~~
Consider

	v1 = [| x * 7 + x |]

We get a wodge of syntax tree for v1.  But consider this:

	w = x*7 + x
	v2 = [| w |]

We get a tiny syntax tree for v2.   How has this happened?  Because
every client of v2 (who might say $v) can also "see" w. 

We can't do this so easily if x is local:

	v3 x = [| x*7 + x |]

Now if we say $(v3 5) we'll get a syntax tree.  We can still try lifting
the body out:

	v3 x = let w = x*7 + x in [| w |]

Actually this will work if w is "liftable".  When we call $(v3 5), we'll
build the code [| 40 |].  Hmm. 
	
===> Observation 2.  Something I didn't realise: this forces w.  
	So substituting for w might change strictness behaviour.  Sigh.

However, if w's type is not liftable (e.g. it's Exp, or Tree), then the
program will be rejected as ill typed.  

====> Observation 3. I wonder if it's a coincidence that you got away
with the lift-to-top-level trick.  In a more general setting you might
not be able to, and then you'd get very heavy encoding.


Double encoding
~~~~~~~~~~
You ask for a function 
	enc :: Exp t -> Exp (Exp t)

Well that's easy, isn't it?

	enc x = return x

I'm not sure if that's what you want.  Indeed, you say " So, the thing
that concerns me is the double encoding. I can't intuitively see the
need for it..."  Maybe you can give the intuition for why you can't see
the need for it?  

I don't think I understand all the issues nearly well enough to comment
further.  Maybe you can chew on it with Ian, who is pretty au fait with
TH.

Simon