Revert a CAF?

[wren ng thornton]

On 12/7/11 5:03 AM, Simon Marlow wrote:
> It would be possible, but it's not quite as straightforward as you might
> think. Suppose you have a program like this:
>
> xs = [1..100000]
> evens = filter ((==0) . (`mod` 2)) xs
>
> and you fully evaluate "evens". Now, GHC will garbage collect "xs",
> because it isn't required any more. However, if you revert "evens" to a
> CAF, now we require "xs" again, so we have to either revert that to a
> CAF or arrange to retain it in the first place on the grounds that we
> might need it again if some other CAF is reverted.
>
> Reverting xs to a CAF is not hard - we could have the GC revert CAFs as
> soon as they become unreachable. Arranging to retain it is harder.

Right. I know it isn't easy to do in the general case, which is why I 
was looking for an API hook rather than a Haskell function. Luckily my 
case is like xs. The only free variables are common functions 
---functions used by list comprehension notation, methods of Num Int, 
Enum Int, Ord Int, (||), and (&&)--- which are almost surely inlined 
away and would never be reverted anyways.


> GHCi gets around this by reverting *all* CAFs at the same time when you
> say :load.
>
> There's one other thing: GHC doesn't support reverting CAFs in
> interpreted code at the moment, you have to reload the module.
>
> So you need the following things:
>
> - modify the GC to revert CAFs when they become garbage
>
> - add a primop to revert a single CAF
>
> not too hard, I would think...

Good to know. I'll take a look at it over the break and see if I can 
come up with something.

-- 
Live well,
~wren