[Haskell-cafe] idea for avoiding temporaries

[David Roundy]

Ah, I was missing your point, I've heard something called copy-on-write,
which wasn't what you describe (or I also misunderstood it when I heard it
before).

I see.  But how would one manage these handles? What's to keep me from
accidentally copying a handle? It sounds like it'd require explicit memory
management, in order to avoid ever copying a handle, if I were to implment
this myself.

Or are you suggesting that if the simons implemented a copy-on-write scheme
in ghc's RTS, then I'd be all set?

In short, managing the reader count is exactly the problem that sounds
hard, and I still don't have any idea how one would go about it.

David

On Thu, Mar 08, 2007 at 01:31:19PM -0800, Dan Weston wrote:
> I might be missing the point, but I think you are missing mine.
> 
> The copy-on-write I am talking about means that it's no longer "your 
> data", so you don't need any knowledge of who has access to it because 
> you don't own it or have a pointer to it. It is owned by some broker 
> from which you request a read-only or write access handle as needed. 
> Requested changes to underlying data already shared by others triggers a 
> copy and reassignment of pointers to it for your handle alone.
> 
> The copy cost appears only when there is more than one handle to the 
> same data and one of them changes it.
> 
> All this can be wrapped up and hidden away. If you want to escape this 
> broker business and steal back your data, just ask: the broker will 
> duplicate shared data needed by others, change their pointers to it, 
> then disown the pointer it returns to you.
> 
> This is copying without writing (unnecessarily). Or am I missing something?
> 
> Dan
> 
> David Roundy wrote:
> >I'm thinking you're missing the point.  The point is to copy without
> >writing, and that requires some knowledge (whether static or runtime) of
> >whether anyone else has a reference to my data--which copy-on-write won't
> >give me.
> >
> >David
> >
> >On Thu, Mar 08, 2007 at 11:15:25AM -0800, Dan Weston wrote:
> >>Or possibly more generally copy-on-write, which requires one more level 
> >>of indirection (handle instead of ptr). Since you are talking about 
> >>using ForeignPtr, this is already within your power to prototype, I 
> >>should think.
> >>
> >>Dan
> >>
> >>Dan Piponi wrote:
> >>>On 3/8/07, David Roundy <droundy at darcs.net> wrote:
> >>>
> >>>>I started wondering whether there's a solution that would allow us to
> >>>>write pretty high-level pure functional code, while the RTS can realize
> >>>>at run-time that we have the only reference to the input argument and
> >>>>that it is therefore safe to consume it destructively.
> >>>I think you're talking about uniqueness typing which is supported by
> >>>the programming language Clean.