Feasibility of native RTS support for continuations?

[Ben Gamari]

Simon Peyton Jones via ghc-devs <ghc-devs at haskell.org> writes:

> |  Now that is very interesting, and certainly not something I would have
> |  expected! Why would asynchronous exceptions need to capture any portion of
> |  the stack? Exceptions obviously trigger stack unwinding, so I assumed the
> |  “abort to the current prompt” part of my implementation would already
> |  exist, but not the “capture a slice of the stack” part. Could you say a
> |  little more about this, or point me to some relevant code?
>
> Suppose a thread happens to be evaluating a pure thunk for (factorial 200). Then it gets an asynchronous exception from another thread.  That asynch exn is nothing to do with (factorial 200). So we could either 
>
> A. revert the thunk to (factorial 200), abandoning all
>    the work done so far, or
> B. capture the stack and attach it to the thunk, so that ie any other
>    thread enters that thunk, it'll just run that stack.
>
> Now (A) means that every thunk has to be revertible, which means keeping its original free variables, which leads to space leaks.  And extra work to avoid losing any info you need for reversion.  Extra work is painful; we want to put all of the extra work on the asynch exn.
>
> So we do (B).
>
> See Section 8 of "Asynchronous exceptions in Haskell".
> https://www.microsoft.com/en-us/research/publication/asynchronous-exceptions-haskell-3/
>
> And "An implementation of resumable black holes" (Reid).
> https://alastairreid.github.io/papers/IFL_98/
>
> This stack-freezing stuff is definitely implemented. I'm not quite
> sure where, but I'm cc'ing Simon Marlow who can point you at it.
>
For the record, runtime system captures the stack state in an AP_STACK
closure. This is done in rts/RaiseAsync.c:raiseAsync and some of this is
described in the comment attached to that function.

As Simon PJ points out, this is all very tricky stuff, especially in a
concurrent context. If you make any changes in this area do be sure to
keep in mind the considerations described in Note [AP_STACKs must be
eagerly blackholed], which arose out of the very-nast #13615.

Cheers and good luck!

- Ben
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 487 bytes
Desc: not available
URL: <http://mail.haskell.org/pipermail/ghc-devs/attachments/20200130/a9d626ba/attachment.sig>