Re: Alternative generic hash function

From: Ken Wakita (wakita@is.titech.ac.jp)
Date: Tue May 30 2000 - 01:45:34 MET DST

  • Next message: Jacques Garrigue: "Re: ocamlc command line suggested features"

    In message (<20000529151357.24569@pauillac.inria.fr>)
    from Francois Pottier <Francois.Pottier@inria.fr>,
    talking about "Re: Alternative generic hash function",
    on Mon, 29 May 2000 15:13:57 +0200

    > I was thinking of a recording mutability per-field, rather than
    > per-object. I agree that your proposal seems economical enough.
    >
    > > there are plenty of clever tricks you can put in the runtime system
    > > that exploit the immutability of objects.
    >
    > Would you care to elaborate? I am no compiler expert, but I'm not sure
    > why it would be so interesting to have this information at runtime,
    > where it is already too late to do code optimization.

    Mutability information is useful in distributed computing. We have
    been implementing a distributed Caml, which comes with distributed
    shared memory. To preserve assignment semantics over the network
    (basically) we need to replace references to mutables with remote
    pointers when they are remotely referenced. In general, it is
    impossible to tell remote/local reference at compile time. Therefore
    we need mutability information at runtime.

    We modified the runtime representations for Caml mutable objects so
    that the communication substrate can detect mutables (vectors and
    mutable record fields). For this purpose, our modifed compiler
    assigns Mutable tag to vectors and arrays. As for mutable record
    field unboxed representation was replaced by boxed representation.
    With the latter design, access to record fields can be slower but it
    is negligible.

    Ken Wakita



    This archive was generated by hypermail 2b29 : Tue May 30 2000 - 22:29:26 MET DST