* A Few Questions
@ 2006-12-18 1:51 Jonathan T Bryant
2006-12-18 3:18 ` [Caml-list] " skaller
2006-12-19 22:31 ` Nathaniel Gray
0 siblings, 2 replies; 3+ messages in thread
From: Jonathan T Bryant @ 2006-12-18 1:51 UTC (permalink / raw)
To: caml-list
I've been reasearching into a parallel extension to OCaml (based on
Reppy's CML and the OCaml Event
module). Mostly it's extending the semantics of channels and using
CamlP4 to add syntactic constructs
for concurrency, but there are a few extensions I'm having trouble with
because of OCaml language
"problems", and so I have a few questions.
1) Why can't a functorized module be used as a functor to another
module? I don't know if this is
semantically not possible or if I am just doing it wrong. I've played
around with the syntax and
reread the manuals, but I can't seem to find a solution. For example:
module type AModule =
sig
type t
end
module type BModule =
functor (A : AModule) ->
sig
type t
type s = A.t
end
module S : sig type t end
module Make = functor (B : BModule) -> S with type t = B.t (* This is
what fails *)
module AImpl =
struct
type t = int
end
module BImpl =
functor (A : AModule) ->
struct
type t = int
type s = A.t
end
module X = Make (BImpl (AImpl))
2) Why, in general, are there not first class modules? I've looked at
Russo's paper on this and it
shouldn't conflict with static typing or type inference. Again, for
example, why isn't this allowed:
val make : ('a -> 'b) -> ('a -> 'c) -> sig type t val x : 'a -> 'b val
y : 'a -> 'c end
let make a b =
struct
type t = int
let x = a
let y = b
end
module X = make (fun x -> x) (fun y -> y)
It doesn't seem like it would be very different from the already
allowed immediate objects and local
module bindings.
3) Since CML's threads are implemented via continuations, speculative
computation is allowed because
threads are simply GCed once they are not referenced any more. Since
OCaml's threads are implemeted
via system calls, is this still the case or do threads need to be
manually joined? I've run into
some instances where I can't create any more threads because the
"Thread limit" of 1024 has been
reached, but in code where nowhere near that many threads should be
left active. Is this limit an
OCaml limit or a system limit? Does it make a difference whether using
native code and system
threads vs. bytecode and vmthreads? Also, what about threads that are
not referenced anymore but
should still be running (i.e., "background services" and the like)? Is
there any way to keep the GC
from collecting them?
4) I've found that in sending functions across sockets, I can only send
them between copies of the
exact same binary image. Is it possible to marshal functions to
different binaries of the same code,
i.e., different platforms? Again, does native vs. bytecode make a
difference?
5) One possible extension is a vector type. Is it possible as is to
make the type inference
engine "as is" include the size of the underlying array as part of the
type information or does that
require modifications to the type system? Adding to the type
information allows runtime size checks
to be avoided and allows code generation to take advantage of external
vector processors and/or GPUs.
The ideal setup is vectors that are unboxed arrays of fixed length,
similar to tuples.
Thanks,
--Jonathan Bryant
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [Caml-list] A Few Questions
2006-12-18 1:51 A Few Questions Jonathan T Bryant
@ 2006-12-18 3:18 ` skaller
2006-12-19 22:31 ` Nathaniel Gray
1 sibling, 0 replies; 3+ messages in thread
From: skaller @ 2006-12-18 3:18 UTC (permalink / raw)
To: Jonathan T Bryant; +Cc: caml-list
On Sun, 2006-12-17 at 20:51 -0500, Jonathan T Bryant wrote:
> 3) Since CML's threads are implemented via continuations, speculative
> computation is allowed because
> threads are simply GCed once they are not referenced any more. Since
> OCaml's threads are implemeted
> via system calls, is this still the case or do threads need to be
> manually joined? I've run into
> some instances where I can't create any more threads because the
> "Thread limit" of 1024 has been
> reached, but in code where nowhere near that many threads should be
> left active. Is this limit an
> OCaml limit or a system limit?
system limit.
> Also, what about threads that are
> not referenced anymore but
> should still be running (i.e., "background services" and the like)? Is
> there any way to keep the GC
> from collecting them?
The gc doesn't collect threads in the first place.
It may collect a data structure identifying the thread if it
is unreachable.
Under Posix, if a thread is joinable it will not die until
joined. If it is detached, it will not die until it choses to.
(Assuming you don't kill it).
> 5) One possible extension is a vector type. Is it possible as is to
> make the type inference
> engine "as is" include the size of the underlying array as part of the
> type information or does that
> require modifications to the type system?
You would not want to do this. Use tuples instead.
Felix has arrays with known length as a data type,
in fact they're precisely tuples. The length type
is an anonymous sum of n units, written just 'n'.
However full manipulation is not possible. For example
you cannot have
concat: array[t,n] * array[t,m] -> array[t,add(n,m)]
because add(n,m) cannot work with an unconstrained type variable.
In Felix, whilst
(x:array[t,20]) + (y:array[t,30])
works correctly, you cannot concatenate arrays inside
a routine with polymorphic array bounds, even if the
instantiator would fix the bounds to a constant later.
In Ocaml there's no instantiator in the first place,
so the calculation would have to be done at run time
anyhow and require dependent typing support.
But the existing Array types in Ocaml simply drop the
length information. This avoids the problem at the
cost of not permitting enforced array bounds checks
based on type.
Note that in theory a subscript to an array of length n
is a value of type n, which is a unit sum, so the subscript
is a value of that unit sum, and such an index never requires
any bounds checking.
oleg showed a cute way to do the dependent typing on
existing arrays, and avoid many unnecessary checks,
in existing Ocaml. Maybe he can post the URL to that again.
--
John Skaller <skaller at users dot sf dot net>
Felix, successor to C++: http://felix.sf.net
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [Caml-list] A Few Questions
2006-12-18 1:51 A Few Questions Jonathan T Bryant
2006-12-18 3:18 ` [Caml-list] " skaller
@ 2006-12-19 22:31 ` Nathaniel Gray
1 sibling, 0 replies; 3+ messages in thread
From: Nathaniel Gray @ 2006-12-19 22:31 UTC (permalink / raw)
To: Jonathan T Bryant; +Cc: caml-list
On 12/17/06, Jonathan T Bryant <jtbryant@valdosta.edu> wrote:
> 4) I've found that in sending functions across sockets, I can only send
> them between copies of the
> exact same binary image. Is it possible to marshal functions to
> different binaries of the same code,
> i.e., different platforms? Again, does native vs. bytecode make a
> difference?
Nope. From the docs for Marshal.to_channel:
If flags contains Marshal.Closures, functional values will be
marshaled as a position in the code of the program. In this case, the
output of marshaling can only be read back in processes that run
exactly the same program, with exactly the same compiled code. (This
is checked at un-marshaling time, using an MD5 digest of the code
transmitted along with the code position.)
Cheers,
-n8
--
>>>-- Nathaniel Gray -- Caltech Computer Science ------>
>>>-- Mojave Project -- http://mojave.cs.caltech.edu -->
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2006-12-19 22:31 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2006-12-18 1:51 A Few Questions Jonathan T Bryant
2006-12-18 3:18 ` [Caml-list] " skaller
2006-12-19 22:31 ` Nathaniel Gray
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox