foo:bar
  foo.i
  foo!bar
  (foo i) ; has potential problems with macro bashing though

  given and givens
  scanners
  call* and defcall

> They were once but it turned out I never used this fact. The things you want to strings turn out to be different.

It'll even support mutation of objects within a single process, although it will gain copy-on-write semantics when an object is shared (via a global or message passing) to other processes.

The usability of the VM is also predicated on the usability of arc2c, for the very simple reason that I intend to just replace the final code generator step with a bytecode generator.

Joe == Joe Armstrong?

I think I'd rather keep Arc's ability to mutate everything, though; although certainly I'd avoid using mutations in any code samples if at all possible.

My problem is with this: http://arclanguage.com/item?id=7066

To summarize: Suppose Process A has an object a. It sends this object to process B. This object is a copy, a' (conceptually, in the language level). Now Process A sends object a to process B again. This object is a second copy, a'':

  (def process-A (process-B a)
    (==> process-B a)
    (==> process-B a))

  (def process-B ()
    (givens a
              (<==
                any any)
            a2
              (<==
                any any)
      (if (is a a2)
          (err "I somehow received the same object twice!"))))

Edit: It is important for them to be separate because process B expects two separate objects, and it might mutate one without expecting the other to be mutated.

There was a solution in my old shared concept, but it involved an O(M log N) search, where M is the message size in number of objects and N is the number of shared objects the receiver has access to, which I decided was simply suboptimal compared to an O(M) copy (where M is number of bytes in the message, but still...).

Edit2: re: passing stuff around. My intent is to not integrate the actual passing-around-stuff-via-network server in the C++-side code; instead, the VM will provide access to plain sockets. Instead an Arc-side server would be defined, with its semantics standardized (but not implementation, so that you can, say, implement a "more secure" server, or use some other protocol, etc.).

Instead, if a server receives a serialized PID over its connection from another computer, it checks a table matching the PID computer and id number. If the table fails to match, it launches a "reflector process" which simply sends every message it receives to the server, for serialization, and inserts that into the table. The PID of the reflector proces is then used as the message sent by the server to the destination process. All PIDs then get this translation (obviously the table has to be bidirectional. so that a PID can be matched to the serialization of the PID.).

And you should probably allow the user to require copying if they need to, so that there aren't any surprises.

As an aside, you mentioned a O(M log N) search to determine whether to share or not. Now maybe I don't understand, but you could use a Bloom filter to determine if the object is present on the other process. The only shortcoming is the false positives, which mean you would copy unnecessarily, but they happen so infrequently that it probably isn't an issue. I believe bloom filters were mentioned on Hacker News yesterday. The link on wikipedia is: http://en.wikipedia.org/wiki/Bloom_filter

Using a bloom filter, you'd be able to in constant time, low memory usage, determine if an item is in a set. In this case the items are shared objects and the set is the objects owned by the target process. Now, maybe I don't understand what you're trying to do, but you can clarify that shortly.

How much are you implementing on the c++ side? Couldn't you get a few basic necessities built first, and then just abstract towards usability in arc? I don't know exactly what you're doing, but I don't see why you have to make everything at once.

Except this doesn't take any shared structure into account---as soon as you send something over the network, all the objects which share it are fragmented. So you need some sort of "versioning." Perhaps you instead have every reference have a version flag, and every copy between threads increments version for the new copy; modifying something checks the reference you are modifying it through, makes a new copy if necessary for that version, and then modifies it. The problem is that I can't see a good way to do versioned indexing, especially if references are really just raw pointers.

Though note that if you do come up with a working "copy-on-write-and-other-things" scheme, the Arc copy function becomes essentially

  (def copy (arg)
    (==> (current-pid) arg)
    (<== (current-pid) any))

.

This was indeed my original idea: http://arclanguage.com/item?id=7037

Incidentally, 'lset is very, very hard to implement if closures are not very mutable, as in the arc2c output.

> and then make the reference through which you modified it point instead to this new object.

  (def hmm (x)
    (let y (cdr x)
      (scar y 42))) ; is x mutated? the reference I used here was *just* y...

Yes, this is true, although I'd prefer to add some sort of 'uniq tag so that other pending messages don't accidentally get taken ('<== does pattern matchimg, obviously implemented via my p-m macro) ^^. It will even correctly handle recursive and shared structure, which I think the current 'copy does not ^^.

  (def copy (arg)
    (w/uniq tag
      (==> (mypid) (list tag arg))
      (<==
        (,tag cpy) cpy)))

> That's a good point. But what about having copy-on-write-if-more-than-one-logical-copy semantics? Then every Arc object on the heap would have a shared flag. Then if you send an object to another thread, the object is shared, and it has its internal shared field set to true (or 1, if this is C).

The problem here is: what if the object to be sent is already shared and the receiver already has it? Say:

  (def process-A ()
    (let v (list 42)
       (==> B v)
       (==> C v)))

  (def process-B ()
    (<==
      any  (==> C any)))

  (def process-C ()
    (givens v1 (<== any any)
            v2 (<== any any)
      (if (is v1 v2)
          (err "I somehow received the same object twice!"))))

Instead, you should only transfer information over messages. That way you don't need to worry about keeping shared objects up to date for all users.

Good idea about the shared flag; maybe it should be a ref counter? That way you can tell if the other process(es) have stopped using it. So, if the ref count is 1, than just straight up mutate, otherwise, copy?

Also, couldn't the send function be:

  (-> pid message)

  (givens str1 "message"        ; 1
          str2 str1             ; 2
    (do                         ; 3
      (-> another-process str1) ; 4
      (= (str1 0) #\M)          ; 5
      (prn str1)                ; 6
      (prn str2))               ; 7

  Message
  message

Versioning would mean that str1 and str2 would both be pointers to "version 0" of some value, which had an internal max-version of 0. Line 4 sends the object over, so max-version is incremented (to 1), and the other process is sent a pointer to "version 1" of the same object (1 because that's the new max-version). However, since no modifications have been made, somehow (I don't know how) both "version 0" and "version 1" resolve to the same underlying object. Then, on line 5, sref is called on "version 0" of this object, so the object is copied: "version 0" is copied away and modified, but "version 1" is left alone. When we print the objects, their reference to "version 0" makes them look at this new, copied objects, so we see

  Message
  Message

  ; Available values
  A --> "a string or another object"
  B --> "A string or another object"
  C --> "A string."
  
  ; Version --> value mapping
  0 --> [A]
  1 --> [A]
  2 --> [B]
  3 --> [A]
  4 --> [C]

Hmm, thinking about the underlying structure of this and writing it out makes me think this makes less and less sense, and that either (1) this is actually trivial/a restatement of another scheme/both and thus won't work, or, if I'm lucky, that (2) there's some more general and less clunky scheme hiding inside it (hey, I can dream, right?). My money's on (1) right now, but I do feel like I understand the problem better.

Oh, and I like your suggestion of -> / <-. Nice and short.

So, since I believe almkglor said he was planning to use arc2c for this VM, only changing the code generation step, then the solution seems to be to just use the shared flag on values (probably the reference-counting version, as shader suggested). It seems like since we have sharedvars, the problem solves itself: when you send an object to another thread, you don't send the sharedvar, you send the value the sharedvar is wrapping. Then, when you modify something, then if it's a sent copy, that copy is modified, not affecting the original (because of the shared flag); if it's a local with aliases, then it's modified through the sharedvar, and since the sharedvar is still aliased and only its internals are modified, there's no problem!

You never actually send sharedvars - they are not accessible at the Arc level. However, they are necessary in order to send functions across processes; the generated code expects to see sharedvars, so it has to have (a copy of) sharedvars.

Edit:

> ...if a local variable is ever mutated, a 'sharedvar' C-struct is used to represent that variable. These are constructed by %sharedvar primitives, and are accessed by %sharedvar-write and %sharedvar-read primitives.

This means this case:

  (let foo 1
    (set foo 42))

  (let foo "hello"
    (sref foo #\H 0))

Also, I agree about the whole perusal / editing bit. I'm probably even less experienced than absz, but no less interested.

And I again recommend moving to a new thread. The window's getting a bit cramped, and we aren't even talking on the original subject of the post.

Not yet, the only code I've started writing on is the memory management code. There are two versions, one which uses a mark-and-sweep for local heaps and has shared heaps implemented via boost::shared_ptr, and my newer version which uses stop-and-copy and does (will do?) copy-on-message-pass.

I can send you (and anyone interested) some initial tarballs though ^^. WARNING: C++. With RAII. Be scared.

My solution for the multiple-version thing was to simply have the thread copy all its shared objects and drop all references to shared.

Then I began to feel uncomfortable with it and decided that just copying the darn thing might be better after all ^^

My main reason for using ==> and <== was to make them stand out a bit ^^. -> and <- just look so skinny. Although it's not carved into stone yet.

Currently my (unimplemented) plan is, memory is kept in managed heaps, which are handled by a mark-and-sweep collector (can't use semispace or variants for reasons below; at least, I don't think I can use semispace). If a message is passed to another thread, the memory of that message (and every other managed area it refers to) is moved from "managed" space to "shared" space, which is reference counted (the references counted are from threads to shared objects, so circular references are impossible). Objects in "shared" space are immutable (and can thus be accessed lock-free). If a thread attempts to mutate a shared object, instead all shared objects still referenced by that thread are copied to managed space (copy-on-write), the shared objects are released (decrement refcount by 1, free if zero), and the managed copy is mutated (we copy all of them because other shared objects might refer to the shared object being mutated).

Note that "managed" space and "shared" space are not separate areas of memory: they are actually two sets of pointers, so moving an area of memory is just moving a pointer from one set to the other (very cheap). The reason I think I can't use a copy collector is that a shared memory object might end up being located between two managed objects, and the shared object can't be moved without disturbing all the other threads that might be referring to it.

Shared objects are OK to refcount because we're refcounting from threads to shared objects. A thread might have 1000 references to a shared object, but it just counts as 1. Only when the thread GC's and determines that the shared object is not marked does it actually decrement the refcount.

Marks are normally kept in a bool with the memory area, if the memory is managed. If it's shared, then the marking information is kept by membership in a marked_and_shared set while marking (in order to preserve the invariant that shared objects are immutable). Sweeping involves traversing both the managed and the shared set; if the managed set memory area is not marked, then it is freed and removed from the set, etc. If the shared set memory area is not in the marked_and_shared set, then it is released (decrement refcount, free if zero) and removed from the set, etc.

There is also a "light" GC which just sweeps the managed set. Crucially, because shared objects are immutable, they cannot point to managed objects, so mark-and-sweep on only managed objects won't have the overhead of referring to marked_and_shared objects.

(and, err, that's pretty much most of my thinking with regards to having separate heaps being used by various threads. I actually have some of the messier details written down in a text file, anyone interested can post their email addresses ^^)

Basically the main thing I'm trying to solve is the copying of memory of messages when an object is sent across processes. Hence my use of shared immutable data, which is only copied if necessary.

However the problem is really something like this:

  ; process A
  (let v (some-value)
    (==> B v)
    (==> B v))

This can actually happen more often than you think. Consider the case where a process serves as an STM-style container:

  (def container (state)
    (while t
      (<==
        ('put obj)
          (= state obj)
        ('query pid)
          (==> pid state))))

   A ==> object ==> B

   A ==> query ==> B

   A <== object <== B

The bit that's bothering me about copy collectors is that it's somewhat difficult to handle cleanup/finalization. For example, it would be neat if the VM automatically closed any file handles you happened to have dropped, or to automatically terminate processes that are waiting for a message (without a timeout) and whose pid's have been dropped (so that you can return a pid as a "data structure" with impunity, and still have the process be garbage collected when the data structure loses scope).

I think I've hit upon a possible way of handling this though. A copy collector needs to put up to-pointers anyway; I think it may be possible to have the to-pointers as part of the object. The to-pointers default to NULL. When a heap is released (i.e. after its contents have been copied to another heap), the heap inspects its held objects. Any objects whose to-pointers are NULL are finalized/deleted; in effect, the to-pointers become really expensive mark bits.

> You could even use different GC strategies on a per-thread basis (I don't know how much this would be helpful, though...).

(the vtable also serves as my "type code", and is probably better thought of that way)

So I can't safely replace a Generic object on the heap with a ToPointer object, because I'd lose access to the size (i.e. I lose the type code, which I'm using to determine the size). This means I can't sweep and finalize afterwards. The alternative is to keep the size separate from the type code/vtable, but then for each object I end up storing a size_t, which on most systems is probably the same size as a Generic* - so I don't save it either.

So basically my options are:

  type code (automagically inserted by C++ as the vtable)
  to-pointer
  data

  size
  type code
  data

However I intend to use the same copy-algorithm for message passing, so I can't overwrite the tag/type code and the data just because I've copied it, since the original could be reused.

In this case, as you've already said, you can't overwrite the orginal type and you have to reserve extra space for the to-pointer (this means one extra word for every allocated object). This way another problem arises: you have to be sure that the to-pointer is clear every time you start a copy this is automatic after a GC, but not after a copy for message passing. This means that after every message passing, the whole heap must be scanned.

> Or maybe I should actually use a different routine for message passing and GC.

Or again use a mark and sweep with copying instead of sharing of objects ^^. Well, not much different.

Such as?

I know we need to have more Erlang-style VM's. In particular I think we need VM's that pretty much enforce shared-nothing by causing all messages to be copied at the VM level. Hence my current project, which will also tie in some code from the AST-based compiler of arc2c.

- Could we make pattern matching through unification a part of the core language?

- Could we make some tasks easier (I mean terser) using Icon or Prolog-like goal-oriented programming?

- Could Haskell typeclasses be adapted to a dynamic language? (I know that is not possible in general, but maybe in the most common cases). CLOS-like oop is the way?

- Easy defstruct as in Haskell is also a win.

- Which would be the best module system? Python's first order modules are cool, but I also Haskell's module system seems better to do ADTs.

- Could we use generators using a lazy list-like interface?

- To which point could the system be reflective? In the past it was considered that "flambda" (user defined, first class special forms) and redefition of eval with "reflective towers" are not that good ideas, because of efficiency. Should we spend more cycles to allow such things?

- Should the read s-expressions also have associated meta-information?

- Could Arc live in an image?

- Could Arc be just the specification of its own compiler, making the bytecode an official part of the language (as pg once said)?

- Could Arc be blazingly fast?

- Could Arc be prepared for multiprocessor architectures like the guys (pun intended ;)) in Fortress want it to be? Or to intensive scalability like Erlang seems to have?

Of course all that questions are implementable. There probably is a library for every one of them in Cliki.

I am just asking which of them we should take and encourage to be seen as part of "Arc's core". Commonality and practice of an idiom, more than its mere possibility, is what actually defines what a language is.

Currently, Arc's core is not much more than sugar for Scheme (modulo defcall, mostly unseen and great). This doesn't want to be insultive, I am happy that people find Arc useful.

IMO yes. It might even be possible to have the meta-information attached to the nearest cons cell instead, so a cons cell might have, say, (cons a d (o line-number) (o file-name)). This makes symbols always equal other symbols without having to worry about attached meta-information for each "symbol".

Might be useful to add in my VM then ^^.

> - Could Arc be just the specification of its own compiler, making the bytecode an official part of the language (as pg once said)?

Interesting. Might be useful to define Arc as a set of macros which just expand to a bunch of bytecode.

Anyway I'm thinking my VM should be "bytecode" based. The bytecode won't have a numeric representation, but it will have a symbolic one, like ((localvar 0) (globalvar foo) (plus))

> - Could Arc be prepared for multiprocessor architectures like the guys (pun intended ;)) in Fortress want it to be? Or to intensive scalability like Erlang seems to have?

Well, this is what I want to do ^^.

> - Could we use generators using a lazy list-like interface?

Anyway, you could build a Prolog-Like DSL with pat-m and amb

  http://arclanguage.org/item?id=2556
  http://arclanguage.org/item?id=6669

- How should you call functional functions from logical functions?

- How should you call logical functions from functional functions without violating referential transparency?

- Do closures over logical variables make any sense?

  import module

  x = module
  x.function(...)