Abstractly, a continuation is a representation of the state of an execution, and especially its "execution stack" [L1 ].
There's a delightful article explaining continuations at [L2 ].
LV I recall a very similar concept from back in my ancient days writing BAL code on IBM/MVS systems. We called it concurrent programming back then.
DKF 17-Apr-2005: Interesting. I just wish it was easy to apply in Tcl, but our openness to extension, especially by C libraries, makes this very tricky indeed.
WHD 17-Apr-2005: Damn. I want this. The idea of User Interface Continuations makes me drool.
jcw - Yeah, me too, same reaction. DKF, yes the C stack can cause trouble, but only when C calls back into Tcl. There are probably many cases when this need not be a show stopper. I think there are ways to get there. Might be something to discuss at the tcl2005e conf - are you going to be there?
NEM 18-Apr-2005: You could sort of get this to work, if we went a similar route to Stackless Python (from what little I know of that project [L3 ] and [L4 ]). You'd have to change all the C-API so that instead of re-calling into the interpreter with Tcl_Eval* APIs, they instead return a TCL_CALL or similar code with some code of where to go next in the interpreter result (i.e. trampoline). This is very similar to the recent proposals for tail-call optimisation (which is also very connected to continuations, through continuation-passing style, CPS). Obviously, it'd be an enormous amount of work to restructure the core and many many C extensions to do things this way. Continuations rock. (I should be at tcl2005e, and would love to discuss these ideas with people).
The core is not that far from doing it. At one point I had an experimental patch working, which copied the C stack away and back. Not sure it covered all the cases, but it was definitely very close to full contins across all of Tcl and C. -jcw
WHD: I don't see why the C stack is an issue, based on what the article says. A closure consists of a stack frame and its parent stack frame, and so on recursively, it says; but a continuation, so far as I can tell, is a single stack frame. Suppose we made the rule that you could "yield" only from the body of a normal proc.... Oh. If you do it from within a while loop, say, you've got the C stack involved. Never mind. (Or is there something else I'm missing?)
Will, the stack becomes a tree, because processing (and further nesting) can proceed at any continuation point. Yes, C loops and C re-entering Tcl (as foreach and proc do now when evaluating their bodies) is where things get tricky. -jcw
RHS 17Apr2005 One of the most interesting (and informative) descriptions I've seen of the use of continuations was in the area of web programming. The programmers used them to use web pages as if they were stateful. The "process" that the web pages are involved with work linearly. Each time it needs input from the user it saves its current continuation (which contains everything about where it is in the process) in a hashtable. It then returns a web page with a form to the user, and one of the elements of the form (a hidden one, I presume) contains the key into the hash table for the current continuation. Once the user fills in the form and hits submit, the process can look up the continuation in the hash table and continue right where it left off... Only, now, it has the data it needs to continue, as provided by the form.
SS 18Apr2005' Using setcontext and getcontext calls that are POSIX and AFAIK also available in some form on every platform Tcl supports allows for continuations without a stackless interpreter (basically you save the C stack too). This is how it's in plan to implement continuations into Jim. (this is also how Ruby implements continuations IIRC).
jcw - Setcontext/getcontext (SC/GC) work by switching stacks, which is not 100% as general as being able to resume any stack frame in what can become a tree of stack frames in the general case. There's also the issue of whether lower stack frames are treated as copy-on-write or not, i.e. whether changes are isolated, when these frames are shared by multiple continuations.
SS Sure jcw, what SC/GC does is to make you able to implement continuations without to write a stackless interpreter. Of course the Tcl stackframes are to be copied (in the trivial implementation at least) but that's not a too big problem as it is as simple as to copy some hashtable and linked data structure. The real problem is to write a stackless interpreter and given that it's a major work and that there are other systems to reach the goal why don't use them.
chi 19Apr2005: I mean to remember, that Ruby does its continuation and thread thingy by simply copying the whole C stack from some point remembered during initialization of the interpreter up to the context of the calling C function. Switching back to that continuation simply copies back that remembered stack. So it works within C functions. Unfortunately setcontext and getcontext seem not to be widely available :-( Perhaps the same implementation as in Lua could be facilitated? In Lua, C context will not be remembered, AFAIK. So a C function has to come back. But a Lua function could be suspended. This restriction seems not to be too hurting, would it?
SS 19Apr2005: Indeed the Ruby trick may work well on many platforms where there is no support for SC/GC, about Lua, maybe the ability to suspend a function allows for coroutine and generators? Btw given that there are ways to copy the whole C stack why not :) In Tcl the ability to reactivate a single procedure may be more complex than what happens in other languages because there is upvar/uplevel that may notice the difference. Not that this is a big problem after all... the user can just avoid them. Well an interesting alternative btw. Oh.. well also there is another problem with the Lua approach, in Tcl the interpreter calls itself recursively even in the context of a single procedure! While in languages like Lua it only happens when a function calls another one I guess.
chi 19Apr2005: That Tcl interpreter calls itself recursively should not matter at all. We would being restoring the whole C stack. So the interpreter would itself find back in the C function it was called to save the context. I had once written an extension for another language that worked pretty well. Then I had simplified the extension to make a C module out of it so that it could be used even in plain C programms. But to use such an extension in Tcl or Jim, we would need a mechanism to store the Tcl/Jim stack as well to protect it against freeing by refcount dropping to zero, though. And we would have to restore the Tcl/Jim stack as well. Further I would not like to keep secret, that there is some performance penalty taking this approach for implementing continuations, as the whole C stack has to be copied over and over again. But it should be doable.
SS Agreed. What I mean is that "without to copy the C stack" some language is still able to do generators (ala Python), just activating the same function every time it's needed. This is sometimes possible in languages that have recursive calls to the interpreter only when functions are called, and not for every thing like if, while and so on like Tcl. Btw this is poor anyway, you can't even write recursive generators. Copying the C stack, there is no problem at all of course. Also to copy the stack frames should not be hard, for instance in Jim there is just a linked list of stackframes containing an hashtable of local variables. There is just to duplicate the list and hash tables just incr-ref-counting every object representing a variable, this should be pretty easy but as you said is very slow, you don't only need to copy the C stack but the whole Tcl stack. Still I think it's nice enough as a first try just to experiment with it. After all in Jim at least I create an hash table with local vars at every function call... and still it's not as slow as you may think, so maybe even to copy the C stack and the stackframes is viable. A bit slow... bust still useful, and requires very little work. So maybe we should try ;) In Jim there are tailcalls, anonymous functions, closures. It seems like that the other big missing abstraction is continuations.
DKF: I get very uneasy about tinkering around with the C stack. I know that it is definitely asking for trouble in Tcl's implementation as there are a number of places where we use pointers into the stack, which means that copying will not work. In theory, we could arrange for the byte-compiling of all core commands (we're actually pretty close) but that would massively raise the bar for anyone wanting to produce a third-party block-structured command in C, and there are lots of those about. Possible? Perhaps. Easy? No. Cheap to execute? Very unlikely (the stack can be multiple megabytes long). Worth it? You be the judge...
SS: It always works even if you have pointers on the stack. You save it, don't use it while it's saved. When you restore it, it gets copied at the same place it used to be. About complexity, I think I can implement this for Jim as an extension without to touch the core at all.
chi 19Apr2005: If you do, perhaps you should add refcounting to the Jim stack as well? Then it would be unnecessary to copy it. Simply IncrRefCount'ing every interesting element and then the Jim stack as well. The continuation could store a pointer to that shared stack then. Should be a bit cheaper then, no?
SS: Yes, I think my first try will be very simple, copying everything, just to avoid complexity that may make debugging of the C-level continuations harder, for the next step I was thinking about to add the refcount to individual stackframes instead to add it to the whole stack, this way a continuation is cheap to create and even if there is some copy-on-write occurring it's likely to interest only a minor part of the stack (few stackframes) in many uses. Btw if to copy the whole stack is reasonably fast I'll be more happy to have slow continuations as an extension that does not make the core more complex, so after I've a working "vanilla" version there will be to do some benchmarking to check if it's acceptable or too slow to be used.
JR 18Apr2005 What I think would be a nifty use for continuations would be to unnest vwait. As everyone knows, if you call vwait from inside an event handler then the outer vwait cannot return until the inner one does. This gets messy when things that you may not expect call vwait internally, with the general conclusion of vwait considered harmful. However, it is also very useful in some circumstances which makes this all a pain. If we had continuations, then vwait could simply be reimplemented as something like
proc vwait {var} {
uplevel trace variable $var w [info continuation 1]
return -code return
}Where info continuation level returns a command that executes the continuation. Then make the stack frames into Tcl_Objs and increment the reference counts to keep them around and it all works naturally. (if only it were that simple)
Lars H, 20 April 2005: I think this more script-level point of view deserves more attention. Discussing implementation techniques is important, but at the same time a bit premature before one knows what should be implemented. Sometimes what looks like a problem resolves itself when specs are being drafted.
One way of handling e.g. vwait continuations could be to have a pair of primitives suspend and resume. A usage example could look something like
proc A {data args} {
# ...
foreach datum $data {
# ...
if {[undecided $datum]} then {
puts "Going into suspended animation."
suspend
puts "But now I'm resumed!"
}
set result [getResultFor $datum]
# ...
}
# ...
}When suspend is evaluated, the entire local context is somehow recorded and put in suspension. Lets say it is all encoded into a Tcl_Obj which is put in the variable state_of_A. At some later time, something (e.g. an after or fileevent script) evaluates the command
resume $state_of_A
and this causes A to resume where it was suspended, thus printing "But now I'm resumed!" and continuing at precisely the right iteration of the foreach loop, with all variable values restored. I think this can serve as a first draft for continuation semantics; even if there is some feature that is missing, it should then be a basis for explaining what that feature is.
How would this kind of suspend/resume work, then? Recording all local variables and their values isn't too hard (that can probably even be done entirely on the script level), but recording the point in the program where execution was suspended is an entirely different matter. AFAICT there's no script level mechanism of determining that the call to a command (be that command suspend or whatever) was made in the second command of the then part of an if command that itself was the Nth command of a foreach body etc. As I understand it even Tcl doesn't know this too well, and this information can be presented in an errorInfo only because the call stack gets unwinded when an error occurs. But maybe this is how suspend should work too!
What if suspend was implemented as a sixth standard return code, that all commands would know and react to by recording their complete state in the interpreter result (or whereever)? This would certainly make suspend halt execution in precisely the expected way. resume has it a bit harder. Although it has all the necessary information, it will also need to be able to restart commands in a nonstandard position. This could require something as deep as a sibling of the Tcl_ObjCmdProc: a C function that is called when the command is to be started in a nondefault position. This naturally leads to the concept of resumable commands -- only those commands for which a resumeProc was supplied upon creation are resumable, and consequently all other commands should convert suspend returns to errors ("I'm not suspendable")!
And yes, suspend returns should be catchable. Indeed, that would be how to get the state into the above state_of_A variable in the first place:
if {[catch {A $data -someOption 1} state_of_A] == 5} then {
after 1000 {resume $state_of_A}
}NEM I'm not sure I understand this. You'd have to wait for the new exception to propagate all the way up to the toplevel in order to properly capture everything (catch only catches information up to this point). If you then immediately restore the context from the toplevel, it sort of works (though, of course, you have to restore to the next instruction -- i.e., the current continuation at the point that suspend was called). The big problem I see though, is that there is no mutation and no parameterization: when you restore, you seem to restore everything (all vars) to their original values, and can't pass an argument. Unless you rely on external state (e.g. in a database, or filesystem, or user input), then this is essentially pointless: you just repeat an exact same computation that you've already done. This isn't to say that there aren't alternatives to call/cc -- there are several. (e.g. shift/reset [L5 ] is similar to what you propose, I think).
Lars H: I can't comment on the shift/reset reference, since that seems to be available only to ACM members. Regarding the "no mutation, no parametrization":
But then again, what I wrote above was mostly thinking loud; I don't claim it's a finished solution. Indeed, upon reading the call-cc reference SS gives below, it occurred to me that there should be argument passing between suspend and resume. For passing in data, we can allow resume to take an extra (optional?) argument, and specify that this argument will be the value returned by the suspend that is resumed. Passing out data is in principle no problem either, since any arguments of suspend can become embedded into the state in such a way that they are easily inspected (e.g., they could be the first list element of the state), but one probably needs some convention for how to interpret them. Ideally, suspend/resume should be possible to use for several different high level constructions (e.g. vwait) within the same context, and they probably need to be treated slightly differently at the catching end.
SS: I think the only good way to implement continuations in Tcl is ala Scheme via a call-cc procedure [L6 ]. That's at least how I'll implement it in Jim at some point. This way is very general and powerful, and starting from call-cc one can create more high-level abstraction in Tcl itself.
Lars H: The reference you give isn't all that clear (and for a large part very Algol-language-family-centric, but then again, so was I before I discovered Tcl :-), but AFAICT this call/cc is very similar to the suspend I sketched above. The main difference is that call/cc isn't catchable, or rather can only be catched by the main event loop of the program, but that's not much of an advantage. Instead of having an explicit resume command the call/cc semantics seems to be to create an anonymous command for resuming at a particular point in the program, but that's just a matter of style (and what degree of obfuscation one desires). A cool use for really returning a frozen state is for more complete introspection (maybe even debugging).
AM (21 april 2005) Here is a small experiment with continuations or (more precisely) generators in pure Tcl - at least as I understand them to work.
There are lots of caveats and the implementation is not as elegant perhaps as it ideally would be - too much gets exposed at the user level, but, well, it is doing its intended job :).
# continuation.tcl --
# Experiment with continuations or, rather, generators
#
# generator --
# Create a generator procedure
# Arguments:
# name Name of the procedure
# arglist List of arguments
# _init_ Dummy (keyword init)
# initbody Body for the initialisation
# _repeat_ Dummy (keyword repeat)
# repeatbody Body for the main loop
# Result:
# None
# Side effect:
# Procedure created by name $name
#
proc generator {name arglist _init_ initbody _repeat_ repeatbody} {
set initbody [string map {yield gensave;return} $initbody]
set repeatbody [string map {yield gensave;return} $repeatbody]
proc $name [concat _ $arglist] \
[string map [list INITBODY $initbody REPEATBODY $repeatbody] {
upvar 1 $_ __
if { [lindex $__ 1] eq "init" } {
INITBODY
} else {
eval [lindex $__ 2]
REPEATBODY
}}]
}
# gencreate --
# Create an actual generator (or the context for running it)
# Arguments:
# name Name of the generator
# Result:
# Context for running the generator
#
proc gencreate {name} {
list $name init {}
}
# genrun --
# Run the generator
# Arguments:
# name Name of the _variable_ holding the context
# Result:
# Whatever the generator gives
#
proc genrun {name} {
upvar 1 $name context
[lindex $context 0] context
}
# gensave --
# Save the visible variables
# Arguments:
# None
# Result:
# None
# Side effect:
# The context is updated (the variable in the caller proc)
#
proc gensave {} {
upvar 1 __ context
set save {}
foreach v [uplevel 1 {info vars}] {
if { $v ne "__" && $v ne "_" } {
lappend save "set $v [uplevel [list set $v]]"
}
}
lset context 1 repeat
lset context 2 [join $save "\;"]
return
}
# Example --
# As an example: use a Fibonacci numbers generator twice
#
generator fib {} init {
set i 0
set j 1
yield 1
} repeat {
foreach {i j} [list $j [expr {$i+$j}]] {break}
yield $j
}
puts [info body fib]
set first [gencreate fib]
set second $first
for {set i 0} {$i < 20} {incr i} {
puts "First: [genrun first]"
if { $i > 10 } {
puts "Second: [genrun second]"
}
}SS the above code shows once again the flexibility of Tcl, but every case covered above is much more conveniently handled with a closure. For instance in Jim this is fully equivalent (but much more clear) of the above:
proc make-fib-generator {} {
lambda {} {{i 0} {j 1}} {
foreach {i j} [list $j [expr {$i+$j}]] break
return $j
}
}
set first [make-fib-generator]
set second [make-fib-generator]
for {set i 0} {$i < 20} {incr i} {
puts "First: [$first]"
if { $i > 10 } {
puts "Second: [$second]"
}
}The point of generators implemented with continuations is that they are similar to coroutines, i.e. you don't need to serialize by hand but instead you can write something like:
generator foobar {
set x 0
while {$x < 1000} {
incr x
yield $x
}
}Because there is no way to leave and re-enter a function without to lost the state there is no way to do this in pure Tcl. This is *very* important with some kind of algorithms. For instance to write a generator that yields an element of a binary tree at every iteration is trivial this way, because you can write the vanilla recursive procedure to print a tree in order but substitute puts with yield, and you are done.
AM Actually, with a bit extra code even that can be done - just add a list of variables in the lexical scope to the argument list of the generator proc. When generator is called it adds code in the init part to set variables with the same name to the value of those listed variables. (Oh, this sounds confusing, but the idea is just the same as in Jim closures)
MAKR I have certain bad feelings about this. Having just read a couple of articels about continuation I recognized the concept already dates back into the late 70s of the last century (an overview of publications: [L7 ]). It has not come very far since and there might be good reasons for this...
BUT, as pointed out in the vwait page, many people meeting vwait for the first time assume that it introduces parallelism and need to have it pointed out that "Multiple vwaits nest". If we could introduce continuations here we could bring vwait's behaviour into line with intuitive expectations. Colin Macleod
NEM The same arguments against continuations can also be made against eval, and various other features that make Tcl powerful. It took me a little while to grok how to use the event loop, and until I did it was hard for me to read event-oriented code. The same can be said for continuations: it takes a little while to get used to them, but you can build some incredibly powerful abstractions with them. If your organisation thinks that their developers won't be able to understand them, then you can always use coding standards and code reviews to outlaw their use. What exactly is an "ordinary software engineer"? Your statement seems dangerously close to saying that a software engineer should never be introduced to a new concept, for fear of upsetting their delicate minds! :) Regarding "goto": continuations are in some ways like goto, but they are first-class (goto labels are not), and they are parameterized (they are closures/procedures). So, you can think of them as a much more controlled goto.
SS: MAKR wrote It has not come very far since and there might be good reasons for this...: sure, most programmers are not good. The same applies to anonymous functions, closures, and many non trivially algol-like things in programming. This is also the same reason why Java is a very used language this days.