TIP 328: Coroutines

Login
Author:         Miguel Sofer <[email protected]>
Author:         Neil Madden <[email protected]>
State:          Final
Type:           Project
Vote:           Done
Created:        07-Sep-2008
Post-History:   
Keywords:       Coroutine,continuation,event-loop,NRE
Tcl-Version:    8.6

Abstract

This TIP recommends adding a coroutine mechanism to Tcl, loosely modelled on those present in Lua.

Rationale

The new Non-Recursive Engine (NRE) implemented in Tcl 8.6 allows support for a number of interesting features which have previously been difficult or impossible to implement efficiently in Tcl. One such feature is support for coroutines, which generalise the notion of subroutine by allowing a command to return multiple times. Conceptually, a coroutine allows a command to suspend its current execution and return (or yield) a value to its caller. The caller may later resume the coroutine at the point where it previously yielded, allowing it to perform further work and potentially yield further values. Coroutines have applications to a number of areas, in particular they allow a more natural representation of certain programming tasks when using the event loop, such as interacting with a Tk graphical user interface, performing asynchronous remote procedure calls, communicating between threads, and lightweight cooperative multitasking. These use-cases are discussed in more detail in the next section, along with a look at the limitations of the proposed approach, particularly in comparison to the closely related concept of continuations.

It should be noted that coroutines (or continuations) add no extra power to the language. Every computation that can be expressed with corouines can be expressed using existing features of the language by writing code in so-called continuation-passing style (CPS), which is the style of code used in loops and other control constructs. Rather, the benefit of coroutines is that it allows certain problems to be expressed more succinctly in cases where the CPS form is tedious to construct, or would require extensive refactoring of existing code. For instance, we can consider the following simple program that asks a user for two numbers and then displays their sum:

 proc tell msg { puts $msg }
 proc ask question {
     puts -nonewline "$question "
     flush stdout
     gets stdin
 }
 proc main {} {
     set x [ask "First number:"]
     set y [ask "Second number:"]
     tell "Sum = [expr {$x + $y}]"
 }
 main

In converting this code to work with a Tk GUI or some other asynchronous interface (such as a Web application) would generally require restructing the ask procedure to take a callback, and then restructuring the main application logic to supply these callbacks. This restructuring when moving from a synchronous to an asynchronous interface hinders such refactoring and also causes the application logic to become fragmented into a number of callback procedures. While anonymous procedures introduced in Tcl 8.5 can mitigate these drawbacks to some extent, there is still some obfuscation of the original code. The coroutine mechanism proposed would allow the interface of the tell and ask procedures to remain unchanged, and the main application logic to remain identical (beyond some trivial wrapping). It is this ability of coroutines to convert synchronous code to work in an asynchronous manner with minimal changes to the application logic that is the main motivating use-case for their consideration. Coroutines can also benefit new asynchronous code that is written from scratch, as the familiar direct style of coding that they enable is clearer to understand in many cases, and to some extent hides the complexity of the underlying asynchronous model. This added expressiveness helps the Tcl programmer to separate logic from control (Kowalksi, 1979).

The particular style of coroutines proposed are loosely based on the equivalent mechanism implemented in the Lua programming language (Lua, 2004). In particular, the coroutines implemented are asymmetric (yield and resume are separate commands), and stackful (yield can be called from any stack depth in the coroutine body). Asymmetric coroutines (similar to generators) can simulate symmetric coroutines and vice-versa.

New Commands and Subcommands

coroutine coroCmdName cmd ?arg ...?

yield ?value?

coroCmdName ?value?

info coroutine

coroutine evaluates the Tcl command

uplevel #0 [list cmd ?arg ...?]

until it returns or a yield is encountered. If yield is found then a command named coroCmdName will be created with special behaviour as described below.

yield suspends execution of the coroutine and returns returnValue to the coroutine caller (not the creator, which may be long dead) as the return value of the coroutine or coroCmdName invocation. It is an error to invoke yield outside of a coroutine's body. It is possible to yield from within a nested call, but under some circumstances yield can return an error (see Limitations).

coroCmdName resumes execution of the suspended coroutine. Execution of the suspended coroutine resumes with returnValue being the return value of the yield that last suspended execution. coroCmdName is garbage-collected: the command and all internal structs are deleted when the coroutine returns. A suspended coroutine is properly cleaned up when its command is rename'd to the empty string.

info coroutine returns the fully qualified name of the command that will resume the currently executing coroutine if it yields. In other words, it returns coroCmdName when invoked in a suspendable environment, and the empty string otherwise.

Whenever coroCmdName is invoked and the coroutine itself is running, Tcl's call stack looks exactly as if

uplevel #0 [list cmd ?arg ...?]

had been invoked instead. This structure is properly reflected in info level, info frame and the error stack.

Use Cases

In this section we review a number of motivating use-cases for coroutines.

Tk Graphical User Interfaces

As mentioned in the introduction, one particularly useful application of continuations in Tcl is to allow essentially asynchronous operations, using the event loop, to provide a familiar interface, much like synchronous operations. The example tell and ask interface can be written using yields to return control temporarily to the event loop until a result is available. This is similar to use of vwait, but avoids creating a nested event loop, with all the problems that entails:

 proc tell msg { tk_messageBox -message $msg }
 proc ask question {
     toplevel .ask
     pack [label .ask.l -text $question] [entry .ask.e]
     raise .ask; focus .ask.e
     bind .ask.e <Return> [list apply {cb {
         set ans [.ask.e get]
         destroy .ask
         $cb $ans
     }} [info coroutine]]
     yield
 }
 # ... main as before ...
 coroutine main-coro main

The original main procedure can be reused as-is. The only constraint is that it must be launched as a coroutine using the coroutine command. The operation of the code should be straight-forward to understand, as it is mostly standard Tk code. The ask command simply creates a Tk dialog box and registers the current coroutine as an event callback for when the user enters a number. Finally the procedure calls yield which suspends the current coroutine (in this case the main procedure) and allows the event loop to run. Once the user has entered a value and hit Return the coroutine is invoked with the value. This resumes the main procedure, returing this value as the return value of ask. From the point of view of the main routine it is as if the ask procedure worked exactly like the synchronous, console-based example.

Complex Iterators

Another common use of coroutines is to support efficient traversal of complex data structures. Here the coroutine is used to implement a stateful iterator for the data structure. In Tcl, one of the most natural and simple ways of writing a traversal interface for a data structure is as a custom control structure or loop. For instance, we can write a simple command for traversing a binary search tree in-order, applying a function to each value:

 # Constructor functions for our BST
 proc cons {name args} { proc $name $args { info level 0 } }
 cons Empty
 cons Branch left val right

 proc bst-map {f tree} {
     if {[lindex $tree 0] eq "Branch"} {
         bst-map $f [lindex $tree 1]
         {*}$f [lindex $tree 2]
         bst-map $f [lindex $tree 3]
     }
 }
 # Print every value in the tree in order
 bst-map puts $tree

While such interfaces are convenient for traversing a single data structure, they are more difficult to use when traversing multiple structures simultaneously, as when merging trees. A coroutine interface allows such custom loops to be easily converted into stateful iterators, allowing a merge to be written in a reasonably straight-forward fashion:

 proc bst-merge-map {f t1 t2} {
     set a [coroutine l bst-map yield $t1]
     set b [coroutine r bst-map yield $t2]
     while {[valid l] || [valid r]} {
         if {[valid l] && (![valid r] || $a < $b)} {
             {*}$f $a; set a [l]
         } else {
             {*}$f $b; set b [r]
         }
     }
 }
 proc valid cmd { llength [info commands $cmd] }
 bst-merge-map puts $t1 $t2

The benefit here is that the simple single-tree traversal function can be reused for merging multiple trees, simply by wrapping it in a coroutine.

Asynchronous Communications

Similar to the use of coroutines for event-based GUI programming, we can also use the mechanism for asynchronous networking, such as fetching data over HTTP, making remote procedure calls, and message-passing between threads. For example, asynchronous HTTP requests become as simple as synchronous ones:

 proc get url {
     http::geturl $url -command [info coroutine]
     yield
 }
 proc main {} {
     set t [get http://wiki.tcl.tk/4]
     puts [http::data $t]
     http::cleanup $t
 }
 coroutine m main

This common pattern of registering the current coroutine as a callback and then yielding can be used to accomplish a wide variety of such tasks.

Limitations

The coroutine implementation depends on the NRE enhancements recently made to the Tcl core. In particular, in order to capture a coroutine it is essential that all commands currently on the evaluation stack are NRE-aware. This is the case for most core Tcl commands, but at the time of writing few extensions will have made this transition. In these cases, trying to capture a coroutine while in an evaluation context containing a non-NRE-aware C command will result in an error. This situation is currently unavoidable, but we believe that it will improve over time as extensions are adapted to take advantage of the new features that NRE enables. In the meantime, the coroutine mechanism is still useful in a wide variety of situations, and the cases where it is not applicable should be easy to detect, as the code will fail immediately on trying to capture a coroutine. Nevertheless, library writers should be aware of the situation and avoid over reliance on coroutines. It is easy to wrap coroutine interfaces around existing callback-based library routines (as in the HTTP example).

Coroutines are roughly equivalent in expressive power to one-shot continuations. A continuation is simply a function that represents the rest of a computation. Continuations can be explicitly created, as in continuation- passing style (CPS) code, but some languages (notably Lisp and Scheme) allow the current execution context to be automatically captured as a continuation with a similar effect to coroutines. Such continuations can either be invoked once (a one-shot continuation) or several times (multi-shot). One-shot continuations can be used for much the same tasks that we have identified in this TIP, and can also be used to implement coroutines. Coroutines can likewise implement one-shot continuations. Multi-shot continuations, however, are strictly more expressive than either coroutines or one-shot continuations as they allow the same continuation to be resumed multiple times. In contrast, when a coroutine yields it can only be resumed once, and then must call yield again. An example of a construct that cannot be implemented using coroutines is a nondeterministic choice operator:

 set x [choose 1 2 3 4]
 set y [choose 3 9 7]
 if {$x**2 != $y} { fail }
 puts "x=$x,y=$y"

Intuitively, we would like this code to eventually succeed with values x = 3 and y = 9. However it is not possible to implement choose using coroutines as we can only yield from inside the choose statement once, and then are forced to yield instead from within fail, which is incorrect. A multi-shot continuation is capable of implementing choose as it can resume the same continuation multiple times with different values, essentially allowing the code to jump back to the appropriate choose statement for each back-track.

While a non-deterministic choice operator is an interesting use-case, it is not considered a primary motivation for this TIP. Such nondeterministic searches can be implemented using loops or custom control structures and exceptions (this essentially amounts to using CPS). For instance:

 foreach x {1 2 3 4} {
     foreach y {3 9 7} {
         if {$x**2 == $y} { puts "x=$x,y=$z" }
     }
 }

Furthermore, support for multi-shot continuations is believed to be more expensive to implement than coroutines, as the execution environment has to be copied for each continuation point, whereas a coroutine reuses the same environment for multiple yield/resume pairs. It is believed that the vast majority of useful use-cases will fall into the expressive power of coroutines. However, this limitation is real and should be taken into account when considering this TIP.

A final apparent limitation of the proposed mechanism is that it only supports passing a single argument when resuming a coroutine. The reason for this is simply that yield only returns a single result, and so a single argument is all that is required. It is believed that this will be sufficient for the majority of use-cases. However, in cases where multiple arguments are required, it is straight-forward to wrap the coroutine resume command so that these are passed as a list:

 proc resume {coro args} { $coro $args }

The coroutine body can then use lassign, lindex or some other means to extract the arguments. This can be used, for instance, when using a coroutine for variable tracing, such as in Colin Macleod's coroutine-enabled version of vwait that avoids nesting event loops http://wiki.tcl.tk/21555 :

 proc co_vwait varName {
     upvar $varName var
     set callback [list resume [info coroutine]]
     trace add    variable var write $callback
     yield
     trace remove variable var write $callback
 }

Along similar lines, the yield and coroCmdName commands currently do not offer support for communicating exceptions to/from coroutines. As with multiple arguments, this can be addressed by passing a dictionary of options (as produced by catch) as well as a value:

 proc exyield {value args} {
     lassign [yield [list $value $args]] value opts
     return -options $opts $value
 }
 proc exresume {coro value args} {
     lassign [$coro [list $value $args]] value args
     return -options $opts $value
 }
 # Usage
 proc mycoro {} { exyield $val -code error -errorcode $somecode ... }

It is expected that such wrappers can be added to Tcllib for now. Adding support for these options directly to the yield and resume command interfaces could be done in the future by a further TIP, if the functionality is deemed sufficiently critical.

Alternatives

An alternative syntax for coroutines is described in (Lua, 2004) based on a symmetric yield command. In this approach, yield takes as an extra argument the name of a coroutine to pass control to, rather than implicitly transferring control back to the caller. This eliminates the need for a separate resume interface (the coroCmdName in the current TIP). However, we believe that the asymmetric interface is more intuitive for most tasks, and the fact that coroutines are stateful entities requires them to have some named representation in any case, so this may as well be used as a resume command. It is possible to simulate symmetric coroutines with the current proposal using a simple loop:

 proc run-coros {coro value} {
     while 1 { lassign [$coro $value] coro value }
 }
 proc symyield {coro value} { yield [list $coro $value] }

An alternative to a coroutine mechanism would be to adopt some form of continuations into Tcl. As described previously, a continuation is a command that captures the current state of a computation (i.e. the current control stack and execution environment) and saves it so that it can be resumed later. In this respect, a continuation is similar to a coroutine. There are a number of varieties of continuations described in the literature (see http://library.readscheme.org/page6.html for a collection of references), with different expressive powers. Escape continuations only allow jumping to a context which is still on the stack. Such continuations are equivalent in power to exceptions, which Tcl already has. One-shot continuations can only be resumed once and are then discarded. As previously stated, such continuations are equivalent to coroutines as any particular yield can only be resumed once. Multi-shot continuations allow the same continuation to be resumed multiple times. These continuations go beyond the power of coroutines, and allow examples such as the nondeterministic choice operator to be implemented. Multi-shot continuations are however rather more expensive to implement, as each yield point requires creating a fresh copy of the execution environment, whereas a coroutine can reuse the same environment copy for multiple yield/resumes as it knows it cannot be resumed more than once for a single yield. A further refinement is the idea of delimited continuations which capture only a certain portion of the execution context (stack), rather than everything. Delimited continuations can simulate normal continuations by simply capturing the entire dynamic extent, but they have the advantage that they can also return a value, allowing them to be composed. Coroutines are also of this form, where the extent of the coroutine is limited to where the coroutine command was originally called from. The coroutine mechanism proposed is therefore of roughly equivalent power to a one-shot delimited continuation. Such a mechanism is strictly less powerful than a multi-shot continuation implementation, but it is believed that it is sufficient to cover the vast majority of useful use-cases, while remaining relatively simple to understand, and efficient to implement.

Lars Hellström has also proposed an alternative control mechanism for Tcl (Hellström, 2008), which is essentially a form of multi-shot full (non delimited) continuation, based on an interface of two commands: suspend and resume. The suspend command captures the current evaluation context into a continuation and then throws this as an exception with a special TCL_SUSPEND return code, allowing it to be caught and later resumed. The interface supports communicating values and exceptions to the resumed continuation. While the interface is potentially more expressive than the coroutine mechanism here, the proposed implementation involves destroying the stack as it is captured when the TCL_SUSPEND exception propagates. This presents a number of practical problems, such as the destruction of local variables (and thus the firing of unset traces), and the overall inefficiency of the approach. As coroutines thus captured are stateless objects they can be implemented as plain Tcl values (i.e., strings), albeit potentially quite large and complex ones. This provides the usual advantages of easy serialisation and transfer, while suffering the usual drawbacks of lack of encapsulation (potentially exposing implementation details) and potential inefficiencies due to shimmering or excess copying of data (a general problem of multi-shot continuations). While some of these problems can be limited or overcome entirely, it is our view that the current coroutines proposal covers a great deal of the expected use-cases in a much simpler and more efficient manner.

Implementation

An experimental implementation of the coroutine mechanism is available in Tcl 8.6a2 in CVS on sourceforge. The implementation is available in the ::tcl::unsupported namespace, exposing the coroutine, yield, and infoCoroutine commands.

References

(Kowalski, 1979) Robert Kowalski, "Algorithm = logic + control", Communications of the ACM 22(7), pp. 424-436, 1979.

(Lua, 2004) Ana Lûcia De Moura, Roberto Ierusalimschy, "Revisting Coroutines", Tech Report, 2004. http://www.inf.puc-rio.br/~roberto/docs/MCC15-04.pdf

(Hellström, 2008) Lars Hellström, "Suspend and Resume", http://wiki.tcl.tk/21537

Copyright

This document has been placed in the public domain.