Stack-switching

This proposal adds typed stack-switching to WebAssembly, enabling a single WebAssembly instance to manage multiple execution stacks concurrently. The primary use-case for stack-switching is to add direct support for modular compilation of advanced non-local control flow idioms, e.g. coroutines, async/await, generators, lightweight threads, and so forth. This document outlines the new instructions and validation rules to facilitate stack-switching.

Motivation

Non-local control flow features provide the ability to suspend the current execution context and later resume it. Many industrial-strength programming languages feature a wealth of non-local control flow features such as async/await, coroutines, generators/iterators, effect handlers, and so forth. For some programming languages non-local control flow is central to their identity, meaning that they rely on non-local control flow for efficiency, e.g. to support massively scalable concurrency.

Rather than build specific control flow mechanisms for all possible varieties of non-local control flow, our strategy is to build a single mechanism, continuations, that can be used by language providers to construct their own language specific features.

Continuations

A continuation represents a snapshot of execution on a particular stack. Stack-switching is realised by instructions for suspending and resuming continuations. Continuations are composable, meaning that when a suspended continuation is resumed it is spliced onto the current continuation. This splicing establishes a parent-child relationship between the current and resumed continuation. In this respect the design provides a form of asymmetric switching.

When suspending, we provide a tag and payload, much like when raising an exception. Correspondingly, when resuming a suspended continuation a handler is installed which specifies different behaviours for the different kinds of tag with which the resumed continuation may subsequently be suspended. Unlike for a normal exception handler the handler is passed the suspended continuation as well as a payload.

We also offer an alternative to the interface based on suspending and resuming continuations by way of an instruction for direct switching between continuations. Direct switching combines suspending the current continuation with resuming a previously suspended peer continuation. Direct switching establishes a peer-to-peer relationship between the current continuation and its peer. In this respect the design provides a form of symmetric switching.

Introduction to continuation-based stack-switching

We illustrate the proposed stack-switching mechanism using two examples: generators and task scheduling. The generators example uses asymmetric stack-switching and the task scheduling example uses symmetric stack-switching.

Generators

The first example illustrates a generator-consumer pattern. Execution switches back and forth between a generator and a consumer execution stack. Whenever execution switches from the generator to the consumer the generator also passes a value to the consumer.

The stack-switching proposal reuses tags from the exception handling proposal. The following tag is used to coordinate between the generator and the consumer.

(tag $gen (param i32))

To switch execution to the consumer, the generator must suspend execution. The suspend instruction takes a tag, here $gen. The tag is used at runtime to determine how to continue execution, by identifying the active suspend handler for that tag. In our example, this handler will be provided by the consumer.

Much like with exceptions parameter types specify the types of values passed from the suspend site to the corresponding handler. In our example, the tag's single i32 parameter value is the value created by the generator that is passed to the consumer.

The proposal also extends tags with result types (not used in this example). These allow the types of values passed from a resume site to a previously suspended continuation when resuming it to be specified.

The overall module implementing our example has the following shape.

(module $generator
  (type $ft (func))
  ;; Types of continuations used by the generator:
  ;; No need for param or result types: No data passed back to the
  ;; generator when resuming it, and $generator function has no return
  ;; values.
  (type $ct (cont $ft))

  ;; Tag used to coordinate between generator and consumer: The i32 param
  ;; corresponds to the generated values passed; no values passed back from
  ;; generator to consumer.
  (tag $gen (param i32))


  (func $print (import "spectest" "print_i32") (param i32))

  ;; Simple generator yielding values from 100 down to 1
  (func $generator ...)
  (elem declare func $generator)

  (func $consumer ...)

)

The module defines continuation type $ct based on function type $ft. Suspended continuations of type (ref $ct) take no arguments and return no results. As we shall see, the generator and consumer manipulate suspended continuations of type (ref $ct).

The generator is defined as follows.

;; Simple generator yielding values from 100 down to 1
(func $generator
  (local $i i32)
  (local.set $i (i32.const 100))
  (loop $l
    ;; Suspend execution, pass current value of $i to consumer
    (suspend $gen (local.get $i))
    ;; Decrement $i and exit loop once $i reaches 0
    (local.tee $i (i32.sub (local.get $i) (i32.const 1)))
    (br_if $l)
  )
)

It executes 100 iterations of a loop and returns afterwards. Execution is suspended on each iteration using the $gen tag. The value passed from the suspend instruction to the handler (i.e., the value produced by the generator) is just the current value of the loop counter.

The consumer is defined as follows.

(func $consumer
 (local $c (ref $ct))
 ;; Create continuation executing function $generator.
 ;; Execution only starts when resumed for the first time.
 (local.set $c (cont.new $ct (ref.func $generator)))

 (loop $loop
   (block $on_gen (result i32 (ref $ct))
     ;; Resume continuation $c
     (resume $ct (on $gen $on_gen) (local.get $c))
     ;; $generator returned: no more data
     (return)
   )
   ;; Generator suspended, stack now contains [i32 (ref $ct)]
   ;; Save continuation to resume it in the next iteration
   (local.set $c)
   ;; Stack now contains the i32 value produced by $generator
   (call $print)

   (br $loop)
 )
)

It uses cont.new to create a continuation executing the generator. This instruction creates a value of reference type (ref $ct), saved in $c. It then runs a loop, where the resume instruction is used to continue execution of the continuation currently saved in $c on each iteration.

In general, a resume instruction may not only take a suspended continuation as an argument, but also additional values to be passed to the suspended continuation when it is resumed. These are specified in the parameters of the continuation's type. In our example, $ct has no parameters, indicating that no data is passed from the consumer to the generator.

When a suspended continuation is resumed it is spliced onto the current continuation (which may in fact be the top-level continuation corresponding to the main stack). This splicing establishes a parent-child relationship between the current and resumed continuation. This asymmetric relationship affects execution in two ways, which we now discuss.

First, in the resume instruction the handler clause (on $gen $on_gen) installs a suspend handler for that tag while executing the continuation. This means that if during execution of $c, the continuation executes the instruction suspend $gen, execution continues in the block $on_gen. In general, executing an instruction suspend $e for some tag $e means that execution continues at the innermost ancestor whose resume instruction installed a suspend handler for $e. This behaviour is directly analogous to the search for a matching exception handler after raising an exception. However, it is more general in that the handler is also passed the suspended current continuation. The extent of a suspended continuation captures execution from the instruction immediately following suspend $e up to the resume instruction that handles $e. In other words, as well as resuming a suspended computation and installing a handler, the resume instruction also acts as a delimiter for new suspended continuations created by performing suspend $e in the scope of the resumed continuation.

When the generator executes suspend $gen, execution continues in the $on_gen block in $consumer. In that case, two values are pushed onto the Wasm value stack. The topmost value is a new suspended continuation. It is the continuation of executing the generator following the suspend instruction (up to the handler). The other value is the i32 value passed from the generator to the consumer, as required by the tag's definition. In our example, the consumer simply prints the generated value and saves the new continuation in $c to be resumed in the next iteration.

Second, the parent-child relationship determines where execution continues after a continuation returns. Control simply transfers to the next instruction after the resume instruction that resumed the continuation in the parent, just as a normal function call returns to the instruction after its call site. For instance, in our example once the loop counter $i reaches 0, the $generator function returns and we have reached the end of the continuation. Execution then continues at the parent immediately after the resume instruction called by the consumer, and the consumer also returns.

The full definition of this module can be found here.

Task scheduling

The second example demonstrates how to implement task scheduling with the stack-switching instructions. Specifically, suppose we want to schedule a number of tasks, represented by functions $task_0 to $task_n, to be executed concurrently. Scheduling is cooperative, meaning that tasks explicitly yield execution so that a scheduler may pick the next task to run.

One approach is to use the asymmetric stack-switching approach we used for the generator example. We define a function $entry that resumes the initial task and installs a handler for a $yield tag inside an event loop. In order to yield execution, tasks simply perform (suspend $yield), transferring control back to the parent continuation, which is the event loop. The event loop then selects the next task (if any) from a task queue and resumes it.

This approach is illustrated by the following skeleton code.

(module $scheduler1

  (type $ft (func))
  ;; Continuation type of all tasks
  (type $ct (cont $ft))


  ;; Tag used to yield execution in one task and resume another one.
  (tag $yield)

  ;; Used by scheduler to manage task continuations
  (table $task_queue 1000 (ref null $ct))

  ;; Entry point, becomes parent of all tasks.
  ;; Also acts as scheduler when tasks yield or finish.
  (func $entry (param $initial_task (ref $ft))
    ;; initialise $task_queue with $initial_task
    ...
    (loop $resume_next
      ;; pick $next_task from queue, or return if no more tasks.
      ...
      (block $on_yield (result (ref $ct))
        (resume $ct (on $yield $on_yield) (local.get $next_task))
        ;; task finished execution
        (br $resume_next)
      )
      ;; task suspended: put continuation in queue, then loop to determine next
      ;; one to resume.
      ...
    )
  )

  (func $task_0
    ...
    ;; To yield execution, simply suspend to scheduling logic in $entry.
    (suspend $yield)
    ...
  )

  ...

  (func $task_n ...)

)

Note that $entry performs all scheduling; it is responsible for picking the next task to resume in two different circumstances: a) if the most recent task suspended itself, and b) if it simply ran to completion and returned. However, notice that this asymmetric approach requires two stack switches in order to change execution from one task to another: first when suspending from the a task to the event loop, and second when the event loop resumes the next task.

Our proposal provides a mechanism to optimise the particular pattern shown here, where suspending one continuation is followed by a handler resuming another continuation $c. We can use the switch instruction, which also relies on tags, to transfer control from the original continuation directly to $c, thus avoiding the need for an intermediate stack switch to the parent.

Concretely, executing switch $ct $yield (local.get $c) in our example behaves equivalently to (suspend $yield), assuming that the active (ordinary) handler for $yield immediately resumes $c and additionally passes the continuation obtained from handling $yield along as an argument to $c. However, as mentioned above, using a switch instruction here has the advantage that a Wasm engine can implement it directly using only a single stack switch. Each switch instruction is annotated with the type of the continuation switched to.

The key idea is to inline scheduling logic in the tasks themselves in order to reduce (or avoid altogether) the need to switch stacks to the event loop in order to implement switching tasks.

This alternative approach is illustrated by the following skeleton code.

(module $scheduler2
  (rec
    (type $ft (func (param (ref null $ct))))
    ;; Continuation type of all tasks
    (type $ct (cont $ft))
  )

  ;; Tag used to yield execution in one task and resume another one.
  (tag $yield)

  ;; Used by scheduler to manage task continuations
  (table $task_queue 1000 (ref null $ct))

  ;; Entry point, becomes parent of all tasks.
  ;; Only acts as scheduler when tasks finish.
  (func $entry (param $initial_task (ref $ft))
    ;; initialise $task_queue with $initial_task
    ...
    (loop $resume_next
      ;; pick $next_task from queue, or return if no more tasks.
      ;; Note that there is no suspend handler for $yield
      ...
      (resume $ct (on $yield switch) (ref.null $ct) (local.get $next_task))
      ;; task finished execution: loop to pick next one
      (br $resume_next)
      ...
    )
  )

  (func $task_0 (type $ft)
    ;; If $c is not null, put in task_queue.
    ...
    ;; To yield execution, call $yield_to_next
    (call $yield_to_next)
    ...
  )

  ...

  (func $task_n (type $ft) ...)

  ;; Determines next task to switch to directly.
  (func $yield_to_next
    ;; determine $next_task
    ...
    (block $done
      (br_if $done (ref.is_null (local.get $next_task)))
      ;; Switch to $next_task.
      ;; The switch instruction implicitly passes a reference to the currently
      ;; executing continuation as an argument to $next_task.
      (switch $ct $yield (local.get $next_task))
      ;; If we get here, some other continuation switch-ed directly to us, or
      ;; $entry resumed us.
      ;; In the first case, we receive the continuation that switched to us here
      ;; and we need to enqueue it in the task list.
      ;; In the second case, the passed continuation reference will be null.
      ...
    )
    ;; Just return if no other task in queue, making the $yield_to_next call
    ;; a noop.
  )

)

Here, the event loop is still responsible for resuming tasks from the task queue, starting with some initial task. Thus, it will still be the parent of all task continuations, as in the previous version. However, it is no longer responsible for handling suspensions of tasks. Instead, it only resumes the next task from the queue whenever the previously running task returns. Yielding execution from one task that merely wishes to suspend itself to another will be handled by the tasks themselves, using the switch instruction.

The fact that the event loop does not handle suspensions is reflected by the fact that its resume instruction does not install a suspend handler for the yield tag. Instead, the resume instruction installs a switch handler for tag yield. In order to yield execution a task calls a separate function $yield_to_next. The scheduling logic picks the next task $next_task and switches directly to it. Here, the target continuation (i.e., $next_task in $yield_to_next) receives the suspended current continuation (i.e., the one that just called $switch_to_next) as an argument. The payload passing mechanism used for integer values in the generator example is now used to pass continuation references. The task that we switched to is now responsible for enqueuing the previous continuation (i.e., the one received as a payload) in the task list.

As a minor complication, we need to encode the fact that the continuation switched to receives the current one as an argument in the type of the continuations handled by all scheduling logic. This means the type $ct must be recursive: a continuation of this type takes a value of type (ref null $ct) as a parameter. In order to give the same type to continuations that have yielded execution (those created by switch) and those continuations that correspond to beginning the execution of a $task_i function (those created by cont.new), we add a (ref null $ct) parameter to all of the $task_i functions. Finally, observe that the event loop passes a null continuation to any continuation it resumes, indicating to the resumed continuation that there is no previous continuation to enqueue in the task list.

Note that installing a switch handler for $yield in entry is strictly necessary. It acts as a delimiter, determining the extent of the suspended continuation created when performing switch with tag $yield. This form of stack-switching is symmetric in the following sense. Rather than switching back to the parent (as suspend would), switch effectively replaces the continuation under the handler for yield in the event loop with a different continuation.

The proposal also allows passing additional payloads when performing a switch instruction, besides the suspended current continuation. For simplicity, our example does not make use of this feature, as we can see from the type $ct, which has no further parameters besides the continuation argument required by switch. However, this mechanism could be used to optimise the implementation of task scheduling further.

In $scheduler2, if a $task_i function finishes and therefore returns, two stack switches are required to continue execution in the next task in the queue. This is due to the fact that the returning continuation switches to the parent (i.e., the event loop), which then resumes the next task. To avoid this additional stack switch, we could add boilerplate code to all of our task functions. Immediately before a task function would ordinarily return, it should instead switch to the next task. When doing so, it should pass a new flag to the target continuation to indicate that the source continuation should not be enqueued in the task list, but should instead be cancelled. Cancellation can be implemented using another instruction, resume_throw, which is described later in the document.

Full versions of $scheduler1 and $scheduler2 can be found here and here.

Instruction set extension

Here we give an informal account of the proposed instruction set extension. In the specification changes we give a more formal account of the validation rules and changes to the binary format.

For simplicity we ignore subtyping in this section, but in the specification changes we take full account of subtyping.

The proposal adds a new reference type for continuations.

  (cont $ft)

A continuation type is specified in terms of a function type $ft, whose parameter types t1* describe the expected stack shape prior to resuming/starting the continuation, and whose return types t2* describe the stack shape after the continuation has run to completion.

As a shorthand, we will often write the function type inline and write a continuation type as

  (cont [t1*] -> [t2*])

Declaring control tags

Control tags generalise exception tags to include result types. Operationally, a control tag may be thought of as a resumable exception. A tag declaration provides the type signature of a control tag.

  (tag $t (param t1*) (result t2*))

The $t is the symbolic index of the control tag in the index space of tags. The parameter types t1* describe the expected stack layout prior to invoking the tag, and the result types t2* describe the stack layout following an invocation of the operation.

We will often write $t : [t1*] -> [t2*] as shorthand for indicating that such a declaration is in scope.

Creating continuations

The following instruction creates a suspended continuation from a function.

  cont.new $ct : [(ref $ft)] -> [(ref $ct)]
  where:
  - $ft = func [t1*] -> [t2*]
  - $ct = cont $ft

It takes a reference to a function of type [t1*] -> [t2*] whose body may perform non-local control flow.

Invoking continuations

There are three ways to invoke a suspended continuation.

The first way to invoke a continuation is to resume the suspended continuation under a handler. The handler specifies what to do when control is subsequently suspended again.

  resume $ct hdl* : [t1* (ref $ct)] -> [t2*]
  where:
  - $ct = cont [t1*] -> [t2*]

The resume instruction is parameterised by a continuation type and a handler dispatch table hdl. The shape of hdl can be either:

(on $e $l) mapping the control tag $e to the label $l. Intercepting $e causes a branch to $l.
(on $e switch) allowing a direct switch with control tag $e.

The resume instruction consumes its continuation argument, meaning that a continuation may be resumed only once.

The second way to invoke a continuation is to raise an exception at the control tag invocation site which causes the stack to be unwound.

  resume_throw $ct $exn hdl* : [te* (ref $ct)])] -> [t2*]
  where:
  - $ct = cont [t1*] -> [t2*]
  - $exn : [te*] -> []

The resume_throw instruction is parameterised by a continuation type, the exception to be raised at the control tag invocation site, and a handler dispatch table. As with resume, this instruction also fully consumes its continuation argument. This instruction raises the exception $exn with parameters of type te* at the control tag invocation point in the context of the supplied continuation. As an exception is being raised (the continuation is not actually being supplied a value) the parameter types for the continuation t1* are unconstrained.

The third way to invoke a continuation is to perform a symmetric switch.

  switch $ct1 $e : [t1* (ref $ct1)] -> [t2*]
  where:
  - $e : [] -> [t*]
  - $ct1 = cont [t1* (ref $ct2)] -> [t*]
  - $ct2 = cont [t2*] -> [t*]

The switch instruction is parameterised by the type of the continuation argument ($ct1) and a control tag ($e). It suspends the current continuation (of type $ct2), then performs a direct switch to the suspended peer continuation (of type $ct1), passing in the required parameters (including the just suspended current continuation, in order to allow the peer to switch back again). As with resume and resume_throw, the switch instruction fully consumes its suspended continuation argument.

Suspending continuations

The current continuation can be suspended.

  suspend $e : [t1*] -> [t2*]
  where:
  - $e : [t1*] -> [t2*]

The suspend instruction invokes the control tag $e with arguments of types t1*. It suspends the current continuation up to the nearest enclosing handler for $e. This behaviour is similar to how raising an exception transfers control to the nearest exception handler that handles the exception. The key difference is that the continuation at the suspension point expects to be resumed later with arguments of types t2*.

Partial continuation application

A suspended continuation can be partially applied to a prefix of its arguments yielding another suspended continuation.

  cont.bind $ct1 $ct2 : [t1* (ref $ct1)] -> [(ref $ct2)]
  where:
  - $ct1 = cont [t1* t3*] -> [t2*]
  - $ct2 = cont [t3*] -> [t2*]

The cont.bind instruction binds a prefix of its arguments of type t1* to a suspended continuation of type $ct1, yielding a modified suspended continuation of type $ct2. The cont.bind instruction also consumes its continuation argument, and yields a new continuation that can be supplied to resume,resume_throw, switch or cont.bind.

Continuation lifetime

Producing continuations

There are four different ways in which continuations may be produced (cont.new,suspend,cont.bind,switch). A fresh continuation object is allocated with cont.new and the current continuation is reused with suspend, cont.bind, and switch.

The cont.bind instruction is similar to the func.bind instruction that was initially part of the function references proposal. However, whereas the latter necessitates the allocation of a new closure, as continuations are single-shot no allocation is necessary: all allocation happens when the original continuation is created by preallocating one slot for each continuation argument.

Consuming continuations

There are four different ways in which suspended continuations are consumed (resume,resume_throw,switch,cont.bind). A suspended continuation may be resumed with a particular handler with resume; aborted with resume_throw; directly switched to via switch; or partially applied with cont.bind.

In order to ensure that continuations are one-shot, resume, resume_throw, switch, and cont.bind destructively modify the suspended continuation such that any subsequent use of the same suspended continuation will result in a trap.

Design considerations

In this section we discuss some key design considerations.

Asymmetric switching

Resuming a suspended continuation establishes a parent-child relationship which aligns with the caller-callee relationship for standard function calls meaning that no special plumbing is needed in order to compose the non-local control features we define with built-in non-local control features such as traps, exceptions, and embedder integration.

Symmetric switching

Direct switching to a suspended peer continuation is semantically equivalent to suspending the current continuation with a special switch tag whose payload is the suspended peer continuation in the context of a handler which resumes the peer continuation. However, direct switching can (and should) be optimised to avoid the need to switch control to the handler before switching control to the peer.

Partial application

Partial application can be important in practice due to the block and type structure of Wasm, as in order to return a continuation from a block all branches within the block must agree on the type of continuation. Using cont.bind, a producer can ensure that the branches within a block each produce a continuation with the same type.

One-shot continuations

Continuations in the current proposal are single-shot (aka linear), meaning that they should be invoked exactly once. A continuation can be invoked either by resuming it (with resume); by aborting it (with resume_throw); or by switching to it (with switch). An attempt to invoke a continuation more than once results in a trap. Some applications such as backtracking, probabilistic programming, and process duplication exploit multi-shot continuations, but none of our critical use-cases requires multi-shot continuations.

Specification changes

This proposal is based on the function references proposal and exception handling proposal.

Types

We extend the structure of composite types and heap types as follows.

cont <typeidx> is a new form of composite type
- (cont $ft) ok iff $ft ok and $ft = [t1*] -> [t2*]

We add two new continuation heap types and their subtyping hierarchy:

heaptypes ::= ... | cont | nocont
nocont ok and cont ok always
nocont is the bottom type of continuation types, whereas cont is the top type, i.e. nocont <: cont

Instructions

The new instructions and their validation rules are as follows. To simplify the presentation, we write this:

C.types[$ct] ~~ cont [t1*] -> [t2*]

where we really mean this:

C.types[$ct] ~~ cont $ft
C.types[$ft] ~~ func [t1*] -> [t2*]

This abbreviation will be formalised with an auxiliary function or other means in the spec.

cont.new <typeidx>
- Create a new continuation from a given typed funcref.
- cont.new $ct : [(ref null $ft)] -> [(ref $ct)]
  - iff C.types[$ct] ~~ cont [t1*] -> [t2*]
cont.bind <typeidx> <typeidx>
- Partially apply a continuation.
- cont.bind $ct $ct' : [t3* (ref null $ct)] -> [(ref $ct')]
  - iff C.types[$ct] ~~ cont [t3* t1*] -> [t2*]
  - and C.types[$ct'] ~~ cont [t1'*] -> [t2'*]
  - and [t1*] -> [t2*] <: [t1'*] -> [t2'*]
resume <typeidx> hdl*
- Execute a given continuation.
  - If the executed continuation suspends with a control tag $t, the corresponding handler (on $t H) is executed.
- resume $ct hdl* : [t1* (ref null $ct)] -> [t2*]
  - iff C.types[$ct] ~~ cont [t1*] -> [t2*]
  - and (hdl : t2*)*
resume_throw <typeidx> <exnidx> hdl*
- Execute a given continuation, but force it to immediately throw the annotated exception.
- Used to abort a continuation.
- resume_throw $ct $e hdl* : [te* (ref null $ct)] -> [t2*]
  - iff C.types[$ct] ~~ cont [t1*] -> [t2*]
  - and C.tags[$e] : tag $ft
  - and C.types[$ft] ~~ func [te*] -> []
  - and (hdl : t2*)*
hdl = (on <tagidx> <labelidx>) | (on <tagidx> switch)
- Handlers attached to resume and resume_throw, handling control tags for suspend and switch, respectively.
- (on $e $l) : t*
  - iff C.tags[$e] = tag $ft
  - and C.types[$ft] ~~ func [t1*] -> [t2*]
  - and C.labels[$l] = [t1'* (ref null? $ct)]
  - and t1* <: t1'*
  - and C.types[$ct] ~~ cont [t2'*] -> [t'*]
  - and [t2*] -> [t*] <: [t2'*] -> [t'*]
- (on $e switch) : t*
  - iff C.tags[$e] = tag $ft
  - and C.types[$ft] ~~ func [] -> [t*]
suspend <tagidx>
- Use a control tag to suspend the current computation.
- suspend $t : [t1*] -> [t2*]
  - iff C.tags[$t] = tag $ft
  - and C.types[$ft] ~~ func [t1*] -> [t2*]
switch <typeidx> <tagidx>
- Switch to executing a given continuation directly, suspending the current execution.
- The suspension and switch are performed from the perspective of a parent (on $e switch) handler, determined by the annotated control tag.
- switch $ct1 $e : [t1* (ref null $ct1)] -> [t2*]
  - iff C.tags[$e] = tag $ft
  - and C.types[$ft] ~~ func [] -> [t*]
  - and C.types[$ct1] ~~ cont [t1* (ref null? $ct2)] -> [te1*]
  - and te1* <: t*
  - and C.types[$ct2] ~~ cont [t2*] -> [te2*]
  - and t* <: te2*

Execution

The same control tag may be used simultaneously by throw, suspend, switch, and their associated handlers. When searching for a handler for an event, only handlers for the matching kind of event are considered, e.g. only (on $e $l) handlers can handle suspend events and only (on $e switch) handlers can handle switch events. The handler search continues past handlers for the wrong kind of event, even if they use the correct tag.

Binary format

We extend the binary format of composite types, heap types, and instructions.

Composite types

Opcode	Type	Parameters	Note
-0x20	`func t1* t2*`	`t1* : vec(valtype)` `t2* : vec(valtype)`	from Wasm 1.0
-0x23	`cont $ft`	`$ft : typeidx`	new

Heap Types

The opcode for heap types is encoded as an s33.

Opcode	Type	Note
i >= 0	i	from function-references
-0x0b	`nocont`	new
-0x18	`cont`	new

Instructions

We use the use the opcode space 0xe0-0xe5 for the seven new instructions.

Opcode	Instruction	Immediates
0xe0	`cont.new $ct`	`$ct : u32`
0xe1	`cont.bind $ct $ct'`	`$ct : u32`, `$ct' : u32`
0xe2	`suspend $t`	`$t : u32`
0xe3	`resume $ct hdl*`	`$ct : u32` (for hdl see below)
0xe4	`resume_throw $ct $e hdl*`	`$ct : u32`, `$e : u32` (for hdl see below)
0xe5	`switch $ct1 $e`	`$ct1 : u32`, `$e : u32`

In the case of resume and resume_throw we use a leading byte to indicate the shape of hdl as follows.

Opcode	On clause shape	Immediates
0x00	`(on $t $h)`	`$t : u32`, `$h : u32`
0x01	`(on $t switch)`	`$t : u32`

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Explainer.md

Explainer.md

Stack-switching

Table of contents

Motivation

Continuations

Introduction to continuation-based stack-switching

Generators

Task scheduling

Instruction set extension

Declaring control tags

Creating continuations

Invoking continuations

Suspending continuations

Partial continuation application

Continuation lifetime

Producing continuations

Consuming continuations

Design considerations

Asymmetric switching

Symmetric switching

Partial application

One-shot continuations

Specification changes

Types

Tags

Instructions

Execution

Binary format

Composite types

Heap Types

Instructions

Files

Explainer.md

Latest commit

History

Explainer.md

File metadata and controls

Stack-switching

Table of contents

Motivation

Continuations

Introduction to continuation-based stack-switching

Generators

Task scheduling

Instruction set extension

Declaring control tags

Creating continuations

Invoking continuations

Suspending continuations

Partial continuation application

Continuation lifetime

Producing continuations

Consuming continuations

Design considerations

Asymmetric switching

Symmetric switching

Partial application

One-shot continuations

Specification changes

Types

Tags

Instructions

Execution

Binary format

Composite types

Heap Types

Instructions