Continuation Passing Style macro compiler for Clojure
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.

What is pulley?

pulley is the Positronic utility libraries. It is a collection of relatively small, simple libraries developed by Positronic Solutions, LLC. It is our pleasure to make them available to the public.

What is pulley.cps?

pulley.cps is part of the pulley collection of libraries. It provides a source-to-source compiler for transforming normal Clojure code to Clojure code in Continuation Passing Style (CPS), as well as runtime support for executing the transformed code. Here “normal Clojure code”, means code that is not written in CPS. It also means that most (but not quite all) code that can be compiled by Clojure can be transformed by the CPS compiler.

pulley.cps is licensed under the GNU Lesser General Public License (LGPL).


pulley.cps is currently very much in the alpha / pre-alpha stage. While the features that have been implemented are functional, neither the interfaces nor their implementation have undergone significant field testing. Interfaces are subject to change without notice.


pulley.cps allows full interaction between CPS and non-CPS code — CPS-transformed functions can call regular Clojure functions, and vice versa.

Currently supported features include:

  • application of both “native” and CPS functions (i.e., you can call a function regardless of whether it is a regular Clojure function or one that has been run through the CPS compiler)
  • calls back and forth between native and CPS functions
  • function definitions (you can create a fn)
  • exception support (throw, try, etc.),
  • special forms: . (dot), catch, def, do, finally, fn*, if, let*, letfn*, new, quote, throw, try, and var
  • any and all macros that expand to transformable code
  • collection literals

In other words, if there’s a reasoable way to transform the code, it either is supported or will be in the future.

pulley.cps also provides some features that are not normally available in Clojure:

  • Full tail-call optimization (TCO) — all tail calls are done in constant space (including binding forms).
  • Continuations (with some caveats)

Planned/Future Features

The following are currently under consideration for future work:

  • Common Lisp-style restarts
  • A way to “escape” transformation (for example, to use an unsupported form)
  • Smarter transformation (the current compiler uses what Matt Might calls the ”naive transformation”)
  • ClojureScript support

Unsupported Forms

It is the general goal of pulley.cps to support any and all forms that can reasonably be translated to CPS. However, there are some forms that are impossible to support directly, because they are deeply involved in Java interop.

For example, forms that define new Java types (deftype, definterface, genclass, etc.) are difficult to transform in a reasonable manner. Even though the method bodies (of, for example, a deftype) could be transformed, invocation of the method would involve the invocation of native code, thus defeating the ability to keep the same CPS context. It might be possible to work around this, but unless and until a suitable solution is developed such forms will not be supported.


To use pulley.cps in your project, simply include the following dependency in your project.clj:

[com.positronic-solutions/pulley.cps "0.2.2"]

Now you just need to require pulley.cps:

(require '[pulley.cps :as cps])

Basic Usage — Invoking the Compiler / Transforming Forms

The compiler is invoked via two macros:

  • cps-fn is used to define a function. It’s used just like clojure.core/fn, but the body(ies) of the function will be CPS transformed. It returns an IFn that executes the CPS code on a trampoline if invoked from a non-CPS context.
  • cps transforms a sequence of expressions. (cps <body...>) is equivalent to ((cps-fn [] <body...>)).


Defining a function

There are a couple options for defining a CPS function. The first way is to use def and cps-fn. cps-fn is the same as Clojure’s fn, except the function is CPS transformed.

(def factorial
  (cps/cps-fn [n]
    (if (> n 0)
      (* n (factorial (dec n)))

You can also use cps to wrap the entire definition. Then you can use defn:

(cps/cps (defn factorial [n]
           (if (> n 0)
             (* n (factorial (dec n)))

Tail Recursion

The previous example does not use tail recursion. While calls to factorial will use a constant amount of Java stack space, they will still consume heap space linearly with respect to n.

If, however, we transform factorial to use accumulator passing style, we can turn factorial into a tail-recursive function. Then calls to factorial will consume a constant amount of space with respect to n (ignoring any growth in the size of acc).

(def factorial
  (cps/cps-fn [n]
    (letfn [(factorial-aps [n acc]
              (if (> n 0)
                (factorial-aps (dec n) (* n acc))
      (factorial-aps n 1))))

State Machine

Of course, Tail Call Optimization (TCO) does not end with simple recursion. Another common application of TCO is implementing a state machine. Instead of encoding the machine as an explicit loop, we encode the machine as a set of mutually-recursive functions.

For example, suppose we have the following states and transitions:

even<empty><return :even>
odd<empty><return :odd>

If even is the initial state, then this machine will determine whether a given input sequence contains an even or odd number of zeroes.

We can impelement this as:

(def even-or-odd-number-of-zeros?
  (cps/cps-fn [inputs]
    (letfn [(even [s]
              (if (empty? s)
                (let [input (first s)]
                  (if (= input 0)
                    (odd (rest s))
                    (even (rest s))))))
            (odd [s]
              (if (empty? s)
                (let [input (first s)]
                  (if (= input 0)
                    (even (rest s))
                    (odd (rest s))))))]
      (even inputs))))


pulley.cps implements exceptions in a manner that is semantically compatible with native Clojure/Java. You can use throw, try, catch, etc. just like you normally would. Unhandled exceptions cross CPS boundaries, as expected.

In addition, the following functions and macros are provided and may be considered to be part of the public API:

Functional equivalent of throw.
Macro in the spirit of the homonymous Common Lisp construct. Provides a slight short-hand for (try ... (finally ...)) patterns.
Macro in the spirit of the homonymous Common Lisp macro.

Note that exception handling relies on an implicit stack of exception handling functions. In fact, try and friends will capture the current dynamic environment. This means forms that introduce an exception handler (try, handler-case, unwind-protect, etc.) will consume heap space throughout their entire extent. I.e., they are not tail-call optimizable. This really isn’t any different from Clojure/Java semantics (other than the storage place is the heap rather than the hardware stack), but it’s something to bear in mind.


pulley.cps supports “full” continuations with some caveats:

  • Within a particular CPS context, continuations can be considered “full” continuations with respect to that context.
  • However, continuations are implicitly delimited by the current CPS context. That is, when a continuation is captured, the continuation is implicitly delimited by the “top-most” transition from non-CPS to CPS code.

In other words, we try to capture full continuations to the extent we are able. However, since we can’t capture the continuation of non-CPS code, the captured continuation can’t cross certain boundaries.

Capturing the current continuation can be accomplished with call-cc. There’s also a macro version, let-cc, which may be more convenient in some cases. They are basically equivalent to Scheme’s call/cc and let/cc respectively.

CPS Overrides

In order to enhance the interoperability between CPS code and existing code, pulley.cps provides the ability to “override” an existing Clojure function or macro with a CPS implementation. Although it is possible to do this directly via low-level interfaces, it is recommended to use one of the following:

Overrides the given function with the provided implementation
A convenience macro that attempts to automatically override a function from its source. Success depends on the following conditions. Failure to meet either condition will result in an exception at compile time.
  1. clojure.repl/source-fn must be able to find the source of the function.
  2. The CPS compiler must be able to transform the code.
Given a function, will generate a CPS override for that function that prevents the function from being called from CPS code.
Overrides the given macro with the provided implementation when it is expanded by the CPS compiler.

Overridden Core Functions

pulley.cps provides CPS overrides for the following clojure.core functions:

  • apply
  • bound-fn*
  • get-thread-bindings
  • with-bindings*

These functions can be called within a CPS context without causing a transition to a non-CPS context per se. For example, if apply is called with a CPS context and is passed a CPS function, then that function will be called within the same CPS context. On the other hand, if apply is passed a non-CPS function then, while the call to apply itself will be within the same context, the actual application of the non-CPS function will necessarily cause a transition to a non-CPS context.

Forbidden Core Functions

There are currently two core functions that are forbidden:

  • push-thread-bindings
  • pop-thread-bindings

These are low-level functions that should not be used directly anyway. A higher-level construct, such as with-bindings should be used instead.

Setting and Testing Limits

By default, pulley.cps does not restrict transitions between CPS and non-CPS code in any way. However, sometimes you may want to limit the amount of interaction between native Clojure and CPS-transformed code. For example, if your code relies heavily on continuations, you may want to restrict function calls to CPS functions only. To this end, pulley.cps provides a number of dynamic vars:

When bound to a truthy value, only calls to CPS functions are allowed. If a non-CPS function is called, an IllegalStateException is thrown.
When bound to a falsey value, then attempts to spawn a trampoline when there is at least one other trampoline on the stack will fail. If a call to a CPS function is made from non-CPS code, then an IllegalStateException will be thrown if *trampoline-depth* >= 1.
Records the number of trampolines currently active on the stack.

While these vars can be used directly, it is recommended to use the following macros if possible:

Executes the body in an environment that does not allow CPS code to call non-CPS code (i.e., *strict-cps* is true).
Executes the body in an environment that does not allow recursive trampolines to be invoked (i.e., *allow-recursive-trampolines* is false).

Although everything in this section carries a fairly high risk of change, the macros will likely prove to be more stable.

Low-Level Implementation, Hooks, and Interfaces

Trampolines, Thunks, and Callables (runtime implementation details)

This section explains in some detail the fundamental runtime implementation of CPS code, as well as the protocols and other interfaces involved. This should be relied upon as little as possible, but the information is provided in case you absolutely need it, want to develop pulley.cps further, or are just plain curious in how pulley.cps operates.

CPS functions defined by cps-fn implement the IFn interface. This means they can be called as a normal function from non-CPS code. The implemented invoke methods in turn invoke a trampoline, via pulley.cps/trampoline.

The trampoline function accepts an ICallable object and the function arguments. ICallable is a protocol that serves as the CPS analog to IFn. ICallable defines a single method, with-continuation. with-continuation accepts a continuation and a dynamic environment, and must return an IFn which, when invoked, will execute the CPS code for the function with the continuation and environment provided via with-continuation. This IFn may return an IThunk, which represents the remainder of the computation. Otherwise, if anything other than an IThunk is returned, the trampoline terminates and returns that value.

A continuation is simply an IFn that takes a single argument — the result of the previous computation. A dynamic environment is a clojure.lang.Var$Frame object.

The call function provides a convenient way to call an ICallable. If you need to manually transform some code, it is recommended to use call instead of with-continuation directly. Note that invocations of call should almost certainly be delayed via an IThunk. The thunk macro provides a convenient way to construct an IThunk for a given expression(s).

The CPS compiler transforms all function calls to invocations of the call function. Thus all function calls from CPS code go through the ICallable interface.

ICallable is extended to IFn, providing a default implementation that invokes the IFn on the native stack. This allows non-CPS functions to be invoked by CPS code without any special handling — it’s all handled by ICallable.

Exception Handling

Exceptions are handled by dynamically binding *exception-handler* to a function for handling exceptions. When an exception is thrown, *exception-handler* is called with the thrown exception passed as the lone parameter.

with-exception-handler is a low-level macro for installing an exception handler. It accepts a form evaluating to a handler function and any number of body forms. with-exception-handler is CPS agnostic, meaning you can use it with equal effect from both CPS-transformed and non-transformed code (this is accomplished via override-macro!). When expanded by the CPS compiler, the handler function will first be bound to the current dynamic environment (via $bound-fn*), then dynamically bound to *exception-handler*. Finally, the body forms are executed. The non-cps version is similar, except it simply uses a try=/=catch block.

default-exception-handler is an exception handler that simply performs a native throw. trampoline binds *exception-handler* to default-exception-handler in the initial dynamic environment, ensuring any exceptions which are unhandled by CPS code are thrown natively.

The heart of call is wrapped in a try block. This allows us to catch exceptions thrown from native code. Exceptions caught in this manner are re-cast into the CPS domain by calling (rather than invoking) raise.

try, catch, and finally are implemented in terms of handler-case and unwind-protect.


Although most input forms are transformed either as macros or function applications, sometimes a form needs to be handled specially. This applies not only to Clojure special forms, but the occasional macro as well. For example, pulley.cps handles binding forms specially because CPS code uses a different mechanism to pass the dynamic environment than native Clojure does.

The *special-form-handlers* dynamic var provides hook to identify handlers for these forms. Since it is dynamic, the behaviour of the CPS compiler can be affected by altering the compile-time dynamic environment.

When a form is translated, and its operator is a symbol, the compiler performs the following checks.

  1. If *special-form-handlers* contains an entry for the symbol, then the form is passed to the associated function. This should be used to handle Clojure special forms, since special form symbols do not resolve to a Var.
  2. If the symbol resolves to a Var contained in *special-form-handlers*, then the form is passed to the associated function. This should be used to provide special handling of a Clojure macro.

So the keys of *special-form-handlers* are either symbols (for special forms) or Vars (for everything else). The values are functions that handle the form. In addition to the input form itself, the handler functions take a number of additional arguments. See the docstring for *special-form-handlers* for details.


We sincerely hope you enjoy using pulley.cps and are able to use it to your advantage. If you should find it lacking in some area, we hope you will consider contributing in one of the following ways:

  • Reporting bugs — If you think you’ve found a bug, don’t hesitate to open an issue on Github.
  • Requesting new features — We won’t know what you want unless you tell us. If you see we are lacking a feature you would like, please feel free to open an issue on Github or open a discussion on an appropriate channel.
  • Contributing code — As always, pull requests and patches are welcome. However, before investing a large portion of you time fixing a bug or implementing a new feature, you may wish to drop us a line so we can coordinate our efforts.