Parameterizable return-from in Common Lisp

Question

I'm learning blocks in Common lisp and did this example to see how blocks and the return-from command work:

 (block b1 
            (print 1)
            (print 2)
            (print 3)
            (block b2 
                   (print 4)
                   (print 5)
                   (return-from b1)
                   (print 6)

               )
            (print 7))

It will print 1, 2, 3, 4, and 5, as expected. Changing the return-from to (return-from b2) it'll print 1, 2, 3, 4, 5, and 7, as one would expect.

Then I tried turn this into a function and paremetrize the label on the return-from:

 (defun test-block (arg)  (block b1 
            (print 1)
            (print 2)
            (print 3)
            (block b2 
                   (print 4)
                   (print 5)
                   (return-from (eval arg))
                   (print 6)

               )
            (print 7)))

and using (test-block 'b1) to see if it works, but it doesn't. Is there a way to do this without conditionals?

Answer 1

Using a conditional like CASE to select a block to return from

The recommended way to do it is using case or similar. Common Lisp does not support computed returns from blocks. It also does not support computed gos.

Using a case conditional expression:

(defun test-block (arg)
  (block b1 
    (print 1)
    (print 2)
    (print 3)
    (block b2 
      (print 4)
      (print 5)
      (case arg
        (b1 (return-from b1))
        (b2 (return-from b2)))
      (print 6))
    (print 7)))

One can't compute lexical go tags, return blocks or local functions from names

CLTL2 says about the restriction for the go construct:

Compatibility note: The ``computed go'' feature of MacLisp is not supported. The syntax of a computed go is idiosyncratic, and the feature is not supported by Lisp Machine Lisp, NIL (New Implementation of Lisp), or Interlisp. The computed go has been infrequently used in MacLisp anyway and is easily simulated with no loss of efficiency by using a case statement each of whose clauses performs a (non-computed) go.

Since features like go and return-from are lexically scoped constructs, computing the targets is not supported. Common Lisp has no way to access lexical environments at runtime and query those. This is for example also not supported for local functions. One can't take a name and ask for a function object with that name in some lexical environment.

Dynamic alternative: CATCH and THROW

The typically less efficient and dynamically scoped alternative is catch and throw. There the tags are computed.

Answer 2

I think these sorts of things boils down to the different types of namespaces bindings and environments in Common Lisp.

One first point is that a slightly more experienced novice learning Lisp might try to modify your attempted function to say (eval (list 'return-from ,arg)) instead. This seems to make more sense but still does not work.

Namespaces

A common beginner mistake in a language like scheme is having a variable called list as this shadows the top level definition of this as a function and stops the programmer from being able to make lists inside the scope for this binding. The corresponding mistake in Common Lisp is trying to use a symbol as a function when it is only bound as a variable.

In Common Lisp there are namespaces which are mappings from names to things. Some namespaces are:

• The functions. To get the corresponding thing either call it: (foo a b c ...), or get the function for a static symbol (function foo) (aka #'foo) or for a dynamic symbol (fdefinition 'foo). Function names are either symbols or lists of setf and one symbol (e.g. (serf bar)). Symbols may alternatively be bound to macros in this namespace in which case function and fdefinition signal errors.
• The variables. This maps symbols to the values in the corresponding variable. This also maps symbols to constants. Get the value of a variable by writing it down, foo or dynamically as (symbol-value). A symbol may also be bound as a symbol-macro in which case special macro expansion rules apply.
• Go tags. This maps symbols to labels to which one can go (like goto in other languages).
• Blocks. This maps symbols to places you can return from.
• Catch tags. This maps objects to the places which catch them. When you throw to an object, the implementation effectively looks up the corresponding catch in this namespace and unwinds the stack to it.
• classes (and structs, conditions). Every class has a name which is a symbol (so different packages may have a point class)
• packages. Each package is named by a string and possibly some nicknames. This string is normally the name of a symbol and therefore usually in uppercase
• types. Every type has a name which is a symbol. Naturally a class definition also defines a type.
• declarations. Introduced with declare, declaim, proclaim
• there might be more. These are all the ones I can think of.

The catch-tag and declarations namespaces aren’t like the others as they don’t really map symbols to things but they do have bindings and environments in the ways described below (note that I have used declarations to refer to the things that have been declared, like the optimisation policy or which variables are special, rather than the namespace in which e.g. optimize, special, and indeed declaration live which seems too small to include).

Now let’s talk about the different ways that this mapping may happen.

The binding of a name to a thing in a namespace is the way in which they are associated, in particular, how it may come to be and how it may be inspected.

The environment of a binding is the place where the binding lives. It says how long the binding lives for and where it may be accessed from. Environments are searched for to find the thing associated with some name in some namespace.

static and dynamic bindings

We say a binding is static if the name that is bound is fixed in the source code and a binding is dynamic if the name can be determined at run time. For example let, block and tags in a tagbody all introduce static bindings whereas catch and progv introduce dynamic bindings.

Note that my definition for dynamic binding is different from the one in the spec. The spec definition corresponds to my dynamic environment below.

Top level environment

This is the environment where names are searched for last and it is where toplevel definitions go to, for example defvar, defun, defclass operate at this level. This is where names are looked up last after all other applicable environments have been searched, e.g. if a function or variable binding can not be found at a closer level then this level is searched. References can sometimes be made to bindings at this level before they are defined, although they may signal warnings. That is, you may define a function bar which calls foo before you have defined foo. In other cases references are not allowed, for example you can’t try to intern or read a symbol foo::bar before the package FOO has been defined. Many namespaces only allow bindings in the top level environment. These are

• constants (within the variables namespace)
• classes
• packages
• types

Although (excepting proclaim) all bindings are static, they can effectively be made dynamic by calling eval which evaluates forms at the top level.

Functions (and [compiler] macros) and special variables (and symbol macros) may also be defined top level. Declarations can be defined toplevel either statically with the macro declaim or dynamically with the function proclaim.

Dynamic environment

A dynamic environment exists for a region of time during the programs execution. In particular, a dynamic environment begins when control flow enters some (specific type of) form and ends when control flow leaves it, either by returning normally or by some nonlocal transfer of control like a return-from or go. To look up a dynamically bound name in a namespace, the currently active dynamic environments are searched (effectively, ie a real system wouldn’t be implemented this way) from most recent to oldest for that name and the first binding wins.

Special variables and catch tags are bound in dynamic environments. Catch tags are bound dynamically using catch while special variables are bound statically using let and dynamically using progv. As we shall discuss later, let can make two different kinds of binding and it knows to treat a symbol as special if it has been defined with defvar or ‘defparameteror if it has been declared asspecial`.

Lexical environment

A lexical environment corresponds to a region of source code as it is written and a specific runtime instantiation of it. It (slightly loosely) begins at an opening parenthesis and ends at the corresponding closing parenthesis, and is instantiated when control flow hits the opening parenthesis. This description is a little complicated so let’s have an example with variables which are bound in a lexically environment (unless they are special. By convention the names special variables are wrapped in * symbols)

(defun foo ()
  (let ((x 10))
    (bar (lambda () x))))
(defun bar (f)
  (let ((x 20))
    (funcall f)))

Now what happens when we call (foo)? Well if x were bound in a dynamic environment (in foo and bar) then the anonymous function would be called in bar and the first dynamic environment with a binding for x would have it bound to 20.

But this call returns 10 because x is bound in a lexical environment so even though the anonymous function gets passed to bar, it remembers the lexical environment corresponding to the application of foo which created it and in that lexical environment, x is bound to 10. Let’s now have another example to show what I mean by ‘specific runtime instantiation’ above.

(defun baz (islast)
  (let ((x (if islast 10 20)))
    (let ((lx (lambda () x)))
      (if islast
        lx
        (frob lx (baz t))))))
(defun frob (a b)
  (list (funcall a) (funcall b)))

Now running (baz nil) will give us (20 10) because the first function passed to frob remembers the lexical environment for the outer call to baz (where islast is nil) whilst the second remembers the environment for the inner call.

For variables which are not special, let creates static lexical bindings. Block names (introduced statically by block), go tags (scopes inside a tagbody), functions (by felt or labels), macros (macrolet), and symbol macros (symbol-macrolet) are all bound statically in lexical environments. Bindings from a lambda list are also lexically bound. Declarations can be created lexically using (declare ...) in one of the allowed places or by using (locally (declare ...) ...) anywhere.

We note that all lexical bindings are static. The eval trick described above does not work because eval happens in the toplevel environment but references to lexical names happen in the lexical environment. This allows the compiler to optimise references to them to know exactly where they are without running code having to carry around a list of bindings or accessing global state (e.g. lexical variables can live in registers and the stack). It also allows the compiler to work out which bindings can escape or be captured in closures or not and optimise accordingly. The one exception is that the (symbol-)macro bindings can be dynamically inspected in a sense as all macros may take an &environment parameter which should be passed to macroexpand (and other expansion related functions) to allow the macroexpander to search the compile-time lexical environment for the macro definitions.

Another thing to note is that without lambda-expressions, lexical and dynamic environments would behave the same way. But note that if there were only a top level environment then recursion would not work as bindings would not be restored as control flow leaves their scope.

Closure

What happens to a lexical binding captured by an anonymous function when that function escapes the scope it was created in? Well there are two things that can happen

1. Trying to access the binding results in an error
2. The anonymous function keeps the lexical environment alive for as long as the functions referencing it are alive and they can read and write it as they please.

The second case is called a closure and happens for functions and variables. The first case happens for control flow related bindings because you can’t return from a form that has already returned. Neither happens for macro bindings as they cannot be accessed at run time.

Nonlocal control flow

In a language like Java, control (that is, program execution) flows from one statement to the next, branching for if and switch statements, looping for others with special statements like break and return for certain kinds of jumping. For functions control flow goes into the function until it eventually comes out again when the function returns. The one nonlocal way to transfer control is by using throw and try/catch where if you execute a throw then the stack is unwound piece by piece until a suitable catch is found.

In C there are is no throw or try/catch but there is goto. The structure of C programs is secretly flat with the nesting just specifying that “blocks” end in the opposite order to the order they start. What I mean by this is that it is perfectly legal to have a while loop in the middle of a switch with cases inside the loop and it is legal to goto the middle of a loop from outside of that loop. There is a way to do nonlocal control transfer in C: you use setjmp to save the current control state somewhere (with the return value indicating whether you have successfully saved the state or just nonlocally returned there) and longjmp to return control flow to a previously saved state. No real cleanup or freeing of memory happens as the stack unwinds and there needn’t be checks that you still have the function which called setjmp on the callstack so the whole thing can be quite dangerous.

In Common Lisp there’s a range of ways to do nonlocal control transfer but the rules are more strict. Lisp doesn’t really have statements but rather everything is built out of a tree of expressions and so the first rule is that you can’t nonlocally transfer control into a deeper expression, you may only transfer out. Let’s look at how these different methods of control transfer work.

block and return-from

You’ve already seen how these work inside a single function but recall that I said block names are lexically scoped. So how does this interact with anonymous functions?

Well suppose you want to search some big nested data structure for something. If you were writing this function in Java or C then you might implement a special search function to recurse through your data structure until it finds the right thing and then return it all the way up. If you were implementing it in Haskell then you would probably want to do it as some kind of fold and rely on lazy evaluation to not do too much work. In Common Lisp you might have a function which applies some other function passed as a parameter to each item in the data structure. And now you can call that with a searching function. How might you get the result out? Well just return-from to the outer block.

tagbody and go

A tagbody is like a progn but instead of evaluating single symbols in the body, they are called tags and any expression within the tagbody can go to them to transfer control to it. This is partly like goto, if you’re still in the same function but if your go expression happens inside some anonymous function then it’s like a safe lexically scoped longjmp.

catch and throw

These are most similar to the Java model. The key difference between block and catch is that block uses lexical scoping and catch uses dynamic scoping. Therefore their relationship is like that between special and regular variables.

Finally

In Java one can execute code to tidy things up if the stack has to unwind through it as an exception is thrown. This is done with try/finally. The Common Lisp equivalent is called unwind-protect which ensures a form is executed however control flow may leave it.

Errors

It’s perhaps worth looking a little at how errors work in Common Lisp. Which of these methods do they use?

Well it turns out that the answer is that errors instead of generally unwinding the stack start by calling functions. First they look up all the possible restarts (ways to deal with an error) and save them somewhere. Next they look up all applicable handlers (a list of handlers could, for example, be stored in a special variable as handlers have dynamic scope) and try each one at a time. A handler is just a function so it might return (ie not want to handle the error) or it might not return. A handler might not return if it invokes a restart. But restarts are just normal functions so why might these not return? Well restarts are created in a dynamic environment below the one where the error was raised and so they can transfer control straight out of the handler and the code that threw the error to some code to try to do something and then carry on. Restarts can transfer control using go or return-from. It is worth noting that it is important here that we have lexical scope. A recursive function could define a restart on each successive call and so it is necessary to have lexical scope for variables and tags/block names so that we can make sure we transfer control to the right level on the call stack with the right state.