Lazy Evaluation Of Function Arguments

Lazy evaluation is the technique of not evaluating an expression unless and until the result of the expression is required. The &&, || and ?: operators are the conventional way to do lazy evaluation:

void test(int* p)
{
    if (p && p[0])
    ...
}

The second expression p[0] is not evaluated unless p is not null. If the second expression was not lazily evaluated, it would generate a runtime fault if p was null.

While invaluable, the lazy evaluation operators have significant limitations. Consider a logging function, which logs a message, and can be turned on and off at runtime based on a global value:

void log(char[] message)
{
    if (logging)
        fwritefln(logfile, message);
}

Often, the message string will be constructed at runtime:

void foo(int i)
{
    log("Entering foo() with i set to " ~ toString(i));
}

While this works, the problem is that the building of the message string happens regardless of whether logging is enabled or not. With applications that make heavy use of logging, this can become a terrible drain on performance.

One way to fix it is by using lazy evaluation:

void foo(int i)
{
    if (logging)
        log("Entering foo() with i set to " ~ toString(i));
}

but this violates encapsulation principles by exposing the details of logging to the user. In C, this problem is often worked around by using a macro:

#define LOG(string)  (logging && log(string))

but that just papers over the problem. Preprocessor macros have well known shortcomings:

• The logging variable is exposed in the user's namespace.
• Macros are invisible to symbolic debuggers.
• Macros are global only, and cannot be scoped.
• Macros cannot be class members.
• Macros cannot have their address taken, so cannot be passed indirectly like functions can.

A robust solution would be a way to do lazy evaluation of function parameters. Such a way is possible in the D programming language using a delegate parameter:

void log(char[] delegate() dg)
{
    if (logging)
        fwritefln(logfile, dg());
}

void foo(int i)
{
    log( { return "Entering foo() with i set to " ~ toString(i); });
}

Now, the string building expression only gets evaluated if logging is true, and encapsulation is maintained. The only trouble is that few are going to want to wrap expressions with { return exp; }.

So D takes it one small, but crucial, step further (suggested by Andrei Alexandrescu). Any expression can be implicitly converted to a delegate that returns either void or the type of the expression. The delegate declaration is replaced by the lazy storage class (suggested by Tomasz Stachowiak). The functions then become:

void log(lazy char[] dg)
{
    if (logging)
        fwritefln(logfile, dg());
}

void foo(int i)
{
    log("Entering foo() with i set to " ~ toString(i));
}

which is our original version, except that now the string is not constructed unless logging is turned on.

Any time there is a repeating pattern seen in code, being able to abstract out that pattern and encapsulate it means we can reduce the complexity of the code, and hence bugs. The most common example of this is the function itself. Lazy evaluation enables encapsulation of a host of other patterns.

For a simple example, suppose an expression is to be evaluated count times. The pattern is:

for (int i = 0; i < count; i++)
    exp;

This pattern can be encapsulated in a function using lazy evaluation:

void dotimes(int count, lazy void exp)
{
    for (int i = 0; i < count; i++)
        exp();
}

It can be used like:

void foo()
{
    int x = 0;
    dotimes(10, write(x++));
}

which will print:

0123456789

More complex user-defined control structures are possible. Here's a method to create a switch like structure:

bool scase(bool b, lazy void dg)
{
    if (b)
        dg();
    return b;
}

/* Here the variadic arguments are converted to delegates in this
   special case.
 */
void cond(bool delegate()[] cases ...)
{
    foreach (c; cases)
    {
        if (c())
            break;
    }
}

which can be used like:

void foo()
{
    int v = 2;
    cond
    (
    scase(v == 1, writeln("it is 1")),
    scase(v == 2, writeln("it is 2")),
    scase(v == 3, writeln("it is 3")),
    scase(true,   writeln("it is the default"))
    );
}

which will print:

it is 2

Those familiar with the Lisp programming language will notice some intriguing parallels with Lisp macros.

For a last example, there is the common pattern:

Abc p;
p = foo();
if (!p)
    throw new Exception("foo() failed");
p.bar();    // now use p

Because throw is a statement, not an expression, expressions that need to do this need to be broken up into multiple statements, and extra variables are introduced. (For a thorough treatment of this issue, see Andrei Alexandrescu and Petru Marginean's paper Enforcements). With lazy evaluation, this can all be encapsulated into a single function:

Abc Enforce(Abc p, lazy char[] msg)
{
    if (!p)
        throw new Exception(msg());
    return p;
}

and the opening example above becomes simply:

Enforce(foo(), "foo() failed").bar();

and 5 lines of code become one. Enforce can be improved by making it a template function:

T Enforce(T)(T p,  lazy char[] msg)
{
    if (!p)
        throw new Exception(msg());
    return p;
}

Conclusion

Lazy evaluation of function arguments dramatically extends the expressive power of functions. It enables the encapsulation into functions of many common coding patterns and idioms that previously were too clumsy or impractical to do.

Acknowledgements

I gratefully acknowledge the inspiration and assistance of Andrei Alexandrescu, Bartosz Milewski, and David Held. The D community helped a lot with much constructive criticism, such as the thread starting with Tomasz Stachowiak in D/41633.