13.7. The Semantics Of Parameters

13.7.1. Pass-By-Value
13.7.2. What's Wrong?
13.7.3. Call-By-Reference

So far we've dealt with the syntax of parameter passing, and we've got the parsing mechanisms in place to handle it. Next, we have to look at the semantics, i.e., the actions to be taken when we encounter parameters. This brings us square up against the issue of the different ways parameters can be passed.

There is more than one way to pass a parameter, and the way we do it can have a profound effect on the character of the language. So this is another of those areas where I can't just give you my solution. Rather, it's important that we spend some time looking at the alternatives so that you can go another route if you choose to.

There are two main ways parameters are passed:

The differences are best seen in the light of a little history.

The old FORTRAN compilers passed all parameters by reference. In other words, what was actually passed was the address of the parameter. This meant that the called subroutine was free to either read or write that parameter, as often as it chose to, just as though it were a global variable. This was actually quite an efficient way to do things, and it was pretty simple since the same mechanism was used in all cases, with one exception that I'll get to shortly.

There were problems, though. Many people felt that this method created entirely too much coupling between the called subroutine and its caller. In effect, it gave the subroutine complete access to all variables that appeared in the parameter list.

Many times, we didn't want to actually change a parameter, but only use it as an input. For example, we might pass an element count to a subroutine, and wish we could then use that count within a DO-loop. To avoid changing the value in the calling program, we had to make a local copy of the input parameter, and operate only on the copy. Some FORTRAN programmers, in fact, made it a practice to copy all parameters except those that were to be used as return values. Needless to say, all this copying defeated a good bit of the efficiency associated with the approach.

There was, however, an even more insidious problem, which was not really just the fault of the “pass by reference” convention, but a bad convergence of several implementation decisions.

Suppose we have a subroutine:

        SUBROUTINE FOO(X, Y, N)

where N is some kind of input count or flag. Many times, we'd like to be able to pass a literal or even an expression in place of avariable, such as:

        CALL FOO(A, B, J + 1)

Here the third parameter is not a variable, and so it has no address. The earliest FORTRAN compilers did not allow such things, so we had to resort to subterfuges like:

        K = J + 1
        CALL FOO(A, B, K)

Here again, there was copying required, and the burden was on the programmer to do it. Not good.

Later FORTRAN implementations got rid of this by allowing expressions as parameters. What they did was to assign a compiler-generated variable, store the value of the expression in the variable, and then pass the address of the expression.

So far, so good. Even if the subroutine mistakenly altered the anonymous variable, who was to know or care? On the next call, it would be recalculated anyway.

The problem arose when someone decided to make things more efficient. They reasoned, rightly enough, that the most common kind of “expression” was a single integer value, as in:

        CALL FOO(A, B, 4)

It seemed inefficient to go to the trouble of “computing” such an integer and storing it in a temporary variable, just to pass it through the calling list. Since we had to pass the address of the thing anyway, it seemed to make lots of sense to just pass the address of the literal integer, 4 in the example above.

To make matters more interesting, most compilers, then and now, identify all literals and store them separately in a “literal pool,” so that we only have to store one value for each unique literal. That combination of design decisions: passing expressions, optimization for literals as a special case, and use of a literal pool, is what led to disaster.

To see how it works, imagine that we call subroutine FOO as in the example above, passing it a literal 4. Actually, what gets passed is the address of the literal 4, which is stored in the literal pool. This address corresponds to the formal parameter, K, in the subroutine itself.

Now suppose that, unbeknownst to the programmer, subroutine FOO actually modifies K to be, say, -7. Suddenly, that literal 4 in the literal pool gets changed, to a -7. From then on, every expression that uses a 4 and every subroutine that passes a 4 will be using the value of -7 instead! Needless to say, this can lead to some bizarre and difficult-to-find behavior. The whole thing gave the concept of pass-by-reference a bad name, although as we have seen, it was really a combination of design decisions that led to the problem.

In spite of the problem, the FORTRAN approach had its good points. Chief among them is the fact that we don't have to support multiple mechanisms. The same scheme, passing the address of the argument, works for every case, including arrays. So the size of the compiler can be reduced.

Partly because of the FORTRAN gotcha, and partly just because of the reduced coupling involved, modern languages like C, Pascal, Ada, and Modula 2 generally pass scalars by value.

This means that the value of the scalar is copied into a separate value used only for the call. Since the value passed is a copy, the called procedure can use it as a local variable and modify it any way it likes. The value in the caller will not be changed.

It may seem at first that this is a bit inefficient, because of the need to copy the parameter. But remember that we're going to have to fetch some value to pass anyway, whether it be the parameter itself or an address for it. Inside the subroutine, using pass-by-value is definitely more efficient, since we eliminate one level of indirection. Finally, we saw earlier that with FORTRAN, it was often necessary to make copies within the subroutine anyway, so pass-by-value reduces the number of local variables. All in all, pass-by-value is better.

Except for one small little detail: if all parameters are passed by value, there is no way for a called to procedure to return a result to its caller! The parameter passed is not altered in the caller, only in the called procedure. Clearly, that won't get the job done.

There have been two answers to this problem, which are equivalent. In Pascal, Wirth provides for VAR parameters, which are passed-by-reference. What a VAR parameter is, in fact, is none other than our old friend the FORTRAN parameter, with a new name and paint job for disguise. Wirth neatly gets around the “changing a literal” problem as well as the “address of an expression” problem, by the simple expedient of allowing only a variable to be the actual parameter. In other words, it's the same restriction that the earliest FORTRANs imposed.

C does the same thing, but explicitly. In C, all parameters are passed by value. One kind of variable that C supports, however, is the pointer. So by passing a pointer by value, you in effect pass what it points to by reference. In some ways this works even better yet, because even though you can change the variable pointed to all you like, you still can't change the pointer itself. In a function such as strcpy, for example, where the pointers are incremented as the string is copied, we are really only incrementing copies of the pointers, so the values of those pointers in the calling procedure still remain as they were. To modify a pointer, you must pass a pointer to the pointer.

Since we are simply performing experiments here, we'll look at both pass-by-value and pass-by-reference. That way, we'll be able to use either one as we need to. It's worth mentioning that it's going to be tough to use the C approach to pointers here, since a pointer is a different type and we haven't studied types yet!