Lock Options

Locks and Types

First let's talk about the way to encode locking options in the function's signature. I took inspiration from Scott Meyers's article "Enforcing Code Feature Requirements in C++" (http://www.artima.com/cppsource/codefeatures.html). His idea was to encode compile-time information about a function in the type of a dummy argument. In my scheme, a dummy parameter is used to encode the function's locking options. This parameter is forwarded from callers to callees. The compiler makes sure that, for every call, the type of the actual dummy argument is compatible (convertible) to the type of the formal dummy parameter. Such a set of types with appropriate conversions is best implemented as a hierarchy of interfaces in D (classes in C++).

For simplicity, assume that locks are given names, and the global sequence is sorted alphabetically. There should be a separate type for "I may take locks A, B, and D" and another for "I may take lock D", and so on. Let's call the first type A_B_D and the second type D.

A derived interface is implicitly convertible to any of its base interfaces (and their bases, transitively). Figure 1 shows the interface hierarchy for three locks—A, B, and C. The top of the hierarchy is the special interface called NoLocks. It can be used to mark a function that doesn't take any locks. Interface A_B describes the option to take locks A and B (in this order), and so on. Listing One presents the actual declaration of such a hierarchy.

It's clear from Figure 1 that a type describing wider lock options can be implicitly converted to any type describing narrower lock options. In other words, it is okay for a function that has the option to take a larger set of locks to call a function that has the option to take a subset of those locks, but not the other way around.

[Click image to view at full size]

Figure 1: Interface hierarchy.

In what follows, I call the dummy variable lockOptions. Below is an example of a function f that declares the option to take locks A, B, and C. It calls another function, g, which declares the option to take a subset of those locks, namely B and C. The call compiles because the type A_B_C is implicitly convertible to the type B_C:

int g(B_C lockOptions); 
int f(int x, A_B_C lockOptions) { 
  return g(lockOptions); // OK 
} 

The following calls, however, will not compile, because the appropriate conversions are missing:

void g1(A lockOptions); 
void g2(A_B_C lockOptions); 
void f1(B_C lockOptions) { 
  g1(lockOptions); // type error 
  g2(lockOptions); // type error 
}