LCA: Lessons from 30 years of Sendmail

First of all, big-O is the worst-case upper bound, not the average case.

> O(m * n), where m = number of source files and n = number of ubiquitously
> included header files, is rather closer to the mark.

Header files tend to include other header files. This is a consequence of one of the other design features of C++, the fact that the definition of a class cannot be spread across multiple files.

So you often see stuff like this:
> #include "private_helper.h"
>
> class MyClass {
> ...
> private:
> PrivateHelper myPrivateHelper;
> };

PrivateHelper is not be part of the public API of the class (hopefully), but even so, it will require you to include private_helper.h. That class itself might have its own private helpers... and so it goes, on and on.

What's that, you say? I can use the pImpl idiom? Sure, if I can live with reduced performance and more boilerplate code to initialize and destroy the pImpl. Sigh.

It would be interesting to graph, say, WebKit compilation times as a function of the number of files. I really doubt it's anywhere close to linear.