23.3.2 Testing system features at application runtime

When pondering how to handle a difficult portability problem or configurable option, consider whether the problem is better solved by performing tests at runtime or by providing a configuration file to customize the application. Keep in mind that the results of tests that Autoconf can perform will ultimately affect how the program will be built--and can limit the number of machines that the program can be moved to without recompiling it. Here is an example where this consideration had to be made in a real life project:

The pthreads for Win32 project has sought to provide a standards compliant implementation for the POSIX threads API. It does so by mapping the POSIX API functions into small functions which achieve the desired result using the Win32 thread API. Windows 95, Windows 98 and Windows NT have different levels of support for a system call primitive that attempts to enter a critical section without blocking. The TryEnterCriticalSection function is missing on Windows 95, is an inoperative stub on Windows 98, and works as expected on Windows NT. If this behavior was to be checked by `configure' at compile time, then the resultant library would only work on the variant of Windows that it was compiled for. Because it's more common to distribute packages for Windows in binary form, this would be an unfortunate situation. Instead, it is sometimes preferable to handle this kind of portability problem with a test, performed by your code at runtime.