Why use a PSR-0 or PSR-4 autoload in composer if classmap is actually faster?
I understand that you can use either a PSR standard to locate files, or tell composer a directory to scan for classes. The documentation recommends using the PSR-4 standard. There is also an option for composer to create an optimized autoloader, which basically generates a full classmap. So why should one use PSR-4 at all if the best way to load is with a classmap?
It makes sense to me to keep the directory structure, since that is a good way to organize anyway. However, it seems like the logical option would be to use PSR-4 loading on development machines, and then classmap for the production environment. That way, you don't have to rebuild your classmap every time you create a new class, but the production environment creates a complete one as a part of the deployment process without an additional call to
./composer.phar dump-autoload -o
Answer
The problem is that the classmap is NOT actually faster in every case!
The speed of the classmap comes from not having to check the filesystem if a file exists before doing the always necessary work of loading it, parsing it (opcode caches will help here) and then executing it.
But the downside of the classmap is that you possibly generate a huge amount of data for every single class, interface and trait included in the libraries you use, without you actually using it in your production code. Loading huge arrays does not come for free - while the code need not be parsed again and again (opcode cache), it still has to be executed, the array data structure has to be put into memory, filled with lots of strings, and then eats up some amount of memory that might have been usable for something else.
I found two resources discussing this topic: First of all there is github issue #1529 suggesting further improvements for the composer autoloader using a bunch of symlinks to avoid having to scan multiple directories.
The discussion there also reveals that you should actually try to use the best possible namespace- or classname-prefix in the PSR-0 autoload declaration, i.e. the longest one possible. You can also use more than one prefix in the declaration.
Then there is a blog post linked in that issue that documents some xhprof benchmarks using a stock EZPublish 5 and fiddling with the settings, including APC Caching and classmap dumping.
Money quote:
This command created a 662KiB vendor/composer/autoload_classmap.php file containing an array that is a hash composed of the class name as index and the path to the file containing the class definition as value. At the time I am writing this post, this array is composed of 4168 entries. [...] Although it should give us the most efficiant autoloading mechanism, it actually slows things down (from 254.53 reqs/second to 197.95). The reason being that even if the file is cached by APC, the PHP array containing the map with more than 4100 entries needs to be re-created at every single request.
Will a classmap be fast? Certainly. Fastest in every case? Of course not - it depends on the ratio used vs. unused classes per request. So even if on average your application actually uses ALL classes in the map, a classmap might still be slower if you only use about 10% of the classes per request, and you'd be better off optimizing the autoload declarations of the libraries you use. In fact, every classname prefix should only ever point to exactly one directory.
Note that the performance gain you'd achieve only is in the area of about low single digit milliseconds per request. Your application surely is awesome if that figure is a significant performance boost in the range of 5 to 10%. But if you really are in that performance range, blindly believing that a classmap is ALWAYS faster probably wastes a lot of unnecessary CPU cycles.
If you optimize something: Measure it! How would you know if it actually becomes better if you cannot measure it?