There seems to still be confusion about this topic, however in most cases it comes down to ease vs performance.
A good mailing list thread to read would be this one on Zend Frameworks mailing list:
http://n4.nabble.com/ZF-and-Autoloading-td640085i20.html
Now, the correlation is here because if you inherit from not-yet-defined
class, you might rely on autoload to define it (though you might also
rely on include), and actually the presence of the autoload facility may
encourage you to use such inheritance. But this is not the autoload
which brings trouble (see after Ramus' "it's not just autoload" in the
blog for some examples of troublesome things).
So the right phrase would be "people which tend to rely on autoload tend
also to use code which defies compile-time binding". Which can't be seen
as autoload fault, of course, and just avoiding autoload won't help a
bit with that - you would also have to rewrite your code so that
compile-time binding could happen. And it has nothing to do with uses of
autoload with "new", for example.
As for the slowdown from the effects described above - i.e., absence of
the compile-time binding - the code indeed becomes a bit slower and such
code can lead in some obscure cases to some trouble to opcode caches
(not in the autoload cases - but in cases where classes are defined
inside conditions, or, God forbid, different definition is created
depending on condition) - but it has next to nothing to do with using
autoload by itself.
The amount of slowdown, however, seem to be greatly exagerrated by
people - it is nothing (and I repeat to be clear - NOTHING) compared
to the performance benefit given by the opcode cache due to the absence
of the disk operations and compilation stage. You could probably
compose an artificial benchmark that would show some significant
slowdown, but I do not believe any real application would even notice.
source: http://n4.nabble.com/ZF-and-Autoloading-td640085i20.html#a640092