When you use CGI.pm 700 lines of code get 'compiled' plus 2400 lines of code also get put in the subs hash. While the code in the subs hash is not compiled a fairly large hash must be generated which takes both memory and time. CGI::Simple uses SelfLoader to avoid compiling methods that are rarely used. You do this by placing these methods below a __DATA__ token. At compile time compilation stops at the __DATA__ token. As a result when you use CGI::Simple only 300 lines of code actually get compiled.
With SelfLoader if you call one of the methods below the data token then (from the docs):
The SelfLoader will read from the CGI::Simple::DATA filehandle to load in the data after __DATA__, and load in any subroutine when it is called. The costs are the one-time parsing of the data after __DATA__, and a load delay for the _first_ call of any autoloaded function. The benefits (hopefully) are a speeded up compilation phase, with no need to load functions which are never used.
One of the neat things about SelfLoader is that if you know that you will regularly use methods x, y, and z you can easily tune the module by placing these above the data token. As a result they will be available without using SelfLoader and the runtime overhead of using SelfLoader need never be paid.
One of the not so neat things is that you have to load SelfLoader to use it, so there is a compile time penalty that you must pay. Fortunately SelfLoader.pm is only 100 lines of code. I was sorely tempted to 'roll my own' as you can do this with much less code provided you do not have to 'cover all the bases' as the module does. This, however, went against the perl concept of using modular code when available. Similarly CGI::Simple uses IO::File->new_tmpfile() to generate a self destructing temp file for file uploads leaving all the details to this module. IO::File is called via a require so you only load it if and when you need it.