>> The difference is that while Emacs is being built, especially 
>> bootstrapped, calls to Lisp code that isn't preloaded by autoload.el 
>> could fail.  So such calls need to be careful not to call a symbol if 
>> the call could fail, or at least use internal_condition_case* functions 
>> to protect themselves against failure.
>
> Basically, you should use safe_call and its ilk.
>

Thank you for that clarification, done.

>
> Even then, some people will frown on 20% slowdown of the startup.
>

It is not a slowdown of the startup, it is a slowdown of setq, which 
becomes visible for the user when say 100 million setq's are executed.

Anyway, given your remark, I improved the patch.  With this optimized 
version, there is no noticeable difference for variables without a plist 
(~40 nanoseconds for each call to setq, with and without the patch), and a 
small difference for variables with a plist (~40 nanoseconds for each call 
to setq without the patch, ~42 nanoseconds with the patch).