🚀 go-pugleaf

Andrue Cope  wrote:

>Malcolm Smith
>>  I have a class that improved by 13% in terms of performance
>> when I removed the __fastcall.

Yeah, classes can be tricky.  Remember that member functions have a
hidden parameter *this, so fastcalls should only be on 1 or 2
parameter functions.

>Our experience is the reverse. We use __fastcall for any frequently called 
>but simple function.

Yep.  Functions of 3 or less parameters. I you have 4, then you don't
gain much. I never use fastcall for 4 or more parameters.

Also, functions that mostly only use the 3 parameters, and maybe one
or two locals.

Wrapper functions can benefit.
char* _fastcall Wrapper(char*dest,char*src)
{	return lstrcpyn(dest,src,7);
}
And especially wrapper wrappers:
char* _fastcall Wrapper2(char*dest)
{	return Wrapper(dest,"Default");
}

> It may be that __fastcall makes things worse for 
>complex functions due to the contraints it imposes on the optimiser by 
>reserving registers.

If a parameter is used late in a (complex) function, or is used
repeatedly to initalize a local, it would probably benefit from being
on the stack, thus not a fastcall candidate.
If you have several locals, it's also probably not a fastcall
candidate.

Some of these constraints might ease if you are compiling for the
Itanium with it's many registers.