Re: [Devel] suggested improvement to FT_Add64

[Nathan Hurst]

Follow up:

I just tried (for kicks :):

void njh(long long* x,
         long long* y,  
         long long* z) {
  *z = *x + *y;
}

And it won handsomely with 
        lwz 9,0(3)
        lwz 10,4(3)
        lwz 11,0(4)
        lwz 12,4(4)
        addc 8,10,12
        adde 7,9,11
        stw 7,0(5)
        stw 8,4(5)
        blr

And I thought "this is mighty nice, I wonder if not using pass by pointer 
would be any faster", so I tried:

void njh2(const long long x,
         const long long y,  
         long long& z) {
  z = x + y;
}

and it produced 5 instructions.

        addc 10,4,6
        adde 9,3,5
        stw 9,0(7)
        stw 10,4(7)
        blr

Finally, using long long add in a stream of code produces only 2 instructions 
(the two adds) and nicely moves them in between two stores of unrelated data 
to remove the bubbles...

        stw 17,28(30)
        addc 10,13,15
        adde 9,12,14
        stw 18,32(30)

njh