import std.stdio;
import std.math;
import std.perf;

struct Old {
    float a;
    float b;
}

struct New {
    float a;
    float b;
    float c;
    float d;
}

class Foo {

    void run(Old a)               { exp(a.a); }
    void run(Old a, Old b)        { exp(a.a); }
    void run(Old a, Old b, Old c) {
        exp(a.a);
//        exp(b.a);
  //      exp(c.a);
    }

    void run(New a)               { exp(a.a); }
    void run(New a, New b)        { exp(a.a); }
    void run(New a, New b, New c) {
        exp(a.a);
//        exp(b.a);
  //      exp(c.a);
    }
}
    void run(Old a)               { exp(a.a); }
    void run(Old a, Old b)        { exp(a.a); }
    void run(Old a, Old b, Old c) { exp(a.a); }

    void run(New a)               { exp(a.a); }
    void run(New a, New b)        { exp(a.a); }
    void run(New a, New b, New c) { exp(a.a); }


int main(char[][] args) {
    scope foo = new Foo();

    Old a1,a2,a3,a4;
    New b1,b2,b3;
    a1.a = 5;  a3 = a1;  a2 = a1;
    b1.a = 5; b2 = b1; b3 = b1;

    a1.a = 10;
    a1.b = 12;
    b1.a = 10;
    b1.d = 12;

    auto c1 = new New;
    auto c2 = new New;
    *c1 = b1;
/*
    asm {
        movups XMM0, b1 ; //movdqa , movups, movaps
        movups b2, XMM0;
    }

    writefln(b1.a," ",b2.a);
    writefln(b1.d," ",b2.d);
return 0;
*/

    auto hpc = new HighPerformanceCounter();

    uint count = 100_000;

    long loop_time = long.max;
    long  old_time = long.max;
    long  new_time = long.max;

    for(uint t = 0; t < 100; t++) {
        hpc.start;
        for(uint i = 0; i < count; i++) {
        }
        hpc.stop;
        if( hpc.microseconds < loop_time )
            loop_time = hpc.microseconds;

        hpc.start;
        for(uint i = 0; i < count; i++) {
//*
            a2 = a1;
            a2 = a1;
            a2 = a1;
            a2 = a1;
            a2 = a1;
            a2 = a1;
            a2 = a1;
            a2 = a1;
            a2 = a1;
            a2 = a1;
/*/
            asm { movq XMM0, a1 ; movq a2, XMM0; }
            asm { movq XMM0, a1 ; movq a2, XMM0; }
            asm { movq XMM0, a1 ; movq a2, XMM0; }
            asm { movq XMM0, a1 ; movq a2, XMM0; }
            asm { movq XMM0, a1 ; movq a2, XMM0; }
            asm { movq XMM0, a1 ; movq a2, XMM0; }
            asm { movq XMM0, a1 ; movq a2, XMM0; }
            asm { movq XMM0, a1 ; movq a2, XMM0; }
            asm { movq XMM0, a1 ; movq a2, XMM0; }
            asm { movq XMM0, a1 ; movq a2, XMM0; }
            asm { movq XMM0, a1 ; movq a2, XMM0; }
//*/
//            foo.run(a1);//,a2,a3);
        }
        hpc.stop;
        if( hpc.microseconds < old_time )
            old_time = hpc.microseconds;


        hpc.start;
        for(uint i = 0; i < count; i++) {

/*
            asm { movq XMM0, b1 ; movq b2, XMM0; }
            asm { movq XMM0, b1 ; movq b2, XMM0; }
            asm { movq XMM0, b1 ; movq b2, XMM0; }
            asm { movq XMM0, b1 ; movq b2, XMM0; }
            asm { movq XMM0, b1 ; movq b2, XMM0; }
            asm { movq XMM0, b1 ; movq b2, XMM0; }
            asm { movq XMM0, b1 ; movq b2, XMM0; }
            asm { movq XMM0, b1 ; movq b2, XMM0; }
            asm { movq XMM0, b1 ; movq b2, XMM0; }
            asm { movq XMM0, b1 ; movq b2, XMM0; }
            asm { movq XMM0, b1 ; movq b2, XMM0; }
//*/
            asm { movq XMM0, b1 ; movq XMM1, b1 ; movq b2, XMM0;  movq b2, XMM1;}
            asm { movq XMM0, b1 ; movq XMM1, b1 ; movq b2, XMM0;  movq b2, XMM1;}
            asm { movq XMM0, b1 ; movq XMM1, b1 ; movq b2, XMM0;  movq b2, XMM1;}
            asm { movq XMM0, b1 ; movq XMM1, b1 ; movq b2, XMM0;  movq b2, XMM1;}
            asm { movq XMM0, b1 ; movq XMM1, b1 ; movq b2, XMM0;  movq b2, XMM1;}
            asm { movq XMM0, b1 ; movq XMM1, b1 ; movq b2, XMM0;  movq b2, XMM1;}
            asm { movq XMM0, b1 ; movq XMM1, b1 ; movq b2, XMM0;  movq b2, XMM1;}
            asm { movq XMM0, b1 ; movq XMM1, b1 ; movq b2, XMM0;  movq b2, XMM1;}
            asm { movq XMM0, b1 ; movq XMM1, b1 ; movq b2, XMM0;  movq b2, XMM1;}
            asm { movq XMM0, b1 ; movq XMM1, b1 ; movq b2, XMM0;  movq b2, XMM1;}
            asm { movq XMM0, b1 ; movq XMM1, b1 ; movq b2, XMM0;  movq b2, XMM1;}
//*/
/*
            asm { movdqu XMM0, b1 ; movdqu b2, XMM0; }
            asm { movdqu XMM0, b1 ; movdqu b2, XMM0; }
            asm { movdqu XMM0, b1 ; movdqu b2, XMM0; }
            asm { movdqu XMM0, b1 ; movdqu b2, XMM0; }
            asm { movdqu XMM0, b1 ; movdqu b2, XMM0; }
            asm { movdqu XMM0, b1 ; movdqu b2, XMM0; }
            asm { movdqu XMM0, b1 ; movdqu b2, XMM0; }
            asm { movdqu XMM0, b1 ; movdqu b2, XMM0; }
            asm { movdqu XMM0, b1 ; movdqu b2, XMM0; }
            asm { movdqu XMM0, b1 ; movdqu b2, XMM0; }
//* /
            asm { movdqa XMM0, b1 ; movdqa b2, XMM0; }
            asm { movdqa XMM0, b1 ; movdqa b2, XMM0; }
            asm { movdqa XMM0, b1 ; movdqa b2, XMM0; }
            asm { movdqa XMM0, b1 ; movdqa b2, XMM0; }
            asm { movdqa XMM0, b1 ; movdqa b2, XMM0; }
            asm { movdqa XMM0, b1 ; movdqa b2, XMM0; }
            asm { movdqa XMM0, b1 ; movdqa b2, XMM0; }
            asm { movdqa XMM0, b1 ; movdqa b2, XMM0; }
            asm { movdqa XMM0, b1 ; movdqa b2, XMM0; }
            asm { movdqa XMM0, b1 ; movdqa b2, XMM0; }
//*/
//            foo.run(b1);//,b2,b3);
//*/
        }
        hpc.stop;
        if( hpc.microseconds < new_time )
            new_time = hpc.microseconds;
    }
    writefln(a1.a," ",a2.a," ",b1.a," ",b2.a);
    writefln(a1.b," ",a2.b," ",b1.d," ",b2.d);

    writefln( loop_time );
    writefln( old_time - loop_time );
    writefln( new_time - loop_time," ",(new_time - loop_time) * 100.0 /(old_time - loop_time) - 100 );


    return 0;
}