#include <stdio.h>

struct regs {
    unsigned char c_regs[8];
    unsigned s_regs[128];
};

unsigned add_32 (unsigned a, unsigned b) {
    return a + b;
}

unsigned sub_32 (unsigned a, unsigned b) {
    return a - b;
}

unsigned ld_32 (unsigned * const b) {
    return *b;
}

void st_32 (unsigned a, unsigned * const b) {
    *b = a;
}

unsigned char cmp_lt_32 (unsigned a, unsigned b) {
    return a < b;
}

void fib (unsigned * const r, unsigned * const n) {

    struct regs gpu_reg_p;

    gpu_reg_p.s_regs[1] = ld_32(n);
    gpu_reg_p.c_regs[1] = cmp_lt_32(gpu_reg_p.s_regs[1], 3);
    if (gpu_reg_p.c_regs[1]) {
        goto ret;
    }

    unsigned p;
    {
        unsigned nm2;
        unsigned res;
        gpu_reg_p.s_regs[2] = sub_32(gpu_reg_p.s_regs[1], 2);
        st_32(gpu_reg_p.s_regs[2], &nm2);
        fib(&res, &nm2);
        gpu_reg_p.s_regs[2] = ld_32(&res);
    }

    st_32(gpu_reg_p.s_regs[2], &p);

    {
        unsigned nm2;
        unsigned res;
        gpu_reg_p.s_regs[2] = sub_32(gpu_reg_p.s_regs[1], 1);
        st_32(gpu_reg_p.s_regs[2], &nm2);
        fib(&res, &nm2);
        gpu_reg_p.s_regs[2] = ld_32(&res);
    }

    gpu_reg_p.s_regs[3] = ld_32(&p);
    gpu_reg_p.s_regs[2] = add_32(gpu_reg_p.s_regs[2], gpu_reg_p.s_regs[3]);
    st_32(gpu_reg_p.s_regs[2], r);
    return;

ret:
    st_32(1, r);
    return;
}


void fibKernel (unsigned * const r_ptr, unsigned * const n_ptr) {
    struct regs gpu_reg_p;
    {
        unsigned r;
        unsigned n;

        gpu_reg_p.s_regs[0] = *n_ptr;
        st_32(gpu_reg_p.s_regs[0], &n);
        fib(&r, &n);
        gpu_reg_p.s_regs[0] = ld_32(&r);
        //s1 = *r_ptr;
        //*((unsigned *)s1) = s0;
        *r_ptr = gpu_reg_p.s_regs[0];
    }
    return;
}


int main (void) {
    unsigned r;
    unsigned i;


    for (i = 1; i < 20; i++) {
        fibKernel(&r, &i);
        printf("%u ", r);
    }

    return 0;
}