#include <stdio.h>
#include <stdint.h>
#include <stdlib.h>
#include <string.h>

void
memset1(void *d, uint32_t p, size_t n)
{
        size_t i;
        uint32_t *ptr = (uint32_t *)d;

        for (i=0; i<n; i++) {
                ptr[i] = p;
        }
}

void
memset2(void *d, uint32_t p, size_t n)
{
        memset(d, p, n*sizeof(uint32_t));
}

void
memset3(void *d, uint32_t p, size_t n)
{
        asm volatile (
                "cld\n\t"
                "rep stosl"
                : "+D" (d), "+c" (n)
                : "a" (p)
                : "memory");
}

void
memset4(void *start, uint32_t colour, size_t count)
{

    asm volatile(
        "cld\n\t"
        "rep\n\t"
        "stosl"
        : /* no output registers */
        : "c" (count), "a" (colour), "D" (start));
}

int
main()
{
        uint32_t *data;
        size_t i, n = 1024*1024;
        uint32_t fill = 0xdeadbeaf;

        data = malloc(n*sizeof(uint32_t));

        for (i=0; i<1000; i++) {
                /*memset1(data, fill, n);*/
                /*memset2(data, fill, n);*/
                /*memset3(data, fill, n);*/
                memset4(data, fill, n);
        }

        for (i=0; i<n; i++) {
                if (data[i] != fill) {
                        printf("error! %x at %u\n", data[i], i);
                        break;
                }
        }

        free(data);
        return 0;
}