strcpy vs strcpy_s

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

#pragma intrinsic(__rdtsc)

typedef unsigned char u8;
typedef unsigned long long u64;

int main() {
    constexpr u64 W = 128; // warmup iterations
    constexpr u64 M = 1024; // test iterations
    constexpr u64 N = 1024 * 64; // bytes to copy per string

    char* input = (char*)malloc(N);
    char* output = (char*)malloc(N);

    if (input == 0|| output == 0) {
        fprintf(stderr, "Failed to allocate testing memory\n");
        return 1;
    }

    // build string to be copied
    for (u64 i = 0; i < N - 1; i++) {
        u8 x = 0;
        do {
            x = (u8)(rand() & 0xFF);
        } while (x == 0);

        input[i] = x;
    }
    input[N - 1] = 0;

    // strcpy warmup
    for (u64 i = 0; i < W; i++) {
        strcpy(output, input);
    }

    // strcpy perf test
    u64 strcpy_total = 0;
    for (u64 i = 0; i < M; i++) {
        u64 start = __rdtsc();
        strcpy(output, input);
        strcpy_total += __rdtsc() - start;
    }

    // strcpy_s warmup
    for (u64 i = 0; i < W; i++) {
        strcpy_s(output, N, input);
    }

    // strcpy_s perf test
    u64 strcpy_s_total = 0;
    for (u64 i = 0; i < M; i++) {
        u64 start = __rdtsc();
        strcpy_s(output, N, input);
        strcpy_s_total += __rdtsc() - start;
    }

    printf("strcpy:   %llu %lf\n", strcpy_total, (double)strcpy_total / (double)M);
    printf("strcpy_s: %llu %lf\n", strcpy_s_total, (double)strcpy_s_total / (double)M);

    return 0;
}