vowel.c

#define _GNU_SOURCE /*for CPU affinity macros/routine*/
#include <stdio.h>
#include <string.h>
#include <stdlib.h>
#include <ctype.h>
#include <assert.h>
#include <sched.h>
/*
 * poor mans profiler. cpu tsc for x86 though would be skewed with cpufreq sched,etc.
 * but still better for approximations.
 */
#define rdtsc(x) do { asm __volatile__("rdtsc":"=A"(x)::); }while(0)
/*
 * flush instr. cacheline before rdtsc
 */
#define barrier() do { asm __volatile__("cpuid":::"memory"); }while(0) 

#ifdef STR

#define vowel_count(s)  vowel_count_slow(s)

int vowel_count_slow(const char *s)
{
    const char *vowel = "aeiou";
    int c = 0;
    while(*s)
            c += strchr(vowel, tolower(*s++)) ? 1 : 0;
    return c;
}

#else

#define vowel_count(s) vowel_count_fast(s)

/*
 * non-static to make it thread-safe, otherwise even faster
 * its faster to load complimentary set with 0's coz of a rep; stosb. zeroing off 256 elements below.
 * otherwise we can set the map[ 0 ... 255 ] = &unused and have the map increment unconditionally
 */
int vowel_count_fast(const char *s)
{
    int c = 0,t;
    int *map[256] = { [0 ... 255]=0,['a']=&c,['e']=&c,['i']=&c,['o']=&c,['u']=&c,
                      ['A']=&c,['E']=&c,['I']=&c,['O']=&c,['U']=&c};
    while((t=*s++)) map[t] ? ++*map[t] : 0;
    return c;
}

#endif

int main(int x,char **y) { 
    cpu_set_t set;
    long long A=0,B=0;
    x == 2 ? 0 : exit(0);
    /* lock proc. to cpu 0 for reasonably accurate rdtsc though cpufreq could skew it in which case
    * clock_gettime(CLOCK_MONOTONIC, &ts); should be okay 
    */
    CPU_ZERO(&set);
    CPU_SET(0, &set);
    assert(sched_setaffinity(0, sizeof set, &set) == 0);
    barrier();
    rdtsc(A);
    vowel_count(y[1]);
    rdtsc(B);
    printf("s = [%s], vowels = [%d], cpu ticks = [%lld]\n", y[1], vowel_count(y[1]), B - A);
    return 0;
}


/*
 * Local variables:
 * c-file-style: "linux"
 * c-basic-offset: 4 
 * tab-width: 4
 * compile-command: "gcc -Wall -g -o vowel vowel.c"
 * End:
*/