glibc wmemchr aarch64 optimization test

Makefile

BIN=wchar-tst
SRC=${BIN}.c

all: no o1 o2 o3

no: ${SRC}
        gcc -o no-${BIN} ${SRC}

o1: ${SRC}
        gcc -O1 -o o1-${BIN} ${SRC}

o2: ${SRC}
        gcc -O2 -o o2-${BIN} ${SRC}

o3: ${SRC}
        gcc -O3 -o o3-${BIN} ${SRC}

clean:
        rm {no,o1,o2,o3}-${BIN}

debug: ${SRC}
        gcc -g -O2 -o o2-${BIN} ${SRC}

wchar-tst.c

#include <wchar.h>
#include <stdio.h>
#include <stdlib.h>
#include <sys/time.h>
#include <string.h>




wchar_t *
my__wmemchr (const wchar_t *s, wchar_t c, size_t n)
{
  /* For performance reasons unfold the loop four times.  */
  while (n >= 4)
    {
      if (s[0] == c)
    return (wchar_t *) s;
      if (s[1] == c)
    return (wchar_t *) &s[1];
      if (s[2] == c)
    return (wchar_t *) &s[2];
      if (s[3] == c)
    return (wchar_t *) &s[3];
      s += 4;
      n -= 4;
    }

  if (n > 0)
    {
      if (*s == c)
    return (wchar_t *) s;
      ++s;
      --n;
    }
  if (n > 0)
    {
      if (*s == c)
    return (wchar_t *) s;
      ++s;
      --n;
    }
  if (n > 0)
    if (*s == c)
      return (wchar_t *) s;

  return NULL;
}

wchar_t *
mynew__wmemchr (const wchar_t *s, wchar_t c, size_t n)
{

	while( n>=4 ){
	
		static unsigned int set;

		__asm__ __volatile__
				("DUP v1.4s, %w[c]\n\t"
				"LD1 {v0.4s}, [%[s]]\n\t"

				//if a lane is all 1s, it matches the character
				"CMEQ v2.4s, v0.4s, v1.4s\n\t"
				
				//"check" if there were any matches
				"UMAXV S3, v2.4s\n\t"
        
				//move v3.4s[0] to set
				"UMOV %w[set], v3.4s[0]\n\t"				
				
				:[set]"=r"(set)
				:[s] "r" (s), [c] "r" (c)//input
				:);
		
		if (set){
			//found match in set
			if (s[0] == c)
			    return (wchar_t *) s;
		    if (s[1] == c)
			    return (wchar_t *) &s[1];
		    if (s[2] == c)
			    return (wchar_t *) &s[2];
		    if (s[3] == c)
			    return (wchar_t *) &s[3];
		}

      s += 4;
      n -= 4;
    }

  if (n > 0)
    {
      if (*s == c)
    return (wchar_t *) s;
      ++s;
      --n;
    }
  if (n > 0)
    {
      if (*s == c)
    return (wchar_t *) s;
      ++s;
      --n;
    }
  if (n > 0)
    if (*s == c)
      return (wchar_t *) s;

  return NULL;

}


#define CHRNUM 1000000000

struct timeval t1, t2;

int main(){

	//wchar_t * wcs = L"abcdefghijklmnopqrstuvwxyz1234567890ABCDEFGHIJKLMNOPQRSTUVWXYZ";

	wchar_t* wcs = (wchar_t*)malloc(CHRNUM * sizeof(wchar_t));

	//generate long string with all chrs as 'a' and last as 'b'
	unsigned long i;
	for(i = 0; i<CHRNUM; i++)
		wcs[i] = L'a';

	//wcs[6] = L'b';
	wcs[CHRNUM-1] = L'b';


	wprintf(L"\nwchar_t size: %d\n", sizeof(wchar_t));
	wprintf(L"%lc - %lc, %lc\n", wcs[0], wcs[CHRNUM-2], wcs[CHRNUM-1]);

	wchar_t wc = L'b';



	//find wc in wcs using the system dynamic library
  	gettimeofday(&t1, NULL);

  	wchar_t* wcp = wmemchr(wcs,wc,CHRNUM);

 	 gettimeofday(&t2, NULL);
  	////////////////////////
  	wprintf(L"\nbase: %p\tplus %lu\twchar: %lc\n", wcs, (unsigned long)(wcp-wcs), *wcp);

  	float elapsed_ms = (t2.tv_sec - t1.tv_sec) *1000 + (t2.tv_usec - t1.tv_usec) / 1000;
  	wprintf(L"so time: %f ms\n",elapsed_ms);


	//find wc in wcs the "old" way
	gettimeofday(&t1, NULL);
	
	wcp = my__wmemchr(wcs,wc,CHRNUM);

	gettimeofday(&t2, NULL);
	////////////////////////
	wprintf(L"\nbase: %p\tplus %lu\twchar: %lc\n", wcs, (unsigned long)(wcp-wcs), *wcp);

	elapsed_ms = (t2.tv_sec - t1.tv_sec) *1000 + (t2.tv_usec - t1.tv_usec) / 1000;
	wprintf(L"Old way time: %f ms\n",elapsed_ms);


	//find wc in wcs the "new" way
  	gettimeofday(&t1, NULL);

  	wcp = mynew__wmemchr(wcs,wc,CHRNUM);

  	gettimeofday(&t2, NULL);
  	////////////////////////
  	wprintf(L"\nbase: %p\tplus %lu\twchar: %lc\n", wcs, (unsigned long)(wcp-wcs), *wcp);
  	elapsed_ms = (t2.tv_sec - t1.tv_sec) *1000 + (t2.tv_usec - t1.tv_usec) / 1000;
  	wprintf(L"new way time: %f ms\n",elapsed_ms);
	
	free(wcs);
	return 0;

}