Implementation of some string.h function using SSE2 instructions

[Wojciech Muła]

PaweÂł Sikora wrote:
> $ cat sse2_strings.c
> #include "sse2_strings.h"
> #include <emmintrin.h>
> 
> static inline __m128i not( __m128i x )
> {
>         __m128i zero = { 0 };
>         __m128i ones = _mm_cmpeq_epi8( zero, zero );
>         return _mm_xor_si128( x, ones );
> }
> 
> int sse2_strcmp( sse2_byte_buffer s1, sse2_byte_buffer s2 )
> {
>         for ( int mask = 0; ; s1 += sizeof( __m128i ), s2 += sizeof( 
> __m128i ) )
>         {
>                 __m128i m1 = *( __m128i* )( s1 );
>                 __m128i m2 = *( __m128i* )( s2 );
>                 __m128i r1 = not( _mm_cmpeq_epi8( m1, m2 ) );
>                 __m128i zero = { 0 };
>                 __m128i r2 = _mm_cmpeq_epi8( m1, zero );
>                 __m128i r3 = _mm_cmpeq_epi8( m2, zero );
>                 __m128i r = _mm_or_si128( r1, _mm_or_si128( r2, r3 ) );
>                 mask = _mm_movemask_epi8( r );
>                 if ( mask )
>                 {
>                         unsigned index = __builtin_ffs( mask ) - 1;
>                         return ( s1[ index ] - s2[ index ] );
>                 }
>         }
> }

Yesterday I found a bit faster method, however tests I've done
show that this function could be **slower** than plain i686 code,
specially for short string, i.e. when inner loop is exected 1-3 times.
Some speedup, around 2x, appear for relative long strings (inner loop
executed 30+ times).  Thus I don't think that strcmp could be simply
replaced.

Also I've done some test with strlen: for short strings speedup is
around 6x, for longer strings up to 13x.


w.