1  /* { dg-do compile { target i?86-*-* x86_64-*-* } } */
       2  /* { dg-options "-msse4" } */
       3  #include <nmmintrin.h>
       4  __m128i load (char *);
       5  char *
       6  foo (const unsigned char *s1, const unsigned char *s2,
       7       int bmsk, __m128i frag2)
       8  {
       9    int len = 0;
      10    char *p1 = (char  *) s1;
      11    char *p2 = (char  *) s2;
      12    __m128i frag1, fruc1, fruc2, mask1, mask2;
      13    int cmp_c, cmp_s;
      14    if( !p2[0]) return (char *) s1;
      15    if( !p1[0] ) return NULL;
      16    if( p2[1]) frag2 = load (p2); 
      17    frag1 = load (p1);
      18    fruc1 = _mm_loadu_si128 ((__m128i *) s1);
      19    fruc2 = _mm_loadu_si128 ((__m128i *) s2);
      20    mask1 = _mm_cmpistrm(fruc1, frag2, 0x44);
      21    mask2 = _mm_cmpistrm(fruc2, frag1, 0x14);
      22    frag2 = _mm_blendv_epi8(frag2, mask1, mask2);
      23    frag1 = _mm_blendv_epi8(frag1, mask2, mask1);
      24    cmp_c = _mm_cmpistrc(frag2, frag1, 0x0c);
      25    cmp_s = _mm_cmpistrs(frag2, frag1, 0x0c);
      26    if( cmp_s  & cmp_c  )
      27      __asm("bsfl %[bmsk], %[len]" : [len] "=r" (len) : [bmsk] "r" (bmsk) );
      28    return p2 + len;
      29  }