[RFC] possible strcmp speedup

[OndÅej BÃlka <neleai at seznam dot cz>]

Hello, 
strcmp is most of time decided by first 16 characters. 
This could be tested faster with following implementation.

I do not know how this is applicable for loop. A posible way could be
read a aligned and b unaligned and handle b page boundaries separately.

#define USE_SSE2
#include "vector.h"
int strcmp2(uchar *a,uchar *b){
  /*Test if we cross page. Assuming addresses are uniformly random 
   probability of that a|b is false positive is 1/16*/
  if (__builtin_expect((((size_t)a)|((size_t)b))%4096>=4096-sizeof(tp_vector),0))
    return strcmp(a,b);
  /* For pair of characters following cases can happen.
      xi==yi!=0    xi==yi==0   xi!=yi
       si=255       si=255      si=0
       mi!=0        mi=0        mi=0
       ri=0         ri=255      ri=255
  */
  tp_vector x=LOAD_UNALIGNED(a);
  tp_vector y=LOAD_UNALIGNED(b);
  tp_vector s=TEST_EQ(x,y);
  tp_vector m=AND(s,x);
  tp_vector r=TEST_EQ(m,BROADCAST(0));
  tp_mask   mask =get_mask(r);
  int       first=__builtin_ctz(mask);
  if(mask){
#ifdef AS_STRCASECMP
    /*If characters differ they are probably caselessly diferent.*/
    uchar ac=tolower(a[first]),bc=tolower(b[first]);
    if(!ac)    return ac-bc;
    if(ac!=bc) return ac-bc;
    return strcmp(a+first,b+first);
#else
    return a[first]-b[first];
#endif
  }
  return strcmp(a+sizeof(tp_vector),b+sizeof(tp_vector));
}