Re: [PATCH 2/4] Improve generic strspn performance

[Wilco Dijkstra <Wilco dot Dijkstra at arm dot com>]

Adhemerval Zanella wrote:

> +  if (accept[0] == '\0')
> +    return 0;
> +  if (accept[1] == '\0')
> +    { 

GCC doesn't get the static branch prediction correct for the 2nd if,
so it would be useful to add __glibc_unlikely given that single-character
accepts are rare.

> +  s = (unsigned char *) ((size_t)(s) & ~3);
> +  unsigned int c0, c1, c2, c3; 
> +  do {
> +      s += 4;
> +      c0 = p[s[0]];
> +      c1 = p[s[1]];
> +      c2 = p[s[2]];
> +      c3 = p[s[3]];
> +  } while ((c0 && c1 && c2 && c3) == 1);

That should use '&' rather than '&&' and '!= 0' similar to how I did it in strcspn.
This will use 3 AND(S) instructions and a single branch.

> +
> +  size_t count = s - (unsigned char *) str;
> +  return (c0 && c1) == 0 ? count - !c0 + 1 : count - !c2 + 3;

Again, c0 & c1 is better and allows CSE with the while expression above.
Also -!c0 +1 is equivalent to c0, -!c2 + 3 is equivalent to c2 + 2 - this is simpler
and faster.

Otherwise it looks good, and thanks for doing this one too!

Cheers,
Wilco