Hi,

At Sun, 2 Mar 2003 03:05:36 +0900,
Mauricio FernáÏdez wrote:
> What about the following?
> I get a speed increase of 3% (compared to the non-inlined version) 
> with the second test of the GCLS.
> I eliminate 2 instructions per iteration (on x86) out of 10 in the loop.
> It's OK to use compiler extensions if we detect them in configure, isn't
> it?

Instead, use -funroll-loops option for gcc.

#ifdef __GNUC__
> +#define HAVE_COMPUTED_GOTO 1 /* this should be detected by configure */
#endif

> +#elif HAVE_COMPUTED_GOTO

And you don't need compiler extensions at all in this case:

    switch (len) {
      case 31: key = key*65599 + *p++;
      case 30: key = key*65599 + *p++;
      case 29: key = key*65599 + *p++;
        /* snip */
      case 1: key = key*65599 + *p++;
      case 0: break;
      default:
        key = key*65599 + *p++;
        while (len--) {
            key = key*65599 + *p;
            p++;
        }
    }

Or famous tricky code:

    if (len > 0) {
        switch (len & 31) {
            do {
              case 0: key = key*65599 + *p++;
              case 31: key = key*65599 + *p++;
              case 30: key = key*65599 + *p++;
              case 29: key = key*65599 + *p++;
              case 28: key = key*65599 + *p++;
                /* snip */
              case 1: key = key*65599 + *p++;
            } while ((len -= 32) > 0);
        }
    }

-- 
Nobu Nakada