On Sat, Oct 27, 2001 at 02:56:07PM +0400, Alexander V. Lukyanov wrote:
> I've rewrote crlf_to_lf so it is more clear and easier to understand.
>
> int crlf_to_lf(char *buf, int s)
> {
> char *store=mem_crlf(buf,s);
> if(!store)
> return s;
>
> int retsize=s-1;
> s-=store+1-buf;
> buf=store+1;
>
> while(s>1)
> {
> char *crlf=mem_crlf(buf,s);
> if(!crlf)
> break;
>
> memmove(store, buf, crlf-buf);
> store+=crlf-buf;
> retsize--;
> s-=crlf+1-buf;
> buf=crlf+1;
> }
> memmove(store,buf,s);
> return retsize;
> }
Both of ours are easier to understand for each of us, only because we
wrote them. I can't quickly make sense of that, either. :)
Originally, I used memmove/memchr for speed--they're well-optimized. In
retrospect, that's also iterating over the input buffer twice (though in
a cache-friendly way.)
Here's one that anyone can understand:
int crlf_to_lf(char *buf, int s)
{
char *store = buf;
int retsiz = s;
while(s) {
if(s-- > 1 && buf[0] == '\r' && buf[1] == '\n') {
buf++;
retsiz--;
continue;
}
*store++ = *buf++;
}
return retsiz;
}
It's about twice as fast when the buffer is a lot of short lines--which is
the typical case of NLIST. The memchr/memmove approach is about 30% faster
for longer strings (ie. 80 characters/line, which would be the case if
it was used for LIST); apparently the word-reading optimizations used
by memchr and memmove offset the penalty of iterating twice.
At 64 bytes/line for 10 megs of data, on my system, it's still 300ms.
The speed difference doesn't really matter.
--
Glenn Maynard
