On 11/3/18 9:25 PM, Norihiro Tanaka wrote:

>   $ seq -f '%040g' 0 9999 | sed '1s/$/\\(0\\)\\1/' >pat

Thanks for the test case and sorry about the delay. And thanks for spotting the
speedup opportunity. I found a few problems with the proposed patch, though:

> +        if (keys[1] == '\\')
> +          keys++;

This mishandles the case where the input byte sequence contains '\', '\', '1'
where the first '\' is the last byte of a multibyte character. Such a byte
sequence can contain backreferences but the function will say it doesn't.

> +      if (backref && prev < p)
> +        {
> +          len = p - prev;
> +          buf = xrealloc (buf, (buflen + len) * sizeof *buf);
> +          memcpy (buf + buflen, p, len);
> +          buflen += len;
> +        }

This seems to have three problems. First, the memcpy copies from P, but it
should copy from PREV. Second, this code assigns to LEN, which breaks a later
use of LEN. Third, if there are many patterns with backreferences, repeated use
of realloc will have O(N**2) behavior.

> +      for (size_t i = 0; i < dc->pcount; i++)
> +        dc->patterns[i + 1] = dc->patterns[i];

This copies dc->patterns[0] to the later values in that array, when a memmove
was intended.

So, after installing the patch, I immediately installed the attached patch,
which should address the abovementioned issues.

Thanks again. You did the hard work - I merely proofread it.