I wonder who has written the vmod.rst documentation clearly pointing out
that this needs to be done... :|

Ref Varnish-Cache ff18a3c5bf7cf02acaa5bc5672cc90c8bb805336

When we have a partial match and no buffer, we copy all of ptr/len and
thus must not touch the startoffset.

We could also copy from startoffset, but that could break lookbehinds,
so don't.

Patch best viewed with -b

We can now iterate over bodies, calling a subroutine for each
match.

Previously, we would only look for the first ever match.
To support multiple matches, we now save the start offset
relative to the ptr argument for the next invocation.

... or the highest match

The main reason is an upcoming improvement which needs workspace.

But also using workspace here probably does not come with too relevant
an advantage, because regular expression matching and copying are
already expensive, and the need to copy should be the exception.

This is in preparation of an upcoming refactor to heap space.

Add filters: regex objects to perform substitutions on bodies

See merge request !3

If the optional ``asfilter`` parameter to ``re.regex()`` is true, the
vmod registers itself as a Varnish Fetch Processor (VFP) for use in
`beresp.filters` and as a Varnish Delivery Processor (VDP) for use in
`resp.filters`. In this setup, the `xregex.substitute_match()` and
`xregex.substitute_all()` methods can be used to define replacements
for matches on the body.

Example:

	    sub vcl_init {
		new reiher = re.regex("r(ei)h(er)", asfilter = true);
	    }
	    sub vcl_deliver {
		unset req.http.Accept-Encoding;
		set resp.filters += " reiher";
		reiher.substitute_match(1, "czapla");
		reiher.substitute_match(2, "\1\2");
		reiher.substitute_match(0, "heron");
	    }

Ref varnish-cache 478e310a7d2c168e1c361d6f34291177cbc2a889

Code taken from varnish-cache