This should not have any impact on the "normal" deliveries, as these
will as before be done through a single writev operation.

For deliveries involving chunked encoding (e.g. ESI or gunzip), this
should prevent an unnecessary delay where the Nagle algorithm kicks in
on the end-chunk write. This should give higher throughput and better
utilization of a single connection, especially for clients not doing
pipelining.

buy 512TB ram to run Varnish on ppc64.

Separate out the text describing the storage backends so it can be resued in both the man page and users guide

Some of them are rather pedantic:  "NULL" vs "(char*)0" -- Really ?

libtoolize was overwriting m4/libtool.m4, which aclocal.m4 depended on
and needed aclocal to update. This caused the autotools files to be
updated, triggering a second configure run on first build.

The JIT compiler is broken on some versions of PCRE, at least on i386,
so disable it by default.  It can be enabled using --enable-pcre-jit
to configure.

Fixes #1191

Move the predictive vary from the req->ws to busyobj->ws.

predictive vary could do.

XXX: doc-change needed.

there.  These bits are legitimate feature-ettes which might make sense
in a production environment.

XXX: need doc-change

XXX: doc-update needed

XXX: doc-update needed

XXX: doc-update needed

XXX: doc-update needed

XXX: doc-update needed