2) in the curl_dns_cache_entry structure, timestamp should be
a time_t instead of an integer (although I doubt it matters).

The default cache timeout for this is 60 seconds, which is arbitrary and
completely subject to change :)

Set-Cookie: line was ignored if they didn't end with a trailing
semicolon. This is indeed wrong syntax, but there are high-profile web sites
out there sending cookies like that so we must make a best-effort to parse
them.

platforms, unrelated to what they might prefer by default

    - Use a global dns cache (via setting the tentatively named,
    CURLOPT_DNS_USE_GLOBAL_CACHE option to true)
    - Use a per-handle dns cache, by default
    - Use a pooled dns cache when in the "multi" interface

expect (nor read) any response-body

cookie list, even though they're not supposed to do that...

I also replaced all printf() calls with calls to Curl_failf()

platforms that have such a function.
This affects multi-thread running libcurls on IPv4 systems that have VERBOSE
switched on. The previous version was risking that another thread overwrote
the data before it was read out in this thread. There could possibly also
be a slight risk that the data isn't zero terminated for a short while and
thus could cause the thread to crash...

header directory but the function *is* present!

to become properly 8-byte aligned on 64-bit archs. Philip Gladstone reported.

headers.

we return a set of headers not more. This enables FTP operations that don't
transfer any data, only perform FTP commands.

set.upload to FALSE or else we might still get an upload if the previous
operation was an upload!

is not dealt with when we find an end-of-response line, as there might be
important stuff even after the correct line. So on subsequent invokes, the
cached data must be used!