With CURL_DISABLE_PROXY or CURL_DISABLE_HTTP, GCC complained about a
missing prototype for Curl_connect_free.

... also updated the CURLOPT_PREQUOTE.3 man page to mention the correct
protocol support.

Closes #1514

... as it really cannot have reached this far with config being NULL,
thus this is unnecesary and misleading.

Bug: https://news.ycombinator.com/item?id=14577585 and
https://daniel.haxx.se/blog/2017/06/17/curl-doesnt-spew-binary-anymore/comment-page-1/#comment-18356

Forwarded-to-us-by: Jakub Wilk

... unless "--output -" is used. Binary detection is done by simply
checking for a binary zero in early data.

Added test 1425 1426 to verify.

Closes #1512

The configure-based build also has this in addition to -Wall.

Closes https://github.com/curl/curl/pull/1578

It usually warns when using commas instead of semicolons or other
operators by accident.

Closes https://github.com/curl/curl/pull/1578

It usually warns when forgetting to declare TU-local variables static.

Closes https://github.com/curl/curl/pull/1578

Enable -Wdouble-promotion for both GCC and clang. It warns on implicit
promotion from float to double.

Closes https://github.com/curl/curl/pull/1578

Previously, that warning was only implicitly active in C90 mode.
Enable it unconditionally as already done for GCC.

Closes https://github.com/curl/curl/pull/1578

Regression since 5113ad04.

... and remove 'flaky' from test 1061 again

Closes #1579

... the previous code would reset the header length wrongly (since
5113ad04). This makes test 1060 reliable again.

Also: make sws send even smaller chunks of data to increase the
likeliness of this happening.

Bug: https://github.com/curl/curl/issues/1254

Closes #1546

Both these tests run the same underlying test code: libntlmconnect.c -
this test code made some assumptions about socket ordering when it used
curl_easy_fdset() and when we changed timing or got accidental changes
in libcurl the tests would fail.

The tests verify that the different transfers keep using the same
connections, which I now instead made sure by adding the number of bytes
each transfer gets and then verifies that they always get the same
amount as when these tests worked.

Closes #1576

Closes #1569

Fails intermittently on travis builds since a few days. Likely due to
5113ad04.

- Move the logic to detect a Windows drive letter prefix
  (eg c: in c:foo) into a function-like macro.

Closes https://github.com/curl/curl/pull/1571

- No longer allow partial downloads of certdata.

Prior to this change partial downloads were (erroneously?) allowed since
only the server code was checked to be 200.

Bug: https://github.com/curl/curl/pull/1577
Reported-by: Matteo B.

Reported-by: Dan Fandrich

Fixes #1476

... all other non-HTTP protocol schemes are now defaulting to "tunnel
trough" mode if a HTTP proxy is specified. In reality there are no HTTP
proxies out there that allow those other schemes.

Assisted-by: Ray Satiro, Michael Kaufmann

Closes #1505

... since commit 73a2fcea

... to make it really apparent if there's any user using this on purpose.

Suggested-by: Jay Satiro

Closes #1542

When this define was set, libcurl would check the environment variable
named CURL_CA_BUNDLE at run-time and use that CA cert bundle. This
feature was only defined by the watcom and m32 makefiles and caused
inconsistent behaviours among libcurls built on different platforms.

The curl tool does already feature its own similar logic and the library
does not really need it, and it isn't documented libcurl behavior. So
this change removes it.

Ref: #1538

So many headers can be provided as @filename.

Suggested-by: Timothe Litt

Closes #1486

Mentioned as a problem since 2007 (8f87c15b) and of course it
existed even before that.

Closes #1547

Bug #1556
Reported-by: Paul Harris
Closes #1559

CURLSYS_PULL_WS2TCPIP_H got renamed to CURL_PULL_WS2TCPIP_H in commit
73a2fcea.

This gives us accurate precision and it allows us to avoid storing "no
time" for systems with too low timer resolution as we then bump the time
up to 1 microsecond. Should fix test 573 on windows.

Remove the now unused curlx_tvdiff_secs() function.

Maintains the external getinfo() API with using doubles.

Fixes #1531

The maketgz script now makes sure the generated hugehelp.c file in the
tarball is newer than the generated curl.1 man page, so that it doesn't
have to get unnecessarily rebuilt first thing in a typical build. It
thus also removes the need for perl to build off a plain release
tarball.

Fixes #1565

Rely entirely on curl/system.h now.

Introduced in Aug 2008 with commit 14240e9e. Now gone.

Fixes #1456