Build system rewrite

[0x5c]

This has dramatic effects on the build system complexity, and in the process exposing lots of dead weight.

Split header files (.h1 .h2) are merged, and header files that only contained a single define have been fully replaced with -D. The makefile is no longer generating scripts, and their files that hardcoded cc/ld/ar calls are gone.

The build system is now fully compatible with standard variables like CC, PREFIX, DESTDIR, and such. The compiler is autodetected when unspecified, build flags are checked for compatibility, and installation directories can now be specified with configure flags.

The configure script was heavily inspired by XBPS's.

The "compile" script is gone along with the original top-level Makefile, both replaced with straightforward and more conventional makefiles. The "man" and "completions" subdirs are now integrated into the build system and use configure-time variables for the paths.

No sed calls or patches should be required to set any paths or toolchain settings.

Usage of .gitattributes and $Id$ strings is now replaced with a VERSION define in the configure script, which optionally uses git to append the 8-digit commit hash to the version string. This also makes -V more meaningful.

Multiple configure checks were removed, some were unused, while others required to be run on the host at build time and were not compatible with cross. This means we are now assuming C99 compiler and availability of poll(); both seem reasonable.

After ~17 years of being deprecated, excluded from builds, and de-documented, svwait{up,down} and runsv{ctrl,stat} were removed from the codebase.

Closes #11
Closes #12

Things left to do

• make check is broken has to be fixed. Long term there's probably room to improve/modernise the tests.

Going further

• Should the docs be integrated in the build system now or as a later time (or not at all)?

• The selection of default CFLAGS, warning flags, and LDFLAGS is mostly taken from XBPS's configure script. Are they adequate here, and are there any to add?

• Lots of checks/etc seem to assume old and exotic platforms, but it's unclear to me which checks only apply to platforms excluded by the new requirements (C99, presence of poll(), $CC as linker) or that we can safely stop supporting.

  While trying to identify those, I found that there's too much variance with even the mainstream BSDs for that to be done in this PR. Anyways, the important part is that they're no longer tangled with the rest of the build system, and should be much easier to tackle individually later.

• The configure check for the socket library checks for linker flags to add for socket.h to link properly. The only relevant results when searching online for libxnet and libnsl are about Solaris and a possibly unrelated libnsl.so missing on RHEL; does it make sense to keep this? A build on void does not require these.

  Decided to drop the check, details in commit

• The old build system would check for the operating system and change the ar behaviour based on that, does this make sense nowadays? It seems the platforms it special-cases are all business unixes or obsolete. (It would run ranlib after ar on all platforms but some, currently replaced with adding the s modifier on ar)

  Decided to drop the special-casing, details in commit

• PREFIX/EPREFIX usage was made to match XBPS, however it seems that XBPS does not properly use them according to the configure help text, is this correct? (EPREFIX should be used for "architecture-dependent" files, but is used or documentation but not executables)

  Turned into an xbps "bug?" report

---

Most of this was a learning experience, so I may have missed obvious things; feedback welcome.

About that complexity...

The makefile in src/ before

... and after.

.dots here.
I recommend opening original.dot in an interactive dot viewer to better see the node relationships.

[0x5c]

Long term there's probably room to improve/modernise the tests.

Hum yeah, the mechanism is clunky and most of these don't seem very useful at all

[0x5c]

-std=c99 was added to the CFLAGS to get uint64_t, however this now means that many files generate -Wimplicit-function-declaration, only solved by the definition of _DEFAULT_SOURCE or _POSIX_SOURCE in those files.

Are the feature macros an ok solution? Should they be defined everywhere? Is there a flag that does it better? y understanding is that -std=c99 does deactivate some defaults, but it's not obvious to me if there's a flag to reactivate them.

[nekopsykose]

Should they be defined everywhere?

unless there's something you're gaining by selectively defining them (you would know; generally it's to selectively expose a different standard of a function in some places intentionally), yes, just -D global cflags.