CMAKE_<LANG>_COMPILER_LAUNCHER was added around cmake 3.4 and, usefully,
can be enabled/disabled via cmake invocation flags (unlike the older
RULE_LAUNCH_COMPILE property).
Refactors many things in cmake to improve and simplify:
- don't use variable indirection for target names; target names are
*already* a variable of sorts. (e.g. ${UTIL_LIB} is now just
lokinet-util). cmake/basic_definitions.cmake is now gone.
- fix LTO enabling to use the standard cmake (3.9+) LTO mechanism rather
than shoving a bunch of flag hacks through link_libraries and
add_compile_options. This also now enables LTO when building a shared
library (because previously the -flto hacks were only turned on in the
static code for some reason).
- build liblokinet as *either* shared library or static library, but not
both. Building both makes things more complicated because they had
different names (lokinet-shared or lokinet-static) and seems pointless:
you generally want one or the other. Now there is just the liblokinet
target, which will be shared or static depending on the value of
BUILD_SHARED_LIBS.
- Simplify lokinet-cryptography AVX2 code: just build *one* library, and
add in the additional AVX2 files when possible, rather than building two
and needing to merge them.
- Compress STATIC_LINK and STATIC_LINK_RUNTIME into just STATIC_LINK.
It makes no sense to use one of these (_RUNTIME) on Windows and the
other on non-Windows when they appear to try to do the same thing.
- remove a bunch of annotations from `endif(FOO)` -> `endif()`.
- move all the tuntap compilation code (including OS-specific source
file selection) into vendor/CMakeLists.txt and build tuntap as an
intermediate OBJECT library rather than keeping a global variable in 5
different files.
- move release motto define to root cmake; it made no sense being
duplicated in both unix.cmake and win32.cmake
- fix add_log_tag to not stomp on any existing source compile flags with
its definition. Also use proper compile definition property instead of
cramming it into compile flags.
- make optimization/linker flags less hacky. There's no reason for us
to force particular optimization flags because the cmake build type
already does that (e.g. -DCMAKE_BUILD_TYPE=Release does -O3). Not doing
that also silences a bunch of cmake warnings because it thinks "-O0 -g3"
etc. are link libraries (which is reasonable: that's what the code was
telling cmake they are).
- sets the default build type to RelWithDebInfo which gives us `-O2 -g`
if you don't specify a build type.
- Move PIC up (so that the things loaded in unix.cmake, notably libuv,
have it set).
- Add a custom `curl` interface library that carries the correct link
target and include paths for curl (system or bundled).
Right after project() which contains the version, so that version +
motto updates are basically in one place (and so that non-Makefile cmake
invocations get it).
CMake will set version variables itself if you give the version in the
project(), which is cleaner. Also removes some (nearly) duplicate
definitions and settings added in basic_definitions.cmake for unknown
reasons.
Removes some redundant settings (name, description, version) from the
cpack settings which already default to the values from the project()
call.
Right after project() which contains the version, so that version +
motto updates are basically in one place (and so that non-Makefile cmake
invocations get it).
CMake will set version variables itself if you give the version in the
project(), which is cleaner. Also removes some (nearly) duplicate
definitions and settings added in basic_definitions.cmake for unknown
reasons.
Removes some redundant settings (name, description, version) from the
cpack settings which already default to the values from the project()
call.
* update travis ci clang-format to clang-format-9 because we use that now
add python3-dev package becuase something inside ci thinks it has python but really does not
* try using python3.6
* try working around hot garbage that is travis-ci
* add deadsnakes repo for python3.8
* prevent nullptr deference when running in unit tests
* move python3.8 to main dependancy matrix and add python3.8 to homebrew deps
* add deadsnake apt repo
* add deadsnakes and python3.8 back to previous matrix
* dev package for python
* toggle hive build in ci
* dont add pybind11 if not bulding hive
* revert setting pyenv shim for travis ci
* make native builds on by default except for windows ci
* only apply native build being off for windows release target becuase that broke macos
Howard Hinnart's date.h is the library that was accepted as C++20
date/calendar support, so this is essentially a backport of C++20 date
time support.
(It does support timezone support, but requires more of the library and
that seems like overkill for what we need; this just prints UTC
timestamps instead, which need only a header-only include).
Step 1 of removing abseil from lokinet.
For the most part this is a drop-in replacement, but there are also a
few changes here to the JSONRPC layer that were needed to work around
current gcc 10 dev snapshot:
- JSONRPC returns a json now instead of an optional<json>. It doesn't
make any sense to have a json rpc call that just closes the connection
with returning anything. Invoked functions can return a null (default
constructed) result now if they don't have anything to return (such a
null value won't be added as "result").
Our current abseil won't build with gcc 10 (its `optional`
implementation appears broken), and spews warnings under slightly older
compilers; updating to the latest stable 2019 branch fixes both issues.
We currently spend a bunch of time setting CRYPTO_FLAGS and then just
completely wipe it out if NATIVE_BUILD is specified.
The AMD_RYZEN_HACK is really not needed; it was only some early CPU
microcode and compiler combinations that had trouble detecting Ryzen's
FMA3 support (and Ryzen *doesn't* properly support FMA4--it can be
forced on and runs, but apparently it can give wrong results).
Tuning to an ancient architecture doesn't make a lot of sense; we want
to support the ancient architecture, but don't want to optimize for it.
Also change the AVX2 tuning to use haswell so that optimizations don't
depend on the CPU in the system the build runs on.
This rewrites the version info using lokid's approach of compiling it
into a .cpp file that gets generated as part of the build (*not* during
the configure stage).
Among other things, this means that changing the version no longer
invalidates ccache or cmake dependencies, and because it depends on
`.git/index` git commits will cause the version to be regenerated,
making the commit tag more reliable (currently if you rebuild without
running cmake your git commit tag doesn't update).
- xenial's cmake version (3.5.1) builds everything fine and test suite
passes, so lower the minimum to that.
- add a hack for xenial's kernel header & glibc version breaking if both
net/if.h and linux/if.h get included. The only thing we actually need
from net/if.h that linux/if.h doesn't have is `if_nametoindex`, so
just hack that definition in for xenial's specific glibc/kernel header
versions.
link libatomic if we're targeting 486
link libatomic.a if we're targeting windows
idk what the hell MSVC does for -arch:IA32
we already set the c++14 flag early on
strip target selection flags from MSVC builds and clang-cl
c++14 does not propagate to compile tests
