Tracking master subjects us to breakages (as happened when
luacheck-scm-1.rockspec got renamed to luacheck-dev-1.rockspec) and
makes older releases unbuildable when that happens.
LuaRocks 2.3 and onwards changed the /P option to no longer include the
version number which made newer releases of LuaRocks fail when compiling
on Windows.
For CI builds unibilium is provided through msys2 packages, and
libtermkey is built from source in third-party from equalsraf/libtermkey.
In Windows we cannot read terminal input from the stdin file descriptor,
instead use libuv's uv_tty API. It should handle key input and encoding.
The UI suspend is not implemented for Windows, because the
SIGSTP/SIGCONT do not exist in windows. Currently this is a NOOP.
Closes#3902Closes#6640
https://github.com/mpeterv/luacheck/pull/81#issuecomment-261099606
> If you really want to use bleeding-edge version you should get the
> rockspec from master branch, not a fixed commit ...
> The correct way to install from a specific commit is cloning that
> commit and running "luarocks make" from project directory. The reason
> is that running "install" or "build" on an scm rockspec fetches
> sources from master but uses build description from the rockspec
> itself, which may be outdated.
Using /MT was causing issues when building luarocks, revert it, use the
dynammic runtime and generate release DLLs for the dependencies.
Some refactoring was required because for linking cmake looks for the
import libraries (.lib) but on runtime executables we need the .dll files
to be in the same folder.
The DLLs are placed in the bin/ folder in order for nvim.exe to run
during the build and tests. The install target installs the DLLs with
the nvim binary - uses GetPrerequisites to find runtime DLLs.
Some minor issues that required adjustments:
- [MSVC] FindMsgpack.cmake now looks for msgpack_import.lib instead of
msgpack.lib
- The lua-client fails to find libuv.lib, instead it looks for uv.lib,
added second copy of the file to the install command.
- [MSVC] CMAKE_BUILD_TYPE affects the output paths, default to Release.
Part of these changes are credited to @jasonwilliams200OK who fixed the
third-party recipes to consistently use the same build type.
Luv is a simple lua binding to libuv, which is now used by neovim lua client.
The bundled luv installation a bit different from other dependencies in that it
is installed two times:
- The "BuildLuv.cmake" script downloads and installs a static version of luv
using its normal cmake build script. This static version will be used later.
- Luv default rockspec is replaced with the alternate under the "rockspecs"
directory(the alternate rockspec plays nicer with neovim build system)
- The alternate rockspec is used to build/install the lua module and make it
available to lua scripts.
Otherwise the dynamic library is built also and find_library will prefer
that over the static one. That results in linking against the dynamic
library which will not be found after install.
This code:
8b3c399b6d/third-party/CMakeLists.txt (L130)
should prevent the above problem, but it doesn't hurt to be explicit.
So far luacheck's rockspec specified only the git protocol. Hence people
behind firewalls/proxies, that block port 9814, had trouble fetching this
dependency via luarocks.
The latest commit updated the rockspec to use either git or https. Thus common
workarounds like this are not needed anymore:
git config --global url."https://".insteadOf git://
References #3769.
Libuv and LuaJIT like to hide the actual compilation and linking
commands behind nice text. This change makes them spit out the actual
command line to help us with debugging issues that people are seeing.
Introduce ALLOW_EXISTING_SRC_DIR option, turned off by default.
The Homebrew formula, which downloads and extracts the third-party
dependency sources before starting the build, would turn this option
ON.
This effectively reverts 585e5d32a3
The pinning was done at a time when `lua_cliargs` caused test failures,
so an older version which didn't was pinned. We're now using the latest
version (2.5-1), so the cause of those failures were presumably fixed.
- Update recipes to build with MSVC or cross compile in Unix with Mingw
- For recipes that need to be reused, wrap recipe in CMake function
using cmake_parse_arguments
- New directory .deps/host is the install root for HOST targets, the old
.deps/usr is used for TARGET
- In windows disable builds for terminal libraries and jemalloc
- Added cmake script CopyFilesGlob.cmake to copy files using glob
cmake -DFROM_GLOB=*.h -DTO=/usr/include -P CopyFilesGlob.cmake
- New CMake variables HOSTDEPS_* can be used in cross compile recipes.
Except when the target is UNIX, since that would break 32bit builds
in 64bit Unix systems using the Travis 32bit toolchain
We didn't have ansicolors pegged, and several others were below the
minimum required versions causing busted to go out and grab newer
versions anyways. Let's peg them all to useful versions.
Libtermkey can be linked against unibilium or curses. For the bundled
dependencies Neovim links against static versions of libtermkey and
unibilium, after building both libraries.
However libtermkey requires pkg-config to be installed in order to detect
and link against unibilium, otherwise it falls back to curses by default.
In systems where pkg-config is not installed building Neovim against the
bundled libtermkey caused a linking error (#2484).
So pkg-config needs to be installed for the bundled libtermkey to build
properly.
* Set JEMALLOC_NO_DEMANGLE to be able to use `je_*` functions,
regardless of how jemalloc was compiled (--with-jemalloc-prefix)
* Show jemalloc information in Neovim's version output.
Resolve#2449.
- When not running the unit tests, building the third-party busted,
nvim-client and other dependencies can be skipped by passing
-DUSE_BUNDLED_BUSTED=FALSE to the third-party project.
Jemalloc will be used if the cmake option `USE_JEMALLOC` is enabled(which is the
default). To avoid trouble with clang's ASAN, it is disabled by default if the
`SANITIZE` option is enabled.
Since jemalloc has thread cache for small objects, it fills the gap created by
removing klib memory pools.
The `xstrdup` funciton(memory.c) had to be reimplemented on top of `xmalloc` to
make it work with a custom allocator.
This requires a couple of extra modules that are not installed by
default, and it requires capturing stdout of the tests--otherwise CMake
output is intermixed with the XML output of busted.
This is a variant of the utfTerminal output handler that will:
- Output the file name before each suite is executed
- Output the test name before each test is executed
This will make it simpler to identify crashing/hanging tests.
It turns out that `file(DOWNLOAD ...)` is not very user friendly with
it's error message, and only supports MD5 on v2.8.10 of CMake (the
default for Ubuntu 12.04). If CMake is built without SSL support,
users are left hanging with a message that the hashes don't match.
It turns out that `file(SHA1 ...)` exists in v2.8.10, and we
can use that to compute the hash ourselves. So this splits the hash
checking into a separate step, where we can provide some additional
advice if the SHA1 is the hash for an empty file. Additionally, it also
allows us to drop the MD5 hashes and maintain only SHA1 hashes for our
dependencies.
This is useful when trying to bisect an issue in a dependency, and we
want to pull from a specific commit but don't want to have to download
and determine the sha1sum and md5sum of the tarball.
- If possble try to abstract away from Make, and use cmake --build
- third-party still needs to find Make to build some components
- Removed search for Make from CMakeLists.txt
* for CMake < 3.0 --build has no color output
When building under Homebrew, we want to let Homebrew manage
downloading and extracting the tarballs. See PR #1411.
Also make sure to skip only if directory is not empty. Fix#1433.
Also, the command was removing only .so versions of the library, and not
the the `.dylib` under Mac OS X, and in a way that it would fail if the
files weren't present.
Instead, let's delegate to a CMake script--to get the portability--and
use a glob to detect and remove the shared versions of the library.
Hopefully, this will become unnecessary as msgpack's build becomes more
full-featured, and we can just tell it to build the static version
instead.
Underneath the hood, CMake uses libcurl and libcurl has had a number of
issues regarding progress feedback. In one sample run against Travis
CI, we ended up with nearly 3,000 lines of progress output for a single
download.
Unfortunately, CMake doesn't have the download and extract steps
separate, so we have some extra work that we have to do. Much of the
content was taken from the ExternalProject.cmake and it's template for
generating the content of the download and extract CMake files.
