GitLab

Bug: https://github.com/android-ndk/ndk/issues/222 : CMake toolchain
file's `CMAKE_CXX_COMPILER_ID_RUN=true` breaks compiler feature
detection
Bug: https://github.com/android-ndk/ndk/issues/253 : cmake toolchain
file does not set CMAKE_CXX_COMPILER_VERSION
Bug: https://gitlab.kitware.com/cmake/cmake/issues/16587 : Adding
"-march=" to flags breaks determining compiler ID when cross-compiling
with Clang

The current comment reads:

> Clang can fail to compile if CMake doesn't correctly supply the target
> and external toolchain, but to do so, CMake needs to already know
> that the compiler is clang. Tell CMake that the compiler is really
> clang, but don't use CMakeForceCompiler, since we still want
> compile checks. We only want to skip the compiler ID detection
> step.

The issue arises in CMakeDetermineCompilerId.cmake, which passes along
user-specified flags like `-march=armv5te` but not the specified
compiler target. On Clang, this fails, since the validity of `-march`
depends on the current target.

I've filed a bug for CMake to fix this (allow determining the compiler
ID without using user flags). In the meantime, I think a more robust
solution than overriding CMAKE_CXX_COMPILER_ID_RUN is to run the
compiler ID checks before passing along any flags. The
CMakeDetermineCCompiler.cmake and CMakeDetermineCXXCompiler.cmake
modules are not marked for internal-use-only, and this method takes a
more proactive approach to the problem.

Change-Id: I2c250ebc2999b3f251417a9e3730b5e093e446d1

* changes:
  Use the linker scripts for libc++ in cmake.
  Set more clang CMake compiler info.

We had libc++_shared specified incorrectly here (wasn't using
libandroid_support). Just use the linker scripts that were created to
avoid this problem.

Test: ./validate.py
Bug: None
Change-Id: Ifcc87954d2b889c63c618c7f70f57a5d747b00f2

Necessary for projects performing checks against the compiler
version. We don't need this for GCC because cmake will compute it,
but we have to hack around that for clang.

Test: No useful test to add since this number has to be updated
      manually anyway.
Bug: https://github.com/android-ndk/ndk/issues/253

Change-Id: I314c01628fe038a3fcade379b6a6119dce094572

* changes:
  Remove support for awk tests.
  Remove awk from the NDK!
  Replace gen-windows-host-path.awk with Python.
  Remove gen-cywgin-deps-converter.awk.
  Remove unused awk scripts.
  Replace extract-debuggable.awk with Python.
  Replace extract-minsdkversion.awk with Python.
  Replace extract-platform.awk with Python.

We don't have any awk in the NDK any more.

Test: ./validate.py
Bug: None
Change-Id: I44ea7e10b1c25ca4c7b0050d2d00e6b426ecd1d1

Test: ./validate.py
Bug: None
Change-Id: I5528663072f30a6def6656bad5113a3253fb9213

No more awk dependencies!

Test: nose2
Bug: None
Change-Id: I518f9542ca336d2e606c3088c9a4f53b65fb5418

As far as I can tell we don't need this at all. For one, Cygwin users
can invoke ndk-build.cmd (and in fact, at the moment that's their
only choice out of the box) and in that situation HOST_OS is
"windows" rather than "cygwin" and this code doesn't get run at all.
If you manually create the ndk-build (shell) shortcut for cygwin,
this isn't needed anyway since all the paths in dep files seem to be
relative paths.

At some point we should figure out if we need the cygwin support at
all. As far as I can tell it's just a bunch of hacks that we don't
need.

Test: ./run_tests.py --abi armeabi-v7a  # In cygwin
Bug: None
Change-Id: Iff080a0bfdf318f781d0a4ca6455b36c0dbaea06

Test: ./checkbuild.py
Bug: None
Change-Id: I5805d9456ec91f601eeb7253ca731a90f9cae73e

Test: nose2
Bug: None
Change-Id: I5eb5bdbab09734aa3597541ca958feb5490a1858

Test: nose2
Bug: None
Change-Id: I259d7adbd5ffe88b637a2ed6b6bf19506c7e8a73

Test: nose2 && ./run_tests.py --filter project-properties
Bug: None
Change-Id: I4479a66881f21f97cec6a213d560f190029f5a2c

This was only being cleaned up on a clean exit, meaning the directory
had to be manually removed before the next run.

Test: ./validate.py  # Wait for tests to start
      ^C
      ./validate.py  # Doesn't abort
Bug: None

Change-Id: I507c10045ff69657743aecd840cb6ddabf2cce1b

We don't really care about stlport any more, and these tests make up
about 1/3 of the entire validate.py running time (from ~28 minutes to
~18 minutes).

There are still some stlport tests here to make sure it doesn't break
entirely, but we certainly don't need to be running the unit tests
four different ways.

Test: ./validate.py
Bug: None
Change-Id: I6734815450b8cab39d7432759d488b600d41862d

This was generating things like "build\ndk-build\ndk-build..cmd".

Test: ./checkbuild --module ndk-build-shortcut --system windows64
Bug: None
Change-Id: Ic1dbafef6bf6a372f7421cfe63560d99e7fc711c

adb pushing each file individually is a lot slower than pushing the
whole test directory. The downside is that we have to push the whole
directory even if a test filter means that we only need one
executable (we can fix this in a follow up by looking at the test
filter).

This saves us a couple minutes in a validate.py run (22 minutes down
from 25).

Test: ./validate.py
Bug: None
Change-Id: I8b7e34106e84389cb63ff260af7019d98f4de277

Having an exception thrown in a corner case of a test 75 minutes into
a test run and bringing down the whole thing is terrible. Catch any
exceptions in tests and translate them into failures.

Test: ./validate.py --filter standalone-toolchain
Bug: None
Change-Id: Ib468e4639378c17a8851d7868eb62d18d95753ca

Test: run checkbuild.py.

Change-Id: I33ca1dbfef4e61cfa586e3187ff67a6b89467ed0

Test: None, markdown only.
Bug: https://github.com/android-ndk/ndk/issues/270
Change-Id: I53c7d65a83a39c9e2600b162330782238579458a

* changes:
  Include configuration in the test build dir.
  Use a temp dir within out for validate.py.
  Improve g++ use detection.

Some people expect this to default to the latest, some expect the
earliest, others expect it to be a required argument. The latter
seems like the best choice, but we're stuck because it's historical
and people don't like it when we change things, so just print a
warning.

Test: build/tools/make_standalone_toolchain.py --arch arm
Bug: https://github.com/android-ndk/ndk/issues/270
Change-Id: Idb4cd6d90eeabb788778540e52c2c2eb809308cc

* changes:
  Better Clang LTO: install LLVMgold to bfd-plugins.
  Support GCC LTO in cmake.

* changes:
  Remove intermediate packaging step for GCC.
  Remove intermediate packaging for most builds.

ar will automatically load any plugins that are installed to
../lib/bfd-plugins. Rather than needing to fix all of our build
systems, we can make LTO Just Work if we copy the gold plugin (and
its dependencies) into the binutils directory.

I've also added a test for cmake LTO.

Test: ./validate.py --filter flto
Bug: https://github.com/android-ndk/ndk/issues/108
Change-Id: I1db61b8ef3405474c64ff20478daa86ca41ab3b3

Not that we actually care since GCC is deprecated, but it was easier
to fix than add the ability to disable tests based on build system.

Test: ./validate.py --filter flto # added in next commit
Bug: None
Change-Id: I18aebf9252a392e248b5960d2e14bc2f335f2e31

Test: ./checkbuild.py --module gcc # diffed with prior install
Bug: None
Change-Id: Ie1405046ecdae1bcc6301fb920c355009947e138

There's no reason for us to zip these just to unzip them now that the
modular NDK is long gone. Now that we have a build system that can
work with that, just install them.

Test: ./checkbuild.py
Bug: None
Change-Id: Ic75d888e2ad15720a149fbe270d54a4f29b32ab0

This gets LTO working for Clang on the supported hosts (no Windows
support yet) and targets (we don't have gold for MIPS).

This also switches us to the small code model when building
executables for non-Intel targets (to match the PIC options).

Test: ./validate.py --filter flto
Bug: https://github.com/android-ndk/ndk/issues/108
Change-Id: I465e475282d17b2137dcebcbab53e53b9372c837