GitLab

Only one real ld.mcld file is under llvm directory, but we duplicate it
for other GCC versions and other targets. Hence, we have 19 duplicate
copies of mcld, which consumes about 350MB for linux-x86 system.

That I use soft-link instead of duplicate copies can shrink on-host
toolchain size greatly.

Change-Id: Ie6ae3144ff56c397f93aa23e75ac0d61120d3c4e

Made the description format use RawTextHelpFormatter
to prevent line breaks in the filename on Windows.

Also replaced \ with / in NDK so it looks more tidy.

Move the determination of NDK_ROOT and NDK_WIN_HOST for ndk-build.cmd
and ndk-gdb-py,cmd into a common batch file - find-ndk-root-win-host.cmd

This also fixes a bug in ndk-gdb-py.cmd on 64bit Windows where it
was hard-coded to assume 32bit Windows.

This is really useful to reduce the size of .zip archives
(about 7 MB / toolchain saved in the final .zip).

For Linux / Darwin, tar automatically detects the hard links
used by the toolchain install script and already copies a single
copy of each file in the archive (and will recreate the hard links
at install time), so this is less useful, but shouldn't hurt.

+ Remove the "simulator" binary from each toolchain bin/
  directory (e.g. arm-linux-androideabi-run) since it's not
  needed (and probably doesn't work anyway).

Change-Id: I9fd4309c6b7bd3c2cac7737c545b6e597d08ca8a

This remove the 4.4.3 toolchain from the NDK, and slightly
updates documentation to not mention it anymore.

Change-Id: Ia42cd54e3a4e88ba5ee3ea35d9ee3cc4b9237ca2

In general it's better for clang to use more recent as/ld, and the
ones comes from binutils-2.23 in gcc4.8 are newer than the ones in
binutils-2.22/gcc4.7

It also help to split NDK package into "primary" which include
gcc4.8/4.6 and clang3.3/3.2, and an "add-on" which includes the rest.

Change-Id: I7f61c370e248a96476722942b822715c6da11b59

* commit 'ba0c25cb':
  Push gabi++ runtime before running the tests.

* commit '9ce7d76c':
  Fix tests for unknown arch.

* commit '2d756b94':
  build-host-gdb.sh stub and Python fixes.

* commit 'c7a57af9':
  Update the wchar-related tests.

Change-Id: I6931a408959e47beb6ced0471985251569d43d02

Change-Id: I7ca92a54aa7f48305e8f0e812f9f912cbbd66557

This updates all wchar_t related definitions to match the fix
currently uploaded at:

  https://android-review.googlesource.com/#/c/62320/

+ Fix a typo in <minitest/minitest.h>

See http://b.android.com/57749

Change-Id: Iaa30eb026c27bac1d3cb4e1870f787f9ca22414e

Both clang3.3 and clang3.2 claim __has_feature(cxx_constexpr) and
turns off _LIBCPP_HAS_NO_CONSTEXPR, but clang3.2 fails to compile
sources/cxx-stl/llvm-libc++/test/../libcxx/include/__mutex_base
line #42 reads

   constexpr mutex() _NOEXCEPT : __m_(PTHREAD_MUTEX_INITIALIZER) {}

clang3.2 will be phased out when 3.4 is released likely the same
time llvm-libc++ is ready for NDK release too.  Let's remove this
test from 3.2

Change-Id: I10a6b0c5f1f6875bcebdcf0bf860cce33c598ce1

* commit 'ab6cfc11':
  Fix test-cupfeatures

am f115098d: Merge "Fixed tests merge-string-literals and prebuilt-copy to observe APP_ABI in env var"

* commit 'f115098d':
  Fixed tests merge-string-literals and prebuilt-copy to observe APP_ABI in env var

-march=armv7-a set earlier in TARGET_CFLAGS isn't compatible with
-mcpu=cortex-a15.  When both exists, GCC 4.7/4.8 uses -march=armv7-a
for codegen (including putting ".arch armv7-a" in the *.s file) and
-mcpu=cortex-a15 for tuning.  Unfortunately, assembler in
binutils-2.22/2.23 (with GCC 4.7/4.8) fails to compile "sdiv"
with ".arch armv7-a" in the *.s file.

GCC 4.6 compiles fine because the older assembler doesn't check

Change-Id: I47eac88b4d2a4d8e8aa8d7fd210d6f2e869d2c9a

* commit '81c782a0':
  Switch x86*-4.7 toolchain to binutils-2.23

* commit 'c9dc9043':
  Fix x86 multi-target gcc build with 32-bit host

Purpose is to enable multi-target compiler with gold linker. x32 as one
of the targets is supported starting from binutils-2.23.

Change-Id: I9199c733cc77adbee908f19de8d626fd728c2b26
Signed-off-by: Pavel Chupin <pavel.v.chupin@intel.com>

It enables 32-bit host compiler build which is able to generate
x86/x86_64/x32 code with -m32/-m64/-mx32 accordingly.

Change-Id: Ia50d10d134078ed9eddea279a50f2affed438922
Signed-off-by: Pavel Chupin <pavel.v.chupin@intel.com>

* commit '8c0e1371':
  cpu-features: Don't assume all VFPv4 devices support IDIV.

Turns out that there are cheap VFPv4 devices which do not support it.
Instead, only add the IDIV feature flags for white-listed devices. For
now, this only means the Nexus 4 kernel.

BUG=57637

Change-Id: If722af2035f63e35241fee112dd9ed064a93c287

* commit '63f93612':
  Fix building mclinker for -fexceptions.