When invoking clang for the host to assemble .S files, the -B flag
would not be provided, which allowed the host prebuilt clang to
use an 'as' from the native environment. Most of the time this
"just works", but some newer 'as' versions cause problems with
the older prebuilt toolchain, for example by generating
unsupported relocation types.
To avoid this problem, simply use the -B flag to tell clang to
invoke the assembler from the correct prebuilt gcc prefix.
(cherry-picked from AOSP commit 5ff9c81b6c)
Change-Id: I18ea4f37ae637b652cfd7321c41929f8be075342
Signed-off-by: Alistair Strachan <alistair.strachan@imgtec.com>
http://b/26987366
- Use clang-2690385 to build RenderScript toolchain.
- Use armv7-linux-androideabi as the triple used to build RenderScript
runtime libraries.
(cherry picked from commit 41f4b242e3)
Change-Id: I50751f410a2f7c83532cec9b96c50d9afe1224c1
Bug: http://b/27482631
This fixes an issue where krait compiles were not passing the proper
flags to the assembler when -no-integrated-as is used. This was
preventing ART from switching to clang in some cases.
(cherry picked from commit c503d8ef58)
Change-Id: I4a0efb78f37fee73bf61841f0347328b8f61c192
* When WITH_STATIC_ANALYZER is set and non-zero, and clang compiler is used,
call new clang ccc-analyzer or c++-analyzer.
* Otherwise, if WITH_SYNTAX_CHECK is set and non-zero,
call compiler with -fsyntax-only.
* Replace "--sysroot=path" with "--sysroot path", to work with ccc-analyzer.
* ccc-analyzer executes the original compilation command to generate
object files before calling clang with --analyze to do static analysis.
* When clang is called with --analyze, macro __clang_analyzer__ is defined.
BUG: 13287788
(cherry picked from commit 765c1ea6d7)
Change-Id: I6e51e51ff4ed3ce514f60d090494dcdf6e520b04
This is mostly the same as the existing 2ND_HOST / HOST_CROSS support.
The interesting thing I did here was make x86 the 'first' architecture,
and x86_64 the second. This way LOCAL_MULTILIB := first defaults to
32-bit windows modules.
windows-x86/bin <- defaults to 32-bit executables
windows-x86/lib <- 32-bit libraries, like before
windows-x86/lib64 <- 64-bit libraries
windows-x86/obj <- 32-bit intermediates
windows-x86/obj64 <- 64-bit intermediates
Then modules are registered with the names:
host_cross_liblog <- 32-bit, like before
host_cross_liblog_64 <- 64-bit
Bug: 26957718
(cherry picked from commit 9ecbf83259)
Change-Id: Iab2d36e20d7002568c3534e7781e9ac8ab21a4a9
This moved from lib/ to lib64/, but wasn't noticed because no one builds
with the profiler on by default.
Change-Id: I0155263b4a50437ea0864338fb34baefc3df59d2
This change adds new variables for RS_LLVM_* tools, which correspond
to the version of LLVM checked in currently to external/, instead of
mapping to the latest host prebuilts. This will help prevent version
differences for newer IR generated by the updated host prebuilts, which
then might not be readable with the on-device LLVM bits we build from
external/.
Change-Id: I6c7b0fa9082e9c8244cc508cb23960569a057740
Building with USE_NINJA=false and piping build output to a file will
result in ANSI color codes in the build output. Only force ANSI codes
on when building with ninja, which strips the ANSI codes when writing
build output to a non-TTY.
Change-Id: Ifd2fb53799fd99c23b7ad91cdb84d943bf707db0
This was marked as unsupported only on arm[64], even though it works.
For soong, we want to keep a single list of unsupported clang flags
for all architectures.
Change-Id: I9c4ae0f255c549d2346ba7abae57e22c86ddcc61
RPATH was used in order for ASan executables to prefer ASan libraries
under /data. Now ASan executables use a special loader (linker_asan),
which implements this logic. RPATH is no longer needed.
Change-Id: Ic7a39b022267b80bd0ce3e95a2e822eb308a1fba
Some projects are still built with our host GCC 4.8, which doesn't
support -fstack-protector-strong. The combo .mk files are used by
GCC and clang, so it's not safe to turn on -fstack-protector-strong
there. Instead, do it in the clang-specific .mk for now.
We can clean this up when elfutils (the last code built for the host
with GCC that I'm aware of) is built by clang. We'll be able to
remove the host GCC prebuilts too!
Change-Id: I314b9eab071c132a8e2cb8cc779a75ae8abb12e2
Not sure why my previous testing was working without this, but kati is
a bit pickier about ?= vs :=.
Change-Id: I80243a3c93dec7fd9b8e6f92e2c6cc956706ace0
The automated build will need to do a two stage build, and we want to
do that without altering the source directory. Building with
`make LLVM_PREBUILTS_BASE=/path/to/prebuilts` will build using an
arbitrary toolchain instead of the installed one.
Bug: http://b/17441393
Change-Id: I4ff5c5224fbb8ebeb3a752907d6c4cd0a925aa1e
This reverts commit f7dbab16ff.
Bug: 23396112
Switch from "-fsanitize-undefined-trap-on-error" to
"-fsanitize-trap=all". The former ends up accidentally leaving
unresolved calls to __ubsan* helper functions in the object file with
clang 3.8. The latter is used when we don't include address sanitizer,
and replaces any misbehavior with a direct call to abort().
GCC & Clang encode the current working directory into the symbol files.
Even if the binaries are stripped before being installed, the build-id
hash will be different. This means that two different checkouts will
produce different installed binaries, even if the sources are identical.
To prevent this, set PWD=/proc/self/cwd, so that it replaces the
absolute path with the static path /proc/self/cwd. With GCC, use
-fdebug-prefix-map to remove that static path. Our clang does not yet
support this flag, but it was merged in r250094.
This does change the behavior when using gdb -- you now need to be in
$ANDROID_BUILD_TOP before gdb will load the sources.
Bug: 24204119
Change-Id: I792d65b042b7a3fdc78d567c5a6e93cdf0141b9b
Bug: 23396112
Switch from "-fsanitize-undefined-trap-on-error" to
"-fsanitize-trap=all". The former ends up accidentally leaving
unresolved calls to __ubsan* helper functions in the object file with
clang 3.8. The latter is used when we don't include address sanitizer,
and replaces any misbehavior with a direct call to abort().
Change-Id: I8a67461b45f5f1dd9f2d179b6b64a4ca905e999f
Force gcc and clang to always output color diagnostics. Ninja will
strip the ANSI color codes if it is not running in a terminal.
Bug: 24273983
Change-Id: Ic9252549bfacf3bbb3f4b9d20f63010fc24ac21d
This doesn't seem to be a problem any more, and mac builds
that differ from linux builds is causing other problems.
This reverts commit 7c40fb95dc.
Bug: 24339250
Change-Id: I8a0da11df42d58ba9c3ded4571f1db870f400a21
Instead of using recursive make to change the HOST_OS when building the
windows SDK under linux, add the concept of cross-building to another
host os.
Bug: 23566667
Change-Id: I6dc525b601b6251d458d197c30bf4660d7485502
We have deferred the CC_WRAPPER prepending to binary.mk per module.
Bug: 22612634
Change-Id: If2c0549503989d17ff7be619fdd49c26fd78f0a5
(cherry-pick from commit 66c8e8c756)
Another change in bionic/linker adds linker_asan/linker_asan64 that
know where to find ASan shared libraries.
Also, include linker_asan to the required packages list when building
for ASan.
Change-Id: I8ebe7c0091bbeb0c135708a891d33d9844373d37
Upstream clang r239152 (http://reviews.llvm.org/D10239) caused a pretty
significant change in behavior. Passing an FPU feature via -mfpu
disables any feature not supported by those flags (in addition to the
old behavior of enabling features supported by the flags). For e.g.,
-mfpu=neon used to just pass +neon, +vfp3 as target features to the
backend. Now, -mfpu=neon also passes -vfp4, -fp16, -fp-armv8.
The backend has always disabled implied feature bits if a feature is
disabled. Upon seeing the target feature -vfp4, it will disable any
processor/feature that implies vfp4, including the bit that the
processor is a Krait. Since Krait has both Neon and vfp4, it is safe to
pass '-mfpu=neon-vfpv4'.
Change-Id: Ibbb992e80b8cbc8dc36d5df556885d0912baea22
