[LLVMdev] Building sanitizers for Android

Alexey Samsonov samsonov at google.com
Wed Apr 2 07:56:48 PDT 2014 

Hi Greg,

On Wed, Apr 2, 2014 at 2:50 AM, Greg Fitzgerald <garious at gmail.com> wrote:

> Alexey, Evgeniy,
>
> I propose the following steps to unify multi-arch support in compiler-rt:
>
> 1) The compiler-rt test suite adds "-L${CMAKE_CURRENT_BINARY_DIR}/lib"
> to its 'clang' variables.  This way we can test the sanitizers without
> installing any libs to the just-built-clang install directory.
>

This is possible, but please keep in mind that:
1) we would still want to use compiler-rt test-suite in a standalone mode,
to test fully built/installed toolchains,
and even GCC.
2) Adding -L... wouldn't work: Clang driver links the static xsan runtimes
from a hardcoded paths in Clang resource directory,
and doesn't add flags like "-lasan -L/path/to/clang/resource/dir". I find
this behavior reasonable.

I don't see a problem with the current approach - we can make "run
sanitizer test suite" command in the
top-level build tree depend on "build/install compiler-rt ExternalProject"
steps (see how it's done currently).


>
> 2) The "clang/runtime" build calls ExternalProject once for each arch
> it needs of compiler-rt.  So once to create x86_64 libs and once for
> i386 libs.
>
> 3) The compiler-rt build drops the ${arch} suffix from its libs, and
> the "clang/runtime" build uses CMAKE_INSTALL_PREFIX to control where
> the compiler-rt libs go.
>
> "CMAKE_INSTALL_PREFIX=${CMAKE_CURRENT_BINARY_DIR}/lib/clang/${CLANG_VERSION}/x86_64-linux"
>

> 4) Remove multi-arch support from the compiler-rt build.  Instead,
> declare compiler-rt as an "any-arch" build, configured with:
>     CMAKE_CXX_COMPILER (defaults to just-built-clang)
>     CMAKE_CXX_FLAGS (defaults to llvm's default target)
>     COMPILER_RT_TEST_COMPILER (defaults to CMAKE_CXX_COMPILER)
>     COMPILER_RT_TEST_FLAGS (defaults to CMAKE_CXX_FLAGS)
>

Unfortunately, this could be problematic for Mac - we use "universal
binaries" there - a single static/dynamic runtime
which contains runtimes for all the possible architectures. We might work
around this, however, by adding a custom
Mac-specific step that would merge all single-arch binaries into a
multi-arch one - I think this is fine as long as we
greatly simplify the rest of the ugly build system.

Do you want to give it a try?


>
> Thoughts?
>
> Thanks,
> Greg
>
> On Tue, Apr 1, 2014 at 12:58 AM, Evgeniy Stepanov
> <eugeni.stepanov at gmail.com> wrote:
> > It does sound like Android is better suited for "honest"
> > cross-compilation, rather than "build compiler-rt for all targets we
> > can find" model.
> >
> > I'm still not convinced that we must require the "ninja install" step.
> > Could we just "ninja clang" and then build the second stage against
> > the first stage build directory? Will this "find_package" thing not
> > work that way, or do you mean it as a way to copy target asan runtime
> > where the first stage compiler will find it (or probably both)?
> >
> > On Mon, Mar 31, 2014 at 6:46 PM, Alexey Samsonov <samsonov at google.com>
> wrote:
> >>
> >> On Fri, Mar 28, 2014 at 10:56 PM, Greg Fitzgerald <garious at gmail.com>
> wrote:
> >>>
> >>> > Note that ASan tests on Android require llvm-symbolizer binary.
> >>>
> >>> That's a really good point.  And I see that llvm-symbolizer can't just
> >>> be pulled into compiler-rt because it has dependencies on DebugInfo,
> >>> Object, and Support libraries.
> >>>
> >>>
> >>> This throws a big wrench in Alexey's plan to have the native
> >>> compiler-rt build generate the cross-compiled binaries for all
> >>> supported targets.  We would need to do the same in the llvm repo.
> >>> The alternative is to run a separate cross-compiled build for each
> >>> architecture.  :-/
> >>
> >>
> >> Yes, this llvm-symbolizer part is frustrating... I though we can get
> away
> >> with it at first, as llvm-symbolizer is optional, and we generally
> won't be
> >> able
> >> to produce working *binaries* for architectures different from the host
> one
> >> (we might not have a linker, for instance).
> >> Android with its NDK looks like an exception, though.
> >>
> >>>
> >>>
> >>>
> >>> > Yes, so "ninja install" is optional.
> >>>
> >>> But the 'install' depends on the 'all' target.
> >>>
> >>> -Greg
> >>>
> >>>
> >>>
> >>>
> >>>
> >>>
> >>> On Fri, Mar 28, 2014 at 12:39 AM, Evgeniy Stepanov
> >>> <eugeni.stepanov at gmail.com> wrote:
> >>> > On Thu, Mar 27, 2014 at 10:18 PM, Greg Fitzgerald <garious at gmail.com
> >
> >>> > wrote:
> >>> >>> Alexey's approach with CMake sub-projects.
> >>> >>
> >>> >> I prefer that direction as well, but what I've proposed is a
> solution
> >>> >> that works today.  To support cross-compilation, we'll need to loop
> >>> >> over each supported arch (llvm-config --targets-built), then loop
> over
> >>> >> each supported triple for each arch (hard-coded map?), and then pair
> >>> >> up each triple with a sysroot (system paths provided by the user).
> >>> >>
> >>> >>
> >>> >>> I think it already copies
> >>> >>> compiler-rt build products to the parent build directory, right?
> >>> >>
> >>> >> The Android compiler-rt build does this, but in an unnecessarily
> >>> >> complex way.  My proposal is a simplification of that process.  What
> >>> >> is described in "llvm/cmake/platforms/Android.cmake" is to build
> >>> >> llvm+clang for the host architecture and then run a second build to
> >>> >> cross-compile llvm+compiler-rt to get a the ASan shared object and
> >>> >> test suite.  Instead, I propose building only llvm for the host
> >>> >> architecture (for the purpose of building llvm-config) and then only
> >>> >> compile compiler-rt for Android.  No need to build clang or to
> >>> >> cross-compile llvm.
> >>> >>
> >>> >>
> >>> >>> Requiring that llvm is installed complicates things.
> >>> >>
> >>> >> Sorry for the confusion.  By "llvm install directory", I don't mean
> to
> >>> >> install llvm to the root directory.  I'm referring to the
> intermediary
> >>> >> directory pointed to by CMAKE_INSTALL_PREFIX.  When you run "ninja
> >>> >> install", the build populates this directory with only the libs,
> >>> >> headers, docs, and executables that the LLVM build intends customers
> >>> >> to use directly.  In the case of compiler-rt, we need LLVM's
> >>> >> "bin/llvm-config" to determine the build mode, flags, and targets
> >>> >> available.
> >>> >
> >>> > Yes, so "ninja install" is optional. This should work with LLVM build
> >>> > directory just as well.
> >>> >
> >>> >>
> >>> >>
> >>> >>> This should not be too hard to fix.
> >>> >>
> >>> >> Porting the LLVM build to Android?  That seems like a more
> challenging
> >>> >> solution than to cut out the unused cross-compiled llvm build.
> >>> >
> >>> > Note that ASan tests on Android require llvm-symbolizer binary.
> >>> > Building just compiler-rt for the target would not be enough.
> >>> >
> >>> >>
> >>> >> -Greg
> >>> >>
> >>> >> On Thu, Mar 27, 2014 at 1:14 AM, Evgeniy Stepanov
> >>> >> <eugeni.stepanov at gmail.com> wrote:
> >>> >>> I'd prefer something based on Alexey's approach with CMake
> >>> >>> sub-projects. I think it already copies compiler-rt build products
> to
> >>> >>> the parent build directory, right?
> >>> >>>
> >>> >>> Requiring that llvm is installed complicates things.
> >>> >>>
> >>> >>> On Thu, Mar 27, 2014 at 5:00 AM, Greg Fitzgerald <
> garious at gmail.com>
> >>> >>> wrote:
> >>> >>>> The build for the Android sanitizers is unique in that it needs to
> >>> >>>> link against the Android system libraries to create a shared
> object
> >>> >>>> and its test suites.  The current solution to build ASan is to
> drop
> >>> >>>> the compiler-rt repo into the llvm source tree and cross-compile
> the
> >>> >>>> llvm build for Android.  This is a bit awkward for few reasons:
> >>> >>>>
> >>> >>>> 1) Not all of llvm can be cross-compiled for Android.  "ninja all"
> >>> >>>> causes expected build errors.
> >>> >>>
> >>> >>> This should not be too hard to fix.
> >>> >>>
> >>> >>>> 2) The sanitizers don't depend on any LLVM libraries, only
> >>> >>>> llvm-config
> >>> >>>> for its build configuration.
> >>> >>>> 3) No "install" rule.  Instead, you cherry-pick files from the
> build
> >>> >>>> directory.
> >>> >>>>
> >>> >>>> Building against the LLVM install directory cleans this up nicely
> and
> >>> >>>> as it turns out, this mostly works today (see CMake configuration
> >>> >>>> below).  The only missing pieces are that the shared object is not
> >>> >>>> added to the install directory and the test suites are not built.
>  Is
> >>> >>>> this a build configuration you'd consider using?
> >>> >>>>
> >>> >>>> $ cmake -G Ninja .. \
> >>> >>>>     -DCMAKE_INSTALL_PREFIX=ship \
> >>> >>>>     -DCMAKE_C_COMPILER=arm-linux-androideabi-gcc \
> >>> >>>>     -DCMAKE_CXX_COMPILER=arm-linux-androideabi-g++ \
> >>> >>>>     -DCMAKE_PREFIX_PATH=`pwd`/../../llvm/out/ship \
> >>> >>>>     -DANDROID=1 \
> >>> >>>>     -DCMAKE_SYSTEM_NAME=Linux \
> >>> >>>>
> -DCMAKE_C_FLAGS=--sysroot=$(ndkDir)/platforms/android-19/arch-arm
> >>> >>>> \
> >>> >>>>
> >>> >>>>
> -DCMAKE_CXX_FLAGS=--sysroot=$(ndkDir)/platforms/android-19/arch-arm
> >>> >>>> ...
> >>> >>>> $ ninja install
> >>> >>>> ...
> >>> >>>> [68/68]
> >>> >>>> -- Install configuration: "Release"
> >>> >>>> -- Installing: ship/include/sanitizer/asan_interface.h
> >>> >>>> -- Installing: ship/include/sanitizer/common_interface_defs.h
> >>> >>>> -- Installing: ship/include/sanitizer/dfsan_interface.h
> >>> >>>> -- Installing: ship/include/sanitizer/linux_syscall_hooks.h
> >>> >>>> -- Installing: ship/include/sanitizer/lsan_interface.h
> >>> >>>> -- Installing: ship/include/sanitizer/msan_interface.h
> >>> >>>> -- Installing: ship/include/sanitizer/tsan_interface_atomic.h
> >>> >>>> -- Installing: ship/asan_blacklist.txt
> >>> >>>> -- Installing: ship/bin/asan_device_setup
> >>> >>>>
> >>> >>>> Thanks,
> >>> >>>> Greg
> >>
> >>
> >>
> >>
> >> --
> >> Alexey Samsonov, MSK
>



-- 
Alexey Samsonov, MSK
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20140402/961c5eac/attachment.html>

More information about the llvm-dev mailing list