This change fixes following problem:
1) a native lib having stubs is defined.
2) the lib is included in an APEX.
3) a static binary is linking the lib from outside of the APEX.
4) then, the dependency from the binary to the lib is vanishing.
This is happening because cc.depsToPaths() mistakely does not
distinguish static lib deps from shared lib deps. For shared lib deps,
it creates two dependencies (one for stubs variant and the other for
non-stubs variant) and choose the stubs variant when the lib and the
current module is not in the same APEX (i.e. dependency to the non-stubs
variant is discarded). However, since we don't have stubs variant for
static library, it ends up having no dependency to the library if the
link is static.
Fixing the issue by skipping the variant selection routine when the link
is static.
Test: m (apex_test added)
Test: build with https://android-review.googlesource.com/c/platform/bionic/+/849044
Change-Id: I21102a31cc5c0b105da2affdd035bd5cc571a6ab
Ie42edc5184f315f998db953594e425214b810e0e added system_shared_libs to
static libraries so that their exported headers can be referenced.
However, it also added unrequired dependencies, which is an
error-triggering issue for static executables.
This change addresses it by adding a condition to shared libs handling code
in binary.go.
Bug: 121152570
Test: cc_test.go, library_test.go
Change-Id: I1828442c4e496f8d815fccaeca970cd5766bdf5d
When dist'ing a library or binary that has use_version_lib set, always
distribute the stamped version, even for the device.
Test: m test_build_version_test dist
Change-Id: I2995ec516b1d182ce18f099aeaa4d186ffbcf01f
In order to simplify the wrapper function, and stop using a linker
script, generate a set of flags to pass to LLD. Then run
host_bionic_inject on the linked binary in order to verify the
embedding, and give the wrapper function the address of the original
entry point (_start).
Bug: 31559095
Test: build host bionic with prebuilts/build-tools/build-prebuilts.sh
Change-Id: I53e326050e0f9caa562c6cf6f76c4d0337bb6faf
Clang is always used now, so we can remove all the GCC checks. Removing
GCC-specific configuration will happen in the next CL.
Test: m
Change-Id: I4835ecf6062159315d0dfb07b098e60bff033a8a
Passing -pie to clang for Windows binaries causes a warning:
clang-7: warning: argument unused during compilation: '-pie' [-Wunused-command-line-argument]
Bug: 115776263
Test: m checkbuild
Change-Id: Ie502a05dac32c192c38953c684dc80d1db24b792
Export Soong cc modules to Make using a new soong_cc_prebuilt.mk that
bypasses all of prebuilt_internal.mk, dynamic_binary.mk and binary.mk.
This also means that stripping is handled in Soong instead of Make.
Relands If9008c50920779048480f5eeeb0084f26006c998 with fixes for
mac builds.
Bug: 113936524
Test: m checkbuild
Change-Id: I9710ff57f0793f36eb889eabd08bd60a365a88dd
Export Soong cc modules to Make using a new soong_cc_prebuilt.mk that
bypasses all of prebuilt_internal.mk, dynamic_binary.mk and binary.mk.
This also means that stripping is handled in Soong instead of Make.
Bug: 113936524
Test: m checkbuild
Change-Id: If9008c50920779048480f5eeeb0084f26006c998
This is common to binaries and libraries, so move it from library.link
and binary.link to baseLinker.linkerFlags and baseLinker.linkerDeps.
Test: make checkbuild
Bug: None
Change-Id: I5fb24118e601673ae0713a6adc773a1565749be8
Bug: http://b/62424007
This breaks 64-bit Windows executables. Probably due to the Clang
migration, the entry-point need not be set even for 32-bit executables.
Test: 32-bit and 64-bit aapt2_tests.exe and ziparchive-tests.exe can
start. (There are some failed tests, though).
Change-Id: Id3a7d3d2802f6b671332afff4558840bdabddeb5
Recovery partition has same layout as system. In other words,
executables are installed to <recovery_root>/system/bin and libs are
installed to ../system/lib.
This is made possible because the recovery partition is now
self-contained, i.e., shell, adbd, etc. are hosted in the recovery
partition and thus we no longer has to mount the real system.img
to /system. So /system is now available to executables built for
recovery mode.
Bug: 63673171
Test: `adb reboot recovery; adb devices` shows the device ID
Test: Select 'mount /system' in the recovery mode, then `adb shell`.
$ lsof -p `pidof adbd` shows that libm.so, libc.so, etc. are loaded from
the /lib directory.
Change-Id: I6c4f903ad21fd3dd10a07c4588be85d3d678c099
* Pass LLVM_{OBJCOPY,STRIP} to makefile rules.
* strip.sh is used only in soong-only mode.
It will use llvm-strip and and llvm-objcopy
when --use-llvm-strip is given.
Keep flags of strip.sh in alphabetic order.
* Tested build/soong/scripts/build-ndk-prebuilts.sh
by adding "UseClangLld": true, to soong.variables file.
Bug: 80093681
Test: build/soong/scripts/build-ndk-prebuilts.sh
Change-Id: I612267304eea434c7a33cc086b27b577d5f64094
`recovery: true` installs a module to the recovery partition.
`recovery_available: true` makes a module to be available to other
`recovery:true` or `recovery_available: true` modules.
These to are very similar to vendor, vendor_available properties, except
for the target partition.
Bug: 67916654
Bug: 64960723
Test: m -j, toybox_recovery is installed to the recovery/root/sbin
Change-Id: Iaebe0593de16c69fa70de251a61f4d018a251509
An upcoming change will stop exporting ProductVariables from Config, so
switch to using existing accessor functions, and add more when they're
missing.
Bug: 76168832
Test: out/soong/build.ninja is identical
Change-Id: Ie0135bdbd2df3258ef3ddb53e5f8fc00aa9b97f7
Allow native modules to specify use_version_lib, which will make
an android::build::GetBuildNumber() function available. For host
builds, the function will return the build number at the time that
the module was linked. For device modules it will return the
value of the ro.build.version.incremental property.
Bug: 71719137
Test: build_version_test
Test: m build_version_test && touch build/make/core/Makefile build/soong/cc/libbuildversion/tests/build_version_test.cpp && m build_version_test shows different build numbers for binary and library tests.
Change-Id: I6f7d40b7574bb8206866c4e39bad9c710c796e32
AConfig() now duplicates Config(). Replace the uses of AConfig()
with Config(). Leave AConfig() for now until code in other
projects is cleaned up.
Test: m checkbuild
Change-Id: Ic88be643049d21dba45dbd1a65588ed94bf43bdc
there's no use case for prepending/appending to bool, and string
properties within module struct. Declearing "*bool" and "*string" almost
cover everything user need.
I did see one case that user specify relative_install_path as
path prefix in cc_defaults, and concatenate with the one in real module
to get the final relative install path in Android.bp <bionic/tests/libs>.
Test: m -j checkbuild
Bug: b/68853585
Change-Id: If3a7a2689c3fc307aae136af6bc9c57f27a1e1a0
Bug: http://b/63768402
Bug: http://b/65598278
Add support for the 'pgo' property to specify how a module is processed
under PGO. A sample property is below:
pgo: {
instrumentation: true, // could be "sampling: true" when supported
profile_file: "pgo_simple.profdata",
benchmarks: ["pgo_simple"],
}
1. Runtime profiles can be gathered using "sampling" or
"instrumentation". Sampling is not supported initially.
2. If 'toolchain/pgo-profiles' project is found,
'toolchain/pgo-profiles/${profile_file}' is passed to the compiler and
linker when building this module.
3. If ANDROID_PGO_INSTRUMENT environment variable is set, and includes a
benchmark in the 'benchmarks' list, appropriate flags (for e.g.
-fprofile-generate for instrumentation) are passed to the compiler and
linker when building this module.
Test: Add example modules that specify the pgo property and verify
appropriate flags and dependencies in the Ninja file. Some
tests/examples are in https://android-review.googlesource.com/474805
Change-Id: I6242e0c904497a115e367dea6927ba1c4b906355
The linux kernel requires that the ELF interpreter (runtime linker)
that's referenced by PT_INTERP be either an absolute path, or a relative
path from the current working directory. We'd prefer a relative path
from the binary, similarly to how we handle looking up shared libraries,
but that's not supported.
Instead, extract the load sections from the runtime linker ELF binary
and embed them into each host bionic binary, omitting the PT_INTERP
declaration. The kernel will treat it as a static binary, and we'll use
a special entry point (linker_wrapper) to fix up the arguments passed by
the kernel before jumping to the embedded linker. From the linker's
point of view, it looks like the kernel loaded the linker like normal.
Bug: 31559095
Test: Enable host bionic,
out/soong/host/linux_bionic-x86/nativetest64/libdemangle_test/libdemangle_test
Change-Id: I8d0aea9790b5e86fcc3ea6e2d00cfa33907e2853
Change module factories from returning a blueprint.Module and a list
of property structs to returning an android.Module, which holds the
list of property structs.
Test: build.ninja identical except for Factory: comment lines
Change-Id: Ica1d823f009db812c518f271a386fbff39c9766f
The build system currently uses -fPIC for position independent
libraries, and -fpie for position independent executables. However,
these levels are incompatible and conflict when an executable includes
a static lib with the -fPIC flag - such as when building with
LTO. This CL changes the flag for executables to the compatible -fPIE
instead, which is equivalent for all architectures except PowerPC.
Bug: 30227045
Test: ENABLE_CFI=true m -j40 checkbuild # builds and boots
Change-Id: I95b585b553bc00bc1d4f52f4271c5e30e1007d9b
Nothing changes if BOARD_VNDK_VERSION isn't set.
When the VNDK is enabled (BOARD_VNDK_VERSION in Make), this will split
/system and /vendor modules into two different variant spaces that can't
link to each other. There are a few interfaces between the two variant
spaces:
The `llndk_library` stubs will be available in the /vendor variant, but
won't be installed, so at runtime the /system variant will be used.
Setting `vendor_available: true` will split a module into both variants.
The /system (or "core") variant will compile just like today. The
/vendor ("vendor") variant will compile against everything else in the
vendor space (so LL-NDK instead of libc/liblog/etc). There will be two
copies of these libraries installed onto the final device.
Since the available runtime interfaces for vendor modules may be
reduced, and your dependencies may not expose their private interfaces,
we allow the vendor variants to reduce their compilation set, and export
a different set of headers:
cc_library {
name: "libfoo",
srcs: ["common.cpp", "private_impl.cpp"],
export_include_dirs: ["include"],
target: {
vendor: {
export_include_dirs: ["include_vndk"],
exclude_srcs: ["private_impl.cpp"],
srcs: ["vendor_only.cpp"],
},
},
}
So the "core" variant would compile with both "common.cpp" and
"private_impl.cpp", and export "include".
The "vendor" variant would compile "common.cpp" and "vendor_only.cpp",
and export "include_vndk".
Bug: 36426473
Bug: 36079834
Test: out/soong/build.ninja, out/soong/Android- only changes due to _core addition and
.llndk -> .vendor
Test: attempt to compile with BOARD_VNDK_VERSION:=current
Change-Id: Idef28764043bf6c33dc0d2e7e2026c38867ff769
This adds a toolchain definition for LinuxBionic that only supports
Clang/64-bit. It pulls pieces from the x86_linux_host and x86_64_device
configs, and uses the android clang triple, with some manual overrides.
To enable building this, set your soong.config file to:
{"Host_bionic": true}
Bug: 31559095
Test: out/soong/{Android,make-vars}-aosp_arm64.mk the same with or
without host bionic turned on
Test: No change to out/soong/build.ninja before/after this change
Change-Id: Id97dda8bd9aa670c32aed31fbe6aaa8175e70b59
This is configured the same as make -- a global NATIVE_COVERAGE=true
flag to allow native coverage, then COVERAGE_PATHS=path1,path2,... to
turn it on for certain paths.
There are .gcnodir files exported to Make and saved in $OUT/coverage/...
files which are `ar` archives containing all of the compiler-produced
.gcno files for a particular executable / shared library.
Unlike the Make implementation, this only passes links the helper
library (automatically through --coverage) when one of the object files
or static libraries being used actually has coverage enabled.
Host support is currently disabled, since we set -nodefaultlibs, which
prevents libclang_rt.profile-*.a from being picked up automatically.
Bug: 32749731
Test: NATIVE_COVERAGE=true COVERAGE_PATHS=system/core/libcutils m -j libbacktrace libutils tombstoned
$OUT/coverage/system/lib*/libcutils.gcnodir looks correct (self)
$OUT/coverage/system/lib*/libbacktrace.gcnodir looks correct (static)
$OUT/coverage/system/lib*/libutils.gcnodir doesn't exist (shared)
$OUT/coverage/system/bin/tombstoned.gcnodir looks correct (executable)
Test: NATIVE_COVERAGE=true COVERAGE_PATHS=external/libcxxabi m -j libc++
Confirm that $OUT/coverage/system/lib*/libc++.gcnodir looks correct (whole_static_libs)
Change-Id: I48aaa0ba8d76e50e9c2d1151421c0c6dc8ed79a9
-pie triggers a bug in glibc's linker when used with goma
(https://sourceware.org/bugzilla/show_bug.cgi?id=16381). Allow
the clang build to disable -pie for host modules through the
DISABLE_HOST_PIE environment variable so it can produce a toolchain
that works with goma.
Bug: 15814177
Bug: 34722791
Change-Id: Ic664a1b821aaeaf2bde14b0afa1a1975e31300cb
The linker uses libc++_static but needs to avoid the libc.a
dependency. It does this by setting `stl: "none"` and manually
linking libc++. This behavior matches make.
A better approach would probably be to generalize system_shared_libs
to system_libs and apply those to static executables the same way we
do for dynamic ones, but that's a patch for another day.
Test: make checkbuild
Bug: http://b/34740564
Change-Id: Ie9da0d49a453a220593e8ec2ee721e9af9378007
Pass a DepsContext that embeds android.BottomUpMutatorContext
instead of android.BaseContext so that dependency methods can
directly add dependencies.
Test: m -j
Change-Id: Id4c157975d3d6f03efd99785d217bef486a76139
Get the binary extension directly from the toolchain instead of
computing it from the target.
Test: mmma -j external/clang
Change-Id: Ib181bc60d0617ad50529791e7b9fc61268376819
Move symlink installation into the binaryDecorator so that it can
access the suffix and extension when installing symlinks. Also
consolidates the symlink and symlink_preferred_arch handling.
Test: m -j clang, ls -l out/host/windows-x86/bin/clang++.exe
Change-Id: I1204afb71fac87b276bd6b625b52ee21274855a0
Export a list of libraries in the VNDK, which is currently empty.
Take in Make's global BOARD_VNDK_VERSION, and use that as the SDK
version for modules that specify use_vndk: true.
Modules that use the vndk have some configuration as if they were
building against the NDK (the absence of globally defined headers), but
in other cases look like platform modules (using the platform libc++,
for now).
This change does not attempt to enforce any linking constraints, that
will come in a later patch.
Test: out/soong/build.ninja doesn't change
Change-Id: I3be206b67015ac5177b7eef4a451c579e3dc903f
Move some cc ctx.Host() / ctx.Device() checks over to using
ctx.toolchain().Bionic(). There will be more changes, these are just the
obvious ones dealing with host_ldlibs / crt / system libraries.
Bug: 31559095
Test: out/soong/build.ninja is identical
Change-Id: Ibba60483b4ab0e01f6996eb7d883120e4acc1830
We need to fall back to the old CRT objects since we aren't generating
those yet.
Test: Created a test module using "current", checked that it linked
the libs from current.
Bug: None
Change-Id: I5fe170d7b26154da8877672ac2acb7da0262fe38
For every file which we can run clang-tidy (C/C++ clang-built), we add a
new build node that depends on the object file (since clang-tidy does
not export a depfile), and is depended on by the link step. This is
better than how we're doing it in make, since calling tidy can be turned
on or off without needing to rebuild the object files.
This does not attempt to port WITH_TIDY_ONLY from Make, since the way
that it works is broken (due to the lack of a depfile).
Bug: 32244182
Test: WITH_TIDY=true mmma -j bionic/libc
Test: ./soong (Setting ClangTidy: true)
Change-Id: I40bbb5bb00d292d72bf1c293b93080b5f9f6d8ea
So that we can represent other files that get generated along with the
objects, like the gcno coverage information, and per-file clang-tidy
runs.
Test: Soong's build.ninja identical before/after
Change-Id: I5c553a153c436d5403549f62c73fe79c5f101779
Support prebuilt shared libraries as an initial proof-of-concept of
prebuilts. Future changes will support binaries and static libraries,
and the ability to select which to use based on something besides
blueprint properties.
Test: TestPrebuilts run during m -j
Change-Id: I6e84da667e9005ae11844bad01d25cbe4ced1ce3
Mutator registration is tightly coupled with the android package, move
all registration from the soong package to the android package.
Test: build.ninja identical
Change-Id: Ie183d0b52cc7431c9e05b231934d189208ef1efe
Track the primary architecture selected for each class based on the
module's multilib setting and the global config. Fixes building
binaries with multlib set to first and DevicePrefer32BitExecutables set,
and fixes symlinks to binaries with multilib set to prefer32.
Bug: 31452121
Test: mmma -j art HOST_PREFER_32_BIT=true
Change-Id: I75094df42f3273f6d613e4058eaa565957174c28
Allow CustomizeProperties to override the dynamic linker by setting a
DynamicLinker property. Needed by art, which overrides the linker for
device-on-host testing.
Change-Id: Ia4cb5a85b9a995d8138da33eb13543addf3b38cc
HOST_PREFER_32_BIT was used during the switch to 64-bit host tools to
keep the SDK building as 32-bit, but is never set any more.
Change-Id: I7c2db269d3f7fa1f4e0c022cbced789755a62d81
Add a symlink_preferred_arch property to binaries to allow compiling the
binary for multiple architectures and then creating a symlink to the
preferred archicture, for example dalvikvm32 and dalvikvm64, with
dalvikvm symlinked to dalvikvm64.
Test: mmma -j art/dalvikvm
Change-Id: Ied15f2be9d52c01006fe8ac207c175b78558eab1
