Background: `min_sdk_version` of a crt object is the oldest SDK version
that the crt object supports. When it's set to for example 16, Soong
creates a number of versioned variants of the crt object starting from
16 to the latest version. The variant for version X is provided to NDK
clients having `min_sdk_version` set to X.
Problem: all versioned variants of a crt object were built with `-target
<arch>-linux-android16`. Therefore they all have been with `#define __ANDROID_API__
16`. This is because the mutated variants still have the same
min_sdk_version property and the clang triple follows min_sdk_version,
not sdk_version. This is too conservative and against our intention to
provide the latest crt object that matches with the min_sdk_version of
the client.
In the other hand, the platform(non-sdk) variant of the crt object
doesn't have such a problem. min_sdk_version is completely ignored.
However, this is a bug and will be fixed by aosp/1640364. As a side
effect of the fixing, the platform variant will begin to built with a
very old __ANDROID_API__ which unnecessarily turns off the new platform
features like the TLS segment over-alignment.
This change fixes the problems:
* For the versioned variants of crt objects, we set both
`min_sdk_version` and `sdk_versio` to the version that the variant is
created for.
* For the platform variant of crt objects, `min_sdk_version` is force
reset to "current".
Bug: 183191008
Test: m
Change-Id: I8c9d0fcea816de8cd1532dac4a47eee4f726c037
Background:
When compiling cc_* modules, min_sdk_version determines the "version"
part of the clang triple: -target <arch>-linux-android<version>. The
version part is used to make sure that calls to the APIs that are added
after the version are guarded with a runtime check (i.e.
__builtin_available).
Previously, min_sdk_version was used as the version part only when the
cc_* module is unbundled (e.g. built for an APEX or built with SDK). In
other words, min_sdk_version has been ignored for the platform variants.
They were built with the version number 10000.
This was problematic for Mainline module tests. Since they are neither
part of an APEX nor built with SDK (because they need to have access to
some of the module-only APIs), they are built with the version number
10000. As a side effect, __builtin_available macro are expanded to 1 at
build time - because 10000 is higher than any API versions. When such a
test built in the latest platform source tree runs on a device running
an old platform, it tries to call an API that might not be available on
the platform and experience a crash, due to the lack of the runtime
check.
This change fixes the problem by using min_sdk_version as the "version"
part of the clang triple, regardless of the module's variant. Then
__builtin_available macro in the tests doesn't expand to 1, but to a
function call to the libclang_rt.builtin library which checks the system
property ro.build.version.sdk.
Bug: N/A
Test: run resolv_stress_test
Change-Id: I88f64c5a35f1b5276c3350e177b116932011a940
Java build components invariably require android build components such
as the arch mutator (to create the common variant) so include them
together.
Bug: 181070625
Test: m nothing
Change-Id: I435b09a34c48c67b890368b37b8da27e8de97c9a
This allows the testCustomizers to be switched to FixturePreparers
incrementally rather than in one go.
Bug: 181070625
Test: m nothing
Change-Id: Idd9d2e28abf9b17fc46b5566ab8d3affa330287e
With clang r377782, llvm binutils implements all the necessary
funtionaility for strip.sh. We can finally get rid of all the fallback
GNU binutils usage.
Test: m
Bug: 141010852
Bug: 135627985
Change-Id: I110f6028dab7f599decf59a5cb1b927b35e11857
Run lint actions in sbox with RuleBuilder.SandboxInputs. This
copies all input files into the sandbox, which prevents the lint
tool from finding nearby source files that were not presented to it.
Using SandboxInputs requires use of PathForInput or PathForOutput
anywhere a path is used outside of the RuleBuilderCommand methods
that take paths so that they can be translated to the paths that
will be used in the sandbox.
Bug: 181681346
Test: lint_test.go
Test: m lint-check dist
Test: m USE_RBE=true RBE_LINT=true lint-check dist
Test: m USE_RBE=true RBE_LINT=true RBE_LINT_EXEC_STRATEGY=remote lint-check dist
Change-Id: Iab4e09d961891ef182643583d4d456e413bc5e39
Allow passing a remoteexec.REParams to RuleBuilder to configure it to
run the rule remotely through RBE. Requires the rule to use
SandboxInputs, which ensures that RuleBuilder is aware of all of the
inputs and outputs of the rule.
Running sbox in RBE initially seems unnecessary, as RBE is already a
good sandbox, but reproxy can execute RBE actions locally when configured
for local execution, local fallback or racing. Using sbox in RBE ensures
that these local actions are also sandboxed, giving consistent results
between directly executed actions, local RBE actions, and remote RBE
actions.
Bug: 182612695
Test: manual
Change-Id: Icf2f24dde8dee833eb680ba22566a8e1c0143b15
Remove the references to the android package in remotexec so that
the android package can reference the remoteexec package. This
will allow RuleBuilder to integrate directly with remoteexec.
Bug: 182612695
Test: m checkbuild
Change-Id: I15be5ef126d8aacbd605518638f341daf6f31bb3
When RuleBuilder.SandboxInputs() is called configure sbox to copy
all the input files into the sandbox directory and then change the
working directory there when running the command.
Copying input files into the sandbox directory gets tricky when
the input file is the output file from another rule, and could
be at an arbitrary, possibly absolute path based on the value
of OUT_DIR. They will need to be copied to a directory in the
sandbox using the path relative to OUT_DIR.
RSP files need special handling, they need to both be copied into
the sandbox as an input, rewritten to contain paths as seen in the
sandbox, and references to them on the command line need to use
sandbox paths.
Bug: 182612695
Test: rule_builder_test.go
Change-Id: Ic0db961961b186e4ed9b76246881e3f04971825c
Jiyong Park