|author||Chitti Babu Theegala <email@example.com>||Fri Dec 16 02:13:28 2016 +0530|
|committer||Christopher Ferris <firstname.lastname@example.org>||Wed Jan 11 15:59:10 2017 -0800|
Fix streaming(memcpy) performance on Cortex-A7 Stream-mode detection for L1 in A7-core is failing for non cache-line-size (non 64 byte) aligned addresses. This leads to destination data getting cached unnecessarily. This A7 issue is confirmed by ARM This issue is solved by aligning destination address to 64 byte before entering the loop in memcpy routine. Though we get lower score for micro_bench memcpy when L1 cache is bypassed, it is desirable since it avoids unnecessary eviction of other process data from L1 which is good for overall system performance. Higher micro_bench memcpy numbers for < 64byte alignment shows good numbers but this is at the cost of L1 cache pollution. During memcpy/memset, unnecessary data is filled in L1 cache, this causes eviction of other process data from L1. For example during msmset(0), L1 cache gets filled with 0s which should be avoided. Additionally, there is another issue with cortex A7 that impacts performance for all alignments / all Android Wear versions: Store Buffer on A7 is 32 byte which limits the 32-byte back to back stores. In the current implementation back to back 32bytes writes is causing CPU stalls. This issue can be solved by interleaved Loads and Stores. This helps in avoiding CPU stalls during memcpy by utilizing efficiently the A7 internal load and store buffers. Bug: 31917556 Change-Id: Ie5f12f2bb5d86f627686730416279057e4f5f6d0 (cherry picked from commit cbfdc7f9054e0bc8071aaf3a70afc00273a8a869)
The C library. Stuff like
The math library. Traditionally Unix systems kept stuff like
cos(3) in a separate library to save space in the days before shared libraries.
The dynamic linker interface library. This is actually just a bunch of stubs that the dynamic linker replaces with pointers to its own implementation at runtime. This is where stuff like
The C++ ABI support functions. The C++ compiler doesn't know how to implement thread-safe static initialization and the like, so it just calls functions that are supplied by the system. Stuff like
__cxa_pure_virtual live here.
The dynamic linker. When you run a dynamically-linked executable, its ELF file has a
DT_INTERP entry that says “use the following program to start me”. On Android, that‘s either
linker64 (depending on whether it’s a 32-bit or 64-bit executable). It's responsible for loading the ELF executable into memory and resolving references to symbols (so that when your code tries to jump to
fopen(3), say, it lands in the right place).
tests/ directory contains unit tests. Roughly arranged as one file per publicly-exported header file.
benchmarks/ directory contains benchmarks.
Adding a system call usually involves:
As mentioned above, this is currently a two-step process:
This is fully automated (and these days handled by the libcore team, because they own icu, and that needs to be updated in sync with bionic):
If you make a change that is likely to have a wide effect on the tree (such as a libc header change), you should run
make checkbuild. A regular
make will not build the entire tree; just the minimum number of projects that are required for the device. Tests, additional developer tools, and various other modules will not be built. Note that
make checkbuild will not be complete either, as
make tests covers a few additional modules, but generally speaking
make checkbuild is enough.
The tests are all built from the tests/ directory.
$ mma $ adb remount $ adb sync $ adb shell /data/nativetest/bionic-unit-tests/bionic-unit-tests32 $ adb shell \ /data/nativetest/bionic-unit-tests-static/bionic-unit-tests-static32 # Only for 64-bit targets $ adb shell /data/nativetest64/bionic-unit-tests/bionic-unit-tests64 $ adb shell \ /data/nativetest64/bionic-unit-tests-static/bionic-unit-tests-static64
The host tests require that you have
lunched either an x86 or x86_64 target.
$ mma $ mm bionic-unit-tests-run-on-host32 $ mm bionic-unit-tests-run-on-host64 # For 64-bit *targets* only.
As a way to check that our tests do in fact test the correct behavior (and not just the behavior we think is correct), it is possible to run the tests against the host's glibc. The executables are already in your path.
$ mma $ bionic-unit-tests-glibc32 $ bionic-unit-tests-glibc64
For either host or target coverage, you must first:
$ export NATIVE_COVERAGE=true
$ mma $ adb sync $ adb shell \ GCOV_PREFIX=/data/local/tmp/gcov \ GCOV_PREFIX_STRIP=`echo $ANDROID_BUILD_TOP | grep -o / | wc -l` \ /data/nativetest/bionic-unit-tests/bionic-unit-tests32 $ acov
acov will pull all coverage information from the device, push it to the right directories, run
lcov, and open the coverage report in your browser.
First, build and run the host tests as usual (see above).
$ croot $ lcov -c -d $ANDROID_PRODUCT_OUT -o coverage.info $ genhtml -o covreport coverage.info # or lcov --list coverage.info
The coverage report is now available at
Bionic's test runner will run each test in its own process by default to prevent tests failures from impacting other tests. This also has the added benefit of running them in parallel, so they are much faster.
However, this also makes it difficult to run the tests under GDB. To prevent each test from being forked, run the tests with the flag
This probably belongs in the NDK documentation rather than here, but these are the known ABI bugs in the 32-bit ABI:
time_t is 32-bit. http://b/5819737. In the 64-bit ABI, time_t is 64-bit.
off_t is 32-bit. There is
off64_t, and in newer releases there is almost-complete support for
_FILE_OFFSET_BITS. Unfortunately our stdio implementation uses 32-bit offsets and -- worse -- function pointers to functions that use 32-bit offsets, so there's no good way to implement the last few pieces http://b/24807045. In the 64-bit ABI, off_t is off64_t.
sigset_t is too small on ARM and x86 (but correct on MIPS), so support for real-time signals is broken. http://b/5828899 In the 64-bit ABI,
sigset_t is the correct size for every architecture.