Reland "ARM64: Enable implicit suspend checks after CCCR."
This reverts commit 2de08ed613903526279512c7f6114b52cbf5e993.
Reason for revert: Reland with app startup performance
impact mitigation. Replace `MADV_FREE` with `MADV_DONTNEED`
in `Thread::MadviseAwayAlternateSignalStack()`.
Some regressions are expected but this change should provide
a good trade-off overall. The benefits for memory use thanks
to reduced code size can be seen in b/263509016 that tracked
the system server RSS regression when we disabled implicit
suspend checks the last time.
The startup regression b/263319992 has been partially
mitigated by
https://android-review.googlesource.com/2375891
and is being partially mitigated by using the `MADV_FREE`,
see above. Based on some test runs, the `MADV_FREE` roughly
halves the regression we would incur with `MADV_DONTNEED`.
On the flip side, the `MADV_FREE` does not immediatelly
reclaim the committed memory of the alternate signal stacks,
so we may see regressions in some memory benchmarks. These
regressions would not be true regressions, however, because
that memory is indeed reclaimable under memory pressure, so
it should be considered as any other temporary allocation.
Some Golem micro-benchmarks show significant improvements on
Odroid-C2 while showing significant regressions on Raspberry
Pi 4 on the very same benchmark. The implicit suspend check
is a single LDR while the explicit check is LDR+TST+BNE and
it seems reasonable to expect better performance on in-order
CPUs due to fewer instructions being executed. However,
out-of-order CPUs may prefer the LDR for explicit checks as
it uses a "constant" address over the LDR for implicit check
which uses the previously loaded value as the the address
for the next check, presumably preventing scheduling these
LDRs closely together, for example in a tight loop.
Test: run-gtests.sh
Test: testrunner.py --target --64 --optimizing --jit
Bug: 38383823
Bug: 215310343
Change-Id: I274a5417293e6d926579ec97190085387137d7c2
5 files changed