Rewrite thread root flip synchronization.

Mark pending "flip function" with a `ThreadFlag` for faster
JNI transitions.

Use two more `ThreadFlag`s to fix potential race conditions.
Some checkpoints were previously running the flip function
on behalf of a suspended thread because they relied on the
thread roots being flipped, but doing that without any
synchronization meant that they could race and one could
have executed its own code while another was still running
the flip function. Other checkpoints that were peforming
a stack walk did not run the flip function at all, so they
could have seen from-space references. We now check for a
pending or running flip function at the start of the
`Thread::RunCheckPointFunction()` and proceed only after
it has completed; holding a mutator lock for the duration
of the whole function prevents a new flip function from
being installed until the checkpoint finishes.

Golem results for art-opt-cc (higher is better):
linux-ia32                       before after
NativeDowncallStaticNormal       46.581 46.813 (+0.4980%)
NativeDowncallStaticNormal6      42.247 42.268 (+0.0497%)
NativeDowncallStaticNormalRefs6  40.918 41.355 (+1.068%)
NativeDowncallVirtualNormal      46.292 46.361 (+0.1496%)
NativeDowncallVirtualNormal6     41.791 41.791 (0%)
NativeDowncallVirtualNormalRefs6 40.500 40.500 (0%)
linux-x64                        before after
NativeDowncallStaticNormal       44.169 43.956 (-0.4815%)
NativeDowncallStaticNormal6      43.198 43.198 (0%)
NativeDowncallStaticNormalRefs6  38.481 38.481 (0%)
NativeDowncallVirtualNormal      43.672 43.672 (0%)
NativeDowncallVirtualNormal6     42.247 42.268 (+0.0479%)
NativeDowncallVirtualNormalRefs6 41.355 41.355 (0%)
linux-armv7                      before after
NativeDowncallStaticNormal       9.9701 10.443 (+4.739%)
NativeDowncallStaticNormal6      9.2457 9.6525 (+4.400%)
NativeDowncallStaticNormalRefs6  8.3868 8.7209 (+3.984%)
NativeDowncallVirtualNormal      9.8377 10.304 (+4.742%)
NativeDowncallVirtualNormal6     9.3596 9.7752 (+4.440%)
NativeDowncallVirtualNormalRefs6 8.4367 8.7719 (+3.973%)
linux-armv8                      before after
NativeDowncallStaticNormal       9.8571 10.685 (+8.397%)
NativeDowncallStaticNormal6      9.4905 10.249 (+7.991%)
NativeDowncallStaticNormalRefs6  8.6705 9.3000 (+7.261%)
NativeDowncallVirtualNormal      9.3183 10.053 (+7.881%)
NativeDowncallVirtualNormal6     9.2638 9.9850 (+7.786%)
NativeDowncallVirtualNormalRefs6 8.2967 8.8714 (+6.926%)
(The x86 and x86-64 differences seem to be lost in noise.)

Test: m test-art-host-gtest
Test: testrunner.py --host --optimizing
Bug: 172332525
Change-Id: I9c2227142010f7fe6ecf07e92273bc65d728c5c6

0 files changed, 0 insertions, 0 deletions