diff options
| author | Alex Light <allight@google.com> | 2018-07-11 15:32:33 -0700 |
|---|---|---|
| committer | Alex Light <allight@google.com> | 2018-07-12 13:45:19 -0700 |
| commit | 9491694c36f673149c4eed6918d4a58c328ef103 (patch) | |
| tree | 34abab8951f980ac9d955101fa8635fa34ecc033 /compiler/optimizing/stack_map_test.cc | |
| parent | dca6840055c5255ad05f7516b0227e44ef6310a2 (diff) | |
Move GCStress BacktraceHelper to libbacktrace
On host x86_64-linux the libgcc based BacktraceHelper could get
confused by our trampoline instrumentation. This replaces it with the
libbacktrace based stack walker that we use elsewhere in the runtime.
This has the advantage of being generally supported better.
For test 001-HelloWorld, this has an ~300% measured time cost as
compared to libgcc (from 9.063s to 28.587s). On the other hand it
seems better at actually differentiating distinct stack traces. Total
time spent doing gcs is about doubled (from 4107.91ms to 7036.97ms).
Total number of gcs is about doubled (from 175 to 320).
Test: ./test/testrunner/testrunner.py --host --gcstress --jit --64 -t 989
Test: ./test/testrunner/testrunner.py --host --gcstress
Change-Id: I745acb0ae92f312f86c3d31d81cfd179dbf9ae7e
Diffstat (limited to 'compiler/optimizing/stack_map_test.cc')
0 files changed, 0 insertions, 0 deletions