Lower JIT threshold when running JIT mode

When running test cases with ART JIT mode with default auto calibration,
the performance actually replies on whether performance critical functions
are JIT compiled during the auto calibration phase (warm up phase).

Currently, the default thresholds that impacts benchmark JIT mode:
- In art-testing framework, default minimum calibration time is 50ms.
- In ART JIT, default threshold to trigger jit is 10000 (hotness count).

For some of the test cases, for example benchmarks/stanford/Oscar,
above thresholds are not enough to trigger JIT compilation for some critical functions,
which results in much slower scores than AOT mode.

This patch proposes to change jit threshold when running jit mode, to solve this problem.

Change-Id: I8178969b6fc48524fa47682ffdf3e9921d637b12

1 files changed, 4 insertions, 1 deletions

diff --git a/tools/benchmarks/run.py b/tools/benchmarks/run.py
index 6d0cc6b..15583fd 100755
--- a/tools/benchmarks/run.py
+++ b/tools/benchmarks/run.py
@@ -113,7 +113,10 @@ def RunBenchADB(mode, compiler_mode, android_root, auto_calibrate, apk,
         # boot-image.
         dalvikvm_options += ' -Ximage-compiler-option %s' % opt
     if compiler_mode == 'jit':
-        dalvikvm_options += ' -Xusejit:true -Xnodex2oat'
+        # For JIT mode benchmarking, enable following options to make sure:
+        # - No dex2oat compilation triggered.
+        # - Performance critical funtions are JIT compiled as soon as possible.
+        dalvikvm_options += ' -Xusejit:true -Xnodex2oat -Xjitthreshold:100'
 
     command = 'cd {workdir} && ' + \
         ' '.join([environment_config, dalvikvm,