diff options
| author | 2018-09-08 18:02:36 +0100 | |
|---|---|---|
| committer | 2018-09-27 20:21:38 +0100 | |
| commit | ef79aa317437d5ba2bc2e3dd197cd893a57b0d19 (patch) | |
| tree | 9f580f5e337c1a9efb61b09ada50ef378c6d147e /runtime/interpreter/interpreter_cache.h | |
| parent | 233b572a940431a94a1790750afdceab2d6f4fde (diff) | |
Use the small thread-local cache for mterp field accesses.
This reduces the overhead of non-quickened code from 10% to 7.5%.
(measured on golem benchmarks for arm64)
Test: ./art/test.py -b -r --interpreter
Change-Id: Icce9183eb60c62ac30a0c6ff57e32c796c807f03
Diffstat (limited to 'runtime/interpreter/interpreter_cache.h')
| -rw-r--r-- | runtime/interpreter/interpreter_cache.h | 4 |
1 files changed, 4 insertions, 0 deletions
diff --git a/runtime/interpreter/interpreter_cache.h b/runtime/interpreter/interpreter_cache.h index c25222eff3..b4966fd615 100644 --- a/runtime/interpreter/interpreter_cache.h +++ b/runtime/interpreter/interpreter_cache.h @@ -35,6 +35,10 @@ class Thread; // All operations must be done from the owning thread, // or at a point when the owning thread is suspended. // +// The values stored for opcodes in the cache currently are: +// iget/iput: The field offset. The field must be non-volatile. +// sget/sput: The ArtField* pointer. The field must be non-volitile. +// // Aligned to 16-bytes to make it easier to get the address of the cache // from assembly (it ensures that the offset is valid immediate value). class ALIGNED(16) InterpreterCache { |