Running the benchmark from JDK-8325269, we found that when we use -XX:-UseStoreStoreForCtor, -XX:-AlwaysMergeDMB significantly improved the the throughput. This experiment shows that "dmb ishst + dmb ishld" is faster than "dmb ish" on Ampere CPU. Thus, we prefer to turn off AlwaysMergeDMB for Ampere CPU by default.
- relates to
-
JDK-8355233 Add a DMB related benchmark
-
- Resolved
-
-
JDK-8325269 Quadruple dmbs could be emitted after JDK-8324186
-
- Closed
-
-
-
- Resolved
-
- links to
-
Commit(master)
openjdk/jdk/128f2d1c
-
Review(master)
openjdk/jdk/24652