Reorder interpreter jump targets for performance #4133
Conversation
|
@pdbain-ibm Can you add the signed-off by tag to your commit? Please also add context on how this was found, the performance difference seen with & without it, etc. I understand there may be details that can't be shared due to the test cases used, but that shouldn't prevent sharing the improvement seen by making this change and the platforms its been tested on. |
pdbain-ibm
force-pushed
the
reorder_iincw
branch
3 times, most recently
from
c12f139 to
b8dd888
Compare
|
@DanHeidinga Done. Let me know if you want more details. |
|
@pdbain-ibm Has this change been tested on any other platforms? |
|
Jenkins test sanity plinux,zlinux,win jdk8 |
|
@DanHeidinga do you mean for performance? If so, no. |
Make the order of the JUMP_TARGETs, specifically the custom "iincw" bytecode, match the bytecode numbers better. This causes an approximately 2% improvement in Liberty startup time with -Xint on Linux PPC. Signed-off-by: Peter Bain <peter_bain@ca.ibm.com>
pdbain-ibm
force-pushed
the
reorder_iincw
branch
from
b8dd888 to
21fa283
Compare
|
Jenkins test sanity plinux,zlinux,win jdk8 |
|
@pdbain-ibm Did you get the performance results with this change from the other platforms? Would be good to record that here if possible |
|
Our internal performance team reports:
|
Make the order of the JUMP_TARGETs match the bytecode numbers
better.
This causes an approximately 2% improvement in Liberty startup time with -Xint
on Linux PPC.
Signed-off-by: Peter Bain peter_bain@ca.ibm.com