Hi,
thanks Mandy and Jim for reviewing!
However, I've found enough evidence now that we should this
one step further and eliminating the allocation in
BasicImageReader::findLocation(String, String), which completely
gets rid of the regressions we're seeing:
http://cr.openjdk.java.net/~redestad/8175561/jdk.02/
http://cr.openjdk.java.net/~redestad/8175561/jdk.01.02.diff/
Working through the details on how to do this it became clear
that stringing together ImageStringsReader.hashCode - as done in
ImageLocationWriter - is likely a bug in the making since it's not
guaranteed to give equivalent results to applying the same method
to the whole string. I took the liberty to fix this, and ran the final
patch through several JDK and HotSpot test lists without any surprises.
Thanks!
/Claes
On 02/23/2017 10:57 PM, Claes Redestad wrote:
Hi,
various resource encapsulation changes in 9+148 meant an uptick in
footprint and startup times for certain applications.
While some of this regression is hard to avoid[1] (opening readers,
touching more memory mapped pages etc), a large part is due to simple
java allocation churn, some of which can be optimized away by reducing
the number of objects we create when scanning modules for resources.
Bug: https://bugs.openjdk.java.net/browse/JDK-8175561
Webrev: http://cr.openjdk.java.net/~redestad/8175561/jdk.01/
Thanks!
/Claes
[1] Things we could do in the future to improve further include adding
an index to the jimage (too late for 9) or marking modules with no
non-encapsulated resources as such so that they can be skipped.
