RFR: 8301869: Regression ~14% in J2dBench-bimg_misc-* in 21-b5 only on linux-aarch64

[Erik Joelsson]

On Mon, 27 Feb 2023 11:49:27 GMT, Jayathirth D V <jdv at openjdk.org> wrote:

> Under https://bugs.openjdk.org/browse/JDK-8264846 we moved to -O3 level of gcc optimizations from -O1 level for libawt build. This improved our J2DBench performance numbers in some options considerably.
> 
> Recent changes done under https://bugs.openjdk.org/browse/JDK-8299337 causes difference in generated code by gcc and this is resulting in performance regression for bimg_misc-* J2DBench options in our performance servers. Under https://bugs.openjdk.org/browse/JDK-8299337 we have just removed unused variables and it is a cleanup task.
> 
> We can force gcc to generate position independent code by using -fpic option.Also i have removed -fgcse-after-reload option for gcc, because this is by default covered under -O3 level of optimization introduced under https://bugs.openjdk.org/browse/JDK-8264846.
> 
> With this change bimg_misc-* J2DBench option performance regression is resolved and there are no regression in other options of J2DBench or SwingMark and it is verified in our performance servers.

make/modules/java.desktop/lib/Awt2dLibraries.gmk line 132:

> 130:   # that any cleanup of code related to AWT will not change the
> 131:   # generated code and cause performance issues in J2DBench
> 132:   LIBAWT_CFLAGS += -fpic

We compile with `-fPIC` as a standard flag for position independent code (I just verified that LIBAWT also gets this flag today). My understanding is that `-fpic` may result in smaller/faster code. Overriding with `-fpic` seems like its helping here, but the comment needs to be fixed to explain the actual difference and reason.

-------------

PR: https://git.openjdk.org/jdk/pull/12761