RFR (XS): 8218192: Remove copy constructor for MemRegion
kim.barrett at oracle.com
Tue Feb 5 18:42:05 UTC 2019
> On Feb 5, 2019, at 11:25 AM, Man Cao <manc at google.com> wrote:
> Hi all,
> Could I have reviews for this small patch?
> Webrev: https://cr.openjdk.java.net/~manc/8218192/webrev.00/
> RFE: https://bugs.openjdk.java.net/browse/JDK-8218192
> This will improve G1 write barrier performance for jython in DaCapo, when building HotSpot with LLVM. More Background and performance numbers are included in the JBS.
The only thing I might suggest is adding a comment that the destructor
and copy constructor must be trivial, to support optimization of pass
by value. I don't need a new webrev if you add such a comment.
The explicit copy constructor is clearly interfering with the expected
optimization on Linux-x64, and likely other platforms where I didn't
examine the generated assembly. The SysV ABI describes passing such
an object in a pair of registers for the members, but only when the
destructor and copy constructor are trivial. (An older version of the
ABI required the class to be POD, which is unnecessarily restrictive;
gcc through 4.9 (I think) used that.)
There are a handful of places that are passing MemRegions by const-ref
rather than by value that probably ought to be looked at, but that can
be done separately.
I did some performance testing of this change as well. Seemed pretty
much in the noise.
More information about the hotspot-gc-dev