RFR 8201785: Provide default implementations for ModRefBarrierSetAssembler::oop_store_at for all platforms (original) (raw)

Aleksey Shipilev shade at redhat.com
Wed Apr 18 10:36:00 UTC 2018

On 04/18/2018 12:30 PM, Erik Österlund wrote:

On 2018-04-18 12:22, Aleksey Shipilev wrote:

On 04/18/2018 12:12 PM, Erik Österlund wrote:

Why are you instantiating a ModRefBarrierSetAssembler as opposed to a BarrierSetAssembler? ModRef is meant as an abstract class for GCs that have a notion of write barriers. I don't believe your GC has such a notion. To answer the Epsilon-related question: in order to reduce code duplication, current implementation tries to delegate to existing barrier sets, where possible. ModRefBSAssembler is nearly there for interpreter GC barriers, because even though its subclasses implement write-barrier-bearing BS, its default implementation may do nothing for write-barriers -- and it does nothing for AArch64 and PPC, which is convenient. The same should go for BarrierSetAssembler - you should be able to just instantiate a BarrierSetAssembler if you don't want to make your own assembler that inherits from BarrierSetAssembler.

Oh! I overlooked that BarrierSetAssembler is not abstract! Thanks, it seems to work :)

At very least, it makes ModRefBSAssembler consistently concrete in all arches. The alternative may be the reverse: make ModRefBSAssembler consistently abstract in all arches, but purging AArch64 and PPC default implementations. I think ModRefBarrierSetAssembler should be abstract on all architectures. It is really only a helper for GCs with a notion of write barriers.

Okay, let me withdraw this patch, and propose the reverse: make it abstract all the way.

Thanks, -Aleksey

More information about the hotspot-dev mailing list