On Thu, 25 Jun 2026 09:43:06 GMT, Joel Sikström <[email protected]> wrote:
> Hello, > > This is (yet another) cleanup to remove redundant memory barriers after the > new ValuePayload abstraction API in the runtime. The ValuePayload API's > (including FlatValuePayload::read) does the necessary calls to order stores > of the copy before the oop is published (OrderAccess::storestore), making the > platform-specific memory barriers redundant. > > I don't have hardware access to RISCV or PPC, but looking at the runtime > implementations of OrderAccess::storestore, they achieve the same thing as > the platform-specific memory barriers that I suggest we remove in this PR. > GHA will show that this compiles at least. > > `membar(Assembler::StoreStore | Assembler::LoadLoad);` and > `OrderAccess::storestore()` both boil down to `lwsync`. > > The C1 call to `load_flat_array` already omits the memory barrier, so we > don't need to remove it here. > > Testing: > * Oracle's tier1-3 linux-aarch64 > * GHA > > --------- > - [x] I confirm that I make this contribution in accordance with the [OpenJDK > Interim AI Policy](https://openjdk.org/legal/ai). This makes sense. Thanks for taking care of all platforms! I prefer a small change for PPC64. src/hotspot/cpu/ppc/templateTable_ppc_64.cpp line 3420: > 3418: __ read_flat_field(Rcache, R17_tos); > 3419: __ verify_oop(R17_tos); > 3420: // memory barrier in read_flat_field I think we should better add ```C++ __ twi_0(R17_tos); __ isync(); to be on the safe side. ------------- PR Review: https://git.openjdk.org/valhalla/pull/2581#pullrequestreview-4570240969 PR Review Comment: https://git.openjdk.org/valhalla/pull/2581#discussion_r3473815894
