[foreign-memaccess+abi] RFR: 8274285: By reference parameters should be kept alive during downcalls [v2]

[Jorn Vernee]

On Fri, 24 Sep 2021 13:33:28 GMT, Maurizio Cimadamore <mcimadamore at openjdk.org> wrote:

>> This patch adds an adapter around the downcall method handle which keeps alive all by-reference arguments by acquiring/releasing their associated scopes.
>> 
>> I've implemented the adapted in a separate method in `SharedUtils`, because `SharedUtils::wrapWithAllocator` is relatively complex, and has mostly to do with creating a scope and closing it after the call.
>> Also, `wrapWithAllocator` is only called when specialized method handles are active, which means adding logic there doesn't fix the programmable fallback case.
>> 
>> Since we plan to do more implementation restructuring in this area, I preferred to keep things simple for now.
>> 
>> Also note that the acquire/release logic isn't too smart - if multiple arguments backed by the same scoped are passed, we do separate acquire/release for each of them. Again, I didn't want to spend too many cycles given that the implementation of this area might change, so my aim for now was to make things safe; we can improve efficiency later.
>> 
>> Here are some benchmarks, for now, which compare the difference between passing a struct by reference using the struct's address (e.g. calling `MemorySegment::address` before the downcall), vs. passing the struct directly. In the first case, the scope is the global scope, so acquiring/relesing is a no-op. In the latter case we have a real scope, which we need to acquire and release.
>> 
>> 
>> Benchmark                                                           Mode  Cnt   Score   Error  Units
>> CallOverheadConstant.panama_identity_memory_address_confined        avgt   30  13.094 ? 0.220  ns/op
>> CallOverheadConstant.panama_identity_memory_address_confined_3      avgt   30  16.051 ? 0.230  ns/op
>> CallOverheadConstant.panama_identity_memory_address_shared          avgt   30  13.287 ? 0.172  ns/op
>> CallOverheadConstant.panama_identity_memory_address_shared_3        avgt   30  15.812 ? 0.365  ns/op
>> 
>> Benchmark
>> CallOverheadConstant.panama_identity_struct_ref_confined    avgt   30  11.544 ? 0.170  ns/op
>> CallOverheadConstant.panama_identity_struct_ref_confined_3  avgt   30  12.213 ? 0.228  ns/op
>> CallOverheadConstant.panama_identity_struct_ref_shared      avgt   30  17.214 ? 0.560  ns/op
>> CallOverheadConstant.panama_identity_struct_ref_shared_3    avgt   30  33.132 ? 0.934  ns/op
>> 
>> 
>> As it can be seen, in the confined case the numbers are good. In the shared case there's more overhead - that's the price for safety. We will be able to make this faster (by only acquiring scopes which have not already been acquired) - but note that users can always opt out of safety by projecting the `Addressable` down to a raw memory address, which would remove all the costs.
>
> Maurizio Cimadamore has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Rename benchamrk method to avoid confusion

Marked as reviewed by jvernee (Committer).

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

PR: https://git.openjdk.java.net/panama-foreign/pull/581