[foreign-memaccess+abi] RFR: 8311594: Avoid GlobalSession liveness check

[Maurizio Cimadamore]

On Fri, 7 Jul 2023 18:14:41 GMT, Jorn Vernee <jvernee at openjdk.org> wrote:

> This patch overrides `checkValidStateRaw` in GlobalSession, to do nothing. The current `checkValidStateRaw` in MemorySessionImpl checks the owner thread, and liveness, which are both not needed for the global session.
> 
> Since the state field is mutable, I found that the JIT can not fold away the liveness check. This has a very marginal effect on performance, but since the fix is also very non-intrusive, I think it's okay to do.
> 
> I've added 2 new benchmarks that I've used to measure the difference. One measures the performance of MS::copy, and the other measures MS::get. Both use a special `ALL` segment which spans `0` to `Long.MAX_VALUE`, which means that the target address can be used as an offset when doing the accesses. This helps the bounds checks being eliminated, and makes the result of overriding `checkValidStateRaw` in GlobalSession easier to see.
> 
> Here are some of the interesting numbers:
> 
> Before:
> 
> 
> Benchmark                                  Mode  Cnt  Score   Error  Units
> MemorySegmentCopyALL.panama_ALL_to_ALL     avgt   30  7.788 ± 0.029  ns/op
> MemorySegmentCopyALL.unsafe_array_to_addr  avgt   30  7.543 ± 0.018  ns/op
> 
> Benchmark                               Mode  Cnt  Score   Error  Units
> MemorySegmentGet.segment_ALL_unaligned  avgt   30  0.486 ± 0.002  ns/op
> MemorySegmentGet.unsafe                 avgt   30  0.398 ± 0.003  ns/op
> 
> 
> After:
> 
> 
> Benchmark                                  Mode  Cnt  Score   Error  Units
> MemorySegmentCopyALL.panama_ALL_to_ALL     avgt   30  7.571 ± 0.017  ns/op
> MemorySegmentCopyALL.unsafe_array_to_addr  avgt   30  7.550 ± 0.019  ns/op
> 
> Benchmark                               Mode  Cnt  Score   Error  Units
> MemorySegmentGet.segment_ALL_unaligned  avgt   30  0.409 ± 0.002  ns/op
> MemorySegmentGet.unsafe                 avgt   30  0.397 ± 0.003  ns/op

src/java.base/share/classes/jdk/internal/foreign/GlobalSession.java line 74:

> 72:     @ForceInline
> 73:     public void checkValidStateRaw() {
> 74:         // do nothing, avoid liveness check

This looks ok, although (see comments on MemorySessionImpl::checkValidStateRaw) sometimes having too many versions of this method can cause profile pollution issues (which is why we consolidated the impl at some point). I suggest trying other benchmarks, such as LoopOverPollutedXYZ

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

PR Review Comment: https://git.openjdk.org/panama-foreign/pull/844#discussion_r1257963606