Integrated: 8252990: Intrinsify Unsafe.storeStoreFence

[Aleksey Shipilev]

On Wed, 27 Oct 2021 11:53:47 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

> `Unsafe.storeStoreFence` currently delegates to stronger `Unsafe.storeFence`. We can teach compilers to map this directly to already existing rules that handle `MemBarStoreStore`. Like explicit `LoadFence`/`StoreFence`, we introduce the special node to differentiate explicit fence and implicit store-store barriers. `storeStoreFence` is usually used to simulate safe `final`-field like constructions in special JDK classes, like `ConstantCallSite` and friends.
> 
> Motivational performance difference on benchmarks from JDK-8276054 on ARM32 (Raspberry Pi 4):
> 
> 
> Benchmark                      Mode  Cnt   Score    Error  Units
> Multiple.plain                 avgt    3   2.669 ±  0.004  ns/op
> Multiple.release               avgt    3  16.688 ±  0.057  ns/op
> Multiple.storeStore            avgt    3  14.021 ±  0.144  ns/op // Better
> 
> MultipleWithLoads.plain        avgt    3   4.672 ±  0.053  ns/op
> MultipleWithLoads.release      avgt    3  16.689 ±  0.044  ns/op
> MultipleWithLoads.storeStore   avgt    3  14.012 ±  0.010  ns/op // Better
> 
> MultipleWithStores.plain       avgt    3  14.687 ±  0.009  ns/op
> MultipleWithStores.release     avgt    3  45.393 ±  0.192  ns/op
> MultipleWithStores.storeStore  avgt    3  38.048 ±  0.033  ns/op // Better
> 
> Publishing.plain               avgt    3  27.079 ±  0.201  ns/op
> Publishing.release             avgt    3  27.088 ±  0.241  ns/op
> Publishing.storeStore          avgt    3  27.009 ±  0.259  ns/op // Within error, hidden by allocation
> 
> Single.plain                   avgt    3   2.670 ± 0.002  ns/op
> Single.releaseFence            avgt    3   6.675 ± 0.001  ns/op
> Single.storeStoreFence         avgt    3   8.012 ± 0.027  ns/op  // Worse, seems to be ARM32 implementation artifact
> 
> 
> The same thing on AArch64 (Raspberry Pi 3):
> 
> 
> Benchmark                      Mode  Cnt   Score   Error  Units
> 
> Multiple.plain                 avgt    3   5.914 ± 0.115  ns/op
> Multiple.release               avgt    3  10.149 ± 0.059  ns/op
> Multiple.storeStore            avgt    3   6.757 ± 0.138  ns/op // Better
> 
> MultipleWithLoads.plain        avgt    3  11.849 ± 0.331  ns/op
> MultipleWithLoads.release      avgt    3  35.565 ± 1.144  ns/op
> MultipleWithLoads.storeStore   avgt    3  19.441 ± 0.471  ns/op // Better
> 
> MultipleWithStores.plain       avgt    3   5.920 ± 0.213  ns/op
> MultipleWithStores.release     avgt    3  20.286 ± 0.347  ns/op
> MultipleWithStores.storeStore  avgt    3  12.686 ± 0.230  ns/op // Better
> 
> Publishing.plain               avgt    3  22.261 ± 1.630  ns/op
> Publishing.release             avgt    3  22.269 ± 0.576  ns/op
> Publishing.storeStore          avgt    3  17.464 ± 0.397  ns/op // Better
> 
> Single.plain                   avgt    3   5.916 ± 0.063  ns/op
> Single.release                 avgt    3  10.148 ± 0.401  ns/op
> Single.storeStore              avgt    3   6.767 ± 0.164  ns/op // Better
> 
> 
> As expected, this does not affect x86_64 at all, because both `release` and `storeStore` are effectively no-ops, only affecting compiler optimizations:
> 
> 
> Benchmark                      Mode  Cnt  Score   Error  Units
> 
> Multiple.plain                 avgt    3  0.406 ± 0.002  ns/op
> Multiple.release               avgt    3  0.409 ± 0.018  ns/op
> Multiple.storeStore            avgt    3  0.406 ± 0.001  ns/op
> 
> MultipleWithLoads.plain        avgt    3  4.328 ± 0.006  ns/op
> MultipleWithLoads.release      avgt    3  4.600 ± 0.014  ns/op
> MultipleWithLoads.storeStore   avgt    3  4.602 ± 0.006  ns/op
> 
> MultipleWithStores.plain       avgt    3  0.812 ± 0.001  ns/op
> MultipleWithStores.release     avgt    3  0.812 ± 0.002  ns/op
> MultipleWithStores.storeStore  avgt    3  0.812 ± 0.002  ns/op
> 
> Publishing.plain               avgt    3  6.370 ± 0.059  ns/op
> Publishing.release             avgt    3  6.358 ± 0.436  ns/op
> Publishing.storeStore          avgt    3  6.367 ± 0.054  ns/op
> 
> Single.plain                   avgt    3  0.407 ± 0.039  ns/op
> Single.releaseFence            avgt    3  0.406 ± 0.001  ns/op
> Single.storeStoreFence         avgt    3  0.406 ± 0.001  ns/op
> 
> 
> Additional testing:
>  - [x] Linux x86_64 fastdebug `tier1`
>  - [x] Linux AArch64 fastdebug `tier1`
>  - [x] Linux x86_64 Fences benchmark
>  - [x] Linux AArch64 Fences benchmark
>  - [x] Linux ARM32 Fences benchmark
>  - [x] Linux AArch64 jcstress `quick` run

This pull request has now been integrated.

Changeset: b7a06be9
Author:    Aleksey Shipilev <shade at openjdk.org>
URL:       https://git.openjdk.java.net/jdk/commit/b7a06be98d3057dac4adbb7f4071ac62cf88fe52
Stats:     38 lines in 16 files changed: 32 ins; 5 del; 1 mod

8252990: Intrinsify Unsafe.storeStoreFence

Reviewed-by: dholmes, thartmann, whuang

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

PR: https://git.openjdk.java.net/jdk/pull/6136