On Wed, 21 Aug 2024 14:21:39 GMT, Markus Grönlund <mgron...@openjdk.org> wrote:
>> Thread.currentThread() has an intrinsic, and isVirtual is just a type check.
>> ContinuationSupport.isSupported reads a static final so will disappear once
>> compiled. The pattern we are using in other areas is for the pin to return a
>> boolean (like David suggested).
>
> I looked into this in more detail. The current suggestion:
>
> mov r10,QWORD PTR [r15+0x388] ; _vthread OopHandle
> mov r10,QWORD PTR [r10] ; dereference OopHandle <<--
> Thread.currentThread() intrinsic gives 2 instructions
> mov r11d,DWORD PTR [r10+0x8] ; InstanceKlass to r11 <-- isVirtual()
> mov r10d,r11d ; InstanceKlass to r10
> mov r8,QWORD PTR [r10+0x40] ; Load slot in InstanceKlass primary
> supers array to r8
> movabs r10,0x2d0481a8 ; InstanceKlass for
> {metadata('java/lang/BaseVirtualThread')} to r10
> cmp r8,r10 ; compare if superklass
> is java/lang/BaseVirtualThread
> jne 0x0000018571e0baf9 ; 6 instructions for isVirtual()
> type check, 8 instructions in total
>
> This gives a prologue of eight instructions.
>
> For JFR, we already have much of this information resolved when loading up
> the EventWriter instance using the existing intrinsic getEventWriter().
> Hence, we could extend that to mark the event writer with a field to say if
> pinning should be performed. This results in only a two instruction prologue:
>
> test r8d,r8d ; pinVirtualThread?
> je 0x0000012580a0f6c9 ; 2 instructions for test
>
> This is an x4 speedup, although slightly less because of an additional store
> instruction for loading the event writer.
>
> Further, I looked into the Continuation.pin() and Continuation.unpin()
> methods. They are currently not intrinsics, but lend themselves well to
> intrinsification. I have created such intrinsics, and the results are quite
> good.
>
> Continuation.pin() or Continuation.unpin() without intrinsics = 112
> instructions each
> Continuation.pin() or Continuation.unpin() with intrinsics = 8 instructions
> each
>
> This is an x14 speedup for virtual threads.
I plan to fix the event writer under this PR (to be updated) and file a
separate tracking enhancement for the intrinsification of Continuation.pin()
and Continuation.unpin().
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/20588#discussion_r1725150866
