RFR (S): 8148994: Replacing MH::invokeBasic with a direct call breaks LF customization
claes.redestad at oracle.com
Fri Feb 12 23:12:11 UTC 2016
this seems OK to me, and I've verified that this gets octane performance
back to 9-b102 levels across a number of platforms.
On 2016-02-12 18:15, Vladimir Ivanov wrote:
> Any reviews, please?
> Best regards,
> Vladimir Ivanov
> On 2/5/16 7:29 PM, Vladimir Ivanov wrote:
>> Thanks, John.
>> On 2/4/16 9:52 PM, John Rose wrote:
>>> Looks good. Except I don't recall how is_inline interacts, if at all,
>>> with customization. It's not clear what is being gated, and which
>>> optimizations get let through.
>> Inlining through MH::invokeBasic doesn't happen when there's @DontInline
>> on the target. It is the case for GWT which has CountingWrapper on both
>> paths. It starts with a reinvoker.dontInline LF and then switches to a
>> reinvoker LF (after 30 invocations). If a compilation happens when
>> reinvoker.dontInline is used, a direct call is generated and even if the
>> branch becomes hot, all the calls goes through reinvoker.dontInline.
>> Update to reinvoker LF isn't visible unless the dispatch goes through
>> MH::invokeBasic linker.
>> I had an impression that I introduced LF customization logic to the
>> non-inlined case in GWT, but I can't find it in the code. So, it seems
>> the regression is caused by repeated checks performed by
>> reinvoker.dontInline LF and not by LF customization missed in generated
>> Also, I stumbled upon a quirk in inlining policy: inlining through
>> linkers can't be delayed. It happens when JIT hits some limit on IR
>> size. So, right now inlining can stop in the middle of MH chain and a
>> linker call is generated.
>> But after JDK-8072008 there's no problem with delaying inlining. C2 can
>> decide whether to keep the direct call or inline through it. So, I
>> enabled late inlining for all linkers. (Surprisingly, no significant
>> performance difference on nashorn.)
>> So, I decided to keep invokeBasic linker call replacement for now, but
>> piggypack LF customization on the stuck counting LF.
>> Updated webrev:
>> * enabled late inlining for linkers;
>> * add LF customization check in CountingWrapper::countDown, so when
>> counting is stuck, the target will be customized.
>> Testing: octane, JPRT.
>> I'm working on JDK-8071793  and will reconsider how LF customization
>> on non-inlined branch in GWTs works given there is a more JVM-friendly
>> mechanism to control inlining. With the proposed fix
>> reinvoker.dontInline is never customized (only the target it points to)
>> and serves as a trampoline between GWT & target LF. It would be better
>> to jump right to the customized LF, if non-inlined branch becomes hot.
>> Best regards,
>> Vladimir Ivanov
>>  https://bugs.openjdk.java.net/browse/JDK-8071793
>> "Prune rarely called invokedynamic call sites during inlining"
>>> Maybe a comment about that would be useful
>>> – John
>>>> On Feb 4, 2016, at 7:51 AM, Vladimir Ivanov
>>>> <vladimir.x.ivanov at oracle.com> wrote:
>>>> JDK-8072008  introduced an optimization to bypass linker calls in
>>>> generated code when target method can't be inlined.
>>>> Unfortunately, it doesn't work well with MH::invokeBasic(), because
>>>> it clashes with LambdaForm customization: if a call is not inlined,
>>>> it is important for performance to update (customize) corresponding
>>>> LambdaForm on-the-fly. MH::invokeBasic() linker does MH -> LF ->
>>>> Method* -> _from_compiled_entry traversal on every invocation, while
>>>> direct call points right to the _from_compiled_entry. So, any
>>>> LambdaForm updates aren't visible until recompilation.
>>>> The fix is to keep MH::invokeBasic linker, so up-to-date LambdaForm
>>>> instance is used on every invocation.
>>>> Also, cleaned up relevant tests a bit.
>>>> Testing: octane, JPRT.
>>>> Best regards,
>>>> Vladimir Ivanov
>>>>  https://bugs.openjdk.java.net/browse/JDK-8072008
>>>> "Emit direct call instead of linkTo* for recursive
>>>> indy/MH.invoke* calls"
More information about the core-libs-dev