RFR: 8300926: Several startup regressions ~6-70% in 21-b6 all platforms [v6]

[Axel Boldt-Christmas]

On Thu, 23 Feb 2023 20:14:32 GMT, Robbin Ehn <rehn at openjdk.org> wrote:

>> Hi all, please consider.
>> 
>> The original issue was when thread 1 asked to deopt nmethod set X and thread 2 asked for the same or a subset of X.
>> All method will already be marked, but the actual deoptimizing, not entrant, patching PC on stacks and patching post call nops, was not done yet. Which meant thread 2 could 'pass' thread 1.
>> Most places did deopt under Compile_lock, thus this is not an issue, but WB and clearCallSiteContext do not.
>> 
>> Since a handshakes may take long before completion and Compile_lock is used for so much more than deopts.
>> The fix in https://bugs.openjdk.org/browse/JDK-8299074 instead always emitted a handshake even when everything was already marked. (instead of adding Compile_lock to all places)
>> 
>> This turnout to be problematic in the startup, for example the number of deopt handshakes in jetty dry run startup went from 5 to 39 handshakes.
>> 
>> This fix first adds a barrier for which you do not pass until the requested deopts have happened and it coalesces the handshakes.
>> Secondly it moves handshakes part out of the Compile_lock where it is possible.
>> 
>> Which means we fix the performance bug and we reduce the contention on Compile_lock, meaning higher throughput in compiler and things such as class-loading.
>> 
>> It passes t1-t7 with flying colours! t8 still not completed and I'm redoing some testing due to last minute simplifications.
>> 
>> Thanks, Robbin
>
> Robbin Ehn has updated the pull request incrementally with two additional commits since the last revision:
> 
>  - Comment fixes
>  - Include/fwd fixes

src/hotspot/share/prims/whitebox.cpp line 798:

> 796:       result += mh->method_holder()->mark_osr_nmethods(&deopt_scope, mh());
> 797:     } else if (mh->code() != nullptr) {
> 798:       deopt_scope.mark(mh->code());

I was working on a patch based on top of this one and had a crash here.

The second call to `mh->code()` returned a `nullptr`. It looks racy to me and seems like the only thing protecting `CompileMethod::code()` is the `CompiledMethod_lock` which is not held here. 

Regardless `mh->code()` should probably only be loaded once. (Or at least not re-loaded after the nullptr check)

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

PR: https://git.openjdk.org/jdk/pull/12585