RFR(XL): 8224675: Late GC barrier insertion for ZGC

[Nils Eliasson]

Hi Roman,

I removed the method. I have verified that Shenandoah builds and isn't 
obviously broken.

Webrev updated in place.

Regards,

Nils

On 2019-05-23 16:31, Roman Kennke wrote:
> Quick glance showed a problem: you are renaming/moving
> BarrierSetC2::add_users_to_worklist() but ShenandoahBarrierSetC2 is not
> updated accordingly.
>
> Roman
>
>
>> Hi,
>>
>> In ZGC we use load barriers on references. In the original
>> implementation these where added as macro nodes at parse time. The load
>> barrier node consumes and produces control flow in order to be able to
>> be lowered into a check with a slow path late. The load barrier nodes
>> are fixed in the control flow, and extensions to different optimizations
>> are need the barriers out of loop and past other unrelated control flow.
>>
>> With this patch the barriers are instead added after the loop
>> optimizations, before macro node expansion. This makes the entire
>> pipeline until that point oblivious about the barriers. A dump of the IR
>> with ZGC or EpsilonGC will be basically identical at that point, and the
>> diff compared to serialGC or ParallelGC that use write barriers is
>> really small.
>>
>> Benefits
>>
>> - A major complexity reduction. One can reason about and implement loop
>> optimization without caring about the barriers. The escape analysis
>> doesn't need to know about the barriers. Loads float freely like they
>> are supposed to.
>>
>> - Less nodes early. The inlining will become more deterministic. A
>> barrier heavy GC will not run into node limits earlier. Also node limit
>> bounded optimization like unrolling and peeling will not be penalized by
>> barriers.
>>
>> - Better test coverage, or reduce testing cost when the same
>> optimization doesn't need to be verified with every GC.
>>
>> - Better control on where barriers end up. It is trivial to guarantee
>> that the load and barriers are not separated by a safepoint.
>>
>> Design
>>
>> The implementation uses an extra phase that piggy back on PhaseIdealLoop
>> which provides control and dominator information for all loads. This
>> extra phase is needed because we need to splice the control flow when
>> adding the load barriers.
>>
>> Barriers are inserted on the loads nodes in post order (any successor
>> first). This is to guarantee the dominator information above every
>> insertion is correct. This is also important within blocks. Two loads in
>> the same block can float in relation to each other. The addition of
>> barriers serializes their order. Any def-use relationship is upheld by
>> expanding them post order.
>>
>> Barrier insertion is done in stages. In this first stage a single macro
>> node that represents the barrier is added with all dependencies that is
>> required. In the macro expansion phase the barrier nodes is expanded
>> into the final shape, adding nodes that represent the conditional load
>> barrier check. (Write barriers in other GCs could possibly be expanded
>> here directly)
>>
>> All the barriers that are needed for unsafe reference operations (cas,
>> swap, cmpx) are also expanded late. They already have control flow, so
>> the expansion is straight forward.
>>
>> The barriers for the unsafe reference operations (cas, getandset, cmpx)
>> have also been simplified. The cas-load-cas dance have been replaced by
>> a pre-load. The pre-load is a load with a barrier, that is kept alive by
>> an extra (required) edge on the unsafe-primitive-nodes (specialized as
>> ZCompareAndSwap, ZGetAndSet, ZCompareAndExchange).
>>
>> One challenge that was encountered early and that have caused
>> considerable work is that nodes (like loads) can end up between calls
>> and their catch projections. This is usually handled after matching, in
>> PhaseCFG::call_catch_cleanup, where the nodes after the call are cloned
>> to all catch blocks. At this stage they are in an ordered list, so that
>> is a straight forward process. For late barrier insertion we need to
>> splice in control earlier, before matching, and control flow between
>> calls and catches is not allowed. This requires us to add a
>> transformation pass where all loads and their dependent instructions are
>> cloned out to the catch blocks before we can start splicing in control
>> flow. This transformation doesn't replace the legacy call_catch_cleanup
>> fully, but it could be a future goal.
>>
>> In the original barrier implementation there where two different load
>> barrier implementations: the basic and the optimized. With the new
>> approach to barriers on unsafe, the basic is no longer required and has
>> been removed. (It provided options for skipping the self healing, and
>> passed the ref in a register, guaranteeing that the oop wasn't reloaded.)
>>
>> The wart that was fixup_partial_loads in zHeap has also been made
>> redundant.
>>
>> Dominating barriers are no longer removed on weak loads. Weak barriers
>> doesn't guarantee self-healing.
>>
>> Follow up work:
>>
>> - Consolidate all uses of GrowableArray::insert_sorted to use the new
>> version
>>
>> - Refactor the phases. There are a lot of simplifications and
>> verification that can be done with more well defined phases.
>>
>> - Simplify the remaining barrier optimizations. There might still be
>> code paths that are no longer needed.
>>
>>
>> Testing:
>>
>> Hotspot tier 1-6, CTW, jcstress, micros, runthese, kitchensink, and then
>> some. All with -XX:+ZVerifyViews.
>>
>> Bug: https://bugs.openjdk.java.net/browse/JDK-8224675
>>
>> Webrev: http://cr.openjdk.java.net/~neliasso/8224675/webrev.01/
>>
>>
>> Please review,
>>
>> Regards,
>>
>> Nils
>>