RFR: 8324969: C2: prevent elimination of unbalanced coarsened locking regions [v5]

[Vladimir Kozlov]

On Wed, 14 Feb 2024 05:24:20 GMT, Vladimir Kozlov <kvn at openjdk.org> wrote:

>> The issue is that when we do nested locks elimination we don't check that synchronized region has both `Lock` and `Unlock` nodes.
>> Which can happen if locks coarsening optimization eliminated pair of Unlock/Lock nodes from adjacent locking regions before we check for nested locks.
>> 
>> Consider this code (all locks/unlocks use the same object):
>> 
>>     lock1(obj) { // outer synchronized region
>>          lock2(obj) {
>>              // nested synchronized region 1
>>          } unlock2(obj);
>>          lock3(obj) {
>>              // nested synchronized region 2
>>          } unlock3(obj);
>>     } unlock1(obj);
>> 
>> If `lock3` directly follows `unlock2` (no branches or safepoints) locks coarsening optimization will remove them:
>> 
>> 
>>     lock1(obj) { // outer synchronized region
>>          lock2(obj) {
>>              // nested synchronized region 1
>>          };
>>          {
>>              // nested synchronized region 2
>>          } unlock3(obj);
>>     } unlock1(obj);
>> 
>> 
>> Nested locks elimination code checks only `Lock` node in one region to find if it is nested (inside other lock region which use the same object) and then eliminate it. So we end up with not eliminated `Unlock` node in second nested region.
>> 
>> Why we don't hit this issue before? Normally nested locks elimination executed first and only then we do locks coarsening elimination. In the example above we eliminate all nested `Lock` and `Unlock` nodes, leaving only outer `Lock` and `Unlock`.
>> 
>> The additional factors which leads to the failure is fully unrolled loop around nested sync regions and some allocation to trigger Escape Analysis:
>> 
>> 
>>     lock1(obj) { // outer synchronized region
>>          Test var = new Test(); // Triggers EA
>>          for (I = 0; I < 3; i++) { // small iteration number to fully unroll
>>              lock2(obj) {
>>                  // nested synchronized region 1
>>              } unlock2(obj);
>>              lock3(obj) {
>>                  // nested synchronized region 2
>>              } unlock3(obj);
>>          }
>>     } unlock1(obj);
>> 
>> Before executing Escape Analysis we do loops optimization to simplify graph: [compile.cpp#L2332](https://github.com/openjdk/jdk/blob/master/src/hotspot/share/opto/compile.cpp#L2332)
>> We also allow to fully unroll short loops (LoopOptsMaxUnroll) to remove merges from graph. It helps EA eliminate allocations.
>> Such unrolling creates several `Lock` and `Unlock` nodes per synchronized region. But nested locks eliminat...
>
> Vladimir Kozlov has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Additional changes

Looking on cases for locks coarsening [callnode.cpp#L1667](https://github.com/openjdk/jdk/blob/master/src/hotspot/share/opto/callnode.cpp#L1667) I don't see how it can become **balanced**.  In all cases we have separate synchronization regions (`s()`) with different BoxLock nodes. Even if one branch of code is gone the rest will still have separate regions.

I can do one more experiment. Do BoxLock marking in `add_coarsened_locks()` as you suggested and compare it in `mark_unbalanced_boxes()`.

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

PR Comment: https://git.openjdk.org/jdk/pull/17697#issuecomment-1944802308