RFR: JDK-8272570: C2: crash in PhaseCFG::global_code_motion [v3]

[Tobias Hartmann]

On Tue, 17 Aug 2021 13:00:06 GMT, 王超 <github.com+25214855+casparcwang at openjdk.org> wrote:

>> In PhaseCFG::global_code_motion, it allocate recalc_pressure_nodes array. condition is at least one block has more than 10 nodes:
>> 
>>   if (OptoRegScheduling) {
>>     for (uint i = 0; i < number_of_blocks(); i++) {
>>       Block* block = get_block(i);
>>       if (block->number_of_nodes() > 10) {
>>         block_size_threshold_ok = true;
>>         break;
>>       }
>>     }
>>   }
>> 
>> 
>> In PhaseCFG::select, it uses recalc_pressure_nodes array, if block's node exceed 10:
>> 
>> bool block_size_threshold_ok = (block->number_of_nodes() > 10) ? true : false;
>> 
>> 
>> In this case, block#4's node size is 10 at GCM begin, but increased to 11 later. PhaseCFG::select use null recalc_pressure_nodes array, this leads to crash.
>> 
>> # --- schedule_local B4, before: ---
>> #   31: Region              31    6
>> #   22: MergeMem          NULL    0   17    0   23   25 NULL NULL
>> #   25: storeImmB            3   17   24 NULL
>> #   23: storeImmB            3   17   24 NULL
>> #    3: MachProj             4
>> #    4: CallLeafDirect      31    0   17   18    0   19    0 NULL
>> #    5: MachProj             4
>> #   19: loadConD          NULL   20 NULL NULL
>> #   20: MachConstantBaseNode   11 NULL NULL NULL
>> #    2: Ret                  3   21   22   18   26 NULL NULL NULL
>> #
>> #   ready cnt: -2  latency:  0    31: Region
>> #   ready cnt:  2  latency:  0    22: MergeMem
>> #   ready cnt:  1  latency:  0    25: storeImmB
>> #   ready cnt:  1  latency:  0    23: storeImmB
>> #   ready cnt:  1  latency:  0     3: MachProj
>> #   ready cnt:  1  latency:  0     4: CallLeafDirect
>> #   ready cnt:  1  latency:  0     5: MachProj
>> #   ready cnt:  1  latency:  0    19: loadConD
>> #   ready cnt:  0  latency:  0    20: MachConstantBaseNode
>> #   ready cnt:  0  latency:  0     2: Ret
>> #   ready list: 20
>> #    select 20: MachConstantBaseNode, latency:0 20  MachConstantBaseNode  ===  11  [[ 19 ]] 
>> #   ready list: 19
>> #    select 19: loadConD, latency:0 19  loadConD  === _  20  [[ 4 ]] dblcon:1.000000
>> #   ready list: 4
>> #    select 4: CallLeafDirect, latency:0 4  CallLeafDirect  ===  31  0  17  18  0  19  0  [[ 5  3 ]] dtan # double/half ( double, half ) !jvms: SDD0::foo @ bci:23 (line 10)
>> #   ready list: 23 25
>> select node_size # B4, 11
>> #   31: Region              31    6
>> #   20: MachConstantBaseNode   11 NULL NULL NULL
>> #   19: loadConD          NULL   20 NULL NULL
>> #    4: CallLeafDirect      31    0   17   18    0   19    0 NULL
>> #    5: MachProj             4
>> #    3: MachProj             4
>> #   35: MachProj             4
>> #    5: MachProj             4
>> #   19: loadConD          NULL   20 NULL NULL
>> #   20: MachConstantBaseNode   11 NULL NULL NULL
>> #    2: Ret                  3   21   22   18   26 NULL NULL NULL
>> #
>> #
>> # A fatal error has been detected by the Java Runtime Environment:
>> #
>> #  SIGSEGV (0xb) at pc=0x00007f14bf3ccf81, pid=31936, tid=31949
>
> 王超 has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Rename test name

So this code assumes that the number of nodes per block does not change. Where did we add a new node?

test/hotspot/jtreg/compiler/c2/TestGCMRecalcPressureNodes.java line 31:

> 29:  * @requires vm.compiler2.enabled
> 30:  *
> 31:  * @run main/othervm TestGCMRecalcPressureNodes

Test can be executed in agent VM mode.

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

Changes requested by thartmann (Reviewer).

PR: https://git.openjdk.java.net/jdk/pull/5140