RFR: 8279219: [REDO] C2 crash when allocating array of size too large [v2]

[Tobias Hartmann]

On Tue, 11 Jan 2022 13:48:13 GMT, Roland Westrelin <roland at openjdk.org> wrote:

>> The new fix is largely similar to the previous one. 3 bugs were filed
>> because of the previous change but there are only really 2 issues:
>> 
>> - attaching the valid length condition at expansion time to the new
>>   array runtime call causes issues. With Shenandoah, passes of loop
>>   opts are executed after macro expansion but precedence edges are
>>   ignored when assigning controls to nodes which causes loop opts to
>>   incorrectly attempt to eliminate the node pointed to by the
>>   precedence edge. A similar issue occurs when the runtime call ends
>>   up in a subgraph that dies after macro expansion because the
>>   precedence edge is not cleared by dead code elimination which causes
>>   the runtime call to still be reachable. In the new patch, this is
>>   fixed by appending an extra input to the runtime call instead of
>>   using a precedence edge.
>> 
>> - In the previous patch, a top valid length input is used when there's
>>   no valid length input that needs to be recorded. That can cause an
>>   assert failure during CCP. If valid length initially has type top,
>>   the CatchNode out of the AllocateArray then have type (control,
>>   control). If next, the valid length input becomes constant 0, the
>>   CatchNode has type (top, control). An assert catches that the type
>>   of the CatchNode doesn't widen. This fixed by using 1 by default as
>>   valid length input and tweaking CatchNode::Value.
>> 
>> The new patch includes test cases for both issues.
>
> Roland Westrelin has updated the pull request incrementally with one additional commit since the last revision:
> 
>   review

Thanks for making these changes, looks good!

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

Marked as reviewed by thartmann (Reviewer).

PR: https://git.openjdk.java.net/jdk/pull/6952