[11] RFR(S): 8202565: C1 compilation crashes with "assert(is_double_stack() && !is_virtual()) failed: type check"

[Tobias Hartmann]

Hi Erik,

thanks for the review!

Best regards,
Tobias

On 04.05.2018 10:07, Erik Österlund wrote:
> Hi Tobias,
> 
> Looks good.
> 
> Thanks,
> /Erik
> 
> On 2018-05-04 09:16, Tobias Hartmann wrote:
>> Hi,
>>
>> please review the following patch:
>> https://bugs.openjdk.java.net/browse/JDK-8202565
>> http://cr.openjdk.java.net/~thartmann/8202565/webrev.00/
>>
>> C1 crashes because the LIR contains an instruction moving a T_OBJECT from the stack to a T_LONG
>> double-register ("move [stack:19|L] [rsirsi|J]"). This code is part of a G1 post barrier for an
>> object field store (stack:19 is the base oop).
>>
>> This problem was introduced by the modularization of the C1 GC barriers [1].
>> ModRefBarrierSetC1::resolve_address() may eagerly resolve the store address into a register,
>> assuming that the post write barrier needs the address in a register anyway. However, in the failing
>> case of a field store, the post barrier does not use precise marking and therefore only uses the
>> base address. The unnecessary leal increases register pressure around the barrier code and causes a
>> spill of the base address which then needs to be (re-)loaded from the stack.
>>
>> The patch restores pre-JDK-8201543 behavior by only eagerly resolving the address into a register if
>> the store is precise. More details (including relevant parts of the LIR) are in the bug comments.
>>
>> Thanks,
>> Tobias
>>
>> [1] https://bugs.openjdk.java.net/browse/JDK-8201543
>> http://cr.openjdk.java.net/~eosterlund/8201543/webrev.02/
>