RFR(S): 8239367: RunThese30M.java failed due to "assert(false) failed: graph should be schedulable"

[Nils Eliasson]

+1

This solution looks more robust.

Best regards,
Nils

On 2020-03-03 08:55, Tobias Hartmann wrote:
> Hi Roland,
>
> that looks good to me and I'm actually glad to get rid of find_bottom_mem.
>
> Best regards,
> Tobias
>
> On 02.03.20 10:15, Roland Westrelin wrote:
>> http://cr.openjdk.java.net/~roland/8239367/webrev.00/
>>
>> The code to wire memory when the new SubTypeCheck node is expanded is
>> broken. It assumes that the first bottom memory it encounters when
>> walking through the graph dominates the point where the SubTypeCheck is
>> expanded. That doesn't hold. In this particular case, the SubTypeCheck
>> node is expanded right after a region that has several memory Phis all
>> feeding into a MergeMem. The find_bottom_mem() method doesn't see bottom
>> memory at the region so picks one of the Region's edge and looks for
>> bottom memory there. It finds a bottom memory Phi which doesn't dominate
>> the SubTypeCheck node. I think fixing find_bottom_mem() is too
>> complicated and propose we go with Vladimir I's suggestion of using
>> immutable memory instead:
>>
>> https://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/2020-February/037022.html
>>
>> Roland.
>>