RFR(S): 8219517: assert(false) failed: infinite loop in PhaseIterGVN::optimize

[Nils Eliasson]

Thank you, Vladimir!

// Nils

On 2019-03-11 19:16, Vladimir Kozlov wrote:
> On 3/11/19 6:33 AM, Nils Eliasson wrote:
>> On 2019-03-08 19:27, Vladimir Kozlov wrote:
>>> Hi Nils,
>>>
>>> What if it is normal diamond shaped graph and no loop? 
>>
>> It can't be. The loads memory edge is 'mem'. 'mem' is a phi, and we 
>> loop over its in-edges . The optimization only triggers when: mem 
>> ==opt_mem_chain(mem->in(x)). That can only happen when there is a loop.
>
> Okay, I got it.
>
>>
>> (And my addition is to block the optimization when both are loops: 
>> mem == opt_mem_chain(mem->in(1)) == opt_mem_chain(mem->in(2)).
>
> Good. Reviewed.
>
> Thanks,
> Vladimir
>
>>
>> Regards,
>>
>> / Nils
>>
>>> You can have case when both branches point to the same memory too. 
>>> May be add is_Loop() check.
>>>
>>> Thanks,
>>> Vladimir
>>>
>>>
>>> On 3/8/19 2:52 AM, Nils Eliasson wrote:
>>>> Hi all,
>>>>
>>>> Background:
>>>>
>>>> We can get stuck in an infinite loop in IGVN. The method 
>>>> reproducing the problem is quite a big graph, and after some 
>>>> optimization, a huge loop will die. But since it is so big, it 
>>>> takes a while before it has been pruned.
>>>>
>>>> In an inner loop there is a phi on memory that gets reduced to a 
>>>> self looping heart, with a membar on each in edge. There is also a 
>>>> connected region that keeps it alive. (From the start there is 
>>>> other memory state coming into this loop, but it gets disconnected 
>>>> early when the loop dies.)
>>>>
>>>> +---+                 +---+
>>>> |   v                 v   |
>>>> | Membar +-+ +---+ Membar |
>>>> |          | |            |
>>>> |          v v            |
>>>> |          Phi            |
>>>> |          + + +          |
>>>> |          | | |          |
>>>> +----------+ | +----------+
>>>>               |
>>>>               v
>>>>             LoadN
>>>>
>>>> In IGVN, Ideal() will be called on the Load.
>>>>
>>>> On iteration 1 - A split_through_phi on one edge will be performed, 
>>>> because we can prove that other edge of the phi is a loop. Now the 
>>>> Load hangs of one of the membars.
>>>>
>>>> On iteration 2 - Optimize_memory_chain will suggest the in to the 
>>>> membar as a more ideal memory, and then the load get the phi back 
>>>> as the memory input.
>>>>
>>>> Repeat.
>>>>
>>>> I have gone great lengths to show that this code is part of a huge 
>>>> loop, that is dead, and will be eliminated in due time.
>>>>
>>>> My suggested solution to breaking the infinite loop, is to change 
>>>> the first case, by simply not perform the memory replacement when 
>>>> both inputs are self loops.
>>>>
>>>> https://bugs.openjdk.java.net/browse/JDK-8219517
>>>>
>>>> http://cr.openjdk.java.net/~neliasso/8219517/webrev.01/
>>>>
>>>> Regards,
>>>>
>>>> Nils
>>>>
>>>>
>>>>
>>>>