RFR: 8269820: C2 PhaseIdealLoop::do_unroll get wrong opaque node

[Roland Westrelin]

On Mon, 28 Nov 2022 18:46:11 GMT, Vladimir Kozlov <kvn at openjdk.org> wrote:

>> A main loop loses its pre loop. The Opaque1 node for the zero trip
>> guard of the main loop is assigned control at a Region through which
>> an If is split. As a result, the Opaque1 is cloned and the zero trip
>> guard takes a Phi that merges Opaque1 nodes. One of the branch dies
>> next and as, a result, the zero trip guard has an Opaque1 as input but
>> at the wrong CmpI input. The assert fires next.
>> 
>> The fix I propose is that if an Opaque1 node that is part of a zero
>> trip guard is encountered during split if, rather than split if up or
>> down, instead, assign it the control of the zero trip guard's
>> control. This way the pattern of the zero trip guard is unaffected and
>> split if can proceed. I believe it's safe to assign it a later
>> control:
>> 
>> - an Opaque1 can't be shared
>> 
>> - the zero trip guard can't be the If that's being split
>> 
>> As Vladimir noted, this bug used to not reproduce with loop strip
>> mining disabled but now always reproduces because the loop
>> strip mining nest is always constructed. The reason is that the
>> main loop in this test is kept alive by the LSM safepoint. If the
>> LSM loop nest is not constructed, the loop is optimized out. I
>> filed:
>> 
>> https://bugs.openjdk.org/browse/JDK-8297724
>> 
>> for this issue.
>
> src/hotspot/share/opto/split_if.cpp line 242:
> 
>> 240:       set_ctrl(n, ctrl->in(0)->in(0));
>> 241:       set_ctrl(cmp, ctrl->in(0)->in(0));
>> 242:       set_ctrl(bol, ctrl->in(0)->in(0));
> 
> Why you assign control to `cmp` and `bol` too?

The subgraph is` (Bool (CmpI (Opaque1...))`. So if only the control of `Opaque1` is updated then the `CmpI`/`Bool` could end up with a control that strictly dominates the control of the `Opaque1`. It's quite possible that that wouldn't break anything but wouldn't it be inconsistent and quite ugly?

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

PR: https://git.openjdk.org/jdk/pull/11391