RFR: 8318049: C2: assert(!failure) failed: Missed optimization opportunity in PhaseIterGVN

Emanuel Peter epeter at openjdk.org
Thu Oct 19 09:56:41 UTC 2023


On Thu, 19 Oct 2023 09:48:39 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>> This patch fixes [JDK-8318049](https://bugs.openjdk.org/browse/JDK-8318049) by informing XorI and XorL nodes in addition to sub and phi nodes when Cast nodes are updated during IGVN. 
>> 
>> The bug manifests when loop optimizations change the IR:
>> ![image](https://github.com/openjdk/jdk/assets/25208576/5274f2cc-a6fb-4ce6-89f4-4a638577901c)
>> 
>> To the following form:
>> ![image](https://github.com/openjdk/jdk/assets/25208576/07bcbec3-f0bb-4d6a-b312-6a459e6b8c0c)
>> 
>> XorI nodes check for that form in their `Value` call, to do the `x ^ x => 0` transformation, checking through casts. As the cast node didn't inform the Xor when it changes, the transformation doesn't have a chance to run.
>> https://github.com/openjdk/jdk/blob/5bd10521eb5e51e76b20e955addd45f76abba6f7/src/hotspot/share/opto/addnode.cpp#L947-L950
>> 
>> I've attached a reduced version of the test found in the bug. Reviews would be appreciated!
>
> src/hotspot/share/opto/phaseX.cpp line 1603:
> 
>> 1601:     // ConstraintCast+ -> Xor
>> 1602:     if (use->is_ConstraintCast()) {
>> 1603:       auto push_phi_or_sub_uses_to_worklist = [&](Node* n){
> 
> Suggestion:
> 
>       auto push_the_uses_to_worklist = [&](Node* n){

Change the name now. Maybe something more generic is good, since we do not know how many more of these examples we will eventually add to this.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/16214#discussion_r1365231067


More information about the hotspot-compiler-dev mailing list