RFR: 8051725: Improve expansion of Conv2B nodes in the middle-end [v7]

[Jasmine Karthikeyan]

On Thu, 18 May 2023 04:17:48 GMT, Jasmine Karthikeyan <jkarthikeyan at openjdk.org> wrote:

>> Hi, I've created optimizations for the expansion of `Conv2B` nodes, especially when followed immediately by an xor of 1. This pattern is fairly common, and can arise from both [cmov idealization](https://github.com/openjdk/jdk/blob/master/src/hotspot/share/opto/movenode.cpp#L241) and [diamond-phi optimization](https://github.com/openjdk/jdk/blob/master/src/hotspot/share/opto/cfgnode.cpp#L1571). This change replaces `Conv2B` nodes in the middle-end during post loop opts IGVN with conditional moves on supported platforms (x86_64, aarch64, arm32), allowing the bit flip with `xor` to be subsumed with an inversion of the comparison instead. This change also reduces the overhead of the matcher in the backends, as fewer rules need to be traversed in order to match an ideal node. Performance results from my (Zen 2) machine:
>> 
>> 
>>                                     Baseline                           Patch              Improvement
>> Benchmark                      Mode  Cnt  Score    Error Units     Score    Error  Units
>> Conv2BRules.testEquals0        avgt   10  47.566 ± 0.346 ns/op  /  34.130 ± 0.177  ns/op  + 28.2%
>> Conv2BRules.testNotEquals0     avgt   10  37.167 ± 0.211 ns/op  /  34.185 ± 0.258  ns/op  + 8.0%
>> Conv2BRules.testEquals1        avgt   10  35.059 ± 0.280 ns/op  /  34.847 ± 0.160  ns/op  (unchanged)
>> Conv2BRules.testEqualsNull     avgt   10  56.768 ± 2.600 ns/op  /  34.330 ± 0.625  ns/op  + 39.5%
>> Conv2BRules.testNotEqualsNull  avgt   10  47.447 ± 1.193 ns/op  /  34.142 ± 0.303  ns/op  + 28.0%
>> 
>> Reviews would be greatly appreciated!
>> 
>> Testing: tier1-2 on linux x64, GHA
>
> Jasmine Karthikeyan has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Changes from code review

Apologies for the delayed update, and thanks for the reviews!

I have aarch64 performance results, from an M1 mac:

                                    Baseline                           Patch              Improvement
Benchmark                      Mode  Cnt  Score    Error Units     Score    Error  Units
Conv2BRules.testEquals0        avgt   12  41.697 ± 0.127 ns/op  /  40.724 ± 0.086  ns/op  + 2.4%
Conv2BRules.testNotEquals0     avgt   12  39.522 ± 0.143 ns/op  /  40.608 ± 0.046  ns/op  - 2.7%
Conv2BRules.testEquals1        avgt   12  40.168 ± 0.136 ns/op  /  40.679 ± 0.044  ns/op  (unchanged)
Conv2BRules.testEqualsNull     avgt   12  48.922 ± 0.498 ns/op  /  42.046 ± 0.018  ns/op  + 15.1%
Conv2BRules.testNotEqualsNull  avgt   12  41.725 ± 0.264 ns/op  /  42.063 ± 0.043  ns/op  - 0.8%

It seems like the patch doesn't have much of an impact other than `testEqualsNull`, which would make sense as the Conv2B rule is using the same `cset` instruction as the 0 and 1 rule for CMoveI. I was unfortunately not able to test for arm32, but I think it should still be beneficial as the Conv2B rules there used two cmoves and had a fixme, whereas with this patch it would only use one cmove.

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

PR Comment: https://git.openjdk.org/jdk/pull/13345#issuecomment-1552411461