RFR: 8325520: Vector loads with offsets incorrectly compiled

[Emanuel Peter]

On Tue, 9 Apr 2024 07:20:39 GMT, Damon Fenacci <dfenacci at openjdk.org> wrote:

>>> It seems that your code would now disallow such a case, because you always check that the Ideal node you get back is of the same type as the this node. Am I right about that? Is that intended?
>> 
>> If I am right that you would have made such a optimization impossible, that probably means that our tests don't have an IR test that cover this case. You would definitely need to add such IR tests, otherwise we don't know if we are getting regressions. You will probably also have to run this patch through performance testing eventually.
>> 
>>> I've tried to use the IR framework to check for folded/non folded nodes but couldn't make it reliable enough (in the end it didn't test more than the current test). So I decided to go back to the actual "regression" test which reproduces the original issue.
>> 
>> Do you know what is the issue with reliability for the IR rules? Why did it not always work?
>
>> If I am right that you would have made such a optimization impossible, that probably means that our tests don't have an IR test that cover this case. You would definitely need to add such IR tests, otherwise we don't know if we are getting regressions. You will probably also have to run this patch through performance testing eventually.
> 
> I've transformed the tests to add IR tests as well. The issue with them seems to be related with [JDK-8302459](https://bugs.openjdk.org/browse/JDK-8302459) (basically there are some missing cleanups when performing late inlining). So, for now the tests force a cleanup at every step (`-XX:+IncrementalInlineForceCleanup`).

@dafedafe Nice, I think this already looks much better. Let me know if/when you want me to look at it again ;)

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

PR Comment: https://git.openjdk.org/jdk/pull/18347#issuecomment-2055553058