RFR: 8342387: C2 SuperWord: refactor and improve compiler/loopopts/superword/TestDependencyOffsets.java [v2]

[Tobias Hartmann]

On Tue, 22 Oct 2024 13:08:49 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>> I want to refactor `TestDependencyOffsets.java` using the `CompileFramework`.
>> 
>> Reasons:
>> - I soon need to modify the IR rules in this test soon anyway (https://github.com/openjdk/jdk/pull/21521), and so a refactor might be good to do first.
>> - The generator script used to be a `generator.py`, stored not on `git` but `JBS`. Not great. Now we have it in Java code, and maintenance is easier.
>> - Since I first wrote that test, I have now introduced the `IRNode.VECTOR_SIZE`. This allows:
>>   - Simplifying the logic for the IR rules (removed the `IRRange` and `IRBool`, and the `Platform`).
>>   - Strengthening the rules.
>> - I was able to add `random` offsets. This means we have better coverage, and do not rely on just hand-crafted values.
>> 
>> I extensively use `String.format` and `StringBuilder`... would be nicer to have string-templates but they don't exist yet.
>> 
>> Recommendation for review: the old file was huge. Finding the new code in the diff can be hard. I would start by only reading the new file.
>> 
>> Ah. And about runtime of the test. On my machine I get this (in ms):
>> 
>> Generate: 27
>> Compile:  5845
>> Run:      23435
>> 
>> Test generation is negligible. 6sec on compilation, 20+sec on execution. I think that is an ok balance, at least we can say that generation and compilation only take about 1/6 of total time - an acceptable overhead I would say.
>
> Emanuel Peter has updated the pull request incrementally with one additional commit since the last revision:
> 
>   add @compile for IR Framework

This is hard to review due to the scattered view in the diff. As you suggested, I had a look at the new file and it's nice how compact and simple the test now is. The changes look good to me.

> I think that is an ok balance, at least we can say that generation and compilation only take about 1/6 of total time - an acceptable overhead I would say.

Yes, I agree. The new version is much more maintainable and there is no risk anymore that the test generation script and the test start to diverge.

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

Marked as reviewed by thartmann (Reviewer).

PR Review: https://git.openjdk.org/jdk/pull/21541#pullrequestreview-2385358021