[lworld] RFR: 8372697: [lworld] compiler/c2/irTests/scalarReplacement/ScalarReplacementWithGCBarrierTests.java fails with --enable-preview

[Paul Hübner]

On Tue, 2 Dec 2025 08:50:02 GMT, Paul Hübner <phubner at openjdk.org> wrote:

>> In
>> https://github.com/openjdk/valhalla/blob/acb511a9f5c7b750b41e1ce77aab3d1a59613097/test/hotspot/jtreg/compiler/c2/irTests/scalarReplacement/ScalarReplacementWithGCBarrierTests.java#L98-L116
>> 
>> we have 2 allocations: `Iter` and `Integer`. Scalar replacement allows to eliminate the allocation of `Iter`, but we still had the allocation of `Integer`. With value classes, we can also save the allocation of `Integer` since it is a value class. Adapting expectations is enough.
>> 
>> To keep the test robust, I prefered to expect the exact amount, and not something like
>> 
>> @IR(phase = { CompilePhase.PHASEIDEAL_BEFORE_EA }, counts = { IRNode.ALLOC, "<= 2" })
>> @IR(phase = { CompilePhase.ITER_GVN_AFTER_ELIMINATION }, counts = { IRNode.ALLOC, "<= 1" })
>> 
>> Since that would allow the respective counts 1 and 1 (for instance, if `Integer` allocation is being removed as a value class, but `Iter` is not because EA would be broken with Valhalla).
>> 
>> To allow the test to work precisely with and without Valhalla, I propose a fake flag `enable-valhalla` that checks whether `Integer` is a value class. I preferred that over more generally "enable-preview" because it lacks granularity with respect to other preview features, and `PreviewFeatures.isEnabled()` is internal and not accessible anyway. It's a little hack, but I think the usage is very natural. It would be good if @chhagedorn could take a look at it.
>> 
>> Thanks,
>> Marc
>
> test/hotspot/jtreg/compiler/c2/irTests/scalarReplacement/ScalarReplacementWithGCBarrierTests.java line 108:
> 
>> 106:     // could not be eliminated.
>> 107:     @Test
>> 108:     @IR(applyIf = {"enable-valhalla", "false"}, phase = { CompilePhase.PHASEIDEAL_BEFORE_EA }, counts = { IRNode.ALLOC, "2" })
> 
> With `EnableValhalla` being removed in favour of `--enable-preview`, I think it would make sense if we transition to a preview on/off state rather than the fine-grained one we have right now.

FWIW there's also `jdk.internal.misc.PreviewFeatures` that can be used to check if preview features are enabled at runtime.

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

PR Review Comment: https://git.openjdk.org/valhalla/pull/1767#discussion_r2580232418