RFR: 8313672: C2: PhaseCCP does not correctly track analysis dependencies involving shift, convert, and mask [v3]

[Emanuel Peter]

On Tue, 31 Oct 2023 15:07:02 GMT, Daniel Lundén <duke at openjdk.org> wrote:

>> Yes, I need to tweak this. The failure is, for some reason, not appearing when run through jtreg. I'll get back to you (also with answers to your other questions).
>
> Fixed now. `-XX:RepeatCompilation=20` is fast and seems enough to practically always trigger the failure (although it is non-deterministic due to `-XX:+StressIGVN`).
> 
> I have not managed to reproduce the bug without `-Xcomp`.
> 
> Running the test with only `-Xcomp` is quite slow (~1 minute on my system). I added the below instead (the test itself completes in <1 second).
> 
> /*
>  * @test
>  * @bug 8313672
>  * @summary Test CCP notification for value update of AndL through LShiftI and
>  *          ConvI2L (reduced set of flags).
>  * @run main/othervm -XX:CompileCommand=compileonly,compiler.ccp.TestShiftConvertAndNotification::test
>  *                   -Xcomp
>  *                   compiler.ccp.TestShiftConvertAndNotification
>  *
>  */

`-Xcomp` is always very slow (after all it compiles everything), unless you restrict compilation.
So here, I would remove both `-Xcomp` and the compilation restriction. That will make the run fast again. But then it is important to have enough repetitions of your test method, see comment lower down.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/16429#discussion_r1377861479