RFR: 8280510: AArch64: Vectorize operations with loop induction variable [v2]

[Pengfei Li]

On Tue, 15 Mar 2022 09:53:28 GMT, Andrew Dinn <adinn at openjdk.org> wrote:

> The code looks ok to me modulo the cut-and-paste error I highlighted.

Thanks for your careful review. The cut-and-paste error has been fixed in my updated patch.

> Have you any way of ensuring that the generated code is correct (other than the fact that it runs quicker).
>
> Also have you checked this does not invalidate any other SVE tests. I would expect some failures thanks to the cut-and-paste error.

I've run almost full jtreg on an SVE featured machine and no issue is found. I noticed that existing hotspot jtregs cases in `compiler/c2/cr7192963/Test*Vect.java` have this kind of loop patterns. (Unfortunately, there is no related failure found even with that cut-and-paste error.)

> It would be a good idea if someone who knows the superword code better than me could comment on this extension to the range of cases covered by superword transformation. This looks like a very useful extension to the range of cases that can be optimized using superword analysis. However, I am not familiar enough with the code to know if this is the best way to proceed or if some other alternative has been considered and might provide a better alternative, especially wrt architectures other than aarch64.

I agree. I think people who know superword well are mostly in Oracle so I'd like to see if anyone in Oracle could help look at this. I've considered enabling this for x86. But as I'm not familiar with x86 instruction set enough, I haven't found any instruction(s) to implement such index vector. I would be very glad if someone from Intel can give me some suggestions.

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

PR: https://git.openjdk.java.net/jdk/pull/7491