Investigate codegen for Vector128.GetUpper

[tannergooding]

As per #53450 (comment), it would like be beneficial to investigate changing the codegen of Vector128.GetUpper to use DUP rather than EXT on ARM64.

category:cq
theme:codegen

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT
See info in area-owners.md if you want to be subscribed.

Issue Details

---

As per #53450 (comment), it would like be beneficial to investigate changing the codegen of Vector128.GetUpper to use DUP rather than EXT on ARM64.

Author:	tannergooding
Assignees:	tannergooding
Labels:	arch-arm64, tenet-performance, area-CodeGen-coreclr, untriaged
Milestone:	-

[JulieLeeMSFT]

Cc @dotnet/jit-contrib

[TIHan]

@TamarChristinaArm , I have a question, looking at the A78 Optimization Guide, EXT D-form and DUP D-form both have the same latency and throughput, is this indeed true? If so, then there wouldn't be a benefit to emitting DUP instead of EXT here.

[ghost]

This issue has been marked needs-author-action and may be missing some important information.

[TamarChristinaArm]

@TIHan yeah they have the same latency and throughput across all Arm Cortex and Neoverse cores.
The only improvement here is if one does Vector128.GetUpper().AsScalar() in which case the UMOV is faster than EXT + UMOV.

[tannergooding]

GetUpper/Lower() followed be a GetElement (including ToScalar()) seems like a general morph optimization.

[TamarChristinaArm]

GetUpper/Lower() followed be a GetElement (including ToScalar()) seems like a general morph optimization.

Indeed, we also do some of the following, like Vector128<uint>.GetUpper().ToScalar() >> 16 would instead generate an UMOV on shorts instead by remapping the index. I don't know how common that is in .NET code, we typically see it during some image/video processing workloads. Just mentioning in case useful.

[TIHan]

Thank you for the information @TamarChristinaArm .

Closing this issue as we confirmed the optimization won't benefit. I'll leave it up to @tannergooding to create another issue regarding Vector128.GetUpper().AsScalar() if we plan to do that.