Use double VectorTableLookup on ARM in ProbabilisticMap

[MihaZupan]

Contributes to #84328

Before

Method	Mean	Error
IndexOfAnyProb	3.776 µs	0.0024 µs

After

Method	Mean	Error
IndexOfAnyProb	2.576 µs	0.0011 µs

[ghost]

Tagging subscribers to this area: @dotnet/area-system-buffers
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Contributes to #84328

When I vectorized this path in #80963, I ran benchmarks on ARM with the initial version of the PR, but later regressed it with 2295370 by using the >>> operator, which apparently isn't turned into AdvSimd.ShiftRightLogical. The first commit 3bfd7d3 fixes that, changing perf from ~36 µs back to ~3.7 µs on this benchmark. cc: @tannergooding

The change to use VectorTableLookup with two table vectors further improves performance:

Before

Method	Mean	Error
IndexOfAnyProb	3.776 µs	0.0024 µs

After

Method	Mean	Error
IndexOfAnyProb	2.576 µs	0.0011 µs

Author:	MihaZupan
Assignees:	MihaZupan
Labels:	area-System.Buffers
Milestone:	8.0.0

[danmoseley]

the >>> operator, which apparently isn't turned into AdvSimd.ShiftRightLogical.

Bug?

[MihaZupan]

Likely just not implemented yet, #75770 may be the right tracking issue for that.

[kunalspathak]

cc: @a74nh

[a74nh]

Code looks good to me. Out of curiosity, do you know roughly how much of the performance gain comes from the VectorTableLookup and how much comes from the ShiftRightLogical ?

[tannergooding]

but later regressed it with 2295370 by using the >>> operator, which apparently isn't turned into AdvSimd.ShiftRightLogical

Can you log a bug for this. Looks like a simple fix as we should just need to recognize the operators in addition to the named methods: https://github.com/dotnet/runtime/blob/main/src/coreclr/jit/hwintrinsicarm64.cpp#L1520-L1555

[MihaZupan]

how much of the performance gain comes from the VectorTableLookup

The 3.776 µs => 2.576 µs change on this benchmark is entirely due to VectorTableLookup.

Can you log a bug for this.

#85257

[kunalspathak]

The 3.776 µs => 2.576 µs change on this benchmark is entirely due to VectorTableLookup.

That's 30% gain. Good to see that.

[kunalspathak]

LGTM

[MihaZupan]

Failures are known according to Build Analysis