[release/8.0-staging] Ensure that the Create(Dot(...)) optimization doesn't kick in for Vector2 pre SSE4.1

[github-actions]

Backport of #96951 to release/8.0-staging

/cc @tannergooding

Customer Impact

• Customer reported
• Found internally

This can result in incorrect computations if you are on a machine that pre-dates SSE4.1 support (a machine with a CPU from before ~Nov 2007)

Regression

• Yes
• No

On older hardware (pre-SSE4.1) use of the Create(Dot(...)) optimization would kick in for Vector2. However, since there are two elements, a multiply followed by a single pairwise addition is emitted and the operation isn't naturally broadcasting on this older hardware. This is in contrast to newer hardware where it can utilize the dot product instruction which can always broadcast the result.

Testing

A regression test covering the scenario was added.

Risk

Low. The issue is well understood and the hardware this impacts is increasingly small. For cloud providers, most default to AVX2 capabable hardware or newer (~Jun 2013). For consumer hardware, places such as the Steam Hardware Survey (see under Other Settings) report 99.57% of customers have SSE4.1 support.

This wasn't surfacing in scenarios such as crossgen/r2r because SSE4.1 support is opportunistically enabled by default.

[ghost]

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

Issue Details

---

Backport of #96951 to release/8.0-staging

/cc @tannergooding

Customer Impact

Testing

Risk

IMPORTANT: If this backport is for a servicing release, please verify that:

• The PR target branch is release/X.0-staging, not release/X.0.

• If the change touches code that ships in a NuGet package, you have added the necessary package authoring and gotten it explicitly reviewed.

Author:	github-actions[bot]
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[tannergooding]

CC. @EgorBo

[danmoseley]

Does this affect Arm?

[tannergooding]

Does this affect Arm?

No.

x86/x64 only supports V128 and therefore you get HorizontalAdd(vector, vector) producing the following:

[
    vector[0] + vector[1],
    vector[2] + vector[3],
    vector[0] + vector[1],
    vector[2] + vector[3]
]

This is problematic for Vector2 since it only uses elements 0 and 1 and so therefore the second element will be incorrect (typically 0).

However, Arm64 supports V64 and so you get AddPairwise(vector, vector) producing the following, so the result remains correct:

[
    vector[0] + vector[1],
    vector[0] + vector[1]
]

[tannergooding]

@jeffschwMSFT, @JulieLeeMSFT was this not approved in servicing review or was there some other issue blocking it?

This was a customer reported codegen bug as indicated in the summary.

[jeffschwMSFT]

This just needs a code review and then we can take it to tactics. I removed the servicing-consider as it was hanging out for multiple potential reviews.

[tannergooding]

@EgorBo could you give the review here since you signed off on the original fix?

[jeffschwMSFT]

approved. we will take for consideration in 8.0.x

[tannergooding]

CC. @EgorBo, @dotnet/jit-contrib for merging. Monday February 12th is the Code Complete deadline for the March Release, so it needs to be merged by then (Julie asked that I let one of the JIT reviewers do the actual merge here).