Used inline SIMD vectors if they are constants

[gfoidl]

Prerequisites

• I have written a descriptive pull-request title
• I have verified that there are no overlapping pull-requests open
• I have verified that I am following the existing coding patterns and practice as demonstrated in the repository. These follow strict Stylecop rules 👮.
• I have provided test coverage for my change (where applicable)

Description

Starting with .NET 5 constant SIMD vectors can be used / created "inline", so that the JIT can read the values from the data section. Thus the need of static readonly fields for these kind of vectors is gone.

Codegen (x64) was spot-checked. Basically it's

-      mov       rax,19D98002F38
-      mov       rax,[rax]
-      vmovupd   xmm8,[rax+8]
+      vmovupd   xmm0,[7FFF734BC5F0]

For loops or other code that isn't tiered compiled by the JIT, the static init check needed by the runtime is also elided with this change.

In micro-benchmarks that change allone I expect to be just below noise, as all data is in the cache and the branch-predictor (static init check) will be trainined. The difference could (hopefully be should) be visible in real-world apps.

Note: for micro-benchmarks #2121 will have an effect, see machine code comparison over there.

Fixes #1762
Contributes to #2121

Patterns applied

• If a constant vector is used only once outside a loop, then Vector{128,256}.Create is called in-place of the argument. That way the JIT is able to produce code like

  vmulps reg, reg, [mem]

  and no additional (logical) register is needed.

• If a constant vector is used once inside a loop, then the Vector{128,256}.Create is hoisted above the loop(s), as the JIT (currently) won't do this for us.

• If a constant vector is used several times outside a loop, then Vector{128,256}.Create is assigned to a vector variable as early as possible (to hide the latency of loading from mem), but not too far away to keep the logical context of the code.

[gfoidl]

Instead of the static constructor (cctor) this method is now called from the static readonly field.
Avoiding a cctor and using beforefieldinit (IL) is the prefered way in dotnet/runtime too, as there are several pro reasons for it, liking timing, static init check, etc.

That's the only other code change not mentioned in PR's description.

[CLAassistant]

All committers have signed the CLA.

[JimBobSquarePants]

@gfoidl

Yep, see #2117

Not sure what the cause is yet. It's only happening with the latest .NET 7 preview 4.

If you update your branch from main those tests will be skipped for now on that platform.

[gfoidl]

The CI leg for windows-latest is failing with

...ImageDifferenceIsOverThresholdException : Image difference is over threshold!
Test Environment OS : Windows
Test Environment is CI : True
Test Environment is .NET Core : True
Report ImageFrame 0:
Total difference: 29.9761%
...

Local tests are passing (VS, standard checkout as given in ReadMe).
Please give me some advice on how to investigate this.

[gfoidl]

@JimBobSquarePants thanks, now we had a race condition.
The short comment (now deleted) GH created while copying the error message (don't know why).

Thanks!

[JimBobSquarePants]

Sorry for the delay in reviewing. Been focusing on one of the other libraries.

This all looks good to me. I think I'll still follow up with a PR to introduce a complete set of constants for MMShuffle though.

[JimBobSquarePants]

I'm merging this. It's the same Skew test failing each time as noted in #2117