Fix Avx2 helpers

[sbomer]

Unblocks dotnet-optimization flow and #84711. Includes a dotnet-optimization update to confirm this it avoids the issue discussed in #84635. This fixes that issue by preventing R2R for any Avx2 helpers which assume they are called with Avx2 enabled.

@davidwrighton taking a shot at fixing this by skipping R2R for the instances that I was able to find by inspection [edit: using the platform analyzer with some tweaks]. Just a draft for now since I'm not sure if this is the approach we want to take - I just wanted to see what happens in ci.

[MihaZupan]

Just to confirm I understand the rules for S.P.CoreLib, a method like Vector128.ShuffleUnsafe may end up calling the Shuffle fallback when running on ARM/WASM, but it will never call the fallback on a platform with Ssse3 support, because that is treated differently?

I'm asking about ShuffleUnsafe specifically because the callers do expect different behavior depending on the platform.

[sbomer]

That matches my read of the rules, since they mention R2R code checks for Ssse3 support at runtime (whereas Avx2 is assumed to be unavailable), but @davidwrighton would know best.

[davidwrighton]

@sbomer I believe this looks good for now. I'm working on a more permanent fix, that will likely result in changes to all of these places to a slightly different form, but getting a fix in before the snap would be better.

@MihaZupan The "rules" indicate that the ShuffleUnsafe method has a problem, but as @sbomer notes, there are also some implementation choices in the current default behavior of Crossgen2 which make the ShuffleUnsafe method reliably functional.

[davidwrighton]

Looks good for now. I'm working on an analyzer we can turn on all the time, and some crossgen2 tweaks to make it easier to write these things. As it is the BypassReadyToRun changes will pessimize some of the work done for the ASP.NET composite container scenario, but I prioritize working over not working. @dotnet/crossgen-contrib

[sbomer]

Wasm failures look unrelated. I filed #85354 for these.

[MihaZupan]

Is something like this also problematic?

runtime/src/libraries/System.Private.CoreLib/src/System/IndexOfAnyValues/IndexOfAnyAsciiSearcher.cs

Lines 822 to 825 in 4ab5265

	Vector128<byte> source =
	Sse2.IsSupported ? Sse2.PackUnsignedSaturate(source0, source1) :
	AdvSimd.IsSupported ? AdvSimd.ExtractNarrowingSaturateUpper(AdvSimd.ExtractNarrowingSaturateLower(source0.AsUInt16()), source1.AsUInt16()) :
	PackedSimd.ConvertNarrowingUnsignedSaturate(source0, source1);

Could we end up calling into PackedSimd on ARM for example?

[davidwrighton]

@MihaZupan, technically ... yes, except you'd have to pass a very surprising set of switches to crossgen2, which nobody is likely to do. I've a followon fix (#85481) which should make this very hard to mess up, and will protect against that kind of case.