[Mono] Intrinsify API's of System.Numerics.Vector and new API's of System.Runtime.Intrinsics.Vector{64, 128} on Arm64

[fanyang-mono]

System.Runtime.Intrinsics.Vector{64, 128} and System.Numerics.Vector:

• IsHardwareAccelerated
  • arm64
• Abs - @mkhamoyan
  • arm64
• Add
  • arm64
• AndNot
  • arm64
• BitwiseAnd
  • arm64
• BitwiseOr
  • arm64
• Ceiling - @simonrozsival
  • arm64
• ConditionalSelect - @simonrozsival
  • arm64
• ConvertToDouble
  • arm64
• ConvertToInt32
  • arm64
• ConvertToInt64
  • arm64
• ConvertToSingle
  • arm64
• ConvertToUInt32
  • arm64
• ConvertToUInt64
  • arm64
• Divide
  • arm64
• Dot - @simonrozsival
  • arm64
• Equals - @simonrozsival
  • arm64
• EqualsAll - @simonrozsival
  • arm64
• EqualsAny - @simonrozsival
  • arm64
• Floor - @simonrozsival
  • arm64
• GreaterThan - @simonrozsival
  • arm64
• GreaterThanAll - @simonrozsival
  • arm64
• GreaterThanAny - @simonrozsival
  • arm64
• GreaterThanOrEqual - @simonrozsival
  • arm64
• GreaterThanOrEqualAll - @simonrozsival
  • arm64
• GreaterThanOrEqualAny - @simonrozsival
  • arm64
• LessThan - @simonrozsival
  • arm64
• LessThanAll - @simonrozsival
  • arm64
• LessThanAny - @simonrozsival
  • arm64
• LessThanOrEqual - @simonrozsival
  • arm64
• LessThanOrEqualAll - @simonrozsival
  • arm64
• LessThanOrEqualAny - @simonrozsival
  • arm64
• Max
  • arm64
• Min
  • arm64
• Multiply
  • arm64
• Narrow
  • arm64
• Negate - @mkhamoyan
  • arm64
• OnesComplement - @mkhamoyan
  • arm64
• Sqrt - @mkhamoyan
  • arm64
• Subtract
  • arm64
• Sum - @simonrozsival
  • arm64
• Widen - @simonrozsival
  • arm64
• Xor
  • arm64

System.Runtime.Intrinsics.Vector{64, 128}<T> and System.Numerics.Vector<T>:

• Operator + (one operand) - @simonrozsival
  • arm64
• Operator - (one operand)
  • arm64
• Operator ~
  • arm64
• Operator + @mkhamoyan
  • arm64
• Operator - @mkhamoyan
  • arm64
• Operator * @mkhamoyan
  • arm64
• Operator / @mkhamoyan
  • arm64
• Operator == - @simonrozsival
  • arm64
• Operator != - @simonrozsival
  • arm64
• Operator &
  • arm64
• Operator |
  • arm64
• Operator ^
  • arm64

I couldn't figure out the best area label to add to this issue. If you have write-permissions please help me learn by adding exactly one area label.

[fanyang-mono]

Motivation of this work:

There is a recent change to System.Numerics.Vector in the library code, which, as a result, caused regressions in Mono runtime (#53450). The regression could be observed via TechEmpower benchmark

and a few microbenchmarks, for example System.Tests.Perf_String.IndexOfAny

When comparing the generated code for the above microbenchmark test, I noticed the big generated code change for System.Numerics.Vector:Equals. Before the change, the IL for it was very small and simple. Mono was able to replace the deferred function System.Numerics.Vector<T>:Equals with an instrinsic. After the change, the IL for this same method became much longer with a few more calls to various complex methods, which ended up becoming way longer assembly code. Here are the codegen logs of a simplified app running with Mono runtime before and after this change

• Before https://gist.github.com/fanyang-mono/9641dcb7cb29b791d307552183999049
• After https://gist.github.com/fanyang-mono/950000e88a94d61a11ffab4ac26fb5f1
• The simplified app https://gist.github.com/fanyang-mono/c1b689ce6ba7def3cd01ad496e1fc7ab

The affected Mono flavors: JIT and AOT
The affected platforms: amd64 and arm64

The solution to this regression is to intrinsify the methods of System.Numerics.Vector.

[fanyang-mono]

Implementation wise, look at SN_Equals case for System.Numerics.Vector<T> for inspiration.

[tannergooding]

This should likely also extend to the APIs of the same names on Vector64, Vector128, and Vector256 (in the System.Runtime.Intrinsics namespace). It "should" hopefully be easy to add the support for both simultaneously.

[fanyang-mono]

This should likely also extend to the APIs of the same names on Vector64, Vector128, and Vector256 (in the System.Runtime.Intrinsics namespace). It "should" hopefully be easy to add the support for both simultaneously.

That work has been tracked here #42350. Most of the intrinsics for Vector64, Vector128 have been implemented in .NET6. The ones for Vector256 will be implemented in .NET7.

[tannergooding]

@fanyang-mono, these are new APIs introduced in .NET 7, so they wouldn't be covered by #42350