So this optimization just repeats what we do in morph, but this time with help of global assertions.

Diffs - no size improvements on arm64 due to fixed size encoding, but perf wise mov should be better than sxtw (the latter can't be handled via register renaming, also, on some hw it has worse TP).

cc @dotnet/jit-contrib

@MihuBot

@EgorBo Will this enable the removal of tricks to force zero-extension like the following:

@EgorBo Will this enable the removal of tricks to force zero-extension like the following:

runtime/src/libraries/System.Private.CoreLib/src/System/MemoryExtensions.cs

Line 38 in 670db5c

return new Span<T>(ref Unsafe.Add(ref MemoryMarshal.GetArrayDataReference(array), (nint)(uint)start /* force zero-extension */), array.Length - start);

I would expect so

@EgorBo Will this enable the removal of tricks to force zero-extension like the following:

runtime/src/libraries/System.Private.CoreLib/src/System/MemoryExtensions.cs

Line 38 in 670db5c

return new Span<T>(ref Unsafe.Add(ref MemoryMarshal.GetArrayDataReference(array), (nint)(uint)start /* force zero-extension */), array.Length - start);

I would expect so

@EgorBo Not quite: https://csharp.godbolt.org/z/fq15ch7EM