Optimize IND<SIMD>(RVA) - #81651

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

// Mainly for:
static Vector128<byte> Test1() => Vector128.Create("0123456789ABCDEF"u8);


// But various ROS<> patterns should work too:
static ReadOnlySpan<byte> rva => new byte[] {1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16};
static Vector128<byte> Test2() => Unsafe.ReadUnaligned<Vector128<byte>>(ref MemoryMarshal.GetReference(rva));

; Method P:Test1():System.Runtime.Intrinsics.Vector128`1[ubyte]
       vzeroupper 
       vmovups  xmm0, xmmword ptr [reloc @RWD00]
       vmovups  xmmword ptr [rcx], xmm0
       mov      rax, rcx
       ret      
; Total bytes of code: 19


; Method P:Test2():System.Runtime.Intrinsics.Vector128`1[ubyte]
       vzeroupper 
       vmovups  xmm0, xmmword ptr [reloc @RWD00]
       vmovups  xmmword ptr [rcx], xmm0
       mov      rax, rcx
       ret      
; Total bytes of code: 19

@jakobbotsch @SingleAccretion PTAL (since you reviewed previous PRs in this function)

Does this optimization need to be implemented in Mono as well to be something one can depend on?

Does this optimization need to be implemented in Mono as well to be something one can depend on?

From my understanding - it's not necessary if it's more complicated than this (I spent like 15 minutes on this). We replace a load from the RVA data with a load from the Data Section. It's just that for RyuJIT it's now imported as GT_CNS_VEC so it can participate in various foldings and be better understandable for CSE.

I need to check Mono's codegen, but perhaps LLVM can do the same with constant global data itself? (depends on how LLVM IR is formed I assume)

Failure is #75244