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

static ReadOnlySpan<byte> RVA => new byte[] { 1, 2, 3, 4, 5, 6, 7, 8 };

// Case1: read byte at a const position (2):
byte ReadByte() => RVA[2];


// Case2: Read Int32 at the give const offset (4)
int ReadInt() =>
    Unsafe.ReadUnaligned<int>(
        ref Unsafe.Add(ref MemoryMarshal.GetReference(RVA), 4));


// Case3: IsWhiteSpace is implemented using RVA table under its hood
bool IsWhiteSpace() => char.IsWhiteSpace(' ');

; Assembly listing for method Prog:ReadByte():ubyte:this
       48B8AA2B11DB96010000 mov      rax, 0x196DB112BAA
       0FB600               movzx    rax, byte  ptr [rax]
       C3                   ret      
 

; Assembly listing for method Prog:ReadInt():int:this
       48B8AC2B11DB96010000 mov      rax, 0x196DB112BAC
       8B00                 mov      eax, dword ptr [rax]
       C3                   ret      


; Assembly listing for method Prog:IsWhiteSpace():bool:this
       33C0                 xor      eax, eax
       F6057723AD5E80       test     byte  ptr [(reloc 0x7ffadf84d5f0)], 128
       0F95C0               setne    al
       C3                   ret      

; Method Prog:ReadByte():ubyte
       B803000000           mov      eax, 3
       C3                   ret      


; Method Prog:ReadInt():int
       B805060708           mov      eax, 0x8070605
       C3                   ret      


; Method Prog:IsWhiteSpace():bool
       B801000000           mov      eax, 1
       C3                   ret     

@jkotas does VM side look good otherwise?

VM/AOT side looks good.

@jakobbotsch @SingleAccretion when you have time can you please take a look at the JIT side? A couple of notes:
I enabled GTF_ICON_STATIC_HDL local propagation for AOT - I basically repeated Andy's change here (thanks to Jakob who found that change) - just a few negative diffs in the crossgen collection (see here)

I needed that because currently we can't obtain FieldSeq* out of a VN (it might be a good idea to encode that info to VN like Single suggested but it's more work for little value IMO).

The jit-diffs are small from this change (just 8 methods improved) but I plan a couple of optimizations on top of this in C#. Also, e.g. a lot of Char apis can now be folded into a constant for constant input (e.g. after inlining).

/azp list

/azp run runtime-coreclr outerloop, runtime-coreclr jitstress, runtime-coreclr jitstressregs

Fun fact: if we intrinsify BitOperations.Log2 (to fold constant input) we can fold this thing to a const 🙂: