Redundant `movsx` in span indexing

[xtqqczze]

I expect M1 and M2 to have exactly the same codegen in the following code snippet, but M1 has a redundant sign-extension.

static int M1(ref byte reference, int length)
{
    ReadOnlySpan<byte> s = MemoryMarshal.CreateReadOnlySpan(ref reference, length);
    return s.Length > 0 ? (sbyte)s[0] : -1;
}

static int M2(ref byte reference, int length)
{
    ReadOnlySpan<byte> s = MemoryMarshal.CreateReadOnlySpan(ref reference, length);
    return s.Length > 0 ? (sbyte)MemoryMarshal.GetReference(s) : -1;
}

// coreclr trunk-20250612+15c514180982522386424c7bface57564ad6724d

C:M1(byref,int):int (FullOpts):
       test     esi, esi
       jle      SHORT G_M3362_IG05
       movzx    rax, byte  ptr [rdi]
       movsx    rax, al
       ret      
G_M3362_IG05:  ;; offset=0x000C
       mov      eax, -1
       ret      

C:M2(byref,int):int (FullOpts):
       test     esi, esi
       jle      SHORT G_M35265_IG05
       movsx    rax, byte  ptr [rdi]
       ret      
G_M35265_IG05:  ;; offset=0x0009
       mov      eax, -1
       ret      

https://csharp.godbolt.org/z/6P4W3PPGf

[dotnet-policy-service]

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

[xtqqczze]

For benchmark data showing a measurable performance impact, see: #116612 (comment).

[JulieLeeMSFT]

cc @dotnet/jit-contrib .

[tannergooding]

For benchmark data showing a measurable performance impact,

It'd be great to remove the redundant movzx here, but it's worth noting that I wouldn't really call this "measurable".

The difference shown is around 1ns or less and is more likely general noise of the system than anything else. movzx is also "free" on most CPUs in the past 15+ years, barring some CPUs where register renaming was disabled, and so the general cost difference tends to come in minor form of decoding cost and impact to things like loop alignment or similar instead.

[xtqqczze]

The difference shown is around 1ns or less and is more likely general noise of the system than anything else. movzx is also "free" on most CPUs in the past 15+ years, barring some CPUs where register renaming was disabled, and so the general cost difference tends to come in minor form of decoding cost and impact to things like loop alignment or similar instead.

It's really the additional movsx after the memory load that is problematic, which unlike movzx cannot be eliminated by register renaming.

I've updated the issue description.