JIT perf: ldind.u1 followed by conv.i1 results in suboptimal codegen

[GrabYourPitchforks]

Repro:

public static int Foo(int i)
{
    ReadOnlySpan<byte> span = new byte[] { 1, 2, 0xFF };
    return (sbyte)span[i];
}

x64 codegen:

C.Foo(Int32)
    L0000: sub rsp, 0x28
    L0004: cmp ecx, 3
    L0007: jae short L0023
    L0009: movsxd rax, ecx
    L000c: mov rdx, 0x1a13a680930
    L0016: movzx eax, byte ptr [rax+rdx]
    L001a: movsx rax, al
    L001e: add rsp, 0x28
    L0022: ret
    L0023: call 0x00007ffc4814ec40
    L0028: int3

Note the movsx instruction on L001a. Ideally this would be folded into the instruction above to produce the single instruction movsx eax, byte ptr [rax + rdx].

category:performance
theme:codegen
skill-level:intermediate
cost:medium
impact:medium

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

Issue Details

---

Repro:

public static int Foo(int i)
{
    ReadOnlySpan<byte> span = new byte[] { 1, 2, 0xFF };
    return (sbyte)span[i];
}

x64 codegen:

C.Foo(Int32)
    L0000: sub rsp, 0x28
    L0004: cmp ecx, 3
    L0007: jae short L0023
    L0009: movsxd rax, ecx
    L000c: mov rdx, 0x1a13a680930
    L0016: movzx eax, byte ptr [rax+rdx]
    L001a: movsx rax, al
    L001e: add rsp, 0x28
    L0022: ret
    L0023: call 0x00007ffc4814ec40
    L0028: int3

Note the movsx instruction on L001a. Ideally this would be folded into the instruction above to produce the single instruction movsx eax, byte ptr [rax + rdx].

Author:	GrabYourPitchforks
Assignees:	-
Labels:	tenet-performance, area-CodeGen-coreclr
Milestone:	-

[JulieLeeMSFT]

cc @dotnet/jit-contrib.
Moving to future since we have less than 2 months to .NET 7 code complete.