JIT: suboptimal generated code for struct overlapped field manipulation

[hez2010]

Description

Repro (compile with roslyn features/ref-fields branch):

using System.Diagnostics.CodeAnalysis;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;

var color = new Color { R = 1, G = 2, B = 3, A = 4 };
color.Raw[2] = 3;
color.SRaw[1] = 4;
Console.WriteLine(color.G);

[StructLayout(LayoutKind.Explicit)]
struct Color
{
    [FieldOffset(0)] public byte R;
    [FieldOffset(1)] public byte G;
    [FieldOffset(2)] public byte B;
    [FieldOffset(3)] public byte A;

    [FieldOffset(0)] public int Rgba;

    [UnscopedRef] public ColorView<byte> Raw => new(ref this);
    [UnscopedRef] public ColorView<short> SRaw => new(ref this);
}

ref struct ColorView<T> where T : unmanaged
{
    private ref Color color;

    public ColorView(ref Color color)
    {
        this.color = ref color;
    }
    private static ref T Throw() => throw new IndexOutOfRangeException();

    public ref T this[uint index]
    {
        [MethodImpl(MethodImplOptions.AggressiveInlining)]
        get
        {
            unsafe
            {
                return ref sizeof(T) * index >= sizeof(Color) ?
                    ref Throw() :
                    ref Unsafe.Add(ref Unsafe.As<Color, T>(ref color), index);
            }
        }
    }
}

Codegen for Main (tier1):

G_M61237_IG01:  ;; offset=0000H
       sub      rsp, 56
						;; size=4 bbWeight=1 PerfScore 0.25

G_M61237_IG02:  ;; offset=0004H
       xor      ecx, ecx
       mov      dword ptr [rsp+28H], ecx
       mov      byte  ptr [rsp+28H], 1
       mov      byte  ptr [rsp+29H], 2
       mov      byte  ptr [rsp+2AH], 3
       mov      byte  ptr [rsp+2BH], 4
       mov      ecx, dword ptr [rsp+28H]
       mov      dword ptr [rsp+30H], ecx
       lea      rcx, bword ptr [rsp+30H]
       add      rcx, 2
       mov      byte  ptr [rcx], 3
       lea      rcx, bword ptr [rsp+30H]
       mov      word  ptr [rcx+02H], 4
       movzx    rcx, byte  ptr [rsp+31H]
       call     [System.Console:WriteLine(int)]
       nop      
						;; size=69 bbWeight=1 PerfScore 14.75

G_M61237_IG03:  ;; offset=0049H
       add      rsp, 56
       ret      
						;; size=5 bbWeight=1 PerfScore 1.25
; Total bytes of code: 78

G_M56640_IG02 can be folded into a constant.

Expected codegen of G_M56640_IG02:

mov ecx, 2
call [System.Console:WriteLine(int)]

Clang codegen: https://godbolt.org/z/W7416EsrT

Configuration

.NET: build from main

category:cq
theme:structs
skill-level:expert
cost:large
impact:medium

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

Issue Details

---

Description

Repro (compile with roslyn features/ref-fields branch):

using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;

var color = new Color { R = 1, G = 2, B = 3, A = 4 };
color.Raw[2] = 3;
color.SRaw[1] = 4;
Console.WriteLine(color.G);

[StructLayout(LayoutKind.Explicit)]
struct Color
{
    [FieldOffset(0)] public byte R;
    [FieldOffset(1)] public byte G;
    [FieldOffset(2)] public byte B;
    [FieldOffset(3)] public byte A;

    [FieldOffset(0)] public int Rgba;

    public ColorView<byte> Raw => new(ref this);
    public ColorView<short> SRaw => new(ref this);
}

ref struct ColorView<T> where T : unmanaged
{
    private ref Color color;
    public ColorView(ref Color color)
    {
        this.color = ref color;
    }
    private static ref T Throw() => throw new IndexOutOfRangeException();
    
    public ref T this[uint index]
    {
        [MethodImpl(MethodImplOptions.AggressiveInlining)]
        get
        {
            unsafe
            {
                return ref sizeof(T) * index >= sizeof(Color) ? 
                    ref Throw() : 
                    ref Unsafe.Add(ref Unsafe.AsRef<T>(Unsafe.AsPointer(ref color)), index);
            }
        }
    }
}

Codegen for Main (tier1):

G_M56640_IG01:
       sub      rsp, 56
                                                ;; size=4 bbWeight=1    PerfScore 0.25
G_M56640_IG02:
       xor      ecx, ecx
       mov      dword ptr [rsp+28H], ecx
       mov      byte  ptr [rsp+28H], 1
       mov      byte  ptr [rsp+29H], 2
       mov      byte  ptr [rsp+2AH], 3
       mov      byte  ptr [rsp+2BH], 4
       mov      ecx, dword ptr [rsp+28H]
       mov      dword ptr [rsp+30H], ecx
       lea      rcx, bword ptr [rsp+30H]
       add      rcx, 2
       mov      byte  ptr [rcx], 3
       lea      rcx, bword ptr [rsp+30H]
       add      rcx, 2
       mov      word  ptr [rcx], 4
       movzx    rcx, byte  ptr [rsp+31H]
       call     [System.Console:WriteLine(int)]
       nop
                                                ;; size=72 bbWeight=1    PerfScore 15.00
G_M56640_IG03:
       add      rsp, 56
       ret

Expected codegen of G_M56640_IG02:

mov ecx, 2
call [System.Console:WriteLine(int)]

Clang codegen: https://godbolt.org/z/W7416EsrT

Configuration

.NET: build from main

Author:	hez2010
Assignees:	-
Labels:	tenet-performance, area-CodeGen-coreclr, untriaged
Milestone:	-

[AndyAyersMS]

The jit historically has been super-cautious about structs with overlapping fields because they are a hallmark of aliasing. But now that we have the beginnings of a physical alias model (thanks to #68712) we might be able to revisit some of this.

cc @dotnet/jit-contrib

[jakobbotsch]

#76928 would solve the basic issues around this by making promotion more "physical" instead of based upon struct layout.

[jakobbotsch]

One major blocker here is that we end up address exposing color because we end up with separate statements for all the block copies between different temps for ColorView, and the inlined constructor that assigns the field to the address of it. We would somehow need to collapse all of this at an early point to realize that it ends up immediately dereferenced. That seems quite hard.

[MichalPetryka]

ref Unsafe.AsRef<T>(Unsafe.AsPointer(ref color))

FYI this is a GC hole, use Unsafe.As<Color, T> instead.

[hez2010]

ref Unsafe.AsRef<T>(Unsafe.AsPointer(ref color))

FYI this is a GC hole, use Unsafe.As<Color, T> instead.

Thanks for pointing out. I have fixed the code in the original issue, while the codegen is still not changed.