Arm64: Better addressing mode for array access whose elements are accessed byref

[kunalspathak]

Continuing on #64819, we should also improve the array accesses if the element is accessed using ref.

private int[] _srcData = new int[20];

public int[] Test(int m, int n, byte v)
{
    for (int i = 0; i < n; i++) {
        Consume(ref _srcData[i]);
    }
    return _srcData;
}

G_M11249_IG03:
            mov     x0, x19
            ldr     x0, [x0,#8]
            ldr     w1, [x0,#8]
            cmp     w21, w1
            bhs     G_M11249_IG06
            add     x0, x0, #16
            ubfiz   x1, x21, #2, #32   
            add     x1, x0, x1           ; <-- can compact with previous instruction?
            mov     x0, x19
            movz    x2, #0xd1ffab1e
            movk    x2, #0xd1ffab1e LSL #16
            movk    x2, #0xd1ffab1e LSL #32
            ldr     x2, [x2]
            blr     x2
            add     w21, w21, #1
            cmp     w21, w20
            blt     G_M11249_IG03

Without ref:

private int[] _srcData = new int[20];

public int[] Test(int m, int n, byte v)
{
    for (int i = 0; i < n; i++) {
        Consume(_srcData[i]);
    }
    return _srcData;
}

G_M11249_IG03:
            mov     x0, x19
            ldr     x0, [x0,#8]
            ldr     w1, [x0,#8]
            cmp     w21, w1
            bhs     G_M11249_IG06
            add     x0, x0, #16
            ldr     w1, [x0, w21, UXTW #2] ; <-- compacted
            mov     x0, x19
            movz    x2, #0xd1ffab1e
            movk    x2, #0xd1ffab1e LSL #16
            movk    x2, #0xd1ffab1e LSL #32
            ldr     x2, [x2]
            blr     x2
            add     w21, w21, #1
            cmp     w21, w20
            blt     G_M11249_IG03

[ghost]

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

Issue Details

---

Continuing on #64819, we should also improve the array accesses if the element is accessed using ref.

private int[] _srcData = new int[20];

public int[] Test(int m, int n, byte v)
{
    for (int i = 0; i < n; i++) {
        Consume(ref _srcData[i]);
    }
    return _srcData;
}

G_M11249_IG03:
            mov     x0, x19
            ldr     x0, [x0,#8]
            ldr     w1, [x0,#8]
            cmp     w21, w1
            bhs     G_M11249_IG06
            add     x0, x0, #16
            ubfiz   x1, x21, #2, #32
            add     x1, x0, x1
            mov     x0, x19
            movz    x2, #0xd1ffab1e
            movk    x2, #0xd1ffab1e LSL #16
            movk    x2, #0xd1ffab1e LSL #32
            ldr     x2, [x2]
            blr     x2
            add     w21, w21, #1
            cmp     w21, w20
            blt     G_M11249_IG03

Without ref:

private int[] _srcData = new int[20];

public int[] Test(int m, int n, byte v)
{
    for (int i = 0; i < n; i++) {
        Consume(ref _srcData[i]);
    }
    return _srcData;
}

G_M11249_IG03:
            mov     x0, x19
            ldr     x0, [x0,#8]
            ldr     w1, [x0,#8]
            cmp     w21, w1
            bhs     G_M11249_IG06
            add     x0, x0, #16
            ldr     w1, [x0, w21, UXTW #2] ; <-- compact
            mov     x0, x19
            movz    x2, #0xd1ffab1e
            movk    x2, #0xd1ffab1e LSL #16
            movk    x2, #0xd1ffab1e LSL #32
            ldr     x2, [x2]
            blr     x2
            add     w21, w21, #1
            cmp     w21, w20
            blt     G_M11249_IG03

Author:	kunalspathak
Assignees:	-
Labels:	area-CodeGen-coreclr, untriaged
Milestone:	-

[kunalspathak]

@dotnet/jit-contrib

[EgorBo]

I don't understand the expected codegen - the first codegen doesn't contain any memory loads (because we don't dereference _srcData)

[kunalspathak]

I don't understand the expected codegen - the first codegen doesn't contain any memory loads (because we don't dereference _srcData)

Ah, I think the 2nd version had to be Consume(_srcData[i]);. In the first codegen, I was referring to the following (if that's what you are asking):

            ubfiz   x1, x21, #2, #32
            add     x1, x0, x1

[EgorBo]

@kunalspathak so can you tell what's the expected codegen here? I don't see loads and I don't see how we can combine these ubfix and add

[kunalspathak]

I was misreading that the ref example has memory load in ldr x2, [x1], but it doesn't need it.