Suboptimal struct zeroing in case of gc pointers

[EgorBo]

public struct MyStruct
{
    public long a;
    public long b;
    public long c;
    public long d;
    public long e;
    public long f;
    public long g;
    public long h;
    public long i;
    public long j;
    public long k;
    public long l;
    public long m;
    public long n;
    public long o;
    public string s;
}

public MyStruct Foo()
{
    MyStruct s = default;
    return s;
}

; Method Program:Foo():Program+MyStruct:this
       33C0                 xor      eax, eax
       488902               mov      qword ptr [rdx], rax
       48894208             mov      qword ptr [rdx+08H], rax
       48894210             mov      qword ptr [rdx+10H], rax
       48894218             mov      qword ptr [rdx+18H], rax
       48894220             mov      qword ptr [rdx+20H], rax
       48894228             mov      qword ptr [rdx+28H], rax
       48894230             mov      qword ptr [rdx+30H], rax
       48894238             mov      qword ptr [rdx+38H], rax
       48894240             mov      qword ptr [rdx+40H], rax
       48894248             mov      qword ptr [rdx+48H], rax
       48894250             mov      qword ptr [rdx+50H], rax
       48894258             mov      qword ptr [rdx+58H], rax
       48894260             mov      qword ptr [rdx+60H], rax
       48894268             mov      qword ptr [rdx+68H], rax
       48894270             mov      qword ptr [rdx+70H], rax
       48894278             mov      qword ptr [rdx+78H], rax
       488BC2               mov      rax, rdx
       C3                   ret      
; Total bytes of code: 69

JIT gives up on SIMDs because of gc pointers in a struct while it could use SIMD only for unmanaged part

[ghost]

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

Issue Details

---

public struct MyStruct
{
    public long a;
    public long b;
    public long c;
    public long d;
    public long e;
    public long f;
    public long g;
    public long h;
    public long i;
    public long j;
    public long k;
    public long l;
    public long m;
    public long n;
    public long o;
    public string s;
}

public MyStruct Foo()
{
    MyStruct s = default;
    return s;
}

; Method Program:Foo():Program+MyStruct:this
       33C0                 xor      eax, eax
       488902               mov      qword ptr [rdx], rax
       48894208             mov      qword ptr [rdx+08H], rax
       48894210             mov      qword ptr [rdx+10H], rax
       48894218             mov      qword ptr [rdx+18H], rax
       48894220             mov      qword ptr [rdx+20H], rax
       48894228             mov      qword ptr [rdx+28H], rax
       48894230             mov      qword ptr [rdx+30H], rax
       48894238             mov      qword ptr [rdx+38H], rax
       48894240             mov      qword ptr [rdx+40H], rax
       48894248             mov      qword ptr [rdx+48H], rax
       48894250             mov      qword ptr [rdx+50H], rax
       48894258             mov      qword ptr [rdx+58H], rax
       48894260             mov      qword ptr [rdx+60H], rax
       48894268             mov      qword ptr [rdx+68H], rax
       48894270             mov      qword ptr [rdx+70H], rax
       48894278             mov      qword ptr [rdx+78H], rax
       488BC2               mov      rax, rdx
       C3                   ret      
; Total bytes of code: 69

JIT gives up on SIMDs because of gc pointers in a struct while it could use SIMD only for unmanaged part

Author:	EgorBo
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[EgorBo]

Arm64 does use SIMD in this case

; Method MyBenchmarks:Foo():MyBenchmarks+MyStruct:this
G_M35785_IG01:              ;; offset=0000H
        A9BF7BFD          stp     fp, lr, [sp, #-0x10]!
        910003FD          mov     fp, sp
        4F00E410          movi    v16.16b, #0
        AD004110          stp     q16, q16, [x8]
        AD014110          stp     q16, q16, [x8, #0x20]
        AD024110          stp     q16, q16, [x8, #0x40]
        AD034110          stp     q16, q16, [x8, #0x60]
        F900411F          str     xzr, [x8, #0x80]
        A8C17BFD          ldp     fp, lr, [sp], #0x10
        D65F03C0          ret     lr

Presumably because stp is atomic

[jakobbotsch]

Is this one fixed @EgorBo?

[EgorBo]

Is this one fixed @EgorBo?

Nope, the codegen is the same