Suboptimal codegen for memset/memcpy unrolling

[EgorBo]

When JIT unrolls memset/memcpy it does suboptimal decisions for certain sizes, e.g. to memset 30 bytes:

struct MyStruct
{
    fixed byte a[30];
}

MyStruct Test()
{
    MyStruct s = default; 
    return s;
}

  xor      eax, eax
  vxorps   xmm0, xmm0
  vmovdqu  xmmword ptr [rdx], xmm0
  mov      qword ptr [rdx+10H], rax
  mov      qword ptr [rdx+16H], rax

so to zero 30 bytes it uses GPR twice. It's better to keep using SIMD and overlap with previously zeroed part:

  vxorps   xmm0, xmm0, xmm0
  vmovups  xmmword ptr [rdx+ 14], xmm0
  vmovups  xmmword ptr [rdx], xmm0

Etc for other sizes.

PS: it seems that arm64 is doing the right thing here

[ghost]

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

Issue Details

---

When JIT unrolls memset/memcpy it does suboptimal decisions for certain sizes, e.g. to memset 30 bytes:

struct MyStruct
{
    fixed byte a[30];
}

MyStruct Test()
{
    MyStruct s = default; 
    return s;
}

  xor      eax, eax
  vxorps   xmm0, xmm0
  vmovdqu  xmmword ptr [rdx], xmm0
  mov      qword ptr [rdx+10H], rax
  mov      qword ptr [rdx+16H], rax

so to zero 30 bytes it uses GPR twice. It's better to keep using SIMD and overlap with previously zeroed part:

  vxorps   xmm0, xmm0, xmm0
  vmovups  xmmword ptr [rdx+ 14], xmm0
  vmovups  xmmword ptr [rdx], xmm0

Etc for other sizes

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