Span.Fill(0) vs Span.Clear()

[EgorBo]

static readonly int[] data = new int[1024];

[Benchmark]
[Arguments(1)]
[Arguments(8)]
[Arguments(32)]
[Arguments(100)]
public void Clear(int len) => data.AsSpan(0, len).Clear();

[Benchmark]
[Arguments(1)]
[Arguments(8)]
[Arguments(32)]
[Arguments(100)]
public void Fill(int len) => data.AsSpan(0, len).Fill(0);

| Method |  len |      Mean |
|------- |----- |----------:|
|  Clear |    1 |  1.898 ns |
|   Fill |    1 |  1.351 ns |

|  Clear |    8 |  1.614 ns |
|   Fill |    8 |  1.108 ns |

|  Clear |   32 |  2.949 ns |
|   Fill |   32 |  1.319 ns |

|  Clear |  100 |  3.099 ns |
|   Fill |  100 |  8.019 ns |

A couple of notes:

1. Clear() calls either InitBlockUnaligned (which is lowered to memset when span's size is not known at jit time) or pinvokes into memset (+ pinvoke machinery in this case). So even for small buffers we end up calling memset.
2. Fill(0) despite not using any external calls 2.5 times slower than memset for len=100
3. Fill(0) is never unrolled for constant spans, e.g.:

void Clear1(int[] data, int len) => data.AsSpan(0, 20).Clear();
void Clear2(int[] data, int len) => data.AsSpan(0, 20).Fill(0);

; Method Program:Clear1(int[],int):this
       4883EC28             sub      rsp, 40
       C5F877               vzeroupper 
       4885D2               test     rdx, rdx
       7421                 je       SHORT G_M58891_IG04
       837A0814             cmp      dword ptr [rdx+08H], 20
       721B                 jb       SHORT G_M58891_IG04
       4883C210             add      rdx, 16
       C5FC57C0             vxorps   ymm0, ymm0
       C5FE7F02             vmovdqu  ymmword ptr[rdx], ymm0 ;; unrolled & vectorized with AVX !
       C5FE7F4220           vmovdqu  ymmword ptr[rdx+20H], ymm0
       C5FA7F4240           vmovdqu  xmmword ptr [rdx+40H], xmm0
       4883C428             add      rsp, 40
       C3                   ret      
G_M58891_IG04:              
       FF15AD4D0E00         call     [System.ThrowHelper:ThrowArgumentOutOfRangeException()]
       CC                   int3     
; Total bytes of code: 52


; Method Program:Clear2(int[],int):this
       4883EC28             sub      rsp, 40
       4885D2               test     rdx, rdx
       741E                 je       SHORT G_M47240_IG04
       837A0814             cmp      dword ptr [rdx+08H], 20
       7218                 jb       SHORT G_M47240_IG04
       488D4A10             lea      rcx, bword ptr [rdx+10H]
       BA14000000           mov      edx, 20
       4533C0               xor      r8d, r8d
       FF15B7968000         call     [System.SpanHelpers:Fill[int](byref,ulong,int)]
       90                   nop      
       4883C428             add      rsp, 40
       C3                   ret      
G_M47240_IG04:              
       FF15634D0E00         call     [System.ThrowHelper:ThrowArgumentOutOfRangeException()]
       CC                   int3     
; Total bytes of code: 46

[Benchmark()]
public void Clear_32() => data.AsSpan(0, 32).Clear();

[Benchmark()]
public void Fill_32() => data.AsSpan(0, 32).Fill(0);

|   Method |      Mean |
|--------- |----------:|
| Clear_32 | 0.3312 ns |
|  Fill_32 | 1.2964 ns |

[ghost]

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

Issue Details

---

static readonly int[] data = new int[1024];

[Benchmark]
[Arguments(1)]
[Arguments(8)]
[Arguments(32)]
[Arguments(100)]
public void Clear(int len) => data.AsSpan(0, len).Clear();

[Benchmark]
[Arguments(1)]
[Arguments(8)]
[Arguments(32)]
[Arguments(100)]
public void Fill(int len) => data.AsSpan(0, len).Fill(0);

| Method |  len |      Mean |
|------- |----- |----------:|
|  Clear |    1 |  1.898 ns |
|   Fill |    1 |  1.351 ns |

|  Clear |    8 |  1.614 ns |
|   Fill |    8 |  1.108 ns |

|  Clear |   32 |  2.949 ns |
|   Fill |   32 |  1.319 ns |

|  Clear |  100 |  3.099 ns |
|   Fill |  100 |  8.019 ns |

A couple of notes:

1. Clear() calls either InitBlockUnaligned (which is lowered to memset when span's size is not known at jit time) or pinvokes into memset (+ pinvoke machinery in this case). So even for small buffers we end up calling memset.
2. Fill(0) despite not using any external calls 2.5 times slower than memset for len=100
3. Fill(0) is never unrolled for constant spans, e.g.:

void Clear1(int[] data, int len) => data.AsSpan(0, 20).Clear();
void Clear2(int[] data, int len) => data.AsSpan(0, 20).Fill(0);

; Method Program:Clear1(int[],int):this
       4883EC28             sub      rsp, 40
       C5F877               vzeroupper 
       4885D2               test     rdx, rdx
       7421                 je       SHORT G_M58891_IG04
       837A0814             cmp      dword ptr [rdx+08H], 20
       721B                 jb       SHORT G_M58891_IG04
       4883C210             add      rdx, 16
       C5FC57C0             vxorps   ymm0, ymm0
       C5FE7F02             vmovdqu  ymmword ptr[rdx], ymm0 ;; unrolled & vectorized with AVX !
       C5FE7F4220           vmovdqu  ymmword ptr[rdx+20H], ymm0
       C5FA7F4240           vmovdqu  xmmword ptr [rdx+40H], xmm0
       4883C428             add      rsp, 40
       C3                   ret      
G_M58891_IG04:              
       FF15AD4D0E00         call     [System.ThrowHelper:ThrowArgumentOutOfRangeException()]
       CC                   int3     
; Total bytes of code: 52


; Method Program:Clear2(int[],int):this
       4883EC28             sub      rsp, 40
       4885D2               test     rdx, rdx
       741E                 je       SHORT G_M47240_IG04
       837A0814             cmp      dword ptr [rdx+08H], 20
       7218                 jb       SHORT G_M47240_IG04
       488D4A10             lea      rcx, bword ptr [rdx+10H]
       BA14000000           mov      edx, 20
       4533C0               xor      r8d, r8d
       FF15B7968000         call     [System.SpanHelpers:Fill[int](byref,ulong,int)]
       90                   nop      
       4883C428             add      rsp, 40
       C3                   ret      
G_M47240_IG04:              
       FF15634D0E00         call     [System.ThrowHelper:ThrowArgumentOutOfRangeException()]
       CC                   int3     
; Total bytes of code: 46

[Benchmark()]
public void Clear_32() => data.AsSpan(0, 32).Clear();

[Benchmark()]
public void Fill_32() => data.AsSpan(0, 32).Fill(0);

|   Method |      Mean |
|--------- |----------:|
| Clear_32 | 0.3312 ns |
|  Fill_32 | 1.2964 ns |

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

[JulieLeeMSFT]

Pushing out to Future for this optimization work.