Improve codegen for Vector128.Shift* operations where a direct intrinsic is not available

[MihaZupan]

(applies to Vector256 as well)

Consider Vector128.ShiftRightLogical(ref byte) where X86 does not have a ShiftRightLogical instruction that operates on bytes:

Vector128<byte> v0 = Vector128.LoadUnsafe(ref source);
Vector128<byte> v1 = Vector128.ShiftRightLogical(v0, 4);

Which currently emits a scalar fallback

TestClass.Foo(Byte ByRef)
    L0000: push rsi
    L0001: sub rsp, 0x40
    L0005: vzeroupper
    L0008: vmovdqu xmm0, [rcx]
    L000c: vmovapd [rsp+0x20], xmm0
    L0012: xor esi, esi
    L0014: lea rcx, [rsp+0x20]
    L0019: movsxd rdx, esi
    L001c: movzx ecx, byte ptr [rcx+rdx]
    L0020: mov edx, 4
    L0025: mov rax, 0x7ffa0845bc60
    L002f: call qword ptr [rax]
    L0031: lea rdx, [rsp+0x30]
    L0036: movsxd rcx, esi
    L0039: mov [rdx+rcx], al
    L003c: inc esi
    L003e: cmp esi, 0x10
    L0041: jl short L0014
    L0043: vmovapd xmm0, [rsp+0x30]
    L0049: vpmovmskb eax, xmm0
    L004d: add rsp, 0x40
    L0051: pop rsi
    L0052: ret

where it could instead emit a 32-bit shift and an AND to clear the overlapping bits

Vector128<byte> v0 = Vector128.LoadUnsafe(ref source);
Vector128<byte> v1 = Vector128.ShiftRightLogical(v0.AsInt32(), 4).AsByte() & Vector128.Create((byte)0xF);

TestClass.Bar(Byte ByRef)
    L0000: vzeroupper
    L0003: vmovdqu xmm0, [rcx]
    L0007: vpsrld xmm0, xmm0, 4
    L000c: vpand xmm0, xmm0, [0x7ffa087600d0]
    L0014: vpmovmskb eax, xmm0
    L0018: ret

We have a few places in runtime that are aware of this issue and employ workarounds, e.g.:

• runtime/src/libraries/System.Private.CoreLib/src/System/IndexOfAnyValues/IndexOfAnyAsciiSearcher.cs

  Line 875 in c1abf87

  : Sse2.ShiftRightLogical(source.AsInt32(), 4).AsByte() & Vector128.Create((byte)0xF);

• runtime/src/libraries/System.Private.CoreLib/src/System/Buffers/Text/Base64Decoder.cs

  Line 594 in dc6ad37

  Vector128<byte> hiNibbles = Vector128.ShiftRightLogical(str.AsInt32(), 4).AsByte() & mask2F;

• https://github.com/dotnet/runtime/blob/8482f562a8b5d96bb0a0fb201bfabea7e5e6b115/src/libraries/System.Private.CoreLib/src/System/IndexOfAnyValues/ProbabilisticMap.cs#L168-L170

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

Issue Details

---

Consider Vector128.ShiftRightLogical(ref byte) where X86 does not have a ShiftRightLogical instruction that operates on bytes:

Vector128<byte> v0 = Vector128.LoadUnsafe(ref source);
Vector128<byte> v1 = Vector128.ShiftRightLogical(v0, 4);

Which currently emits a scalar fallback

TestClass.Foo(Byte ByRef)
    L0000: push rsi
    L0001: sub rsp, 0x40
    L0005: vzeroupper
    L0008: vmovdqu xmm0, [rcx]
    L000c: vmovapd [rsp+0x20], xmm0
    L0012: xor esi, esi
    L0014: lea rcx, [rsp+0x20]
    L0019: movsxd rdx, esi
    L001c: movzx ecx, byte ptr [rcx+rdx]
    L0020: mov edx, 4
    L0025: mov rax, 0x7ffa0845bc60
    L002f: call qword ptr [rax]
    L0031: lea rdx, [rsp+0x30]
    L0036: movsxd rcx, esi
    L0039: mov [rdx+rcx], al
    L003c: inc esi
    L003e: cmp esi, 0x10
    L0041: jl short L0014
    L0043: vmovapd xmm0, [rsp+0x30]
    L0049: vpmovmskb eax, xmm0
    L004d: add rsp, 0x40
    L0051: pop rsi
    L0052: ret

where it could instead emit a 32-bit shift and an AND to clear the overlapping bits

Vector128<byte> v0 = Vector128.LoadUnsafe(ref source);
Vector128<byte> v1 = Vector128.ShiftRightLogical(v0.AsInt32(), 4).AsByte() & Vector128.Create((byte)0xF);

TestClass.Bar(Byte ByRef)
    L0000: vzeroupper
    L0003: vmovdqu xmm0, [rcx]
    L0007: vpsrld xmm0, xmm0, 4
    L000c: vpand xmm0, xmm0, [0x7ffa087600d0]
    L0014: vpmovmskb eax, xmm0
    L0018: ret

We have a few places in runtime that are aware of this issue and employ workarounds, e.g.:

• runtime/src/libraries/System.Private.CoreLib/src/System/IndexOfAnyValues/IndexOfAnyAsciiSearcher.cs

  Line 875 in c1abf87

  : Sse2.ShiftRightLogical(source.AsInt32(), 4).AsByte() & Vector128.Create((byte)0xF);

• runtime/src/libraries/System.Private.CoreLib/src/System/Buffers/Text/Base64Decoder.cs

  Line 594 in dc6ad37

  Vector128<byte> hiNibbles = Vector128.ShiftRightLogical(str.AsInt32(), 4).AsByte() & mask2F;

• https://github.com/dotnet/runtime/blob/5b6f7a27868e3a72ae1f993df16f41686fcc589b/src/libraries/System.Private.CoreLib/src/System/IndexOfAnyValues/ProbabilisticMap.cs#L166-L168

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

[gfoidl]

For

runtime/src/libraries/System.Private.CoreLib/src/System/Buffers/Text/Base64Decoder.cs

Line 594 in dc6ad37

Vector128<byte> hiNibbles = Vector128.ShiftRightLogical(str.AsInt32(), 4).AsByte() & mask2F;

just noting that with the proposed codegen a extra register for Vector128.Create((byte)0xF) is needed which was avoided by intention by re-using the already present 0x2F which has effectively the same masking effect.

I don't think any compiler is smart enough these days to have that knowledge / information in order to do such optimizations.
On the other hand I don't know how much perf would regress by using the simpler C#-code that this issue proposes.

[JulieLeeMSFT]

Assigning to @tannergooding to respond to the request.

[JulieLeeMSFT]

We will not have time to implement this code optimization in .NET8.
Cc @kunalspathak @TIHan.

[tannergooding]

Resolved with #111108