new Vector<T> for byte, short emits vpbroadcastd thus needs an imul

[gfoidl]

public Vector<byte> Bad(byte val) => new Vector<byte>(val);

generates on x64 with AVX2 supported:

C.Bad(Byte)
    L0000: vzeroupper
    L0003: movzx eax, r8b
    L0007: imul eax, 0x1010101
    L000d: vmovd xmm0, eax
    L0011: vpbroadcastd ymm0, xmm0
    L0016: vmovupd [rdx], ymm0
    L001a: mov rax, rdx
    L001d: vzeroupper
    L0020: ret

The imul isn't needed when instead of vpbroadcastd the vpbroadcastb is used, as it's done by

public Vector256<byte> Good(byte val) => Vector256.Create(val);

C.Good(Byte)
    L0000: vzeroupper
    L0003: movzx eax, r8b
    L0007: vmovd xmm0, eax
    L000b: vpbroadcastb ymm0, xmm0
    L0010: vmovupd [rdx], ymm0
    L0014: mov rax, rdx
    L0017: vzeroupper
    L001a: ret

Type	actual	should be
byte, sbyte	imul + vpbroadcastd	vpbroadcastb
ushort, short	imul + vpbroadcastd	vpbroadcastw
uint, int	vpbroadcastd	✔️
ulong, long	vpbroadcastq	✔️

For more see sharplab
I see the same on .NET 7 Preview 2.

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

Issue Details

---

public Vector<byte> Bad(byte val) => new Vector<byte>(val);

generates on x64 with AVX2 supported:

C.Bad(Byte)
    L0000: vzeroupper
    L0003: movzx eax, r8b
    L0007: imul eax, 0x1010101
    L000d: vmovd xmm0, eax
    L0011: vpbroadcastd ymm0, xmm0
    L0016: vmovupd [rdx], ymm0
    L001a: mov rax, rdx
    L001d: vzeroupper
    L0020: ret

The imul isn't needed when instead of vpbroadcastd the vpbroadcastb is used, as it's done by

public Vector256<byte> Good(byte val) => Vector256.Create(val);

C.Good(Byte)
    L0000: vzeroupper
    L0003: movzx eax, r8b
    L0007: vmovd xmm0, eax
    L000b: vpbroadcastb ymm0, xmm0
    L0010: vmovupd [rdx], ymm0
    L0014: mov rax, rdx
    L0017: vzeroupper
    L001a: ret

Type	actual	should be
byte, sbyte	imul + vpbroadcastd	vpbroadcastb
ushort, short	imul + vpbroadcastd	vpbroadcastw
uint, int	vpbroadcastd	✔️
ulong, long	vpbroadcastq	✔️

For more see sharplab
I see the same on .NET 7 Preview 2.

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

[tannergooding]

Hmmm. These should be going down the same path. I'd guess there is something missing in simdashwintrinsic.cpp during import.