Optimize Vector<>.op_Division with constant divisor

[EgorBo]

From twitter: https://twitter.com/nietras1/status/1319546076756643842

static Vector<int> DivTest(Vector<int> vec) => vec / new Vector<int>(2);

Current codegen:

G_M28418_IG01:
       push     rdi
       push     rsi
       push     rbx
       sub      rsp, 144
       vzeroupper 
       mov      rsi, rcx
G_M28418_IG02:
       mov      ecx, 2
       vmovd    xmm0, ecx
       vpbroadcastd ymm0, ymm0
       vmovupd  ymm1, ymmword ptr[rdx]
       vmovupd  ymmword ptr[rsp+50H], ymm1
       vmovupd  ymmword ptr[rsp+30H], ymm0
       vxorps   ymm0, ymm0
       vmovupd  ymmword ptr[rsp+70H], ymm0
       xor      rdi, rdi
G_M28418_IG03:
       test     rdi, rdi
       jl       SHORT G_M28418_IG05
G_M28418_IG04:
       cmp      rdi, 8
       setl     cl
       movzx    rcx, cl
       jmp      SHORT G_M28418_IG06
G_M28418_IG05:
       xor      ecx, ecx
G_M28418_IG06:
       mov      rdx, 0xD1FFAB1E
       mov      rbx, gword ptr [rdx]
       mov      rdx, rbx
       mov      gword ptr [rsp+28H], rdx
       mov      rax, rdx
       test     cl, cl
       jne      SHORT G_M28418_IG08
G_M28418_IG07:
       mov      rcx, rax
       mov      rdx, rax
       call     System.Diagnostics.Debug:Fail(System.String,System.String)  ;; Checked config.
G_M28418_IG08:
       lea      rcx, bword ptr [rsp+50H]
       mov      ebx, dword ptr [rcx+4*rdi]
       test     rdi, rdi
       jl       SHORT G_M28418_IG10
G_M28418_IG09:
       cmp      rdi, 8
       setl     cl
       movzx    rcx, cl
       jmp      SHORT G_M28418_IG11
G_M28418_IG10:
       xor      ecx, ecx
G_M28418_IG11:
       mov      rdx, gword ptr [rsp+28H]
       mov      rax, rdx
       test     cl, cl
       jne      SHORT G_M28418_IG13
G_M28418_IG12:
       mov      rcx, rax
       mov      rdx, rax
       call     System.Diagnostics.Debug:Fail(System.String,System.String)
G_M28418_IG13:
       lea      rcx, bword ptr [rsp+30H]
       mov      edx, dword ptr [rcx+4*rdi]
       mov      ecx, ebx
       call     System.Numerics.Vector`1[Int32][System.Int32]:ScalarDivide(int,int):int ;; not inlined?
       mov      ebx, eax
       test     rdi, rdi
       jl       SHORT G_M28418_IG15
G_M28418_IG14:
       cmp      rdi, 8
       setl     cl
       movzx    rcx, cl
       jmp      SHORT G_M28418_IG16
G_M28418_IG15:
       xor      ecx, ecx
G_M28418_IG16:
       mov      rdx, gword ptr [rsp+28H]
       test     cl, cl
       jne      SHORT G_M28418_IG18
G_M28418_IG17:
       mov      rcx, rdx
       call     System.Diagnostics.Debug:Fail(System.String,System.String)
G_M28418_IG18:
       lea      rax, bword ptr [rsp+70H]
       mov      dword ptr [rax+4*rdi], ebx
       inc      rdi
       cmp      rdi, 8
       jl       G_M28418_IG03
G_M28418_IG19:
       vmovupd  ymm0, ymmword ptr[rsp+70H]
       vmovupd  ymmword ptr[rsi], ymm0
       mov      rax, rsi
G_M28418_IG20:
       vzeroupper 
       add      rsp, 144
       pop      rbx
       pop      rsi
       pop      rdi
       ret      
; Total bytes of code 267

Expected codegen:

DivTest: 
        vpsrld  ymm1, ymm0, 31
        vpaddd  ymm0, ymm0, ymm1
        vpsrad  ymm0, ymm0, 1
        ret

There are no instructions in x86 to perform a division for vectors of integers, but it makes sense to share the "Magic Division" logic we have for scalars for these Vectors since we expose operator / (Users assume it's accelerated just like other Vector APIs)
At least for power-of-two numbers.

Bonus points:

1. Revise current implementation, it looks less efficient than https://godbolt.org/z/81r66h (Loops are not unrolled?)
2. File an API proposal for operator / scalar (e.g. operator * has a scalar overload, division does not)

/cc @tannergooding

[ghost]

Tagging subscribers to this area: @tannergooding, @pgovind, @jeffhandley
See info in area-owners.md if you want to be subscribed.