All Vector<T> operators use software fallbacks when AVX is not presented

[EgorBo]

Filing an issue on @stephentoub's behalf:

int Count() => Vector<int>.Count;

Vector<int> Add(Vector<int> a, Vector<int> b) => a + b;

Codegen on a CPU with AVX2:

; Method Tests:Count():int:this (FullOpts)
 mov eax, 8
 ret 

; Method Tests:Add
 vmovups ymm0, ymmword ptr [r8]
 vpaddd ymm0, ymm0, ymmword ptr [r9]
 vmovups ymmword ptr [rdx], ymm0
 mov rax, rdx
 vzeroupper 
 ret 

Looks perfect, Vector<> is 256bit as expected. Now we if compile it without AVX (e.g. DOTNET_EnableAVX=0):

; Method Tests:Count():int:this (FullOpts)
 mov eax, 4
 ret 

; Method Tests:Add
G_M64079_IG01: ;; offset=0000H
 push rdi
 push rsi
 push rbp
 push rbx
 sub rsp, 88
 mov rsi, rdx
G_M64079_IG02: ;; offset=000BH
 movups xmm0, xmmword ptr [r8]
 movups xmmword ptr [rsp+38H], xmm0
 movups xmm0, xmmword ptr [r9]
 movups xmmword ptr [rsp+28H], xmm0
 xor edi, edi
G_M64079_IG03: ;; offset=001FH
 test edi, edi
 jl SHORT G_M64079_IG05
G_M64079_IG04: ;; offset=0023H
 xor ecx, ecx
 cmp edi, 4
 setl cl
 test cl, cl
 jne SHORT G_M64079_IG06
G_M64079_IG05: ;; offset=002FH
 mov rdx, 0x1EB80200008 ; ''
 mov rcx, rdx
 call [System.Diagnostics.Debug:Fail(System.String,System.String)]
G_M64079_IG06: ;; offset=0042H
 lea rcx, bword ptr [rsp+38H]
 movsxd rbx, edi
 mov ebp, dword ptr [rcx+4*rbx]
 test edi, edi
 jl SHORT G_M64079_IG08
G_M64079_IG07: ;; offset=0051H
 xor ecx, ecx
 cmp edi, 4
 setl cl
 test cl, cl
 jne SHORT G_M64079_IG09
G_M64079_IG08: ;; offset=005DH
 mov rdx, 0x1EB80200008 ; ''
 mov rcx, rdx
 call [System.Diagnostics.Debug:Fail(System.String,System.String)]
G_M64079_IG09: ;; offset=0070H
 lea rcx, bword ptr [rsp+28H]
 add ebp, dword ptr [rcx+4*rbx]
 test edi, edi
 jl SHORT G_M64079_IG11
G_M64079_IG10: ;; offset=007CH
 xor ecx, ecx
 cmp edi, 4
 setl cl
 test cl, cl
 jne SHORT G_M64079_IG12
G_M64079_IG11: ;; offset=0088H
 mov rdx, 0x1EB80200008 ; ''
 mov rcx, rdx
 call [System.Diagnostics.Debug:Fail(System.String,System.String)]
G_M64079_IG12: ;; offset=009BH
 lea rax, bword ptr [rsp+48H]
 mov dword ptr [rax+4*rbx], ebp
 inc edi
 cmp edi, 4
 jl G_M64079_IG03
G_M64079_IG13: ;; offset=00AEH
 movups xmm0, xmmword ptr [rsp+48H]
 movups xmmword ptr [rsi], xmm0
 mov rax, rsi
G_M64079_IG14: ;; offset=00B9H
 add rsp, 88
 pop rbx
 pop rbp
 pop rsi
 pop rdi
 ret 

we get 128-bit Vector<T> (expected) and software fallback for all operators - it feels like they're marked as "needs avx2".

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

Issue Details

---

Filing an issue on @stephentoub's behalf:

int Count() => Vector<int>.Count;

Vector<int> Add(Vector<int> a, Vector<int> b) => a + b;

Codegen on a CPU with AVX2:

; Method Tests:Count():int:this (FullOpts)
 mov eax, 8
 ret 

; Method Tests:Add
 vmovups ymm0, ymmword ptr [r8]
 vpaddd ymm0, ymm0, ymmword ptr [r9]
 vmovups ymmword ptr [rdx], ymm0
 mov rax, rdx
 vzeroupper 
 ret 

Looks perfect, Vector<> is 256bit as expected. Now we if compile it without AVX (e.g. DOTNET_EnableAVX=0):

; Method Tests:Count():int:this (FullOpts)
 mov eax, 4
 ret 

; Method Tests:Add
G_M64079_IG01: ;; offset=0000H
 push rdi
 push rsi
 push rbp
 push rbx
 sub rsp, 88
 mov rsi, rdx
G_M64079_IG02: ;; offset=000BH
 movups xmm0, xmmword ptr [r8]
 movups xmmword ptr [rsp+38H], xmm0
 movups xmm0, xmmword ptr [r9]
 movups xmmword ptr [rsp+28H], xmm0
 xor edi, edi
G_M64079_IG03: ;; offset=001FH
 test edi, edi
 jl SHORT G_M64079_IG05
G_M64079_IG04: ;; offset=0023H
 xor ecx, ecx
 cmp edi, 4
 setl cl
 test cl, cl
 jne SHORT G_M64079_IG06
G_M64079_IG05: ;; offset=002FH
 mov rdx, 0x1EB80200008 ; ''
 mov rcx, rdx
 call [System.Diagnostics.Debug:Fail(System.String,System.String)]
G_M64079_IG06: ;; offset=0042H
 lea rcx, bword ptr [rsp+38H]
 movsxd rbx, edi
 mov ebp, dword ptr [rcx+4*rbx]
 test edi, edi
 jl SHORT G_M64079_IG08
G_M64079_IG07: ;; offset=0051H
 xor ecx, ecx
 cmp edi, 4
 setl cl
 test cl, cl
 jne SHORT G_M64079_IG09
G_M64079_IG08: ;; offset=005DH
 mov rdx, 0x1EB80200008 ; ''
 mov rcx, rdx
 call [System.Diagnostics.Debug:Fail(System.String,System.String)]
G_M64079_IG09: ;; offset=0070H
 lea rcx, bword ptr [rsp+28H]
 add ebp, dword ptr [rcx+4*rbx]
 test edi, edi
 jl SHORT G_M64079_IG11
G_M64079_IG10: ;; offset=007CH
 xor ecx, ecx
 cmp edi, 4
 setl cl
 test cl, cl
 jne SHORT G_M64079_IG12
G_M64079_IG11: ;; offset=0088H
 mov rdx, 0x1EB80200008 ; ''
 mov rcx, rdx
 call [System.Diagnostics.Debug:Fail(System.String,System.String)]
G_M64079_IG12: ;; offset=009BH
 lea rax, bword ptr [rsp+48H]
 mov dword ptr [rax+4*rbx], ebp
 inc edi
 cmp edi, 4
 jl G_M64079_IG03
G_M64079_IG13: ;; offset=00AEH
 movups xmm0, xmmword ptr [rsp+48H]
 movups xmmword ptr [rsi], xmm0
 mov rax, rsi
G_M64079_IG14: ;; offset=00B9H
 add rsp, 88
 pop rbx
 pop rbp
 pop rsi
 pop rdi
 ret 

we get 128-bit Vector<T> (expected) and software fallback for all operators - it feels like they're marked as "needs avx2".

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