Use crossplat vectors in GetPointerToFirstInvalidChar

[EgorBo]

This whole path is currently not used on ARM64 (it uses Vector<> instead).

Sse4.1 requirement was lifted to Sse2

[ghost]

Tagging subscribers to this area: @dotnet/area-system-runtime-intrinsics
See info in area-owners.md if you want to be subscribed.

Issue Details

---

This whole path is currently not used on ARM64 (it uses Vector<> instead).

Sse4.1 requirement was lifted to Sse2 because of Vector128.Min which is not bad for Sse2, enough for R2R:

Vector128<ushort> Min(Vector128<ushort> a, Vector128<ushort> b)
    => Vector128.Min(a, b);

; Method mytest:Min
       movups   xmm0, xmmword ptr [rdx]
       movups   xmm1, xmmword ptr [reloc @RWD00]
       paddw    xmm0, xmm1
       movups   xmm2, xmmword ptr [r8]
       paddw    xmm2, xmm1
       pminsw   xmm0, xmm2
       psubw    xmm0, xmm1
       movups   xmmword ptr [rcx], xmm0
       mov      rax, rcx
       ret      
RWD00  	dq	8000800080008000h, 8000800080008000h
; Total bytes of code: 37

Author:	EgorBo
Assignees:	-
Labels:	area-System.Runtime.Intrinsics
Milestone:	-

[EgorBo]

To be exposed as a public API with #82559

[stephentoub]

Why is Vector128.Min slower than Sse41.Min?

[EgorBo]

Vector128.Min uses fallback without SSE4.1:

Vector128<ushort> Min(Vector128<ushort> a, Vector128<ushort> b)
    => Vector128.Min(a, b);

; Method mytest:Min
       movups   xmm0, xmmword ptr [rdx]
       movups   xmm1, xmmword ptr [reloc @RWD00]
       paddw    xmm0, xmm1
       movups   xmm2, xmmword ptr [r8]
       paddw    xmm2, xmm1
       pminsw   xmm0, xmm2
       psubw    xmm0, xmm1
       movups   xmmword ptr [rcx], xmm0
       mov      rax, rcx
       ret      
RWD00  	dq	8000800080008000h, 8000800080008000h
; Total bytes of code: 37

it probably can be improved but Sse4.1 Min is a single instruction under a quick opportunistic check branch.

[EgorBo]

I am not sure R2R codegen that important though

[stephentoub]

Is that a bug in Vector128.Min then?

[EgorBo]

This pattern is only for R2R/NAOT which is compiled as SSE2, for JIT everything is great as is.
the idea to emit:

if (runtimeIsaCheck == 1)
{
    // use single-instruction for Min
}
else
{
    // slow fallback because there is no SSE4.1
}

but I decided to drop it because it looks like we have this problem already in many places where we only check Vector128.IsHardwareAccelerated and don't care what exactly Vector128.* APIs need under the hood.

[stephentoub]

Shouldn't we be changing the fallback Vector<> path to use Vector128<> instead, too?

EDIT: #90391

[EgorBo]

Shouldn't we be changing the fallback Vector<> path to use Vector128<> instead, too?

EDIT: #90391

Ah, I didn't notice you already ported it, will revert mine

[EgorBo]

wasm job is failing with dotnet/dnceng#450