Investigate emitting `movq` for the `CreateScalarUnsafe` helper intrinsics that take a `long/ulong` on x86

[tannergooding]

For the CreateScalarUnsafe helper intrinsics that take a long/ulong, it may be beneficial to emit the movq xmm, [mem] instruction on x86, rather than falling back to emulating the behavior in software.

Given that the input is a regular long/ulong, but on x86 the instruction requires the second operand come from memory, this may require some additional work to make happen or we may find it not worthwhile to do at all.

category:cq
theme:hardware-intrinsics
skill-level:beginner
cost:small

[tannergooding]

CC. @CarolEidt, @fiigii

As per https://github.com/dotnet/coreclr/pull/21351/files/ff31bc6a186a15cd9395928c546e82de077f7246#r239645439

[CarolEidt]

@tannergooding - thanks for capturing this.

[anthonycanino]

@tannergooding I think this is no longer applicable.

On SharpLab

using System;

using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.X86;

public class C {
    public Vector128<long> Foo(long u1) {
        return Vector128.CreateScalarUnsafe(u1);
    }
}

gives

C.Foo(Int64)
    L0000: vzeroupper
    L0003: vmovq xmm0, r8
    L0008: vmovupd [rdx], xmm0
    L000c: mov rax, rdx
    L000f: ret

[saucecontrol]

mov rax, rdx

that's x64 code.

here's x86

C.Foo(Int64)
    L0000: push dword ptr [esp+8]
    L0004: push dword ptr [esp+8]
    L0008: mov ecx, edx
    L000a: call System.Runtime.Intrinsics.Vector128.CreateScalarUnsafe(Int64)
    L000f: ret 8

[anthonycanino]

Got it --- thanks for that. My mistake.