Suboptimal codgen for `Vector128.NarrowWithSaturation`

[xtqqczze]

I expect PackSources0 and PackSources1 to have exactly the same codegen in the following code snippet:

static Vector128<byte> PackSources1(Vector128<ushort> lower, Vector128<ushort> upper)
    => Vector128.NarrowWithSaturation(lower, upper);

static Vector128<byte> PackSources0(Vector128<ushort> lower, Vector128<ushort> upper)
    => Sse2.PackUnsignedSaturate(
        Vector128.Min(lower, Vector128.Create((ushort)255)).AsInt16(),
        Vector128.Min(upper, Vector128.Create((ushort)255)).AsInt16());

// coreclr trunk-20250611+5415b7342d44af9c974905760539f198fad13682

C:PackSources1(System.Runtime.Intrinsics.Vector128`1[ushort],System.Runtime.Intrinsics.Vector128`1[ushort]):System.Runtime.Intrinsics.Vector128`1[byte] (FullOpts):
       vbroadcastss xmm0, dword ptr [reloc @RWD00]
       vpminuw  xmm1, xmm0, xmmword ptr [rsp+0x08]
       vpand    xmm1, xmm1, xmm0
       vpminuw  xmm2, xmm0, xmmword ptr [rsp+0x18]
       vpand    xmm0, xmm2, xmm0
       vpackuswb xmm0, xmm1, xmm0
       vmovups  xmmword ptr [rdi], xmm0
       mov      rax, rdi
       ret      
RWD00  	dd	00FF00FFh		; 2.34184e-38

C:PackSources0(System.Runtime.Intrinsics.Vector128`1[ushort],System.Runtime.Intrinsics.Vector128`1[ushort]):System.Runtime.Intrinsics.Vector128`1[byte] (FullOpts):
       vbroadcastss xmm0, dword ptr [reloc @RWD00]
       vpminuw  xmm1, xmm0, xmmword ptr [rsp+0x08]
       vpminuw  xmm0, xmm0, xmmword ptr [rsp+0x18]
       vpackuswb xmm0, xmm1, xmm0
       vmovups  xmmword ptr [rdi], xmm0
       mov      rax, rdi
       ret      
RWD00  	dd	00FF00FFh		; 2.34184e-38

https://csharp.godbolt.org/z/o39G8GP9T

Related: #115525

cc: @tannergooding

[dotnet-policy-service]

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

[MihaZupan]

Example use case:

runtime/src/libraries/System.Private.CoreLib/src/System/SearchValues/IndexOfAnyAsciiSearcher.cs

Lines 1362 to 1374 in 8bcc9e2

	// Replace with Vector128.NarrowWithSaturation once https://github.com/dotnet/runtime/issues/75724 is implemented.
	[MethodImpl(MethodImplOptions.AggressiveInlining)]
	[CompExactlyDependsOn(typeof(Sse2))]
	[CompExactlyDependsOn(typeof(PackedSimd))]
	public static Vector128<byte> PackSources(Vector128<ushort> lower, Vector128<ushort> upper)
	{
	Vector128<short> lowerMin = Vector128.Min(lower, Vector128.Create((ushort)255)).AsInt16();
	Vector128<short> upperMin = Vector128.Min(upper, Vector128.Create((ushort)255)).AsInt16();
	if (Sse2.IsSupported)
	{
	return Sse2.PackUnsignedSaturate(lowerMin, upperMin);
	}

[tannergooding]

This has to do with deferring down to gtNewSimdNarrowNode and it not understanding that the value is already restricted to being within range.

We'd either need to modify that method to take in a new parameter indicating that additional logic should be skipped -or- duplicate the underlying narrowing instruction selection instead.

[tannergooding]

I likely won't get to this in the next couple weeks. Contributions are welcome.

[TechPizzaDev]

What's the point of the Min(x, 255) here when using Sse2.PackUnsignedSaturate? I thought SSE2 already did the correct thing without further clamping?
I realized right after asking that these are unsigned overloads, and SSE2 only takes signed int16...