[lewing]

PackedSpanHelpers narrows two Vector128<ushort> into one Vector128<byte> to search 16 chars per compare, but was gated on Sse2 only, so WebAssembly always took the non-packed path. PackedSimd.ConvertNarrowingSaturateUnsigned (i8x16.narrow_i16x8_u) has the same saturation semantics as Sse2.PackUnsignedSaturate, so the existing CanUsePackedIndexOf rules apply unchanged.

• PackedIndexOfIsSupported becomes Sse2 || PackedSimd; PackSources/LoadPacked128 narrow with PackedSimd, using explicit IsSupported chains so ILLink removes dead branches.
• string.Split and SingleStringSearchValuesPackedThreeChars use the packed path on wasm.
• [CompExactlyDependsOn(typeof(PackedSimd))] lets CoreCLR wasm R2R precompile these methods.

This applies to Mono (interpreter and AOT) and CoreCLR R2R; the CoreCLR interpreter reports PackedSimd.IsSupported == false.

CoreCLR browser-wasm R2R, new/base	len=16	len=1024
span.IndexOf(char)	1.07	0.60
span.IndexOfAny(c0, c1, c2)	1.17	0.71
SearchValues<string>	1.05	0.61

Short inputs cost about 1–2 ns per call on Arm hosts, where V8 and Cranelift lower the narrow and bitmask to multi-instruction sequences; on x64 they map to packuswb/pmovmskb (not measured).

Full benchmark table and Arm details

Apple M-series host under Node/V8, five interleaved rounds per build, medians; ratio = new / baseline.

Benchmark	len=16	len=64	len=1024
span.IndexOf(char)	1.07	1.01	0.60
span.IndexOfAny(c0, c1)	1.10	1.02	0.70
span.IndexOfAny(c0, c1, c2)	1.17	1.01	0.71
span.IndexOfAnyExcept(char)	1.11	1.00	0.60
span.Contains(char)	1.04	1.11	0.63
string.IndexOf(char, OrdinalIgnoreCase)	1.05	0.96	0.64
SearchValues<char> (3 values)	1.13	1.05	0.73
SearchValues<string> (1 value)	1.05	1.04	0.61
string.Split (3 separators, dense)	1.07	1.06	1.03
string.Replace(string, string) (1-char old value, dense)	1.04	1.07	1.15

Mono browser-wasm shows the same shape: AOT 0.66–0.80 at len=1024 and 1.09–1.27 at len=16; the interpreter 0.65–0.85 at most sizes. Results break even around 64–96 chars.

On aarch64, V8 (d8 --print-wasm-code) and Cranelift lower i8x16.narrow_i16x8_u to a dependent sqxtun/sqxtun2 pair and i8x16.bitmask to a two-pmull + trn2 sequence; i16x8.bitmask needs one pmull. A possible follow-up keeps the packed any_true loop but computes the match index from the unpacked 16-bit vectors, as MakeSeparatorListVectorized does.

Related: #135869 (other CompExactlyDependsOn(PackedSimd) gaps), #135892 (crossgen2 doesn't fold box T; isinst X).

Note

This PR description was drafted with GitHub Copilot assistance.

[azure-pipelines]

Azure Pipelines:
Successfully started running 3 pipeline(s).
13 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[dotnet-policy-service]

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