SIMD: AVX-512 VBMI as primary path for all byte/sbyte sizes

[jonathanpeppers]

Summary

Make AVX-512 VBMI the primary SIMD path for all byte/sbyte sizes (8-64), with AVX2 as a fallback for sizes 8-32 on hardware without VBMI support.

Fixes #29

Motivation

Previously, byte/sbyte sizes ≤32 only had an AVX2 path using complex cross-lane shuffling (Permute2x128 + dual vpshufb + Or). Sizes 33-64 already used the much simpler VBMI PermuteVar64x8. This change makes VBMI the preferred path for all sizes, following the same primary/fallback pattern already used for:

• short/ushort/char: AVX-512 BW primary → AVX2 fallback
• double: AVX-512F primary → AVX2 fallback

Changes

All changes are in SimdX86Emitter.cs:

Area	Change
GetGuardCondition	Always returns Avx512Vbmi.IsSupported for 1-byte types (was conditional on size > 32)
CanEmitAvx2Fallback	Added byte/sbyte support for sizes 8-32
EmitAvx2Fallback	Routes byte types to new EmitByteAvx2 method
Emit()	Routes all byte sizes to EmitByteAvx512Vbmi
EmitByte → EmitByteAvx2	Renamed; generates SortSimdAvx2_ fallback methods
EmitByteAvx512Vbmi	Extended to handle sizes 8-32 (Vector128/256 → Vector512 zero-extension)

Generated dispatch (example)

if (Avx512Vbmi.IsSupported) {
    if (n == 8)  { SortSimd8_byte(span); return; }   // VBMI Vector512
    if (n == 16) { SortSimd16_byte(span); return; }  // VBMI Vector512
    if (n == 48) { SortSimd48_byte(span); return; }  // VBMI Vector512
}
else if (Avx2.IsSupported) {
    if (n == 8)  { SortSimdAvx2_8_byte(span); return; }   // AVX2 Vector256
    if (n == 16) { SortSimdAvx2_16_byte(span); return; }  // AVX2 Vector256
}
if (AdvSimd.Arm64.IsSupported) { ... }
// scalar fallback

Benchmark Results (AMD EPYC 9V74, AVX-512 VBMI)

byte

Size	ArraySort	GeneratedSort	Speedup
23	1,028 ns	55 ns	19x
27	1,250 ns	53 ns	24x
28	1,415 ns	54 ns	26x
32	1,516 ns	54 ns	28x
34	1,759 ns	64 ns	27x

sbyte

Size	ArraySort	GeneratedSort	Speedup
27	1,355 ns	57 ns	24x
28	1,495 ns	58 ns	26x
32	1,598 ns	58 ns	28x
38	2,160 ns	68 ns	32x

All elements fit in a single Vector512<byte> with PermuteVar64x8 shuffles. Zero allocations. On CPUs without VBMI, sizes 8-32 fall back to AVX2.

Testing

All 455 tests pass across all four CI platforms (ubuntu x64, ubuntu ARM, windows, macOS).

[Copilot]

Pull request overview

This PR updates the x86 SIMD code generator so byte/sbyte sorting networks prefer AVX-512 VBMI across the full supported size range, while still generating AVX2 fallbacks for smaller sizes on machines without VBMI. It fits into the generator’s existing pattern of “newer AVX-512 primary path, older AVX2 fallback” used for other element widths.

Changes:

• Switched byte/sbyte primary x86 guard/dispatch from size-dependent AVX2-or-VBMI logic to VBMI for all supported byte widths.
• Added AVX2 fallback emission for byte/sbyte sizes 8-32, including distinct SortSimdAvx2_* method generation.
• Extended the VBMI byte emitter to load/store sizes 8-32 by zero-extending smaller vectors into Vector512<byte>.