FindFirstNonZeroByte (AVX)

FindFirstNonZeroByteAVX.cs

// Ugly pure ASM version of https://github.com/aspnet/KestrelHttpServer/pull/1138 and gist https://gist.github.com/benaadams/2dd3f99230757111e91915f638067a09
// Not performing better though...

[Benchmark]
public int TestFindFirstByteAVX()
{
	var vector = _vectors[ByteSet];
	return FindFirstByteAVX(ref vector);
}

[MethodImpl(MethodImplOptions.NoInlining)]
public static int FindFirstByteAVX(ref Vector<byte> byteEquals)
{
	throw new InvalidOperationException();
}

static unsafe FirstByteBenchmark()
{
	// ASSUME THAT Vector<byte>.Count == 32 (AVX)
	var method = typeof(FirstByteBenchmark).GetMethod("FindFirstByteAVX", BindingFlags.Public | BindingFlags.Static);
	var methodHandle = method.MethodHandle;
	RuntimeHelpers.PrepareMethod(methodHandle);
	var functionPtr = methodHandle.GetFunctionPointer();
	
	//auto mask = _mm256_setzero_si256();
	//auto val = _mm256_loadu_si256((const __m256i*)byteEquals);
	//auto result = _mm256_cmpeq_epi8(val, mask);
	//auto a = ~_mm256_movemask_epi8(result);
	//auto index = _tzcnt_u32(a);
	//return index == 32 ? -1 : index;

	//   0000f	c5 f1 ef c9	 vpxor	 xmm1, xmm1, xmm1
	//   00013	c5 f5 74 09	 vpcmpeqb ymm1, ymm1, YMMWORD PTR [rcx]
	//   00017	c5 fd d7 c1	 vpmovmskb eax, ymm1
	//   0001b	f7 d0		 not	 eax
	//   0001d	83 c9 ff	 or	 ecx, -1			; ffffffffH
	//   00020	f3 0f bc c0	 tzcnt	 eax, eax
	//   00024	83 f8 20	 cmp	 eax, 32			; 00000020H
	//   00027	0f 44 c1	 cmove	 eax, ecx
	//   00032	c3           ret	 0 
	byte[] patch = new byte[]
	{
		0xc5, 0xf1, 0xef, 0xc9,
		0xc5, 0xf5, 0x74, 0x09,
		0xc5, 0xfd, 0xd7, 0xc1,
		0xf7, 0xd0,
		0x83, 0xc9, 0xff,
		0xf3, 0x0f, 0xbc, 0xc0,
		0x83, 0xf8, 0x20,
		0x0f, 0x44, 0xc1,
		0xc3
	};
	for (int i = 0; i < patch.Length; i++)
	{
		((byte*)functionPtr)[i] = patch[i];
	}
}