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

public class SIMDStringToLower
{
    public static string ToLowerCase(string input)
    {
        if (input == null)
        {
            throw new ArgumentNullException(nameof(input));
        }

        int length = input.Length;
        char[] result = new char[length];

        if (Sse2.IsSupported)
        {
            int simdLength = Vector128<ushort>.Count;
            int simdIterations = length / simdLength;
            Vector128<ushort> upperCaseMask = Vector128.Create((ushort)0x0020);

            for (int i = 0; i < simdIterations * simdLength; i += simdLength)
            {
                Vector128<ushort> inputVector = Unsafe.ReadUnaligned<Vector128<ushort>>(ref Unsafe.As<char, byte>(ref input[i]));
                Vector128<ushort> upperCaseMaskApplied = Sse2.Add(inputVector, upperCaseMask);
                Vector128<ushort> isInRange = Sse2.And(Sse2.CompareGreaterThan(inputVector, Vector128.Create((ushort)0x0040)), Sse2.CompareLessThan(inputVector, Vector128.Create((ushort)0x005b)));

                Vector128<ushort> resultVector = Sse2.Or(Sse2.And(isInRange, upperCaseMaskApplied), Sse2.AndNot(isInRange, inputVector));

                Unsafe.WriteUnaligned(ref Unsafe.As<char, byte>(ref result[i]), resultVector);
            }

            // Process the remaining characters
            for (int i = simdIterations * simdLength; i < length; i++)
            {
                char c = input[i];
                result[i] = (c >= 'A' && c <= 'Z') ? (char)(c | 0x20) : c;
            }
        }
        else
        {
            for (int i = 0; i < length; i++)
            {
                char c = input[i];
                result[i] = (c >= 'A' && c <= 'Z') ? (char)(c | 0x20) : c;
            }
        }

        return new string(result);
    }
}