#include <Columns/ColumnString.h>


namespace DB
{

template <char not_case_lower_bound, char not_case_upper_bound>
struct LowerUpperImpl
{
    static void vector(const ColumnString::Chars_t & data,
        const ColumnString::Offsets & offsets,
        ColumnString::Chars_t & res_data,
        ColumnString::Offsets & res_offsets)
    {
        res_data.resize(data.size());
        res_offsets.assign(offsets);
        array(data.data(), data.data() + data.size(), res_data.data());
    }

    static void vector_fixed(const ColumnString::Chars_t & data, size_t /*n*/, ColumnString::Chars_t & res_data)
    {
        res_data.resize(data.size());
        array(data.data(), data.data() + data.size(), res_data.data());
    }

private:
    static void array(const UInt8 * src, const UInt8 * src_end, UInt8 * dst)
    {
        const auto flip_case_mask = 'A' ^ 'a';

#if __SSE2__
        const auto bytes_sse = sizeof(__m128i);
        const auto src_end_sse = src_end - (src_end - src) % bytes_sse;

        const auto v_not_case_lower_bound = _mm_set1_epi8(not_case_lower_bound - 1);
        const auto v_not_case_upper_bound = _mm_set1_epi8(not_case_upper_bound + 1);
        const auto v_flip_case_mask = _mm_set1_epi8(flip_case_mask);

        for (; src < src_end_sse; src += bytes_sse, dst += bytes_sse)
        {
            /// load 16 sequential 8-bit characters
            const auto chars = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src));

            /// find which 8-bit sequences belong to range [case_lower_bound, case_upper_bound]
            const auto is_not_case
                = _mm_and_si128(_mm_cmpgt_epi8(chars, v_not_case_lower_bound), _mm_cmplt_epi8(chars, v_not_case_upper_bound));

            /// keep `flip_case_mask` only where necessary, zero out elsewhere
            const auto xor_mask = _mm_and_si128(v_flip_case_mask, is_not_case);

            /// flip case by applying calculated mask
            const auto cased_chars = _mm_xor_si128(chars, xor_mask);

            /// store result back to destination
            _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), cased_chars);
        }
#endif

        for (; src < src_end; ++src, ++dst)
            if (*src >= not_case_lower_bound && *src <= not_case_upper_bound)
                *dst = *src ^ flip_case_mask;
            else
                *dst = *src;
    }
};

}