Primitive AVX2 strlen implementation in C++

strlen.h

#pragma once

#include "common.h" // https://github.com/oliora/habr-switches-perf-test/blob/main/common.h
#include <cstdint>
#include <string_view>
#include <utility>
#include <immintrin.h>

template <size_t StepSize>
requires((StepSize <= PageSize)
    && (PageSize % StepSize == 0)
    && (StepSize % sizeof(long long) == 0))
inline size_t strlenVectorized(const char *input) noexcept {
    constexpr size_t WordsPerStep = StepSize / sizeof(__m256i);

    const auto start = input;

    // Process unaligned preamble in naive way
    {
        const auto headEnd = alignedAfter<StepSize>(input);
        while (input != headEnd && *input) {
            ++input;
        }
        if (input != headEnd) [[unlikely]] {
            return input - start;
        }
    }

    const __m256i null_c = _mm256_set1_epi8(0); // Fill all packed 8-bit integers with 0

    while (true) {
        __m256i eq_null;
        forEach<WordsPerStep>([&]<size_t Idx>(IndexConstant<Idx>) {
            const __m256i block = _mm256_load_si256(reinterpret_cast<const __m256i*>(input) + Idx);
            if constexpr (Idx == 0) {
                eq_null =  _mm256_cmpeq_epi8(null_c, block);
            } else {
                eq_null = _mm256_or_si256(eq_null, _mm256_cmpeq_epi8(null_c, block));
            }
        });

        __m128i step_null = _mm_or_si128(_mm256_extracti128_si256(eq_null, 1), *reinterpret_cast<const __m128i*>(&eq_null));
        step_null = _mm_or_si128(_mm_bsrli_si128(step_null, 8), step_null);
        const auto anyNull = _mm_extract_epi64(step_null, 0);
        if (anyNull) [[unlikely]] {
            // Found null character, calculate the remaining values char by char until we hit the null character
            while (*input) {
                ++input;
            }
            return input - start;
        }

        input += StepSize;
    }
}