On Sat, Oct 4, 2025 at 4:49 AM Walt Karas wrote:
>
> Response: Sorry, I still don't understand. Can you show me how you would rework this code?


I threw this code together quickly so there's a few oversights in it
(you can get it to crash with different numbers so there's a bug
somewhere), but anyway what you want might be something along the
following lines:

     https://godbolt.org/z/q8oMobhrx

And here it is copy-pasted:

#include <cassert> // assert
#include <array> // array
#include <atomic> // atomic
#include <optional> // optional
#include <set> // set
#include <stdexcept> // runtime_error
#include <thread> // thread
#include <utility> // forward
#include <vector> // vector
#include <mutex> // mutex, lock_guard

template<typename T, unsigned max_threads>
class ThreadLocal final {

    struct FreeNode {
        unsigned index;
        FreeNode *next;
    };

    static inline std::array< std::optional<T>, max_threads > storage{};
    static inline std::atomic<unsigned> nextIndex{0};
    static inline std::atomic<FreeNode*> freeList{nullptr};
    static inline std::array<FreeNode, max_threads> nodes{};

    static unsigned acquireSlot(void) noexcept(false)
    {
        FreeNode *head = freeList.load(std::memory_order_acquire);

        while ( head )
        {
            FreeNode *const next = head->next;
            if ( false == freeList.compare_exchange_weak(head, next,
std::memory_order_acq_rel) ) continue;
            return head->index;
        }

        unsigned const index = nextIndex.fetch_add(1u,
std::memory_order_relaxed);
        if ( index >= max_threads ) throw std::runtime_error("Exceeded
max_threads limit!");
        return index;
    }

    static void releaseSlot(unsigned const index) noexcept
    {
        FreeNode *const node = &nodes[index];
        FreeNode *oldHead = freeList.load(std::memory_order_relaxed);

        do node->next = oldHead;
        while ( false == freeList.compare_exchange_weak(oldHead, node,
std::memory_order_release, std::memory_order_relaxed) );

        storage[index].reset();
    }

    template<typename... Params>
    unsigned registerThread(Params&&... args) const noexcept(false)
    {
        static thread_local unsigned const myIndex = acquireSlot();

        // Automatically recycle the slot when the thread exits
        struct Reclaimer {
            unsigned idx;
            ~Reclaimer(void) noexcept { releaseSlot(idx); }
        };

        static thread_local Reclaimer const reclaimer{myIndex};

        if ( false == storage[myIndex].has_value() )
storage[myIndex].emplace(std::forward<Params>(args)...);
        return myIndex;
    }

public:
    template<typename... Params>
    T &operator()(Params&&... args) noexcept(false)
    {
        static thread_local unsigned slot = registerThread(
std::forward<Params>(args)... );
        return *storage[slot];
    }

    template<typename... Params>
    T const &operator()(Params&&... args) const noexcept(false)
    {
        return const_cast<ThreadLocal*>(this)->operator()(std::forward<Params>(args)...);
    }
};

// =============
// Example usage
// =============

class Counter {
    std::atomic<unsigned> total{0u};
    std::atomic<unsigned> num_threads{0u};
    std::mutex mtx;

    struct PerThread {
        std::atomic<unsigned> count{0u};
        Counter &parent;

        explicit PerThread(Counter *const c) noexcept : parent(*c)
        {
            std::lock_guard lock{parent.mtx};
            parent.pt_active.insert(this);
            parent.num_threads.fetch_add(1u, std::memory_order_relaxed);
        }

        ~PerThread(void) noexcept
        {
            {
                std::lock_guard lock{parent.mtx};
                parent.pt_active.erase(this);
            }

            if ( unsigned const leftover =
count.load(std::memory_order_relaxed) )
                parent.total.fetch_add(leftover, std::memory_order_relaxed);

            parent.num_threads.fetch_sub(1u, std::memory_order_relaxed);
        }
    };

    std::set<PerThread*> pt_active;
    ThreadLocal<PerThread, 32u> pt;

public:
    void incr(void)
    {
        pt(this).count.fetch_add(1u, std::memory_order_relaxed);
    }

    unsigned collect(void)
    {
        unsigned sum = 0u;

        {
            std::lock_guard lock{mtx};
            for ( PerThread *const p : pt_active ) sum +=
p->count.exchange(0u, std::memory_order_relaxed);
        }

        if (sum) total.fetch_add(sum, std::memory_order_relaxed);
        return sum;
    }
};

#include <iostream> // cout

auto main(void) -> int
{
    Counter counter;

    constexpr unsigned numThreads = 6u;
    constexpr unsigned iterations = 200u;

    auto work = [&](unsigned id)
    {
        for (unsigned i = 0; i < iterations; ++i) counter.incr();
    };

    std::vector<std::thread> threads;
    threads.reserve(numThreads);

    for ( unsigned i = 0; i < numThreads; ++i ) threads.emplace_back(work, i);

    for ( auto &t : threads ) t.join();

    std::cout << "Collected total = " << counter.collect() << std::endl;
}