For learning purposes, I've written an iterative version of a recursive algorithm for generating the set of permutations for a given set of unique integers.

Can I make any improvements to increase readability and performance?

I've also read in various sources that when transforming a recursive algorithm into an iterative one, the stack data-structure is often used to 'substitute' the call stack used in the recursive case. I'm not using a std::stack here but could I have used one to simplify my implementation?

using std::vector;

vector<vector<int>> permutations(vector<int> integers) {
    vector<vector<int>> curr_perms;

    if (integers.size() <= 1) {
        return curr_perms;

    size_t sz = integers.size();

    vector<vector<int>> next_perms;

    // initialise with the one permutation of the last single element
    vector<int> t = { integers[sz - 1] };

    // backwards iterating a sequence of at least two elements
    for (int i = sz - 2; i >= 0; --i) {
        auto first = integers[i];

        for (const auto &s : curr_perms) {
            for (auto j = 0U; j < s.size() + 1; ++j) {
                vector<int> p(s.begin(), s.end());  // make a copy 
                p.insert(p.begin() + j, first);     // shuffle in 'first' element
                next_perms.push_back(std::move(p)); // and add this as a new permutation

        // permutations found in this iteration are the input for the next
        std::swap(next_perms, curr_perms); // <-- is this needlessly copying all the elements?

    return curr_perms;

1 Answer 1


As often, the recursive algorithm is readable and beautiful, but not very efficient: stack consumption can be overwhelming, memory allocations are plentiful, and you're forced to compute the whole set of permutations at once.

When you mechanically translate the recursive algorithm into its iterative form, you often lose most of its beauty and readability, without gaining much on the front of performance: you just emulate with a data structure what the compiler would have done. It might be useful sometimes but you can generally assume that the compiler will do a better job than yourself.

Hence I would recommend looking into two alternative approaches:

  • the first is to do it the standard library way, and provide a std::next_permutation function which return the next permutation in the lexicographical order (see for instance my answer to this question). It can be improved upon though, with the SJT algorithm, and that would be an exercise of choice;

  • the second is to keep the general, incremental reasoning of the recursive algorithm, but use the additional knowledge you get in the iterative form: indeed, on the contrary to the recursive form, you know upfront how many permutations you'll return in the end, and thus can avoid all but one allocation.

For instance:

#include <vector>

std::vector<std::vector<int>> permutations_set(const std::vector<int>& set) {

    if (set.empty()) return {};

    auto current = std::begin(set);

    // the unique memory allocation
    std::vector<std::vector<int>> permutations(factorial(set.size()), {*current});

    for (++current; current != std::end(set); ++current) {
        std::size_t offset = 0; // offset plays the central role here; you could say it replaces the stack
        for (auto& permutation : permutations) {
            const auto initial_size = permutation.size();
            permutation.insert(std::next(std::begin(permutation), offset), *current); 
            if (offset++ == initial_size) offset = 0;

    return permutations;           
  • \$\begingroup\$ When you say 'you just emulate with a data structure what the compiler would have done', does that mean that you don't have to care about stack overflows anymore from deeply recursive call stacks when using a modern compiler? Can a compiler always transform a recursive algorithm into iterative form? That sounds pretty advanced. \$\endgroup\$
    – yorztif
    Commented Apr 27, 2019 at 1:15
  • \$\begingroup\$ @yorztif: you can configure recursion depth on compilers, and it is limited only by memory availability. But what I meant is that the compiler build an execution stack to execute the program. Execution contexts are then pushed / popped. That's what you also do when translating from recursive to iterative with a stack. Some compilers are able to translate some forms of recursion ( tail recursion) but I'm not sure C++ compilers do it. \$\endgroup\$
    – papagaga
    Commented Apr 27, 2019 at 5:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.