OpenMP directive in constexpr functions

Question

According to OpenMP specification (2.1. Directive Format)

Directives may not appear in constexpr functions or in constant expressions. Variadic parameter packs cannot be expanded into a directive or its clauses except as part of an expression argument to be evaluated by the base language, such as into a function call inside an if clause.

I extensively use OpenMP in constexpr functions until g++14.2 without compiler made a warning. But using g++15 is an error.

The following is not a g++14.2 error but it is a g++15 error.

error: OpenMP directives may not appear in 'constexpr' functions [-Wtemplate-body]

constexpr void func(size_t sz) noexcept
{
     #ifdef _OPENMP
     #pragma omp parallel for
     #endif
     for (size_t i = 0; i < sz; ++i) something(i);
}

Which is the reason for this restriction? Compiler has the ability to execute OpenMP-aware code without enabling OpenMP. Did I miss something?

I am trying to patch my code with the following workaround. g++15 does not forbid this but I don't know if it breaks again the OpenMP standard and in e.g. g++16 I must again patch my code.

Is the following accepted from OpenMP specification? It is accepted from g++15 (and previous).

constexpr void func(size_t sz) noexcept
{
    auto fcommon = [](size_t i) {};
    auto fomp    = [sz]()
    {
         #ifdef _OPENMP
         #pragma omp parallel for
         #endif
         for (size_t i = 0; i < sz; ++i) fcommon(i);
    };
    auto fseq    = [sz]() { for (size_t i = 0; i < sz; ++i) fcommon(i); };

    if (std::is_constant_evaluated()) fseq(); else fomp();
}

@kiner_shah error: OpenMP directives may not appear in 'constexpr' functions [-Wtemplate-body] — Chameleon
– Chameleon, Commented Apr 17 at 7:49
clarifications should be added to the question not only in comments — 463035818_is_not_an_ai
– 463035818_is_not_an_ai, Commented Apr 17 at 7:53
as you are already making the branch, why not write two functions, one constexpr without OpenMP and one non-constexpr with openmp. Then you are in line with the openmp specs (and I expect the code to be even a bit more readable) — 463035818_is_not_an_ai
– 463035818_is_not_an_ai, Commented Apr 17 at 7:56
Why do you expect the compiler can call parallelized code at compile time? — Pepijn Kramer
– Pepijn Kramer, Commented Apr 17 at 7:57

463035818_is_not_an_ai · Accepted Answer · 2025-04-17 08:06:10Z

3

Generally, an implementation can give you more than what the specs guarantee. A problem it is when it gives you less. That there was no error before shouldn't come as a surprise. You broke the rules, so the specs cannot help you to tell if the code is ok or not.

On the other hand, as you used it and it worked as expected I suppose also your new variant would work as expected. You just cannot count on it. It may break on the next release of openmp or the compiler.

In your new code you are already branching to call two different functions. You can go one step further by writing one function with the OpenMp directives that is not constexpr and another function that is constexpr without OpenMp. In the actual function you branch depending on std::is_constant_evaluated() to call either of the two. Then you are in line with the OpenMp specifications.

edited Apr 17 at 8:06

answered Apr 17 at 8:01

463035818_is_not_an_ai

128k11 gold badges115 silver badges240 bronze badges

Sign up to request clarification or add additional context in comments.

Collectives™ on Stack Overflow

OpenMP directive in constexpr functions

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related