OpenMP target region inside parallel region causes internal compiler error

Accelerated Computing HPC Compilers nvc, nvc++ and nvfortran
1

While trying to get some benchmark examples for testing the OMPT interface to run on my system, I encountered a small issue when trying to start a target region from inside a parallel and for region.

Take the following code as an example:

int main(void)
{
    int num_iterations = 100;
    #pragma omp parallel
    {
        #pragma omp for
        for(int i = 0; i < num_iterations; ++i) {
            #pragma omp target teams distribute parallel for nowait
            for(int j = 0; j < 10; ++j){}
        }
    }
}

Trying to compile this code with NVHPC 23.3 causes the following error message to appear:

> nvc++ -fopenmp -mp=multicore,gpu -Minfo=all error.cpp
NVC++-F-0000-Internal compiler error. targetinfo symbol not found in uplevel struct    1079  (error.cpp: 8)
NVC++/x86-64 Linux 23.3-0: compilation aborted

Both the nowait clause and the num_iterations variable are important to this issue. Exchanging the variable by a constant value or removing the nowait causes the program to compile perfectly fine.

2

Hi Jan,

I’m not sure if we support nowait in this context as of yet, but even so, the compiler shouldn’t be giving an ICE.

Filed TPR #33521 and sent it to engineering for review.

Thanks!
Mat