Bug report: OpenMP atomic with -O0 or -O1 optimization levels (PGI19 or PGI20)


I have a problem using OpenMP atomic with PGI compiler versions 19 and 20.

Here is a simple reproducer:

#include <iostream>
#include <vector>
#include <random>
#include <omp.h>

int main()
  const int SIZE = 10;
  std::vector<double> vect(SIZE);

#pragma omp parallel
    auto rank = omp_get_thread_num();
    std::mt19937 rand;
    std::uniform_real_distribution<double> double_dist;
    std::uniform_int_distribution<int> int_dist(0,SIZE-1);
    auto index = int_dist(rand);
    auto val = double_dist(rand);
#pragma omp atomic
    vect[index] += val;

  for (const auto& v : vect)
    std::cout << v << " ";
  std::cout << std::endl;

  return 0;

This compiles with PGI 18.7 but not with PGI 19.10 nor PGI 20.10 with -O0 or -O1 optimization levels:

$ pgc++ --version
pgc++ 20.1-0 LLVM 64-bit target on x86-64 Linux -tp skylake
PGI Compilers and Tools
Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.

$ pgc++ -O1 -std=c++11 -mp test2.cpp
PGC++-S-0000-Internal compiler error. get_dtype_from_ilm: unexpected ILM opc       2  (test2.cpp: 21)
PGC++-S-0155-Invalid atomic update statement.  (test2.cpp: 21)
PGC++/x86-64 Linux 20.1-0: compilation completed with severe errors

$ pgc++ -O2 -std=c++11 -mp test2.cpp
$ ./a.out
6.37603 3.54772 5.52872 4.60111 2.51587 6.62821 1.46953 2.05598 3.59274 4.26629

Thanks in advance.

Thanks for the report Emeric. This is a new one that we haven’t seen before and I’m not sure what’s triggering the compiler error, nor see an easy work-around.

I’ve created a new problem report (TPR #28315) and sent it our engineers for further investigation.