[Bug libgomp/83590] [nvptx] openacc reduction C regressions

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83590

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
With that change, we end up with weird array, TYPE_SIZE_UNIT/TYPE_SIZE on its
type is constant, but DECL_SIZE_UNIT/DECL_SIZE on the decl is non-constant
(pedantically it is a VLA, but we know the sizes at compile time).
So, gimplify_vla_decl is invoked on it.  Perhaps just omp-low.c
(is_variable_sized) should be changed:
 static inline bool
 is_variable_sized (const_tree expr)
 {
+  if (DECL_P (expr))
+    return TREE_CODE (DECL_SIZE_UNIT (expr)) != INTEGER_CST;
   return !TREE_CONSTANT (TYPE_SIZE_UNIT (TREE_TYPE (expr)));
 }               

I don't have nvptx offloading setup yet on my WS after reinstalling it, can you
check if following works with -fopenmp -O2 or not?

const int n = 100;

int
main ()
{
  int s = 0;
  int array[n];

  for (int i = 0; i < n; i++)
    array[i] = i + 1;

#pragma omp target parallel for map(tofrom:s) reduction (+:s)
  for (int i = 0; i < n; i++)
    s += array[i];

#pragma omp target parallel for map(tofrom:s) reduction (+:s) map(to:array)
  for (int i = 0; i < n; i++)
    s += array[i];

  if (s != n * (n + 1))
    __builtin_abort ();

  return 0;
}