[Flang][OpenMP] Use the ultimate symbol for allocatable check

[peixin]

This reminds me another case as follows:

program main
  real(4), pointer::a
  allocate(a)
  a=1.0
!$omp parallel private(a)
  if (.not. associated(a)) then
     print *,'ng1: a'
  endif
  a=2.0
!$omp end parallel
  if (a /= 1.0) then
     print *, 'ng: a = ', a
  endif
  print *,'pass'
end program main

$ flang-new -fc1 tmp.f90 -fopenmp
error: Semantic errors in tmp.f90
./tmp.f90:6:24: error: POINTER= argument of ASSOCIATED() must be a POINTER
    if (.not. associated(a)) then
                         ^
./tmp.f90:2:21: Declaration of 'a'
    real(4), pointer::a
                      ^

Since we use host-association for privitized variables, is there one better way to handle these cases? I guess there are still some bugs if we make the changes in non-OpenMP function code.

[kiranchandramohan]

In D144383#4140211, @peixin wrote:

This reminds me another case as follows:

program main
  real(4), pointer::a
  allocate(a)
  a=1.0
!$omp parallel private(a)
  if (.not. associated(a)) then
     print *,'ng1: a'
  endif
  a=2.0
!$omp end parallel
  if (a /= 1.0) then
     print *, 'ng: a = ', a
  endif
  print *,'pass'
end program main

$ flang-new -fc1 tmp.f90 -fopenmp
error: Semantic errors in tmp.f90
./tmp.f90:6:24: error: POINTER= argument of ASSOCIATED() must be a POINTER
    if (.not. associated(a)) then
                         ^
./tmp.f90:2:21: Declaration of 'a'
    real(4), pointer::a
                      ^

Since we use host-association for privitized variables, is there one better way to handle these cases? I guess there are still some bugs if we make the changes in non-OpenMP function code.

Thanks @peixin, one option is to copy the attributes while making the associated symbol. But I was advised to instead use the UltimateSymbol way in https://reviews.llvm.org/D112876. I don't know of other general ways to handle these. You are right that there are some bugs. I hope we can catch these with widespread testing and careful reviews like yours. :)

Do you want me to make the change for pointer in this patch?

[peixin]

Thanks @peixin, one option is to copy the attributes while making the associated symbol. But I was advised to instead use the UltimateSymbol way in https://reviews.llvm.org/D112876. I don't know of other general ways to handle these. You are right that there are some bugs. I hope we can catch these with widespread testing and careful reviews like yours. :)

Thanks. Got it. That fix is too long ago and I forgot about it. Now this sounds reasonable to me.

Do you want me to make the change for pointer in this patch?

You can fix it in another patch, which may look more clear. Either way is OK to me.

flang/lib/Evaluate/tools.cpp

bool IsAllocatableOrPointerObject(
   return (sym && semantics::IsAllocatableOrPointer(*sym)) ||
       evaluate::IsObjectPointer(expr, context);
 }
 
 bool IsAllocatableDesignator(const Expr<SomeType> &expr) {
   // Allocatable sub-objects are not themselves allocatable (9.5.3.1 NOTE 2).
   if (const semantics::Symbol *
       sym{UnwrapWholeSymbolOrComponentOrCoarrayRef(expr)}) {
-    return semantics::IsAllocatable(*sym);
+    return semantics::IsAllocatable(sym->GetUltimate());
   }
   return false;
 }
 
 bool MayBePassedAsAbsentOptional(
     const Expr<SomeType> &expr, FoldingContext &context) {
   const semantics::Symbol *sym{UnwrapWholeSymbolOrComponentDataRef(expr)};
   // 15.5.2.12 1. is pretty clear that an unallocated allocatable/pointer actual

flang/test/Semantics/OpenMP/omp-private-is-pointer-allocatable-check.f90

 ! RUN: %flang_fc1 -fopenmp -fsyntax-only %s
 
 subroutine s
   integer, pointer :: p
   integer, target :: t
+  real(4), allocatable :: arr
 
   !$omp parallel private(p)
     p=>t
   !$omp end parallel
+
+  allocate(arr)
+  !$omp parallel private(arr)
+  if (.not. allocated(arr)) then
+     print *, 'not allocated'
+  endif
+  !$omp end parallel
 end subroutine