[flang] [llvm] [flang][OpenMP] Overhaul implementation of ATOMIC construct (PR #137852)

[Tom Eccles via llvm-commits]

================
@@ -0,0 +1,83 @@
+!RUN: %python %S/../test_errors.py %s %flang -fopenmp -fopenmp-version=60
+
+subroutine f00
+  integer :: x, y
+
+  ! The x is a direct argument of the + operator. Expect no diagnostics.
+  !$omp atomic update
+  x = x + (y - 1)
+end
+
+subroutine f01
+  integer :: x
+
+  ! x + 0 is unusual, but legal. Expect no diagnostics.
+  !$omp atomic update
+  x = x + 0
+end
+
+subroutine f02
+  integer :: x
+
+  ! This is formally not allowed by the syntax restrictions of the spec,
+  ! but it's equivalent to either x+0 or x*1, both of which are legal.
+  ! Allow this case. Expect no diagnostics.
+  !$omp atomic update
+  x = x
+end
+
+subroutine f03
+  integer :: x, y
+
+  !$omp atomic update
+  !ERROR: The atomic variable x should occur exactly once among the arguments of the top-level + operator
+  x = (x + y) + 1
+end
+
+subroutine f04
+  integer :: x
+  real :: y
+
+  !$omp atomic update
+  !ERROR: This intrinsic function is not a valid ATOMIC UPDATE operation
+  x = floor(x + y)
+end
+
+subroutine f05
+  integer :: x
+  real :: y
+
+  ! An explicit conversion is accepted as an extension.
+  !$omp atomic update
+  x = int(x + y)
----------------
tblah wrote:

Please could you document this extension (maybe flang/docs/Extensions.md or at least explain in the commit message about these consequences of using evaluate::Expr).

https://github.com/llvm/llvm-project/pull/137852