[flang][cuda] Avoid I/O error in block inside a kernel (llvm#160599)

Make sure we get the program unit to check the device context. The scope
would be the block otherwise.

Author: clementval

flang/lib/Semantics/check-cuda.cpp

@@ -777,7 +777,8 @@ void CUDAChecker::Enter(const parser::AssignmentStmt &x) {
 void CUDAChecker::Enter(const parser::PrintStmt &x) {
   CHECK(context_.location());
   const Scope &scope{context_.FindScope(*context_.location())};
-  if (IsCUDADeviceContext(&scope) || deviceConstructDepth_ > 0) {
+  const Scope &progUnit{GetProgramUnitContaining(scope)};
+  if (IsCUDADeviceContext(&progUnit) || deviceConstructDepth_ > 0) {
     return;
   }
 

flang/test/Semantics/cuf23.cuf

@@ -46,3 +46,10 @@ attributes(device) subroutine device1()
   real, device :: a(10)
   print*, a ! ok
 end subroutine
+
+attributes(global) subroutine global_with_block()
+  block
+    real, device :: a(10)
+    print*, a ! ok
+  end block
+end subroutine