[flang] Fixed regression with CDEFINED linkage

#162722 introduced a regression that
started creating initializers for CDEFINED variables. CDEFINED variables
cannot have initializers, because their storage is expected come from
elsewhere, likely outside of Fortran. Fixed the regression and improved the
regression test to catch the incorrect initialization case.

Author: eugeneepshteyn

flang/lib/Semantics/resolve-names.cpp

@@ -9194,11 +9194,12 @@ bool DeclarationVisitor::CheckNonPointerInitialization(
               "'%s' has already been initialized"_err_en_US);
         } else if (IsAllocatable(ultimate)) {
           Say(name, "Allocatable object '%s' cannot be initialized"_err_en_US);
+        } else if (details->isCDefined()) {
+          // CDEFINED variables cannot have initializer, because their storage
+          // may come outside of Fortran.
+          context().Warn(common::UsageWarning::CdefinedInit, name.source,
+              "CDEFINED variable cannot be initialized"_warn_en_US);
         } else {
-          if (details->isCDefined()) {
-            context().Warn(common::UsageWarning::CdefinedInit, name.source,
-                "CDEFINED variable should not have an initializer"_warn_en_US);
-          }
           return true;
         }
       } else {

flang/test/Lower/cdefined.f90

@@ -6,4 +6,5 @@ module m
   integer(c_int), bind(C, name='c_global', CDEFINED) :: c  = 42
   ! CHECK: fir.global @c_global : i32
   ! CHECK-NOT: fir.zero_bits 
+  ! CHECK-NOT: arith.constant 42
 end

flang/test/Semantics/cdefined.f90

@@ -1,6 +1,6 @@
 ! RUN: %python %S/test_errors.py %s %flang_fc1 -pedantic -Werror
 module m
   use iso_c_binding
-  !WARNING: CDEFINED variable should not have an initializer [-Wcdefined-init]
+  !WARNING: CDEFINED variable cannot be initialized [-Wcdefined-init]
   integer(c_int), bind(C, name='c_global', CDEFINED) :: c  = 42
 end