[clang] [flang] [flang] Add support for -finit-local-zero (PR #159788)

[via cfe-commits]

================
@@ -1317,6 +1366,48 @@ static void instantiateLocal(Fortran::lower::AbstractConverter &converter,
       Fortran::lower::genUnpackArray(*converterPtr, loc, *varDef, *sym);
     });
   }
+
+  /// These options do not initialize:
+  ///   1) Any variable already initialized
+  ///   2) objects with the POINTER attribute
+  ///   3) allocatable arrays
+  ///   4) variables that appear in an EQUIVALENCE statement
+
+  auto isEligibleForImplicitAssignment = [&var]() -> bool {
+    if (!var.hasSymbol())
+      return false;
+
+    const Fortran::semantics::Symbol &sym = var.getSymbol();
+    if (const auto *details =
+            sym.detailsIf<Fortran::semantics::ObjectEntityDetails>()) {
+      if (details->init())
+        return false;
+    }
+
+    if (sym.attrs().test(Fortran::semantics::Attr::POINTER))
+      return false;
+
+    if (sym.Rank() > 0 &&
+        sym.attrs().test(Fortran::semantics::Attr::ALLOCATABLE))
+      return false;
----------------
jeanPerier wrote:

The code that you have will segfault (you can test it end to end, this is doing a plain write to null.

However, where is the requirement that `-finit-local-zero` should initialize local scalar allocatable, as far as I see, gfortran is not doing it and it seems a bit odd to me. I do not think we need to allocate/initialize allocatable on entry, this could change the behavior of valid programs (given this changes the result of ALLOCATED initrinsic), and any usage of unitialized allocatable will be caught by a segfault since the base address is set to null.

I would advise rejecting all allocatables.

https://github.com/llvm/llvm-project/pull/159788