[Bug debug/55541] [4.6/4.7/4.8 Regression] unable to see local variables due extra lexical block was generated

["jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55541

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jason at gcc dot gnu.org

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-11-30 13:03:40 UTC ---
I've tried to do:
--- gcc/cp/decl.c.jj    2012-11-19 14:41:16.000000000 +0100
+++ gcc/cp/decl.c    2012-11-30 12:24:42.255349700 +0100
@@ -13700,7 +13700,27 @@ finish_function (int flags)
     {
       /* Make it so that `main' always returns 0 by default.  */
       if (DECL_MAIN_P (current_function_decl))
-    finish_return_stmt (integer_zero_node);
+    {
+      /* If there is so far just a single statement, BIND_EXPR
+         with the whole main function body in it, push
+         return 0 stmt at the end of its body instead of after
+         the BIND_EXPR, so that variables declared in the function
+         scope are still in scope on the return stmt.  */
+      tree bind = NULL_TREE;
+      if (TREE_CODE (cur_stmt_list) == STATEMENT_LIST)
+        {
+          tree_stmt_iterator i = tsi_start (cur_stmt_list);
+          if (tsi_one_before_end_p (i)
+          && TREE_CODE (tsi_stmt (i)) == BIND_EXPR)
+        {
+          bind = tsi_stmt (i);
+          vec_safe_push (stmt_list_stack, BIND_EXPR_BODY (bind));
+        }
+        }
+      finish_return_stmt (integer_zero_node);
+      if (bind)
+        BIND_EXPR_BODY (bind) = stmt_list_stack->pop ();
+    }

       if (use_eh_spec_block (current_function_decl))
     finish_eh_spec_block (TYPE_RAISES_EXCEPTIONS

which makes sure that the return 0; has the right locus (for C99 similar change
would be needed), unfortunately it doesn't fix this.
Unlike C, the C++ FE creates an extra BLOCK around the BLOCK with BLOCK_VARS i
in this case, supposedly for the argument scope, so DECL_INITIAL
(current_function_decl) is in C the block with i var, but in C++ a block whose
BLOCK_SUBBLOCK is block with i var.  While the return stmt with the above patch
still has correct block, it seems to be actually completely ignored in the end,
because what matters actually is epilogue_location (in 4.7 and earlier
epilogue_locator).  cfgexpand.c and expand_function_end now set
epilogue_location to the input_location, when outside of particular stmt
expansion, which doesn't have a block, and I assume that DECL_INITIAL is then
used.  So the question is, can the C++ FE (when?) get rid of the outermost
BLOCK and change DECL_INITIAL (perhaps during genericization or so) to the real
function body scope instead.