[Bug middle-end/53696] [4.7/4.8 Regression] ICE: SIGSEGV in gimplify_decl_expr (gimplify.c:1454) with -fkeep-inline-functions on invalid use of lambda

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

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53696

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

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

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-12-07 15:54:22 UTC ---
And
struct A;

template <int N>
void foo(A& a)
{
  [=](){a;};
}

void
bar (A&a)
{
  foo<6>(a);
}

ICEs even without -fkeep-inline-functions elsewhere.
I wonder if
  else
    /* Capture by copy requires a complete type.  */
    type = complete_type (type);
in semantics.c (add_capture) shouldn't be something like:
--- cp/semantics.c.jj    2012-12-06 21:33:57.000000000 +0100
+++ cp/semantics.c    2012-12-07 16:45:30.000000000 +0100
@@ -9164,7 +9164,14 @@ add_capture (tree lambda, tree id, tree
     }
   else
     /* Capture by copy requires a complete type.  */
-    type = complete_type (type);
+    {
+      if (processing_template_decl)
+    type = complete_type (type);
+      else
+    type = complete_type_or_else (type, initializer);
+      if (type == NULL_TREE)
+    type = error_mark_node;
+    }

   /* Add __ to the beginning of the field name so that user code
      won't find the field with name lookup.  We can't just leave the name

This fixes the ICE in #c0 (but we error twice for the incomplete type, once
within the lambda, once within the parent function).
For templates it of course doesn't change anything, we I think generally can't
require the type to be complete there, so it must be done later during pt.c.