Reproduced with vanilla 4.8; 4.9 is fine.  Bisecting...

http://gcc.gnu.org/r218205

And that commit had a follow-up:
http://gcc.gnu.org/r218316

Reduced.  Only -O is needed to show the ICE.

class A
{
    int _M_p;
};
template < typename _Key, typename _Tp > class B
{
public:
    _Tp operator[](_Key);
};

union U
{
    long double ld;
    void *v;
};
A a;
void
fn1 (U p1)
{
    if (p1.v)
      *reinterpret_cast < A * >(p1.v) = a;
}

class C
{
public:
    C (A)
    {
        m_data.ld = 0;
        fn1 (m_data);
    }
    U m_data;
};

class D
{
public:
    A a1;
    A a2;
    void m_fn1 ()
    {
        arr[2][a1] = a2;
    } B < int, B < A, C > >arr;
};

void
fn2 ()
{
    D b;
    b.m_fn1 ();
}

Slightly more cleaned up:

// { dg-options "-O -Wno-psabi" }
// { dg-do compile }

struct A { int a; };
template <typename T, typename V> struct B { V operator[] (T); };
union U { long double ld; void *v; };
A a;

void
bar (U &x)
{
  if (x.v) *reinterpret_cast <A *>(x.v) = a;
}

struct C { C (A) { c.ld = 0; bar (c); } U c; };
struct D { A d, e; void foo () { f[0][d] = e; } B <int, B <A, C> > f; };

void
baz ()
{
  D d;
  d.foo ();
}

I've backported the relevant fixes to the 4.8 branch.

(In reply to Marek Polacek from comment #7)
> I've backported the relevant fixes to the 4.8 branch.

You guys are awesome!    Thank you.    Will this be out in a patch to the compiler sometime soon?

Most likely in RHEL 7.2.

*** Bug 1222631 has been marked as a duplicate of this bug. ***

Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.

For information on the advisory, and where to find the updated
files, follow the link below.

If the solution does not work for you, open a new bug report.

https://rhn.redhat.com/errata/RHBA-2015-2080.html