This is probably bug 93951.

Yes, it looks very much related - the same "fix" works.  Confirmed.

I'm reducing what looks like the same ICE for 19 hours already and am at 3.5MB so far.

(In reply to Jakub Jelinek from comment #3)
> I'm reducing what looks like the same ICE for 19 hours already and am at
> 3.5MB so far.

Btw. are you using C-Vise or C-Reduce?

cvise.

Reduced testcase (-g -std=c++17 -O1 -flto):
namespace N {
  template <typename, typename>
  bool c;
}
namespace {
class A {};
class B {};
using D = class C {};
class E {};
}
namespace O::P::Q {
struct F {};
}
namespace O::R {
template <typename>
struct G;
template <typename T>
struct H {
  using I = G<decltype (T::foo)>;
  constexpr static unsigned h = I::cr;
};
template <typename T>
A
bar ()
{
  if (N::c<T, P::Q::F>) {
    struct J {
      static void foo();
    };
    H<J>();
  }
  return A ();
}
template <typename T>
A
baz ()
{
  bar<T>;
  return A ();
}
}
namespace O::P {
template <typename T>
A
baz (B, D, C, E, T)
{
  R::baz<T>;
  return A ();
}
A
baz (B cx, D cy, C cz, E d)
{
  baz (cx, cy, cz, d, Q::F{});
  return A ();
}
}

Started with r8-2599-g1ea85365b430b5ade1d0b824e5406c006e6ffdb7 too.

The master branch has been updated by Jason Merrill <jason@gcc.gnu.org>:

https://gcc.gnu.org/g:89d9c634dc5c10b499c23297ef70133066946790

commit r11-5214-g89d9c634dc5c10b499c23297ef70133066946790
Author: Jason Merrill <jason@redhat.com>
Date:   Fri Nov 20 15:20:45 2020 -0500

    dwarf2: ICE with local class in unused function [PR97918]
    
    Here, since we only mention bar<B>, we never emit debug information for it.
    But we do emit debug information for H<J>::h, so we need to refer to the
    debug info for bar<B>::J even though there is no bar<B>.  We deal with this
    sort of thing in dwarf2out with the limbo_die_list; parentless dies like J
    get attached to the CU at EOF.  But here, we were flushing the limbo list,
    then generating the template argument DIE for H<J> that refers to J, which
    adds J to the limbo list, too late to be flushed.  So let's flush a little
    later.
    
    gcc/ChangeLog:
    
            PR c++/97918
            * dwarf2out.c (dwarf2out_early_finish): flush_limbo_die_list
            after gen_scheduled_generic_parms_dies.
    
    gcc/testsuite/ChangeLog:
    
            PR c++/97918
            * g++.dg/debug/localclass2.C: New test.

The new g++.dg/debug/localclass2.C test file you added is marked "target c++11", but it seems I can only cause the crash with -std=c++17 or later standards. Should the test file perhaps be amended?

For debug/ tests I guess it isn't really used, just documents that the testcase is valid C++11 or later, because debug/ tests iterate over various debug options rather than C++ standard options.

And the default is -std=gnu++17.

The releases/gcc-10 branch has been updated by Jason Merrill <jason@gcc.gnu.org>:

https://gcc.gnu.org/g:8157b74114f2ba8a6495244f3e171a818a86436a

commit r10-9071-g8157b74114f2ba8a6495244f3e171a818a86436a
Author: Jason Merrill <jason@redhat.com>
Date:   Fri Nov 20 15:20:45 2020 -0500

    dwarf2: ICE with local class in unused function [PR97918]
    
    Here, since we only mention bar<B>, we never emit debug information for it.
    But we do emit debug information for H<J>::h, so we need to refer to the
    debug info for bar<B>::J even though there is no bar<B>.  We deal with this
    sort of thing in dwarf2out with the limbo_die_list; parentless dies like J
    get attached to the CU at EOF.  But here, we were flushing the limbo list,
    then generating the template argument DIE for H<J> that refers to J, which
    adds J to the limbo list, too late to be flushed.  So let's flush a little
    later.
    
    gcc/ChangeLog:
    
            PR c++/97918
            * dwarf2out.c (dwarf2out_early_finish): flush_limbo_die_list
            after gen_scheduled_generic_parms_dies.
    
    gcc/testsuite/ChangeLog:
    
            PR c++/97918
            * g++.dg/debug/localclass2.C: New test.

The releases/gcc-8 branch has been updated by Jason Merrill <jason@gcc.gnu.org>:

https://gcc.gnu.org/g:ca8325441a6bb06292db9f165607d4e395f46c4b

commit r8-10638-gca8325441a6bb06292db9f165607d4e395f46c4b
Author: Jason Merrill <jason@redhat.com>
Date:   Fri Nov 20 15:20:45 2020 -0500

    dwarf2: ICE with local class in unused function [PR97918]
    
    Here, since we only mention bar<B>, we never emit debug information for it.
    But we do emit debug information for H<J>::h, so we need to refer to the
    debug info for bar<B>::J even though there is no bar<B>.  We deal with this
    sort of thing in dwarf2out with the limbo_die_list; parentless dies like J
    get attached to the CU at EOF.  But here, we were flushing the limbo list,
    then generating the template argument DIE for H<J> that refers to J, which
    adds J to the limbo list, too late to be flushed.  So let's flush a little
    later.
    
    gcc/ChangeLog:
    
            PR c++/97918
            * dwarf2out.c (dwarf2out_early_finish): flush_limbo_die_list
            after gen_scheduled_generic_parms_dies.
    
    gcc/testsuite/ChangeLog:
    
            PR c++/97918
            * g++.dg/debug/localclass2.C: New test.

The releases/gcc-9 branch has been updated by Jason Merrill <jason@gcc.gnu.org>:

https://gcc.gnu.org/g:196716c10bcd4074c404cc8f13bf8d9b31c76238

commit r9-9067-g196716c10bcd4074c404cc8f13bf8d9b31c76238
Author: Jason Merrill <jason@redhat.com>
Date:   Fri Nov 20 15:20:45 2020 -0500

    dwarf2: ICE with local class in unused function [PR97918]
    
    Here, since we only mention bar<B>, we never emit debug information for it.
    But we do emit debug information for H<J>::h, so we need to refer to the
    debug info for bar<B>::J even though there is no bar<B>.  We deal with this
    sort of thing in dwarf2out with the limbo_die_list; parentless dies like J
    get attached to the CU at EOF.  But here, we were flushing the limbo list,
    then generating the template argument DIE for H<J> that refers to J, which
    adds J to the limbo list, too late to be flushed.  So let's flush a little
    later.
    
    gcc/ChangeLog:
    
            PR c++/97918
            * dwarf2out.c (dwarf2out_early_finish): flush_limbo_die_list
            after gen_scheduled_generic_parms_dies.
    
    gcc/testsuite/ChangeLog:
    
            PR c++/97918
            * g++.dg/debug/localclass2.C: New test.

Fixed for 8.5/9.4/10.3/11.