[Bug c++/71827] [5/6/7 Regression] ICE on invalid C++11 code with non constant diff between two labels

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71827

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
__INTPTR_TYPE__
foo () 
{
  m: n: constexpr __INTPTR_TYPE__ a = ((__INTPTR_TYPE__) &&n -
(__INTPTR_TYPE__) &&m) + 23;
  return a;
}

compiles fine, so it is considered an integral constant expression that just
doesn't fold into an INTEGER_CST.  For the template parameter we need an
INTEGER_CST (I bet in various other cases).

constexpr int
foo (const int *a)
{
  m: n: constexpr __INTPTR_TYPE__ b = ((__INTPTR_TYPE__) &&n -
(__INTPTR_TYPE__) &&m) + 23;
  return a[b];
}

const int c[10] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 };
constexpr int d = foo (c);

is rejected though:
pr71827-4.C:9:23:   in constexpr expansion of ‘foo(((const int*)(& c)))’
pr71827-4.C:5:13: error: ‘(((long unsigned int)((((long int)(&& n)) - ((long
int)(&& m))) + 23)) * 4)’ is not a constant expression
   return a[b];
             ^

clang++ rejects all of these.
So perhaps replace the assert with an error message if the expression is not
INTEGER_CST?