Consider this testcase:

  class locale {
  private:
    void _M_add_reference() const throw() {
    }
  };
  class ios_base {
    locale _M_ios_locale;
  public:
    class Init {
    };
  };
  static ios_base::Init __ioinit;

When built without debug info, GCC and clang agree, the resulting assembly is:
        .local  _ZL8__ioinit
        .comm   _ZL8__ioinit,1,1

but where things get exciting is the debug info. Besides the filename/directory/compiler identifier, the strings that GCC emits are

.LASF4:
        .string "Init"
.LASF5:
        .string "__ioinit"

which makes sense, there's one object named __ioinit of type Init (technically "ios_base::Init", so emitting that would be fine too). Clang emits:

.Lstring3:
        .ascii   "__ioinit"
        .zero   1
.Lstring4:
        .ascii   "_ZL8__ioinit"
        .zero   1
.Lstring5:
        .ascii   "Init"
        .zero   1
.Lstring6:
        .ascii   "_M_ios_locale"
        .zero   1
.Lstring7:
        .ascii   "_ZNK6locale16_M_add_referenceEv"
        .zero   1
.Lstring8:
        .ascii   "_M_add_reference"
        .zero   1
.Lstring9:
        .ascii   "locale"
        .zero   1
.Lstring10:
        .ascii   "ios_base"
        .zero   1

Yow! We never emitted code for any of those functions at the very least. My estimate is that this is causing about 60% of the extra strings. (Only an estimate because there are other sources of differences, such as clang emitting all case names for a given enum, while gcc only emits the used ones. I consider clang's behaviour better.)