This is the mail archive of the gcc-bugs@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[Bug middle-end/35056] [4.3 Regression] ICE in copy_to_mode_reg, at explow.c:621



------- Comment #6 from matz at gcc dot gnu dot org  2008-02-02 16:08 -------
(written before richis comment, essentially the same info)

The compare routine doesn't need to be a template to show the bug.  But it
needs to take reference parameters.  The difference is in the call.  E.g.
with this testcase:

% cat x.cc
enum EBorderStyle { bla = 1 };
inline bool compare_ref(const unsigned int &t, const EBorderStyle &u)
{ return t == u; }
inline bool compare_val(const unsigned int t, const EBorderStyle u)
{ return t == u; }
struct S {
  unsigned m_style : 4;
};
void call_ref (S *s, EBorderStyle v)
{ if (!compare_ref(s->m_style, v)) s->m_style = v; }
void call_val (S *s, EBorderStyle v)
{ if (!compare_val(s->m_style, v)) s->m_style = v; }

The bodies of the two compare routines is sensible (after all nothing
fancy happens as not bitfield types are involved there anymore):

bool compare_ref(const unsigned int&, const EBorderStyle&) (t, u)
{
  bool D.1655;
  unsigned int D.1656;
  EBorderStyle D.1657;

  D.1656 = *t;
  D.1657 = *u;
  D.1655 = D.1656 == D.1657;
  return D.1655;
}
bool compare_val(unsigned int, EBorderStyle) (t, u)
{
  bool D.1662;

  D.1662 = t == u;
  return D.1662;
}

But the difference in the call setup is revealing:

void call_ref(S*, EBorderStyle) (s, v)
  <unnamed-unsigned:4> D.1672;
  unsigned int D.1670;
  ...
  D.1672 = s->m_style;
  D.1670 = D.1672;
  D.1673 = compare_ref (&D.1670, &v);
}

void call_val(S*, EBorderStyle) (s, v)
{
  ...
  <unnamed-unsigned:4> D.1682;
  unsigned int D.1683;
  ...
  D.1682 = s->m_style;
  D.1683 = (unsigned int) D.1682;
  D.1684 = compare_val (D.1683, v);
}

Note how the call with the reference parameter copies the value from the 
bitfield into an unsigned temp without a cast, and the call by value copies
it into a unsigned temp with a cast.  That latter cast will stay there until
130.final_cleanup, and hence the compare instruction will be expanded
correctly.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35056


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]