squeeze_notes vs. g++.other/loop2.C (resend)

[Jason Merrill <jason at redhat dot com>]

loop2.C has been failing for me on i686-linux for months now.  I finally
got around to investigating it, er, a month ago when I sent this note
before.  A simplified version:

struct A {
  ~A();
};

struct B : public A {
  void f ();
};

bool t;

void q()
{
  while (true)
    {
      B b;
      if (t)
	break;
      b.f();
    }
}

There are two main EH regions: the one around the loop body for cleaning
up b, and inside it the one around the jump fixup to avoid destroying b
twice.  The loop optimizer appropriately notices that it can move the fixup
block outside the loop.  The region start note is in the batch of notes
immediately after the jump_insn, and the region end note is in the batch of
notes after the barrier.  insns_safe_to_move_p sees that the block contains
a complete eh region, and says OK.  So far, so good.

But then we call squeeze_notes, which moves the start note out of the group
of insns to be moved, but leaves the end note because it assumes that its
end parm can't be a movable note.  This causes a crash in scan_regions, and
seems a simple bug.

But it also seems to me that squeeze_notes shouldn't be messing with EH
notes at all; how can it possibly be right to remove EH region notes from
around a block of instructions?  Won't that screw up EH semantics?

What did you intend when you changed squeeze_notes to move EH notes?

Perhaps you're aware of these issues, and plan to ignore this problem until
we've done away with the EH notes entirely, as you've proposed.

Jason