[Bug debug/55094] [4.7/4.8 Regression] ICE in maybe_record_trace_start, at dwarf2cfi.c:2224

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

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55094

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2012-11-13
                 CC|                            |jakub at gcc dot gnu.org,
                   |                            |rth at gcc dot gnu.org,
                   |                            |vmakarov at gcc dot gnu.org
     Ever Confirmed|0                           |1

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-11-13 14:31:11 UTC ---
Doesn't ICE on the trunk anymore since LRA merge, apparently because LRA
decides for some reason to use frame pointer anyway, despite
-fomit-frame-pointer -Os.
CCing Vlad to see why.  Apparently it happens in:
665          if (ep->from_rtx == XEXP (x, 0)
666              || (ep->to_rtx == XEXP (x, 0)
667              && ep->to_rtx != hard_frame_pointer_rtx))
668            setup_can_eliminate (ep, false);
where ep->from_rtx is (frame), ep->to_rtx is (sp) and x is (pre_dec:SI
(reg/f:SI 7 sp)).
Is that intentional?

Anyway, testcase that also ICEs on 4.7 branch is e.g.
extern int puts (const char *);
int x;

int
main (int argc, char **argv)
{
  if (argc)
    {
      puts (argv[0]);
      x = 1;
      __builtin_unreachable ();
    }
  x = 1;
  __builtin_unreachable ();
}
with the same options - similar reason, crossjumping during jump2 crossjumps
bbs with different ARGS_SIZE final depth.  For noreturn calls we've handled it
by adding REG_ARGS_SIZE note on the noreturn call I think, for unconditional
TRAP_IF we perhaps could do the same, but not sure what we can do for
__builtin_unreachable which expands to nothing..., with no outgoing edge from
the bb at all.