gcc HEAD, bad DW_CFA_offset register

[mec dot gnu at mindspring dot com (Michael Elizabeth Chastain)]

I'm getting a lot of gdb test suite regressions with gcc HEAD 2004-07-01
compared to gcc HEAD 2004-06-17.  stabs+ is okay but dwarf-2 has about
1000 differences, almost all bad.

I think there's a bug in gcc's dwarf-2 emitter.  Can Daniel J, or
somebody else who knows dwarf-2, check me on this?

Here is a function from arch/i386-unwind.c:

  void
  trap (void)
  {
    asm ("int $0x03");
  }

Here is the good-looking CFA information from gcc HEAD 2004-06-17:

  00000028 00000014 00000014 FDE cie=00000014 pc=08048318..0804831e
    DW_CFA_advance_loc: 1 to 08048319
    DW_CFA_def_cfa_offset: 8
    DW_CFA_offset: r5 at cfa-8
    DW_CFA_advance_loc: 2 to 0804831b
    DW_CFA_def_cfa_reg: r5

Looks okay to me.  This is native i686-pc-linux-gnu so r5 is $ebp.

Here is the strange CFA information from gcc HEAD 2004-07-01:

  00000028 00000014 00000014 FDE cie=00000014 pc=08048318..0804831e
    DW_CFA_advance_loc: 1 to 08048319
    DW_CFA_def_cfa_offset: 8
    DW_CFA_offset: r6 at cfa-8
    DW_CFA_advance_loc: 2 to 0804831b
    DW_CFA_def_cfa_reg: r5

This is strange: the r5 changed to an r6!  (r5=$ebp, r6=$esi)
This confuses gdb when it's locating frames; I get a bad backtrace
and a FAIL.

This is r5->r6 change is the *only* difference (besides version strings)
from 20040617.s to 20040701.s.

I just want to check that I am reading the CFA's correctly and
that the second version of the FDE is messed up.  Is that right?
If so, I'll narrow it down to a particular gcc patch and file
a PR against gcc HEAD.

Michael C