Register variables in stabs

[Mark Kettenis <kettenis at wins dot uva dot nl>]

The run-time dynamic linker (/lib/ld-linux.so.2) on my systems
contains the following seemingly bogus stab (output from objdump
--stabs):

Symnum n_type n_othr n_desc n_value  n_strx String

2074   RSYM   0      1496   ffffffff 27890  buf:r(0,40)=*(0,41)=ar(0,1);0;-1;(0,2)

which means that buf is a pointer to an array of some type which lives
in register number 0xffffffff or -1.  Of course -1 isn't a valid
register number, and my recent register renumbering patches for the
i386 map this bogus number onto NUM_REGS + NUM_PSEUDO_REGS, which
forces GDB to complain about the symbol.  Before my change GDB would
keep the register number at -1 which resulted in no complaint.  My
question now is why the register number -1 is used.  Is this a
compiler bug, or has -1 a special meaning.

Depending on the answer to the question above: should we leave
negative register numbers alone (as GDB does for most other targets)
and not warn about them, or is it OK to complain?

Mark