gdbserver and stepping by instruction on PowerPC

[Ron McCall <ronald dot mccall at snet dot net>]

Hi,

I have a program that has some floating point code in
it.  If I run it in gdbserver (powerpc-linux native)
in conjunction with gdb (also powerpc-linux native on
the same system) and then step the program repeatedly
by instruction (after reaching main, ), and print some
of the floating point results in gdb, I get NANs.  If
I run it entirely under gdb (no gdbserver), the same
exact program, using the same exact commands (except
for the target remote and using run instead of
continue initially) the results are fine.  This is on
Linux kernel 2.6.15 on an MPC7457 (SBS VG5 board using
ELDK 4.0).  The problem occurs both with gdb 6.3 and
gdb 6.5.  Note that by using breakpoints and
continuing and stopping does not make the problem
occur, only when doing a lot of instruction-level
stepping.

Running the same test on an Apple PowerBook (MPC7447A)
with Ubuntu Breezy Badger (kernel 2.6.12), the results
are good under both gdb and gdbserver.  I built the
2.6.15 kernel from ELDK 4.0 on the PowerBook, booted
it, and re-ran the test.  The problem recurred while
running under gdbserver but still worked fine under
gdb by itself.

I built kernel 2.6.15.7 and 2.6.17.1 for the SBS VG5
(porting SBS's Linux BSP to those kernels) and the
gdbserver problem still occurs.  I have been
unsuccessful getting a 2.6.14 series kernel to boot on
the SBS VG5 so I can't tell where between 2.6.12 and
2.6.15 the problem originated.

I upgraded the PowerBook to Ubuntu Dapper Drake
(kernel 2.6.15) and re-ran the test.  The problem with
gdbserver recurs.

The question is whether this would seem to be a Linux
kernel problem or could it be a gdbserver problem?  It
seems like maybe a floating point context switching
problem but I'm not really sure.

Thanks!

Ron McCall