Re: gdb in CVS is totally broken on Linux/i386

[Jim Blandy <jimb at cygnus dot com>]

> > What is happening inside the `if (GETXFPREGS_SUPPLIES (regno))'
> > clause?  Why isn't fetch_xfpregs providing the information accurately?
> > 
> 
> GETFPREGS_SUPPLIES is defined and never used. Should it be a little
> faster to check GETFPREGS_SUPPLIES first to get the normal fp
> registers?

I think it's faster to use PTRACE_GETXFPREGS if it's available, since
that gives you both the FP and the SSE registers in one system call.

> > If you are running a very bleeding-edge kernel, Doug Ledford may have
> > changed the format of `struct i387_hard_fxsave'.  That would
> > definitely give you garbage values for your FP registers.  My changes
> > are only meant to work with the kernel patches posted on
> > sourceware.cygnus.com.
> 
> I am using Doug's P3 patch. Apparently he didn't handle
> i387_hard_fxsave right between kernel and user space. I
> have sent him detailed bug report. I hope he will fix it
> soon.

See, I think there are several versions of Doug's patches flying
around.  In the one that GDB expects at the moment, the kernel's
struct i387_hard_fxsave contains both an FSTENV-format and an
FXSAVE-format record, because Doug's patch uses both instructions
during a context switch --- the FP control registers are saved twice.
PTRACE_GETXFPREGS gives you the whole thing.

However, in a more recent version of Doug's patch (which I have not
yet gotten to work), he uses only the FXSAVE instruction, which is
more efficient.  But this means that the data accessed by
PTRACE_GETXFPREGS is different from what GDB expects.