Re: FLOAT INFO QUESTION (fwd)

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

   Date: Thu, 22 Jun 2000 06:47:32 -0700 (PDT)
   From: takis@xfree86.org

   Hello,
   I had a look in the new float info routines
   of gdb 5.0 and I am a little confused. If you look
   in <linux/processor.h> you will find that the 387 
   structure is like 
	   cwd (fctrl)
	   swd (fstat)
	   twd (ftag)
	   fip
	   fcs
	   foo
	   80 bytes = 8*10  
	   status

Yes, this is the layout used by the fsave and frstor floating point
instructions.  Most systems that support x86 floating point use this
layout.

   Now in the operating system that I run (aka DG/UX unix)
   and want to do a port of 5.0, I find the structure:

   struct fpregset {

		union{

		       struct fpchip_state fpchip_state;
		       emulation struct
		       int f_fpreg[62];
		     }

		     long f_wregs[33]; /* weitek */
		    }

   Where: struct fpchip_state {
		     int state[27];
		     int status;
		    }

   So the first member reserves 108 bytes = 7 4 bytes fpregs + 8 *10.
   Tah agrees with the linux idea.

Great, that means that support the FPU is probably not too difficult!
The stuff marked with /* weitek */ is probably for systems that have a
Weitek FPU instead of an Intel i387 FPU.  Supporting those may be
difficult.  But I don't think there are a lot of systems that have one
around anymore.

   Now in GDB-5.0 we find in tm-i386.h a new layout as

      fctrl, fstat, ftag, fiseg, fioff, foseg, fooff, fop

   Which is clearly different from the struct of linux. Basically          
   if I want to unpack fpchip_state above to floating regs I dont know
   if the order is fctrl, fstat, ftag , fiseg(=fcs), fioff(=?) , ????
   or (what I always thought is)

     fctrl,fstat,ftag,fip,fcs,fopoff,fopsel,status

   And in this new layout in tm-i386.h which one is the status? I
   clearly have 108 bytes so how I would obtain the fop 
   mentioned in tm-i386.h?

This stuff is a bit complicated.  GDB sees all FPU control registers
as being 4 bytes long.  But some of them are actually shorter.  You
should be able to determine how things work by looking at the code in
i386-linux-nat.c.  You might also want to consult the Intel ia32
processor manuals (you can get those in PDF from developer.intel.com).

Anyway, I recently added the file i387-nat.c to GDB (you can find it
in the CVS tree, but not in GDB 5.0).  It defines two functions
i387_supply_fsave() and i387_fill_fsave() that convert back and forth
between the fsave format used by the FPU and the format used by GDB.
I'd like to encourage you to use these functions if possible.

Suppose you have the fpu registers in a struct fpregset `fpregs'.
Then you can simply call

   i386_supply_fsave (&fpregs);

to let GDB know about the FPU registers.  Take a look at i386bsd.c for
a real-life example on how to use these functions.

Mark