Re: [RFA]: x86_64 target - multiarch

[Eli Zaretskii <eliz at is dot elta dot co dot il>]

On 5 Sep 2001, Jiri Smid wrote:

> 	* config/djgpp/fnchange.lst: Add entries for x86_64-linux-tdep.c
> 	and x86_64-linux-nat.c

These changes are approved.

> + int
> + x86_64_insert_watchpoint (ptid_t ptid, CORE_ADDR addr, int len, int rw)
> + {
> +   return x86_64_insert_aligned_watchpoint (ptid, addr, addr, len, rw);
> + }
> + 
> + static int
> + x86_64_insert_aligned_watchpoint (ptid_t ptid, CORE_ADDR waddr,
> + 				  CORE_ADDR addr, int len, int rw)
> + {
> +   int i;
> +   int read_write_bits, len_bits;
> +   int free_debug_register;
> +   int register_number;
> + 
> +   /* Look for a free debug register.  */
> +   for (i = DR_FIRSTADDR; i <= DR_LASTADDR; i++)
> +     {
> +       if (address_lookup[i - DR_FIRSTADDR] == 0)
> + 	break;
> +     }
> + 
> +   /* No more debug registers!  */
> +   if (i > DR_LASTADDR)
> +     return -1;
> + 
> +   read_write_bits = (rw & 1) ? DR_RW_READ : DR_RW_WRITE;
> + 
> +   if (len == 1)
> +     len_bits = DR_LEN_1;
> +   else if (len == 2)
> +     {
> +       if (addr % 2)
> + 	return x86_64_insert_nonaligned_watchpoint (ptid, waddr, addr, len,
> + 						    rw);
> +       len_bits = DR_LEN_2;
> +     }

I think I already mentioned this in earlier discussions: why did you base 
the watchpoint support on the old SysV stuff instead of the newer code in 
i386-nat.c?  Don't you want the additional features supported by 
i386-nat.c, like debug register sharing between watchpoints?  Try this, 
for example:

	(gdb) watch a == 1
	(gdb) watch a == 2
	(gdb) watch a == 3
	(gdb) watch a == 4

In your implementation, this takes 4 debug registers, whereas i386-nat.c 
will only use one.  Also, with your implementation, even if you are ready 
to waste 3 debug registers, GDB might become confused if you decide to 
remove one of these watchpoints (since the remove_watchpoint code will 
zero out all the debug registers because they watch the same address).

Also, I think this implementation doesn't support watching regions larger 
than 8 bytes very well, because the code which removes watchpoints 
doesn't break the region into smaller regions the same way 
x86_64_insert_nonaligned_watchpoint does.

So I suggest to use the code in i386-nat.c as the starting point.