Paul Brook wrote:
>> 2) The current watchpoint code is not implemented correctly so it is not
>> safe to improve it using the same system (IMHO it should not have been
>> commited in its current state). A correct implementation should not
>> delay the DEBUG exception. It should be implemented like the "normal"
>> MMU exceptions.
> 
> On most targets watchpoint traps occur after the instruction completes, so you 
> have to defer the DEBUG exception.
> Normal MMU faults occur before the instruction completes.

OK, that explains why it is not an immediately raised exception.

Still, instead of changing the TB stop condition as you suggested, I
would prefer to pay the price only on watchpoint hit. Any show stopper
to reuse the pattern of tb_invalidate_phys_page_range (find pc,
[invalidate TB?,] regenerate TB @PC with CF_SINGLE_INSN, resume
execution)? Or is you suggestion simpler and not as costly as I assume?

Jan