Re: question about ARM watchpoints

On 09/05/2014 04:51 AM, Joel Brobecker wrote:
>> So sounds like this line should be skipped on ARMv7-M:
>>
>>   arm-tdep.c:  set_gdbarch_have_nonsteppable_watchpoint (gdbarch, 1);
>>
>> Could you try removing it?
>>
>> But then I wonder why we never heard of this before.  Are there
>> ARMv7-M's that behave differently?  And what about ARMv6-M ?
> 
> That's what I wanted to try too, and will do soon. As to why we never
> heard of this before - the only affirmative answer would be from someone
> better able to undertand the docs than me.  But here's a wild guess: the
> fact that GDB stopped one instruction too late is invisible to the user
> 99% of the time. What triggered me seeing it was a change in code
> generation which caused the update to be at the penultimate instruction
> of a function. I wouldn't have seen it if the update was anywhere before
> that.

Ah, indeed.

Hmm.  Sounds like we need a test to catch this.  :-)

Most preferably, a portable one, which may be tricky.

I've written something like this before:

volatile int i1;
volatile int i2;
volatile int i3;

int main ()
{
	i1 = 1;
	i2 = 2;
	i3 = 3;
}

assuming each write would be one insn, but, then that turned
out to be a wrong assumption, for ARM even, I think.

Hmm...  I had an idea.  How about:

volatile int global;

set_global (int val)
{
	global = val;
}

main ()
{
   set_global (1);
   set_global (2);
}

#01 - run to main
#02 - set a _software_ watchpoint on i1 (w/ hw watchpoints force-disabled)
#03 - continue
#04 - (GDB single-steps the inferior, and checks the watchpoint's
     value at each step.)
#05 - watchpoint triggers
#06 - the PC now points to the instruction right after the instruction that
     actually caused the memory write.  So this is the address a hardware
     watchpoint should present the stop to the user too.  Store the PC address.
#07 - delete the software watchpoint
#08 - set a _hardware_ watchpoint on i1.
#09 - continue
#10 - hardware watchpoint triggers.  PC should point to the address
      stored in #6.

Thanks,
Pedro Alves