Re: gcc warning in my trace_benchmark() code

[David Daney <ddaney.cavm@gmail.com>]

On 06/05/2014 11:53 AM, Steven Rostedt wrote:
> On Thu, 05 Jun 2014 11:44:45 -0700
> David Daney <ddaney.cavm@gmail.com> wrote:
>
>>> But stddev is s64. Ah, but the compare is:
>>>
>>> (void)(((typeof((n)) *)0) == ((uint64_t *)0));
>>>
>>> so it's complaining about a signed verses unsigned compare, not length.
>>> I think I can ignore this warning then.
>>
>> The pedant in me thinks that you should fix your code if using do_div()
>> on a signed object is undefined.  But if you aren't planning on merging
>> the code, then it probably doesn't matter.
>
> It's undefined on signed 64 numbers?

Evidently it is.

The top of asm-generic/div64.h has:

.
.
.
  * The semantics of do_div() are:
  *
  * uint32_t do_div(uint64_t *n, uint32_t base)
  * {
.
.
.

do_div() really passes the first parameter by reference, and C doesn't 
have by reference parameters, so the example is not quite right.  But it 
does seem to imply the thing should be an *unsigned* 64-bit wide variable.

It has been like this since the beginning of the git epoch.

> Where is that documented.

The code is the documentation.

> I don't
> see it in the comments, and I don't see anything in the Documentation
> directory. It only states that n must be 64bit. It doesn't say unsigned
> 64 bit.

The handful of call sites I examined, seem to all use u64 or unsigned 
long long.

I get:

   $ grep -r do_div Documentation | wc
       0       0       0

So it would seem that most of the do_div() documentation actually is the 
code.

David Daney