Re: [boost] Re: Formal Review for Interval Library beginning

From: Guillaume Melquiond (gmelquio_at_[hidden])
Date: 2002-09-03 04:01:26

On Mon, 2 Sep 2002, Douglas Gregor wrote:

> I mispoke. If the return value were tribool then comparisons would act 
> similarly to those with bool return values and the compare_certainly policy 
> in boolean contexts, because with something like:
> 
>   interval<int> i,j; // init i, j
>   if (i < j) {
>     // ...
>   }
> 
> would act the same. tribool returns true in a boolean context when the value 
> is definitely/certainly true. Unfortunately, this would act differently:
> 
>   interval<int> i,j; // init i, j
>   if (!(i < j)) {
>     // ...
>   }
> 
> compare_certainly would enter the if when it is not certain that i < j, 
> whereas tribool would enter the if when it is certain that !(i < j). 

Yes, comparisons with intervals are really a strange thing. Here are
some equalities: (the left terms use tribool)

i < j    == compare_certainly::(i < j) == cerlt(i, j)
!(i < j) == !compare_possibly::(i < j) == !poslt(i, j)

However, if compare_full is modified in order for it to return a tribool,
then you will have:

i < j    == compare_full::(i < j)
!(i < j) == !compare_full::(i < j)

> Essentially, the effect of the policy disappears once the result is converted 
> to a bool. So compare_certainly can't state that 'the result of a boolean 
> expression involving a comparison is true if the expression is certainly 
> true', but is required to state the weaker 'the result of a comparison is 
> true if the comparison is certainly true'.

Yes. And it's the reason why the explicit comparison functions where added
to the library: to avoid this kind of misunderstanding.

You could also use compare_full and made it throw an exception in the
intermediary case. And you would be able to state:
- the result of the boolean expression [...] is true if the expression is
  certainly true;
- the result of the boolean expression [...] is false if the expression is
  certainly false;
- an exception is thrown if the expression is uncertain.

Guillaume