Re: [boost] Boost.Test feature request

From: Gennadiy Rozental (gennadiy.rozental_at_[hidden])
Date: 2006-11-09 15:42:02

"Paul A Bristow" <pbristow_at_[hidden]> wrote in message 
news:E1GiE1n-000204-61_at_he304war.uk.vianw.net...
> |  > 2) If the type has no numeric_limits support, say if it's
> |  a composite
> |  > type,
> |  > like std::complex<double> then you don't get enough digits
> |  to tell what
> |  > the
> |  > problem is.  Could the code default to whatever digits
> |  long-double has?
> |  > Again a minor change to print_log_value (in my experience
> |  std lib's ignore
> |  > requests for more digits than a type really has anyway, so
> |  it should be
> |  > harmless?).
> |
> |  Could you provide the code?
>
> I've been investigating and experimenting with this a bit.
...

> So I've got as far as changing set_precision to:
>
> void set_precision( std::ostream& ostr, mpl::false_ )
> {
>  ostr.setf(std::ios::showpoint); // Show all significant trailing zeros.
>  if(std::numeric_limits<T>::is_specialized && 
> std::numeric_limits<T>::radix == 10)
>  { // All decimal types, perhaps including NTL::RR where digits10 == 
> runtime set precision.
>    ostr.precision(std::numeric_limits<T>::digits10);
>  }
>  else if (std::numeric_limits<T>::digits == 0)
>  { // User-defined type for which digits (significand bits) value has not 
> been assigned,
>    // even if is_specialized == true, perhaps assigning other 
> numeric_limits values.
>    ostr.precision(std::numeric_limits<long double>::digits10 + 2);
>    // But this may still not be enough to avoid misleading error messages
>    // caused by not enough decimal digits being displayed.
>    // User-defined types shuld define digits, even if approximately, as a 
> workaround.
>  }
>  else if( std::numeric_limits<T>::is_specialized && 
> std::numeric_limits<T>::radix == 2 )
>  { // && std::numeric_limits<T>::digits != 0)
>    ostr.precision( 2 + std::numeric_limits<T>::digits * 301/1000 );
>    // std::numeric_limits<T>::max_digits10; for C++0x
>  }
>  // else default is equivalent to ostr.precision(6);
> } // void set_precision
>
> And this seems to work as I suggest it should.
>
> But the logic of this is more complicated than might appear, so other 
> views are welcome.
>
> (Also what are views on the usefulness of trailing zeros to show the 
> implicit range?  It is certainly helpful for in testing
> precision of floating-point functions).

We will have to return to this a bit later. Also I would like to hear from 
you the rationale for all the conditions and selected precisions. And I 
would also like to hear more opinions on this.

Gennadiy