boost::math::policies::raise_rounding_error returns the wrong min/max values for ignore_error and errno_on_error

[benkerby@…]

In the documentation for the rounding functions, in "Table 6. Possible Actions for Rounding Errors", it says that the round function will return "the largest representable value of the target integer type (or the most negative value if the argument to the function was less than zero).

However, the specializations of the raise_rounding_error function for ignore_error and errno_on_error call std::numeric_limits min() and max() with the source type rather than the result type.

For example, calling

iround(INT_MAX + 1.0)

will return

static_cast<int>(std::numeric_limits<double>::max())

rather than

static_cast<int>(std::numeric_limits<int>::max())

as was specified.

[benkerby@…]

Note that the function reads:

   // This may or may not do the right thing, but the user asked for the error
   // to be ignored so here we go anyway:
   return std::numeric_limits<T>::is_specialized ? (val > 0 ? (std::numeric_limits<T>::max)() : -(std::numeric_limits<T>::max)()): val;

but it should read

   // This may or may not do the right thing, but the user asked for the error
   // to be ignored so here we go anyway:
   return std::numeric_limits<TargetType>::is_specialized ? (val > 0 ? (std::numeric_limits<TargetType>::max)() : (std::numeric_limits<TargetType>::min)()): val;

Two changes:

• template specialized on the TargetType
• return the minimum value rather than the negative version of the max value (important for unsigned types)

[John Maddock]

Confirmed, testing fix.

[John Maddock]

(In [82885]) Fix raise_rounding_error to return the correct result (and type) when an error occurs. Fixes #7905.