In Java, when a long value is cast to double value, the IEEE754 round-to-nearest rounding mode is applied when lossless conversion is not possible.
For example, cast from 1L << 55
or (1L << 55) + 8
to double
are lossless. However, any number in between cannot be accurately represented as double
. The IEEE-754 round-to-nearest (ties to even) rounding mode is applied by default. As a result, (1L << 55) + 4)
or below will round down where as (1L << 55) + 5
or up will round up.
Is there any function in standard library or guava that allows me to cast a long to a double using Round toward +∞
or Round toward -∞
mode (chosen on a per-callsite basis)? Assuming there isn't one, how do I write such two functions myself efficiently?
How about this:
static double longRoundUp(long lng)
{
double d = (double)lng;
return lng <= (long)d ? d : Math.nextUp(d);
}
static double longRoundDown(long lng)
{
double d = (double)lng;
return lng >= (long)d ? d : Math.nextDown(d);
}
Test:
long l1 = (1L << 55);
long l2 = (1L << 55) + 8;
long l3 = (1L << 55) + 4;
long l4 = (1L << 55) + 5;
System.out.format("Up %d %.0f%n", l1, longRoundUp(l1));
System.out.format("Up %d %.0f%n", l2, longRoundUp(l2));
System.out.format("Up %d %.0f%n", l3, longRoundUp(l3));
System.out.format("Up %d %.0f%n", l4, longRoundUp(l4));
System.out.println();
System.out.format("Down %d %.0f%n", l1, longRoundDown(l1));
System.out.format("Down %d %.0f%n", l2, longRoundDown(l2));
System.out.format("Down %d %.0f%n", l3, longRoundDown(l3));
System.out.format("Down %d %.0f%n", l4, longRoundDown(l4));
Output:
Up 36028797018963968 36028797018963968
Up 36028797018963976 36028797018963976
Up 36028797018963972 36028797018963976
Up 36028797018963973 36028797018963976
Down 36028797018963968 36028797018963968
Down 36028797018963976 36028797018963976
Down 36028797018963972 36028797018963968
Down 36028797018963973 36028797018963968