On Saturday, 14 May 2016 at 09:11:49 UTC, QAston wrote:
On Saturday, 14 May 2016 at 05:46:38 UTC, Ola Fosheim Grøstad
wrote:
In Java all compile time constants are done using strict
settings and it provides a keyword «strictfp» to get strict
behaviour for a particular class/function.
In java everything used to be strictfp (and there was no
keyword), but it was changed to do non-strict arithmetic by
default after a backlash from numeric programmers.
Java had a healthy default, but switched in order to not look so
bad in comparison to C on current day hardware. However, they
retained the ability to get stricter floating point.
At the end of the day there is literally no end to how far you
can move down the line of implementation defined floating point.
Take a look at the ARM instruction set, it makes x86 look high
level. You can even choose how many iterations you want for
complex instructions (i.e. choose the approximation level for
faster execution).
However, these days IEEE754-2008 is becoming available in
hardware and therefore one is better off choosing the most
well-defined semantics for the regular case. It means less
optimization opportunities unless you specify relaxed semantics,
but that isn't such a bad trade off as long as specifying relaxed
semantics is easy.
