On Thursday, 17 March 2016 at 09:59:41 UTC, Dominikus Dittes
Scherkl wrote:
Or you can use an improved opCmp implementation in the
compiler, that only add additional runtime cost, if someone is
stupid enough to compare signed with unsigned values - but
yield the correct result:
For the signed/unsigned comparison problem, specifically - I
agree that would be the best solution (and is what D should have
done originally). There's nothing "stupid" about doing mixed
comparisons correctly, either - I doubt the performance hit would
even be measurable in most code.
(`checkedint` solves a lot of other problems; the 30% performance
hit does *not* come from properly handling signed/unsigned
comparisons. A library-level wrapper that only fixed that one
problem, and none of the others, could be super fast with proper
optimizations.)
However, Walter and Andrei already decided that we should just
have a warning for unsafe comparisons, instead. Just trying to
get the warning implemented has dragged on for years already, so
I'm not anxious to go back to square one by arguing for a
different solution entirely...
