On Sunday, 9 February 2014 at 13:04:20 UTC, Dicebot wrote:
You seem to forget that actually writing data to socket takes time and this time is longer than one needed for most simple stack unwinding. Fun fact - even if you copy rendering code to static exception example it will still be ~5-10% faster because it will cancel pending async socket output and terminate fiber early.
Oh, I'm not surprised that it can be faster than the non-exception version, after all, it does less work.
What troubles me is the large speedup if you use both a statically allocated exception, and don't generate a stack trace. Neither of them alone makes that much of a difference. I would have expected the combined performance gain to be the sum of both for themselves, more or less.
But the difference actual difference is so much larger. Where does it come from? It might mean that there is some costly operation that both `new` and `toString()` perform. Or it might actually be a bug, and when fixed, would have worse performance. It seems almost too good to be true...
But yes, those results match my earlier observations - pre-allocated exceptions with no stack trace generation are of fast enough to not be a concern. Impact of both of those factors is huge enough to provide DoS vector.
Yes, if it is really safe to use, vibe.d should introduce a "production" mode, where it neither outputs nor logs stack traces by default. That's a good idea anyway, to keep internal information from leaking out to a potential attacker.
