Re:TerrariaClone – An incomprehensible hellscape of spaghett...
by CydeWeys on Thursday October 12, 2017 @08:03PM
(#15462372)
Why do you say it's faster? It's guaranteed to fail branch prediction one out of ten times. My guess is that'd be a lot slower than using the integer modulo operator, which is not an expensive operation.
[ Parent
| on HN ]
Re:TerrariaClone – An incomprehensible hellscape of spaghett...
by hacking_again on Thursday October 12, 2017 @08:35PM
(#15462513)
On their own, % and / are way slower than +, -, *, <<, >>. Cycle counts depend on your architecture, you can look them up. That mod is so slow and should be avoided is a kind of folklore based in truth - kind of like function calls being slow - but like everything time-sensitive the mistake lies in not profiling before (manual) optimization.
There's an example on SO, I got similar results just now when I replicated it:
https://stackoverflow.com/questions/15596318/is-it-better-to...
It doesn't matter on my machine whether the divisor is 10 or 42 (as in the example), the branching is way faster. Now, maybe if the branching were not in a loop and hence not so easily predicted, it wouldn't make a difference. But if this code is not being used in a loop, optimization may be premature anyway (as indicated in my original comment).
Probably f() has something to do inside the main game loop and gets called on a bunch of objects every frame. I haven't looked at the code enough to know if that's a bottleneck.
Re:TerrariaClone – An incomprehensible hellscape of spaghett...
by 35bge57dtjku on Thursday October 12, 2017 @10:10PM
(#15462819)
> On their own, % and / are way slower than +, -, *, <<, >>.
And the branch instruction is free??
Re:TerrariaClone – An incomprehensible hellscape of spaghett...
by hacking_again on Thursday October 12, 2017 @11:40PM
(#15463117)
More or less, assuming you can predict it. But there's a penalty for misprediction. So it boils down to whether using % frequently (either as a native instruction or as a sequence of instructions) is more or less expensive than predicting a branch frequently, given a certain misprediction rate.
Re:TerrariaClone – An incomprehensible hellscape of spaghett...
by 35bge57dtjku on Saturday October 14, 2017 @05:18PM
(#15474369)
> More or less, assuming you can predict it. But there's a penalty for misprediction.
Yeah, no, that's not free.
Re:TerrariaClone – An incomprehensible hellscape of spaghett...
by hacking_again on Saturday October 14, 2017 @07:13PM
(#15474851)
Ok, so what changes would you make to my explanations?
Re:TerrariaClone – An incomprehensible hellscape of spaghett...
by CUViper on Thursday October 12, 2017 @08:40PM
(#15462531)
At least in C++, GCC and Clang will both use a CMOVE: https://godbolt.org/g/W1GWyP