> On May 22, 2016, at 6:41 PM, Filip Pizlo <fpi...@apple.com> wrote: > > Hi everyone! > > I’d like us to stop using std::chrono and go back to using doubles for time. > First I list the things that I think we wanted to get from std::chrono - the > reasons why we started switching to it in the first place. Then I list some > disadvantages of std::chrono that we've seen from fixing std::chrono-based > code. Finally I propose some options for how to use doubles for time. > > Why we switched to std::chrono > > A year ago we started using std::chrono for measuring time. std::chrono has > a rich typesystem for expressing many different kinds of time. For example, > you can distinguish between an absolute point in time and a relative time. > And you can distinguish between different units, like nanoseconds, > milliseconds, etc. > > Before this, we used doubles for time. std::chrono’s advantages over doubles > are: > > Easy to remember what unit is used: We sometimes used doubles for > milliseconds and sometimes for seconds. std::chrono prevents you from > getting the two confused. > > Easy to remember what kind of clock is used: We sometimes use the monotonic > clock and sometimes the wall clock (aka the real time clock). Bad things > would happen if we passed a time measured using the monotonic clock to > functions that expected time measured using the wall clock, and vice-versa. > I know that I’ve made this mistake in the past, and it can be painful to > debug. > > In short, std::chrono uses compile-time type checking to catch some bugs. > > Disadvantages of using std::chrono > > We’ve seen some problems with std::chrono, and I think that the problems > outweigh the advantages. std::chrono suffers from a heavily templatized API > that results in template creep in our own internal APIs. std::chrono’s > default of integers without overflow protection means that math involving > std::chrono is inherently more dangerous than math involving double. This is > particularly bad when we use time to speak about timeouts. > > Too many templates: std::chrono uses templates heavily. It’s overkill for > measuring time. This leads to verbosity and template creep throughout common > algorithms that take time as an argument. For example if we use doubles, a > method for sleeping for a second might look like sleepForSeconds(double). > This works even if someone wants to sleep for a nanoseconds, since 0.000001 > is easy to represent using a double. Also, multiplying or dividing a double > by a small constant factor (1,000,000,000 is small by double standards) is > virtually guaranteed to avoid any loss of precision. But as soon as such a > utility gets std::chronified, it becomes a template. This is because you > cannot have sleepFor(std::chrono::seconds), since that wouldn’t allow you to > represent fractions of seconds. This brings me to my next point. > > Overflow danger: std::chrono is based on integers and its math methods do not > support overflow protection. This has led to serious bugs like > https://bugs.webkit.org/show_bug.cgi?id=157924 > <https://bugs.webkit.org/show_bug.cgi?id=157924>. This cancels out the > “remember what unit is used” benefit cited above. It’s true that I know what > type of time I have, but as soon as I duration_cast it to another unit, I may > overflow. The type system does not help! This is insane: std::chrono > requires you to do more work when writing multi-unit code, so that you > satisfy the type checker, but you still have to be just as paranoid around > multi-unit scenarios. Forgetting that you have milliseconds and using it as > seconds is trivially fixable. But if std::chrono flags such an error and you > fix it with a duration_cast (as any std::chrono tutorial will tell you to > do), you’ve just introduced an unchecked overflow and such unchecked > overflows are known to cause bugs that manifest as pages not working > correctly. > > I think that doubles are better than std::chrono in multi-unit scenarios. It > may be possible to have std::chrono work with doubles, but this probably > implies us writing our own clocks. std::chrono’s default clocks use > integers, not doubles. It also may be possible to teach std::chrono to do > overflow protection, but that would make me so sad since using double means > not having to worry about overflow at all. > > The overflow issue is interesting because of its implications for how we do > timeouts. The way to have a method with an optional timeout is to do one of > these things: > > - Use 0 to mean no timeout. > - Have one function for timeout and one for no timeout. > - Have some form of +Inf or INT_MAX to mean no timeout. This makes so much > mathematical sense. > > WebKit takes the +Inf/INT_MAX approach. I like this approach the best > because it makes the most mathematical sense: not giving a timeout is exactly > like asking for a timeout at time-like infinity. When used with doubles, > this Just Works. +Inf is greater than any value and it gets preserved > properly in math (+Inf * real = +Inf, so it survives gracefully in unit > conversions; +Inf + real = +Inf, so it also survives absolute-to-relative > conversions). > > But this doesn’t work with std::chrono. The closest thing to +Inf is > duration::max(), i.e. some kind of UINT_MAX, but this is guaranteed to > overflow anytime it’s converted to a more precise unit of time > (nanoseconds::max() converted to milliseconds is something bogus). It > appears that std::chrono doesn’t have a good story for infinite timeout, > which means that anyone writing a function that can optionally have a timeout > is going to have a bad time. We have plenty of such functions in WebKit. > For example, I’m not sure how to come up with a feel-good solution to > https://bugs.webkit.org/show_bug.cgi?id=157937 > <https://bugs.webkit.org/show_bug.cgi?id=157937> so long as we use > std::chrono. > > Going back to doubles > > Considering these facts, I propose that we switch back to using doubles for > time. We can either simply revert to the way we used doubles before, or we > can come up with some more sophisticated approach that blends the best of > both worlds. I prefer plain doubles because I love simplicity. > > Simply revert to our old ways: I like this approach the best because it > involves only very simple changes. Prior to std::chrono, we used a double to > measure time in seconds. It was understood that seconds was the default > unit. We would use both monotonic and wall clocks, and we used double for > both of them. > > Come up with a new type system: Having learned from std::chrono and doubles, > it seems that the best typesystem for time would comprise three classes: > Seconds, WallTime, and MonotonicTime. Seconds would be a class that holds a > double and supports +/+=/-/-=/</<=/>/>=/==/!= operations, as well as > conversions to a raw double for when you really need it. WallTime and > MonotonicTime would be wrappers for Seconds with a more limited set of > available operations. You can convert WallTime or MonotonicTime to Seconds > and vice-versa, but some operators are overloaded to make casts unnecessary > in most cases (WallTime + Seconds = WallTime, WallTime - WallTime = Seconds, > etc). This would save us from forgetting the unit or the clock. The name of > the Seconds class is a dead give-away, and WallTime and MonotonicTime will > not yield you a value that is unit-sensitive unless you say something like > WallTime::toSeconds(). There will be no easy way to convert WallTime to > MonotonicTime and vice-versa, since we want to discourage such conversions. > > Personally I feel very comfortable with doubles for time. I like to put the > word “Seconds” into variable names and function names (waitForSeconds(double) > is a dead give-away). On the other hand, I sort of like the idea of a type > system to protect clock mix-ups. I think that’s the biggest benefit we got > from std::chrono. > > If it was entirely up to me, I’d go for doubles. I think that there needs to > be a high burden of proof for using types to catch semantic bugs. A type > system *will* slow you down when writing code, so the EV (expected value) of > the time savings from bugs caught early needs to be greater than the EV of > the time lost due to spoonfeeding the compiler or having to remember how to > use those classes. Although I know that using doubles sometimes meant we had > bugs, I don’t think they were frequent or severe enough for the odds to be > good for the Seconds/WallTime/MonotonicTime solution. > > Thoughts?
In this description of the State of Time in WebKit™, I fixated on a few key points: 1 - When we used plain doubles, we had one class of subtle bugs. 2 - When we switched to chrono, we had a different class of subtle bugs. (and template creep) 3 - There exists a solution - non-templated custom classes - that removes both classes of subtle bugs, without the template creep. The WebKit project believes in tool-based assistance (regression tests, benchmarking, healthy family of scripts for both common and critical tasks, etc etc) And it sounds like we have the ability to enlist another tool - the compiler - to prevent a very human class of bugs from creeping in to the project. Since we’ve already acclimated ourselves to using a class-based solution for time/durations (We use chromo already), I disagree with the pessimism about the cost of using Seconds/WT/MT. I think the cost to individual programmers in the future is the same as using chrono, which we already agreed to do. I think we should do Seconds/WT/MT. Thanks, ~Brady
_______________________________________________ webkit-dev mailing list webkit-dev@lists.webkit.org https://lists.webkit.org/mailman/listinfo/webkit-dev
