Iterators and Streams in Rust and Haskell

10 Jul 2017 Michael Snoyman

Streaming data is a problem domain I've played with a lot in Haskell. In Haskell, the closest we come to built-in streaming data support is laziness-by-default, which doesn't fully capture streaming data. (I'm not going into those details today, but if you want to understand this better, there's plenty of information in the conduit tutorial.) Real streaming data is handled at the library level in Haskell, with many different options available.

Rust does things differently: it bakes in a concept called iterators not only with the standard library, but the language itself: for loops are built-in syntax for iterators. There are some interesting trade-offs to discuss regarding solving a problem in the language itself versus a library, which I'm not going to get into.

Also, Rust's approach follows a state machine design, as opposed to many Haskell libraries which use coroutines. That choice turns out to be pretty crucial to getting good performance, and applies in the Haskell world as well. In fact, I've already blogged about this concept with my aptly-named Vegito concept. For those familiar with it: you'll see some crossovers in this blog post, but prior knowledge isn't necessary.

While digging into the implementation of iterators in Rust, I found it very enlightening how the design differed from what idiomatic Haskell would look like. Trying to mirror the design from one language in the other really demonstrates some profound differences in the languages, which is what I'm going to try and dive in on today.

To motivate the examples here, we're going to try to perform the same computation many times: stream the numbers from 1 to 1,000,000, filter to just the evens, multiply every number by 2, and then sum them up. You can find all of the code in a Gist. Here's an overview of the benchmark results, with many more details below:

Benchmark results

Also, each function takes an integer argument to tell it the highest value it should count it (which is always 1,000,000). Criterion requires this kind of argument be present to ensure that the Haskell compiler (GHC) doesn't optimize away our function call and give us bogus benchmarking results.

Baseline and cheating

The Gist includes code in Haskell, C, and Rust, with many different implementations of the same kind of function. The Haskell code foreign imports both Rust and C and uses the Criterion benchmarking library to benchmark them. To start off, I implemented a cheating version of each benchmark. Instead of actually filtering and doubling, it just increments the counter by 4 each time and adds it to a total. For example, in C this looks like:

int c_cheating(int high) {
  int total = 0;
  int i;
  high *= 2;
  for (i = 4; i <= high; i += 4) {
    total += i;
  }
  return total;
}

By contrast, the non-cheating loop version in C is:

int c_loop(int high) {
  int total = 0;
  int i;
  for (i = 1; i <= high; ++i) {
    if (i % 2 == 0) {
      total += i * 2;
    }
  }
  return total;
}

Similarly, we have cheating and loop implementations in Rust:

#[no_mangle]
pub extern fn rust_cheating(high: isize) -> isize {
    let mut total = 0;
    let mut i = 4;
    let high = high * 2;
    while i <= high {
        total += i;
        i += 4;
    }
    total
}

#[no_mangle]
pub extern fn rust_loop(mut high: isize) -> isize {
    let mut total = 0;
    while high > 0 {
        if high % 2 == 0 {
            total += high << 1;
        }

        high -= 1;
    }

    total
}

And in Haskell. Haskell uses recursion in place of looping, but under the surface the compiler turns it into a loop at the assembly level.

haskellCheating :: Int -> Int
haskellCheating high' =
  loop 0 4
  where
    loop !total !i
      | i <= high = loop (total + i) (i + 4)
      | otherwise = total
    high = high' * 2

recursion :: Int -> Int
recursion high =
  loop 1 0
  where
    loop !i !total
      | i > high = total
      | even i = loop (i + 1) (total + i * 2)
      | otherwise = loop (i + 1) total

These two sets of tests give us some baseline numbers to compare everything else we're going to look at. First, the cheating results:

benchmarking C cheating
time                 87.13 ns   (86.26 ns .. 87.99 ns)
                     0.999 R²   (0.999 R² .. 1.000 R²)
mean                 86.87 ns   (86.08 ns .. 87.57 ns)
std dev              2.369 ns   (1.909 ns .. 3.127 ns)
variance introduced by outliers: 41% (moderately inflated)

benchmarking Rust cheating
time                 174.7 μs   (172.8 μs .. 176.9 μs)
                     0.998 R²   (0.997 R² .. 0.999 R²)
mean                 175.2 μs   (173.3 μs .. 177.3 μs)
std dev              6.869 μs   (5.791 μs .. 8.762 μs)
variance introduced by outliers: 37% (moderately inflated)

benchmarking Haskell cheating
time                 175.2 μs   (172.2 μs .. 178.9 μs)
                     0.998 R²   (0.995 R² .. 0.999 R²)
mean                 174.6 μs   (172.9 μs .. 176.8 μs)
std dev              6.427 μs   (4.977 μs .. 9.365 μs)
variance introduced by outliers: 34% (moderately inflated)

You may be surprised that C is about twice as fast as Rust and Haskell. But look again: C is taking 87 nanoseconds, while Rust and Haskell both take about 175 microseconds. It turns out that GCC it able to optimize this into a downward-counting loop, which drastically improves the performance. We can do similar things in Rust and Haskell to get down to nanosecond-level performance, but that's not our goal today. I do have to say: well done GCC.

The non-cheating results still favor C, but not to the same extent:

benchmarking C loop
time                 636.3 μs   (631.8 μs .. 640.5 μs)
                     0.999 R²   (0.998 R² .. 0.999 R²)
mean                 636.3 μs   (629.9 μs .. 642.9 μs)
std dev              22.67 μs   (18.76 μs .. 27.87 μs)
variance introduced by outliers: 27% (moderately inflated)

benchmarking Rust loop
time                 632.8 μs   (623.8 μs .. 640.4 μs)
                     0.999 R²   (0.998 R² .. 0.999 R²)
mean                 626.9 μs   (621.4 μs .. 631.9 μs)
std dev              18.45 μs   (14.97 μs .. 23.18 μs)
variance introduced by outliers: 20% (moderately inflated)

benchmarking Haskell recursion
time                 741.9 μs   (733.1 μs .. 755.0 μs)
                     0.998 R²   (0.996 R² .. 0.999 R²)
mean                 748.7 μs   (739.8 μs .. 762.8 μs)
std dev              36.37 μs   (29.18 μs .. 52.40 μs)
variance introduced by outliers: 41% (moderately inflated)

EDIT Originally this article listed a performance number for the C loop as being faster than Rust. However, as pointed out on Reddit, the code in question was mistakenly using int instead of int64_t to match the Rust and Haskell behavior. The numbers have been updated.

All of the results are the same order of magnitude. C and Rust are neck and neck, with Haskell lagging by about 15%. Understanding the differences between the languages' performance would be an interesting topic in and of itself, but our goal today is to compare the higher-level APIs and see how they affect performance within each language. So for the rest of this post, we'll focus on comparing intra-language performance numbers.

Rust's iterators

OK, with that out of the way, let's look at Rust's implementation using iterators. The code is concise, readable, and elegant:

#[no_mangle]
pub extern fn rust_iters(high: isize) -> isize {
    (1..high + 1)
        .filter(|x| x % 2 == 0)
        .map(|x| x * 2)
        .sum()
}

We can compare this pretty directly with a Haskell implementation using lists or vectors:

list :: Int -> Int
list high =
  sum $ map (* 2) $ filter even $ enumFromTo 1 high

unboxedVector :: Int -> Int
unboxedVector high =
  VU.sum $ VU.map (* 2) $ VU.filter even $ VU.enumFromTo 1 high

This is the first interesting API difference between Haskell and Rust. With Haskell, sum, map, and filter are each functions which are applied to an existing list or vector. You'll notice that, in the vector case, we need to use a qualified import VU. to ensure we're getting the correct version of the function. By contrast, in Rust, we're simply calling methods on the Iterator trait. This means that no namespacing is necessary, but on the other hand adding a new iterator adapter would mean the new function would not follow the same function call syntax as the others. (To me, this seems like a watered down version of the expression problem.)

EDIT As pointed out on Reddit, an extension trait can allow new methods to be added to all iterators.

This doesn't seem like a big deal, but it does show an inherent difference in how namespacing is handled in the two languages, and the impact is fairly ubiquitous. I'd argue that this is a fairly surface-level distinction, but an important one to note.

benchmarking Rust iters
time                 919.5 μs   (905.5 μs .. 936.0 μs)
                     0.998 R²   (0.998 R² .. 0.999 R²)
mean                 919.1 μs   (910.4 μs .. 926.7 μs)
std dev              28.63 μs   (24.52 μs .. 33.91 μs)
variance introduced by outliers: 21% (moderately inflated)

benchmarking Haskell unboxed vector
time                 733.3 μs   (722.6 μs .. 745.2 μs)
                     0.998 R²   (0.996 R² .. 0.999 R²)
mean                 742.4 μs   (732.2 μs .. 752.8 μs)
std dev              33.42 μs   (28.01 μs .. 41.24 μs)
variance introduced by outliers: 36% (moderately inflated)

benchmarking Haskell list
time                 714.0 μs   (707.0 μs .. 720.8 μs)
                     0.999 R²   (0.998 R² .. 0.999 R²)
mean                 710.4 μs   (702.7 μs .. 719.4 μs)
std dev              26.49 μs   (21.79 μs .. 33.72 μs)
variance introduced by outliers: 29% (moderately inflated)

Interesting. While the Haskell benchmarks are about the same as the lower-level recursion approach, the Rust iterator implementation is noticeably slower than the low level loop. I have my own theory on what's going on there, and I'll share it below. Unfortunately, my Rust skills are not strong enough to properly test my hypothesis.

Implementing iterators in Haskell

In Rust, there is an Iterator trait with an associated type Item and a method next. Eliding some extra methods we don't care about here, it looks like this:

pub trait Iterator {
    type Item;
    fn next(&mut self) -> Option<Self::Item>;
}

Let's translate this directly to Haskell:

class Iterator iter where
  type Item iter
  next :: iter -> Maybe (Item iter)

That looks remarkably similar. Some basic differences worth noting:

  • In Rust, we use pub to indicate if something is publicly exposed. In Haskell, we use export lists on the module.
  • Instead of Self, Haskell uses type variables (I called it iter here)
  • Function signature syntax is different
  • Rust tracks information about mutability and references. This is a big difference, and will play out a lot in this post, so I won't detail it too much here
  • Rust says Option, Haskell says Maybe

Let's do a simple implementation in Rust:

struct PowerUp {
    curr: u32,
}

impl Iterator for PowerUp {
    type Item = u32;

    fn next(&mut self) -> Option<u32> {
        if self.curr > 9000 {
            None
        } else {
            let res = self.curr;
            self.curr += 1;
            Some(res)
        }
    }
}

This will iterate through all of the numbers between the starting value and 9000. But there's one line in particular I want to draw your attention to:

self.curr += 1;

That is mutation, and for anyone familiar with Haskell, you know we don't like it very much. In fact, our Iterator typeclass above doesn't work at all, since it has no way of mutating a variable. In order to make this work, we'll need to modify our class. Since we'll have lots of these, I'm going to start numbering them:

class Iterator1 iter where
  type Item1 iter
  next1 :: iter -> IO (Maybe (Item1 iter))

The point is that, each time we iterate our value, it can have some side-effect of mutating a variable. This is a crucial distinction between Rust and Haskell. Rust tracks whether individual values can be mutated or not. And it even defaults to (IMO) the right behavior of immutability. Nonetheless, there is no indication in the type signature of a function that it performs side effects.

Let's power up in Haskell:

data PowerUp = PowerUp (IORef Int)

instance Iterator1 PowerUp where
  type Item1 PowerUp = Int
  next1 (PowerUp ref) = do
    curr <- readIORef ref
    if curr > 9000
      then return Nothing
      else do
        writeIORef ref $! curr + 1
        return $ Just curr

Ignoring unimportant syntax differences:

  • In Haskell, we explicitly need to wrap any mutable field with an IORef (or similar mutable variable)
  • Similarly, we need to use explicit readIORef and writeIORef functions to access the value, as opposed to getting and modifying values directly in Rust.
  • You may have noticed the $! before curr + 1. If you were paying close attention above, in the recursion function, I had something similar with loop !i !total. These are special operators and syntax in Haskell to force evaluation. This is because Haskell is lazy by default, whereas Rust is strict by default.

Alright, so I went ahead and implemented everything with this Iterator1 class and ended up with:

iterator1 :: Int -> Int
iterator1 high =
  unsafePerformIO $
  enumFromTo1 1 high >>=
  filter1 even >>=
  map1 (* 2) >>=
  sum1

We're using unsafePerformIO here, since we want to run this function purely, but it's performing side-effects. A better approach in Haskell is using the ST type, but I'm going for simplicity here. I'm not going to copy the implementation of the types here; please take a look at the Gist if you're curious.

Now let's look at performance:

benchmarking Haskell iterator 1
time                 5.181 ms   (5.108 ms .. 5.241 ms)
                     0.997 R²   (0.993 R² .. 0.999 R²)
mean                 5.192 ms   (5.140 ms .. 5.267 ms)
std dev              179.5 μs   (125.3 μs .. 294.5 μs)
variance introduced by outliers: 16% (moderately inflated)

That's 5 milliseconds, or 5000 microseconds. Meaning, a hell of a lot slower than recursion, lists, and vectors. So we've hit three hurdles:

  • The code looks less clean that the list/vector version
  • We've had to pull out unsafePerformIO
  • And performance sucks

I guess idiomatic Rust isn't so idiomatic in Haskell.

Boxed vs unboxed

Haskell aficiandos may have noticed one major performance bottleneck in what I've presented. IORefs are boxed data structures. Meaning: the data they contain is actually a pointer to a heap object. This means that, each time we write a new Int to an IORef, we have to:

  • Allocate a new heap object to hold it. That heap object needs to be big enough to hold the payload (one machine word) and the data constructor (another machine word).
  • Update the pointer inside the IORef to the new heap object.
  • Garbage collect the old heap object. This won't happen immediately, but is an overhead incurred for each iteration.

Fortunately, there's a workaround for this: unboxed references. There's a library providing them, and switching over to them in our implementation drops the runtime to:

benchmarking Haskell iterator 1 unboxed
time                 2.966 ms   (2.938 ms .. 2.995 ms)
                     0.999 R²   (0.999 R² .. 1.000 R²)
mean                 2.974 ms   (2.952 ms .. 3.007 ms)
std dev              84.05 μs   (57.67 μs .. 145.2 μs)
variance introduced by outliers: 13% (moderately inflated)

Better, but still not great. The simple fact is that Haskell is not optimized for dealing with mutable data. There are some use cases that still work well for mutable data in Haskell, but this kind of low level, tight inner loop isn't one of them.

As a side note: as much as I'm implying that boxed references are a terrible thing in Haskell, they have some great advantages. The biggest is atomic operations: an operation like atomicModifyIORef or the entirety of Software Transactional Memory (STM) leverage the fact that they can create a new multi-machine-word data structure on the heap, and then atomically update the one-machine-word pointer. That's pretty nifty.

Immutable

Alright, the mutable variable approach seems like a dead end. Let's get more idiomatic with Haskell: immutable values. We'll take in an iterator state, and return an updated state:

data Step2 iter
  = Done2
  | Yield2 !iter !(Item2 iter)

class Iterator2 iter where
  type Item2 iter
  next2 :: iter -> Step2 iter

We've added a helper data type to capture what's going on here. At each iteration, we can either be done, or yield both a new value and a new state for our iterator. The IO bit has disappeared, since there are no mutation side-effects occuring. This implementation turned out to be wonderfully inefficient:

benchmarking Haskell iterator 2
time                 15.80 ms   (15.06 ms .. 16.64 ms)
                     0.992    (0.987  .. 0.998 )
mean                 15.16 ms   (14.99 ms .. 15.41 ms)
std dev              561.5 μs   (363.6 μs .. 934.6 μs)
variance introduced by outliers: 11% (moderately inflated)

Why the hate? It turns out that we've just exacerbated our previous problem. Before, each iteration caused a new Int heap object to be created and a pointer to be updated. Now, each iteration causes a bunch of new heap objects, namely all of our data types representing the various functions:

data EnumFromTo2 a = EnumFromTo2 !a !a
data Filter2 a iter = Filter2 !(a -> Bool) !iter
data Map2 a b iter = Map2 !(a -> b) !iter

These are built up and torn down each time we iterate, which is pretty pathetic performance wise. If Haskell could inline the embedded iter fields in each of these data constructors (via the UNPACK pragma), life would be better, but GHC can't unpack polymorphic fields. So we're creating 3 new heap objects each time.

I've included Iterator3 in the Gist, which monomorphizes things a whole bunch to allow inlining. As expected, it improves performance significantly:

benchmarking Haskell iterator 3
time                 8.391 ms   (8.161 ms .. 8.638 ms)
                     0.996 R²   (0.994 R² .. 0.999 R²)
mean                 8.397 ms   (8.301 ms .. 8.517 ms)
std dev              300.0 μs   (218.4 μs .. 443.9 μs)
variance introduced by outliers: 14% (moderately inflated)

But it's still bad. Something more fundamental is wrong here.

Functions are data

Until now in Haskell, we've stuck with the Rust approach of:

  • Declare a trait (type class)
  • Define a data type for each operation
  • Implement the trait/instanciate the class for each data type

This seems to work really well for Rust (more on that below). But it's neither idiomatic Haskell code, nor does it play nicely with Haskell's runtime behavior and garbage collector. Let's remember that, in Haskell, functions are data, and completely bypass the typeclass:

data Step4 s a
  = Done4
  | Yield4 s a

data Iterator4 a = forall s. Iterator4 s (s -> Step4 s a)

Our Step4 data type has two type variables: s is the internal state of the iterator, and a is the next value to be yielded. Now the cool part: Iterator4 says "well, the outside world cares about the a type variable, but the internal state is irrelevant." So it uses an existential to say "this works for all possible internal states."

We then have two fields: the current value of the state, and a function that gets the next step from the current state. To really drive this home, we'll need to look at some implementations:

enumFromTo4 :: (Ord a, Num a) => a -> a -> Iterator4 a
enumFromTo4 start high =
  Iterator4 start f
  where
    f i
      | i > high  = Done4
      | otherwise = Yield4 (i + 1) i

We've define a helper function f. This f function remains constant throughout the entire lifetime of enumFromTo4. Only the i value it is passed gets updated. And let's see how we would call one of these Iterator4s:

sum4 :: Num a => Iterator4 a -> a
sum4 (Iterator4 s0 next) =
  loop 0 s0
  where
    loop !total !s1 =
      case next s1 of
        Done4 -> total
        Yield4 s2 x -> loop (total + x) s2

We capture the next function once and then use it throughout the loop. This may not seem too different from previous code: we're still going to need to create a new state value and destroy it each time we iterate. However, that's not the case: GHC is smart enough to realize that our state is just a single machine Int, and ends up storing it in a machine register, bypassing heap allocations entirely.

Don't get too excited yet though. While we've decimated iterator 2 and 3, our performance is still bad:

benchmarking Haskell iterator 4
time                 3.614 ms   (3.559 ms .. 3.669 ms)
                     0.998 R²   (0.997 R² .. 0.999 R²)
mean                 3.590 ms   (3.542 ms .. 3.641 ms)
std dev              151.4 μs   (116.4 μs .. 192.4 μs)
variance introduced by outliers: 24% (moderately inflated)

We've got one more trick up our performance sleeve, after a message from our sponsors.

Why Rust likes data types

We've seen that the Rust implementation uses individual data types for each operation. But surely with its first class functions, it should be able to do the same thing as Haskell, right? I'm not a Rust expert, but I believe the answer is: yes, but the performance will suffer greatly.

To explain why, consider the type of the expression (1..high + 1).filter(|x| x % 2 == 0).map(|x| x * 2):

std::iter::Map<
  std::iter::Filter<
    std::ops::Range<isize>,
    [closure@rust.rs:7:17: 7:31]
  >,
  [closure@rust.rs:8:14: 8:23]
>

As a Haskeller, when I first realized that this type was being employed, I was pretty confused. By contrast, the type of the the equivalent Haskell expression map4 (* 2) $ filter4 even $ enumFromTo4 1 high is just Iterator Int, which seems much more direct.

Here's the rub: Haskell is happy—perhaps even too happy—to just stick data on the heap and forget about it. We don't care about the size of our data as much, since we'll just stick a pointer into a data structure. Rust, by contrast, does really well when data is stored on the stack. In order to make that happen, it needs to know the exact size of the data in question. And therefore, instead of the Filter data structure getting to say "yeah, I just work on any data structure that implements Iterator," it is generic over all possible implementations.

This is similar to the lack of polymorphic unpacking in Haskell that I mentioned aboved, and leads to inherently different coding styles in many cases, including this one. Also, this behavior of Rust is in direct contradiction to the existential we used above to explicitly hide internal state from our type signature, whereas in Rust we're flaunting it.

A single loop

Alright, back to our pessimal performance problems. We could do a bunch of performance analyses right now, look at GHC generated core and assembly, and spend months writing up a paper on how to improve performance. Fortunately, someone else already did it. To understand the problem, let's look at our Iterator4 again. We already saw that there's a loop in the implementation of sum4, as you'd expect. Let's see filter4:

filter4 :: (a -> Bool) -> Iterator4 a -> Iterator4 a
filter4 predicate (Iterator4 s0 next) =
  Iterator4 s0 loop
  where
    loop s1 =
      case next s1 of
        Done4 -> Done4
        Yield4 s2 x
          | predicate x -> Yield4 s2 x
          | otherwise   -> loop s2

Notice the loop here as well: if the predicate fails, we need to drop a value, and therefore need to call next again. It turns out that GHC is really good at optimizing code that has a single loop, but performance degrades terribly when you have two nested loops, like we do here.

The stream fusion paper provides the solution to this problem: extend our Step datatype with a Skip constructor, which indicates "loop again with a new state, but I don't have any new data available."

data Step5 s a
  = Done5
  | Skip5 s
  | Yield5 s a

Then our implementations change a bit. filter5 becomes:

filter5 :: (a -> Bool) -> Iterator5 a -> Iterator5 a
filter5 predicate (Iterator5 s0 next) =
  Iterator5 s0 noloop
  where
    noloop s1 =
      case next s1 of
        Done5 -> Done5
        Skip5 s2 -> Skip5 s2
        Yield5 s2 x
          | predicate x -> Yield5 s2 x
          | otherwise   -> Skip5 s2

Notice the total lack of a loop. If the predicate fails, we simply Skip5. The implementation of sum5 has to change as well:

sum5 :: Num a => Iterator5 a -> a
sum5 (Iterator5 s0 next) =
  loop 0 s0
  where
    loop !total !s1 =
      case next s1 of
        Done5 -> total
        Skip5 s2 -> loop total s2
        Yield5 s2 x -> loop (total + x) s2

Cue the drumroll... and our performance is now:

benchmarking Haskell iterator 5
time                 744.5 μs   (732.1 μs .. 761.7 μs)
                     0.996 R²   (0.994 R² .. 0.998 R²)
mean                 768.6 μs   (757.9 μs .. 780.8 μs)
std dev              38.18 μs   (31.22 μs .. 48.98 μs)
variance introduced by outliers: 41% (moderately inflated)

Whew, we've gone all the way back to recursion-level performance. An astute reader may be wondering why we bothered at all, when lists and vectors got similar performance. A few things:

  • Vectors use this stream fusion technique under the surface
  • Lists use a different fusion framework (build/foldr) which has different tradeoffs than stream fusion, and therefore handles some other cases much worse
  • We can extend this Iterator5 approach to include IO and perform side effects (like reading from a file) between actions, which can't be done with lists

We've ended up with idiomatic Haskell code, involving no unnecessary data types or type classes, leveraging first-class functions, and dealing in immutable data. We've added an optimization specifically tailored for GHC's preferred code structure. And we get relatively simple high level code with great performance.

And finally, along the way, we got to see some places where Rust and Haskell take very different approaches to the same problem. My personal takeaway is that it's pretty astounding that, with its heap-friendly, garbage collected nature, the performance of the Haskell code is competitive with Rust's.

Why are Rust iterators slower than looping?

If you remember, there was a substantial slowdown when going from Rust loops to Rust iterators. This was a bit disappointing to me. I'd like to understand why. Unfortunately, I don't have an answer right now, only a hunch. And that hunch is that the double-inner-loop problem is kicking in. This is just conjecture right now.

I tried implementing a "stream fusion" style implementation in Rust that looks like this:

enum Step<T> {
    Done,
    Skip,
    Yield(T),
}

trait Stream {
    type Item;
    fn next(&mut self) -> Step<Self::Item>;
}

Almost identical to Iterator, except it uses Step instead of Option, allowing the possibility of Skipping. Unfortunately I saw a slowdown there:

benchmarking Rust stream
time                 958.7 μs   (931.2 μs .. 1.007 ms)
                     0.968 R²   (0.925 R² .. 0.999 R²)
mean                 968.0 μs   (944.3 μs .. 1.019 ms)
std dev              124.1 μs   (45.79 μs .. 212.7 μs)
variance introduced by outliers: 82% (severely inflated)

This could be for many reasons, including better optimizations for Option versus my Step enum, or simply my inability to write performant Rust code. (Or that my theory is just dead wrong, and skip only gets in the way.)

Then I decided to try a similar approach, using immutable state values instead of mutable ones, which looked like:

enum StepI<S, T> {
    Done,
    Skip(S),
    Yield(S, T),
}

trait StreamI where Self: Sized {
    type Item;
    fn next(self) -> StepI<Self, Self::Item>;
}

This implementation was a bit faster than the mutable one, most likely due to user error on my part:

benchmarking Rust stream immutable
time                 878.4 μs   (866.9 μs .. 888.9 μs)
                     0.998 R²   (0.997 R² .. 0.999 R²)
mean                 889.1 μs   (878.7 μs .. 906.0 μs)
std dev              44.75 μs   (27.17 μs .. 86.29 μs)
variance introduced by outliers: 41% (moderately inflated)

A big takeaway from me here was the impact of move semantics in Rust. The ability to fully "consume" an input value and prevent it from being used again is the kind of thing I often want to state in Haskell, but am unable to. On the other hand: dealing with moved values feels tricky in Rust, but that's likely just lack of experience speaking.

The final implementation I tried out in Rust was explicitly passing closures around like we do in Haskell (though including mutable variables). I'm not sure I chose the best representation, but ended up with:

struct NoTrait<A> {
    next: Box<(FnMut() -> Option<A>)>,
}

As an example, the range function looked like this:

fn range_nt(mut low: isize, high: isize) -> NoTrait<isize> {
    NoTrait {
        next: Box::new(move || {
            if low >= high {
                None
            } else {
                let res = low;
                low += 1;
                Some(res)
            }
        })
    }
}

This is pretty close in spirit to how we do things in Haskell, and could be modified to be completely non-mutating if desired with explicit state passing. Anyway, the performance turned out to be (as I'd expected) pretty bad:

benchmarking Rust no trait
time                 4.206 ms   (4.148 ms .. 4.265 ms)
                     0.998 R²   (0.998 R² .. 0.999 R²)
mean                 4.197 ms   (4.155 ms .. 4.237 ms)
std dev              134.4 μs   (109.6 μs .. 166.0 μs)
variance introduced by outliers: 15% (moderately inflated)

GHC is optimized for these kinds of cases, since passing around closures and partially applied functions is standard practice in Haskell. In our Iterator5 implementation, GHC will end up inlining all of the intermediate functions, and then see through all of the closure magic to turn our code into a tight inner loop. This is non-idiomatic Rust, and therefore (AFAICT) the compiler is not performing any such optimizations.

Consider the fact that it has to perform explicit function calls at each step of the iteration, I'd say that the fact that the Rust implementation is only an order of magnitude slower than iterators is pretty impressive.

Conclusion

I find the contrasts between these two languages to be very informative. I definitely walked away with a better understanding of Rust after performing this analysis. And at a higher level, I think the Haskell ecosystem can learn from Rust's focus on zero-cost abstractions in our library design a bit more.

I'd love to hear from Rustaceans about why the iterator version of the code is slower than the loop. I'd be especially interested if some of the ideas from stream fusion could be used to help that speed difference disappear.

And finally: GCC deserves a shoutout for optimizing the hell out of its code and confusing me with crazy assembly until Chris Done helped me work through it :).

comments powered by Disqus

Copyright © 2013-2017 FP Complete Corp. All rights reserved