Why are Scala for loops slower than logically identical while loops?

Question

I love the power of Scala's for comprehensions and the way they integrate with any monadic type with map and flatMap. However, I'd also like to do simple integer loops without a major speed penalty. Why doesn't Scala have the following two logically identical loops run with similar runtime performance or even compile into similar byte code?

// This is slow...
for (i <- 0 until n) println(s"for loop with $i")

// This runs much faster. It runs roughly at the same speed as Java code doing an identical while or for loop.
var i = 0;
while (i < n) {
  println(s"while loop with $i")
  i += 1
}

Just to confirm... You *are* compiling with `-optimize`, aren't you? — Kevin Wright, Jan 27 '14 at 08:15
@KevinWright: I'm having a hard time finding any description of what the `-optimize` option actually does. Do you have a pointer to any description of its effects? — Randall Schulz, Jan 27 '14 at 14:54
@RandallSchulz It varies from version to version, but you can get a fair idea of the possibilities here: http://magarciaepfl.github.io/scala/ — Kevin Wright, Jan 27 '14 at 15:01
@KevinWright: I found that, but it seems to be specific to 2.11 (and just now, that site is down). — Randall Schulz, Jan 27 '14 at 15:03

score 6 · Accepted Answer · answered Jan 27 '14 at 15:09

6

The main (but not only) reason why they're different is boxing.

In the code:

for (i <- 0 until n) println(s"for loop with $i")

You're passing an anonymous function println(s"for loop with $i") into the comprehension for (i <- 0 until n). It's equivalent to:

(0 until n) foreach (i =>
  println(s"for loop with $i")
}

That function is erased in the bytecode, which means that i can't be a primitive int, it has to be boxed as an Integer instead. Java doesn't have Fixnum references to avoid this cost in the same way that e.g. Smalltalk does (especially frustrating, given how much older smalltalk is!)

Using -optimize can help in some cases, especially in trunk builds of scalac.

You can also use scalaxy/loops to speed things up :)

answered Jan 27 '14 at 15:09

Kevin Wright

49,540
9
105
155

Be careful with the -optimize option as some libraries have trouble with it. I believe Akka is one of them. – Giovanni Botta Jan 27 '14 at 15:39
This is informative, but it's still just an excuse. The Scala compiler can clearly fix this. Boxing ints into Integers may be easier and cleaner for the compiler implementation, but there is no reason the byte code can't use unboxed Java primitive ints. – user2684301 Jan 27 '14 at 16:20
2

@user2684301 That would be a cross-project optimisation, which the current stable compiler doesn't do. `Range.foreach` takes a *Generic* function, so optimising would need the implementation of `foreach` to be copied from the standard library into the using code and then specialised - which causes all kinds of fun for binary compatibility. – Kevin Wright Jan 27 '14 at 20:17

Why are Scala for loops slower than logically identical while loops?

1 Answers1