Free~> Trampoline:OutOfMemoryError的递归程序崩溃

And*_*kin 6 scala tail-recursion out-of-memory scalaz free-monad

假设我只尝试使用一个操作来实现一个非常简单的特定于域的语言:

printLine(line)
Run Code Online (Sandbox Code Playgroud)

然后我想写一个程序,它接受一个整数n作为输入,打印一些东西,如果n可以被10k整除,然后调用自己n + 1,直到n达到一些最大值N.

省略由for-comprehensions引起的所有语法噪音,我想要的是:

@annotation.tailrec def p(n: Int): Unit = {
  if (n % 10000 == 0) printLine("line")
  if (n > N) () else p(n + 1)
}
Run Code Online (Sandbox Code Playgroud)

从本质上讲,它将是一种"fizzbuzz".

以下是使用Scalaz 7.3.0-M7中的Free monad实现此操作的一些尝试:

import scalaz._

object Demo1 {

  // define operations of a little domain specific language
  sealed trait Lang[X]
  case class PrintLine(line: String) extends Lang[Unit]

  // define the domain specific language as the free monad of operations
  type Prog[X] = Free[Lang, X]

  import Free.{liftF, pure}

  // lift operations into the free monad
  def printLine(l: String): Prog[Unit] = liftF(PrintLine(l))
  def ret: Prog[Unit] = Free.pure(())

  // write a program that is just a loop that prints current index 
  // after every few iteration steps
  val mod =  100000
  val N =   1000000

  // straightforward syntax: deadly slow, exits with OutOfMemoryError
  def p0(i: Int): Prog[Unit] = for {
    _ <- (if (i % mod == 0) printLine("i = " + i) else ret)
    _ <- (if (i > N) ret else p0(i + 1))
  } yield ()

  // Same as above, but written out without `for`
  def p1(i: Int): Prog[Unit] = 
    (if (i % mod == 0) printLine("i = " + i) else ret).flatMap{
      ignore1 =>
      (if (i > N) ret else p1(i + 1)).map{ ignore2 => () }
    }

  // Same as above, with `map` attached to recursive call
  def p2(i: Int): Prog[Unit] = 
    (if (i % mod == 0) printLine("i = " + i) else ret).flatMap{
      ignore1 =>
      (if (i > N) ret else p2(i + 1).map{ ignore2 => () })
    }

  // Same as above, but without the `map`; performs ok.
  def p3(i: Int): Prog[Unit] = {
    (if (i % mod == 0) printLine("i = " + i) else ret).flatMap{ 
      ignore1 =>
      if (i > N) ret else p3(i + 1)
    }
  }

  // Variation of the above; Ok.
  def p4(i: Int): Prog[Unit] = (for {
    _ <- (if (i % mod == 0) printLine("i = " + i) else ret)
  } yield ()).flatMap{ ignored2 => 
    if (i > N) ret else p4(i + 1) 
  }

  // try to use the variable returned by the last generator after yield,
  // hope that the final `map` is optimized away (it's not optimized away...)
  def p5(i: Int): Prog[Unit] = for {
    _ <- (if (i % mod == 0) printLine("i = " + i) else ret)
    stopHere <- (if (i > N) ret else p5(i + 1))
  } yield stopHere

  // define an interpreter that translates the programs into Trampoline
  import scalaz.Trampoline
  type Exec[X] = Free.Trampoline[X]  
  val interpreter = new (Lang ~> Exec) {
    def apply[A](cmd: Lang[A]): Exec[A] = cmd match {
      case PrintLine(l) => Trampoline.delay(println(l))
    }
  }

  // try it out
  def main(args: Array[String]): Unit = {
    println("\n p0")
    p0(0).foldMap(interpreter).run // extremely slow; OutOfMemoryError
    println("\n p1")
    p1(0).foldMap(interpreter).run // extremely slow; OutOfMemoryError
    println("\n p2")
    p2(0).foldMap(interpreter).run // extremely slow; OutOfMemoryError
    println("\n p3")
    p3(0).foldMap(interpreter).run // ok 
    println("\n p4")
    p4(0).foldMap(interpreter).run // ok
    println("\n p5")
    p5(0).foldMap(interpreter).run // OutOfMemory
  }
}
Run Code Online (Sandbox Code Playgroud)

不幸的是,直接的转换(p0)似乎运行某种O(N ^ 2)开销,并与OutOfMemoryError崩溃.问题似乎是for-comprehension map{x => ()}在递归调用之后附加了一个p0,这迫使Freemonad用"提示'完成'p0'然后什么都不做"的提醒来填充整个内存.如果我手动"展开" for理解,并flatMap明确地写出最后一个(如p3p4),那么问题就会消失,一切都会顺利进行.然而,这是一个非常脆弱的解决方法:如果我们只是简单地附加一个程序,程序的行为会发生显着变化map(id),而这map(id)在代码中甚至不可见,因为它是由for-comprehension 自动生成的.

在这篇较老的帖子中:https://apocalisp.wordpress.com/2011/10/26/tail-call-elimination-in-scala-monads/ 已反复建议将递归调用包装成一个suspend.以下是对Applicative实例的尝试suspend:

import scalaz._

// Essentially same as in `Demo1`, but this time with 
// an `Applicative` and an explicit `Suspend` in the 
// `for`-comprehension
object Demo2 {

  sealed trait Lang[H]

  case class Const[H](h: H) extends Lang[H]
  case class PrintLine[H](line: String) extends Lang[H]

  implicit object Lang extends Applicative[Lang] {
    def point[A](a: => A): Lang[A] = Const(a)
    def ap[A, B](a: => Lang[A])(f: => Lang[A => B]): Lang[B] = a match {
      case Const(x) => {
        f match {
          case Const(ab) => Const(ab(x))
          case _ => throw new Error
        }
      }
      case PrintLine(l) => PrintLine(l)
    }
  }

  type Prog[X] = Free[Lang, X]

  import Free.{liftF, pure}
  def printLine(l: String): Prog[Unit] = liftF(PrintLine(l))
  def ret: Prog[Unit] = Free.pure(())

  val mod = 100000
  val N = 2000000

  // try to suspend the entire second generator
  def p7(i: Int): Prog[Unit] = for {
    _ <- (if (i % mod == 0) printLine("i = " + i) else ret)
    _ <- Free.suspend(if (i > N) ret else p7(i + 1))
  } yield ()

  // try to suspend the recursive call
  def p8(i: Int): Prog[Unit] = for {
    _ <- (if (i % mod == 0) printLine("i = " + i) else ret)
    _ <- if (i > N) ret else Free.suspend(p8(i + 1))
  } yield ()

  import scalaz.Trampoline
  type Exec[X] = Free.Trampoline[X]

  val interpreter = new (Lang ~> Exec) {
    def apply[A](cmd: Lang[A]): Exec[A] = cmd match {
      case Const(x) => Trampoline.done(x)
      case PrintLine(l) => 
        (Trampoline.delay(println(l))).asInstanceOf[Exec[A]]
    }
  }

  def main(args: Array[String]): Unit = {
    p7(0).foldMap(interpreter).run // extremely slow; OutOfMemoryError
    p8(0).foldMap(interpreter).run // same...
  }
}
Run Code Online (Sandbox Code Playgroud)

插入suspend并没有真正帮助:它仍然很慢,并且与OutOfMemoryErrors 崩溃.

我应该suspend以某种方式使用不同的方式?

也许有一些纯粹的句法补救措施可以使用for-comprehension而不会产生map最终?

如果有人能指出我在这里做错了什么,以及如何修复它,我真的很感激.

Tom*_*ula 3

Scala 编译器添加的多余内容map将递归从尾部位置移动到非尾部位置。Free monad 仍然使该堆栈安全,但空间复杂度变为O(N)而不是O(1)。(具体来说,它仍然不是O(N 2 )。)

是否有可能进行scalac优化,这map是一个单独的问题(我不知道答案)。

我将尝试说明解释p1与 时发生的情况p3。(我将忽略对 的翻译Trampoline,这是多余的(见下文)。)

p3(即没有额外的map

让我使用以下简写:

def cont(i: Int): Unit => Prg[Unit] =
  ignore1 => if (i > N) ret else p3(i + 1)
Run Code Online (Sandbox Code Playgroud)

p3(0)解释如下

p3(0)
printLine("i = " + 0) flatMap cont(0)
// side-effect: println("i = 0")
cont(0)
p3(1)
ret flatMap cont(1)
cont(1)
p3(2)
ret flatMap cont(2)
cont(2)
Run Code Online (Sandbox Code Playgroud)

等等...您会发现任何时候所需的内存量都不会超过某个恒定的上限。

p1(即有额外的map

我将使用以下简写:

def cont(i: Int): Unit => Prg[Unit] =
  ignore1 => (if (i > N) ret else p1(i + 1)).map{ ignore2 => () }

def cpu: Unit => Prg[Unit] = // constant pure unit
  ignore => Free.pure(())
Run Code Online (Sandbox Code Playgroud)

p1(0)解读如下:

p1(0)
printLine("i = " + 0) flatMap cont(0)
// side-effect: println("i = 0")
cont(0)
p1(1) map { ignore2 => () }
// Free.map is implemented via flatMap
p1(1) flatMap cpu
(ret flatMap cont(1)) flatMap cpu
cont(1) flatMap cpu
(p1(2) map { ignore2 => () }) flatMap cpu
(p1(2) flatMap cpu) flatMap cpu
((ret flatMap cont(2)) flatMap cpu) flatMap cpu
(cont(2) flatMap cpu) flatMap cpu
((p1(3) map { ignore2 => () }) flatMap cpu) flatMap cpu
((p1(3) flatMap cpu) flatMap cpu) flatMap cpu
(((ret flatMap cont(3)) flatMap cpu) flatMap cpu) flatMap cpu
Run Code Online (Sandbox Code Playgroud)

等等...您会看到内存消耗线性取决于N。我们刚刚将计算从堆栈移至堆。

要点:为了保持记忆友好,请将递归保持在“尾部位置”,即在(或)Free的右侧。flatMapmap

旁白:翻译为Trampoline是不必要的,因为Free已经被蹦床了。您可以直接解释Id并用于foldMapRec堆栈安全解释:

val idInterpreter = new (Lang ~> Id) {
  def apply[A](cmd: Lang[A]): Id[A] = cmd match {
    case PrintLine(l) => println(l)
  }
}

p0(0).foldMapRec(idInterpreter)
Run Code Online (Sandbox Code Playgroud)

这将为您重新获得一些记忆(但不会使问题消失)。