泛函编程(29)-泛函实用结构:Trampoline-不再怕StackOverflow

   泛函编程方式其中一个特点就是普遍地使用递归算法,而且有些地方还无法避免使用递归算法。比如说flatMap就是一种推进式的递归算法,没了它就无法使用for-comprehension,那么泛函编程也就无法被称为Monadic Programming了。虽然递归算法能使代码更简洁易明,但同时又以占用堆栈(stack)方式运作。堆栈是软件程序有限资源,所以在使用递归算法对大型数据源进行运算时系统往往会出现StackOverflow错误。如果不想办法解决递归算法带来的StackOverflow问题,泛函编程模式也就失去了实际应用的意义了。

针对StackOverflow问题,Scala compiler能够对某些特别的递归算法模式进行优化:把递归算法转换成while语句运算,但只限于尾递归模式(TCE, Tail Call Elimination),我们先用例子来了解一下TCE吧:

以下是一个右折叠算法例子:

def foldR[A,B](as: List[A], b: B, f: (A,B) => B): B = as match {
	case Nil => b
	case h :: t => f(h,foldR(t,b,f))
}                                                 //> foldR: [A, B](as: List[A], b: B, f: (A, B) => B)B
def add(a: Int, b: Int) = a + b                   //> add: (a: Int, b: Int)Int

foldR((1 to 100).toList, 0, add)                  //> res0: Int = 5050
foldR((1 to 10000).toList, 0, add)                //> java.lang.StackOverflowError

以上的右折叠算法中自引用部分不在最尾部,Scala compiler无法进行TCE,所以处理一个10000元素的List就发生了StackOverflow。

再看看左折叠:

def foldL[A,B](as: List[A], b: B, f: (B,A) => B): B = as match {
	case Nil => b
	case h :: t => foldL(t,f(b,h),f)
}                                                 //> foldL: [A, B](as: List[A], b: B, f: (B, A) => B)B
foldL((1 to 100000).toList, 0, add)               //> res1: Int = 705082704

在这个左折叠例子里自引用foldL出现在尾部位置,Scala compiler可以用TCE来进行while转换:

  def foldl2[A,B](as: List[A], b: B,
                 f: (B,A) => B): B = {
    var z = b
    var az = as
    while (true) {
      az match {
        case Nil => return z
        case x :: xs => {
          z = f(z, x)
          az = xs
        }
      }
    }
    z
  }

经过转换后递归变成Jump,程序不再使用堆栈,所以不会出现StackOverflow。

但在实际编程中,统统把递归算法编写成尾递归是不现实的。有些复杂些的算法是无法用尾递归方式来实现的,加上JVM实现TCE的能力有局限性,只能对本地(Local)尾递归进行优化。

我们先看个稍微复杂点的例子:

def even[A](as: List[A]): Boolean = as match {
	case Nil => true
	case h :: t => odd(t)
}                                                 //> even: [A](as: List[A])Boolean
def odd[A](as: List[A]): Boolean = as match {
	case Nil => false
	case h :: t => even(t)
}                                                 //> odd: [A](as: List[A])Boolean

在上面的例子里even和odd分别为跨函数的各自的尾递归,但Scala compiler无法进行TCE处理,因为JVM不支持跨函数Jump:

even((1 to 100).toList)                           //> res2: Boolean = true
even((1 to 101).toList)                           //> res3: Boolean = false
odd((1 to 100).toList)                            //> res4: Boolean = false
odd((1 to 101).toList)                            //> res5: Boolean = true
even((1 to 10000).toList)                         //> java.lang.StackOverflowError

处理10000个元素的List还是出现了StackOverflowError

我们可以通过设计一种数据结构实现以heap交换stack。Trampoline正是专门为解决StackOverflow问题而设计的数据结构:

trait Trampoline[+A] {
 final def runT: A = this match {
  	case Done(a) => a
  	case More(k) => k().runT
  }
}
case class Done[+A](a: A) extends Trampoline[A]
case class More[+A](k: () => Trampoline[A]) extends Trampoline[A]

Trampoline代表一个可以一步步进行的运算。每步运算都有两种可能:Done(a),直接完成运算并返回结果a,或者More(k)运算k后进入下一步运算;下一步又有可能存在Done和More两种情况。注意Trampoline的runT方法是明显的尾递归,而且runT有final标示,表示Scala可以进行TCE。

有了Trampoline我们可以把even,odd的函数类型换成Trampoline:

def even[A](as: List[A]): Trampoline[Boolean] = as match {
	case Nil => Done(true)
	case h :: t => More(() => odd(t))
}                                                 //> even: [A](as: List[A])ch13.ex1.Trampoline[Boolean]
def odd[A](as: List[A]): Trampoline[Boolean] = as match {
	case Nil => Done(false)
	case h :: t => More(() => even(t))
}                                                 //> odd: [A](as: List[A])ch13.ex1.Trampoline[Boolean]

我们可以用Trampoline的runT来运算结果:

even((1 to 10000).toList).runT                    //> res6: Boolean = true
even((1 to 10001).toList).runT                    //> res7: Boolean = false
odd((1 to 10000).toList).runT                     //> res8: Boolean = false
odd((1 to 10001).toList).runT                     //> res9: Boolean = true

这次我们不但得到了正确结果而且也没有发生StackOverflow错误。就这么简单?

我们再从一个比较实际复杂一点的例子分析。在这个例子中我们遍历一个List并维持一个状态。我们首先需要State类型:

case class State[S,+A](runS: S => (A,S)) {
import State._
	def flatMap[B](f: A => State[S,B]): State[S,B] = State[S,B] {
		s => {
			val (a1,s1) = runS(s)
			f(a1) runS s1
		}
	}
  def map[B](f: A => B): State[S,B] = flatMap( a => unit(f(a)))
}
object State {
	def unit[S,A](a: A) = State[S,A] { s => (a,s) }
	def getState[S]: State[S,S] = State[S,S] { s => (s,s) }
	def setState[S](s: S): State[S,Unit] = State[S,Unit] { _ => ((),s)}
}

再用State类型来写一个对List元素进行序号标注的函数:

def zip[A](as: List[A]): List[(A,Int)] = {
	as.foldLeft(
	  unit[Int,List[(A,Int)]](List()))(
	  (acc,a) => for {
	    xs <- acc
	    n <- getState[Int]
	    _ <- setState[Int](n + 1)
	  } yield (a,n) :: xs
	).runS(0)._1.reverse
}                                                 //> zip: [A](as: List[A])List[(A, Int)]

运行一下这个zip函数:

zip((1 to 10).toList)                             //> res0: List[(Int, Int)] = List((1,0), (2,1), (3,2), (4,3), (5,4), (6,5), (7,6
                                                  //| ), (8,7), (9,8), (10,9))

结果正确。如果针对大型的List呢?

zip((1 to 10000).toList)                          //> java.lang.StackOverflowError

按理来说foldLeft是尾递归的,怎么StackOverflow出现了。这是因为State组件flatMap是一种递归算法,也会导致StackOverflow。那么我们该如何改善呢?我们是不是像上面那样把State转换动作的结果类型改成Trampoline就行了呢?

case class State[S,A](runS: S => Trampoline[(A,S)]) {
	def flatMap[B](f: A => State[S,B]): State[S,B] = State[S,B] {
		s => More(() => {
			val (a1,s1) = runS(s).runT
			More(() => f(a1) runS s1)
		})
	}
  def map[B](f: A => B): State[S,B] = flatMap( a => unit(f(a)))
}
object State {
	def unit[S,A](a: A) = State[S,A] { s => Done((a,s)) }
	def getState[S]: State[S,S] = State[S,S] { s => Done((s,s)) }
	def setState[S](s: S): State[S,Unit] = State[S,Unit] { _ => Done(((),s))}
}
trait Trampoline[+A] {
 final def runT: A = this match {
  	case Done(a) => a
  	case More(k) => k().runT
  }
}
case class Done[+A](a: A) extends Trampoline[A]
case class More[+A](k: () => Trampoline[A]) extends Trampoline[A]

def zip[A](as: List[A]): List[(A,Int)] = {
	as.foldLeft(
	  unit[Int,List[(A,Int)]](List()))(
	  (acc,a) => for {
	    xs <- acc
	    n <- getState[Int]
	    _ <- setState[Int](n + 1)
	  } yield (a,n) :: xs
	).runS(0).runT._1.reverse
}                                                 //> zip: [A](as: List[A])List[(A, Int)]
zip((1 to 10).toList)                             //> res0: List[(Int, Int)] = List((1,0), (2,1), (3,2), (4,3), (5,4), (6,5), (7,
                                                  //| 6), (8,7), (9,8), (10,9))

在这个例子里我们把状态转换函数 S => (A,S) 变成 S => Trampoline[(A,S)]。然后把其它相关函数类型做了相应调整。运行zip再检查结果:结果正确。那么再试试大型List:

zip((1 to 10000).toList)                          //> java.lang.StackOverflowError

还是会出现StackOverflow。这次是因为flatMap中的runT不在尾递归位置。那我们把Trampoline变成Monad看看如何?那我们就得为Trampoline增加一个flatMap函数:

trait Trampoline[+A] {
 final def runT: A = this match {
  	case Done(a) => a
  	case More(k) => k().runT
  }
  def flatMap[B](f: A => Trampoline[B]): Trampoline[B] = {
  	this match {
  		case Done(a) => f(a)
  		case More(k) => f(runT)
  	}
  }
}
case class Done[+A](a: A) extends Trampoline[A]
case class More[+A](k: () => Trampoline[A]) extends Trampoline[A]

这样我们可以把State.flatMap调整成以下这样:

case class State[S,A](runS: S => Trampoline[(A,S)]) {
	def flatMap[B](f: A => State[S,B]): State[S,B] = State[S,B] {
		s => More(() => {
//			val (a1,s1) = runS(s).runT
//			More(() => f(a1) runS s1)
		  runS(s) flatMap {   // runS(s) >>> Trampoline
		    case (a1,s1) => More(() => f(a1) runS s1)
		  }
		})
	}
  def map[B](f: A => B): State[S,B] = flatMap( a => unit(f(a)))
}

现在我们把递归算法都推到了Trampoline.flatMap这儿了。不过Trampoline.flatMap的runT引用f(runT)不在尾递归位置,所以这样调整还不足够。看来核心还是要解决flatMap尾递归问题。我们可以再为Trampoline增加一个状态结构FlatMap然后把flatMap函数引用变成类型实例构建(type construction):

case class Done[+A](a: A) extends Trampoline[A]
case class More[+A](k: () => Trampoline[A]) extends Trampoline[A]
case class FlatMap[A,B](sub: Trampoline[A], k: A => Trampoline[B]) extends Trampoline[B]

case class FlatMap这种Trampoline状态意思是先引用sub然后把结果传递到下一步k再运行k:基本上是沿袭flatMap功能。再调整Trampoline.resume, Trampoline.flatMap把FlatMap这种状态考虑进去:

trait Trampoline[+A] {
 final def runT: A = resume match {
  	case Right(a) => a
  	case Left(k) => k().runT
  }
  def flatMap[B](f: A => Trampoline[B]): Trampoline[B] = {
  	this match {
//  		case Done(a) => f(a)
//  		case More(k) => f(runT)
      case FlatMap(a,g) => FlatMap(a, (x: Any) => g(x) flatMap f)
      case x => FlatMap(x, f)
  	}
  }
  def map[B](f: A => B) = flatMap(a => Done(f(a)))
  def resume: Either[() => Trampoline[A], A] = this match {
    case Done(a) => Right(a)
    case More(k) => Left(k)
    case FlatMap(a,f) => a match {
    	case Done(v) => f(v).resume
    	case More(k) => Left(() => k() flatMap f)
    	case FlatMap(b,g) => FlatMap(b, (x: Any) => g(x) flatMap f).resume
    }
  }
}
case class Done[+A](a: A) extends Trampoline[A]
case class More[+A](k: () => Trampoline[A]) extends Trampoline[A]
case class FlatMap[A,B](sub: Trampoline[A], k: A => Trampoline[B]) extends Trampoline[B]

在以上对Trampoline的调整里我们引用了Monad的结合特性(associativity):

FlatMap(FlatMap(b,g),f) == FlatMap(b,x => FlatMap(g(x),f)

重新右结合后我们可以用FlatMap正确表达复数步骤的运算了。

现在再试着运行zip:

def zip[A](as: List[A]): List[(A,Int)] = {
	as.foldLeft(
	  unit[Int,List[(A,Int)]](List()))(
	  (acc,a) => for {
	    xs <- acc
	    n <- getState[Int]
	    _ <- setState[Int](n + 1)
	  } yield (a,n) :: xs
	).runS(0).runT._1.reverse
}                                                 //> zip: [A](as: List[A])List[(A, Int)]
zip((1 to 10000).toList)                          //> res0: List[(Int, Int)] = List((1,0), (2,1), (3,2), (4,3), (5,4), (6,5), (7,

这次运行正常,再不出现StackOverflowError了。

实际上我们可以考虑把Trampoline当作一种通用的堆栈溢出解决方案。

我们首先可以利用Trampoline的Monad特性来调控函数引用,如下:

val x = f() 
val y = g(x) 
h(y)
//以上这三步函数引用可以写成:
for {
 x <- f()
 y <- g(x)
 z <- h(y) 
} yield z


举个实际例子:

implicit def step[A](a: => A): Trampoline[A] = {
	More(() => Done(a))
}                                                 //> step: [A](a: => A)ch13.ex1.Trampoline[A]
def getNum: Double = 3                            //> getNum: => Double
def addOne(x: Double) = x + 1                     //> addOne: (x: Double)Double
def timesTwo(x: Double) = x * 2                   //> timesTwo: (x: Double)Double
(for {
	x <- getNum
	y <- addOne(x)
	z <- timesTwo(y)
} yield z).runT                                   //> res6: Double = 8.0

又或者:

def fib(n: Int): Trampoline[Int] = {
	if (n <= 1) Done(n) else for {
		x <- More(() => fib(n-1))
		y <- More(() => fib(n-2))
	} yield x + y
}                                                 //> fib: (n: Int)ch13.ex1.Trampoline[Int]
(fib(10)).runT                                    //> res7: Int = 55

从上面得出我们可以用flatMap来对Trampoline运算进行流程控制。另外我们还可以通过把多个Trampoline运算交叉组合来实现并行运算:

trait Trampoline[+A] {
 final def runT: A = resume match {
  	case Right(a) => a
  	case Left(k) => k().runT
  }
  def flatMap[B](f: A => Trampoline[B]): Trampoline[B] = {
  	this match {
//  		case Done(a) => f(a)
//  		case More(k) => f(runT)
      case FlatMap(a,g) => FlatMap(a, (x: Any) => g(x) flatMap f)
      case x => FlatMap(x, f)
  	}
  }
  def map[B](f: A => B) = flatMap(a => Done(f(a)))
  def resume: Either[() => Trampoline[A], A] = this match {
    case Done(a) => Right(a)
    case More(k) => Left(k)
    case FlatMap(a,f) => a match {
    	case Done(v) => f(v).resume
    	case More(k) => Left(() => k() flatMap f)
    	case FlatMap(b,g) => FlatMap(b, (x: Any) => g(x) flatMap f).resume
    }
  }
  def zip[B](tb: Trampoline[B]): Trampoline[(A,B)] = {
    (this.resume, tb.resume) match {
    	case (Right(a),Right(b)) => Done((a,b))
    	case (Left(f),Left(g)) => More(() => f() zip g())
    	case (Right(a),Left(k)) => More(() => Done(a) zip k())
    	case (Left(k),Right(a)) => More(() => k() zip Done(a))
    }
  }
}
case class Done[+A](a: A) extends Trampoline[A]
case class More[+A](k: () => Trampoline[A]) extends Trampoline[A]
case class FlatMap[A,B](sub: Trampoline[A], k: A => Trampoline[B]) extends Trampoline[B]

我们可以用这个zip函数把几个Trampoline运算交叉组合起来实现并行运算:

def hello: Trampoline[Unit] = for {
	_ <- print("Hello ")
	_ <- println("World!")
} yield ()                                        //> hello: => ch13.ex1.Trampoline[Unit]

(hello zip hello zip hello).runT                  //> Hello Hello Hello World!
                                                  //| World!
                                                  //| World!
                                                  //| res8: ((Unit, Unit), Unit) = (((),()),())

用Trampoline可以解决StackOverflow这个大问题。现在我们可以放心地进行泛函编程了。



















评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值