I have used and read about @tailrec
annotation to have a tail recursive method. I have gone through many links that explains it. For example it only works when for self-calling functions and shouldnot be overriden etc.
Everywhere it is mentioned that the compiler optimizes
. But what magic/concept does the compiler do to make it tail recursive. For a simple function below, what does the compiler do:
@tailrec def fact(acc: Int, n: Int): Int = {
if (n <= 1) acc
else fact(n * acc, n - 1)
}
fact(1,10)
I mean does it convert it into a loop where it repeatedly calls it and then returns the final value? Is there any link to paper which explains it
In addition to my comment on your question (repasting the code here):
var acc = 1
var n = 10
start:
if (n <= 1) return acc
else {
acc = n * acc
n = n - 1
goto start
}
I tried compiling the fact
method with a recent build I just happened to have and with scalac -Xprint:all
and somehow the compiler emitted an icode
file. So this really illustrates how it optimizes the tail call:
// methods
def fact(acc: Int (INT), n: Int (INT)): Int {
locals: value acc, value n, value _$this
startBlock: 1
blocks: [1,2,3,4,5]
1:
2 JUMP 2
2: // huynhjl's comment: IF condition is here
3 LOAD_LOCAL(value n)
3 CONSTANT(1)
3 CJUMP (INT)LE ? 3 : 4
3: // huynhjl's comment: first branch of IF, will return acc
3 LOAD_LOCAL(value acc)
3 JUMP 5
5:
2 RETURN(INT)
4: // huynhjl's comment: else branch of IF, update acc and n and jump back
4 LOAD_LOCAL(value n)
4 LOAD_LOCAL(value acc)
4 CALL_PRIMITIVE(Arithmetic(MUL,INT))
4 LOAD_LOCAL(value n)
4 CONSTANT(1)
4 CALL_PRIMITIVE(Arithmetic(SUB,INT))
4 STORE_LOCAL(value n)
4 STORE_LOCAL(value acc)
4 JUMP 2
}
Here is a nice Blog post on the subject
@tailrec
only guarantees that an error is issued if tail call optimization cannot be performed by the compiler. Scala performs tail call optimization by default.
When conditions described by the paper are met, it is possible to keep the last frame, instead of the stack of frames, and to perform a loop. The process is better described here