Tail call optimization: enable by default?
|Target version:||next minor|
Some hours ago, Matz proposed turning on "tail call optimization" by default from Ruby 2.0.
What do you think about it?
Tail call: Method invocation at last of method body.
Tail call optimization: Eliminating the new stack frame creation when method invocation is "tail call".
For exmaple, the method bar() is located at last of method foo(), so bar() is "tail call".
In this case, after invocation of method bar(), foo()'s method frame information (which contains local variables, program counter, stack pointer and so on) is no longer needed because method foo() doesn't work after that (correctly, method foo() only does "return").
Next example, a simple recursion code by foo(). Of course, foo() is "tail call".
Current Ruby causes stack overflow error because such recursion consumes the (VM) stack. However, using tail call optimization, VM doesn't consume stack frame any more.
Such recursion can be converted to simple loop:
Someone calls tail-call opt as "tail recursion optimization" because recursion is famous use-case (*1).
*1: Generally, tail-recursion optimization includes another optimization technique - "call" to "jump" translation. In my opinion, it is difficult to apply this optimization because recognizing "recursion" is difficult in Ruby's world.
Next example. fact() method invocation in "else" clause is not a "tail call".
if n < 2
n * fact(n-1)
If you want to use tail-call optimization on fact() method, you need to change fact() method as follows (continuation passing style).
def fact(n, r)
if n < 2
In this case, fact() is tail-call (and a bit difficult to read/write).
Of course, the following code is easy to understand and short.
Last examples. Recognizing tail-call is a bit difficult.
bar2() # not a tail-call
bar3() # not a tail-call
bar4() # not a tail-call
bar5() # tail-call!
return bar("break") # tail-call? (current CRuby can't handle "break" in eval().
CRuby 1.9 has a code tail-call optimization (not tested yet. maybe there are several bugs). However, it is off by default because of several problems described in next section.
- (1) backtrace: Eliminating method frame means eliminating backtrace.
- (2) settracefunc(): It is difficult to probe "return" event for tail-call methods.
- (3) semantics: It is difficult to define tail-call in document (half is joking, but half is serious)
Maybe (1) has big impact for ordinal users.
In this case, backtrace information only include "baz", because bar() in foo and baz() in bar are "tail-call". Users can't see eliminated frame information in backtrace.
This is why we don't introduce them by default to Ruby 1.9.
Many people ask us that "why don't you introduce tail-call optimization? it is very easy technique." I wrote reasons above.
Matz said "it seems small impact enough. Go ahead". (I doubt it ;P )
Yusuke Endo proposed that introducing special form (for example, send_tail(:foo, ...)) to declare tail call. Users only use this special form when the backtrace information can be eliminated (*2).
(*2) Special form "goto foo()" is nice joking feature :) I like it but I believe Matz will reject it.
Akira Tanaka introduced that special backtrace notation like:
... (eliminated by tail call optimization)
to represent eliminating method invocation information. We can know they were eliminated (good) but we can't know what method frames were eliminated (bad).
Matz wanted to introduce it. However it has several problems. Should we turn on this optimization by default?
Sorry for long (and poor English) article. Comments and proposals are welcome (with short English, long Ruby codes ;p).
#3 Updated by Koichi Sasada almost 2 years ago
(2012/06/18 20:29), trans (Thomas Sawyer) wrote:
However, in the the case of tail-recursion, it could be ok since the stack entry is essentially a facsimile of the previous.
Do you mean the following code isn't applied tail call optimization?
# double recursion (foo -> bar -> foo -> bar -> ...)
I agree it is one option for this problem.
// SASADA Koichi at atdot dot net