Is the ruby operator ||= intelligent?

Question

I have a question regarding the ||= statement in ruby and this is of particular interest to me as I'm using it to write to memcache. What I'm wondering is, does ||= check the receiver first to see if it's set before calling that setter, or is it literally an alias to x = x || y

This wouldn't really matter in the case of a normal variable but using something like:

CACHE[:some_key] ||= "Some String"

could possibly do a memcache write which is more expensive than a simple variable set. I couldn't find anything about ||= in the ruby api oddly enough so I haven't been able to answer this myself.

Of course I know that:

CACHE[:some_key] = "Some String" if CACHE[:some_key].nil?

would achieve this, I'm just looking for the most terse syntax.

Answer 1

This is extremely easy to test:

class MyCache
  def initialize
    @hash = {}
  end

  def []=(key, value)
    puts "Cache key '#{key}' written"
    @hash[key] = value
  end

  def [](key)
    puts "Cache key '#{key}' read"
    @hash[key]
  end
end

Now simply try the ||= syntax:

cache = MyCache.new
cache["my key"] ||= "my value"  # cache value was nil (unset)
# Cache key 'my key' read
# Cache key 'my key' written

cache["my key"] ||= "my value"  # cache value is already set
# Cache key 'my key' read

So we can conclude that no assignment takes place if the cache key already exists.

The following extract from the Rubyspec shows that this is by design and should not be dependent on the Ruby implementation:

describe "Conditional operator assignment 'obj.meth op= expr'" do
  # ...
  it "may not assign at all, depending on the truthiness of lhs" do
    m = mock("object")
    m.should_receive(:foo).and_return(:truthy)
    m.should_not_receive(:foo=)
    m.foo ||= 42

    m.should_receive(:bar).and_return(false)
    m.should_not_receive(:bar=)
    m.bar &&= 42
  end
  # ...
end

In the same file, there is a similar spec for [] and []= that mandates identical behaviour.

Although the Rubyspec is still a work in progress, it has become clear that the major Ruby implementation projects intend to comply with it.

Answer 2

According to §11.3.1.2.2 of the Draft ISO Specification,

CACHE[:some_key] ||= "Some String"

expands to

o = CACHE
*l = :some_key
v = o.[](*l)
w = "Some String"
x = v || w
l << x
o.[]=(*l)
x

Or, in the more general case

primary_expression[indexing_argument_list] ω= expression

(I am using ω here to denote any operator, so it could be ||=, +=, *=, >>=, %=,…)

Expands to:

o = primary_expression
*l = indexing_argument_list
v = o.[](*l)
w = expression
x = v ω w
l << x
o.[]=(*l)
x

So, according to the specification, []= will always get called. But that is actually not the case in current implementations (I tested MRI, YARV, Rubinius, JRuby and IronRuby):

def (h = {}).[]=(k, v) p "Setting #{k} to #{v}"; super end
h[:key] ||= :value # => :value
# "Setting key to value"
h[:key] ||= :value # => :value

So, obviously either the specification is wrong or all five currently released implementations are wrong. And since the purpose of the specification is to describe the behavior of the existing implementations, it's obviously that the specification must be wrong.

In general, as a first approximation

a ||= b

expands to

a || a = b

However, there's all kinds of subleties involved, for example, whether or not a is undefined, whether a is a simple variable or a more complex expression like foo[bar] or foo.bar and so on.

See also some of the other instances of this same question, that have already been asked and answered here on StackOverflow (for example, this one). Also, the question has been discussed so many times on the ruby-talk mailinglist, that there are now discussion threads whose sole purpose it is to summarize the other discussion threads. (Although please note that that list is far from complete.)