Previously the on_path method itself would specify a block to use when a
node was matched. This makes it impossible to customize this behaviour
which is needed when a path is used in an operator or predicate. By
letting the #compile() method specify the block other handlers can
overwrite it.
This also comes with some changes to the specs as the old behaviour of
the Evaluator was incorrect. The Evaluator would bail after matching a
single node but instead it's meant to continue until it runs out of
parent nodes.
This currently supports index predicates (e.g. "foo[10]") and predicates
using paths (e.g. foo[bar/baz]). The usage of boolean operators and more
complex expressions has not yet been tested as these are not yet
supported in the first place.
This currently comes with exactly 0% test coverage. Once I've
implemented all required handler methods I'll be updating the current
evaluator tests to use the compiler instead. This removes the need for
writing an entirely new set of tests.
Currently the compiler is only capable of compiling basic expressions
such as "foo", "foo/bar" and "foo[@x="y"]/bar".
This change is broken up in to two parts:
1. Using a Hash to track if a node is already in a NodeSet
2. Only calling take_ownership when an owner is set
== Using a Hash
Previously various methods such as NodeSet#push and NodeSet#unshift
would call Array#include? (on the internal "nodes" Array) to see if a
node is already present in the set. This is quite problematic
performance wise, especially for large NodeSets. In fact, for the
attached benchmark the vast majority of the time was spent in
Array#include? calls.
Because a NodeSet demands ordering of nodes and must be able to access
them by index (something Set can't do without relying on Enumerable), a
Hash is used to separately keep track of what nodes are in a NodeSet.
This means that checking the presence of a node is simply a matter of
checking a Hash key's presence.
== Calling take_ownership
The if-check for the "owner" variable has been moved out of the
"take_ownership" method and into the methods that call "take_ownership".
This ensures the method isn't called in the first place if no owner is
present, at the cost of slightly more code repetition. The same applies
to the "remove_ownership" method.
== Conclusion
The combined result is a speedup of about 50x when running the attached
concurrent_time_bench.rb benchmark.
While the performance difference between the old and new approach is
pretty much negligible, it's simply not needed to use #shift/#unshift
here.
Thanks to Mon_Ouie from the #ruby IRC channel for suggesting this.
Without this the following could happen:
1. Thread A acquires the lock and sets the ownership to A.
2. Thread A yields and returns
3. Thread B tries to acquire the lock
4. At this exact moment Thread A calls the "synchronize" method again
and sees that the "owner" variable is still set to Thread A
5. Both thread A and B can now access the underlying data in parallel,
possibly leading to corrupted objects
This can be demonstrated using the following script:
require 'oga'
lru = Oga::LRU.new(64)
threads = 50.times.map do
Thread.new do
loop do
number = rand(100)
lru[number] = number
end
end
end
threads.each(&:join)
Run this for a while on either JRuby or Rubinius and you'll end up with
errors such as "ConcurrencyError: Detected invalid array contents due to
unsynchronized modifications with concurrent users" on JRuby or
"ArgumentError: negative array size" on Rubinius.
Resetting the owner variable ensures the above can never happen. Thanks
to @chrisseaton for bringing this up earlier today.
Yorick Peterse
Dan Fockler