module Vool class SendStatement < Statement attr_reader :name , :receiver , :arguments def initialize(name , receiver , arguments ) @name , @receiver , @arguments = name , receiver , arguments @arguments ||= [] end def collect(arr) @receiver.collect(arr) @arguments.each do |arg| arg.collect(arr) end super end # Sending in a dynamic language is off course not as simple as just calling. # The function that needs to be called depends after all on the receiver, # and no guarantees can be made on what that is. # # It helps to know that usually (>99%) the class of the receiver does not change. # Our stategy then is to cache the functions and only dynamically determine it in # case of a miss (the 1%, and first invocation) # # As cache key we must use the type of the object (which is the first word of _every_ object) # as that is constant, and function implementations depend on the type (not class) # # A Send breaks down to 2 steps: # - Setting up the next message, with receiver, arguments, and (importantly) return address # - a CachedCall , or a SimpleCall, depending on wether the receiver type can be determined # # FIXME: we now presume direct (assignable) values for the arguments and receiver. # in a not so distant future, temporary variables will have to be created # and complex statements hoisted to assign to them. pps: same as in conditions def to_mom( in_method ) if(@receiver.ct_type) simple_call(in_method) else cached_call(in_method) end end def message_setup(in_method) setup = [Mom::MessageSetup.new(in_method)] receiver = @receiver.slot_class.new([:message , :next_message , :receiver] , @receiver) arg_target = [:message , :next_message , :arguments] args = [] @arguments.each_with_index do |arg , index| args << arg.slot_class.new( arg_target + [index] , arg) end setup << Mom::ArgumentTransfer.new( receiver , args ) end def simple_call(in_method) type = @receiver.ct_type called_method = type.resolve_method(@name) raise "No method #{@name} for #{type}" unless called_method Mom::Statements.new( message_setup(in_method) << Mom::SimpleCall.new( called_method) ) end # this breaks cleanly into two parts: # - check the cached type and if neccessary update # - call the cached method def cached_call(in_method) create_tmps(in_method) Mom::Statements.new( cache_check(in_method) + call_cached_method(in_method) ) end def flatten raise "flat" end # check that current type is the cached type # if not, change and find method for the type (simple_call to resolve_method) # conceptually easy in ruby, but we have to compile that "easy" ruby def cache_check(in_method) # if cached_type != current_type # cached_type = current_type # cached_method = current_type.resolve_method(method.name) if_true = [build_type_cache_update , build_method_cache_update] #@if_true.to_mom( in_method ) #find and assign [Mom::IfStatement.new( build_condition , if_true )] end # this may look like a simple_call, but the difference is that we don't know # the method until run-time. Alas the setup is the same def call_cached_method(in_method) message_setup(in_method) << Mom::DynamicCall.new( method_var_name) end private # cached type and method are stored in the frame as local variables. # this creates the varables in the frame. Names are method_var_name and type_var_name def create_tmps(in_method) in_method.create_tmp end # we store the (one!) cached mathod in the frame, under the name that this # method returns def method_var_name "cached_method_#{object_id}" end def type_var_name "cached_type_#{object_id}" end private def build_condition cached_type = Mom::SlotDefinition.new(:message , [:frame , type_var_name]) current_type = Mom::SlotDefinition.new(:message , [:self , :type]) Mom::NotSameCheck.new(cached_type , current_type) end def build_type_cache_update end def build_method_cache_update end end end