module Risc # A Builder is used to generate code, either by using it's api, or dsl # # The code is added to the method_compiler. # # Basically this allows to many Risc instructions with extremely readable code. # example: # space << Parfait.object_space # load constant # message[:receiver] << space #make current message's (r0) receiver the space # See http://ruby-x.org/rubyx/builder.html for details # class Builder attr_reader :built , :compiler , :names # pass a compiler, to which instruction are added (usually) # second arg determines weather instructions are added (default true) # call build with a block to build def initialize(compiler, for_source) raise "no compiler" unless compiler raise "no source" unless for_source @compiler = compiler @source = for_source @source_used = false @names = {} end # make the magic: convert incoming names into registers that have the # type set according to the name (using resolve_type) # names are stored, so subsequent calls use the same register def method_missing(name , *args) super if args.length != 0 name = name.to_s return @names[name] if @names.has_key?(name) if name == "message" return Risc.message_reg.set_builder(self) end if name.index("label") reg = Risc.label( @source , "#{name}_#{object_id}") @source_used = true else last_char = name[-1] name = name[0 ... -1] if last_char == "!" or last_char == "?" if @names.has_key?(name) return @names[name] if last_char == "?" raise "Name exists (#{@names.keys})before creating it #{name}#{last_char}" end else raise "Must create (with ! or ?) before using #{name}#{last_char}" end type = infer_type(name ) reg = @compiler.use_reg( type.object_class.name ).set_builder(self) end @names[name] = reg reg end # Infer the type from a symbol. In the simplest case the symbol is the class name. # But in building, sometimes variations are needed, so next_message or caller work # too (and both return "Message") # A general "_reg"/"_obj"/"_const" or "_tmp" at the end of the name will be removed # An error is raised if the symbol/object can not be inferred def infer_type( name ) as_string = name.to_s parts = as_string.split("_") if( ["reg" , "obj" , "tmp" , "self" , "const", "1" , "2"].include?( parts.last ) ) parts.pop as_string = parts.join("_") end as_string = "word" if as_string == "name" as_string = "message" if as_string == "next_message" as_string = "message" if as_string == "caller" as_string = "named_list" if as_string == "arguments" sym = as_string.camelise.to_sym clazz = Parfait.object_space.get_class_by_name(sym) raise "Not implemented/found object #{name}:#{sym}" unless clazz return clazz.instance_type end def if_zero( label ) @source_used = true add_code Risc::IsZero.new(@source , label) end def if_not_zero( label ) @source_used = true add_code Risc::IsNotZero.new(@source , label) end def if_minus( label ) @source_used = true add_code Risc::IsMinus.new(@source , label) end def branch( label ) @source_used = true add_code Risc::Branch.new(@source, label) end # To avoid many an if, it can be handy to swap variable names. # But since the names in the builder are not variables, we need this method. # As it says, swap the two names around. Names must exist def swap_names(left , right) left , right = left.to_s , right.to_s l = @names[left] r = @names[right] raise "No such name #{left}" unless l raise "No such name #{right}" unless r @names[left] = r @names[right] = l end # Reset the names stored by the builder. The names are sort of variables names # that can be used in the build block due to method_missing magic. # # But just as the compiler has reset_regs, the builder has this reset button, to # start fresh. Quite crude for now, and only used in allocate_int # # Compiler regs are reset as well def reset_names @names = {} compiler.reset_regs end # Build code using dsl (see __init__ or MessageSetup for examples). # Names (that ruby would resolve to a variable/method) are converted # to registers. << means assignment and [] is supported both on # L and R values (but only one at a time). R values may also be constants. # # Basically this allows to create LoadConstant, RegToSlot, SlotToReg and # Transfer instructions with extremely readable code. # example: # space << Parfait.object_space # load constant # message[:receiver] << space #make current message's (r0) receiver the space # # build result is added to compiler directly # def build(&block) instance_eval(&block) end # add code straight to the compiler def add_code(ins) @compiler.add_code(ins) return ins end # for some methods that return an integer it is beneficial to pre allocate the # integer and store it in the return value. That is what this function does. # # Those (builtin) methods, mostly syscall wrappers then go on to do this and that # clobbering registers and so the allocate and even move would be difficult. # We sidestep all that by pre-allocating. def prepare_int_return integer_tmp = allocate_int reset_names build do message[:return_value] << integer_tmp end end # allocate int fetches a new int, for sure. It is a builder method, rather than # an inbuilt one, to avoid call overhead for 99.9% # The factories allocate in 1k, so only when that runs out do we really need a call. # Note: # Unfortunately (or so me thinks), this creates code bloat, as the calling is # included in 100%, but only needed in 0.1. Risc-level Blocks or Macros may be needed. # as the calling in (the same) 40-50 instructions for every basic int op. # # The method # - grabs a Integer instance from the Integer factory # - checks for nil and calls (get_more) for more if needed # - returns the RiscValue (Register) where the object is found # # The implicit condition is that the method is called at the entry of a method. # It uses a fair few registers and resets all at the end. The returned object # will always be in r1, because the method resets, and all others will be clobbered. # # Return RegisterValue(:r1) that will be named integer_tmp def allocate_int compiler.reset_regs integer = self.integer! build do factory! << Parfait.object_space.get_factory_for(:Integer) integer << factory[:next_object] object! << Parfait.object_space.nil_object object - integer if_not_zero cont_label integer_2! << factory[:reserve] factory[:next_object] << integer_2 call_get_more integer << factory[:next_object] add_code cont_label integer_2 << integer[:next_integer] factory[:next_object] << integer_2 end reset_names integer_tmp! end # Call_get_more calls the method get_more on the factory (see there). # From the callers perspective the method ensures there is a next_object. # # Calling is three step process # - setting up the next message # - moving receiver (factory) and arguments (none) # - issuing the call # These steps shadow the MomInstructions MessageSetup, ArgumentTransfer and SimpleCall def call_get_more factory = Parfait.object_space.get_factory_for( :Integer ) calling = factory.get_type.get_method( :get_more ) calling = Parfait.object_space.get_method!(:Space,:main) #until we actually parse Factory raise "no main defined" unless calling Mom::MessageSetup.new( calling ).build_with( self ) self.build do factory_reg! << factory message[:receiver] << factory_reg end Mom::SimpleCall.new(calling).to_risc(compiler) end end end