rubyx/lib/virtual/compiled_method_info.rb
2015-06-01 08:33:51 +03:00

160 lines
6.6 KiB
Ruby

require_relative "block"
module Virtual
# the static info of a method (with its compiled code, argument names etc ) is part of the
# runtime, ie found in Parfait::Method
# the info we create here is injected int the method and used only at compile-time
# receiver
# return arg (usually mystery, but for coded ones can be more specific)
#
# Methods are one step up from to VM::Blocks. Where Blocks can be jumped to, Methods can be called.
# Methods also have arguments and a return. These are typed by subclass instances of Value
# They also have local variables.
# Code-wise Methods are made up from a list of Blocks, in a similar way blocks are made up of
# Instructions. The function starts with one block, and that has a start and end (return)
# Blocks can be linked in two ways:
# -linear: flow continues from one to the next as they are sequential both logically and
# "physically" use the block set_next for this.
# This "straight line", there must be a continuous sequence from body to return
# Linear blocks may be created from an existing block with new_block
# - branched: You create new blocks using function.new_block which gets added "after" return
# These (eg if/while) blocks may themselves have linear blocks ,but the last of these
# MUST have an uncoditional branch. And remember, all roads lead to return.
class CompiledMethodInfo
# create method does two things
# first it creates the parfait method, for the given class, with given argument names
# second, it creates CompiledMethodInfo and attaches it to the method
#
# compile code then works with the method, but adds code tot the info
def self.create_method( class_name , method_name , args)
raise "uups #{class_name}.#{class_name.class}" unless class_name.is_a? Symbol
raise "uups #{method_name}.#{method_name.class}" unless class_name.is_a? Symbol
clazz = Machine.instance.space.get_class_by_name class_name
raise "No such class #{class_name}" unless clazz
method = clazz.create_instance_method( method_name , Virtual.new_list(args))
method.info = CompiledMethodInfo.new
method
end
def initialize return_type = Virtual::Mystery
# first block we have to create with .new , as new_block assumes a current
enter = Block.new( "enter" , self ).add_code(MethodEnter.new())
@return_type = return_type
@blocks = [enter]
@current = enter
new_block("return").add_code(MethodReturn.new)
@constants = []
end
attr_reader :blocks , :constants
attr_accessor :return_type , :current , :receiver
# add an instruction after the current (insertion point)
# the added instruction will become the new insertion point
def add_code instruction
unless (instruction.is_a?(Instruction) or instruction.is_a?(Register::Instruction))
raise instruction.inspect
end
@current.add_code(instruction) #insert after current
self
end
# return a list of registers that are still in use after the given block
# a call_site uses pushes and pops these to make them available for code after a call
def locals_at l_block
used =[]
# call assigns the return register, but as it is in l_block, it is not asked.
assigned = [ RegisterReference.new(Virtual::RegisterMachine.instance.return_register) ]
l_block.reachable.each do |b|
b.uses.each {|u|
(used << u) unless assigned.include?(u)
}
assigned += b.assigns
end
used.uniq
end
# control structures need to see blocks as a graph, but they are stored as a list with implict
# branches
# So when creating a new block (with new_block), it is only added to the list, but instructions
# still go to the current one
# With this function one can change the current block, to actually code it.
# This juggling is (unfortunately) neccessary, as all compile functions just keep puring their
# code into the method and don't care what other compiles (like if's) do.
# Example: while, needs 2 extra blocks
# 1 condition code, must be its own blockas we jump back to it
# - the body, can actually be after the condition as we don't need to jump there
# 2 after while block. Condition jumps here
# After block 2, the function is linear again and the calling code does not need to know what
# happened
# But subsequent statements are still using the original block (self) to add code to
# So the while expression creates the extra blocks, adds them and the code and then "moves"
# the insertion point along
def current block
@current = block
self
end
# create a new linear block after the current insertion block.
# Linear means there is no brach needed from that one to the new one.
# Usually the new one just serves as jump address for a control statement
# In code generation , the new_block is written after this one, ie zero runtime cost
# This does _not_ change the insertion point, that has do be done with insert_at(block)
def new_block new_name
new_b = Block.new( new_name , self )
index = @blocks.index( @current )
@blocks.insert( index + 1 , new_b ) # + one because we want the ne after the insert_at
return new_b
end
def get_tmp
name = "__tmp__#{@tmps.length}"
@tmps << name
Ast::NameExpression.new(name)
end
# sugar to create instructions easily.
# any method will be passed on to the RegisterMachine and the result added to the insertion block
# With this trick we can write what looks like assembler,
# Example func.instance_eval
# mov( r1 , r2 )
# add( r1 , r2 , 4)
# end
# mov and add will be called on Machine and generate Instructions that are then added
# to the current block
# also symbols are supported and wrapped as register usages (for bare metal programming)
def method_missing(meth, *arg_names, &block)
add_code ::Arm::ArmMachine.send(meth , *arg_names)
end
def mem_length
l = @blocks.inject(0) { |c , block| c += block.mem_length }
padded(l)
end
def padded len
a = 32 * (1 + (len + 7)/32 )
#puts "#{a} for #{len}"
a
end
# position of the function is the position of the entry block, is where we call
def set_position at
at += 8 #for the 2 header words
@blocks.each do |block|
block.set_position at
at = at + block.mem_length
end
end
end
end