PyNerva 0.0.7__py3-none-any.whl

nervapy/x86_64/function.py

@@ -0,0 +1,2707 @@
+# This file is part of PeachPy package and is licensed under the Simplified BSD license.
+#    See license.rst for the full text of the license.
+
+from __future__ import print_function
+
+import bisect
+import collections
+import operator
+import os
+
+import six
+
+import nervapy
+import nervapy.name
+import nervapy.writer
+import nervapy.x86_64.avx
+import nervapy.x86_64.instructions
+import nervapy.x86_64.meta
+import nervapy.x86_64.options
+import nervapy.x86_64.registers
+
+
+class Function:
+    """Generalized x86-64 assembly function.
+
+    A function consists of C signature and a list of instructions.
+
+    On this level the function is supposed to be compatible with multiple ABIs. In particular, instructions may have
+    virtual registers, and instruction stream may contain pseudo-instructions, such as LOAD.ARGUMENT or RETURN.
+    """
+
+    def __init__(
+        self,
+        name,
+        arguments,
+        result_type=None,
+        package=None,
+        target=None,
+        debug_level=None,
+    ):
+        """
+        :param str name: name of the function without mangling (as in C language).
+        :param tuple arguments: a tuple of :class:`nervapy.Argument` objects.
+        :param Type result_type: the return type of the function. None if the function returns no value (void function).
+        :param str package: the name of the Go package containing this function.
+        :param Microarchitecture target: the target microarchitecture for this function.
+        :param int debug_level: the verbosity level for debug information collected for instructions. 0 means no
+            debug information, 1 and above enables information about the lines of Python code that originated an
+            instruction. Collecting debug information increases processing time by several times.
+        :ivar Label entry: a label that marks the entry point of the function. A user can place the entry point in any
+            place in the function by defining this label with LABEL pseudo-instruction. If this label is not defined
+            by the user, it will be placed automatically before the first instruction of the function.
+
+        :ivar int _indent_level: the level of indentation for this instruction in assembly listings. Indentation level
+            is changed by Loop statements.
+        :ivar list _instructions: the list of :class:`Instruction` objects that comprise the function code.
+
+        :ivar set _label_names: a set of string names of LABEL quasi-instructions in the function. The set is populated
+            as instructions are added and is intended to track duplicate labels.
+
+        :ivar dict _named_constants: a dictionary that maps names of literal constants to Constant objects.
+            As instructions are added the dictionary is used to track constants with same names, but different content.
+        """
+        self.name = name
+        self.arguments = arguments
+        self.result_type = result_type
+        if package is None:
+            self.package = nervapy.x86_64.options.package
+        self.package = package
+        if target is None:
+            target = nervapy.x86_64.options.target
+        if target is None:
+            target = nervapy.x86_64.uarch.default
+        if not isinstance(target, nervapy.x86_64.uarch.Microarchitecture):
+            raise TypeError("%s is not an valid CPU target" % str(target))
+        self.target = target
+        if debug_level is None:
+            self.debug_level = nervapy.x86_64.options.debug_level
+        else:
+            self.debug_level = int(debug_level)
+
+        from nervapy.name import Name
+        from nervapy.x86_64.pseudo import Label
+
+        self.entry = Label((Name("__entry__", None),))
+
+        self._indent_level = 1
+
+        self._instructions = list()
+
+        # This set is only used to ensure that all labels references in branches are defined
+        self._label_names = set()
+        # Map from id of Name objects to their copies.
+        # This ensures that Name objects without name can be compared for equality using id
+        self._names_memo = dict()
+        self._scope = nervapy.name.Namespace(None)
+
+        self._local_variables_count = 0
+        self._virtual_general_purpose_registers_count = 0
+        self._virtual_mmx_registers_count = 0
+        self._virtual_xmm_registers_count = 0
+        self._virtual_mask_registers_count = 0
+
+        from nervapy.x86_64 import m256, m256d, m256i
+
+        avx_types = [m256, m256d, m256i]
+        self.avx_environment = (
+            any([arg.c_type in avx_types for arg in self.arguments])
+            or self.result_type in avx_types
+        )
+        self._avx_prolog = None
+
+        from nervapy.common import RegisterAllocator
+        from nervapy.x86_64.registers import (GeneralPurposeRegister,
+                                              KRegister, MMXRegister,
+                                              XMMRegister)
+
+        self._register_allocators = {
+            GeneralPurposeRegister._kind: RegisterAllocator(),
+            MMXRegister._kind: RegisterAllocator(),
+            XMMRegister._kind: RegisterAllocator(),
+            KRegister._kind: RegisterAllocator(),
+        }
+
+    @property
+    def c_signature(self):
+        """C signature (including parameter names) for the function"""
+
+        signature = "void" if self.result_type is None else str(self.result_type)
+        signature = signature + " " + self.name
+        signature = signature + "(" + ", ".join(map(str, self.arguments)) + ")"
+        return signature
+
+    @property
+    def go_signature(self):
+        """Go signature (including parameter names) for the function.
+
+        None if the function argument or return type is incompatible with Go"""
+
+        def c_to_go_type(c_type):
+            assert isinstance(c_type, nervapy.Type)
+            if c_type.is_pointer:
+                if c_type.base is not None:
+                    return "*" + c_to_go_type(c_type.base)
+                else:
+                    return "uintptr"
+            elif c_type.is_bool:
+                return "boolean"
+            elif c_type.is_size_integer:
+                return "int" if c_type.is_signed_integer else "uint"
+            elif c_type.is_signed_integer:
+                return {1: "int8", 2: "int16", 4: "int32", 8: "int64"}[c_type.size]
+            elif c_type.is_unsigned_integer:
+                return {1: "uint8", 2: "uint16", 4: "uint32", 8: "uint64"}[c_type.size]
+            elif c_type.is_floating_point:
+                return {4: "float32", 8: "float64"}[c_type.size]
+            else:
+                return None
+
+        go_argument_types = list(
+            map(c_to_go_type, map(operator.attrgetter("c_type"), self.arguments))
+        )
+        # Some of the C types doesn't have a Go analog
+        if not (all(map(bool, go_argument_types))):
+            return None
+
+        go_arguments = map(
+            lambda name_gotype: " ".join(name_gotype),
+            zip(map(operator.attrgetter("name"), self.arguments), go_argument_types),
+        )
+        if self.result_type is None:
+            return "func %s(%s)" % (self.name, ", ".join(go_arguments))
+        else:
+            go_result_type = c_to_go_type(self.result_type)
+            if go_result_type is None:
+                return None
+            else:
+                return "func %s(%s) %s" % (
+                    self.name,
+                    ", ".join(go_arguments),
+                    go_result_type,
+                )
+
+    @property
+    def isa_extensions(self):
+        from nervapy.x86_64.isa import Extensions
+
+        extensions = set()
+        for instruction in self._instructions:
+            extensions.update(instruction.isa_extensions)
+        return Extensions(*extensions)
+
+    def __enter__(self):
+        self.attach()
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        self.detach()
+        if exc_type is None:
+            self._add_default_labels()
+            self._check_undefined_labels()
+            self._remove_unused_labels()
+            self._analize()
+            if nervapy.x86_64.options.rtl_dump_file:
+                nervapy.x86_64.options.rtl_dump_file.write(self.format_instructions())
+            self._check_live_registers()
+            self._preallocate_registers()
+            self._bind_registers()
+            self._scope.assign_names()
+            if nervapy.x86_64.options.abi is not None:
+                abi_function = self.finalize(nervapy.x86_64.options.abi)
+
+                for writer in nervapy.writer.active_writers:
+                    writer.add_function(abi_function)
+        else:
+            raise
+
+    def attach(self):
+        """Makes active the function and its associated instruction stream.
+
+        While the instruction stream is active, generated instructions are added to this function.
+
+        While the function is active, generated instructions are checked for compatibility with the function target.
+        """
+        import nervapy.common.function
+        import nervapy.stream
+
+        if nervapy.common.function.active_function is not None:
+            raise ValueError(
+                "Can not attach the function: alternative function %s is active"
+                % nervapy.common.function.active_function.name
+            )
+        if nervapy.stream.active_stream is not None:
+            raise ValueError(
+                "Can not attach the function instruction stream: alternative instruction stream is active"
+            )
+        nervapy.common.function.active_function = self
+        nervapy.stream.active_stream = self
+        return self
+
+    def detach(self):
+        """Make the function and its associated instruction stream no longer active.
+
+        The function and its instruction stream must be active before calling the method.
+        """
+        import nervapy.common.function
+        import nervapy.stream
+
+        if nervapy.common.function.active_function is None:
+            raise ValueError("Can not detach the function: no function is active")
+        if nervapy.common.function.active_function is not self:
+            raise ValueError(
+                "Can not detach the function: a different function is active"
+            )
+        nervapy.common.function.active_function = None
+        nervapy.stream.active_stream = None
+        return self
+
+    @staticmethod
+    def _check_arguments(args):
+        # Check types
+        if not isinstance(args, (list, tuple)):
+            raise TypeError(
+                "Invalid arguments types (%s): a tuple or list of function arguments expected"
+                % str(args)
+            )
+        for i, arg in enumerate(args):
+            if not isinstance(arg, Argument):
+                raise TypeError(
+                    "Invalid argument object for argument #%d (%s): nervapy.Argument expected"
+                    % (i, str(arg))
+                )
+        # mapping from argument name to argument number
+        names = dict()
+
+        # First check argument names for arguments with explicit names
+        for i, arg in enumerate(args):
+            if arg.name:
+                if arg.name in names:
+                    raise ValueError(
+                        "Argument #%d (%s) has the same name as argument #%d (%s)"
+                        % (i, str(arg), names[arg.name], args[names[arg.name]])
+                    )
+                names[arg.name] = i
+
+    def _find_argument(self, argument_target):
+        from nervapy import Argument
+
+        assert isinstance(
+            argument_target, (Argument, str)
+        ), "Either Argument object or argument name expected"
+        if isinstance(argument_target, Argument):
+            if argument_target in self.arguments:
+                return argument_target
+            else:
+                return None
+        else:
+            return next(
+                (
+                    argument
+                    for argument in self.arguments
+                    if argument.name == argument_target
+                ),
+                None,
+            )
+
+    def add_instruction(self, instruction):
+        # If instruction is None, do nothing
+        if instruction is None:
+            return
+
+        from nervapy.x86_64.instructions import Instruction
+
+        if not isinstance(instruction, Instruction):
+            raise TypeError("Instruction object expected")
+
+        from nervapy.x86_64.pseudo import LABEL
+
+        # Check that label with the same name is not added twice
+        if isinstance(instruction, LABEL):
+            self._scope.add_scoped_name(instruction.identifier)
+            self._label_names.add(instruction.identifier)
+
+        constant = instruction.constant
+        if constant is not None:
+            self._scope.add_scoped_name(constant.name)
+
+        # Check that the instruction is supported by the target ISA
+        for extension in instruction.isa_extensions:
+            if self.target is not None and extension not in self.target.extensions:
+                raise ValueError(
+                    "{0} is not supported on the target microarchitecture".format(
+                        extension
+                    )
+                )
+
+        instruction._indent_level = self._indent_level
+        self._instructions.append(instruction)
+
+    def add_instructions(self, instructions):
+        for instruction in instructions:
+            self.add_instruction(instruction)
+
+    def finalize(self, abi):
+        from nervapy.x86_64.abi import ABI
+
+        if not isinstance(abi, ABI):
+            raise TypeError("%s is not an ABI object" % str(abi))
+        return ABIFunction(self, abi)
+
+    def _add_default_labels(self):
+        """Adds default labels if they are not defined"""
+
+        from nervapy.x86_64.pseudo import LABEL
+
+        if self.entry.name not in self._scope.names:
+            self._instructions.insert(0, LABEL(self.entry))
+            self._scope.add_scoped_name(self.entry.name)
+            self._label_names.add(self.entry.name)
+
+    def _check_undefined_labels(self):
+        """Verifies that all labels referenced by branch instructions are defined"""
+
+        from nervapy.x86_64.instructions import BranchInstruction
+
+        referenced_label_names = set()
+        for instruction in self._instructions:
+            if isinstance(instruction, BranchInstruction) and instruction.label_name:
+                referenced_label_names.add(instruction.label_name)
+        if not referenced_label_names.issubset(self._label_names):
+            undefined_label_names = referenced_label_names.difference(self._label_names)
+            raise ValueError(
+                "Undefined labels found: "
+                + ", ".join(
+                    map(lambda name: ".".join(map(str, name)), undefined_label_names)
+                )
+            )
+
+    def _remove_unused_labels(self):
+        """Removes labels that are not referenced by any instruction"""
+
+        from nervapy.x86_64.instructions import BranchInstruction
+        from nervapy.x86_64.pseudo import LABEL
+
+        referenced_label_names = set()
+        for instruction in self._instructions:
+            if isinstance(instruction, BranchInstruction) and instruction.label_name:
+                referenced_label_names.add(instruction.label_name)
+        unreferenced_label_names = self._label_names.difference(referenced_label_names)
+        # Do not remove entry label if it is in the middle of the function
+        if self.entry.name in unreferenced_label_names:
+            if (
+                not isinstance(self._instructions[0], LABEL)
+                or self._instructions[0].identifier != self.entry.name
+            ):
+                unreferenced_label_names.remove(self.entry.name)
+        # Remove LABEL pseudo-instructions with unreferenced label names
+        self._instructions = [
+            instruction
+            for instruction in self._instructions
+            if not isinstance(instruction, LABEL)
+            or instruction.identifier not in unreferenced_label_names
+        ]
+        self._label_names.difference_update(unreferenced_label_names)
+
+    def _analize(self):
+        from nervapy.x86_64.generic import RET
+        from nervapy.x86_64.instructions import BranchInstruction
+        from nervapy.x86_64.pseudo import LABEL, RETURN
+
+        # Query input/output registers for each instruction
+        input_registers = []
+        output_registers = []
+        for instruction in self._instructions:
+            input_registers.append(instruction.input_registers_masks)
+            output_registers.append(instruction.output_registers_masks)
+
+        # Map from label name to its quasi-instruction number in the stream
+        labels = {
+            instruction.identifier: i
+            for (i, instruction) in enumerate(self._instructions)
+            if isinstance(instruction, LABEL)
+        }
+        entry_position = 0
+        if self.entry.name in self._label_names:
+            entry_position = labels[self.entry.name]
+        branch_instructions = [
+            (i, instruction)
+            for (i, instruction) in enumerate(self._instructions)
+            if isinstance(instruction, BranchInstruction) and instruction.label_name
+        ]
+        # Basic blocks start at function entry position or on branch target
+        basic_block_starts = {entry_position}
+        for i, branch_instruction in branch_instructions:
+            basic_block_starts.add(labels[branch_instruction.label_name])
+            if branch_instruction.is_conditional:
+                basic_block_starts.add(i + 1)
+        basic_block_starts = sorted(basic_block_starts)
+        # Basic block ends on a referenced label instruction or right after return/branch instructions
+        basic_block_ends = [
+            i + int(not isinstance(instruction, LABEL))
+            for (i, instruction) in enumerate(self._instructions)
+            if isinstance(instruction, (BranchInstruction, RETURN, RET, LABEL))
+        ]
+        # TODO: check that the last block with an unconditional branch/return instruction
+        basic_block_bounds = [
+            (start, basic_block_ends[bisect.bisect_right(basic_block_ends, start)])
+            for start in basic_block_starts
+        ]
+
+        class BasicBlock:
+            def __init__(
+                self,
+                start_position,
+                end_position,
+                input_registers_list,
+                output_registers_list,
+            ):
+                self.start_position = start_position
+                self.end_position = end_position
+                self.input_registers_list = input_registers_list
+                self.output_registers_list = output_registers_list
+
+                self.consumed_register_masks = collections.defaultdict(int)
+                self.produced_register_masks = collections.defaultdict(int)
+
+                self.live_register_masks = collections.defaultdict(int)
+                self.available_register_masks = collections.defaultdict(int)
+
+                self.is_reachable = False
+
+                self.liveness_analysis_passes = 0
+                self.availability_analysis_passes = 0
+
+                self.input_blocks = list()
+                self.output_blocks = list()
+
+                self.processed_input_blocks = set()
+                self.processed_output_blocks = set()
+
+                # Mark available and consumed registers:
+                # - If a register is consumed by an instruction but not produced by preceding instructions of the basic
+                #   block, the register is consumed by the basic block
+                # - If a register is produced by an instruction, it becomes available for the subsequent instructions
+                #   of the basic block and counts as produced by the basic block
+                for input_registers, output_registers in zip(
+                    input_registers_list, output_registers_list
+                ):
+                    for input_register_id, input_register_mask in six.iteritems(
+                        input_registers
+                    ):
+                        consumed_mask = (
+                            input_register_mask
+                            & ~self.produced_register_masks[input_register_id]
+                        )
+                        if consumed_mask != 0:
+                            self.consumed_register_masks[
+                                input_register_id
+                            ] |= consumed_mask
+                    for output_register_id, output_register_mask in six.iteritems(
+                        output_registers
+                    ):
+                        self.produced_register_masks[
+                            output_register_id
+                        ] |= output_register_mask
+
+            def reset_processed_blocks(self):
+                self.processed_input_blocks = set()
+                self.processed_output_blocks = set()
+
+            @property
+            def available_registers_list(self):
+                from nervapy.x86_64.registers import Register
+
+                available_registers_list = []
+                available_registers_masks = self.available_register_masks.copy()
+                for output_registers in self.output_registers_list:
+                    # Record available registers for current instruction
+                    available_registers_list.append(available_registers_masks.copy())
+                    # Update with output registers for current instruction
+                    for output_register_id, output_register_mask in six.iteritems(
+                        output_registers
+                    ):
+                        available_registers_masks[output_register_id] = (
+                            available_registers_masks.get(output_register_id, 0)
+                            | output_register_mask
+                        )
+                return available_registers_list
+
+            @property
+            def live_registers_list(self):
+                from nervapy.x86_64.registers import Register
+
+                live_registers_list = []
+                live_registers_masks = self.live_register_masks.copy()
+                for input_registers, output_registers in reversed(
+                    list(zip(self.input_registers_list, self.output_registers_list))
+                ):
+                    # Mark register written by the instruction as non-live
+                    for output_register_id, output_register_mask in six.iteritems(
+                        output_registers
+                    ):
+                        if output_register_id in live_registers_masks:
+                            new_live_register_mask = (
+                                live_registers_masks[output_register_id]
+                                & ~output_register_mask
+                            )
+                            if new_live_register_mask != 0:
+                                live_registers_masks[output_register_id] = (
+                                    new_live_register_mask
+                                )
+                            else:
+                                del live_registers_masks[output_register_id]
+                    # Mark registers read by the instruction as live
+                    for input_register_id, input_register_mask in six.iteritems(
+                        input_registers
+                    ):
+                        live_registers_masks[input_register_id] = (
+                            live_registers_masks.get(input_register_id, 0)
+                            | input_register_mask
+                        )
+                    # Record available registers for current instruction
+                    live_registers_list.append(live_registers_masks.copy())
+                live_registers_list.reverse()
+                return live_registers_list
+
+            def __str__(self):
+                return "[%d, %d)" % (self.start_position, self.end_position)
+
+            def __repr__(self):
+                return str(self)
+
+            def analyze_availability(self, extra_available_registers):
+                self.availability_analysis_passes += 1
+
+                if self.availability_analysis_passes == 1:
+                    # First pass: add registers produced by the block and propagate further
+                    self.available_register_masks.update(extra_available_registers)
+                    if self.output_blocks:
+                        # Add registers produced by this block
+                        for produced_reg_id, produced_reg_mask in six.iteritems(
+                            self.produced_register_masks
+                        ):
+                            extra_available_registers[produced_reg_id] = (
+                                extra_available_registers.get(produced_reg_id, 0)
+                                | produced_reg_mask
+                            )
+                else:
+                    # Subsequent passes: compute and propagate only the input registers that were not processed before
+                    for reg_id, extra_reg_mask in list(
+                        six.iteritems(extra_available_registers)
+                    ):
+                        old_reg_mask = self.available_register_masks[reg_id]
+                        update_reg_mask = extra_reg_mask & ~old_reg_mask
+                        if update_reg_mask != 0:
+                            self.available_register_masks[reg_id] |= extra_reg_mask
+                        if self.output_blocks:
+                            if update_reg_mask != 0:
+                                extra_available_registers[reg_id] = update_reg_mask
+                            else:
+                                del extra_available_registers[reg_id]
+
+                if self.output_blocks and (
+                    extra_available_registers or self.availability_analysis_passes == 1
+                ):
+                    for output_block in self.output_blocks[1:]:
+                        # The dict needs to be copied because output blocks can change it
+                        output_block.analyze_availability(
+                            extra_available_registers.copy()
+                        )
+                    # Optimization: do not create a copy of the dict
+                    self.output_blocks[0].analyze_availability(
+                        extra_available_registers
+                    )
+
+            def analyze_liveness(self, extra_live_registers):
+                # Update in liveness analysis consists of three steps:
+                # 1. Update live registers for this basic block.
+                # 2. Mark registers which are produced to by the basic block as non-live.
+                # 3. Mark registers which are consumed by the basic block as live (only on first pass).
+
+                self.liveness_analysis_passes += 1
+
+                # Steps 1 and 2
+                for reg_id, extra_reg_mask in list(six.iteritems(extra_live_registers)):
+                    old_reg_mask = self.live_register_masks[reg_id]
+                    update_reg_mask = extra_reg_mask & ~old_reg_mask
+                    if update_reg_mask != 0:
+                        self.live_register_masks[reg_id] |= extra_reg_mask
+                    if self.input_blocks:
+                        # On the first pass do not modify the extra live registers masks that are passed to input blocks
+                        # On subsequent passes only the novel live registers need to be passed further
+                        if self.liveness_analysis_passes == 1:
+                            update_reg_mask = extra_reg_mask
+                        update_reg_mask &= ~self.produced_register_masks.get(reg_id, 0)
+                        if update_reg_mask != 0:
+                            extra_live_registers[reg_id] = update_reg_mask
+                        else:
+                            del extra_live_registers[reg_id]
+
+                # Step 3
+                if self.input_blocks:
+                    if self.liveness_analysis_passes == 1:
+                        for consumed_reg_id, consumed_reg_mask in six.iteritems(
+                            self.consumed_register_masks
+                        ):
+                            extra_live_registers[consumed_reg_id] = (
+                                extra_live_registers.get(consumed_reg_id, 0)
+                                | consumed_reg_mask
+                            )
+
+                if self.input_blocks and (
+                    extra_live_registers or self.liveness_analysis_passes == 1
+                ):
+                    for input_block in self.input_blocks[1:]:
+                        # The dict needs to be copied because input blocks can change it
+                        input_block.analyze_liveness(extra_live_registers.copy())
+                    # Optimization: do not create a copy of the dict
+                    self.input_blocks[0].analyze_liveness(extra_live_registers)
+
+            def analyze_reachability(self):
+                if not self.is_reachable:
+                    self.is_reachable = True
+                    for output_block in self.output_blocks:
+                        output_block.analyze_reachability()
+
+            def forward_pass(self, processing_function, instructions, input_state):
+                output_state = processing_function(self, instructions, input_state)
+                for output_block in self.output_blocks:
+                    if output_block.start_position not in self.processed_output_blocks:
+                        self.processed_output_blocks.add(output_block.start_position)
+                        output_block.forward_pass(
+                            processing_function, instructions, output_state
+                        )
+
+            def backward_pass(self, processing_function, instructions, input_state):
+                output_state = processing_function(self, instructions, input_state)
+                for input_block in self.input_blocks:
+                    if input_block.start_position not in self.processed_input_blocks:
+                        self.processed_input_blocks.add(input_block.start_position)
+                        input_block.backward_pass(
+                            processing_function, instructions, output_state
+                        )
+
+            def propogate_sse_avx_state_forward(self, instructions, is_avx_environment):
+                from nervapy.x86_64.avx import VZEROALL, VZEROUPPER
+                from nervapy.x86_64.pseudo import LOAD, STORE
+
+                avx_state = True if is_avx_environment else None
+
+                def propogate_forward(block, instructions, avx_state):
+                    for instruction in instructions[
+                        block.start_position : block.end_position
+                    ]:
+                        if isinstance(instruction, (VZEROUPPER, VZEROALL)):
+                            avx_state = None
+                        elif instruction.avx_mode is None:
+                            # Instruction without a mode
+                            if isinstance(
+                                instruction,
+                                (LOAD.ARGUMENT, STORE.RESULT, RETURN, RET, LABEL),
+                            ):
+                                # Some pseudo-instructions need AVX/SSE mode for lowering
+                                instruction.avx_mode = avx_state
+                        elif instruction.avx_mode:
+                            # AVX-mode instruction
+                            avx_state = True
+                        else:
+                            # SSE-mode instruction
+                            if avx_state:
+                                raise TypeError(
+                                    "AVX-mode instruction {0} follows an SSE-mode instruction".format(
+                                        instruction
+                                    )
+                                )
+                            avx_state = False
+                    return avx_state
+
+                self.forward_pass(propogate_forward, instructions, avx_state)
+
+            def propogate_sse_state_backward(self, instructions, is_avx_environment):
+                from nervapy.x86_64.pseudo import LOAD, STORE
+
+                avx_state = True if is_avx_environment else None
+
+                def propogate_sse_backward(block, instructions, avx_state):
+                    for instruction in reversed(
+                        instructions[block.start_position : block.end_position]
+                    ):
+                        if instruction.avx_mode is not None:
+                            avx_state = instruction.avx_mode
+                        elif avx_state is not None and not avx_state:
+                            if isinstance(
+                                instruction,
+                                (LOAD.ARGUMENT, STORE.RESULT, RETURN, RET, LABEL),
+                            ):
+                                instruction.avx_mode = avx_state
+                    return avx_state
+
+                self.backward_pass(propogate_sse_backward, instructions, avx_state)
+
+            def propogate_avx_state_backward(self, instructions, is_avx_environment):
+                from nervapy.x86_64.pseudo import LOAD, STORE
+
+                avx_state = True if is_avx_environment else None
+
+                def propogate_avx_backward(block, instructions, avx_state):
+                    for instruction in reversed(
+                        instructions[block.start_position : block.end_position]
+                    ):
+                        if instruction.avx_mode is not None:
+                            avx_state = instruction.avx_mode
+                        elif avx_state:
+                            if isinstance(
+                                instruction,
+                                (LOAD.ARGUMENT, STORE.RESULT, RETURN, RET, LABEL),
+                            ):
+                                instruction.avx_mode = avx_state
+
+                self.backward_pass(propogate_avx_backward, instructions, avx_state)
+
+        basic_blocks = list(
+            map(
+                lambda start_end: BasicBlock(
+                    start_end[0],
+                    start_end[1],
+                    input_registers[start_end[0] : start_end[1]],
+                    output_registers[start_end[0] : start_end[1]],
+                ),
+                basic_block_bounds,
+            )
+        )
+        # Map from block start position to BasicBlock object
+        basic_blocks_map = {
+            basic_block_start: basic_block
+            for (basic_block_start, basic_block) in zip(
+                basic_block_starts, basic_blocks
+            )
+        }
+        # Set output basic blocks for each basic block object
+        for i, basic_block in enumerate(basic_blocks):
+            # Consider last instruction of the basic block
+            last_instruction = self._instructions[basic_block.end_position - 1]
+            if isinstance(last_instruction, (RET, RETURN)):
+                # Basic block that ends with a return instruction has no output blocks
+                pass
+            elif (
+                isinstance(last_instruction, BranchInstruction)
+                and last_instruction.label_name
+            ):
+                # Basic block that ends with a branch instruction can jump to the block at branch label
+                target_position = labels[last_instruction.label_name]
+                basic_block.output_blocks = [basic_blocks_map[target_position]]
+                if last_instruction.is_conditional:
+                    # Basic blocks that end with a conditional branch instruction can fall through to the next block
+                    basic_block.output_blocks.append(basic_blocks[i + 1])
+            else:
+                # Basic block ends before a label and continues to the next basic block
+                basic_block.output_blocks = [basic_blocks[i + 1]]
+        # Set input basic blocks for each basic block object
+        for basic_block in basic_blocks:
+            basic_block.input_blocks = list(
+                filter(lambda bb: basic_block in bb.output_blocks, basic_blocks)
+            )
+
+        # Analyze which blocks can be reached from the entry point
+        basic_blocks_map[entry_position].analyze_reachability()
+        exit_positions = [
+            block.start_position for block in basic_blocks if not block.output_blocks
+        ]
+
+        # Analyze register lifetime
+        basic_blocks_map[entry_position].analyze_availability(dict())
+        for exit_position in exit_positions:
+            basic_blocks_map[exit_position].analyze_liveness(dict())
+
+        # Analyze SSE/AVX mode
+        basic_blocks_map[entry_position].propogate_sse_avx_state_forward(
+            self._instructions, self.avx_environment
+        )
+        for exit_position in exit_positions:
+            basic_blocks_map[exit_position].propogate_sse_state_backward(
+                self._instructions, self.avx_environment
+            )
+        for basic_block in basic_blocks:
+            basic_block.reset_processed_blocks()
+        for exit_position in exit_positions:
+            basic_blocks_map[exit_position].propogate_avx_state_backward(
+                self._instructions, self.avx_environment
+            )
+        self._avx_prolog = self._instructions[entry_position].avx_mode
+
+        # Reconstruct live and available registers for the whole instruction sequence
+        for basic_block in basic_blocks:
+            for instruction, available_registers, live_registers in zip(
+                self._instructions[
+                    basic_block.start_position : basic_block.end_position
+                ],
+                basic_block.available_registers_list,
+                basic_block.live_registers_list,
+            ):
+                instruction._live_registers = live_registers
+                instruction._available_registers = available_registers
+            # Remove referenced to input/output blocks to avoid memory leaks due to cycles in ref graph
+            basic_block.input_blocks = None
+            basic_block.output_blocks = None
+
+        # Analyze conflicting registers
+        output_registers = set()
+        for instruction in self._instructions:
+            instruction_registers = instruction.input_registers
+            instruction_registers.update(output_registers)
+            for instruction_register in instruction_registers:
+                if instruction_register.is_virtual:
+                    conflict_internal_ids = [
+                        reg_id
+                        for (reg_id, reg_mask) in six.iteritems(
+                            instruction._live_registers
+                        )
+                        if reg_mask & instruction_register.mask != 0
+                    ]
+                    self._register_allocators[instruction_register.kind].add_conflicts(
+                        instruction_register.virtual_id, conflict_internal_ids
+                    )
+            physical_registers = [r for r in instruction_registers if not r.is_virtual]
+            if physical_registers:
+                from nervapy.x86_64.registers import Register
+
+                live_virtual_registers = Register._reconstruct_multiple(
+                    {
+                        reg_id: reg_mask
+                        for (reg_id, reg_mask) in six.iteritems(
+                            instruction._live_registers
+                        )
+                        if reg_id < 0
+                    }
+                )
+                for live_virtual_register in live_virtual_registers:
+                    conflict_internal_ids = [
+                        reg._internal_id
+                        for reg in physical_registers
+                        if reg.mask & live_virtual_register.mask != 0
+                    ]
+                    self._register_allocators[live_virtual_register.kind].add_conflicts(
+                        live_virtual_register.virtual_id, conflict_internal_ids
+                    )
+            output_registers = instruction.output_registers
+
+    def _check_live_registers(self):
+        """Checks that the number of live registers does not exceed the number of physical registers for each insruction"""
+        from nervapy.x86_64.registers import (GeneralPurposeRegister,
+                                              KRegister, MMXRegister,
+                                              XMMRegister)
+
+        max_live_registers = {
+            GeneralPurposeRegister._kind: 15,
+            MMXRegister._kind: 8,
+            XMMRegister._kind: 16,
+            KRegister._kind: 8,
+        }
+        for instruction in self._instructions:
+            live_registers = max_live_registers.copy()
+            for reg in instruction.live_registers:
+                live_registers[reg.kind] -= 1
+            if any(
+                surplus_count < 0 for surplus_count in six.itervalues(live_registers)
+            ):
+                if (
+                    instruction.source_file is not None
+                    and instruction.line_number is not None
+                ):
+                    raise nervapy.RegisterAllocationError(
+                        "The number of live virtual registers exceeds physical constraints %s at %s:%d"
+                        % (
+                            str(instruction),
+                            instruction.source_file,
+                            instruction.line_number,
+                        )
+                    )
+                else:
+                    raise nervapy.RegisterAllocationError(
+                        "The number of live virtual registers exceeds physical constraints %s"
+                        % str(instruction)
+                    )
+
+    def _preallocate_registers(self):
+        """Allocates registers that can be binded only to a single virtual register.
+
+        Several instructions accept only a fixed a register as their operand. If a virtual register is supplied as such
+        operand, it must be binded to the fixed register accepted by instruction encoding.
+
+        These instructions are:
+
+            - BLENDVPS xmm, xmm/m128, xmm0
+            - BLENDVPD xmm, xmm/m128, xmm0
+            - PBLENDVB xmm, xmm/m128, xmm0
+            - SHA256RNDS2 xmm, xmm/m128, xmm0
+            - SHR r/m, cl
+            - SAR r/m, cl
+            - SAL r/m, cl
+            - SHL r/m, cl
+            - ROR r/m, cl
+            - ROL r/m, cl
+            - SHRD r/m, r, cl
+            - SHLD r/m, r, cl
+        """
+
+        from nervapy import RegisterAllocationError
+        from nervapy.x86_64.registers import (GeneralPurposeRegister,
+                                              GeneralPurposeRegister8,
+                                              XMMRegister, cl, xmm0)
+
+        cl_binded_registers = set()
+        xmm0_binded_registers = set()
+        for instruction in self._instructions:
+            if instruction.name in {"BLENDVPD", "BLENDVPS", "PBLENDVB", "SHA256RNDS2"}:
+                assert (
+                    len(instruction.operands) == 3
+                ), "expected 3 operands, got %d (%s)" % (
+                    len(instruction.operands),
+                    ", ".join(map(str, instruction.operands)),
+                )
+                xmm0_operand = instruction.operands[2]
+                assert isinstance(
+                    xmm0_operand, XMMRegister
+                ), "expected xmm registers in the 3rd operand, got %s" % str(
+                    xmm0_operand
+                )
+                if xmm0_operand.is_virtual:
+                    # Check that xmm0 is not live at this instruction
+                    if (
+                        instruction._live_registers.get(xmm0._internal_id, 0)
+                        & XMMRegister._mask
+                        != 0
+                    ):
+                        raise RegisterAllocationError(
+                            (
+                                "Instruction %s requires operand 3 to be allocated to xmm0 register, "
+                                + "but xmm0 is a live register"
+                            )
+                            % str(instruction.name)
+                        )
+                    xmm0_binded_registers.add(xmm0_operand._internal_id)
+            elif instruction.name in {
+                "SAL",
+                "SAR",
+                "SHL",
+                "SHR",
+                "ROL",
+                "ROR",
+                "RCL",
+                "RCR",
+            }:
+                assert (
+                    len(instruction.operands) == 2
+                ), "expected 2 operands, got %d (%s)" % (
+                    len(instruction.operands),
+                    ", ".join(map(str, instruction.operands)),
+                )
+                count_operand = instruction.operands[1]
+                # The count operand can be cl or imm8
+                if (
+                    isinstance(count_operand, GeneralPurposeRegister8)
+                    and count_operand.is_virtual
+                ):
+                    # Check that cl is not live at this instruction
+                    if (
+                        instruction._live_registers.get(cl._internal_id, 0)
+                        & GeneralPurposeRegister8._mask
+                        != 0
+                    ):
+                        raise RegisterAllocationError(
+                            "Instruction %s requires operand 2 to be allocated to cl register, "
+                            + "but cl is a live register" % instruction.name
+                        )
+
+                    cl_binded_registers.add(count_operand._internal_id)
+            elif instruction.name in {"SHLD", "SHRD"}:
+                assert (
+                    len(instruction.operands) == 3
+                ), "expected 3 operands, got %d (%s)" % (
+                    len(instruction.operands),
+                    ", ".join(map(str, instruction.operands)),
+                )
+                count_operand = instruction.operands[2]
+                # The count operand can be cl or imm8
+                if (
+                    isinstance(count_operand, GeneralPurposeRegister8)
+                    and count_operand.is_virtual
+                ):
+                    # Check that cl is not live at this instruction
+                    if (
+                        instruction._live_registers.get(cl._internal_id, 0)
+                        & GeneralPurposeRegister8._mask
+                        != 0
+                    ):
+                        raise RegisterAllocationError(
+                            "Instruction %s requires operand 3 to be allocated to cl register, "
+                            + "but cl is a live register" % instruction.name
+                        )
+
+                    cl_binded_registers.add(count_operand._internal_id)
+
+        # Check that cl-binded registers are not mutually conflicting
+        for cl_register in cl_binded_registers:
+            other_cl_registers = filter(
+                operator.methodcaller("__ne__", cl_register), cl_binded_registers
+            )
+            conflicting_registers = self._conflicting_registers[1][cl_register]
+            if any(
+                [
+                    other_register in conflicting_registers
+                    for other_register in other_cl_registers
+                ]
+            ):
+                raise RegisterAllocationError(
+                    "Two conflicting virtual registers are required to bind to cl"
+                )
+
+        # Check that xmm0-binded registers are not mutually conflicting
+        for xmm0_register in xmm0_binded_registers:
+            other_xmm0_registers = filter(
+                operator.methodcaller("__ne__", xmm0_register), xmm0_binded_registers
+            )
+            conflicting_registers = self._conflicting_registers[3][xmm0_register]
+            if any(
+                [
+                    other_register in conflicting_registers
+                    for other_register in other_xmm0_registers
+                ]
+            ):
+                raise RegisterAllocationError(
+                    "Two conflicting virtual registers are required to bind to xmm0"
+                )
+
+        # Commit register allocations
+        for cl_register in cl_binded_registers:
+            self._register_allocations[GeneralPurposeRegister._kind][
+                cl_register
+            ] = cl.physical_id
+        for xmm0_register in xmm0_binded_registers:
+            self._register_allocations[XMMRegister._kind][
+                xmm0_register
+            ] = xmm0.physical_id
+
+    def _bind_registers(self):
+        """Iterates through the list of instructions and assigns physical IDs to allocated registers"""
+
+        for instruction in self._instructions:
+            for register in instruction.register_objects:
+                if register.is_virtual:
+                    register.physical_id = self._register_allocators[
+                        register.kind
+                    ].register_allocations.get(
+                        register._internal_id, register.physical_id
+                    )
+
+    def _allocate_local_variable(self):
+        """Returns a new unique ID for a local variable"""
+        self._local_variables_count += 1
+        return self._local_variables_count
+
+    def _allocate_mask_register_id(self):
+        """Returns a new unique ID for a virtual mask (k) register"""
+        self._virtual_mask_registers_count += 1
+        return self._virtual_mask_registers_count
+
+    def _allocate_xmm_register_id(self):
+        """Returns a new unique ID for a virtual SSE/AVX/AVX-512 (xmm/ymm/zmm) register"""
+        self._virtual_xmm_registers_count += 1
+        return self._virtual_xmm_registers_count
+
+    def _allocate_mmx_register_id(self):
+        """Returns a new unique ID for a virtual MMX (mm) register"""
+        self._virtual_mmx_registers_count += 1
+        return self._virtual_mmx_registers_count
+
+    def _allocate_general_purpose_register_id(self):
+        """Returns a new unique ID for a virtual general-purpose register"""
+        self._virtual_general_purpose_registers_count += 1
+        return self._virtual_general_purpose_registers_count
+
+    def __str__(self):
+        """Returns string representation of the function signature and instructions"""
+
+        return self.format()
+
+    def format_instructions(self, line_separator=os.linesep):
+        """Formats instruction listing including data on input, output, available and live registers"""
+
+        from nervapy.x86_64.pseudo import ALIGN, LABEL
+
+        code = []
+        tab = " " * 4
+        for instruction in self._instructions:
+            code.append(instruction.format("peachpy", indent=False))
+            if not isinstance(instruction, (LABEL, ALIGN)):
+                code.append(
+                    tab
+                    + "In regs:    "
+                    + ", ".join(sorted(map(str, instruction.input_registers)))
+                )
+                code.append(
+                    tab
+                    + "Out regs:   "
+                    + ", ".join(sorted(map(str, instruction.output_registers)))
+                )
+                code.append(
+                    tab
+                    + "Live regs:  "
+                    + ", ".join(sorted(map(str, instruction.live_registers)))
+                )
+                code.append(
+                    tab
+                    + "Avail regs: "
+                    + ", ".join(sorted(map(str, instruction.available_registers)))
+                )
+            code.append("")
+        if line_separator is None:
+            return code
+        else:
+            return str(line_separator).join(code)
+
+    def format(self, line_separator=os.linesep):
+        """Formats assembly listing of the function according to specified parameters"""
+
+        code = [self.c_signature]
+        for instruction in self._instructions:
+            code.append(instruction.format("peachpy", indent=True))
+        if line_separator is None:
+            return code
+        else:
+            return str(line_separator).join(code)
+
+
+class Argument(nervapy.Argument):
+    def __init__(self, argument, abi):
+        """Extends generic Argument object with x86-64 specific attributes required for stack frame construction
+
+        :ivar nervapy.x86_64.registers.Register register: the register in which the argument is passed to the function
+            or None if the argument is passed on stack.
+        :ivar int stack_offset: offset from the end of return address on the stack to the location of the argument on
+            stack or None if the argument is passed in a register and has no stack location. Note that in Microsoft X64
+            ABI the first four arguments are passed in registers but have stack space reserved for their storage.
+            For these arguments both register and stack_offset are non-null.
+        :ivar nervapy.x86_64.operand.MemoryAddress address: address of the argument on stack, relative to rsp or rbp.
+            The value of this attribute is None until after register allocation. In Golang ABIs this attribute is never
+            initialized because to load arguments from stack Golang uses its own pseudo-register FP, which is not
+            representable in PeachPy (LOAD.ARGUMENT pseudo-instructions use stack_offset instead when formatted as
+            Golang assembly).
+        """
+        assert isinstance(
+            argument, nervapy.Argument
+        ), "Architecture-specific argument must be constructed from generic Argument object"
+        from nervapy.x86_64.abi import ABI
+
+        assert isinstance(abi, ABI), "ABI object expected"
+        from copy import deepcopy
+
+        super(Argument, self).__init__(deepcopy(argument.c_type), argument.name)
+        if self.c_type.size is None:
+            self.c_type.size = self.c_type.get_size(abi)
+        self.abi = abi
+        self.register = None
+        self.address = None
+        self.stack_offset = None
+        self.save_on_stack = False
+
+    @property
+    def passed_on_stack(self):
+        return self.register is None
+
+
+class ABIFunction:
+    """ABI-specific x86-64 assembly function.
+
+    A function consists of C signature, ABI, and a list of instructions without virtual registers.
+    """
+
+    def __init__(self, function, abi):
+        from copy import deepcopy
+
+        from nervapy.x86_64.abi import (ABI, goasm_amd64_abi,
+                                        goasm_amd64p32_abi, gosyso_amd64_abi,
+                                        gosyso_amd64p32_abi, linux_x32_abi,
+                                        microsoft_x64_abi,
+                                        native_client_x86_64_abi,
+                                        system_v_x86_64_abi)
+
+        assert isinstance(function, Function), "Function object expected"
+        assert isinstance(abi, ABI), "ABI object expected"
+        self.name = function.name
+        self.arguments = [Argument(argument, abi) for argument in function.arguments]
+        self.result_type = function.result_type
+        self.result_offset = None
+        self.package = function.package
+        self.target = function.target
+        self.isa_extensions = function.isa_extensions
+        self.abi = abi
+        self.c_signature = function.c_signature
+        self.go_signature = function.go_signature
+
+        self.avx_environment = function.avx_environment
+        self._avx_prolog = function._avx_prolog
+
+        from nervapy.x86_64.registers import rsp
+
+        self._stack_base = rsp
+        self._stack_frame_size = 0
+        self._stack_frame_alignment = self.abi.stack_alignment
+        self._local_variables_size = 0
+
+        self._instructions = deepcopy(function._instructions)
+        self._register_allocators = deepcopy(function._register_allocators)
+
+        if abi == microsoft_x64_abi:
+            self._setup_windows_arguments()
+        elif abi in {system_v_x86_64_abi, linux_x32_abi, native_client_x86_64_abi}:
+            self._setup_unix_arguments()
+        elif abi in {
+            gosyso_amd64_abi,
+            gosyso_amd64p32_abi,
+            goasm_amd64_abi,
+            goasm_amd64p32_abi,
+        }:
+            self._setup_golang_arguments()
+        else:
+            raise ValueError("Unsupported ABI: %s" % str(abi))
+
+        self._update_argument_loads(function.arguments)
+
+        self._layout_local_variables()
+
+        self._allocate_registers()
+        self._bind_registers()
+
+        self._clobbered_registers = self._analyze_clobbered_registers()
+        self._update_stack_frame()
+        self._update_argument_addresses()
+
+        self._lower_argument_loads()
+        self._lower_pseudoinstructions()
+        self._filter_instruction_encodings()
+
+        self.mangled_name = self.mangle_name()
+
+    def _update_argument_loads(self, arguments):
+        from nervapy.x86_64.pseudo import LOAD
+
+        for instruction in self._instructions:
+            if isinstance(instruction, LOAD.ARGUMENT):
+                instruction.operands = (
+                    instruction.operands[0],
+                    self.arguments[arguments.index(instruction.operands[1])],
+                )
+                if instruction.operands[1].register in instruction.available_registers:
+                    instruction.operands[1].save_on_stack = True
+
+    def _setup_windows_arguments(self):
+        from nervapy.x86_64.abi import microsoft_x64_abi
+
+        assert (
+            self.abi == microsoft_x64_abi
+        ), "This function must only be used with Microsoft x64 ABI"
+        # The first 4 arguments are passed in registers, others are on stack.
+        # 8 bytes on stack is reserved for each parameter (regardless of their size).
+        # On-stack space is also reserved, but not initialized, for parameters passed in registers.
+        # Arguments are NOT extended to 8 bytes, and high bytes of registers/stack cells may contain garbage.
+        from nervapy.x86_64 import m64
+        from nervapy.x86_64.registers import (r8, r9, rcx, rdx, xmm0, xmm1,
+                                              xmm2, xmm3)
+
+        floating_point_argument_registers = (xmm0, xmm1, xmm2, xmm3)
+        integer_argument_registers = (rcx, rdx, r8, r9)
+        for index, argument in enumerate(self.arguments):
+            argument.passed_by_reference = argument.is_vector and argument != m64
+            if index < 4:
+                if argument.is_floating_point:
+                    argument.register = floating_point_argument_registers[index]
+                elif (
+                    argument.is_integer
+                    or argument.is_pointer
+                    or argument.is_codeunit
+                    or argument.is_mask
+                    or argument.c_type == m64
+                ):
+                    argument_register = integer_argument_registers[index]
+                    argument.register = {
+                        1: argument_register.as_low_byte,
+                        2: argument_register.as_word,
+                        4: argument_register.as_dword,
+                        8: argument_register,
+                    }[argument.size]
+                elif argument.is_vector:
+                    argument.register = integer_argument_registers[index]
+                else:
+                    assert False
+            # Stack offset does not include return address
+            argument.stack_offset = index * 8
+
+    def _setup_unix_arguments(self):
+        from nervapy.x86_64.abi import (linux_x32_abi,
+                                        native_client_x86_64_abi,
+                                        system_v_x86_64_abi)
+
+        assert self.abi in {
+            system_v_x86_64_abi,
+            linux_x32_abi,
+            native_client_x86_64_abi,
+        }, "This function must only be used with System V x86-64, Linux x32 or Native Client x86-64 SFI ABI"
+
+        from nervapy.x86_64.registers import (r8, r9, rcx, rdi, rdx, rsi, xmm0,
+                                              xmm1, xmm2, xmm3, xmm4, xmm5,
+                                              xmm6, xmm7)
+
+        # The first 6 integer/pointer arguments are passed in general-purpose registers.
+        # The first 8 floating-point arguments are passed in SSE registers.
+        # For all integer arguments in excess of 6 and floating-point arguments in excess the caller reserves
+        # 8 bytes on stack.
+        # Arguments smaller than 4 bytes are extended (with sign-extension, if needed) to 4 bytes.
+        # For 4-byte and smaller arguments passed on stack the high 4 bytes are not initialized.
+        # For 4-byte and smaller arguments passed in registers the high 4 bytes are zero-initialized.
+        # X32 and Native Client ABIs were not much tested, but they seem similar
+        available_floating_point_registers = [
+            xmm0,
+            xmm1,
+            xmm2,
+            xmm3,
+            xmm4,
+            xmm5,
+            xmm6,
+            xmm7,
+        ]
+        available_integer_registers = [rdi, rsi, rdx, rcx, r8, r9]
+
+        # Stack offset does not include return address
+        stack_offset = 0
+        for argument in self.arguments:
+            if (argument.is_floating_point or argument.is_vector) and len(
+                available_floating_point_registers
+            ) > 0:
+                argument.register = available_floating_point_registers.pop(0)
+                if argument.size in {4, 8, 16}:
+                    pass
+                elif argument.size == 32:
+                    argument.register = argument.register.as_ymm
+                elif argument.size == 64:
+                    argument.register = argument.register.as_zmm
+                else:
+                    assert False
+            elif (
+                argument.is_integer or argument.is_pointer or argument.is_codeunit
+            ) and len(available_integer_registers) > 0:
+                argument_register = available_integer_registers.pop(0)
+                argument.register = {
+                    1: argument_register.as_dword,
+                    2: argument_register.as_dword,
+                    4: argument_register.as_dword,
+                    8: argument_register,
+                }[argument.size]
+            elif argument.is_vector or argument.is_mask:
+                assert False
+            else:
+                argument.stack_offset = stack_offset
+                stack_offset += 8
+
+    def _setup_golang_arguments(self):
+        from nervapy.x86_64.abi import (goasm_amd64_abi, goasm_amd64p32_abi,
+                                        gosyso_amd64_abi, gosyso_amd64p32_abi)
+
+        assert self.abi in {
+            gosyso_amd64_abi,
+            gosyso_amd64p32_abi,
+            goasm_amd64_abi,
+            goasm_amd64p32_abi,
+        }, "This function must only be used with Golang AMD64 or AMD64p32 ABI"
+
+        from nervapy.util import roundup
+
+        # All arguments are passed on stack
+        # Stack offset does not include the return address
+        stack_offset = 0
+        for index, argument in enumerate(self.arguments):
+            # Arguments are aligned on stack
+            stack_offset = roundup(stack_offset, argument.size)
+            argument.stack_offset = stack_offset
+            stack_offset += argument.size
+        if self.result_type is not None:
+            self.result_offset = roundup(stack_offset, self.result_type.size)
+
+    def _layout_local_variables(self):
+        from nervapy.x86_64.registers import rsp
+
+        local_variables_set = set()
+        local_variables_list = list()
+        for instruction in self._instructions:
+            local_variable = instruction.local_variable
+            if local_variable is not None:
+                local_variable = local_variable.root
+                if local_variable not in local_variables_set:
+                    local_variables_set.add(local_variable)
+                    local_variables_list.append(local_variable)
+        if local_variables_list:
+            local_variables_list = list(
+                sorted(local_variables_list, key=lambda var: var.size)
+            )
+            self._stack_frame_alignment = max(
+                var.alignment for var in local_variables_list
+            )
+            local_variable_address = 0
+            from nervapy.util import roundup
+
+            for local_variable in local_variables_list:
+                local_variable_address = roundup(
+                    local_variable_address, local_variable.alignment
+                )
+                local_variable._address = local_variable_address
+                local_variable_address += local_variable.size
+            self._local_variables_size = local_variable_address
+
+            for instruction in self._instructions:
+                local_variable = instruction.local_variable
+                if local_variable is not None:
+                    assert local_variable.address is not None
+                    memory_address = instruction.memory_address
+                    assert memory_address is not None
+                    assert memory_address.base == rsp
+                    instruction.memory_address.displacement = local_variable.address
+
+    def _allocate_registers(self):
+        for register_kind, register_allocator in six.iteritems(
+            self._register_allocators
+        ):
+            register_allocator.set_allocation_options(self.abi, register_kind)
+
+        from nervapy.x86_64.pseudo import LOAD
+        from nervapy.x86_64.registers import (GeneralPurposeRegister,
+                                              KRegister, MMXRegister, Register,
+                                              XMMRegister)
+
+        for instruction in self._instructions:
+            if isinstance(instruction, LOAD.ARGUMENT):
+                dst_reg = instruction.operands[0]
+                src_arg = instruction.operands[1]
+                assert isinstance(dst_reg, Register)
+                assert isinstance(src_arg, Argument)
+                if dst_reg.is_virtual and src_arg.register is not None:
+                    self._register_allocators[dst_reg.kind].try_allocate_register(
+                        dst_reg.virtual_id, src_arg.register.physical_id
+                    )
+
+        for register_allocator in six.itervalues(self._register_allocators):
+            register_allocator.allocate_registers()
+
+    def _lower_argument_loads(self):
+        from nervapy.x86_64.abi import goasm_amd64_abi, goasm_amd64p32_abi
+        from nervapy.x86_64.lower import load_memory, load_register
+        from nervapy.x86_64.pseudo import LOAD
+        from nervapy.x86_64.registers import (GeneralPurposeRegister,
+                                              MMXRegister, XMMRegister,
+                                              YMMRegister)
+
+        if self.abi == goasm_amd64_abi or self.abi == goasm_amd64p32_abi:
+            # Like PeachPy, Go assembler uses pseudo-instructions for argument loads
+            return
+        lowered_instructions = []
+        for i, instruction in enumerate(self._instructions):
+            if isinstance(instruction, LOAD.ARGUMENT):
+                assert isinstance(
+                    instruction.operands[0],
+                    (GeneralPurposeRegister, MMXRegister, XMMRegister, YMMRegister),
+                ), "Lowering LOAD.ARGUMENT is supported only for general-purpose, mmx, xmm, and ymm target registers"
+                if instruction.operands[1].register is not None:
+                    # The argument is passed to function in a register
+                    ld_reg = load_register(
+                        instruction.operands[0],
+                        instruction.operands[1].register,
+                        instruction.operands[1].c_type,
+                        prototype=instruction,
+                    )
+                    if ld_reg is not None:
+                        lowered_instructions.append(ld_reg)
+                else:
+                    # The argument is passed to function on stack
+                    ld_mem = load_memory(
+                        instruction.operands[0],
+                        instruction.operands[1].address,
+                        instruction.operands[1].c_type,
+                        prototype=instruction,
+                    )
+                    lowered_instructions.append(ld_mem)
+            else:
+                lowered_instructions.append(instruction)
+        self._instructions = lowered_instructions
+
+    def _lower_pseudoinstructions(self):
+        from nervapy.stream import InstructionStream
+        from nervapy.util import is_int, is_sint32, is_uint32
+        from nervapy.x86_64.abi import (goasm_amd64_abi, goasm_amd64p32_abi,
+                                        gosyso_amd64_abi, gosyso_amd64p32_abi,
+                                        native_client_x86_64_abi)
+        from nervapy.x86_64.avx import VMOVAPS, VZEROUPPER
+        from nervapy.x86_64.generic import (ADD, AND, LEA, MOV, POP, PUSH, RET,
+                                            SUB, XOR)
+        from nervapy.x86_64.lower import load_register
+        from nervapy.x86_64.mmxsse import MOVAPS
+        from nervapy.x86_64.nacl import (NACLASP, NACLJMP, NACLRESTBP,
+                                         NACLRESTSP, NACLSSP)
+        from nervapy.x86_64.pseudo import RETURN, STORE
+        from nervapy.x86_64.registers import (GeneralPurposeRegister64,
+                                              XMMRegister, eax, rbp, rsp)
+
+        # The new list with lowered instructions
+        instructions = list()
+        # Generate prologue
+        cloberred_xmm_registers = list()
+        cloberred_general_purpose_registers = list()
+        with InstructionStream() as prolog_stream:
+            # 1. Save clobbered general-purpose registers with PUSH instruction
+            # 2. If there are clobbered XMM registers, allocate space for them on stack (subtract stack pointer)
+            # 3. Save clobbered XMM registers on stack with (V)MOVAPS instruction
+            for reg in self._clobbered_registers:
+                assert isinstance(
+                    reg, (GeneralPurposeRegister64, XMMRegister)
+                ), "Internal error: unexpected register %s in clobber list" % str(reg)
+                if isinstance(reg, GeneralPurposeRegister64):
+                    cloberred_general_purpose_registers.append(reg)
+                    PUSH(reg)
+                else:
+                    cloberred_xmm_registers.append(reg)
+            # If stack needs to be realigned
+            if self._stack_frame_alignment > self.abi.stack_alignment:
+                cloberred_general_purpose_registers.append(rbp)
+                PUSH(rbp)
+                MOV(rbp, rsp)
+            if cloberred_xmm_registers or self._local_variables_size != 0:
+                # Total size of the stack frame less what is already adjusted with PUSH instructions
+                stack_adjustment = (
+                    self._stack_frame_size
+                    - len(cloberred_general_purpose_registers)
+                    * GeneralPurposeRegister64.size
+                )
+                if self.abi != native_client_x86_64_abi:
+                    SUB(rsp, stack_adjustment + self._local_variables_size)
+                    if self._stack_frame_alignment > self.abi.stack_alignment:
+                        AND(rsp, -self._stack_frame_alignment)
+                else:
+                    if self._stack_frame_alignment > self.abi.stack_alignment:
+                        # Note: do not modify rcx/rdx/r8/r9 as they may contain function arguments
+                        LEA(
+                            eax, [rsp - (stack_adjustment + self._local_variables_size)]
+                        )
+                        AND(eax, -self._stack_frame_alignment)
+                        NACLRESTSP(eax)
+                    else:
+                        NACLSSP(stack_adjustment + self._local_variables_size)
+            for i, xmm_reg in enumerate(cloberred_xmm_registers):
+                movaps = VMOVAPS if self._avx_prolog else MOVAPS
+                movaps(
+                    [rsp + self._local_variables_size + i * XMMRegister.size], xmm_reg
+                )
+
+        # TODO: handle situations when entry point is in the middle of a function
+        instructions.extend(prolog_stream.instructions)
+
+        for instruction in self._instructions:
+            if isinstance(instruction, RETURN):
+                from nervapy.x86_64.registers import (GeneralPurposeRegister,
+                                                      MMXRegister, XMMRegister,
+                                                      YMMRegister, al, ax, eax,
+                                                      ecx, mm0, rax, rcx, xmm0,
+                                                      ymm0)
+
+                is_goasm_abi = self.abi in {goasm_amd64_abi, goasm_amd64p32_abi}
+                is_gosyso_abi = self.abi in {gosyso_amd64_abi, gosyso_amd64p32_abi}
+                with InstructionStream() as epilog_stream:
+                    # Save return value
+                    if instruction.operands:
+                        assert len(instruction.operands) == 1
+                        if is_int(instruction.operands[0]):
+                            assert (
+                                self.result_type.is_integer
+                                or self.result_type.is_pointer
+                            )
+                            # Return immediate constant
+                            if is_goasm_abi:
+                                # Return value must be saved on stack with STORE.RESULT pseudo-instruction
+                                if self.result_type.size <= 4 or is_sint32(
+                                    instruction.operands[0]
+                                ):
+                                    # STORE.RESULT will assemble to one of the forms:
+                                    # - MOV m8, imm8
+                                    # - MOV m16, imm16
+                                    # - MOV m32, imm32
+                                    # - MOV m64, imm32
+                                    STORE.RESULT(
+                                        instruction.operands[0],
+                                        prototype=instruction,
+                                        target_function=self,
+                                    )
+                                else:
+                                    # STORE.RESULT can't be used directly (MOV m64, imm64 doesn't exist), instead use
+                                    # MOV rax, imm64 + MOV m64, rax (STORE.RESULT)
+                                    MOV(
+                                        eax,
+                                        instruction.operands[0],
+                                        prototype=instruction,
+                                    )
+                                    STORE.RESULT(
+                                        eax, prototype=instruction, target_function=self
+                                    )
+                            else:
+                                # Return value is returned in:
+                                # - eax register if result type is not greater than 4 bytes
+                                # - rax register if result type is greater than 8 bytes
+                                if instruction.operands[0] == 0:
+                                    # - Zero eax register (high 32 bits of rax register clear automatically)
+                                    XOR(eax, eax, prototype=instruction)
+                                elif self.result_type.size <= 4 or is_uint32(
+                                    instruction.operands[0]
+                                ):
+                                    # - If the result type is not greater than 4 bytes, directly mov it to eax register
+                                    # - If the result type is greater than 4 bytes, but the result value is
+                                    #   representable as unsigned 32-bit literal, mov it to eax register and the high
+                                    #   32 bits of rax will be cleared automatically
+                                    MOV(
+                                        eax,
+                                        instruction.operands[0],
+                                        prototype=instruction,
+                                    )
+                                else:
+                                    # - Either negative 32-bit constant (would use MOV rax, imm32 form)
+                                    # - Or large 64-bit constant (would use MOV rax, imm64 form)
+                                    MOV(
+                                        rax,
+                                        instruction.operands[0],
+                                        prototype=instruction,
+                                    )
+                        elif isinstance(
+                            instruction.operands[0], GeneralPurposeRegister
+                        ):
+                            if (
+                                is_goasm_abi
+                                and instruction.operands[0].size
+                                == self.result_type.size
+                            ):
+                                STORE.RESULT(
+                                    instruction.operands[0],
+                                    prototype=instruction,
+                                    target_function=self,
+                                )
+                            else:
+                                result_reg = eax if self.result_type.size <= 4 else rax
+                                epilog_stream.add_instruction(
+                                    load_register(
+                                        result_reg,
+                                        instruction.operands[0],
+                                        self.result_type,
+                                        prototype=instruction,
+                                    )
+                                )
+                                if is_goasm_abi:
+                                    result_subreg = {1: al, 2: ax, 4: eax, 8: rax}[
+                                        self.result_type.size
+                                    ]
+                                    STORE.RESULT(
+                                        result_subreg,
+                                        prototype=instruction,
+                                        target_function=self,
+                                    )
+                        elif isinstance(instruction.operands[0], MMXRegister):
+                            epilog_stream.add_instruction(
+                                load_register(
+                                    mm0,
+                                    instruction.operands[0],
+                                    self.result_type,
+                                    prototype=instruction,
+                                )
+                            )
+                        elif isinstance(instruction.operands[0], XMMRegister):
+                            if self.result_type.is_floating_point and is_goasm_abi:
+                                assert self.result_type.size in {4, 8}
+                                STORE.RESULT(
+                                    instruction.operands[0],
+                                    prototype=instruction,
+                                    target_function=self,
+                                )
+                            else:
+                                epilog_stream.add_instruction(
+                                    load_register(
+                                        xmm0,
+                                        instruction.operands[0],
+                                        self.result_type,
+                                        prototype=instruction,
+                                    )
+                                )
+                        elif isinstance(instruction.operands[0], YMMRegister):
+                            epilog_stream.add_instruction(
+                                load_register(
+                                    ymm0,
+                                    instruction.operands[0],
+                                    self.result_type,
+                                    prototype=instruction,
+                                )
+                            )
+                        else:
+                            assert False
+                    if instruction.avx_mode and not self.avx_environment:
+                        VZEROUPPER(prototype=instruction)
+                    # Generate epilog
+                    # 1. Restore clobbered XMM registers on stack with (V)MOVAPS instruction
+                    # 2. If there are clobbered XMM registers, release their space on stack (increment stack pointer)
+                    # 3. Restore clobbered general-purpose registers with PUSH instruction
+                    for i, xmm_reg in enumerate(cloberred_xmm_registers):
+                        movaps = VMOVAPS if self.avx_environment else MOVAPS
+                        movaps(
+                            xmm_reg,
+                            [rsp + self._local_variables_size + i * XMMRegister.size],
+                        )
+                    if self._stack_frame_alignment > self.abi.stack_alignment:
+                        # Restore rsp value from rbp
+                        MOV(rsp, rbp)
+                    elif cloberred_xmm_registers or self._local_variables_size != 0:
+                        # Total size of the stack frame less what will be adjusted with POP instructions
+                        stack_adjustment = (
+                            self._stack_frame_size
+                            - len(cloberred_general_purpose_registers)
+                            * GeneralPurposeRegister64.size
+                        )
+                        if self.abi != native_client_x86_64_abi:
+                            ADD(rsp, stack_adjustment + self._local_variables_size)
+                        else:
+                            NACLASP(stack_adjustment + self._local_variables_size)
+                    # Important: registers must be POPed in reverse order
+                    for reg in reversed(cloberred_general_purpose_registers):
+                        if reg == rbp and self.abi == native_client_x86_64_abi:
+                            POP(rcx)
+                            NACLRESTBP(ecx)
+                        else:
+                            POP(reg)
+                    # Return from the function
+                    if self.abi == native_client_x86_64_abi:
+                        POP(rcx, prototype=instruction)
+                        NACLJMP(ecx)
+                    else:
+                        RET(prototype=instruction)
+                instructions.extend(epilog_stream.instructions)
+            elif isinstance(instruction, STORE.RESULT):
+                instruction.destination_offset = self.result_offset
+                instructions.append(instruction)
+            else:
+                if self.abi == native_client_x86_64_abi and instruction.name != "LEA":
+                    from nervapy.x86_64.operand import is_m
+
+                    memory_operands = list(
+                        filter(lambda op: is_m(op), instruction.operands)
+                    )
+                    if memory_operands:
+                        assert (
+                            len(memory_operands) == 1
+                        ), "x86-64 instructions can not have more than 1 explicit memory operand"
+                        memory_address = memory_operands[0].address
+                        from nervapy.x86_64.operand import MemoryAddress
+
+                        if isinstance(memory_address, MemoryAddress):
+                            if memory_address.index is not None:
+                                raise ValueError("NaCl does not allow index addressing")
+                            from nervapy.x86_64.registers import r15, rbp, rsp
+
+                            if (
+                                memory_address.base is not None
+                                and memory_address.base not in {rbp, rsp, r15}
+                            ):
+                                # Base register is not a restricted register: needs transformation
+                                memory_address.index = memory_address.base
+                                memory_address.scale = 1
+                                memory_address.base = r15
+                instructions.append(instruction)
+        self._instructions = instructions
+
+    def _filter_instruction_encodings(self):
+        for instruction in self._instructions:
+            instruction.encodings = instruction._filter_encodings()
+
+    def _update_argument_addresses(self):
+        for argument in self.arguments:
+            if argument.stack_offset is not None:
+                argument.address = self._argument_stack_base + argument.stack_offset
+
+    def _analyze_clobbered_registers(self):
+        from nervapy.x86_64.registers import (GeneralPurposeRegister,
+                                              XMMRegister, YMMRegister,
+                                              ZMMRegister)
+
+        output_subregisters = set()
+        for instruction in self._instructions:
+            output_subregisters.update(instruction.output_registers)
+        output_registers = set()
+        for subreg in output_subregisters:
+            if isinstance(subreg, GeneralPurposeRegister):
+                output_registers.add(subreg.as_qword)
+            elif isinstance(subreg, (XMMRegister, YMMRegister, ZMMRegister)):
+                output_registers.add(subreg.as_xmm)
+            # Other register types are volatile registers for all x86-64 ABIs
+        return list(
+            sorted(
+                filter(
+                    lambda reg: reg in self.abi.callee_save_registers, output_registers
+                )
+            )
+        )
+
+    def _update_stack_frame(self):
+        from nervapy.x86_64.registers import (GeneralPurposeRegister64,
+                                              XMMRegister, rbp, rsp)
+
+        clobbered_general_purpose_registers = 0
+        clobbered_xmm_registers = 0
+        for reg in self._clobbered_registers:
+            assert isinstance(
+                reg, (GeneralPurposeRegister64, XMMRegister)
+            ), "Internal error: unexpected register %s in clobber list" % str(reg)
+            if isinstance(reg, GeneralPurposeRegister64):
+                clobbered_general_purpose_registers += 1
+            else:
+                clobbered_xmm_registers += 1
+        # If the stack needs to be aligned, rbp register needs to be preserved too
+        if self._stack_frame_alignment > self.abi.stack_alignment:
+            clobbered_general_purpose_registers += 1
+        self._stack_frame_size = (
+            clobbered_general_purpose_registers * GeneralPurposeRegister64.size
+            + clobbered_xmm_registers * XMMRegister.size
+        )
+        # 1. On function entry stack is misaligned by 8
+        # 2. Each clobbered general-purpose register is pushed as 8 bytes
+        # 3. If the number of clobbered general-purpose registers is odd, the stack will be misaligned by 8 after they
+        #    are pushed on stack
+        # 4. If additionally there are clobbered XMM registers, we need to subtract 8 from stack to make it aligned
+        #    by 16 after the general-purpose registers are pushed
+        if (
+            clobbered_xmm_registers != 0 or self._local_variables_size != 0
+        ) and clobbered_general_purpose_registers % 2 == 0:
+            self._stack_frame_size += 8
+
+        # Set stack_argument_base
+        return_address_size = 8
+        if self._stack_frame_alignment > self.abi.stack_alignment:
+            # rsp is realigned, argument addressing uses rbp
+            saved_rbp_size = 8
+            self._argument_stack_base = rbp + return_address_size + saved_rbp_size
+        else:
+            # argument addressing uses rsp
+            self._argument_stack_base = (
+                rsp
+                + return_address_size
+                + self._stack_frame_size
+                + self._local_variables_size
+            )
+
+    def _bind_registers(self):
+        """Iterates through the list of instructions and assigns physical IDs to allocated registers"""
+
+        for instruction in self._instructions:
+            for register in instruction.register_objects:
+                if register.is_virtual:
+                    register.physical_id = self._register_allocators[
+                        register.kind
+                    ].register_allocations[register.virtual_id]
+
+    def format_code(
+        self,
+        assembly_format="peachpy",
+        line_separator=os.linesep,
+        indent=True,
+        line_number=1,
+    ):
+        """Returns code of assembly instructions comprising the function"""
+
+        code = []
+        if assembly_format == "gas":
+            # Pre-assign line number to labels
+            from nervapy.x86_64.pseudo import LABEL
+
+            for i, instruction in enumerate(self._instructions):
+                if isinstance(instruction, LABEL):
+                    instruction.operands[0].line_number = line_number + i
+
+        for i, instruction in enumerate(self._instructions):
+            from nervapy.x86_64.instructions import Instruction
+
+            # if isinstance(instruction, Instruction):
+            #     try:
+            #         hex_string = " ".join("%02X" % byte for byte in instruction.encode())
+            #         code.append("    " + "# " + hex_string)
+            #     except Exception as e:
+            #         import sys
+            #         code.append(e.message)
+            #         # raise
+            code.append(
+                instruction.format(
+                    assembly_format=assembly_format,
+                    indent=indent,
+                    line_number=line_number + i,
+                )
+            )
+        if line_separator is None:
+            return code
+        else:
+            return str(line_separator).join(code)
+
+    def format(
+        self, assembly_format="peachpy", line_separator=os.linesep, line_number=1
+    ):
+        """Formats assembly listing of the function according to specified parameters"""
+
+        if assembly_format == "go":
+            # Arguments for TEXT directive in Go assembler
+            package_string = self.package
+            if package_string is None:
+                package_string = ""
+            if six.PY2:
+                text_arguments = [
+                    package_string + "\xc2\xb7" + self.mangled_name + "(SB)"
+                ]
+            else:
+                text_arguments = [
+                    package_string + "\u00b7" + self.mangled_name + "(SB)"
+                ]
+
+            text_arguments.append("4")
+            stack_size = sum(map(operator.attrgetter("size"), self.arguments))
+            if self.result_offset is not None:
+                stack_size = self.result_offset + self.result_type.size
+            if stack_size == 0:
+                text_arguments.append("$0")
+            else:
+                text_arguments.append("$0-%d" % stack_size)
+
+            code = ["TEXT " + ",".join(text_arguments)]
+            if self.go_signature is not None:
+                code.insert(0, "// " + self.go_signature)
+        elif assembly_format == "gas":
+            from nervapy.util import ilog2
+
+            code_alignment = 16
+            code = [
+                "#ifdef __APPLE__",
+                ".section __TEXT,__text,regular,pure_instructions",
+                ".globl _{name}".format(name=self.mangled_name),
+                ".p2align {ilog2alignment}, 0x90".format(
+                    ilog2alignment=ilog2(code_alignment)
+                ),
+                "_{name}:".format(name=self.mangled_name),
+                "#else /* !__APPLE__ */",
+                ".text",
+                ".p2align {ilog2alignment},,{max_alignment_bytes}".format(
+                    ilog2alignment=ilog2(code_alignment),
+                    max_alignment_bytes=code_alignment - 1,
+                ),
+                ".globl " + self.mangled_name,
+                ".type {name}, @function".format(name=self.mangled_name),
+                "{name}:".format(name=self.mangled_name),
+                "#endif /* !__APPLE */",
+            ]
+        else:
+            code = []
+
+        code += self.format_code(
+            assembly_format,
+            line_separator=None,
+            indent=True,
+            line_number=line_number + len(code),
+        )
+        if assembly_format == "gas":
+            code += [
+                "#ifndef __APPLE__",
+                ".size {name}, .-{name}".format(name=self.mangled_name),
+                "#endif /* !__APPLE__ */",
+            ]
+
+        if assembly_format in ["go", "gas"]:
+            # Add trailing line or assembler will refuse to compile
+            code.append("")
+        if line_separator is None:
+            return code
+        else:
+            return str(line_separator).join(code)
+
+    def encode(self):
+        return EncodedFunction(self)
+
+    @property
+    def metadata(self):
+        metadata = collections.OrderedDict(
+            [
+                ("entry", "function"),
+                ("name", self.name),
+                ("symbol", self.mangled_name),
+                (
+                    "return",
+                    "void" if self.result_type is None else str(self.result_type),
+                ),
+                (
+                    "arguments",
+                    [
+                        collections.OrderedDict(
+                            [("name", argument.name), ("type", str(argument.c_type))]
+                        )
+                        for argument in self.arguments
+                    ],
+                ),
+                ("arch", "x86-64"),
+                ("abi", str(self.abi)),
+                ("uarch", self.target.name),
+                ("isa", [str(extension) for extension in self.isa_extensions.minify()]),
+            ]
+        )
+        return metadata
+
+    def mangle_name(self):
+        import string
+
+        import nervapy.x86_64.options
+
+        name = (
+            nervapy.x86_64.options.name_mangling.replace("${Name}", self.name)
+            .replace("${name}", self.name.lower())
+            .replace("${NAME}", self.name.upper())
+            .replace("${uArch}", self.target.id)
+            .replace("${uarch}", self.target.id.lower())
+            .replace("${UARCH}", self.target.id.upper())
+            .replace(
+                "${ISA}",
+                "_".join(
+                    [extension.safe_name for extension in self.isa_extensions.minify()]
+                ),
+            )
+            .replace(
+                "${isa}",
+                "_".join(
+                    [
+                        extension.safe_name.lower()
+                        for extension in self.isa_extensions.minify()
+                    ]
+                ),
+            )
+        )
+        return name
+
+
+class InstructionBundle:
+    def __init__(self, capacity, address):
+        if capacity not in {16, 32, 64}:
+            raise ValueError("Bundle capacity must be 16, 32, or 64")
+        self.capacity = capacity
+        self.address = address
+        self.size = 0
+        self._instructions = []
+        # Map from instruction position to tuple (label address, long encoding, short range)
+        self.branch_info_map = dict()
+
+    @property
+    def padding(self):
+        return self.capacity - self.size
+
+    def add(self, instructions):
+        from nervapy.x86_64.instructions import Instruction
+
+        assert isinstance(instructions, list)
+        assert all(
+            isinstance(instruction, Instruction) for instruction in instructions
+        ), "Instruction instance expected"
+        bytecode = bytearray().join(
+            [instruction.encode() for instruction in instructions]
+        )
+        if self.size + len(bytecode) <= self.capacity:
+            self.size += len(bytecode)
+            for instruction in instructions:
+                instruction.bytecode = instruction.encode()
+                self._instructions.append(instruction)
+        else:
+            raise BufferError()
+
+    def add_label_branch(self, instruction, label_address=None, long_encoding=False):
+        from nervapy.x86_64.instructions import BranchInstruction
+
+        assert isinstance(
+            instruction, BranchInstruction
+        ), "BranchInstruction instance expected"
+        long_encoding, bytecode = instruction._encode_label_branch(
+            self.address + self.size, label_address, long_encoding
+        )
+        if self.capacity - self.size > len(bytecode):
+            self.size += len(bytecode)
+            if not long_encoding and label_address is not None:
+                # offset = label_address - self.end_address
+                # -> self.end_address = label_address - offset
+                # -> branch_address = label_address - len(bytecode) - offset
+                # -> branch_position = label_address - self.start_address - len(bytecode) - offset
+                #
+                # -> branch_pos >= label_address - self.start_address - len(bytecode) - 127
+                # -> branch_pos <= label_address - self.start_address - len(bytecode) + 128
+                branch_pos = label_address - self.address - len(bytecode)
+                branch_pos_max = min(branch_pos + 128, self.capacity)
+            else:
+                branch_pos_max = self.capacity
+            instruction.bytecode = bytecode
+            self.branch_info_map[len(self._instructions)] = (
+                label_address,
+                long_encoding,
+                branch_pos_max,
+            )
+            self._instructions.append(instruction)
+        else:
+            raise BufferError()
+
+    def optimize(self):
+        from nervapy.x86_64.instructions import BranchInstruction
+        from nervapy.x86_64.pseudo import LABEL
+
+        if any(
+            isinstance(instruction, (BranchInstruction, LABEL))
+            for instruction in self._instructions
+        ):
+            return
+
+        def suitable_encodings(instruction):
+            return [
+                (encoding, length)
+                for (length, encoding) in six.iteritems(
+                    instruction.encode_length_options()
+                )
+                if 0 < length - len(instruction.bytecode) <= self.padding
+            ]
+
+        while self.size < self.capacity:
+            suitable_instructions = [
+                instr for instr in self._instructions if any(suitable_encodings(instr))
+            ]
+            if not suitable_instructions:
+                break
+
+            shortest_suitable_instruction = min(
+                suitable_instructions, key=lambda instr: len(instr.bytecode)
+            )
+            new_encoding, new_length = min(
+                suitable_encodings(shortest_suitable_instruction),
+                key=operator.itemgetter(1),
+            )
+            self.size += new_length - len(shortest_suitable_instruction.bytecode)
+            assert self.size <= self.capacity
+            shortest_suitable_instruction.bytecode = new_encoding
+
+    def finalize(self):
+        from nervapy.x86_64.generic import NOP
+
+        while self.capacity > self.size:
+            self.add([NOP()])
+        self.size = self.capacity
+
+    @property
+    def label_address_map(self):
+        from nervapy.x86_64.pseudo import LABEL
+
+        label_address_map = dict()
+        code_address = self.address
+        for instruction in self._instructions:
+            if isinstance(instruction, LABEL):
+                label_address_map[instruction.identifier] = code_address
+            else:
+                code_address += len(instruction.bytecode)
+        return label_address_map
+
+    def __len__(self):
+        return self.size
+
+
+class EncodedFunction:
+    """ABI-specific x86-64 assembly function.
+
+    A function consists of C signature, ABI, and a list of instructions without virtual registers.
+    """
+
+    def __init__(self, function):
+        from copy import copy, deepcopy
+
+        assert isinstance(function, ABIFunction), "ABIFunction object expected"
+        self.name = function.name
+        self.mangled_name = function.mangled_name
+        self.arguments = list(map(copy, function.arguments))
+        self.result_type = function.result_type
+        self.target = function.target
+        self.abi = function.abi
+
+        from nervapy.x86_64.abi import native_client_x86_64_abi
+        from nervapy.x86_64.meta import Section, SectionType
+
+        if self.abi == native_client_x86_64_abi:
+            # Align with HLT instruction
+            self.code_section = Section(SectionType.code, alignment_byte=0xF4)
+            self.code_section.alignment = 32
+        else:
+            # Align with INT 3 instruction
+            self.code_section = Section(SectionType.code, alignment_byte=0xCC)
+            self.code_section.alignment = 16
+
+        self.const_section = Section(SectionType.const_data)
+
+        self._instructions = deepcopy(function._instructions)
+
+        self._constant_symbol_map = dict()
+        self._layout_literal_constants()
+        self._encode()
+
+    def _layout_literal_constants(self):
+        from nervapy.encoder import Encoder
+        from nervapy.x86_64.meta import Symbol, SymbolType
+
+        encoder = Encoder(self.abi.endianness)
+
+        constants = list()
+        for instruction in self._instructions:
+            constant = instruction.constant
+            if constant is not None:
+                constants.append(constant)
+
+        max_constant_size = 0
+        max_constant_alignment = 0
+        if constants:
+            max_constant_size = max(constant.size for constant in constants)
+            max_constant_alignment = max(constant.alignment for constant in constants)
+        self.const_section.alignment = max_constant_alignment
+
+        # Unsorted list of Symbol objects for constants
+        constant_symbols = list()
+        # This set is used to ensure that each constant is added only once
+        constant_names_set = set()
+        if max_constant_size != 0:
+            # Map from constant value (as bytes) to address in the const data section
+            constants_address_map = dict()
+
+            for instruction in self._instructions:
+                constant = instruction.constant
+                if constant is not None:
+                    constant_value = bytes(constant.encode(encoder))
+                    if constant_value not in constants_address_map:
+                        # Add the new constant to the section
+                        assert (
+                            constant.size == max_constant_size
+                        ), "Handling of functions with constant literals of different size is not implemented"
+                        assert (
+                            constant.alignment == max_constant_alignment
+                        ), "Handling of functions with constant literals of different alignment is not implemented"
+                        constants_address_map[constant_value] = len(self.const_section)
+                        self.const_section.content += constant_value
+                    if constant.name not in constant_names_set:
+                        constant_names_set.add(constant.name)
+                        const_symbol = Symbol(
+                            constants_address_map[constant_value],
+                            SymbolType.literal_constant,
+                            name=constant.name,
+                            size=constant.size,
+                        )
+                        constant_symbols.append(const_symbol)
+                        self._constant_symbol_map[constant.name] = const_symbol
+        for constant_symbol in sorted(
+            constant_symbols, key=lambda sym: (sym.offset, -sym.size)
+        ):
+            self.const_section.add_symbol(constant_symbol)
+
+    def _encode(self):
+        from nervapy.x86_64.instructions import BranchInstruction
+        from nervapy.x86_64.pseudo import LABEL
+
+        label_address_map = dict()
+        long_branches = set()
+
+        # Special post-processing for Native Client SFI
+        from nervapy.x86_64.abi import native_client_x86_64_abi
+
+        if self.abi == native_client_x86_64_abi:
+            has_updated_branches = True
+            has_unresolved_labels = True
+            bundles = list()
+            while has_updated_branches or has_unresolved_labels:
+                code_address = 0
+                has_updated_branches = False
+                has_unresolved_labels = False
+                bundles = list()
+                current_bundle = InstructionBundle(32, code_address)
+                for i, instruction in enumerate(self._instructions):
+                    if isinstance(instruction, LABEL):
+                        label_address_map[instruction.identifier] = code_address
+                        current_bundle.add([instruction])
+                    elif (
+                        isinstance(instruction, BranchInstruction)
+                        and instruction.label_name
+                    ):
+                        label_address = label_address_map.get(instruction.label_name)
+                        if label_address is None:
+                            has_unresolved_labels = True
+                        was_long = i in long_branches
+                        is_long, instruction.bytecode = (
+                            instruction._encode_label_branch(
+                                code_address, label_address, long_encoding=was_long
+                            )
+                        )
+                        if is_long and not was_long:
+                            long_branches.add(i)
+                            has_updated_branches = True
+                        try:
+                            current_bundle.add_label_branch(
+                                instruction, label_address, is_long
+                            )
+                        except BufferError:
+                            bundles.append(current_bundle)
+                            current_bundle = InstructionBundle(
+                                32, current_bundle.address + current_bundle.capacity
+                            )
+                            current_bundle.add_label_branch(
+                                instruction, label_address, is_long
+                            )
+                    else:
+                        instruction_group = [instruction]
+
+                        memory_address = instruction.memory_address
+                        from nervapy.x86_64.operand import MemoryAddress
+
+                        if (
+                            isinstance(memory_address, MemoryAddress)
+                            and memory_address.index is not None
+                        ):
+                            from nervapy.stream import NullStream
+
+                            with NullStream():
+                                from nervapy.x86_64.generic import MOV
+
+                                instruction_group.insert(
+                                    0,
+                                    MOV(
+                                        memory_address.index.as_dword,
+                                        memory_address.index.as_dword,
+                                    ),
+                                )
+                        try:
+                            current_bundle.add(instruction_group)
+                        except BufferError:
+                            bundles.append(current_bundle)
+                            current_bundle = InstructionBundle(
+                                32, current_bundle.address + current_bundle.capacity
+                            )
+                            current_bundle.add(instruction_group)
+                    code_address = current_bundle.address + current_bundle.size
+                bundles.append(current_bundle)
+            self._instructions = list()
+            for bundle in bundles:
+                bundle.optimize()
+                for instruction in bundle._instructions:
+                    constant = instruction.constant
+                    if constant:
+                        relocation = instruction.relocation
+                        for index in range(relocation.offset, relocation.offset + 4):
+                            instruction.bytecode[index] = 0
+                        relocation.offset += len(self.code_section)
+                        relocation.program_counter += len(self.code_section)
+                        relocation.symbol = self._constant_symbol_map[
+                            instruction.constant.name
+                        ]
+                        self.code_section.add_relocation(relocation)
+
+                    if instruction.bytecode:
+                        self.code_section.content += instruction.bytecode
+                if bundle.size < bundle.capacity:
+                    if bundle is not bundles[-1]:
+                        self.code_section.content += self._encode_nops(
+                            bundle.capacity - bundle.size
+                        )
+                    else:
+                        self.code_section.content += self._encode_abort(
+                            bundle.capacity - bundle.size
+                        )
+        else:
+            has_updated_branches = True
+            has_unresolved_labels = True
+            while has_updated_branches or has_unresolved_labels:
+                code_address = 0
+                has_updated_branches = False
+                has_unresolved_labels = False
+                for i, instruction in enumerate(self._instructions):
+                    if isinstance(instruction, LABEL):
+                        label_address_map[instruction.identifier] = code_address
+                    elif (
+                        isinstance(instruction, BranchInstruction)
+                        and instruction.label_name
+                    ):
+                        label_address = label_address_map.get(instruction.label_name)
+                        if label_address is None:
+                            has_unresolved_labels = True
+                        was_long = i in long_branches
+                        is_long, instruction.bytecode = (
+                            instruction._encode_label_branch(
+                                code_address, label_address, long_encoding=was_long
+                            )
+                        )
+                        if is_long and not was_long:
+                            long_branches.add(i)
+                            has_updated_branches = True
+                    else:
+                        instruction.bytecode = instruction.encode()
+                    if instruction.bytecode:
+                        code_address += len(instruction.bytecode)
+
+            for instruction in self._instructions:
+                constant = instruction.constant
+                if constant:
+                    relocation = instruction.relocation
+                    for index in range(relocation.offset, relocation.offset + 4):
+                        instruction.bytecode[index] = 0
+                    relocation.offset += len(self.code_section)
+                    relocation.program_counter += len(self.code_section)
+                    relocation.symbol = self._constant_symbol_map[
+                        instruction.constant.name
+                    ]
+                    self.code_section.add_relocation(relocation)
+
+                if instruction.bytecode:
+                    self.code_section.content += instruction.bytecode
+
+    def _encode_nops(self, length):
+        assert 1 <= length <= 31
+        from nervapy.x86_64.encoding import nop
+
+        if length <= 15:
+            return nop(length)
+        elif length <= 30:
+            return nop(length // 2) + nop(length - length // 2)
+        else:
+            return nop(8) + nop(8) + nop(15)
+
+    def _encode_abort(self, length):
+        from nervapy.x86_64.abi import (goasm_amd64_abi, goasm_amd64p32_abi,
+                                        native_client_x86_64_abi)
+
+        if self.abi == native_client_x86_64_abi:
+            # Use HLT instructions
+            return bytearray([0xF4] * length)
+        elif self.abi in {goasm_amd64_abi, goasm_amd64p32_abi}:
+            # Use a single INT 3 instruction as alignment is not supported anyway
+            return bytearray([0xCD])
+        else:
+            # Use INT 3 instructions
+            return bytearray([0xCD] * length)
+
+    def format_code(
+        self, assembly_format="peachpy", line_separator=os.linesep, indent=True
+    ):
+        """Returns code of assembly instructions comprising the function"""
+
+        code = []
+        for instruction in self._instructions:
+            code.append(instruction.format_encoding(indent=indent))
+            code.append(
+                instruction.format(assembly_format=assembly_format, indent=indent)
+            )
+        if line_separator is None:
+            return code
+        else:
+            return str(line_separator).join(filter(lambda line: line is not None, code))
+
+    def format(self, assembly_format="peachpy", line_separator=os.linesep):
+        """Formats assembly listing of the function according to specified parameters"""
+
+        if assembly_format == "go":
+            # Arguments for TEXT directive in Go assembler
+            text_arguments = ["%s\xc2\xb7%s(SB)" % (self.package_name, self.name)]
+
+            text_arguments.append("4")
+            text_arguments.append("$0")
+
+            code = ["TEXT " + ",".join(text_arguments)]
+        else:
+            code = []
+
+        code.extend(self.format_code(assembly_format, line_separator=None, indent=True))
+        if assembly_format == "go":
+            # Add trailing line or assembler will refuse to compile
+            code.append("")
+        if line_separator is None:
+            return code
+        else:
+            return str(line_separator).join(filter(bool, code))
+
+    def load(self):
+        return ExecutableFuntion(self)
+
+
+class ExecutableFuntion:
+    def __init__(self, function):
+        assert isinstance(function, EncodedFunction), "EncodedFunction object expected"
+        import nervapy.x86_64.abi
+
+        process_abi = nervapy.x86_64.abi.detect()
+        if process_abi != function.abi:
+            raise ValueError(
+                "Function ABI (%s) does not match process ABI (%s)"
+                % (str(function.abi), str(process_abi))
+            )
+
+        self.code_segment = bytearray(function.code_section.content)
+        self.const_segment = bytearray(function.const_section.content)
+
+        import nervapy.loader
+
+        self.loader = nervapy.loader.Loader(
+            len(self.code_segment), len(self.const_segment)
+        )
+
+        # Apply relocations
+        from nervapy.util import is_sint32
+        from nervapy.x86_64.meta import RelocationType
+
+        for relocation in function.code_section.relocations:
+            assert relocation.type == RelocationType.rip_disp32
+            assert relocation.symbol in function.const_section.symbols
+            old_value = (
+                self.code_segment[relocation.offset]
+                | (self.code_segment[relocation.offset + 1] << 8)
+                | (self.code_segment[relocation.offset + 2] << 16)
+                | (self.code_segment[relocation.offset + 3] << 24)
+            )
+            new_value = (
+                old_value
+                + (self.loader.data_address + relocation.symbol.offset)
+                - (self.loader.code_address + relocation.program_counter)
+            )
+            assert is_sint32(new_value)
+            self.code_segment[relocation.offset] = new_value & 0xFF
+            self.code_segment[relocation.offset + 1] = (new_value >> 8) & 0xFF
+            self.code_segment[relocation.offset + 2] = (new_value >> 16) & 0xFF
+            self.code_segment[relocation.offset + 3] = (new_value >> 24) & 0xFF
+        assert not function.const_section.relocations
+
+        self.loader.copy_code(self.code_segment)
+        self.loader.copy_data(self.const_segment)
+
+        import ctypes
+
+        result_type = (
+            None
+            if function.result_type is None
+            else function.result_type.as_ctypes_type
+        )
+        argument_types = [arg.c_type.as_ctypes_type for arg in function.arguments]
+        self.function_type = ctypes.CFUNCTYPE(result_type, *argument_types)
+        self.function_pointer = self.function_type(self.loader.code_address)
+
+    def __call__(self, *args):
+        return self.function_pointer(*args)
+
+    def __del__(self):
+        del self.loader
+        self.loader = None
+        self.function_pointer = None
+
+
+class LocalVariable:
+    def __init__(self, size_option, alignment=None):
+        from nervapy.util import is_int
+
+        if alignment is not None and not is_int(alignment):
+            raise TypeError("alignment %s is not an integer" % str(alignment))
+        if alignment is not None and alignment <= 0:
+            raise ValueError("alignment %d is not a positive integer" % alignment)
+        self.alignment = alignment
+        if is_int(size_option):
+            if size_option <= 0:
+                raise ValueError("size %d is not a positive integer" % size_option)
+            self.size = size_option
+        elif isinstance(size_option, nervapy.x86_64.registers.Register):
+            self.size = size_option.size
+        else:
+            raise TypeError(
+                "Unsupported size specification %s: register or integer expected"
+                % size_option
+            )
+        if self.alignment is None:
+            self.alignment = self.size
+        self._address = None
+        self._offset = 0
+        self.parent = None
+
+    def __eq__(self, other):
+        return (
+            isinstance(other, LocalVariable)
+            and self.root is other.root
+            and self.size == other.size
+            and self.offset == other.offset
+        )
+
+    def __ne__(self, other):
+        return (
+            not isinstance(other, LocalVariable)
+            or self.root is not other.root
+            or self.size != other.size
+            or self.offset != other.offset
+        )
+
+    def __hash__(self):
+        return id(self.root) ^ hash(self.size) ^ hash(self.offset)
+
+    def __str__(self):
+        if self.address is not None:
+            return "[" + str(self.address) + "]"
+        else:
+            return "local-variable<%d[%d:%d]>" % (
+                id(self.root),
+                self.offset,
+                self.offset + self.size,
+            )
+
+    def __repr__(self):
+        return str(self)
+
+    @property
+    def is_subvariable(self):
+        return self.parent is not None
+
+    @property
+    def root(self):
+        root = self
+        while root.parent is not None:
+            root = root.parent
+        return root
+
+    @property
+    def offset(self):
+        node = self
+        offset = 0
+        while node.parent is not None:
+            offset += node._offset
+            node = node.parent
+        return offset
+
+    @property
+    def address(self):
+        if self.is_subvariable:
+            base_address = self.root.address
+            if base_address is not None:
+                return base_address + self.offset
+        else:
+            return self._address
+
+    @property
+    def lo(self):
+        assert self.size % 2 == 0
+        child = LocalVariable(self.size // 2, min(self.size // 2, self.alignment))
+        child.parent = self
+        child._offset = 0
+        return child
+
+    @property
+    def hi(self):
+        assert self.size % 2 == 0
+        child = LocalVariable(self.size // 2, min(self.size // 2, self.alignment))
+        child.parent = self
+        child._offset = self.size // 2
+        return child