modelverse/kernel/modelverse_jit/jit.py

import math
import keyword
from collections import defaultdict
import modelverse_kernel.primitives as primitive_functions
import modelverse_jit.runtime as jit_runtime

# Import JitCompilationFailedException because it used to be defined
# in this module.
JitCompilationFailedException = jit_runtime.JitCompilationFailedException

def map_and_simplify_generator(function, instruction):
    """Applies the given mapping function to every instruction in the tree
       that has the given instruction as root, and simplifies it on-the-fly.

       This is at least as powerful as first mapping and then simplifying, as
       maps and simplifications are interspersed.

       This function assumes that function creates a generator that returns by
       raising a primitive_functions.PrimitiveFinished."""

    # First handle the children by mapping on them and then simplifying them.
    new_children = []
    for inst in instruction.get_children():
        new_inst, = yield [("CALL_ARGS", [map_and_simplify_generator, (function, inst)])]
        new_children.append(new_inst)

    # Then apply the function to the top-level node.
    transformed, = yield [("CALL_ARGS", [function, (instruction.create(new_children),)])]
    # Finally, simplify the transformed top-level node.
    raise primitive_functions.PrimitiveFinished(transformed.simplify_node())

def expand_constant_read(instruction):
    """Tries to replace a read of a constant node by a literal."""
    if isinstance(instruction, tree_ir.ReadValueInstruction) and \
        isinstance(instruction.node_id, tree_ir.LiteralInstruction):
        val, = yield [("RV", [instruction.node_id.literal])]
        raise primitive_functions.PrimitiveFinished(tree_ir.LiteralInstruction(val))
    else:
        raise primitive_functions.PrimitiveFinished(instruction)

def optimize_tree_ir(instruction):
    """Optimizes an IR tree."""
    return map_and_simplify_generator(expand_constant_read, instruction)

def create_bare_function(function_name, parameter_list, function_body):
    """Creates a function definition from the given function name, parameter list
       and function body. No prolog is included."""
    # Wrap the IR in a function definition, give it a unique name.
    return tree_ir.DefineFunctionInstruction(
        function_name,
        parameter_list + ['**' + jit_runtime.KWARGS_PARAMETER_NAME],
        function_body)

def create_function(
        function_name, parameter_list, param_dict,
        body_param_dict, function_body, source_map_name=None,
        compatible_temporary_protects=False):
    """Creates a function from the given function name, parameter list,
       variable-to-parameter name map, variable-to-local name map and
       function body. An optional source map can be included, too."""
    # Write a prologue and prepend it to the generated function body.
    prolog_statements = []
    # If the source map is not None, then we should generate a "DEBUG_INFO"
    # request.
    if source_map_name is not None:
        prolog_statements.append(
            tree_ir.RegisterDebugInfoInstruction(
                tree_ir.LiteralInstruction(function_name),
                tree_ir.LoadGlobalInstruction(source_map_name),
                tree_ir.LiteralInstruction(jit_runtime.BASELINE_JIT_ORIGIN_NAME)))

    # Create a LOCALS_NODE_NAME node, and connect it to the user root.
    prolog_statements.append(
        tree_ir.create_new_local_node(
            jit_runtime.LOCALS_NODE_NAME,
            tree_ir.LoadIndexInstruction(
                tree_ir.LoadLocalInstruction(jit_runtime.KWARGS_PARAMETER_NAME),
                tree_ir.LiteralInstruction('task_root')),
            jit_runtime.LOCALS_EDGE_NAME))
    for (key, val) in list(param_dict.items()):
        arg_ptr = tree_ir.create_new_local_node(
            body_param_dict[key],
            tree_ir.LoadLocalInstruction(jit_runtime.LOCALS_NODE_NAME))
        prolog_statements.append(arg_ptr)
        prolog_statements.append(
            tree_ir.CreateDictionaryEdgeInstruction(
                tree_ir.LoadLocalInstruction(body_param_dict[key]),
                tree_ir.LiteralInstruction('value'),
                tree_ir.LoadLocalInstruction(val)))

    constructed_body = tree_ir.create_block(
        *(prolog_statements + [function_body]))

    # Shield temporaries from the GC.
    constructed_body = tree_ir.protect_temporaries_from_gc(
        constructed_body,
        tree_ir.LoadLocalInstruction(jit_runtime.LOCALS_NODE_NAME),
        compatible_temporary_protects)

    return create_bare_function(function_name, parameter_list, constructed_body)

def print_value(val):
    """A thin wrapper around 'print'."""
    print(val)

class ModelverseJit(object):
    """A high-level interface to the modelverse JIT compiler."""
    def __init__(self, max_instructions=None, compiled_function_lookup=None):
        self.todo_entry_points = set()
        self.no_jit_entry_points = set()
        self.jitted_parameters = {}
        self.jit_globals = {
            'PrimitiveFinished' : primitive_functions.PrimitiveFinished,
        }
        # jitted_entry_points maps body ids to values in jit_globals.
        self.jitted_entry_points = {}
        # global_functions maps global value names to body ids.
        self.global_functions = {}
        # global_functions_inv maps body ids to global value names.
        self.global_functions_inv = {}
        # bytecode_graphs maps body ids to their parsed bytecode graphs.
        self.bytecode_graphs = {}
        # jitted_function_aliases maps body ids to known aliases.
        self.jitted_function_aliases = defaultdict(set)
        self.jit_count = 0
        self.max_instructions = max_instructions
        self.compiled_function_lookup = compiled_function_lookup
        self.compilation_dependencies = {}
        self.jit_enabled = True
        self.direct_calls_allowed = True
        self.tracing_enabled = False
        self.source_maps_enabled = True
        self.input_function_enabled = False
        self.nop_insertion_enabled = True
        self.jit_success_log_function = None
        self.jit_code_log_function = None

    def set_jit_enabled(self, is_enabled=True):
        """Enables or disables the JIT."""
        self.jit_enabled = is_enabled

    def allow_direct_calls(self, is_allowed=True):
        """Allows or disallows direct calls from jitted to jitted code."""
        self.direct_calls_allowed = is_allowed

    def use_input_function(self, is_enabled=True):
        """Configures the JIT to compile 'input' instructions as function calls."""
        self.input_function_enabled = is_enabled

    def enable_tracing(self, is_enabled=True):
        """Enables or disables tracing for jitted code."""
        self.tracing_enabled = is_enabled

    def enable_source_maps(self, is_enabled=True):
        """Enables or disables the creation of source maps for jitted code. Source maps
           convert lines in the generated code to debug information.

           Source maps are enabled by default."""
        self.source_maps_enabled = is_enabled

    def enable_nop_insertion(self, is_enabled=True):
        """Enables or disables nop insertion for jitted code. If enabled, the JIT will
           insert nops at loop back-edges. Inserting nops sacrifices performance to
           keep the jitted code from blocking the thread of execution and consuming
           all resources; nops give the Modelverse server an opportunity to interrupt
           the currently running code."""
        self.nop_insertion_enabled = is_enabled

    def set_jit_success_log(self, log_function=print_value):
        """Configures this JIT instance with a function that prints output to a log.
           Success and failure messages for specific functions are then sent to said log."""
        self.jit_success_log_function = log_function

    def set_jit_code_log(self, log_function=print_value):
        """Configures this JIT instance with a function that prints output to a log.
           Function definitions of jitted functions are then sent to said log."""
        self.jit_code_log_function = log_function

    def set_function_body_compiler(self, compile_function_body):
        """Sets the function that the JIT uses to compile function bodies."""
        self.compile_function_body = compile_function_body

    def mark_entry_point(self, body_id):
        """Marks the node with the given identifier as a function entry point."""
        if body_id not in self.no_jit_entry_points and body_id not in self.jitted_entry_points:
            self.todo_entry_points.add(body_id)

    def is_entry_point(self, body_id):
        """Tells if the node with the given identifier is a function entry point."""
        return body_id in self.todo_entry_points or \
               body_id in self.no_jit_entry_points or \
               body_id in self.jitted_entry_points

    def is_jittable_entry_point(self, body_id):
        """Tells if the node with the given identifier is a function entry point that
           has not been marked as non-jittable. This only returns `True` if the JIT
           is enabled and the function entry point has been marked jittable, or if
           the function has already been compiled."""
        return ((self.jit_enabled and body_id in self.todo_entry_points) or
                self.has_compiled(body_id))

    def has_compiled(self, body_id):
        """Tests if the function belonging to the given body node has been compiled yet."""
        return body_id in self.jitted_entry_points

    def get_compiled_name(self, body_id):
        """Gets the name of the compiled version of the given body node in the JIT
           global state."""
        if body_id in self.jitted_entry_points:
            return self.jitted_entry_points[body_id]
        else:
            return None

    def mark_no_jit(self, body_id):
        """Informs the JIT that the node with the given identifier is a function entry
           point that must never be jitted."""
        self.no_jit_entry_points.add(body_id)
        if body_id in self.todo_entry_points:
            self.todo_entry_points.remove(body_id)

    def generate_name(self, infix, suggested_name=None):
        """Generates a new name or picks the suggested name if it is still
           available."""
        if suggested_name is not None \
            and suggested_name not in self.jit_globals \
            and not keyword.iskeyword(suggested_name):
            self.jit_count += 1
            return suggested_name
        else:
            function_name = 'jit_%s%d' % (infix, self.jit_count)
            self.jit_count += 1
            return function_name

    def generate_function_name(self, body_id, suggested_name=None):
        """Generates a new function name or picks the suggested name if it is still
           available."""
        if suggested_name is None:
            suggested_name = self.get_global_name(body_id)

        return self.generate_name('func', suggested_name)

    def register_global(self, body_id, global_name):
        """Associates the given body id with the given global name."""
        self.global_functions[global_name] = body_id
        self.global_functions_inv[body_id] = global_name

    def get_global_name(self, body_id):
        """Gets the name of the global function with the given body id.
           Returns None if no known global exists with the given id."""
        if body_id in self.global_functions_inv:
            return self.global_functions_inv[body_id]
        else:
            return None

    def get_global_body_id(self, global_name):
        """Gets the body id of the global function with the given name.
           Returns None if no known global exists with the given name."""
        if global_name in self.global_functions:
            return self.global_functions[global_name]
        else:
            return None

    def register_compiled(self, body_id, compiled_function, function_name=None):
        """Registers a compiled entry point with the JIT."""
        # Get the function's name.
        actual_function_name = self.generate_function_name(body_id,
        function_name)
        # Map the body id to the given parameter list.
        self.jitted_entry_points[body_id] = actual_function_name
        self.jit_globals[actual_function_name] = compiled_function
        if function_name is not None:
            self.register_global(body_id, function_name)

        if body_id in self.todo_entry_points:
            self.todo_entry_points.remove(body_id)

    def import_value(self, value, suggested_name=None):
        """Imports the given value into the JIT's global scope, with the given suggested name.
           The actual name of the value (within the JIT's global scope) is returned."""
        actual_name = self.generate_name('import', suggested_name)
        self.jit_globals[actual_name] = value
        return actual_name

    def __lookup_compiled_body_impl(self, body_id):
        """Looks up a compiled function by body id. Returns a matching function,
           or None if no function was found."""
        if body_id is not None and body_id in self.jitted_entry_points:
            return self.jit_globals[self.jitted_entry_points[body_id]]
        else:
            return None

    def __lookup_external_body_impl(self, global_name, body_id):
        """Looks up an external function by global name. Returns a matching function,
           or None if no function was found."""
        if global_name is not None and self.compiled_function_lookup is not None:
            result = self.compiled_function_lookup(global_name)
            if result is not None and body_id is not None:
                self.register_compiled(body_id, result, global_name)

            return result
        else:
            return None

    def lookup_compiled_body(self, body_id):
        """Looks up a compiled function by body id. Returns a matching function,
           or None if no function was found."""
        result = self.__lookup_compiled_body_impl(body_id)
        if result is not None:
            return result
        else:
            global_name = self.get_global_name(body_id)
            return self.__lookup_external_body_impl(global_name, body_id)

    def lookup_compiled_function(self, global_name):
        """Looks up a compiled function by global name. Returns a matching function,
           or None if no function was found."""
        body_id = self.get_global_body_id(global_name)
        result = self.__lookup_compiled_body_impl(body_id)
        if result is not None:
            return result
        else:
            return self.__lookup_external_body_impl(global_name, body_id)

    def jit_signature(self, body_id):
        """Acquires the signature for the given body id node, which consists of the
           parameter variables, parameter name and a flag that tells if the given function
           is mutable."""
        if body_id not in self.jitted_parameters:
            signature_id, = yield [("RRD", [body_id, jit_runtime.FUNCTION_BODY_KEY])]
            signature_id = signature_id[0]
            param_set_id, is_mutable = yield [
                ("RD", [signature_id, "params"]),
                ("RD", [signature_id, jit_runtime.MUTABLE_FUNCTION_KEY])]
            if param_set_id is None:
                self.jitted_parameters[body_id] = ([], [], is_mutable)
            else:
                param_name_ids, = yield [("RDK", [param_set_id])]
                param_names = yield [("RV", [n]) for n in param_name_ids]
                #NOTE Patch up strange links...
                param_names = [i for i in param_names if i is not None]
                param_vars = yield [("RD", [param_set_id, k]) for k in param_names]

                #NOTE that variables might not be in the correct order, as we just read them out!
                lst = sorted([(name, var) for name, var in zip(param_names, param_vars)])
                param_vars = [i[1] for i in lst]
                param_names = [i[0] for i in lst]
                self.jitted_parameters[body_id] = (param_vars, param_names, is_mutable)

        raise primitive_functions.PrimitiveFinished(self.jitted_parameters[body_id])

    def jit_parse_bytecode(self, body_id):
        """Parses the given function body as a bytecode graph."""
        if body_id in self.bytecode_graphs:
            raise primitive_functions.PrimitiveFinished(self.bytecode_graphs[body_id])

        parser = bytecode_parser.BytecodeParser()
        result, = yield [("CALL_ARGS", [parser.parse_instruction, (body_id,)])]
        self.bytecode_graphs[body_id] = result
        raise primitive_functions.PrimitiveFinished(result)

    def check_jittable(self, body_id, suggested_name=None):
        """Checks if the function with the given body id is obviously non-jittable. If it's
           non-jittable, then a `JitCompilationFailedException` exception is thrown."""
        if body_id is None:
            raise ValueError('body_id cannot be None: ' + suggested_name)
        elif body_id in self.no_jit_entry_points:
            # We're not allowed to jit this function or have tried and failed before.
            raise JitCompilationFailedException(
                'Cannot jit function %s at %d because it is marked non-jittable.' % (
                    '' if suggested_name is None else "'" + suggested_name + "'",
                    body_id))
        elif not self.jit_enabled:
            # We're not allowed to jit anything.
            raise JitCompilationFailedException(
                'Cannot jit function %s at %d because the JIT has been disabled.' % (
                    '' if suggested_name is None else "'" + suggested_name + "'",
                    body_id))

    def jit_recompile(self, task_root, body_id, function_name, compile_function_body=None):
        """Replaces the function with the given name by compiling the bytecode at the given
           body id."""
        if compile_function_body is None:
            compile_function_body = self.compile_function_body

        self.check_jittable(body_id, function_name)

        # Generate a name for the function we're about to analyze, and pretend that
        # it already exists. (we need to do this for recursive functions)
        self.jitted_entry_points[body_id] = function_name
        self.jit_globals[function_name] = None

        (_, _, is_mutable), = yield [
            ("CALL_ARGS", [self.jit_signature, (body_id,)])]

        dependencies = set([body_id])
        self.compilation_dependencies[body_id] = dependencies

        def handle_jit_exception(exception):
            # If analysis fails, then a JitCompilationFailedException will be thrown.
            print("EXCEPTION with mutable")
            del self.compilation_dependencies[body_id]
            for dep in dependencies:
                self.mark_no_jit(dep)
                if dep in self.jitted_entry_points:
                    del self.jitted_entry_points[dep]

            failure_message = "%s (function '%s' at %d)" % (
                str(exception), function_name, body_id)
            if self.jit_success_log_function is not None:
                self.jit_success_log_function('JIT compilation failed: %s' % failure_message)
            raise JitCompilationFailedException(failure_message)

        # Try to analyze the function's body.
        yield [("TRY", [])]
        yield [("CATCH", [JitCompilationFailedException, handle_jit_exception])]
        if is_mutable:
            # We can't just JIT mutable functions. That'd be dangerous.
            raise JitCompilationFailedException(
                "Function was marked '%s'." % jit_runtime.MUTABLE_FUNCTION_KEY)

        compiled_function, = yield [
            ("CALL_ARGS", [compile_function_body, (self, function_name, body_id, task_root)])]

        yield [("END_TRY", [])]
        del self.compilation_dependencies[body_id]

        if self.jit_success_log_function is not None:
            assert self.jitted_entry_points[body_id] == function_name
            self.jit_success_log_function(
                "JIT compilation successful: (function '%s' at %d)" % (function_name, body_id))

        raise primitive_functions.PrimitiveFinished(compiled_function)

    def get_source_map_name(self, function_name):
        """Gets the name of the given jitted function's source map. None is returned if source maps
           are disabled."""
        if self.source_maps_enabled:
            return function_name + "_source_map"
        else:
            return None

    def get_can_rejit_name(self, function_name):
        """Gets the name of the given jitted function's can-rejit flag."""
        return function_name + "_can_rejit"

    def jit_define_function(self, function_name, function_def):
        """Converts the given tree-IR function definition to Python code, defines it,
           and extracts the resulting function."""
        # The comment below makes pylint shut up about our (hopefully benign) use of exec here.
        # pylint: disable=I0011,W0122
        if self.jit_code_log_function is not None:
            self.jit_code_log_function(function_def)

        # Convert the function definition to Python code, and compile it.
        code_generator = tree_ir.PythonGenerator()
        function_def.generate_python_def(code_generator)
        source_map_name = self.get_source_map_name(function_name)
        if source_map_name is not None:
            self.jit_globals[source_map_name] = code_generator.source_map_builder.source_map
        exec(str(code_generator), self.jit_globals)

        # Extract the compiled function from the JIT global state.
        return self.jit_globals[function_name]

    def jit_delete_function(self, function_name):
        """Deletes the function with the given function name."""
        del self.jit_globals[function_name]

    def jit_compile(self, task_root, body_id, suggested_name=None):
        """Tries to jit the function defined by the given entry point id and parameter list."""
        if body_id is None:
            raise ValueError('body_id cannot be None: ' + str(suggested_name))
        elif body_id in self.jitted_entry_points:
            raise primitive_functions.PrimitiveFinished(
                self.jit_globals[self.jitted_entry_points[body_id]])

        compiled_func = self.lookup_compiled_body(body_id)
        if compiled_func is not None:
            raise primitive_functions.PrimitiveFinished(compiled_func)

        # Generate a name for the function we're about to analyze, and 're-compile'
        # it for the first time.
        function_name = self.generate_function_name(body_id, suggested_name)
        yield [("TAIL_CALL_ARGS", [self.jit_recompile, (task_root, body_id, function_name)])]

    def jit_rejit(self, task_root, body_id, function_name, compile_function_body=None):
        """Re-compiles the given function. If compilation fails, then the can-rejit
           flag is set to false."""
        old_jitted_func = self.jitted_entry_points[body_id]
        def __handle_jit_failed(_):
            self.jit_globals[self.get_can_rejit_name(function_name)] = False
            self.jitted_entry_points[body_id] = old_jitted_func
            self.no_jit_entry_points.remove(body_id)
            raise primitive_functions.PrimitiveFinished(None)

        yield [("TRY", [])]
        yield [("CATCH", [jit_runtime.JitCompilationFailedException, __handle_jit_failed])]
        jitted_function, = yield [
            ("CALL_ARGS",
             [self.jit_recompile, (task_root, body_id, function_name, compile_function_body)])]
        yield [("END_TRY", [])]

        # Update all aliases.
        for function_alias in self.jitted_function_aliases[body_id]:
            self.jit_globals[function_alias] = jitted_function

    def new_compile(self, body_id):
        print("Compiling body ID " + str(body_id))
        raise JitCompilationFailedException("Function was marked '%s'." % jit_runtime.MUTABLE_FUNCTION_KEY)

        #raise primitive_functions.PrimitiveFinished("pass")