artiq/artiq/compiler/transforms/artiq_ir_generator.py

"""
:class:`ARTIQIRGenerator` transforms typed AST into ARTIQ intermediate
representation. ARTIQ IR is designed to be low-level enough that
its operations are elementary--contain no internal branching--
but without too much detail, such as exposing the reference/value
semantics explicitly.
"""

from collections import OrderedDict, defaultdict
from functools import reduce
from itertools import chain
from pythonparser import algorithm, diagnostic, ast
from .. import types, builtins, asttyped, ir, iodelay

def _readable_name(insn):
    if isinstance(insn, ir.Constant):
        return str(insn.value)
    else:
        return insn.name

def _extract_loc(node):
    if "keyword_loc" in node._locs:
        return node.keyword_loc
    else:
        return node.loc

# We put some effort in keeping generated IR readable,
# i.e. with a more or less linear correspondence to the source.
# This is why basic blocks sometimes seem to be produced in an odd order.
class ARTIQIRGenerator(algorithm.Visitor):
    """
    :class:`ARTIQIRGenerator` contains a lot of internal state,
    which is effectively maintained in a stack--with push/pop
    pairs around any state updates. It is comprised of following:

    :ivar current_loc: (:class:`pythonparser.source.Range`)
        source range of the node being currently visited
    :ivar current_function: (:class:`ir.Function` or None)
        module, def or lambda currently being translated
    :ivar current_globals: (set of string)
        set of variables that will be resolved in global scope
    :ivar current_block: (:class:`ir.BasicBlock`)
        basic block to which any new instruction will be appended
    :ivar current_env: (:class:`ir.Alloc` of type :class:`ir.TEnvironment`)
        the chained function environment, containing variables that
        can become upvalues
    :ivar current_private_env: (:class:`ir.Alloc` of type :class:`ir.TEnvironment`)
        the private function environment, containing internal state
    :ivar current_args: (dict of string to :class:`ir.Argument`)
        the map of Python names of formal arguments to
        the current function to their SSA names
    :ivar current_assign: (:class:`ir.Value` or None)
        the right-hand side of current assignment statement, or
        a component of a composite right-hand side when visiting
        a composite left-hand side, such as, in ``x, y = z``,
        the 2nd tuple element when visting ``y``
    :ivar break_target: (:class:`ir.BasicBlock` or None)
        the basic block to which ``break`` will transfer control
    :ivar continue_target: (:class:`ir.BasicBlock` or None)
        the basic block to which ``continue`` will transfer control
    :ivar return_target: (:class:`ir.BasicBlock` or None)
        the basic block to which ``return`` will transfer control
    :ivar unwind_target: (:class:`ir.BasicBlock` or None)
        the basic block to which unwinding will transfer control
    :ivar catch_clauses: (list of (:class:`ir.BasicBlock`, :class:`types.Type` or None))
        a list of catch clauses that should be appended to inner try block
        landingpad
    :ivar final_branch: (function (target: :class:`ir.BasicBlock`, block: :class:`ir.BasicBlock)
                         or None)
        the function that appends to ``block`` a jump through the ``finally`` statement
        to ``target``

    There is, additionally, some global state that is used to translate
    the results of analyses on AST level to IR level:

    :ivar function_map: (map of :class:`ast.FunctionDefT` to :class:`ir.Function`)
        the map from function definition nodes to IR functions
    :ivar variable_map: (map of :class:`ast.NameT` to :class:`ir.GetLocal`)
        the map from variable name nodes to instructions retrieving
        the variable values
    :ivar method_map: (map of :class:`ast.AttributeT` to :class:`ir.GetAttribute`)
        the map from method resolution nodes to instructions retrieving
        the called function inside a translated :class:`ast.CallT` node

    Finally, functions that implement array operations are instantiated on the fly as
    necessary. They are kept track of in global dictionaries, with a mangled name
    containing types and operations as key:

    :ivar array_op_funcs: the map from mangled name to implementation of
        operations on/between arrays
    """

    _size_type = builtins.TInt32()

    def __init__(self, module_name, engine, ref_period, embedding_map):
        self.engine = engine
        self.embedding_map = embedding_map
        self.functions = []
        self.name = [module_name] if module_name != "" else []
        self.ref_period = ir.Constant(ref_period, builtins.TFloat())
        self.current_loc = None
        self.current_function = None
        self.current_class = None
        self.current_globals = set()
        self.current_block = None
        self.current_env = None
        self.current_private_env = None
        self.current_args = None
        self.current_assign = None
        self.current_exception = None
        self.current_remote_fn = False
        self.break_target = None
        self.continue_target = None
        self.return_target = None
        self.unwind_target = None
        self.catch_clauses = []
        self.final_branch = None
        self.function_map = dict()
        self.variable_map = dict()
        self.method_map = defaultdict(lambda: [])
        self.array_op_funcs = dict()
        self.raise_assert_func = None

    def annotate_calls(self, devirtualization):
        for var_node in devirtualization.variable_map:
            callee_node = devirtualization.variable_map[var_node]
            if callee_node is None:
                continue
            callee = self.function_map[callee_node]

            call_target = self.variable_map[var_node]
            for use in call_target.uses:
                if isinstance(use, (ir.Call, ir.Invoke)) and \
                        use.target_function() == call_target:
                    use.static_target_function = callee

        for type_and_method in devirtualization.method_map:
            callee_node = devirtualization.method_map[type_and_method]
            if callee_node is None:
                continue
            callee = self.function_map[callee_node]

            for call in self.method_map[type_and_method]:
                assert isinstance(call, (ir.Call, ir.Invoke))
                call.static_target_function = callee

    def add_block(self, name=""):
        block = ir.BasicBlock([], name)
        self.current_function.add(block)
        return block

    def append(self, insn, block=None, loc=None):
        if loc is None:
            loc = self.current_loc
        if block is None:
            block = self.current_block

        if insn.loc is None:
            insn.loc = loc
        return block.append(insn)

    def terminate(self, insn):
        if not self.current_block.is_terminated():
            self.append(insn)
        else:
            insn.drop_references()

    def warn_unreachable(self, node):
        diag = diagnostic.Diagnostic("warning",
            "unreachable code", {},
            node.loc.begin())
        self.engine.process(diag)

    # Visitors

    def visit(self, obj):
        if isinstance(obj, list):
            for elt in obj:
                if self.current_block.is_terminated():
                    self.warn_unreachable(elt)
                    break
                self.visit(elt)
        elif isinstance(obj, ast.AST):
            try:
                old_loc, self.current_loc = self.current_loc, _extract_loc(obj)
                return self._visit_one(obj)
            finally:
                self.current_loc = old_loc

    # Module visitor

    def visit_ModuleT(self, node):
        # Treat start of module as synthesized
        self.current_loc = None

        try:
            typ = types.TFunction(OrderedDict(), OrderedDict(), builtins.TNone())
            func = ir.Function(typ, ".".join(self.name + ['__modinit__']), [],
                               loc=node.loc.begin())
            self.functions.append(func)
            old_func, self.current_function = self.current_function, func

            entry = self.add_block("entry")
            old_block, self.current_block = self.current_block, entry

            env_type = ir.TEnvironment(name=func.name, vars=node.typing_env)
            env = self.append(ir.Alloc([], env_type, name="env"))
            old_env, self.current_env = self.current_env, env

            priv_env_type = ir.TEnvironment(name=func.name + ".priv", vars={ "$return": typ.ret })
            priv_env = self.append(ir.Alloc([], priv_env_type, name="privenv"))
            old_priv_env, self.current_private_env = self.current_private_env, priv_env

            self.generic_visit(node)
            self.terminate(ir.Return(ir.Constant(None, builtins.TNone()),
                           remote_return=self.current_remote_fn))

            return self.functions
        finally:
            self.current_function = old_func
            self.current_block = old_block
            self.current_env = old_env
            self.current_private_env = old_priv_env

    # Statement visitors

    def visit_ClassDefT(self, node):
        klass = self.append(ir.Alloc([], node.constructor_type,
                                     name="class.{}".format(node.name)))
        self._set_local(node.name, klass)

        try:
            old_class, self.current_class = self.current_class, klass
            self.visit(node.body)
        finally:
            self.current_class = old_class

    def visit_function(self, node, is_lambda=False, is_internal=False, is_quoted=False,
                       flags={}):
        if is_lambda:
            name = "lambda@{}:{}".format(node.loc.line(), node.loc.column())
            typ = node.type.find()
        else:
            name = node.name
            typ = node.signature_type.find()

        try:
            defaults = []
            if not is_quoted:
                for arg_name, default_node in zip(typ.optargs, node.args.defaults):
                    default = self.visit(default_node)
                    env_default_name = \
                        self.current_env.type.add("$default." + arg_name, default.type)
                    self.append(ir.SetLocal(self.current_env, env_default_name, default))
                    def codegen_default(env_default_name):
                        return lambda: self.append(ir.GetLocal(self.current_env, env_default_name))
                    defaults.append(codegen_default(env_default_name))
            else:
                for default_node in node.args.defaults:
                    def codegen_default(default_node):
                        return lambda: self.visit(default_node)
                    defaults.append(codegen_default(default_node))

            old_name, self.name = self.name, self.name + [name]

            env_arg  = ir.EnvironmentArgument(self.current_env.type, "ARG.ENV")

            old_args, self.current_args = self.current_args, {}

            args = []
            for arg_name in typ.args:
                arg = ir.Argument(typ.args[arg_name], "ARG." + arg_name)
                self.current_args[arg_name] = arg
                args.append(arg)

            optargs = []
            for arg_name in typ.optargs:
                arg = ir.Argument(ir.TOption(typ.optargs[arg_name]), "ARG." + arg_name)
                self.current_args[arg_name] = arg
                optargs.append(arg)

            for (arg, arg_node) in zip(args + optargs, node.args.args):
                arg.loc = arg_node.loc

            func = ir.Function(typ, ".".join(self.name), [env_arg] + args + optargs,
                               loc=node.lambda_loc if is_lambda else node.keyword_loc)
            func.is_internal = is_internal
            func.flags = flags
            self.functions.append(func)
            old_func, self.current_function = self.current_function, func

            if not is_lambda:
                self.function_map[node] = func

            entry = self.add_block("entry")
            old_block, self.current_block = self.current_block, entry

            old_globals, self.current_globals = self.current_globals, node.globals_in_scope
            old_remote_fn = self.current_remote_fn
            self.current_remote_fn = getattr(node, "remote_fn", False)

            env_without_globals = \
                {var: node.typing_env[var]
                 for var in node.typing_env
                  if var not in node.globals_in_scope}
            env_type = ir.TEnvironment(name=func.name,
                                       vars=env_without_globals, outer=self.current_env.type)
            env = self.append(ir.Alloc([], env_type, name="ENV"))
            old_env, self.current_env = self.current_env, env

            if not is_lambda:
                priv_env_type = ir.TEnvironment(name="{}.private".format(func.name),
                                                vars={ "$return": typ.ret })
                priv_env = self.append(ir.Alloc([], priv_env_type, name="PRV"))
                old_priv_env, self.current_private_env = self.current_private_env, priv_env

            self.append(ir.SetLocal(env, "$outer", env_arg))
            for index, arg_name in enumerate(typ.args):
                self.append(ir.SetLocal(env, arg_name, args[index]))
            for index, (arg_name, codegen_default) in enumerate(zip(typ.optargs, defaults)):
                default = codegen_default()
                value = self.append(ir.Builtin("unwrap_or", [optargs[index], default],
                                               typ.optargs[arg_name],
                                               name="DEF.{}".format(arg_name)))
                self.append(ir.SetLocal(env, arg_name, value))

            result = self.visit(node.body)

            if is_lambda:
                self.terminate(ir.Return(result))
            elif builtins.is_none(typ.ret):
                if not self.current_block.is_terminated():
                    self.current_block.append(ir.Return(ir.Constant(None, builtins.TNone()), 
                                                        remote_return=self.current_remote_fn))
            else:
                if not self.current_block.is_terminated():
                    if len(self.current_block.predecessors()) != 0:
                        diag = diagnostic.Diagnostic("error",
                            "this function must return a value of type {typ} explicitly",
                            {"typ": types.TypePrinter().name(typ.ret)},
                            node.keyword_loc)
                        self.engine.process(diag)

                    self.current_block.append(ir.Unreachable())

        finally:
            self.name = old_name
            self.current_args = old_args
            self.current_function = old_func
            self.current_block = old_block
            self.current_globals = old_globals
            self.current_env = old_env
            self.current_remote_fn = old_remote_fn
            if not is_lambda:
                self.current_private_env = old_priv_env

        return self.append(ir.Closure(func, self.current_env))

    def visit_FunctionDefT(self, node):
        func = self.visit_function(node, is_internal=len(self.name) > 0)
        if self.current_class is None:
            self._set_local(node.name, func)
        else:
            self.append(ir.SetAttr(self.current_class, node.name, func))

    def visit_QuotedFunctionDefT(self, node):
        self.visit_function(node, is_internal=True, is_quoted=True, flags=node.flags)

    def visit_Return(self, node):
        if node.value is None:
            return_value = ir.Constant(None, builtins.TNone())
        else:
            return_value = self.visit(node.value)

        if self.return_target is None:
            self.append(ir.Return(return_value, 
                                  remote_return=self.current_remote_fn))
        else:
            self.append(ir.SetLocal(self.current_private_env, "$return", return_value))
            self.append(ir.Branch(self.return_target))

    def visit_Expr(self, node):
        # Ignore the value, do it for side effects.
        result = self.visit(node.value)

        # See comment in visit_Pass.
        if isinstance(result, ir.Constant):
            self.visit_Pass(node)

    def visit_Pass(self, node):
        # Insert a dummy instruction so that analyses which extract
        # locations from CFG have something to use.
        self.append(ir.Builtin("nop", [], builtins.TNone()))

    def visit_Assign(self, node):
        try:
            self.current_assign = self.visit(node.value)
            assert self.current_assign is not None
            for target in node.targets:
                self.visit(target)
        finally:
            self.current_assign = None

    def visit_AugAssign(self, node):
        lhs = self.visit(node.target)
        rhs = self.visit(node.value)

        if builtins.is_array(lhs.type):
            name = type(node.op).__name__
            def make_op(l, r):
                return self.append(ir.Arith(node.op, l, r))
            self._broadcast_binop(name, make_op, lhs.type, lhs, rhs, assign_to_lhs=True)
            return

        value = self.append(ir.Arith(node.op, lhs, rhs))
        try:
            self.current_assign = value
            self.visit(node.target)
        finally:
            self.current_assign = None

    def coerce_to_bool(self, insn, block=None):
        if builtins.is_bool(insn.type):
            return insn
        elif builtins.is_int(insn.type):
            return self.append(ir.Compare(ast.NotEq(loc=None), insn, ir.Constant(0, insn.type)),
                               block=block)
        elif builtins.is_float(insn.type):
            return self.append(ir.Compare(ast.NotEq(loc=None), insn, ir.Constant(0, insn.type)),
                               block=block)
        elif builtins.is_iterable(insn.type):
            length = self.iterable_len(insn)
            return self.append(ir.Compare(ast.NotEq(loc=None), length, ir.Constant(0, length.type)),
                               block=block)
        elif builtins.is_none(insn.type):
            return ir.Constant(False, builtins.TBool())
        else:
            note = diagnostic.Diagnostic("note",
                "this expression has type {type}",
                {"type": types.TypePrinter().name(insn.type)},
                insn.loc)
            diag = diagnostic.Diagnostic("warning",
                "this expression, which is always truthful, is coerced to bool", {},
                insn.loc, notes=[note])
            self.engine.process(diag)
            return ir.Constant(True, builtins.TBool())

    def visit_If(self, node):
        cond = self.visit(node.test)
        cond = self.coerce_to_bool(cond)
        head = self.current_block

        if_true = self.add_block("if.body")
        self.current_block = if_true
        self.visit(node.body)
        post_if_true = self.current_block

        if any(node.orelse):
            if_false = self.add_block("if.else")
            self.current_block = if_false
            self.visit(node.orelse)
            post_if_false = self.current_block

        tail = self.add_block("if.tail")
        self.current_block = tail
        if not post_if_true.is_terminated():
            post_if_true.append(ir.Branch(tail))

        if any(node.orelse):
            if not post_if_false.is_terminated():
                post_if_false.append(ir.Branch(tail))
            self.append(ir.BranchIf(cond, if_true, if_false), block=head)
        else:
            self.append(ir.BranchIf(cond, if_true, tail), block=head)

    def visit_While(self, node):
        try:
            head = self.add_block("while.head")
            self.append(ir.Branch(head))
            self.current_block = head
            old_continue, self.continue_target = self.continue_target, head
            cond = self.visit(node.test)
            cond = self.coerce_to_bool(cond)
            post_head = self.current_block

            break_block = self.add_block("while.break")
            old_break, self.break_target = self.break_target, break_block

            body = self.add_block("while.body")
            self.current_block = body
            self.visit(node.body)
            post_body = self.current_block
        finally:
            self.break_target = old_break
            self.continue_target = old_continue

        if any(node.orelse):
            else_tail = self.add_block("while.else")
            self.current_block = else_tail
            self.visit(node.orelse)
            post_else_tail = self.current_block

        tail = self.add_block("while.tail")
        self.current_block = tail

        if any(node.orelse):
            if not post_else_tail.is_terminated():
                post_else_tail.append(ir.Branch(tail))
        else:
            else_tail = tail

        post_head.append(ir.BranchIf(cond, body, else_tail))
        if not post_body.is_terminated():
            post_body.append(ir.Branch(head))
        break_block.append(ir.Branch(tail))

    def iterable_len(self, value, typ=_size_type):
        if builtins.is_listish(value.type):
            if isinstance(value, ir.Constant):
                name = None
            else:
                name = "{}.len".format(value.name)
            len = self.append(ir.Builtin("len", [value], builtins.TInt32(),
                                         name=name))
            return self.append(ir.Coerce(len, typ))
        elif builtins.is_range(value.type):
            start  = self.append(ir.GetAttr(value, "start"))
            stop   = self.append(ir.GetAttr(value, "stop"))
            step   = self.append(ir.GetAttr(value, "step"))
            spread = self.append(ir.Arith(ast.Sub(loc=None), stop, start))
            return self.append(ir.Arith(ast.FloorDiv(loc=None), spread, step,
                                        name="{}.len".format(value.name)))
        else:
            assert False

    def iterable_get(self, value, index):
        # Assuming the value is within bounds.
        if builtins.is_array(value.type):
            # Scalar indexing into ndarray.
            num_dims = value.type.find()["num_dims"].value
            if num_dims > 1:
                old_shape = self.append(ir.GetAttr(value, "shape"))
                lengths = [self.append(ir.GetAttr(old_shape, i)) for i in range(1, num_dims)]
                new_shape = self._make_array_shape(lengths)

                stride = reduce(
                    lambda l, r: self.append(ir.Arith(ast.Mult(loc=None), l, r)),
                    lengths[1:], lengths[0])
                offset = self.append(ir.Arith(ast.Mult(loc=None), stride, index))
                old_buffer = self.append(ir.GetAttr(value, "buffer"))
                new_buffer = self.append(ir.Offset(old_buffer, offset))

                result_type = builtins.TArray(value.type.find()["elt"],
                    types.TValue(num_dims - 1))
                return self.append(ir.Alloc([new_buffer, new_shape], result_type))
            else:
                buffer = self.append(ir.GetAttr(value, "buffer"))
                return self.append(ir.GetElem(buffer, index))
        elif builtins.is_listish(value.type):
            return self.append(ir.GetElem(value, index))
        elif builtins.is_range(value.type):
            start  = self.append(ir.GetAttr(value, "start"))
            step   = self.append(ir.GetAttr(value, "step"))
            offset = self.append(ir.Arith(ast.Mult(loc=None), step, index))
            return self.append(ir.Arith(ast.Add(loc=None), start, offset))
        else:
            assert False

    def visit_ForT(self, node):
        try:
            iterable = self.visit(node.iter)
            length = self.iterable_len(iterable)
            prehead = self.current_block

            head = self.add_block("for.head")
            self.append(ir.Branch(head))
            self.current_block = head
            phi = self.append(ir.Phi(length.type, name="IND"))
            phi.add_incoming(ir.Constant(0, phi.type), prehead)
            cond = self.append(ir.Compare(ast.Lt(loc=None), phi, length, name="CMP"))

            break_block = self.add_block("for.break")
            old_break, self.break_target = self.break_target, break_block

            continue_block = self.add_block("for.continue")
            old_continue, self.continue_target = self.continue_target, continue_block
            self.current_block = continue_block

            updated_index = self.append(ir.Arith(ast.Add(loc=None), phi, ir.Constant(1, phi.type),
                                                 name="IND.new"))
            phi.add_incoming(updated_index, continue_block)
            self.append(ir.Branch(head))

            body = self.add_block("for.body")
            self.current_block = body
            elt = self.iterable_get(iterable, phi)
            try:
                self.current_assign = elt
                self.visit(node.target)
            finally:
                self.current_assign = None
            self.visit(node.body)
            post_body = self.current_block
        finally:
            self.break_target = old_break
            self.continue_target = old_continue

        if any(node.orelse):
            else_tail = self.add_block("for.else")
            self.current_block = else_tail
            self.visit(node.orelse)
            post_else_tail = self.current_block

        tail = self.add_block("for.tail")
        self.current_block = tail

        if any(node.orelse):
            if not post_else_tail.is_terminated():
                post_else_tail.append(ir.Branch(tail))
        else:
            else_tail = tail

        if node.trip_count is not None:
            head.append(ir.Loop(node.trip_count, phi, cond, body, else_tail))
        else:
            head.append(ir.BranchIf(cond, body, else_tail))
        if not post_body.is_terminated():
            post_body.append(ir.Branch(continue_block))
        break_block.append(ir.Branch(tail))

    def visit_Break(self, node):
        self.append(ir.Branch(self.break_target))

    def visit_Continue(self, node):
        self.append(ir.Branch(self.continue_target))

    def raise_exn(self, exn=None, loc=None):
        if self.final_branch is not None:
            raise_proxy = self.add_block("try.raise")
            self.final_branch(raise_proxy, self.current_block)
            self.current_block = raise_proxy

        if exn is not None:
            # if we need to raise the exception in a final body, we have to
            # lazy-evaluate the exception object to make sure that we generate
            # it in the raise_proxy block
            exn = exn()
        if exn is not None:
            assert loc is not None
            loc_file = ir.Constant(loc.source_buffer.name, builtins.TStr())
            loc_line = ir.Constant(loc.line(), builtins.TInt32())
            loc_column = ir.Constant(loc.column(), builtins.TInt32())
            loc_function = ir.Constant(".".join(self.name), builtins.TStr())

            self.append(ir.SetAttr(exn, "#__file__", loc_file))
            self.append(ir.SetAttr(exn, "#__line__", loc_line))
            self.append(ir.SetAttr(exn, "#__col__", loc_column))
            self.append(ir.SetAttr(exn, "#__func__", loc_function))

            if self.unwind_target is not None:
                self.append(ir.Raise(exn, self.unwind_target))
            else:
                self.append(ir.Raise(exn))
        else:
            if self.unwind_target is not None:
                self.append(ir.Resume(self.unwind_target))
            else:
                self.append(ir.Resume())

    def visit_Raise(self, node):
        if node.exc is not None and types.is_exn_constructor(node.exc.type):
            self.raise_exn(lambda: self.alloc_exn(node.exc.type.instance), loc=self.current_loc)
        else:
            self.raise_exn(lambda: self.visit(node.exc), loc=self.current_loc)

    def visit_Try(self, node):
        dispatcher = self.add_block("try.dispatch")
        cleanup = self.add_block('handler.cleanup')
        landingpad = ir.LandingPad(cleanup)
        dispatcher.append(landingpad)

        if any(node.finalbody):
            # k for continuation
            final_suffix   = ".try@{}:{}".format(node.loc.line(), node.loc.column())
            final_env_type = ir.TEnvironment(name=self.current_function.name + final_suffix,
                                             vars={ "$cont": ir.TBasicBlock() })
            final_state    = self.append(ir.Alloc([], final_env_type))
            final_targets  = []
            final_paths    = []

            def final_branch(target, block):
                block.append(ir.SetLocal(final_state, "$cont", target))
                final_targets.append(target)
                final_paths.append(block)

            if self.break_target is not None:
                break_proxy = self.add_block("try.break")
                old_break, self.break_target = self.break_target, break_proxy
                final_branch(old_break, break_proxy)

            if self.continue_target is not None:
                continue_proxy = self.add_block("try.continue")
                old_continue, self.continue_target = self.continue_target, continue_proxy
                final_branch(old_continue, continue_proxy)

            return_proxy = self.add_block("try.return")
            old_return, self.return_target = self.return_target, return_proxy
            if old_return is not None:
                final_branch(old_return, return_proxy)
            else:
                return_action = self.add_block("try.doreturn")
                value = return_action.append(ir.GetLocal(self.current_private_env, "$return"))
                return_action.append(ir.Return(value))
                final_branch(return_action, return_proxy)
        else:
            landingpad.has_cleanup = False

        # we should propagate the clauses to nested try catch blocks
        # so nested try catch will jump to our clause if the inner one does not
        # match
        # note that the phi instruction here requires some hack, see
        # llvm_ir_generator process_function for details
        clauses = []
        found_catch_all = False
        for handler_node in node.handlers:
            if found_catch_all:
                self.warn_unreachable(handler_node)
                continue
            exn_type = handler_node.name_type.find()
            if handler_node.filter is not None and \
                    not builtins.is_exception(exn_type, 'Exception'):
                handler = self.add_block("handler." + exn_type.name)
                phi = ir.Phi(builtins.TException(), 'exn')
                handler.append(phi)
                clauses.append((handler, exn_type, phi))
            else:
                handler = self.add_block("handler.catchall")
                phi = ir.Phi(builtins.TException(), 'exn')
                handler.append(phi)
                clauses.append((handler, None, phi))
                found_catch_all = True

        all_clauses = clauses[:]
        for clause in self.catch_clauses:
            # if the last clause is accept all, do not add further clauses
            if len(all_clauses) == 0 or all_clauses[-1][1] is not None:
                all_clauses.append(clause)

        body = self.add_block("try.body")
        self.append(ir.Branch(body))
        self.current_block = body

        old_unwind, self.unwind_target = self.unwind_target, dispatcher
        old_clauses, self.catch_clauses = self.catch_clauses, all_clauses
        try:
            self.visit(node.body)
        finally:
            self.unwind_target = old_unwind
            self.catch_clauses = old_clauses

        if not self.current_block.is_terminated():
            self.visit(node.orelse)
        elif any(node.orelse):
            self.warn_unreachable(node.orelse[0])
        body = self.current_block

        if any(node.finalbody):
            # if we have a final block, we should not append clauses to our
            # landingpad or we will skip the finally block.
            # when the finally block calls resume, it will unwind to the outer
            # try catch block automatically
            all_clauses = clauses
            # reset targets
            if self.break_target:
                self.break_target = old_break
            if self.continue_target:
                self.continue_target = old_continue
            self.return_target = old_return

        if any(node.finalbody):
            # create new unwind target for cleanup
            final_dispatcher = self.add_block("try.final.dispatch")
            final_landingpad = ir.LandingPad(cleanup)
            final_dispatcher.append(final_landingpad)

            # make sure that exception clauses are unwinded to the finally block
            old_unwind, self.unwind_target = self.unwind_target, final_dispatcher

        if any(node.finalbody):
            # if we have a while:try/finally continue must execute finally
            # before continuing the while
            redirect = final_branch
        else:
            redirect = lambda dest, proxy: proxy.append(ir.Branch(dest))

        # we need to set break/continue/return to execute end_catch
        if self.break_target is not None:
            break_proxy = self.add_block("try.break")
            break_proxy.append(ir.Builtin("end_catch", [], builtins.TNone()))
            old_break, self.break_target = self.break_target, break_proxy
            redirect(old_break, break_proxy)

        if self.continue_target is not None:
            continue_proxy = self.add_block("try.continue")
            continue_proxy.append(ir.Builtin("end_catch", [],
                                             builtins.TNone()))
            old_continue, self.continue_target = self.continue_target, continue_proxy
            redirect(old_continue, continue_proxy)

        return_proxy = self.add_block("try.return")
        return_proxy.append(ir.Builtin("end_catch", [], builtins.TNone()))
        old_return, self.return_target = self.return_target, return_proxy
        old_return_target = old_return
        if old_return_target is None:
            old_return_target = self.add_block("try.doreturn")
            value = old_return_target.append(ir.GetLocal(self.current_private_env, "$return"))
            old_return_target.append(ir.Return(value))
        redirect(old_return_target, return_proxy)

        handlers = []

        for (handler_node, (handler, exn_type, phi)) in zip(node.handlers, clauses):
            self.current_block = handler
            if handler_node.name is not None:
                exn = self.append(ir.Builtin("exncast", [phi], handler_node.name_type))
                self._set_local(handler_node.name, exn)
            self.visit(handler_node.body)
            # only need to call end_catch if the current block is not terminated
            # other possible paths: break/continue/return/raise
            # we will call end_catch in the first 3 cases, and we should not
            # end_catch in the last case for nested exception
            if not self.current_block.is_terminated():
                self.append(ir.Builtin("end_catch", [], builtins.TNone()))
            post_handler = self.current_block
            handlers.append(post_handler)

        # branch to all possible clauses, including those from outer try catch
        # block
        # if we have a finally block, all_clauses will not include those from
        # the outer block
        for (handler, clause, phi) in all_clauses:
            phi.add_incoming(landingpad, dispatcher)
            landingpad.add_clause(handler, clause)

        if self.break_target:
            self.break_target = old_break
        if self.continue_target:
            self.continue_target = old_continue
        self.return_target = old_return

        if any(node.finalbody):
            # Finalize and continue after try statement.
            self.unwind_target = old_unwind
            # Exception path
            finalizer_reraise = self.add_block("finally.resume")
            self.current_block = finalizer_reraise
            self.visit(node.finalbody)
            self.terminate(ir.Resume(self.unwind_target))
            cleanup.append(ir.Branch(finalizer_reraise))

            # Normal path
            finalizer = self.add_block("finally")
            self.current_block = finalizer
            self.visit(node.finalbody)
            post_finalizer = self.current_block
            self.current_block = tail = self.add_block("try.tail")
            final_targets.append(tail)

            # if final block is not terminated, branch to tail
            if not post_finalizer.is_terminated():
                dest = post_finalizer.append(ir.GetLocal(final_state, "$cont"))
                post_finalizer.append(ir.IndirectBranch(dest, final_targets))
            # make sure proxies will branch to finalizer
            for block in final_paths:
                if finalizer in block.predecessors():
                    # avoid producing irreducible graphs
                    # generate a new finalizer
                    self.current_block = tmp_finalizer = self.add_block("finally.tmp")
                    self.visit(node.finalbody)
                    if not self.current_block.is_terminated():
                        assert isinstance(block.instructions[-1], ir.SetLocal)
                        self.current_block.append(ir.Branch(block.instructions[-1].operands[-1]))
                        block.instructions[-1].erase()
                    block.append(ir.Branch(tmp_finalizer))
                    self.current_block = tail
                else:
                    block.append(ir.Branch(finalizer))
            # if no raise in body/handlers, branch to finalizer
            for block in chain([body], handlers):
                if not block.is_terminated():
                    if finalizer in block.predecessors():
                        # similar to the above case
                        self.current_block = tmp_finalizer = self.add_block("finally.tmp")
                        self.visit(node.finalbody)
                        self.terminate(ir.Branch(tail))
                        block.append(ir.Branch(tmp_finalizer))
                        self.current_block = tail
                    else:
                        block.append(ir.SetLocal(final_state, "$cont", tail))
                        block.append(ir.Branch(finalizer))
        else:
            self.current_block = tail = self.add_block("try.tail")
            if not body.is_terminated():
                body.append(ir.Branch(tail))

            cleanup.append(ir.Resume(self.unwind_target))

            for handler in handlers:
                if not handler.is_terminated():
                    handler.append(ir.Branch(tail))

    def _try_finally(self, body_gen, finally_gen, name):
        dispatcher = self.add_block("{}.dispatch".format(name))

        try:
            old_unwind, self.unwind_target = self.unwind_target, dispatcher
            body_gen()
        finally:
            self.unwind_target = old_unwind

        if not self.current_block.is_terminated():
            finally_gen()

        self.post_body = self.current_block

        self.current_block = self.add_block("{}.cleanup".format(name))
        dispatcher.append(ir.LandingPad(self.current_block))
        finally_gen()
        self.terminate(ir.Resume(self.unwind_target))

        self.current_block = self.post_body

    def visit_With(self, node):
        context_expr_node  = node.items[0].context_expr
        optional_vars_node = node.items[0].optional_vars

        if types.is_builtin(context_expr_node.type, "sequential"):
            self.visit(node.body)
            return
        elif types.is_builtin(context_expr_node.type, "interleave"):
            interleave = self.append(ir.Interleave([]))

            heads, tails = [], []
            for stmt in node.body:
                self.current_block = self.add_block("interleave.branch")
                heads.append(self.current_block)
                self.visit(stmt)
                tails.append(self.current_block)

            for head in heads:
                interleave.add_destination(head)

            self.current_block = self.add_block("interleave.tail")
            for tail in tails:
                if not tail.is_terminated():
                    tail.append(ir.Branch(self.current_block))
            return
        elif types.is_builtin(context_expr_node.type, "parallel"):
            start_mu = self.append(ir.Builtin("now_mu", [], builtins.TInt64()))
            end_mu   = start_mu

            for stmt in node.body:
                self.append(ir.Builtin("at_mu", [start_mu], builtins.TNone()))

                block = self.add_block("parallel.branch")
                if self.current_block.is_terminated():
                    self.warn_unreachable(stmt[0])
                else:
                    self.append(ir.Branch(block))
                self.current_block = block

                self.visit(stmt)

                mid_mu = self.append(ir.Builtin("now_mu", [], builtins.TInt64()))
                gt_mu  = self.append(ir.Compare(ast.Gt(loc=None), mid_mu, end_mu))
                end_mu = self.append(ir.Select(gt_mu, mid_mu, end_mu))

            self.append(ir.Builtin("at_mu", [end_mu], builtins.TNone()))
            return

        cleanup = []
        for item_node in node.items:
            # user-defined context manager
            context_expr_node  = item_node.context_expr
            optional_vars_node = item_node.optional_vars
            context_mgr = self.visit(context_expr_node)
            enter_fn    = self.append(ir.GetAttr(context_mgr, '__enter__'))
            exit_fn     = self.append(ir.GetAttr(context_mgr, '__exit__'))

            try:
                self.current_assign = self._user_call(enter_fn, [], {})
                if optional_vars_node is not None:
                    self.visit(optional_vars_node)
            finally:
                self.current_assign = None

            none = self.append(ir.Alloc([], builtins.TNone()))
            cleanup.append(lambda:
                self._user_call(exit_fn, [none, none, none], {}))

        self._try_finally(
            body_gen=lambda: self.visit(node.body),
            finally_gen=lambda: [thunk() for thunk in cleanup],
            name="with")

    # Expression visitors
    # These visitors return a node in addition to mutating
    # the IR.

    def visit_LambdaT(self, node):
        return self.visit_function(node, is_lambda=True, is_internal=True)

    def visit_IfExpT(self, node):
        cond = self.visit(node.test)
        head = self.current_block

        if_true = self.add_block("ifexp.body")
        self.current_block = if_true
        true_result = self.visit(node.body)
        post_if_true = self.current_block

        if_false = self.add_block("ifexp.else")
        self.current_block = if_false
        false_result = self.visit(node.orelse)
        post_if_false = self.current_block

        tail = self.add_block("ifexp.tail")
        self.current_block = tail

        if not post_if_true.is_terminated():
            post_if_true.append(ir.Branch(tail))
        if not post_if_false.is_terminated():
            post_if_false.append(ir.Branch(tail))
        head.append(ir.BranchIf(cond, if_true, if_false))

        phi = self.append(ir.Phi(node.type))
        phi.add_incoming(true_result, post_if_true)
        phi.add_incoming(false_result, post_if_false)
        return phi

    def visit_NumT(self, node):
        return ir.Constant(node.n, node.type)

    def visit_StrT(self, node):
        return ir.Constant(node.s, node.type)

    def visit_NameConstantT(self, node):
        return ir.Constant(node.value, node.type)

    def _env_for(self, name):
        if name in self.current_globals:
            return self.append(ir.Builtin("globalenv", [self.current_env],
                                          self.current_env.type.outermost()))
        else:
            return self.current_env

    def _get_local(self, name):
        if self.current_class is not None and \
                name in self.current_class.type.attributes:
            return self.append(ir.GetAttr(self.current_class, name,
                                          name="FLD." + name))

        return self.append(ir.GetLocal(self._env_for(name), name,
                                       name="LOC." + name))

    def _set_local(self, name, value):
        if self.current_class is not None and \
                name in self.current_class.type.attributes:
            return self.append(ir.SetAttr(self.current_class, name, value))

        self.append(ir.SetLocal(self._env_for(name), name, value))

    def visit_NameT(self, node):
        if self.current_assign is None:
            insn = self._get_local(node.id)
            self.variable_map[node] = insn
            return insn
        else:
            return self._set_local(node.id, self.current_assign)

    def visit_AttributeT(self, node):
        try:
            old_assign, self.current_assign = self.current_assign, None
            obj = self.visit(node.value)
        finally:
            self.current_assign = old_assign

        if self.current_assign is None:
            return self.append(ir.GetAttr(obj, node.attr,
                                          name="{}.FLD.{}".format(_readable_name(obj), node.attr)))
        else:
            return self.append(ir.SetAttr(obj, node.attr, self.current_assign))

    def _make_check(self, cond, exn_gen, loc=None, params=[]):
        if loc is None:
            loc = self.current_loc

        try:
            name = "check:{}:{}".format(loc.line(), loc.column())
            args = [ir.EnvironmentArgument(self.current_env.type, "ARG.ENV")] + \
                   [ir.Argument(param.type, "ARG.{}".format(index))
                    for index, param in enumerate(params)]
            typ  = types.TFunction(OrderedDict([("arg{}".format(index), param.type)
                                                for index, param in enumerate(params)]),
                                   OrderedDict(),
                                   builtins.TNone())
            func = ir.Function(typ, ".".join(self.name + [name]), args, loc=loc)
            func.is_internal = True
            func.is_cold = True
            func.is_generated = True
            self.functions.append(func)
            old_func, self.current_function = self.current_function, func

            entry = self.add_block("entry")
            old_block, self.current_block = self.current_block, entry

            old_final_branch, self.final_branch = self.final_branch, None
            old_unwind, self.unwind_target = self.unwind_target, None
            self.raise_exn(lambda: exn_gen(*args[1:]), loc=loc)
        finally:
            self.current_function = old_func
            self.current_block = old_block
            self.final_branch = old_final_branch
            self.unwind_target = old_unwind

        # cond:    bool Value, condition
        # exn_gen: lambda()->exn Value, exception if condition not true
        cond_block = self.current_block

        self.current_block = body_block = self.add_block("check.body")
        self._invoke_raising_func(func, params, "check")

        self.current_block = tail_block = self.add_block("check.tail")
        cond_block.append(ir.BranchIf(cond, tail_block, body_block))

    def _invoke_raising_func(self, func, params, block_name):
        """Emit a call/invoke instruction as appropriate to terminte the current
        basic block with a call to a helper function that always raises an
        exception.

        (This is done for compiler-inserted checks and assertions to keep the
        generated code tight for the normal case.)
        """
        closure = self.append(ir.Closure(func,
            ir.Constant(None, ir.TEnvironment("raise", {}))))
        if self.unwind_target is None:
            insn = self.append(ir.Call(closure, params, {}))
        else:
            after_invoke = self.add_block(block_name + ".invoke")
            insn = self.append(ir.Invoke(closure, params, {}, after_invoke, self.unwind_target))
            self.current_block = after_invoke
        insn.is_cold = True
        self.append(ir.Unreachable())

    def _map_index(self, length, index, one_past_the_end=False, loc=None):
        lt_0          = self.append(ir.Compare(ast.Lt(loc=None),
                                               index, ir.Constant(0, index.type)))
        from_end      = self.append(ir.Arith(ast.Add(loc=None), length, index))
        mapped_index  = self.append(ir.Select(lt_0, from_end, index))
        mapped_ge_0   = self.append(ir.Compare(ast.GtE(loc=None),
                                               mapped_index, ir.Constant(0, mapped_index.type)))
        end_cmpop     = ast.LtE(loc=None) if one_past_the_end else ast.Lt(loc=None)
        mapped_lt_len = self.append(ir.Compare(end_cmpop, mapped_index, length))
        in_bounds     = self.append(ir.Select(mapped_ge_0, mapped_lt_len,
                                              ir.Constant(False, builtins.TBool())))
        head = self.current_block

        self._make_check(
            in_bounds,
            lambda index, length: self.alloc_exn(builtins.TException("IndexError"),
                ir.Constant("index {0} out of bounds 0:{1}", builtins.TStr()),
                index, length),
            params=[index, length],
            loc=loc)

        return mapped_index

    def _make_loop(self, init, cond_gen, body_gen, name="loop"):
        # init:     'iter Value, initial loop variable value
        # cond_gen: lambda('iter Value)->bool Value, loop condition
        # body_gen: lambda('iter Value)->'iter Value, loop body,
        #               returns next loop variable value
        init_block = self.current_block

        self.current_block = head_block = self.add_block("{}.head".format(name))
        init_block.append(ir.Branch(head_block))
        phi = self.append(ir.Phi(init.type))
        phi.add_incoming(init, init_block)
        cond = cond_gen(phi)

        self.current_block = body_block = self.add_block("{}.body".format(name))
        body = body_gen(phi)
        self.append(ir.Branch(head_block))
        phi.add_incoming(body, self.current_block)

        self.current_block = tail_block = self.add_block("{}.tail".format(name))
        head_block.append(ir.BranchIf(cond, body_block, tail_block))

        return head_block, body_block, tail_block

    def visit_SubscriptT(self, node):
        try:
            old_assign, self.current_assign = self.current_assign, None
            value = self.visit(node.value)
        finally:
            self.current_assign = old_assign

        if types.is_tuple(node.value.type):
            assert isinstance(node.slice, ast.Index), \
                "Internal compiler error: tuple index should be an Index"
            assert isinstance(node.slice.value, ast.Num), \
                "Internal compiler error: tuple index should be a constant"

            if self.current_assign is not None:
                diag = diagnostic.Diagnostic("error",
                    "cannot assign to a tuple element",
                    {}, node.loc)
                self.engine.process(diag)

            index = node.slice.value.n
            indexed = self.append(
                ir.GetAttr(value, index, name="{}.e{}".format(value.name, index)),
                loc=node.loc
            )

            return indexed

        elif isinstance(node.slice, ast.Index):
            try:
                old_assign, self.current_assign = self.current_assign, None
                index = self.visit(node.slice.value)
            finally:
                self.current_assign = old_assign

            # For multi-dimensional indexes, just apply them sequentially. This
            # works, as they are only supported for types where we do not
            # immediately need to distinguish between the Get and Set cases
            # (i.e. arrays, which are reference types).
            if types.is_tuple(index.type):
                num_idxs = len(index.type.find().elts)
                indices = [
                    self.append(ir.GetAttr(index, i)) for i in range(num_idxs)
                ]
            else:
                indices = [index]
            indexed = value
            for i, idx in enumerate(indices):
                length = self.iterable_len(indexed, idx.type)
                mapped_index = self._map_index(length, idx, loc=node.begin_loc)
                if self.current_assign is None or i < len(indices) - 1:
                    indexed = self.iterable_get(indexed, mapped_index)
                    indexed.set_name("{}.at.{}".format(indexed.name,
                                                       _readable_name(idx)))
                else:
                    self.append(ir.SetElem(indexed, mapped_index, self.current_assign,
                                           name="{}.at.{}".format(value.name,
                                                                  _readable_name(index))))
            if self.current_assign is None:
                return indexed
        else:
            # This is a slice. The endpoint checking logic is the same for both lists
            # and NumPy arrays, but the actual implementations differ – while slices of
            # built-in lists are always copies in Python, they are views sharing the
            # same backing storage in NumPy.
            length = self.iterable_len(value, node.slice.type)

            if node.slice.lower is not None:
                try:
                    old_assign, self.current_assign = self.current_assign, None
                    start_index = self.visit(node.slice.lower)
                finally:
                    self.current_assign = old_assign
            else:
                start_index = ir.Constant(0, node.slice.type)
            mapped_start_index = self._map_index(length, start_index,
                                                 loc=node.begin_loc)

            if node.slice.upper is not None:
                try:
                    old_assign, self.current_assign = self.current_assign, None
                    stop_index = self.visit(node.slice.upper)
                finally:
                    self.current_assign = old_assign
            else:
                stop_index = length
            mapped_stop_index = self._map_index(length, stop_index, one_past_the_end=True,
                                                loc=node.begin_loc)

            if builtins.is_array(node.type):
                # To implement strided slicing with the proper NumPy reference
                # semantics, the pointer/length array representation will need to be
                # extended by another field to hold a variable stride.
                assert node.slice.step is None, (
                    "array slices with non-trivial step "
                    "should have been disallowed during type inference")

                # One-dimensionally slicing an array only affects the outermost
                # dimension.
                shape = self.append(ir.GetAttr(value, "shape"))
                lengths = [
                    self.append(ir.GetAttr(shape, i))
                    for i in range(len(shape.type.elts))
                ]

                # Compute outermost length – zero for "backwards" indices.
                raw_len = self.append(
                    ir.Arith(ast.Sub(loc=None), mapped_stop_index, mapped_start_index))
                is_neg_len = self.append(
                    ir.Compare(ast.Lt(loc=None), raw_len, ir.Constant(0, raw_len.type)))
                outer_len = self.append(
                    ir.Select(is_neg_len, ir.Constant(0, raw_len.type), raw_len))
                new_shape = self._make_array_shape([outer_len] + lengths[1:])

                # Offset buffer pointer by start index (times stride for inner dims).
                stride = reduce(
                    lambda l, r: self.append(ir.Arith(ast.Mult(loc=None), l, r)),
                    lengths[1:], ir.Constant(1, lengths[0].type))
                offset = self.append(
                    ir.Arith(ast.Mult(loc=None), stride, mapped_start_index))
                buffer = self.append(ir.GetAttr(value, "buffer"))
                new_buffer = self.append(ir.Offset(buffer, offset))

                return self.append(ir.Alloc([new_buffer, new_shape], node.type))
            else:
                if node.slice.step is not None:
                    try:
                        old_assign, self.current_assign = self.current_assign, None
                        step = self.visit(node.slice.step)
                    finally:
                        self.current_assign = old_assign

                    self._make_check(
                        self.append(ir.Compare(ast.NotEq(loc=None), step, ir.Constant(0, step.type))),
                        lambda: self.alloc_exn(builtins.TException("ValueError"),
                            ir.Constant("step cannot be zero", builtins.TStr())),
                        loc=node.slice.step.loc)
                else:
                    step = ir.Constant(1, node.slice.type)
                counting_up = self.append(ir.Compare(ast.Gt(loc=None), step,
                                                    ir.Constant(0, step.type)))

                unstepped_size = self.append(ir.Arith(ast.Sub(loc=None),
                                                    mapped_stop_index, mapped_start_index))
                slice_size_a = self.append(ir.Arith(ast.FloorDiv(loc=None), unstepped_size, step))
                slice_size_b = self.append(ir.Arith(ast.Mod(loc=None), unstepped_size, step))
                rem_not_empty = self.append(ir.Compare(ast.NotEq(loc=None), slice_size_b,
                                                    ir.Constant(0, slice_size_b.type)))
                slice_size_c = self.append(ir.Arith(ast.Add(loc=None), slice_size_a,
                                                    ir.Constant(1, slice_size_a.type)))
                slice_size = self.append(ir.Select(rem_not_empty,
                                                slice_size_c, slice_size_a,
                                                name="slice.size"))
                self._make_check(
                    self.append(ir.Compare(ast.LtE(loc=None), slice_size, length)),
                    lambda slice_size, length: self.alloc_exn(builtins.TException("ValueError"),
                        ir.Constant("slice size {0} is larger than iterable length {1}",
                                    builtins.TStr()),
                        slice_size, length),
                    params=[slice_size, length],
                    loc=node.slice.loc)

                if self.current_assign is None:
                    is_neg_size = self.append(ir.Compare(ast.Lt(loc=None),
                                                        slice_size, ir.Constant(0, slice_size.type)))
                    abs_slice_size = self.append(ir.Select(is_neg_size,
                                                        ir.Constant(0, slice_size.type), slice_size))
                    other_value = self.append(ir.Alloc([abs_slice_size], value.type,
                                                    name="slice.result"))
                else:
                    other_value = self.current_assign

                prehead = self.current_block

                head = self.current_block = self.add_block("slice.head")
                prehead.append(ir.Branch(head))

                index = self.append(ir.Phi(node.slice.type,
                                        name="slice.index"))
                index.add_incoming(mapped_start_index, prehead)
                other_index = self.append(ir.Phi(node.slice.type,
                                                name="slice.resindex"))
                other_index.add_incoming(ir.Constant(0, node.slice.type), prehead)

                # Still within bounds?
                bounded_up = self.append(ir.Compare(ast.Lt(loc=None), index, mapped_stop_index))
                bounded_down = self.append(ir.Compare(ast.Gt(loc=None), index, mapped_stop_index))
                within_bounds = self.append(ir.Select(counting_up, bounded_up, bounded_down))

                body = self.current_block = self.add_block("slice.body")

                if self.current_assign is None:
                    elem = self.iterable_get(value, index)
                    self.append(ir.SetElem(other_value, other_index, elem))
                else:
                    elem = self.append(ir.GetElem(self.current_assign, other_index))
                    self.append(ir.SetElem(value, index, elem))

                next_index = self.append(ir.Arith(ast.Add(loc=None), index, step))
                index.add_incoming(next_index, body)
                next_other_index = self.append(ir.Arith(ast.Add(loc=None), other_index,
                                                        ir.Constant(1, node.slice.type)))
                other_index.add_incoming(next_other_index, body)
                self.append(ir.Branch(head))

                tail = self.current_block = self.add_block("slice.tail")
                head.append(ir.BranchIf(within_bounds, body, tail))

                if self.current_assign is None:
                    return other_value

    def visit_TupleT(self, node):
        if self.current_assign is None:
            return self.append(ir.Alloc([self.visit(elt) for elt in node.elts], node.type))
        else:
            try:
                old_assign = self.current_assign
                for index, elt_node in enumerate(node.elts):
                    self.current_assign = \
                        self.append(ir.GetAttr(old_assign, index,
                                               name="{}.e{}".format(old_assign.name, index)),
                                    loc=elt_node.loc)
                    self.visit(elt_node)
            finally:
                self.current_assign = old_assign

    def visit_ListT(self, node):
        if self.current_assign is None:
            elts = [self.visit(elt_node) for elt_node in node.elts]
            lst = self.append(ir.Alloc([ir.Constant(len(node.elts), self._size_type)],
                                       node.type))
            for index, elt_node in enumerate(elts):
                self.append(ir.SetElem(lst, ir.Constant(index, self._size_type), elt_node))
            return lst
        else:
            length = self.iterable_len(self.current_assign)
            self._make_check(
                self.append(ir.Compare(ast.Eq(loc=None), length,
                                       ir.Constant(len(node.elts), self._size_type))),
                lambda length: self.alloc_exn(builtins.TException("ValueError"),
                    ir.Constant("list must be {0} elements long to decompose", builtins.TStr()),
                    length),
                params=[length])

            for index, elt_node in enumerate(node.elts):
                elt = self.append(ir.GetElem(self.current_assign,
                                             ir.Constant(index, self._size_type)))
                try:
                    old_assign, self.current_assign = self.current_assign, elt
                    self.visit(elt_node)
                finally:
                    self.current_assign = old_assign

    def visit_ListCompT(self, node):
        assert len(node.generators) == 1
        comprehension = node.generators[0]
        assert comprehension.ifs == []

        iterable = self.visit(comprehension.iter)
        length = self.iterable_len(iterable)
        result = self.append(ir.Alloc([length], node.type))

        try:
            gen_suffix = ".gen@{}:{}".format(node.loc.line(), node.loc.column())
            env_type = ir.TEnvironment(name=self.current_function.name + gen_suffix,
                                       vars=node.typing_env, outer=self.current_env.type)
            env = self.append(ir.Alloc([], env_type, name="env.gen"))
            old_env, self.current_env = self.current_env, env

            self.append(ir.SetLocal(env, "$outer", old_env))

            def body_gen(index):
                elt = self.iterable_get(iterable, index)
                try:
                    old_assign, self.current_assign = self.current_assign, elt
                    self.visit(comprehension.target)
                finally:
                    self.current_assign = old_assign

                mapped_elt = self.visit(node.elt)
                self.append(ir.SetElem(result, index, mapped_elt))
                return self.append(ir.Arith(ast.Add(loc=None), index,
                                            ir.Constant(1, length.type)))
            self._make_loop(ir.Constant(0, length.type),
                lambda index: self.append(ir.Compare(ast.Lt(loc=None), index, length)),
                body_gen)

            return result
        finally:
            self.current_env = old_env

    def visit_BoolOpT(self, node):
        blocks = []
        for value_node in node.values:
            value_head = self.current_block
            value = self.visit(value_node)
            value_tail = self.current_block

            blocks.append((value, value_head, value_tail))
            self.current_block = self.add_block("boolop.seq")

        tail = self.current_block
        phi = self.append(ir.Phi(node.type))
        for ((value, value_head, value_tail), (next_value_head, next_value_tail)) in \
                    zip(blocks, [(h,t) for (v,h,t) in blocks[1:]] + [(tail, tail)]):
            phi.add_incoming(value, value_tail)
            if next_value_head != tail:
                cond = self.coerce_to_bool(value, block=value_tail)
                if isinstance(node.op, ast.And):
                    value_tail.append(ir.BranchIf(cond, next_value_head, tail))
                else:
                    value_tail.append(ir.BranchIf(cond, tail, next_value_head))
            else:
                value_tail.append(ir.Branch(tail))
        return phi

    def _make_array_unaryop(self, name, make_op, result_type, arg_type):
        try:
            result = ir.Argument(result_type, "result")
            arg = ir.Argument(arg_type, "arg")

            # TODO: We'd like to use a "C function" here to be able to supply
            # specialised implementations in a library in the future (and e.g. avoid
            # passing around the context argument), but the code generator currently
            # doesn't allow emitting them.
            args = [result, arg]
            typ = types.TFunction(args=OrderedDict([(arg.name, arg.type)
                                                    for arg in args]),
                                  optargs=OrderedDict(),
                                  ret=builtins.TNone())
            env_args = [ir.EnvironmentArgument(self.current_env.type, "ARG.ENV")]

            old_loc, self.current_loc = self.current_loc, None
            func = ir.Function(typ, name, env_args + args)
            func.is_internal = True
            func.is_generated = True
            self.functions.append(func)
            old_func, self.current_function = self.current_function, func

            entry = self.add_block("entry")
            old_block, self.current_block = self.current_block, entry

            old_final_branch, self.final_branch = self.final_branch, None
            old_unwind, self.unwind_target = self.unwind_target, None

            shape = self.append(ir.GetAttr(arg, "shape"))

            result_buffer = self.append(ir.GetAttr(result, "buffer"))
            arg_buffer = self.append(ir.GetAttr(arg, "buffer"))
            num_total_elts = self._get_total_array_len(shape)

            def body_gen(index):
                a = self.append(ir.GetElem(arg_buffer, index))
                self.append(
                    ir.SetElem(result_buffer, index, make_op(a)))
                return self.append(
                    ir.Arith(ast.Add(loc=None), index, ir.Constant(1, self._size_type)))

            self._make_loop(
                ir.Constant(0, self._size_type), lambda index: self.append(
                    ir.Compare(ast.Lt(loc=None), index, num_total_elts)), body_gen)

            self.append(ir.Return(ir.Constant(None, builtins.TNone())))
            return func
        finally:
            self.current_loc = old_loc
            self.current_function = old_func
            self.current_block = old_block
            self.final_branch = old_final_branch
            self.unwind_target = old_unwind

    def _get_array_unaryop(self, name, make_op, result_type, arg_type):
        name = "_array_{}_{}".format(
            name, self._mangle_arrayop_types([result_type, arg_type]))
        if name not in self.array_op_funcs:
            self.array_op_funcs[name] = self._make_array_unaryop(
                name, make_op, result_type, arg_type)
        return self.array_op_funcs[name]

    def visit_UnaryOpT(self, node):
        if isinstance(node.op, ast.Not):
            cond = self.coerce_to_bool(self.visit(node.operand))
            return self.append(ir.Select(cond,
                        ir.Constant(False, builtins.TBool()),
                        ir.Constant(True,  builtins.TBool())))
        elif isinstance(node.op, ast.Invert):
            operand = self.visit(node.operand)
            return self.append(ir.Arith(ast.BitXor(loc=None),
                                        ir.Constant(-1, operand.type), operand))
        elif isinstance(node.op, ast.USub):
            def make_sub(val):
                return self.append(ir.Arith(ast.Sub(loc=None),
                                        ir.Constant(0, val.type), val))
            operand = self.visit(node.operand)
            if builtins.is_array(operand.type):
                shape = self.append(ir.GetAttr(operand, "shape"))
                result, _ = self._allocate_new_array(node.type.find()["elt"], shape)
                func = self._get_array_unaryop("USub", make_sub, node.type, operand.type)
                self._invoke_arrayop(func, [result, operand])
                return result
            else:
                return make_sub(operand)
        elif isinstance(node.op, ast.UAdd):
            # No-op.
            return self.visit(node.operand)
        else:
            assert False

    def visit_CoerceT(self, node):
        value = self.visit(node.value)
        if node.type.find() == value.type:
            return value
        else:
            if builtins.is_array(node.type):
                result_elt = node.type.find()["elt"]
                shape = self.append(ir.GetAttr(value, "shape"))
                result, _ = self._allocate_new_array(result_elt, shape)
                func = self._get_array_unaryop(
                    "Coerce", lambda v: self.append(ir.Coerce(v, result_elt)),
                    node.type, value.type)
                self._invoke_arrayop(func, [result, value])
                return result
            else:
                return self.append(
                    ir.Coerce(value,
                              node.type,
                              name="{}.{}".format(_readable_name(value),
                                                  node.type.name)))

    def _get_total_array_len(self, shape):
        lengths = [
            self.append(ir.GetAttr(shape, i)) for i in range(len(shape.type.elts))
        ]
        return reduce(lambda l, r: self.append(ir.Arith(ast.Mult(loc=None), l, r)),
                      lengths[1:], lengths[0])

    def _allocate_new_array(self, elt, shape):
        total_length = self._get_total_array_len(shape)
        buffer = self.append(ir.Alloc([total_length], types._TPointer(elt=elt)))
        result_type = builtins.TArray(elt, types.TValue(len(shape.type.elts)))
        return self.append(ir.Alloc([buffer, shape], result_type)), total_length

    def _make_array_binop(self, name, result_type, lhs_type, rhs_type, body_gen):
        try:
            result = ir.Argument(result_type, "result")
            lhs = ir.Argument(lhs_type, "lhs")
            rhs = ir.Argument(rhs_type, "rhs")

            # TODO: We'd like to use a "C function" here to be able to supply
            # specialised implementations in a library in the future (and e.g. avoid
            # passing around the context argument), but the code generator currently
            # doesn't allow emitting them.
            args = [result, lhs, rhs]
            typ = types.TFunction(args=OrderedDict([(arg.name, arg.type)
                                                    for arg in args]),
                                  optargs=OrderedDict(),
                                  ret=builtins.TNone())
            env_args = [ir.EnvironmentArgument(self.current_env.type, "ARG.ENV")]

            old_loc, self.current_loc = self.current_loc, None
            func = ir.Function(typ, name, env_args + args)
            func.is_internal = True
            func.is_generated = True
            self.functions.append(func)
            old_func, self.current_function = self.current_function, func

            entry = self.add_block("entry")
            old_block, self.current_block = self.current_block, entry

            old_final_branch, self.final_branch = self.final_branch, None
            old_unwind, self.unwind_target = self.unwind_target, None

            body_gen(result, lhs, rhs)

            self.append(ir.Return(ir.Constant(None, builtins.TNone())))
            return func
        finally:
            self.current_loc = old_loc
            self.current_function = old_func
            self.current_block = old_block
            self.final_branch = old_final_branch
            self.unwind_target = old_unwind

    def _make_array_elementwise_binop(self, name, result_type, lhs_type,
                                      rhs_type, make_op):
        def body_gen(result, lhs, rhs):
            # At this point, shapes are assumed to match; could just pass buffer
            # pointer for two of the three arrays as well.
            result_buffer = self.append(ir.GetAttr(result, "buffer"))
            shape = self.append(ir.GetAttr(result, "shape"))
            num_total_elts = self._get_total_array_len(shape)

            if builtins.is_array(lhs.type):
                lhs_buffer = self.append(ir.GetAttr(lhs, "buffer"))
                def get_left(index):
                    return self.append(ir.GetElem(lhs_buffer, index))
            else:
                def get_left(index):
                    return lhs

            if builtins.is_array(rhs.type):
                rhs_buffer = self.append(ir.GetAttr(rhs, "buffer"))
                def get_right(index):
                    return self.append(ir.GetElem(rhs_buffer, index))
            else:
                def get_right(index):
                    return rhs

            def loop_gen(index):
                l = get_left(index)
                r = get_right(index)
                result = make_op(l, r)
                self.append(ir.SetElem(result_buffer, index, result))
                return self.append(
                    ir.Arith(ast.Add(loc=None), index,
                             ir.Constant(1, self._size_type)))

            self._make_loop(
                ir.Constant(0, self._size_type), lambda index: self.append(
                    ir.Compare(ast.Lt(loc=None), index, num_total_elts)),
                loop_gen)

        return self._make_array_binop(name, result_type, lhs_type, rhs_type,
                                      body_gen)

    def _mangle_arrayop_types(self, types):
        def name_error(typ):
            assert False, "Internal compiler error: No RPC tag for {}".format(typ)

        def mangle_name(typ):
            typ = typ.find()
            # rpc_tag is used to turn element types into mangled names for no
            # particularly good reason apart from not having to invent yet another
            # string representation.
            if builtins.is_array(typ):
                return mangle_name(typ["elt"]) + str(typ["num_dims"].find().value)
            return ir.rpc_tag(typ, name_error).decode()

        return "_".join(mangle_name(t) for t in types)

    def _get_array_elementwise_binop(self, name, make_op, result_type, lhs_type, rhs_type):
        # Currently, we always have any type coercions resolved explicitly in the AST.
        # In the future, this might no longer be true and the three types might all
        # differ.
        name = "_array_{}_{}".format(
            name,
            self._mangle_arrayop_types([result_type, lhs_type, rhs_type]))
        if name not in self.array_op_funcs:
            self.array_op_funcs[name] = self._make_array_elementwise_binop(
                name, result_type, lhs_type, rhs_type, make_op)
        return self.array_op_funcs[name]

    def _invoke_arrayop(self, func, params):
        closure = self.append(
            ir.Closure(func, ir.Constant(None, ir.TEnvironment("arrayop", {}))))
        if self.unwind_target is None:
            self.append(ir.Call(closure, params, {}))
        else:
            after_invoke = self.add_block("arrayop.invoke")
            self.append(ir.Invoke(func, params, {}, after_invoke, self.unwind_target))
            self.current_block = after_invoke

    def _get_array_offset(self, shape, indices):
        result = indices[0]
        for dim, index in zip(shape[1:], indices[1:]):
            result = self.append(ir.Arith(ast.Mult(loc=None), result, dim))
            result = self.append(ir.Arith(ast.Add(loc=None), result, index))
        return result

    def _get_matmult(self, result_type, lhs_type, rhs_type):
        name = "_array_MatMult_" + self._mangle_arrayop_types(
            [result_type, lhs_type, rhs_type])
        if name not in self.array_op_funcs:

            def body_gen(result, lhs, rhs):
                assert builtins.is_array(result.type), \
                    "vec @ vec should have been normalised into array result"

                # We assume result has correct shape; could just pass buffer pointer
                # as well.
                result_buffer = self.append(ir.GetAttr(result, "buffer"))
                lhs_buffer = self.append(ir.GetAttr(lhs, "buffer"))
                rhs_buffer = self.append(ir.GetAttr(rhs, "buffer"))

                num_rows, num_summands, _, num_cols = self._get_matmult_shapes(lhs, rhs)

                elt = result.type["elt"].find()
                env_type = ir.TEnvironment(name + ".loop", {"$total": elt})
                env = self.append(ir.Alloc([], env_type))

                def row_loop(row_idx):
                    lhs_base_offset = self.append(
                        ir.Arith(ast.Mult(loc=None), row_idx, num_summands))
                    lhs_base = self.append(ir.Offset(lhs_buffer, lhs_base_offset))
                    result_base_offset = self.append(
                        ir.Arith(ast.Mult(loc=None), row_idx, num_cols))
                    result_base = self.append(
                        ir.Offset(result_buffer, result_base_offset))

                    def col_loop(col_idx):
                        rhs_base = self.append(ir.Offset(rhs_buffer, col_idx))

                        self.append(
                            ir.SetLocal(env, "$total", ir.Constant(elt.zero(), elt)))

                        def sum_loop(sum_idx):
                            lhs_elem = self.append(ir.GetElem(lhs_base, sum_idx))
                            rhs_offset = self.append(
                                ir.Arith(ast.Mult(loc=None), sum_idx, num_cols))
                            rhs_elem = self.append(ir.GetElem(rhs_base, rhs_offset))
                            product = self.append(
                                ir.Arith(ast.Mult(loc=None), lhs_elem, rhs_elem))
                            prev_total = self.append(ir.GetLocal(env, "$total"))
                            total = self.append(
                                ir.Arith(ast.Add(loc=None), prev_total, product))
                            self.append(ir.SetLocal(env, "$total", total))
                            return self.append(
                                ir.Arith(ast.Add(loc=None), sum_idx,
                                         ir.Constant(1, self._size_type)))

                        self._make_loop(
                            ir.Constant(0, self._size_type), lambda index: self.append(
                                ir.Compare(ast.Lt(loc=None), index, num_summands)),
                            sum_loop)

                        total = self.append(ir.GetLocal(env, "$total"))
                        self.append(ir.SetElem(result_base, col_idx, total))

                        return self.append(
                            ir.Arith(ast.Add(loc=None), col_idx,
                                     ir.Constant(1, self._size_type)))

                    self._make_loop(
                        ir.Constant(0, self._size_type), lambda index: self.append(
                            ir.Compare(ast.Lt(loc=None), index, num_cols)), col_loop)
                    return self.append(
                        ir.Arith(ast.Add(loc=None), row_idx,
                                 ir.Constant(1, self._size_type)))

                self._make_loop(
                    ir.Constant(0, self._size_type), lambda index: self.append(
                        ir.Compare(ast.Lt(loc=None), index, num_rows)), row_loop)

            self.array_op_funcs[name] = self._make_array_binop(
                name, result_type, lhs_type, rhs_type, body_gen)
        return self.array_op_funcs[name]

    def _get_matmult_shapes(self, lhs, rhs):
        lhs_shape = self.append(ir.GetAttr(lhs, "shape"))
        if lhs.type["num_dims"].value == 1:
            lhs_shape_outer = ir.Constant(1, self._size_type)
            lhs_shape_inner = self.append(ir.GetAttr(lhs_shape, 0))
        else:
            lhs_shape_outer = self.append(ir.GetAttr(lhs_shape, 0))
            lhs_shape_inner = self.append(ir.GetAttr(lhs_shape, 1))

        rhs_shape = self.append(ir.GetAttr(rhs, "shape"))
        if rhs.type["num_dims"].value == 1:
            rhs_shape_inner = self.append(ir.GetAttr(rhs_shape, 0))
            rhs_shape_outer = ir.Constant(1, self._size_type)
        else:
            rhs_shape_inner = self.append(ir.GetAttr(rhs_shape, 0))
            rhs_shape_outer = self.append(ir.GetAttr(rhs_shape, 1))

        return lhs_shape_outer, lhs_shape_inner, rhs_shape_inner, rhs_shape_outer

    def _make_array_shape(self, dims):
        return self.append(ir.Alloc(dims, types.TTuple([self._size_type] * len(dims))))

    def _emit_matmult(self, node, left, right):
        # TODO: Also expose as numpy.dot.
        lhs = self.visit(left)
        rhs = self.visit(right)

        num_rows, lhs_inner, rhs_inner, num_cols = self._get_matmult_shapes(lhs, rhs)
        self._make_check(
            self.append(ir.Compare(ast.Eq(loc=None), lhs_inner, rhs_inner)),
            lambda lhs_inner, rhs_inner: self.alloc_exn(
                builtins.TException("ValueError"),
                ir.Constant(
                    "inner dimensions for matrix multiplication do not match ({0} vs. {1})",
                    builtins.TStr()), lhs_inner, rhs_inner),
            params=[lhs_inner, rhs_inner],
            loc=node.loc)
        result_shape = self._make_array_shape([num_rows, num_cols])

        final_type = node.type.find()
        if not builtins.is_array(final_type):
            elt = node.type
            result_dims = 0
        else:
            elt = final_type["elt"]
            result_dims = final_type["num_dims"].value

        result, _ = self._allocate_new_array(elt, result_shape)
        func = self._get_matmult(result.type, left.type, right.type)
        self._invoke_arrayop(func, [result, lhs, rhs])

        if result_dims == 2:
            return result
        result_buffer = self.append(ir.GetAttr(result, "buffer"))
        if result_dims == 1:
            shape = self._make_array_shape(
                [num_cols if lhs.type["num_dims"].value == 1 else num_rows])
            return self.append(ir.Alloc([result_buffer, shape], node.type))
        return self.append(ir.GetElem(result_buffer, ir.Constant(0, self._size_type)))

    def _broadcast_binop(self, name, make_op, result_type, lhs, rhs, assign_to_lhs):
        # Broadcast scalars (broadcasting higher dimensions is not yet allowed in the
        # language).
        broadcast = False
        array_arg = lhs
        if not builtins.is_array(lhs.type):
            broadcast = True
            array_arg = rhs
        elif not builtins.is_array(rhs.type):
            broadcast = True

        shape = self.append(ir.GetAttr(array_arg, "shape"))

        if not broadcast:
            rhs_shape = self.append(ir.GetAttr(rhs, "shape"))
            self._make_check(
                self.append(ir.Compare(ast.Eq(loc=None), shape, rhs_shape)),
                lambda: self.alloc_exn(
                    builtins.TException("ValueError"),
                    ir.Constant("operands could not be broadcast together",
                                builtins.TStr())))
        if assign_to_lhs:
            result = lhs
        else:
            elt = result_type.find()["elt"]
            result, _ = self._allocate_new_array(elt, shape)
        func = self._get_array_elementwise_binop(name, make_op, result_type, lhs.type,
            rhs.type)
        self._invoke_arrayop(func, [result, lhs, rhs])
        return result

    def visit_BinOpT(self, node):
        if isinstance(node.op, ast.MatMult):
            return self._emit_matmult(node, node.left, node.right)
        elif builtins.is_array(node.type):
            lhs = self.visit(node.left)
            rhs = self.visit(node.right)
            name = type(node.op).__name__
            def make_op(l, r):
                return self.append(ir.Arith(node.op, l, r))
            return self._broadcast_binop(name, make_op, node.type, lhs, rhs,
                                         assign_to_lhs=False)
        elif builtins.is_numeric(node.type):
            lhs = self.visit(node.left)
            rhs = self.visit(node.right)
            if isinstance(node.op, (ast.LShift, ast.RShift)):
                # Check for negative shift amount.
                self._make_check(
                    self.append(ir.Compare(ast.GtE(loc=None), rhs, ir.Constant(0, rhs.type))),
                    lambda: self.alloc_exn(builtins.TException("ValueError"),
                        ir.Constant("shift amount must be nonnegative", builtins.TStr())),
                    loc=node.right.loc)
            elif isinstance(node.op, (ast.Div, ast.FloorDiv, ast.Mod)):
                self._make_check(
                    self.append(ir.Compare(ast.NotEq(loc=None), rhs, ir.Constant(0, rhs.type))),
                    lambda: self.alloc_exn(builtins.TException("ZeroDivisionError"),
                        ir.Constant("cannot divide by zero", builtins.TStr())),
                    loc=node.right.loc)

            return self.append(ir.Arith(node.op, lhs, rhs))
        elif isinstance(node.op, ast.Add): # list + list, tuple + tuple, str + str
            lhs, rhs = self.visit(node.left), self.visit(node.right)
            if types.is_tuple(node.left.type) and types.is_tuple(node.right.type):
                elts = []
                for index, elt in enumerate(node.left.type.elts):
                    elts.append(self.append(ir.GetAttr(lhs, index)))
                for index, elt in enumerate(node.right.type.elts):
                    elts.append(self.append(ir.GetAttr(rhs, index)))
                return self.append(ir.Alloc(elts, node.type))
            elif builtins.is_listish(node.left.type) and builtins.is_listish(node.right.type):
                lhs_length = self.iterable_len(lhs)
                rhs_length = self.iterable_len(rhs)

                result_length = self.append(ir.Arith(ast.Add(loc=None), lhs_length, rhs_length))
                result = self.append(ir.Alloc([result_length], node.type))

                # Copy lhs
                def body_gen(index):
                    elt = self.append(ir.GetElem(lhs, index))
                    self.append(ir.SetElem(result, index, elt))
                    return self.append(ir.Arith(ast.Add(loc=None), index,
                                                ir.Constant(1, self._size_type)))
                self._make_loop(ir.Constant(0, self._size_type),
                    lambda index: self.append(ir.Compare(ast.Lt(loc=None), index, lhs_length)),
                    body_gen)

                # Copy rhs
                def body_gen(index):
                    elt = self.append(ir.GetElem(rhs, index))
                    result_index = self.append(ir.Arith(ast.Add(loc=None), index, lhs_length))
                    self.append(ir.SetElem(result, result_index, elt))
                    return self.append(ir.Arith(ast.Add(loc=None), index,
                                                ir.Constant(1, self._size_type)))
                self._make_loop(ir.Constant(0, self._size_type),
                    lambda index: self.append(ir.Compare(ast.Lt(loc=None), index, rhs_length)),
                    body_gen)

                return result
            else:
                assert False
        elif isinstance(node.op, ast.Mult): # list * int, int * list
            lhs, rhs = self.visit(node.left), self.visit(node.right)
            if builtins.is_listish(lhs.type) and builtins.is_int(rhs.type):
                lst, num = lhs, rhs
            elif builtins.is_int(lhs.type) and builtins.is_listish(rhs.type):
                lst, num = rhs, lhs
            else:
                assert False

            lst_length = self.iterable_len(lst)

            result_length = self.append(ir.Arith(ast.Mult(loc=None), lst_length, num))
            result = self.append(ir.Alloc([result_length], node.type))

            # num times...
            def body_gen(num_index):
                # ... copy the list
                def body_gen(lst_index):
                    elt = self.append(ir.GetElem(lst, lst_index))
                    base_index = self.append(ir.Arith(ast.Mult(loc=None),
                                                      num_index, lst_length))
                    result_index = self.append(ir.Arith(ast.Add(loc=None),
                                                        base_index, lst_index))
                    self.append(ir.SetElem(result, base_index, elt))
                    return self.append(ir.Arith(ast.Add(loc=None), lst_index,
                                                ir.Constant(1, self._size_type)))
                self._make_loop(ir.Constant(0, self._size_type),
                    lambda index: self.append(ir.Compare(ast.Lt(loc=None), index, lst_length)),
                    body_gen)

                return self.append(ir.Arith(ast.Add(loc=None), num_index,
                                            ir.Constant(1, self._size_type)))
            self._make_loop(ir.Constant(0, self._size_type),
                lambda index: self.append(ir.Compare(ast.Lt(loc=None), index, num)),
                body_gen)

            return result
        else:
            assert False

    def polymorphic_compare_pair_order(self, op, lhs, rhs):
        if builtins.is_none(lhs.type) and builtins.is_none(rhs.type):
            return self.append(ir.Compare(op, lhs, rhs))
        elif builtins.is_numeric(lhs.type) and builtins.is_numeric(rhs.type):
            return self.append(ir.Compare(op, lhs, rhs))
        elif builtins.is_bool(lhs.type) and builtins.is_bool(rhs.type):
            return self.append(ir.Compare(op, lhs, rhs))
        elif types.is_tuple(lhs.type) and types.is_tuple(rhs.type):
            result = None
            for index in range(len(lhs.type.elts)):
                lhs_elt = self.append(ir.GetAttr(lhs, index))
                rhs_elt = self.append(ir.GetAttr(rhs, index))
                elt_result = self.polymorphic_compare_pair(op, lhs_elt, rhs_elt)
                if result is None:
                    result = elt_result
                else:
                    result = self.append(ir.Select(result, elt_result,
                                                   ir.Constant(False, builtins.TBool())))
            return result
        elif builtins.is_listish(lhs.type) and builtins.is_listish(rhs.type):
            head = self.current_block
            lhs_length = self.iterable_len(lhs)
            rhs_length = self.iterable_len(rhs)
            compare_length = self.append(ir.Compare(op, lhs_length, rhs_length))
            eq_length = self.append(ir.Compare(ast.Eq(loc=None), lhs_length, rhs_length))

            # If the length is the same, compare element-by-element
            # and break when the comparison result is false
            loop_head = self.add_block("compare.head")
            self.current_block = loop_head
            index_phi = self.append(ir.Phi(self._size_type))
            index_phi.add_incoming(ir.Constant(0, self._size_type), head)
            loop_cond = self.append(ir.Compare(ast.Lt(loc=None), index_phi, lhs_length))

            loop_body = self.add_block("compare.body")
            self.current_block = loop_body
            lhs_elt = self.append(ir.GetElem(lhs, index_phi))
            rhs_elt = self.append(ir.GetElem(rhs, index_phi))
            body_result = self.polymorphic_compare_pair(op, lhs_elt, rhs_elt)
            body_end = self.current_block

            loop_body2 = self.add_block("compare.body2")
            self.current_block = loop_body2
            index_next = self.append(ir.Arith(ast.Add(loc=None), index_phi,
                                              ir.Constant(1, self._size_type)))
            self.append(ir.Branch(loop_head))
            index_phi.add_incoming(index_next, loop_body2)

            tail = self.add_block("compare.tail")
            self.current_block = tail
            phi = self.append(ir.Phi(builtins.TBool()))
            head.append(ir.BranchIf(eq_length, loop_head, tail))
            phi.add_incoming(compare_length, head)
            loop_head.append(ir.BranchIf(loop_cond, loop_body, tail))
            phi.add_incoming(ir.Constant(True, builtins.TBool()), loop_head)
            body_end.append(ir.BranchIf(body_result, loop_body2, tail))
            phi.add_incoming(body_result, body_end)

            if isinstance(op, ast.NotEq):
                result = self.append(ir.Select(phi,
                    ir.Constant(False, builtins.TBool()), ir.Constant(True, builtins.TBool())))
            else:
                result = phi

            return result
        else:
            loc = lhs.loc
            loc.end = rhs.loc.end
            diag = diagnostic.Diagnostic("error",
                "Custom object comparison is not supported",
                {},
                loc)
            self.engine.process(diag)

    def polymorphic_compare_pair_inclusion(self, needle, haystack):
        if builtins.is_range(haystack.type):
            # Optimized range `in` operator
            start       = self.append(ir.GetAttr(haystack, "start"))
            stop        = self.append(ir.GetAttr(haystack, "stop"))
            step        = self.append(ir.GetAttr(haystack, "step"))
            after_start = self.append(ir.Compare(ast.GtE(loc=None), needle, start))
            after_stop  = self.append(ir.Compare(ast.Lt(loc=None), needle, stop))
            from_start  = self.append(ir.Arith(ast.Sub(loc=None), needle, start))
            mod_step    = self.append(ir.Arith(ast.Mod(loc=None), from_start, step))
            on_step     = self.append(ir.Compare(ast.Eq(loc=None), mod_step,
                                                 ir.Constant(0, mod_step.type)))
            result      = self.append(ir.Select(after_start, after_stop,
                                                ir.Constant(False, builtins.TBool())))
            result      = self.append(ir.Select(result, on_step,
                                                ir.Constant(False, builtins.TBool())))
        elif builtins.is_iterable(haystack.type):
            length = self.iterable_len(haystack)

            cmp_result = loop_body2 = None
            def body_gen(index):
                nonlocal cmp_result, loop_body2

                elt = self.iterable_get(haystack, index)
                cmp_result = self.polymorphic_compare_pair(ast.Eq(loc=None), needle, elt)

                loop_body2 = self.add_block("compare.body")
                self.current_block = loop_body2
                return self.append(ir.Arith(ast.Add(loc=None), index,
                                            ir.Constant(1, length.type)))
            loop_head, loop_body, loop_tail = \
                self._make_loop(ir.Constant(0, length.type),
                    lambda index: self.append(ir.Compare(ast.Lt(loc=None), index, length)),
                    body_gen, name="compare")

            loop_body.append(ir.BranchIf(cmp_result, loop_tail, loop_body2))
            phi = loop_tail.prepend(ir.Phi(builtins.TBool()))
            phi.add_incoming(ir.Constant(False, builtins.TBool()), loop_head)
            phi.add_incoming(ir.Constant(True, builtins.TBool()), loop_body)

            result = phi
        else:
            loc = needle.loc
            loc.end = haystack.loc.end
            diag = diagnostic.Diagnostic("error",
                "Custom object inclusion test is not supported",
                {},
                loc)
            self.engine.process(diag)

        return result

    def invert(self, value):
        return self.append(ir.Select(value,
                    ir.Constant(False, builtins.TBool()),
                    ir.Constant(True, builtins.TBool())))

    def polymorphic_compare_pair(self, op, lhs, rhs):
        if isinstance(op, (ast.Is, ast.IsNot)):
            # The backend will handle equality of aggregates.
            return self.append(ir.Compare(op, lhs, rhs))
        elif isinstance(op, ast.In):
            return self.polymorphic_compare_pair_inclusion(lhs, rhs)
        elif isinstance(op, ast.NotIn):
            result = self.polymorphic_compare_pair_inclusion(lhs, rhs)
            return self.invert(result)
        elif isinstance(op, (ast.Eq, ast.Lt, ast.LtE, ast.Gt, ast.GtE)):
            return self.polymorphic_compare_pair_order(op, lhs, rhs)
        elif isinstance(op, ast.NotEq):
            result = self.polymorphic_compare_pair_order(ast.Eq(loc=op.loc), lhs, rhs)
            return self.invert(result)
        else:
            assert False

    def visit_CompareT(self, node):
        # Essentially a sequence of `and`s performed over results
        # of comparisons.
        blocks = []
        lhs = self.visit(node.left)
        for op, rhs_node in zip(node.ops, node.comparators):
            result_head = self.current_block
            rhs = self.visit(rhs_node)
            result = self.polymorphic_compare_pair(op, lhs, rhs)
            result_tail = self.current_block

            blocks.append((result, result_head, result_tail))
            self.current_block = self.add_block("compare.seq")
            lhs = rhs

        tail = self.current_block
        phi = self.append(ir.Phi(node.type))
        for ((result, result_head, result_tail), (next_result_head, next_result_tail)) in \
                    zip(blocks, [(h,t) for (v,h,t) in blocks[1:]] + [(tail, tail)]):
            phi.add_incoming(result, result_tail)
            if next_result_head != tail:
                result_tail.append(ir.BranchIf(result, next_result_head, tail))
            else:
                result_tail.append(ir.Branch(tail))
        return phi

    # Keep this function with builtins.TException.attributes.
    def alloc_exn(self, typ, message=None, param0=None, param1=None,
                  param2=None, nomsgcheck=False):
        typ = typ.find()
        name = "{}:{}".format(typ.id, typ.name)
        name_id = self.embedding_map.store_str(name)
        attributes = [
            ir.Constant(name_id,        builtins.TInt32()),   # typeinfo
            ir.Constant("<not thrown>", builtins.TStr()),   # file
            ir.Constant(0,              builtins.TInt32()), # line
            ir.Constant(0,              builtins.TInt32()), # column
            ir.Constant("<not thrown>", builtins.TStr()),   # function
        ]

        if message is None:
            attributes.append(ir.Constant(typ.name, builtins.TStr()))
        elif isinstance(message, ir.Constant) or nomsgcheck:
            attributes.append(message)                  # message
        else:
            diag = diagnostic.Diagnostic(
                "error",
                "only constant exception messages are supported",
                {},
                self.current_loc if message.loc is None else message.loc
            )
            self.engine.process(diag)

        param_type = builtins.TInt64()
        for param in [param0, param1, param2]:
            if param is None:
                attributes.append(ir.Constant(0, builtins.TInt64()))
            else:
                if param.type != param_type:
                    param = self.append(ir.Coerce(param, param_type))
                attributes.append(param)                  # paramN, N=0:2

        return self.append(ir.Alloc(attributes, typ))

    def visit_builtin_call(self, node):
        # A builtin by any other name... Ignore node.func, just use the type.
        typ = node.func.type
        if types.is_builtin(typ, "bool"):
            if len(node.args) == 0 and len(node.keywords) == 0:
                return ir.Constant(False, builtins.TBool())
            elif len(node.args) == 1 and len(node.keywords) == 0:
                arg = self.visit(node.args[0])
                return self.coerce_to_bool(arg)
            else:
                assert False
        elif types.is_builtin(typ, "int") or \
                types.is_builtin(typ, "int32") or types.is_builtin(typ, "int64"):
            if len(node.args) == 0 and len(node.keywords) == 0:
                return ir.Constant(0, node.type)
            elif len(node.args) == 1 and \
                    (len(node.keywords) == 0 or \
                     len(node.keywords) == 1 and node.keywords[0].arg == 'width'):
                # The width argument is purely type-level
                arg = self.visit(node.args[0])
                return self.append(ir.Coerce(arg, node.type))
            else:
                assert False
        elif types.is_builtin(typ, "float"):
            if len(node.args) == 0 and len(node.keywords) == 0:
                return ir.Constant(0.0, builtins.TFloat())
            elif len(node.args) == 1 and len(node.keywords) == 0:
                arg = self.visit(node.args[0])
                return self.append(ir.Coerce(arg, node.type))
            else:
                assert False
        elif (types.is_builtin(typ, "list") or
              types.is_builtin(typ, "bytearray") or types.is_builtin(typ, "bytes")):
            if len(node.args) == 0 and len(node.keywords) == 0:
                length = ir.Constant(0, builtins.TInt32())
                return self.append(ir.Alloc([length], node.type))
            elif len(node.args) == 1 and len(node.keywords) == 0:
                arg = self.visit(node.args[0])
                length = self.iterable_len(arg)
                result = self.append(ir.Alloc([length], node.type))

                def body_gen(index):
                    elt = self.iterable_get(arg, index)
                    elt = self.append(ir.Coerce(elt, builtins.get_iterable_elt(node.type)))
                    self.append(ir.SetElem(result, index, elt))
                    return self.append(ir.Arith(ast.Add(loc=None), index,
                                                ir.Constant(1, length.type)))
                self._make_loop(ir.Constant(0, length.type),
                    lambda index: self.append(ir.Compare(ast.Lt(loc=None), index, length)),
                    body_gen)

                return result
            else:
                assert False
        elif types.is_builtin(typ, "array"):
            if len(node.args) == 1 and len(node.keywords) in (0, 1):
                result_type = node.type.find()
                arg = self.visit(node.args[0])

                result_elt = result_type["elt"].find()
                num_dims = result_type["num_dims"].value

                # Derive shape from first element on each level (and fail later if the
                # array is in fact jagged).
                first_elt = None
                lengths = []
                for dim_idx in range(num_dims):
                    if first_elt is None:
                        first_elt = arg
                    else:
                        first_elt = self.iterable_get(first_elt,
                                                      ir.Constant(0, self._size_type))
                    lengths.append(self.iterable_len(first_elt))

                shape = self.append(ir.Alloc(lengths, result_type.attributes["shape"]))
                num_total_elts = self._get_total_array_len(shape)

                # Assign buffer from nested iterables.
                buffer = self.append(
                    ir.Alloc([num_total_elts], result_type.attributes["buffer"]))

                def assign_elems(outer_indices, indexed_arg):
                    if len(outer_indices) == num_dims:
                        dest_idx = self._get_array_offset(lengths, outer_indices)
                        coerced = self.append(ir.Coerce(indexed_arg, result_elt))
                        self.append(ir.SetElem(buffer, dest_idx, coerced))
                    else:
                        this_level_len = self.iterable_len(indexed_arg)
                        dim_idx = len(outer_indices)
                        if dim_idx > 0:
                            # Check for rectangularity (outermost index is never jagged,
                            # by definition).
                            result_len = self.append(ir.GetAttr(shape, dim_idx))
                            self._make_check(
                                self.append(ir.Compare(ast.Eq(loc=None), this_level_len, result_len)),
                                lambda a, b: self.alloc_exn(
                                    builtins.TException("ValueError"),
                                    ir.Constant(
                                        "arrays must be rectangular (lengths were {0} vs. {1})",
                                        builtins.TStr()), a, b),
                                params=[this_level_len, result_len],
                                loc=node.loc)

                        def body_gen(index):
                            elem = self.iterable_get(indexed_arg, index)
                            assign_elems(outer_indices + [index], elem)
                            return self.append(
                                ir.Arith(ast.Add(loc=None), index,
                                        ir.Constant(1, self._size_type)))
                        self._make_loop(
                            ir.Constant(0, self._size_type), lambda index: self.append(
                                ir.Compare(ast.Lt(loc=None), index, this_level_len)), body_gen)
                assign_elems([], arg)
                return self.append(ir.Alloc([buffer, shape], node.type))
            else:
                assert False
        elif types.is_builtin(typ, "range"):
            elt_typ = builtins.get_iterable_elt(node.type)
            if len(node.args) == 1 and len(node.keywords) == 0:
                max_arg = self.visit(node.args[0])
                return self.append(ir.Alloc([
                    ir.Constant(elt_typ.zero(), elt_typ),
                    max_arg,
                    ir.Constant(elt_typ.one(), elt_typ),
                ], node.type))
            elif len(node.args) == 2 and len(node.keywords) == 0:
                min_arg = self.visit(node.args[0])
                max_arg = self.visit(node.args[1])
                return self.append(ir.Alloc([
                    min_arg,
                    max_arg,
                    ir.Constant(elt_typ.one(), elt_typ),
                ], node.type))
            elif len(node.args) == 3 and len(node.keywords) == 0:
                min_arg = self.visit(node.args[0])
                max_arg = self.visit(node.args[1])
                step_arg = self.visit(node.args[2])
                return self.append(ir.Alloc([
                    min_arg,
                    max_arg,
                    step_arg,
                ], node.type))
            else:
                assert False
        elif types.is_builtin(typ, "len"):
            if len(node.args) == 1 and len(node.keywords) == 0:
                arg = self.visit(node.args[0])
                return self.iterable_len(arg)
            else:
                assert False
        elif types.is_builtin(typ, "round"):
            if len(node.args) == 1 and len(node.keywords) == 0:
                arg = self.visit(node.args[0])
                return self.append(ir.Builtin("round", [arg], node.type))
            else:
                assert False
        elif types.is_builtin(typ, "abs"):
            if len(node.args) == 1 and len(node.keywords) == 0:
                arg = self.visit(node.args[0])
                neg = self.append(
                    ir.Arith(ast.Sub(loc=None), ir.Constant(0, arg.type), arg))
                cond = self.append(
                    ir.Compare(ast.Lt(loc=None), arg, ir.Constant(0, arg.type)))
                return self.append(ir.Select(cond, neg, arg))
            else:
                assert False
        elif types.is_builtin(typ, "min"):
            if len(node.args) == 2 and len(node.keywords) == 0:
                arg0, arg1 = map(self.visit, node.args)
                cond = self.append(ir.Compare(ast.Lt(loc=None), arg0, arg1))
                return self.append(ir.Select(cond, arg0, arg1))
            else:
                assert False
        elif types.is_builtin(typ, "max"):
            if len(node.args) == 2 and len(node.keywords) == 0:
                arg0, arg1 = map(self.visit, node.args)
                cond = self.append(ir.Compare(ast.Gt(loc=None), arg0, arg1))
                return self.append(ir.Select(cond, arg0, arg1))
            else:
                assert False
        elif types.is_builtin(typ, "make_array"):
            if len(node.args) == 2 and len(node.keywords) == 0:
                arg0, arg1 = map(self.visit, node.args)

                num_dims = node.type.find()["num_dims"].value
                if types.is_tuple(arg0.type):
                    lens = [self.append(ir.GetAttr(arg0, i)) for i in range(num_dims)]
                else:
                    assert num_dims == 1
                    lens = [arg0]

                shape = self._make_array_shape(lens)
                result, total_len = self._allocate_new_array(node.type.find()["elt"],
                                                             shape)

                def body_gen(index):
                    self.append(ir.SetElem(result, index, arg1))
                    return self.append(
                        ir.Arith(ast.Add(loc=None), index,
                                 ir.Constant(1, self._size_type)))

                self._make_loop(
                    ir.Constant(0, self._size_type), lambda index: self.append(
                        ir.Compare(ast.Lt(loc=None), index, total_len)), body_gen)
                return result
            else:
                assert False
        elif types.is_builtin(typ, "numpy.transpose"):
            if len(node.args) == 1 and len(node.keywords) == 0:
                arg, = map(self.visit, node.args)

                num_dims = arg.type.find()["num_dims"].value
                if num_dims == 1:
                    # No-op as per NumPy semantics.
                    return arg
                assert num_dims == 2
                arg_shape = self.append(ir.GetAttr(arg, "shape"))
                dim0 = self.append(ir.GetAttr(arg_shape, 0))
                dim1 = self.append(ir.GetAttr(arg_shape, 1))
                shape = self._make_array_shape([dim1, dim0])
                result, _ = self._allocate_new_array(node.type.find()["elt"], shape)
                arg_buffer = self.append(ir.GetAttr(arg, "buffer"))
                result_buffer = self.append(ir.GetAttr(result, "buffer"))

                def outer_gen(idx1):
                    arg_base = self.append(ir.Offset(arg_buffer, idx1))
                    result_offset = self.append(ir.Arith(ast.Mult(loc=None), idx1,
                                                         dim0))
                    result_base = self.append(ir.Offset(result_buffer, result_offset))

                    def inner_gen(idx0):
                        arg_offset = self.append(
                            ir.Arith(ast.Mult(loc=None), idx0, dim1))
                        val = self.append(ir.GetElem(arg_base, arg_offset))
                        self.append(ir.SetElem(result_base, idx0, val))
                        return self.append(
                            ir.Arith(ast.Add(loc=None), idx0, ir.Constant(1,
                                                                          idx0.type)))

                    self._make_loop(
                        ir.Constant(0, self._size_type), lambda idx0: self.append(
                            ir.Compare(ast.Lt(loc=None), idx0, dim0)), inner_gen)
                    return self.append(
                        ir.Arith(ast.Add(loc=None), idx1, ir.Constant(1, idx1.type)))

                self._make_loop(
                    ir.Constant(0, self._size_type),
                    lambda idx1: self.append(ir.Compare(ast.Lt(loc=None), idx1, dim1)),
                    outer_gen)
                return result
            else:
                assert False
        elif types.is_builtin(typ, "print"):
            self.polymorphic_print([self.visit(arg) for arg in node.args],
                                   separator=" ", suffix="\n")
            return ir.Constant(None, builtins.TNone())
        elif types.is_builtin(typ, "rtio_log"):
            prefix, *args = node.args
            self.polymorphic_print([self.visit(prefix)],
                                   separator=" ", suffix="\x1E", as_rtio=True)
            self.polymorphic_print([self.visit(arg) for arg in args],
                                   separator=" ", suffix="\x1D", as_rtio=True)
            return ir.Constant(None, builtins.TNone())
        elif types.is_builtin(typ, "delay"):
            if len(node.args) == 1 and len(node.keywords) == 0:
                arg = self.visit(node.args[0])
                arg_mu_float = self.append(ir.Arith(ast.Div(loc=None), arg, self.ref_period))
                arg_mu = self.append(ir.Builtin("round", [arg_mu_float], builtins.TInt64()))
                return self.append(ir.Builtin("delay_mu", [arg_mu], builtins.TNone()))
            else:
                assert False
        elif types.is_builtin(typ, "now_mu") or types.is_builtin(typ, "delay_mu") \
                or types.is_builtin(typ, "at_mu"):
            return self.append(ir.Builtin(typ.name,
                                          [self.visit(arg) for arg in node.args], node.type))
        elif types.is_builtin(typ, "subkernel_await"):
            if len(node.args) == 2 and len(node.keywords) == 0:
                fn = node.args[0].type
                timeout = self.visit(node.args[1])
            elif len(node.args) == 1 and len(node.keywords) == 0:
                fn = node.args[0].type
                timeout = ir.Constant(-1, builtins.TInt64())
            else:
                assert False
            if types.is_method(fn):
                fn = types.get_method_function(fn)
            sid = ir.Constant(fn.sid, builtins.TInt32())
            if not builtins.is_none(fn.ret):
                ret = self.append(ir.Builtin("subkernel_retrieve_return", [sid, timeout], fn.ret))
            else:
                ret = ir.Constant(None, builtins.TNone())
            self.append(ir.Builtin("subkernel_await_finish", [sid, timeout], builtins.TNone()))
            return ret
        elif types.is_builtin(typ, "subkernel_preload"):
            if len(node.args) == 1 and len(node.keywords) == 0:
                fn = node.args[0].type
            else:
                assert False
            if types.is_method(fn):
                fn = types.get_method_function(fn)
            sid = ir.Constant(fn.sid, builtins.TInt32())
            dest = ir.Constant(fn.destination, builtins.TInt32())
            return self.append(ir.Builtin("subkernel_preload", [sid, dest], builtins.TNone()))
        elif types.is_builtin(typ, "subkernel_send"):
            if len(node.args) == 3 and len(node.keywords) == 0:
                dest = self.visit(node.args[0])
                name = node.args[1].s
                value = self.visit(node.args[2])
            else:
                assert False
            msg_id, msg = self.embedding_map.store_subkernel_message(name, value.type, "send", node.loc)
            msg_id = ir.Constant(msg_id, builtins.TInt32())
            if value.type != msg.value_type:
                diag = diagnostic.Diagnostic("error",
                    "type mismatch for subkernel message '{name}', receiver expects {recv} while sending {send}",
                    {"name": name, "recv": msg.value_type, "send": value.type},
                    node.loc)
                self.engine.process(diag)
            return self.append(ir.Builtin("subkernel_send", [msg_id, dest, value], builtins.TNone()))
        elif types.is_builtin(typ, "subkernel_recv"):
            if len(node.args) == 2 and len(node.keywords) == 0:
                name = node.args[0].s
                vartype = node.args[1].value
                timeout = ir.Constant(-1, builtins.TInt64())
            elif len(node.args) == 3 and len(node.keywords) == 0:
                name = node.args[0].s
                vartype = node.args[1].value
                timeout = self.visit(node.args[2])
            else:
                assert False
            msg_id, msg = self.embedding_map.store_subkernel_message(name, vartype, "recv", node.loc)
            msg_id = ir.Constant(msg_id, builtins.TInt32())
            if vartype != msg.value_type:
                diag = diagnostic.Diagnostic("error",
                    "type mismatch for subkernel message '{name}', receiver expects {recv} while sending {send}",
                    {"name": name, "recv": vartype, "send": msg.value_type},
                    node.loc)
                self.engine.process(diag)
            return self.append(ir.Builtin("subkernel_recv", [msg_id, timeout], vartype))
        elif types.is_exn_constructor(typ):
            return self.alloc_exn(node.type, *[self.visit(arg_node) for arg_node in node.args])
        elif types.is_constructor(typ):
            return self.append(ir.Alloc([], typ.instance))
        else:
            diag = diagnostic.Diagnostic("error",
                "builtin function '{name}' cannot be used in this context",
                {"name": typ.find().name},
                node.loc)
            self.engine.process(diag)

    def _user_call(self, callee, positional, keywords, arg_exprs={}, remote_fn=False):
        if types.is_function(callee.type) or types.is_rpc(callee.type) or types.is_subkernel(callee.type):
            func     = callee
            self_arg = None
            fn_typ   = callee.type
            offset   = 0
        elif types.is_method(callee.type):
            func     = self.append(ir.GetAttr(callee, "__func__",
                                              name="{}.ENV".format(callee.name)))
            self_arg = self.append(ir.GetAttr(callee, "__self__",
                                              name="{}.SLF".format(callee.name)))
            fn_typ   = types.get_method_function(callee.type)
            offset   = 1
        else:
            assert False

        if types.is_rpc(fn_typ) or types.is_subkernel(fn_typ):
            if self_arg is None or types.is_subkernel(fn_typ):
                # self is not passed to subkernels by remote
                args = positional
            elif self_arg is not None:
                args = [self_arg] + positional

            for keyword in keywords:
                arg = keywords[keyword]
                args.append(self.append(ir.Alloc([ir.Constant(keyword, builtins.TStr()), arg],
                                                 ir.TKeyword(arg.type))))
        elif remote_fn:
            assert self_arg is None
            assert len(fn_typ.args) >= len(positional)
            assert len(keywords) == 0  # no keyword support
            args = [None] * fn_typ.arity()
            index = 0
            # fill in first available args
            for arg in positional:
                args[index] = arg
                index += 1

            # remaining args are received through DRTIO
            if index < len(args):
                # min/max args received remotely (minus already filled)
                offset = index
                min_args = ir.Constant(len(fn_typ.args)-offset, builtins.TInt8())
                max_args = ir.Constant(fn_typ.arity()-offset, builtins.TInt8())

                arg_types = list(fn_typ.args.items())[offset:] 
                arg_type_list = [a[1] for a in arg_types] + [a[1] for a in fn_typ.optargs.items()]
                rcvd_count = self.append(ir.SubkernelAwaitArgs([min_args, max_args], arg_type_list))
                # obligatory arguments
                for arg_name, arg_type in arg_types:
                    args[index] = self.append(ir.GetArgFromRemote(arg_name, arg_type,
                                            name="ARG.{}".format(arg_name)))
                    index += 1

                # optional arguments
                for optarg_name, optarg_type in fn_typ.optargs.items():
                    idx = ir.Constant(index-offset, builtins.TInt8())
                    args[index] = \
                        self.append(ir.GetOptArgFromRemote(optarg_name, optarg_type, rcvd_count, idx))
                    index += 1

        else:
            args = [None] * (len(fn_typ.args) + len(fn_typ.optargs))

            for index, arg in enumerate(positional):
                if index + offset < len(fn_typ.args):
                    args[index + offset] = arg
                else:
                    args[index + offset] = self.append(ir.Alloc([arg], ir.TOption(arg.type)))

            for keyword in keywords:
                arg = keywords[keyword]
                if keyword in fn_typ.args:
                    for index, arg_name in enumerate(fn_typ.args):
                        if keyword == arg_name:
                            assert args[index] is None
                            args[index] = arg
                            break
                elif keyword in fn_typ.optargs:
                    for index, optarg_name in enumerate(fn_typ.optargs):
                        if keyword == optarg_name:
                            assert args[len(fn_typ.args) + index] is None
                            args[len(fn_typ.args) + index] = \
                                    self.append(ir.Alloc([arg], ir.TOption(arg.type)))
                            break

            for index, optarg_name in enumerate(fn_typ.optargs):
                if args[len(fn_typ.args) + index] is None:
                    args[len(fn_typ.args) + index] = \
                            self.append(ir.Alloc([], ir.TOption(fn_typ.optargs[optarg_name])))

            if self_arg is not None:
                assert args[0] is None
                args[0] = self_arg

            assert None not in args

        if self.unwind_target is None or \
                types.is_external_function(callee.type) and "nounwind" in callee.type.flags:
            insn = self.append(ir.Call(func, args, arg_exprs))
        else:
            after_invoke = self.add_block("invoke")
            insn = self.append(ir.Invoke(func, args, arg_exprs,
                                         after_invoke, self.unwind_target))
            self.current_block = after_invoke

        return insn

    def visit_CallT(self, node):
        if not types.is_builtin(node.func.type):
            callee   = self.visit(node.func)
            args     = [self.visit(arg_node) for arg_node in node.args]
            keywords = {kw_node.arg: self.visit(kw_node.value) for kw_node in node.keywords}

        if node.iodelay is not None and not iodelay.is_const(node.iodelay, 0):
            before_delay = self.current_block
            during_delay = self.add_block("delay.head")
            before_delay.append(ir.Branch(during_delay))
            self.current_block = during_delay

        if types.is_builtin(node.func.type):
            insn = self.visit_builtin_call(node)
        elif (types.is_broadcast_across_arrays(node.func.type) and len(args) >= 1
              and any(builtins.is_array(arg.type) for arg in args)):
            # The iodelay machinery set up in the surrounding code was
            # deprecated/a relic from the past when array broadcasting support
            # was added, so no attempt to keep the delay tracking intact is
            # made.
            def make_call(*args):
                return self._user_call(ir.Constant(None, callee.type), args, {},
                                       node.arg_exprs)
            # TODO: Generate more generically if non-externals are allowed.
            name = node.func.type.find().name

            if len(args) == 1:
                shape = self.append(ir.GetAttr(args[0], "shape"))
                result, _ = self._allocate_new_array(node.type.find()["elt"], shape)
                func = self._get_array_unaryop(name, make_call, node.type, args[0].type)
                self._invoke_arrayop(func, [result, args[0]])
                insn = result
            elif len(args) == 2:
                insn = self._broadcast_binop(name, make_call, node.type, *args,
                                             assign_to_lhs=False)
            else:
                assert False, "Broadcasting for {} arguments not implemented".format(len)
        else:
            remote_fn = getattr(node, "remote_fn", False)
            insn = self._user_call(callee, args, keywords, node.arg_exprs, remote_fn)
            if isinstance(node.func, asttyped.AttributeT):
                attr_node = node.func
                self.method_map[(attr_node.value.type.find(),
                                 attr_node.attr)].append(insn)

        if node.iodelay is not None and not iodelay.is_const(node.iodelay, 0):
            after_delay = self.add_block("delay.tail")
            self.append(ir.Delay(node.iodelay, insn, after_delay))
            self.current_block = after_delay

        return insn

    def visit_QuoteT(self, node):
        return self.append(ir.Quote(node.value, node.type))

    def _get_raise_assert_func(self):
        """Emit the helper function that constructs AssertionErrors and raises
        them, if it does not already exist in the current module.

        A separate function is used for code size reasons. (This could also be
        compiled into a stand-alone support library instead.)
        """
        if self.raise_assert_func:
            return self.raise_assert_func
        try:
            msg = ir.Argument(builtins.TStr(), "msg")
            file = ir.Argument(builtins.TStr(), "file")
            line = ir.Argument(builtins.TInt32(), "line")
            col = ir.Argument(builtins.TInt32(), "col")
            function = ir.Argument(builtins.TStr(), "function")

            args = [msg, file, line, col, function]
            typ = types.TFunction(args=OrderedDict([(arg.name, arg.type)
                                                    for arg in args]),
                                  optargs=OrderedDict(),
                                  ret=builtins.TNone())
            env = ir.TEnvironment(name="raise", vars={})
            env_arg = ir.EnvironmentArgument(env, "ARG.ENV")
            func = ir.Function(typ, "_artiq_raise_assert", [env_arg] + args)
            func.is_internal = True
            func.is_cold = True
            func.is_generated = True
            self.functions.append(func)
            old_func, self.current_function = self.current_function, func

            entry = self.add_block("entry")
            old_block, self.current_block = self.current_block, entry
            old_final_branch, self.final_branch = self.final_branch, None
            old_unwind, self.unwind_target = self.unwind_target, None

            exn = self.alloc_exn(builtins.TException("AssertionError"),
                                 message=msg, nomsgcheck=True)
            self.append(ir.SetAttr(exn, "#__file__", file))
            self.append(ir.SetAttr(exn, "#__line__", line))
            self.append(ir.SetAttr(exn, "#__col__", col))
            self.append(ir.SetAttr(exn, "#__func__", function))
            self.append(ir.Raise(exn))
        finally:
            self.current_function = old_func
            self.current_block = old_block
            self.final_branch = old_final_branch
            self.unwind_target = old_unwind

        self.raise_assert_func = func
        return self.raise_assert_func

    def visit_Assert(self, node):
        cond = self.visit(node.test)
        head = self.current_block

        if_failed = self.current_block = self.add_block("assert.fail")
        text = str(node.msg.s) if node.msg else "AssertionError"
        msg = ir.Constant(text, builtins.TStr())
        loc_file = ir.Constant(node.loc.source_buffer.name, builtins.TStr())
        loc_line = ir.Constant(node.loc.line(), builtins.TInt32())
        loc_column = ir.Constant(node.loc.column(), builtins.TInt32())
        loc_function = ir.Constant(".".join(self.name), builtins.TStr())
        self._invoke_raising_func(self._get_raise_assert_func(), [
            msg, loc_file, loc_line, loc_column, loc_function
        ], "assert.fail")

        tail = self.current_block = self.add_block("assert.tail")
        self.append(ir.BranchIf(cond, tail, if_failed), block=head)

    def polymorphic_print(self, values, separator, suffix="", as_repr=False, as_rtio=False):
        def printf(format_string, *args):
            format = ir.Constant(format_string, builtins.TStr())
            if as_rtio:
                self.append(ir.Builtin("rtio_log", [format, *args], builtins.TNone()))
            else:
                self.append(ir.Builtin("printf", [format, *args], builtins.TNone()))

        format_string = ""
        args = []
        def flush():
            nonlocal format_string, args
            if format_string != "":
                printf(format_string + "\x00", *args)
                format_string = ""
                args = []

        for value in values:
            if format_string != "":
                format_string += separator

            if types.is_tuple(value.type):
                format_string += "("; flush()
                self.polymorphic_print([self.append(ir.GetAttr(value, index))
                                        for index in range(len(value.type.elts))],
                                       separator=", ", as_repr=True, as_rtio=as_rtio)
                if len(value.type.elts) == 1:
                    format_string += ",)"
                else:
                    format_string += ")"
            elif types.is_function(value.type):
                format_string += "<closure %p(%p)>"
                args.append(self.append(ir.GetAttr(value, '__code__')))
                args.append(self.append(ir.GetAttr(value, '__closure__')))
            elif builtins.is_none(value.type):
                format_string += "None"
            elif builtins.is_bool(value.type):
                format_string += "%.*s"
                args.append(self.append(ir.Select(value,
                                                  ir.Constant("True", builtins.TStr()),
                                                  ir.Constant("False", builtins.TStr()))))
            elif builtins.is_int(value.type):
                width = builtins.get_int_width(value.type)
                if width <= 32:
                    format_string += "%d"
                elif width <= 64:
                    format_string += "%lld"
                else:
                    assert False
                args.append(value)
            elif builtins.is_float(value.type):
                format_string += "%g"
                args.append(value)
            elif builtins.is_str(value.type):
                if as_repr:
                    format_string += "\"%.*s\""
                else:
                    format_string += "%.*s"
                args.append(value)
            elif builtins.is_listish(value.type):
                if builtins.is_list(value.type):
                    format_string += "["; flush()
                elif builtins.is_bytes(value.type):
                    format_string += "bytes(["; flush()
                elif builtins.is_bytearray(value.type):
                    format_string += "bytearray(["; flush()
                elif builtins.is_array(value.type):
                    format_string += "array(["; flush()
                else:
                    assert False

                length = self.iterable_len(value)
                last = self.append(ir.Arith(ast.Sub(loc=None), length, ir.Constant(1, length.type)))
                def body_gen(index):
                    elt = self.iterable_get(value, index)
                    self.polymorphic_print([elt], separator="", as_repr=True, as_rtio=as_rtio)
                    is_last = self.append(ir.Compare(ast.Lt(loc=None), index, last))
                    head = self.current_block

                    if_last = self.current_block = self.add_block("print.comma")
                    printf(", \x00")

                    tail = self.current_block = self.add_block("print.tail")
                    if_last.append(ir.Branch(tail))
                    head.append(ir.BranchIf(is_last, if_last, tail))

                    return self.append(ir.Arith(ast.Add(loc=None), index,
                                                ir.Constant(1, length.type)))
                self._make_loop(ir.Constant(0, length.type),
                    lambda index: self.append(ir.Compare(ast.Lt(loc=None), index, length)),
                    body_gen)

                if builtins.is_list(value.type):
                    format_string += "]"
                elif (builtins.is_bytes(value.type) or builtins.is_bytearray(value.type) or
                      builtins.is_array(value.type)):
                    format_string += "])"
            elif builtins.is_range(value.type):
                format_string += "range("; flush()

                start  = self.append(ir.GetAttr(value, "start"))
                stop   = self.append(ir.GetAttr(value, "stop"))
                step   = self.append(ir.GetAttr(value, "step"))
                self.polymorphic_print([start, stop, step], separator=", ", as_rtio=as_rtio)

                format_string += ")"
            elif builtins.is_exception(value.type):
                # message may not be an actual string...
                # so we cannot really print it
                name    = self.append(ir.GetAttr(value, "#__name__"))
                param1  = self.append(ir.GetAttr(value, "#__param0__"))
                param2  = self.append(ir.GetAttr(value, "#__param1__"))
                param3  = self.append(ir.GetAttr(value, "#__param2__"))

                format_string += "%ld(%lld, %lld, %lld)"
                args += [name, param1, param2, param3]
            else:
                assert False

        format_string += suffix
        flush()