Major fixes and new features

venv/lib/python3.12/site-packages/mypyc/codegen/cstring.py

@@ -0,0 +1,54 @@
+"""Encode valid C string literals from Python strings.
+
+If a character is not allowed in C string literals, it is either emitted
+as a simple escape sequence (e.g. '\\n'), or an octal escape sequence
+with exactly three digits ('\\oXXX'). Question marks are escaped to
+prevent trigraphs in the string literal from being interpreted. Note
+that '\\?' is an invalid escape sequence in Python.
+
+Consider the string literal "AB\\xCDEF". As one would expect, Python
+parses it as ['A', 'B', 0xCD, 'E', 'F']. However, the C standard
+specifies that all hexadecimal digits immediately following '\\x' will
+be interpreted as part of the escape sequence. Therefore, it is
+unexpectedly parsed as ['A', 'B', 0xCDEF].
+
+Emitting ("AB\\xCD" "EF") would avoid this behaviour. However, we opt
+for simplicity and use octal escape sequences instead. They do not
+suffer from the same issue as they are defined to parse at most three
+octal digits.
+"""
+
+from __future__ import annotations
+
+import string
+from typing import Final
+
+CHAR_MAP: Final = [f"\\{i:03o}" for i in range(256)]
+
+# It is safe to use string.printable as it always uses the C locale.
+for c in string.printable:
+    CHAR_MAP[ord(c)] = c
+
+# These assignments must come last because we prioritize simple escape
+# sequences over any other representation.
+for c in ("'", '"', "\\", "a", "b", "f", "n", "r", "t", "v"):
+    escaped = f"\\{c}"
+    decoded = escaped.encode("ascii").decode("unicode_escape")
+    CHAR_MAP[ord(decoded)] = escaped
+
+# This escape sequence is invalid in Python.
+CHAR_MAP[ord("?")] = r"\?"
+
+
+def encode_bytes_as_c_string(b: bytes) -> str:
+    """Produce contents of a C string literal for a byte string, without quotes."""
+    escaped = "".join([CHAR_MAP[i] for i in b])
+    return escaped
+
+
+def c_string_initializer(value: bytes) -> str:
+    """Create initializer for a C char[]/ char * variable from a string.
+
+    For example, if value if b'foo', the result would be '"foo"'.
+    """
+    return '"' + encode_bytes_as_c_string(value) + '"'

venv/lib/python3.12/site-packages/mypyc/codegen/emitfunc.py

@@ -0,0 +1,841 @@
+"""Code generation for native function bodies."""
+
+from __future__ import annotations
+
+from typing import Final
+
+from mypyc.analysis.blockfreq import frequently_executed_blocks
+from mypyc.codegen.emit import DEBUG_ERRORS, Emitter, TracebackAndGotoHandler, c_array_initializer
+from mypyc.common import MODULE_PREFIX, NATIVE_PREFIX, REG_PREFIX, STATIC_PREFIX, TYPE_PREFIX
+from mypyc.ir.class_ir import ClassIR
+from mypyc.ir.func_ir import FUNC_CLASSMETHOD, FUNC_STATICMETHOD, FuncDecl, FuncIR, all_values
+from mypyc.ir.ops import (
+    ERR_FALSE,
+    NAMESPACE_MODULE,
+    NAMESPACE_STATIC,
+    NAMESPACE_TYPE,
+    Assign,
+    AssignMulti,
+    BasicBlock,
+    Box,
+    Branch,
+    Call,
+    CallC,
+    Cast,
+    ComparisonOp,
+    ControlOp,
+    DecRef,
+    Extend,
+    Float,
+    FloatComparisonOp,
+    FloatNeg,
+    FloatOp,
+    GetAttr,
+    GetElementPtr,
+    Goto,
+    IncRef,
+    InitStatic,
+    Integer,
+    IntOp,
+    KeepAlive,
+    LoadAddress,
+    LoadErrorValue,
+    LoadGlobal,
+    LoadLiteral,
+    LoadMem,
+    LoadStatic,
+    MethodCall,
+    Op,
+    OpVisitor,
+    RaiseStandardError,
+    Register,
+    Return,
+    SetAttr,
+    SetMem,
+    Truncate,
+    TupleGet,
+    TupleSet,
+    Unbox,
+    Unreachable,
+    Value,
+)
+from mypyc.ir.pprint import generate_names_for_ir
+from mypyc.ir.rtypes import (
+    RArray,
+    RStruct,
+    RTuple,
+    RType,
+    is_int32_rprimitive,
+    is_int64_rprimitive,
+    is_int_rprimitive,
+    is_pointer_rprimitive,
+    is_tagged,
+)
+
+
+def native_function_type(fn: FuncIR, emitter: Emitter) -> str:
+    args = ", ".join(emitter.ctype(arg.type) for arg in fn.args) or "void"
+    ret = emitter.ctype(fn.ret_type)
+    return f"{ret} (*)({args})"
+
+
+def native_function_header(fn: FuncDecl, emitter: Emitter) -> str:
+    args = []
+    for arg in fn.sig.args:
+        args.append(f"{emitter.ctype_spaced(arg.type)}{REG_PREFIX}{arg.name}")
+
+    return "{ret_type}{name}({args})".format(
+        ret_type=emitter.ctype_spaced(fn.sig.ret_type),
+        name=emitter.native_function_name(fn),
+        args=", ".join(args) or "void",
+    )
+
+
+def generate_native_function(
+    fn: FuncIR, emitter: Emitter, source_path: str, module_name: str
+) -> None:
+    declarations = Emitter(emitter.context)
+    names = generate_names_for_ir(fn.arg_regs, fn.blocks)
+    body = Emitter(emitter.context, names)
+    visitor = FunctionEmitterVisitor(body, declarations, source_path, module_name)
+
+    declarations.emit_line(f"{native_function_header(fn.decl, emitter)} {{")
+    body.indent()
+
+    for r in all_values(fn.arg_regs, fn.blocks):
+        if isinstance(r.type, RTuple):
+            emitter.declare_tuple_struct(r.type)
+        if isinstance(r.type, RArray):
+            continue  # Special: declared on first assignment
+
+        if r in fn.arg_regs:
+            continue  # Skip the arguments
+
+        ctype = emitter.ctype_spaced(r.type)
+        init = ""
+        declarations.emit_line(
+            "{ctype}{prefix}{name}{init};".format(
+                ctype=ctype, prefix=REG_PREFIX, name=names[r], init=init
+            )
+        )
+
+    # Before we emit the blocks, give them all labels
+    blocks = fn.blocks
+    for i, block in enumerate(blocks):
+        block.label = i
+
+    # Find blocks that are never jumped to or are only jumped to from the
+    # block directly above it. This allows for more labels and gotos to be
+    # eliminated during code generation.
+    for block in fn.blocks:
+        terminator = block.terminator
+        assert isinstance(terminator, ControlOp)
+
+        for target in terminator.targets():
+            is_next_block = target.label == block.label + 1
+
+            # Always emit labels for GetAttr error checks since the emit code that
+            # generates them will add instructions between the branch and the
+            # next label, causing the label to be wrongly removed. A better
+            # solution would be to change the IR so that it adds a basic block
+            # inbetween the calls.
+            is_problematic_op = isinstance(terminator, Branch) and any(
+                isinstance(s, GetAttr) for s in terminator.sources()
+            )
+
+            if not is_next_block or is_problematic_op:
+                fn.blocks[target.label].referenced = True
+
+    common = frequently_executed_blocks(fn.blocks[0])
+
+    for i in range(len(blocks)):
+        block = blocks[i]
+        visitor.rare = block not in common
+        next_block = None
+        if i + 1 < len(blocks):
+            next_block = blocks[i + 1]
+        body.emit_label(block)
+        visitor.next_block = next_block
+
+        ops = block.ops
+        visitor.ops = ops
+        visitor.op_index = 0
+        while visitor.op_index < len(ops):
+            ops[visitor.op_index].accept(visitor)
+            visitor.op_index += 1
+
+    body.emit_line("}")
+
+    emitter.emit_from_emitter(declarations)
+    emitter.emit_from_emitter(body)
+
+
+class FunctionEmitterVisitor(OpVisitor[None]):
+    def __init__(
+        self, emitter: Emitter, declarations: Emitter, source_path: str, module_name: str
+    ) -> None:
+        self.emitter = emitter
+        self.names = emitter.names
+        self.declarations = declarations
+        self.source_path = source_path
+        self.module_name = module_name
+        self.literals = emitter.context.literals
+        self.rare = False
+        # Next basic block to be processed after the current one (if any), set by caller
+        self.next_block: BasicBlock | None = None
+        # Ops in the basic block currently being processed, set by caller
+        self.ops: list[Op] = []
+        # Current index within ops; visit methods can increment this to skip/merge ops
+        self.op_index = 0
+
+    def temp_name(self) -> str:
+        return self.emitter.temp_name()
+
+    def visit_goto(self, op: Goto) -> None:
+        if op.label is not self.next_block:
+            self.emit_line("goto %s;" % self.label(op.label))
+
+    def visit_branch(self, op: Branch) -> None:
+        true, false = op.true, op.false
+        negated = op.negated
+        negated_rare = False
+        if true is self.next_block and op.traceback_entry is None:
+            # Switch true/false since it avoids an else block.
+            true, false = false, true
+            negated = not negated
+            negated_rare = True
+
+        neg = "!" if negated else ""
+        cond = ""
+        if op.op == Branch.BOOL:
+            expr_result = self.reg(op.value)
+            cond = f"{neg}{expr_result}"
+        elif op.op == Branch.IS_ERROR:
+            typ = op.value.type
+            compare = "!=" if negated else "=="
+            if isinstance(typ, RTuple):
+                # TODO: What about empty tuple?
+                cond = self.emitter.tuple_undefined_check_cond(
+                    typ, self.reg(op.value), self.c_error_value, compare
+                )
+            else:
+                cond = f"{self.reg(op.value)} {compare} {self.c_error_value(typ)}"
+        else:
+            assert False, "Invalid branch"
+
+        # For error checks, tell the compiler the branch is unlikely
+        if op.traceback_entry is not None or op.rare:
+            if not negated_rare:
+                cond = f"unlikely({cond})"
+            else:
+                cond = f"likely({cond})"
+
+        if false is self.next_block:
+            if op.traceback_entry is None:
+                if true is not self.next_block:
+                    self.emit_line(f"if ({cond}) goto {self.label(true)};")
+            else:
+                self.emit_line(f"if ({cond}) {{")
+                self.emit_traceback(op)
+                self.emit_lines("goto %s;" % self.label(true), "}")
+        else:
+            self.emit_line(f"if ({cond}) {{")
+            self.emit_traceback(op)
+
+            if true is not self.next_block:
+                self.emit_line("goto %s;" % self.label(true))
+
+            self.emit_lines("} else", "    goto %s;" % self.label(false))
+
+    def visit_return(self, op: Return) -> None:
+        value_str = self.reg(op.value)
+        self.emit_line("return %s;" % value_str)
+
+    def visit_tuple_set(self, op: TupleSet) -> None:
+        dest = self.reg(op)
+        tuple_type = op.tuple_type
+        self.emitter.declare_tuple_struct(tuple_type)
+        if len(op.items) == 0:  # empty tuple
+            self.emit_line(f"{dest}.empty_struct_error_flag = 0;")
+        else:
+            for i, item in enumerate(op.items):
+                self.emit_line(f"{dest}.f{i} = {self.reg(item)};")
+        self.emit_inc_ref(dest, tuple_type)
+
+    def visit_assign(self, op: Assign) -> None:
+        dest = self.reg(op.dest)
+        src = self.reg(op.src)
+        # clang whines about self assignment (which we might generate
+        # for some casts), so don't emit it.
+        if dest != src:
+            # We sometimes assign from an integer prepresentation of a pointer
+            # to a real pointer, and C compilers insist on a cast.
+            if op.src.type.is_unboxed and not op.dest.type.is_unboxed:
+                src = f"(void *){src}"
+            self.emit_line(f"{dest} = {src};")
+
+    def visit_assign_multi(self, op: AssignMulti) -> None:
+        typ = op.dest.type
+        assert isinstance(typ, RArray)
+        dest = self.reg(op.dest)
+        # RArray values can only be assigned to once, so we can always
+        # declare them on initialization.
+        self.emit_line(
+            "%s%s[%d] = %s;"
+            % (
+                self.emitter.ctype_spaced(typ.item_type),
+                dest,
+                len(op.src),
+                c_array_initializer([self.reg(s) for s in op.src], indented=True),
+            )
+        )
+
+    def visit_load_error_value(self, op: LoadErrorValue) -> None:
+        if isinstance(op.type, RTuple):
+            values = [self.c_undefined_value(item) for item in op.type.types]
+            tmp = self.temp_name()
+            self.emit_line("{} {} = {{ {} }};".format(self.ctype(op.type), tmp, ", ".join(values)))
+            self.emit_line(f"{self.reg(op)} = {tmp};")
+        else:
+            self.emit_line(f"{self.reg(op)} = {self.c_error_value(op.type)};")
+
+    def visit_load_literal(self, op: LoadLiteral) -> None:
+        index = self.literals.literal_index(op.value)
+        if not is_int_rprimitive(op.type):
+            self.emit_line("%s = CPyStatics[%d];" % (self.reg(op), index), ann=op.value)
+        else:
+            self.emit_line(
+                "%s = (CPyTagged)CPyStatics[%d] | 1;" % (self.reg(op), index), ann=op.value
+            )
+
+    def get_attr_expr(self, obj: str, op: GetAttr | SetAttr, decl_cl: ClassIR) -> str:
+        """Generate attribute accessor for normal (non-property) access.
+
+        This either has a form like obj->attr_name for attributes defined in non-trait
+        classes, and *(obj + attr_offset) for attributes defined by traits. We also
+        insert all necessary C casts here.
+        """
+        cast = f"({op.class_type.struct_name(self.emitter.names)} *)"
+        if decl_cl.is_trait and op.class_type.class_ir.is_trait:
+            # For pure trait access find the offset first, offsets
+            # are ordered by attribute position in the cl.attributes dict.
+            # TODO: pre-calculate the mapping to make this faster.
+            trait_attr_index = list(decl_cl.attributes).index(op.attr)
+            # TODO: reuse these names somehow?
+            offset = self.emitter.temp_name()
+            self.declarations.emit_line(f"size_t {offset};")
+            self.emitter.emit_line(
+                "{} = {};".format(
+                    offset,
+                    "CPy_FindAttrOffset({}, {}, {})".format(
+                        self.emitter.type_struct_name(decl_cl),
+                        f"({cast}{obj})->vtable",
+                        trait_attr_index,
+                    ),
+                )
+            )
+            attr_cast = f"({self.ctype(op.class_type.attr_type(op.attr))} *)"
+            return f"*{attr_cast}((char *){obj} + {offset})"
+        else:
+            # Cast to something non-trait. Note: for this to work, all struct
+            # members for non-trait classes must obey monotonic linear growth.
+            if op.class_type.class_ir.is_trait:
+                assert not decl_cl.is_trait
+                cast = f"({decl_cl.struct_name(self.emitter.names)} *)"
+            return f"({cast}{obj})->{self.emitter.attr(op.attr)}"
+
+    def visit_get_attr(self, op: GetAttr) -> None:
+        dest = self.reg(op)
+        obj = self.reg(op.obj)
+        rtype = op.class_type
+        cl = rtype.class_ir
+        attr_rtype, decl_cl = cl.attr_details(op.attr)
+        prefer_method = cl.is_trait and attr_rtype.error_overlap
+        if cl.get_method(op.attr, prefer_method=prefer_method):
+            # Properties are essentially methods, so use vtable access for them.
+            version = "_TRAIT" if cl.is_trait else ""
+            self.emit_line(
+                "%s = CPY_GET_ATTR%s(%s, %s, %d, %s, %s); /* %s */"
+                % (
+                    dest,
+                    version,
+                    obj,
+                    self.emitter.type_struct_name(rtype.class_ir),
+                    rtype.getter_index(op.attr),
+                    rtype.struct_name(self.names),
+                    self.ctype(rtype.attr_type(op.attr)),
+                    op.attr,
+                )
+            )
+        else:
+            # Otherwise, use direct or offset struct access.
+            attr_expr = self.get_attr_expr(obj, op, decl_cl)
+            self.emitter.emit_line(f"{dest} = {attr_expr};")
+            always_defined = cl.is_always_defined(op.attr)
+            merged_branch = None
+            if not always_defined:
+                self.emitter.emit_undefined_attr_check(
+                    attr_rtype, dest, "==", obj, op.attr, cl, unlikely=True
+                )
+                branch = self.next_branch()
+                if branch is not None:
+                    if (
+                        branch.value is op
+                        and branch.op == Branch.IS_ERROR
+                        and branch.traceback_entry is not None
+                        and not branch.negated
+                    ):
+                        # Generate code for the following branch here to avoid
+                        # redundant branches in the generated code.
+                        self.emit_attribute_error(branch, cl.name, op.attr)
+                        self.emit_line("goto %s;" % self.label(branch.true))
+                        merged_branch = branch
+                        self.emitter.emit_line("}")
+                if not merged_branch:
+                    exc_class = "PyExc_AttributeError"
+                    self.emitter.emit_line(
+                        'PyErr_SetString({}, "attribute {} of {} undefined");'.format(
+                            exc_class, repr(op.attr), repr(cl.name)
+                        )
+                    )
+
+            if attr_rtype.is_refcounted and not op.is_borrowed:
+                if not merged_branch and not always_defined:
+                    self.emitter.emit_line("} else {")
+                self.emitter.emit_inc_ref(dest, attr_rtype)
+            if merged_branch:
+                if merged_branch.false is not self.next_block:
+                    self.emit_line("goto %s;" % self.label(merged_branch.false))
+                self.op_index += 1
+            elif not always_defined:
+                self.emitter.emit_line("}")
+
+    def next_branch(self) -> Branch | None:
+        if self.op_index + 1 < len(self.ops):
+            next_op = self.ops[self.op_index + 1]
+            if isinstance(next_op, Branch):
+                return next_op
+        return None
+
+    def visit_set_attr(self, op: SetAttr) -> None:
+        if op.error_kind == ERR_FALSE:
+            dest = self.reg(op)
+        obj = self.reg(op.obj)
+        src = self.reg(op.src)
+        rtype = op.class_type
+        cl = rtype.class_ir
+        attr_rtype, decl_cl = cl.attr_details(op.attr)
+        if cl.get_method(op.attr):
+            # Again, use vtable access for properties...
+            assert not op.is_init and op.error_kind == ERR_FALSE, "%s %d %d %s" % (
+                op.attr,
+                op.is_init,
+                op.error_kind,
+                rtype,
+            )
+            version = "_TRAIT" if cl.is_trait else ""
+            self.emit_line(
+                "%s = CPY_SET_ATTR%s(%s, %s, %d, %s, %s, %s); /* %s */"
+                % (
+                    dest,
+                    version,
+                    obj,
+                    self.emitter.type_struct_name(rtype.class_ir),
+                    rtype.setter_index(op.attr),
+                    src,
+                    rtype.struct_name(self.names),
+                    self.ctype(rtype.attr_type(op.attr)),
+                    op.attr,
+                )
+            )
+        else:
+            # ...and struct access for normal attributes.
+            attr_expr = self.get_attr_expr(obj, op, decl_cl)
+            if not op.is_init and attr_rtype.is_refcounted:
+                # This is not an initialization (where we know that the attribute was
+                # previously undefined), so decref the old value.
+                always_defined = cl.is_always_defined(op.attr)
+                if not always_defined:
+                    self.emitter.emit_undefined_attr_check(
+                        attr_rtype, attr_expr, "!=", obj, op.attr, cl
+                    )
+                self.emitter.emit_dec_ref(attr_expr, attr_rtype)
+                if not always_defined:
+                    self.emitter.emit_line("}")
+            elif attr_rtype.error_overlap and not cl.is_always_defined(op.attr):
+                # If there is overlap with the error value, update bitmap to mark
+                # attribute as defined.
+                self.emitter.emit_attr_bitmap_set(src, obj, attr_rtype, cl, op.attr)
+
+            # This steals the reference to src, so we don't need to increment the arg
+            self.emitter.emit_line(f"{attr_expr} = {src};")
+            if op.error_kind == ERR_FALSE:
+                self.emitter.emit_line(f"{dest} = 1;")
+
+    PREFIX_MAP: Final = {
+        NAMESPACE_STATIC: STATIC_PREFIX,
+        NAMESPACE_TYPE: TYPE_PREFIX,
+        NAMESPACE_MODULE: MODULE_PREFIX,
+    }
+
+    def visit_load_static(self, op: LoadStatic) -> None:
+        dest = self.reg(op)
+        prefix = self.PREFIX_MAP[op.namespace]
+        name = self.emitter.static_name(op.identifier, op.module_name, prefix)
+        if op.namespace == NAMESPACE_TYPE:
+            name = "(PyObject *)%s" % name
+        self.emit_line(f"{dest} = {name};", ann=op.ann)
+
+    def visit_init_static(self, op: InitStatic) -> None:
+        value = self.reg(op.value)
+        prefix = self.PREFIX_MAP[op.namespace]
+        name = self.emitter.static_name(op.identifier, op.module_name, prefix)
+        if op.namespace == NAMESPACE_TYPE:
+            value = "(PyTypeObject *)%s" % value
+        self.emit_line(f"{name} = {value};")
+        self.emit_inc_ref(name, op.value.type)
+
+    def visit_tuple_get(self, op: TupleGet) -> None:
+        dest = self.reg(op)
+        src = self.reg(op.src)
+        self.emit_line(f"{dest} = {src}.f{op.index};")
+        self.emit_inc_ref(dest, op.type)
+
+    def get_dest_assign(self, dest: Value) -> str:
+        if not dest.is_void:
+            return self.reg(dest) + " = "
+        else:
+            return ""
+
+    def visit_call(self, op: Call) -> None:
+        """Call native function."""
+        dest = self.get_dest_assign(op)
+        args = ", ".join(self.reg(arg) for arg in op.args)
+        lib = self.emitter.get_group_prefix(op.fn)
+        cname = op.fn.cname(self.names)
+        self.emit_line(f"{dest}{lib}{NATIVE_PREFIX}{cname}({args});")
+
+    def visit_method_call(self, op: MethodCall) -> None:
+        """Call native method."""
+        dest = self.get_dest_assign(op)
+        obj = self.reg(op.obj)
+
+        rtype = op.receiver_type
+        class_ir = rtype.class_ir
+        name = op.method
+        method = rtype.class_ir.get_method(name)
+        assert method is not None
+
+        # Can we call the method directly, bypassing vtable?
+        is_direct = class_ir.is_method_final(name)
+
+        # The first argument gets omitted for static methods and
+        # turned into the class for class methods
+        obj_args = (
+            []
+            if method.decl.kind == FUNC_STATICMETHOD
+            else [f"(PyObject *)Py_TYPE({obj})"]
+            if method.decl.kind == FUNC_CLASSMETHOD
+            else [obj]
+        )
+        args = ", ".join(obj_args + [self.reg(arg) for arg in op.args])
+        mtype = native_function_type(method, self.emitter)
+        version = "_TRAIT" if rtype.class_ir.is_trait else ""
+        if is_direct:
+            # Directly call method, without going through the vtable.
+            lib = self.emitter.get_group_prefix(method.decl)
+            self.emit_line(f"{dest}{lib}{NATIVE_PREFIX}{method.cname(self.names)}({args});")
+        else:
+            # Call using vtable.
+            method_idx = rtype.method_index(name)
+            self.emit_line(
+                "{}CPY_GET_METHOD{}({}, {}, {}, {}, {})({}); /* {} */".format(
+                    dest,
+                    version,
+                    obj,
+                    self.emitter.type_struct_name(rtype.class_ir),
+                    method_idx,
+                    rtype.struct_name(self.names),
+                    mtype,
+                    args,
+                    op.method,
+                )
+            )
+
+    def visit_inc_ref(self, op: IncRef) -> None:
+        src = self.reg(op.src)
+        self.emit_inc_ref(src, op.src.type)
+
+    def visit_dec_ref(self, op: DecRef) -> None:
+        src = self.reg(op.src)
+        self.emit_dec_ref(src, op.src.type, is_xdec=op.is_xdec)
+
+    def visit_box(self, op: Box) -> None:
+        self.emitter.emit_box(self.reg(op.src), self.reg(op), op.src.type, can_borrow=True)
+
+    def visit_cast(self, op: Cast) -> None:
+        branch = self.next_branch()
+        handler = None
+        if branch is not None:
+            if (
+                branch.value is op
+                and branch.op == Branch.IS_ERROR
+                and branch.traceback_entry is not None
+                and not branch.negated
+                and branch.false is self.next_block
+            ):
+                # Generate code also for the following branch here to avoid
+                # redundant branches in the generated code.
+                handler = TracebackAndGotoHandler(
+                    self.label(branch.true),
+                    self.source_path,
+                    self.module_name,
+                    branch.traceback_entry,
+                )
+                self.op_index += 1
+
+        self.emitter.emit_cast(
+            self.reg(op.src), self.reg(op), op.type, src_type=op.src.type, error=handler
+        )
+
+    def visit_unbox(self, op: Unbox) -> None:
+        self.emitter.emit_unbox(self.reg(op.src), self.reg(op), op.type)
+
+    def visit_unreachable(self, op: Unreachable) -> None:
+        self.emitter.emit_line("CPy_Unreachable();")
+
+    def visit_raise_standard_error(self, op: RaiseStandardError) -> None:
+        # TODO: Better escaping of backspaces and such
+        if op.value is not None:
+            if isinstance(op.value, str):
+                message = op.value.replace('"', '\\"')
+                self.emitter.emit_line(f'PyErr_SetString(PyExc_{op.class_name}, "{message}");')
+            elif isinstance(op.value, Value):
+                self.emitter.emit_line(
+                    "PyErr_SetObject(PyExc_{}, {});".format(
+                        op.class_name, self.emitter.reg(op.value)
+                    )
+                )
+            else:
+                assert False, "op value type must be either str or Value"
+        else:
+            self.emitter.emit_line(f"PyErr_SetNone(PyExc_{op.class_name});")
+        self.emitter.emit_line(f"{self.reg(op)} = 0;")
+
+    def visit_call_c(self, op: CallC) -> None:
+        if op.is_void:
+            dest = ""
+        else:
+            dest = self.get_dest_assign(op)
+        args = ", ".join(self.reg(arg) for arg in op.args)
+        self.emitter.emit_line(f"{dest}{op.function_name}({args});")
+
+    def visit_truncate(self, op: Truncate) -> None:
+        dest = self.reg(op)
+        value = self.reg(op.src)
+        # for C backend the generated code are straight assignments
+        self.emit_line(f"{dest} = {value};")
+
+    def visit_extend(self, op: Extend) -> None:
+        dest = self.reg(op)
+        value = self.reg(op.src)
+        if op.signed:
+            src_cast = self.emit_signed_int_cast(op.src.type)
+        else:
+            src_cast = self.emit_unsigned_int_cast(op.src.type)
+        self.emit_line(f"{dest} = {src_cast}{value};")
+
+    def visit_load_global(self, op: LoadGlobal) -> None:
+        dest = self.reg(op)
+        self.emit_line(f"{dest} = {op.identifier};", ann=op.ann)
+
+    def visit_int_op(self, op: IntOp) -> None:
+        dest = self.reg(op)
+        lhs = self.reg(op.lhs)
+        rhs = self.reg(op.rhs)
+        if op.op == IntOp.RIGHT_SHIFT:
+            # Signed right shift
+            lhs = self.emit_signed_int_cast(op.lhs.type) + lhs
+            rhs = self.emit_signed_int_cast(op.rhs.type) + rhs
+        self.emit_line(f"{dest} = {lhs} {op.op_str[op.op]} {rhs};")
+
+    def visit_comparison_op(self, op: ComparisonOp) -> None:
+        dest = self.reg(op)
+        lhs = self.reg(op.lhs)
+        rhs = self.reg(op.rhs)
+        lhs_cast = ""
+        rhs_cast = ""
+        if op.op in (ComparisonOp.SLT, ComparisonOp.SGT, ComparisonOp.SLE, ComparisonOp.SGE):
+            # Always signed comparison op
+            lhs_cast = self.emit_signed_int_cast(op.lhs.type)
+            rhs_cast = self.emit_signed_int_cast(op.rhs.type)
+        elif op.op in (ComparisonOp.ULT, ComparisonOp.UGT, ComparisonOp.ULE, ComparisonOp.UGE):
+            # Always unsigned comparison op
+            lhs_cast = self.emit_unsigned_int_cast(op.lhs.type)
+            rhs_cast = self.emit_unsigned_int_cast(op.rhs.type)
+        elif isinstance(op.lhs, Integer) and op.lhs.value < 0:
+            # Force signed ==/!= with negative operand
+            rhs_cast = self.emit_signed_int_cast(op.rhs.type)
+        elif isinstance(op.rhs, Integer) and op.rhs.value < 0:
+            # Force signed ==/!= with negative operand
+            lhs_cast = self.emit_signed_int_cast(op.lhs.type)
+        self.emit_line(f"{dest} = {lhs_cast}{lhs} {op.op_str[op.op]} {rhs_cast}{rhs};")
+
+    def visit_float_op(self, op: FloatOp) -> None:
+        dest = self.reg(op)
+        lhs = self.reg(op.lhs)
+        rhs = self.reg(op.rhs)
+        if op.op != FloatOp.MOD:
+            self.emit_line(f"{dest} = {lhs} {op.op_str[op.op]} {rhs};")
+        else:
+            # TODO: This may set errno as a side effect, that is a little sketchy.
+            self.emit_line(f"{dest} = fmod({lhs}, {rhs});")
+
+    def visit_float_neg(self, op: FloatNeg) -> None:
+        dest = self.reg(op)
+        src = self.reg(op.src)
+        self.emit_line(f"{dest} = -{src};")
+
+    def visit_float_comparison_op(self, op: FloatComparisonOp) -> None:
+        dest = self.reg(op)
+        lhs = self.reg(op.lhs)
+        rhs = self.reg(op.rhs)
+        self.emit_line(f"{dest} = {lhs} {op.op_str[op.op]} {rhs};")
+
+    def visit_load_mem(self, op: LoadMem) -> None:
+        dest = self.reg(op)
+        src = self.reg(op.src)
+        # TODO: we shouldn't dereference to type that are pointer type so far
+        type = self.ctype(op.type)
+        self.emit_line(f"{dest} = *({type} *){src};")
+
+    def visit_set_mem(self, op: SetMem) -> None:
+        dest = self.reg(op.dest)
+        src = self.reg(op.src)
+        dest_type = self.ctype(op.dest_type)
+        # clang whines about self assignment (which we might generate
+        # for some casts), so don't emit it.
+        if dest != src:
+            self.emit_line(f"*({dest_type} *){dest} = {src};")
+
+    def visit_get_element_ptr(self, op: GetElementPtr) -> None:
+        dest = self.reg(op)
+        src = self.reg(op.src)
+        # TODO: support tuple type
+        assert isinstance(op.src_type, RStruct)
+        assert op.field in op.src_type.names, "Invalid field name."
+        self.emit_line(
+            "{} = ({})&(({} *){})->{};".format(
+                dest, op.type._ctype, op.src_type.name, src, op.field
+            )
+        )
+
+    def visit_load_address(self, op: LoadAddress) -> None:
+        typ = op.type
+        dest = self.reg(op)
+        if isinstance(op.src, Register):
+            src = self.reg(op.src)
+        elif isinstance(op.src, LoadStatic):
+            prefix = self.PREFIX_MAP[op.src.namespace]
+            src = self.emitter.static_name(op.src.identifier, op.src.module_name, prefix)
+        else:
+            src = op.src
+        self.emit_line(f"{dest} = ({typ._ctype})&{src};")
+
+    def visit_keep_alive(self, op: KeepAlive) -> None:
+        # This is a no-op.
+        pass
+
+    # Helpers
+
+    def label(self, label: BasicBlock) -> str:
+        return self.emitter.label(label)
+
+    def reg(self, reg: Value) -> str:
+        if isinstance(reg, Integer):
+            val = reg.value
+            if val == 0 and is_pointer_rprimitive(reg.type):
+                return "NULL"
+            s = str(val)
+            if val >= (1 << 31):
+                # Avoid overflowing signed 32-bit int
+                if val >= (1 << 63):
+                    s += "ULL"
+                else:
+                    s += "LL"
+            elif val == -(1 << 63):
+                # Avoid overflowing C integer literal
+                s = "(-9223372036854775807LL - 1)"
+            elif val <= -(1 << 31):
+                s += "LL"
+            return s
+        elif isinstance(reg, Float):
+            r = repr(reg.value)
+            if r == "inf":
+                return "INFINITY"
+            elif r == "-inf":
+                return "-INFINITY"
+            elif r == "nan":
+                return "NAN"
+            return r
+        else:
+            return self.emitter.reg(reg)
+
+    def ctype(self, rtype: RType) -> str:
+        return self.emitter.ctype(rtype)
+
+    def c_error_value(self, rtype: RType) -> str:
+        return self.emitter.c_error_value(rtype)
+
+    def c_undefined_value(self, rtype: RType) -> str:
+        return self.emitter.c_undefined_value(rtype)
+
+    def emit_line(self, line: str, *, ann: object = None) -> None:
+        self.emitter.emit_line(line, ann=ann)
+
+    def emit_lines(self, *lines: str) -> None:
+        self.emitter.emit_lines(*lines)
+
+    def emit_inc_ref(self, dest: str, rtype: RType) -> None:
+        self.emitter.emit_inc_ref(dest, rtype, rare=self.rare)
+
+    def emit_dec_ref(self, dest: str, rtype: RType, is_xdec: bool) -> None:
+        self.emitter.emit_dec_ref(dest, rtype, is_xdec=is_xdec, rare=self.rare)
+
+    def emit_declaration(self, line: str) -> None:
+        self.declarations.emit_line(line)
+
+    def emit_traceback(self, op: Branch) -> None:
+        if op.traceback_entry is not None:
+            self.emitter.emit_traceback(self.source_path, self.module_name, op.traceback_entry)
+
+    def emit_attribute_error(self, op: Branch, class_name: str, attr: str) -> None:
+        assert op.traceback_entry is not None
+        globals_static = self.emitter.static_name("globals", self.module_name)
+        self.emit_line(
+            'CPy_AttributeError("%s", "%s", "%s", "%s", %d, %s);'
+            % (
+                self.source_path.replace("\\", "\\\\"),
+                op.traceback_entry[0],
+                class_name,
+                attr,
+                op.traceback_entry[1],
+                globals_static,
+            )
+        )
+        if DEBUG_ERRORS:
+            self.emit_line('assert(PyErr_Occurred() != NULL && "failure w/o err!");')
+
+    def emit_signed_int_cast(self, type: RType) -> str:
+        if is_tagged(type):
+            return "(Py_ssize_t)"
+        else:
+            return ""
+
+    def emit_unsigned_int_cast(self, type: RType) -> str:
+        if is_int32_rprimitive(type):
+            return "(uint32_t)"
+        elif is_int64_rprimitive(type):
+            return "(uint64_t)"
+        else:
+            return ""

venv/lib/python3.12/site-packages/mypyc/codegen/emitwrapper.py

@@ -0,0 +1,979 @@
+"""Generate CPython API wrapper functions for native functions.
+
+The wrapper functions are used by the CPython runtime when calling
+native functions from interpreted code, and when the called function
+can't be determined statically in compiled code. They validate, match,
+unbox and type check function arguments, and box return values as
+needed. All wrappers accept and return 'PyObject *' (boxed) values.
+
+The wrappers aren't used for most calls between two native functions
+or methods in a single compilation unit.
+"""
+
+from __future__ import annotations
+
+from typing import Sequence
+
+from mypy.nodes import ARG_NAMED, ARG_NAMED_OPT, ARG_OPT, ARG_POS, ARG_STAR, ARG_STAR2, ArgKind
+from mypy.operators import op_methods_to_symbols, reverse_op_method_names, reverse_op_methods
+from mypyc.codegen.emit import AssignHandler, Emitter, ErrorHandler, GotoHandler, ReturnHandler
+from mypyc.common import (
+    BITMAP_BITS,
+    BITMAP_TYPE,
+    DUNDER_PREFIX,
+    NATIVE_PREFIX,
+    PREFIX,
+    bitmap_name,
+    use_vectorcall,
+)
+from mypyc.ir.class_ir import ClassIR
+from mypyc.ir.func_ir import FUNC_STATICMETHOD, FuncIR, RuntimeArg
+from mypyc.ir.rtypes import (
+    RInstance,
+    RType,
+    is_bool_rprimitive,
+    is_int_rprimitive,
+    is_object_rprimitive,
+    object_rprimitive,
+)
+from mypyc.namegen import NameGenerator
+
+# Generic vectorcall wrapper functions (Python 3.7+)
+#
+# A wrapper function has a signature like this:
+#
+# PyObject *fn(PyObject *self, PyObject *const *args, Py_ssize_t nargs, PyObject *kwnames)
+#
+# The function takes a self object, pointer to an array of arguments,
+# the number of positional arguments, and a tuple of keyword argument
+# names (that are stored starting in args[nargs]).
+#
+# It returns the returned object, or NULL on an exception.
+#
+# These are more efficient than legacy wrapper functions, since
+# usually no tuple or dict objects need to be created for the
+# arguments. Vectorcalls also use pre-constructed str objects for
+# keyword argument names and other pre-computed information, instead
+# of processing the argument format string on each call.
+
+
+def wrapper_function_header(fn: FuncIR, names: NameGenerator) -> str:
+    """Return header of a vectorcall wrapper function.
+
+    See comment above for a summary of the arguments.
+    """
+    return (
+        "PyObject *{prefix}{name}("
+        "PyObject *self, PyObject *const *args, size_t nargs, PyObject *kwnames)"
+    ).format(prefix=PREFIX, name=fn.cname(names))
+
+
+def generate_traceback_code(
+    fn: FuncIR, emitter: Emitter, source_path: str, module_name: str
+) -> str:
+    # If we hit an error while processing arguments, then we emit a
+    # traceback frame to make it possible to debug where it happened.
+    # Unlike traceback frames added for exceptions seen in IR, we do this
+    # even if there is no `traceback_name`. This is because the error will
+    # have originated here and so we need it in the traceback.
+    globals_static = emitter.static_name("globals", module_name)
+    traceback_code = 'CPy_AddTraceback("%s", "%s", %d, %s);' % (
+        source_path.replace("\\", "\\\\"),
+        fn.traceback_name or fn.name,
+        fn.line,
+        globals_static,
+    )
+    return traceback_code
+
+
+def make_arg_groups(args: list[RuntimeArg]) -> dict[ArgKind, list[RuntimeArg]]:
+    """Group arguments by kind."""
+    return {k: [arg for arg in args if arg.kind == k] for k in ArgKind}
+
+
+def reorder_arg_groups(groups: dict[ArgKind, list[RuntimeArg]]) -> list[RuntimeArg]:
+    """Reorder argument groups to match their order in a format string."""
+    return groups[ARG_POS] + groups[ARG_OPT] + groups[ARG_NAMED_OPT] + groups[ARG_NAMED]
+
+
+def make_static_kwlist(args: list[RuntimeArg]) -> str:
+    arg_names = "".join(f'"{arg.name}", ' for arg in args)
+    return f"static const char * const kwlist[] = {{{arg_names}0}};"
+
+
+def make_format_string(func_name: str | None, groups: dict[ArgKind, list[RuntimeArg]]) -> str:
+    """Return a format string that specifies the accepted arguments.
+
+    The format string is an extended subset of what is supported by
+    PyArg_ParseTupleAndKeywords(). Only the type 'O' is used, and we
+    also support some extensions:
+
+    - Required keyword-only arguments are introduced after '@'
+    - If the function receives *args or **kwargs, we add a '%' prefix
+
+    Each group requires the previous groups' delimiters to be present
+    first.
+
+    These are used by both vectorcall and legacy wrapper functions.
+    """
+    format = ""
+    if groups[ARG_STAR] or groups[ARG_STAR2]:
+        format += "%"
+    format += "O" * len(groups[ARG_POS])
+    if groups[ARG_OPT] or groups[ARG_NAMED_OPT] or groups[ARG_NAMED]:
+        format += "|" + "O" * len(groups[ARG_OPT])
+    if groups[ARG_NAMED_OPT] or groups[ARG_NAMED]:
+        format += "$" + "O" * len(groups[ARG_NAMED_OPT])
+    if groups[ARG_NAMED]:
+        format += "@" + "O" * len(groups[ARG_NAMED])
+    if func_name is not None:
+        format += f":{func_name}"
+    return format
+
+
+def generate_wrapper_function(
+    fn: FuncIR, emitter: Emitter, source_path: str, module_name: str
+) -> None:
+    """Generate a CPython-compatible vectorcall wrapper for a native function.
+
+    In particular, this handles unboxing the arguments, calling the native function, and
+    then boxing the return value.
+    """
+    emitter.emit_line(f"{wrapper_function_header(fn, emitter.names)} {{")
+
+    # If fn is a method, then the first argument is a self param
+    real_args = list(fn.args)
+    if fn.sig.num_bitmap_args:
+        real_args = real_args[: -fn.sig.num_bitmap_args]
+    if fn.class_name and not fn.decl.kind == FUNC_STATICMETHOD:
+        arg = real_args.pop(0)
+        emitter.emit_line(f"PyObject *obj_{arg.name} = self;")
+
+    # Need to order args as: required, optional, kwonly optional, kwonly required
+    # This is because CPyArg_ParseStackAndKeywords format string requires
+    # them grouped in that way.
+    groups = make_arg_groups(real_args)
+    reordered_args = reorder_arg_groups(groups)
+
+    emitter.emit_line(make_static_kwlist(reordered_args))
+    fmt = make_format_string(fn.name, groups)
+    # Define the arguments the function accepts (but no types yet)
+    emitter.emit_line(f'static CPyArg_Parser parser = {{"{fmt}", kwlist, 0}};')
+
+    for arg in real_args:
+        emitter.emit_line(
+            "PyObject *obj_{}{};".format(arg.name, " = NULL" if arg.optional else "")
+        )
+
+    cleanups = [f"CPy_DECREF(obj_{arg.name});" for arg in groups[ARG_STAR] + groups[ARG_STAR2]]
+
+    arg_ptrs: list[str] = []
+    if groups[ARG_STAR] or groups[ARG_STAR2]:
+        arg_ptrs += [f"&obj_{groups[ARG_STAR][0].name}" if groups[ARG_STAR] else "NULL"]
+        arg_ptrs += [f"&obj_{groups[ARG_STAR2][0].name}" if groups[ARG_STAR2] else "NULL"]
+    arg_ptrs += [f"&obj_{arg.name}" for arg in reordered_args]
+
+    if fn.name == "__call__" and use_vectorcall(emitter.capi_version):
+        nargs = "PyVectorcall_NARGS(nargs)"
+    else:
+        nargs = "nargs"
+    parse_fn = "CPyArg_ParseStackAndKeywords"
+    # Special case some common signatures
+    if not real_args:
+        # No args
+        parse_fn = "CPyArg_ParseStackAndKeywordsNoArgs"
+    elif len(real_args) == 1 and len(groups[ARG_POS]) == 1:
+        # Single positional arg
+        parse_fn = "CPyArg_ParseStackAndKeywordsOneArg"
+    elif len(real_args) == len(groups[ARG_POS]) + len(groups[ARG_OPT]):
+        # No keyword-only args, *args or **kwargs
+        parse_fn = "CPyArg_ParseStackAndKeywordsSimple"
+    emitter.emit_lines(
+        "if (!{}(args, {}, kwnames, &parser{})) {{".format(
+            parse_fn, nargs, "".join(", " + n for n in arg_ptrs)
+        ),
+        "return NULL;",
+        "}",
+    )
+    for i in range(fn.sig.num_bitmap_args):
+        name = bitmap_name(i)
+        emitter.emit_line(f"{BITMAP_TYPE} {name} = 0;")
+    traceback_code = generate_traceback_code(fn, emitter, source_path, module_name)
+    generate_wrapper_core(
+        fn,
+        emitter,
+        groups[ARG_OPT] + groups[ARG_NAMED_OPT],
+        cleanups=cleanups,
+        traceback_code=traceback_code,
+    )
+
+    emitter.emit_line("}")
+
+
+# Legacy generic wrapper functions
+#
+# These take a self object, a Python tuple of positional arguments,
+# and a dict of keyword arguments. These are a lot slower than
+# vectorcall wrappers, especially in calls involving keyword
+# arguments.
+
+
+def legacy_wrapper_function_header(fn: FuncIR, names: NameGenerator) -> str:
+    return "PyObject *{prefix}{name}(PyObject *self, PyObject *args, PyObject *kw)".format(
+        prefix=PREFIX, name=fn.cname(names)
+    )
+
+
+def generate_legacy_wrapper_function(
+    fn: FuncIR, emitter: Emitter, source_path: str, module_name: str
+) -> None:
+    """Generates a CPython-compatible legacy wrapper for a native function.
+
+    In particular, this handles unboxing the arguments, calling the native function, and
+    then boxing the return value.
+    """
+    emitter.emit_line(f"{legacy_wrapper_function_header(fn, emitter.names)} {{")
+
+    # If fn is a method, then the first argument is a self param
+    real_args = list(fn.args)
+    if fn.sig.num_bitmap_args:
+        real_args = real_args[: -fn.sig.num_bitmap_args]
+    if fn.class_name and not fn.decl.kind == FUNC_STATICMETHOD:
+        arg = real_args.pop(0)
+        emitter.emit_line(f"PyObject *obj_{arg.name} = self;")
+
+    # Need to order args as: required, optional, kwonly optional, kwonly required
+    # This is because CPyArg_ParseTupleAndKeywords format string requires
+    # them grouped in that way.
+    groups = make_arg_groups(real_args)
+    reordered_args = reorder_arg_groups(groups)
+
+    emitter.emit_line(make_static_kwlist(reordered_args))
+    for arg in real_args:
+        emitter.emit_line(
+            "PyObject *obj_{}{};".format(arg.name, " = NULL" if arg.optional else "")
+        )
+
+    cleanups = [f"CPy_DECREF(obj_{arg.name});" for arg in groups[ARG_STAR] + groups[ARG_STAR2]]
+
+    arg_ptrs: list[str] = []
+    if groups[ARG_STAR] or groups[ARG_STAR2]:
+        arg_ptrs += [f"&obj_{groups[ARG_STAR][0].name}" if groups[ARG_STAR] else "NULL"]
+        arg_ptrs += [f"&obj_{groups[ARG_STAR2][0].name}" if groups[ARG_STAR2] else "NULL"]
+    arg_ptrs += [f"&obj_{arg.name}" for arg in reordered_args]
+
+    emitter.emit_lines(
+        'if (!CPyArg_ParseTupleAndKeywords(args, kw, "{}", "{}", kwlist{})) {{'.format(
+            make_format_string(None, groups), fn.name, "".join(", " + n for n in arg_ptrs)
+        ),
+        "return NULL;",
+        "}",
+    )
+    for i in range(fn.sig.num_bitmap_args):
+        name = bitmap_name(i)
+        emitter.emit_line(f"{BITMAP_TYPE} {name} = 0;")
+    traceback_code = generate_traceback_code(fn, emitter, source_path, module_name)
+    generate_wrapper_core(
+        fn,
+        emitter,
+        groups[ARG_OPT] + groups[ARG_NAMED_OPT],
+        cleanups=cleanups,
+        traceback_code=traceback_code,
+    )
+
+    emitter.emit_line("}")
+
+
+# Specialized wrapper functions
+
+
+def generate_dunder_wrapper(cl: ClassIR, fn: FuncIR, emitter: Emitter) -> str:
+    """Generates a wrapper for native __dunder__ methods to be able to fit into the mapping
+    protocol slot. This specifically means that the arguments are taken as *PyObjects and returned
+    as *PyObjects.
+    """
+    gen = WrapperGenerator(cl, emitter)
+    gen.set_target(fn)
+    gen.emit_header()
+    gen.emit_arg_processing()
+    gen.emit_call()
+    gen.finish()
+    return gen.wrapper_name()
+
+
+def generate_ipow_wrapper(cl: ClassIR, fn: FuncIR, emitter: Emitter) -> str:
+    """Generate a wrapper for native __ipow__.
+
+    Since __ipow__ fills a ternary slot, but almost no one defines __ipow__ to take three
+    arguments, the wrapper needs to tweaked to force it to accept three arguments.
+    """
+    gen = WrapperGenerator(cl, emitter)
+    gen.set_target(fn)
+    assert len(fn.args) in (2, 3), "__ipow__ should only take 2 or 3 arguments"
+    gen.arg_names = ["self", "exp", "mod"]
+    gen.emit_header()
+    gen.emit_arg_processing()
+    handle_third_pow_argument(
+        fn,
+        emitter,
+        gen,
+        if_unsupported=[
+            'PyErr_SetString(PyExc_TypeError, "__ipow__ takes 2 positional arguments but 3 were given");',
+            "return NULL;",
+        ],
+    )
+    gen.emit_call()
+    gen.finish()
+    return gen.wrapper_name()
+
+
+def generate_bin_op_wrapper(cl: ClassIR, fn: FuncIR, emitter: Emitter) -> str:
+    """Generates a wrapper for a native binary dunder method.
+
+    The same wrapper that handles the forward method (e.g. __add__) also handles
+    the corresponding reverse method (e.g. __radd__), if defined.
+
+    Both arguments and the return value are PyObject *.
+    """
+    gen = WrapperGenerator(cl, emitter)
+    gen.set_target(fn)
+    if fn.name in ("__pow__", "__rpow__"):
+        gen.arg_names = ["left", "right", "mod"]
+    else:
+        gen.arg_names = ["left", "right"]
+    wrapper_name = gen.wrapper_name()
+
+    gen.emit_header()
+    if fn.name not in reverse_op_methods and fn.name in reverse_op_method_names:
+        # There's only a reverse operator method.
+        generate_bin_op_reverse_only_wrapper(fn, emitter, gen)
+    else:
+        rmethod = reverse_op_methods[fn.name]
+        fn_rev = cl.get_method(rmethod)
+        if fn_rev is None:
+            # There's only a forward operator method.
+            generate_bin_op_forward_only_wrapper(fn, emitter, gen)
+        else:
+            # There's both a forward and a reverse operator method.
+            generate_bin_op_both_wrappers(cl, fn, fn_rev, emitter, gen)
+    return wrapper_name
+
+
+def generate_bin_op_forward_only_wrapper(
+    fn: FuncIR, emitter: Emitter, gen: WrapperGenerator
+) -> None:
+    gen.emit_arg_processing(error=GotoHandler("typefail"), raise_exception=False)
+    handle_third_pow_argument(fn, emitter, gen, if_unsupported=["goto typefail;"])
+    gen.emit_call(not_implemented_handler="goto typefail;")
+    gen.emit_error_handling()
+    emitter.emit_label("typefail")
+    # If some argument has an incompatible type, treat this the same as
+    # returning NotImplemented, and try to call the reverse operator method.
+    #
+    # Note that in normal Python you'd instead of an explicit
+    # return of NotImplemented, but it doesn't generally work here
+    # the body won't be executed at all if there is an argument
+    # type check failure.
+    #
+    # The recommended way is to still use a type check in the
+    # body. This will only be used in interpreted mode:
+    #
+    #    def __add__(self, other: int) -> Foo:
+    #        if not isinstance(other, int):
+    #            return NotImplemented
+    #        ...
+    generate_bin_op_reverse_dunder_call(fn, emitter, reverse_op_methods[fn.name])
+    gen.finish()
+
+
+def generate_bin_op_reverse_only_wrapper(
+    fn: FuncIR, emitter: Emitter, gen: WrapperGenerator
+) -> None:
+    gen.arg_names = ["right", "left"]
+    gen.emit_arg_processing(error=GotoHandler("typefail"), raise_exception=False)
+    handle_third_pow_argument(fn, emitter, gen, if_unsupported=["goto typefail;"])
+    gen.emit_call()
+    gen.emit_error_handling()
+    emitter.emit_label("typefail")
+    emitter.emit_line("Py_INCREF(Py_NotImplemented);")
+    emitter.emit_line("return Py_NotImplemented;")
+    gen.finish()
+
+
+def generate_bin_op_both_wrappers(
+    cl: ClassIR, fn: FuncIR, fn_rev: FuncIR, emitter: Emitter, gen: WrapperGenerator
+) -> None:
+    # There's both a forward and a reverse operator method. First
+    # check if we should try calling the forward one. If the
+    # argument type check fails, fall back to the reverse method.
+    #
+    # Similar to above, we can't perfectly match Python semantics.
+    # In regular Python code you'd return NotImplemented if the
+    # operand has the wrong type, but in compiled code we'll never
+    # get to execute the type check.
+    emitter.emit_line(
+        "if (PyObject_IsInstance(obj_left, (PyObject *){})) {{".format(
+            emitter.type_struct_name(cl)
+        )
+    )
+    gen.emit_arg_processing(error=GotoHandler("typefail"), raise_exception=False)
+    handle_third_pow_argument(fn, emitter, gen, if_unsupported=["goto typefail2;"])
+    # Ternary __rpow__ calls aren't a thing so immediately bail
+    # if ternary __pow__ returns NotImplemented.
+    if fn.name == "__pow__" and len(fn.args) == 3:
+        fwd_not_implemented_handler = "goto typefail2;"
+    else:
+        fwd_not_implemented_handler = "goto typefail;"
+    gen.emit_call(not_implemented_handler=fwd_not_implemented_handler)
+    gen.emit_error_handling()
+    emitter.emit_line("}")
+    emitter.emit_label("typefail")
+    emitter.emit_line(
+        "if (PyObject_IsInstance(obj_right, (PyObject *){})) {{".format(
+            emitter.type_struct_name(cl)
+        )
+    )
+    gen.set_target(fn_rev)
+    gen.arg_names = ["right", "left"]
+    gen.emit_arg_processing(error=GotoHandler("typefail2"), raise_exception=False)
+    handle_third_pow_argument(fn_rev, emitter, gen, if_unsupported=["goto typefail2;"])
+    gen.emit_call()
+    gen.emit_error_handling()
+    emitter.emit_line("} else {")
+    generate_bin_op_reverse_dunder_call(fn, emitter, fn_rev.name)
+    emitter.emit_line("}")
+    emitter.emit_label("typefail2")
+    emitter.emit_line("Py_INCREF(Py_NotImplemented);")
+    emitter.emit_line("return Py_NotImplemented;")
+    gen.finish()
+
+
+def generate_bin_op_reverse_dunder_call(fn: FuncIR, emitter: Emitter, rmethod: str) -> None:
+    if fn.name in ("__pow__", "__rpow__"):
+        # Ternary pow() will never call the reverse dunder.
+        emitter.emit_line("if (obj_mod == Py_None) {")
+    emitter.emit_line(f"_Py_IDENTIFIER({rmethod});")
+    emitter.emit_line(
+        'return CPy_CallReverseOpMethod(obj_left, obj_right, "{}", &PyId_{});'.format(
+            op_methods_to_symbols[fn.name], rmethod
+        )
+    )
+    if fn.name in ("__pow__", "__rpow__"):
+        emitter.emit_line("} else {")
+        emitter.emit_line("Py_INCREF(Py_NotImplemented);")
+        emitter.emit_line("return Py_NotImplemented;")
+        emitter.emit_line("}")
+
+
+def handle_third_pow_argument(
+    fn: FuncIR, emitter: Emitter, gen: WrapperGenerator, *, if_unsupported: list[str]
+) -> None:
+    if fn.name not in ("__pow__", "__rpow__", "__ipow__"):
+        return
+
+    if (fn.name in ("__pow__", "__ipow__") and len(fn.args) == 2) or fn.name == "__rpow__":
+        # If the power dunder only supports two arguments and the third
+        # argument (AKA mod) is set to a non-default value, simply bail.
+        #
+        # Importantly, this prevents any ternary __rpow__ calls from
+        # happening (as per the language specification).
+        emitter.emit_line("if (obj_mod != Py_None) {")
+        for line in if_unsupported:
+            emitter.emit_line(line)
+        emitter.emit_line("}")
+        # The slot wrapper will receive three arguments, but the call only
+        # supports two so make sure that the third argument isn't passed
+        # along. This is needed as two-argument __(i)pow__ is allowed and
+        # rather common.
+        if len(gen.arg_names) == 3:
+            gen.arg_names.pop()
+
+
+RICHCOMPARE_OPS = {
+    "__lt__": "Py_LT",
+    "__gt__": "Py_GT",
+    "__le__": "Py_LE",
+    "__ge__": "Py_GE",
+    "__eq__": "Py_EQ",
+    "__ne__": "Py_NE",
+}
+
+
+def generate_richcompare_wrapper(cl: ClassIR, emitter: Emitter) -> str | None:
+    """Generates a wrapper for richcompare dunder methods."""
+    # Sort for determinism on Python 3.5
+    matches = sorted(name for name in RICHCOMPARE_OPS if cl.has_method(name))
+    if not matches:
+        return None
+
+    name = f"{DUNDER_PREFIX}_RichCompare_{cl.name_prefix(emitter.names)}"
+    emitter.emit_line(
+        "static PyObject *{name}(PyObject *obj_lhs, PyObject *obj_rhs, int op) {{".format(
+            name=name
+        )
+    )
+    emitter.emit_line("switch (op) {")
+    for func in matches:
+        emitter.emit_line(f"case {RICHCOMPARE_OPS[func]}: {{")
+        method = cl.get_method(func)
+        assert method is not None
+        generate_wrapper_core(method, emitter, arg_names=["lhs", "rhs"])
+        emitter.emit_line("}")
+    emitter.emit_line("}")
+
+    emitter.emit_line("Py_INCREF(Py_NotImplemented);")
+    emitter.emit_line("return Py_NotImplemented;")
+
+    emitter.emit_line("}")
+
+    return name
+
+
+def generate_get_wrapper(cl: ClassIR, fn: FuncIR, emitter: Emitter) -> str:
+    """Generates a wrapper for native __get__ methods."""
+    name = f"{DUNDER_PREFIX}{fn.name}{cl.name_prefix(emitter.names)}"
+    emitter.emit_line(
+        "static PyObject *{name}(PyObject *self, PyObject *instance, PyObject *owner) {{".format(
+            name=name
+        )
+    )
+    emitter.emit_line("instance = instance ? instance : Py_None;")
+    emitter.emit_line(f"return {NATIVE_PREFIX}{fn.cname(emitter.names)}(self, instance, owner);")
+    emitter.emit_line("}")
+
+    return name
+
+
+def generate_hash_wrapper(cl: ClassIR, fn: FuncIR, emitter: Emitter) -> str:
+    """Generates a wrapper for native __hash__ methods."""
+    name = f"{DUNDER_PREFIX}{fn.name}{cl.name_prefix(emitter.names)}"
+    emitter.emit_line(f"static Py_ssize_t {name}(PyObject *self) {{")
+    emitter.emit_line(
+        "{}retval = {}{}{}(self);".format(
+            emitter.ctype_spaced(fn.ret_type),
+            emitter.get_group_prefix(fn.decl),
+            NATIVE_PREFIX,
+            fn.cname(emitter.names),
+        )
+    )
+    emitter.emit_error_check("retval", fn.ret_type, "return -1;")
+    if is_int_rprimitive(fn.ret_type):
+        emitter.emit_line("Py_ssize_t val = CPyTagged_AsSsize_t(retval);")
+    else:
+        emitter.emit_line("Py_ssize_t val = PyLong_AsSsize_t(retval);")
+    emitter.emit_dec_ref("retval", fn.ret_type)
+    emitter.emit_line("if (PyErr_Occurred()) return -1;")
+    # We can't return -1 from a hash function..
+    emitter.emit_line("if (val == -1) return -2;")
+    emitter.emit_line("return val;")
+    emitter.emit_line("}")
+
+    return name
+
+
+def generate_len_wrapper(cl: ClassIR, fn: FuncIR, emitter: Emitter) -> str:
+    """Generates a wrapper for native __len__ methods."""
+    name = f"{DUNDER_PREFIX}{fn.name}{cl.name_prefix(emitter.names)}"
+    emitter.emit_line(f"static Py_ssize_t {name}(PyObject *self) {{")
+    emitter.emit_line(
+        "{}retval = {}{}{}(self);".format(
+            emitter.ctype_spaced(fn.ret_type),
+            emitter.get_group_prefix(fn.decl),
+            NATIVE_PREFIX,
+            fn.cname(emitter.names),
+        )
+    )
+    emitter.emit_error_check("retval", fn.ret_type, "return -1;")
+    if is_int_rprimitive(fn.ret_type):
+        emitter.emit_line("Py_ssize_t val = CPyTagged_AsSsize_t(retval);")
+    else:
+        emitter.emit_line("Py_ssize_t val = PyLong_AsSsize_t(retval);")
+    emitter.emit_dec_ref("retval", fn.ret_type)
+    emitter.emit_line("if (PyErr_Occurred()) return -1;")
+    emitter.emit_line("return val;")
+    emitter.emit_line("}")
+
+    return name
+
+
+def generate_bool_wrapper(cl: ClassIR, fn: FuncIR, emitter: Emitter) -> str:
+    """Generates a wrapper for native __bool__ methods."""
+    name = f"{DUNDER_PREFIX}{fn.name}{cl.name_prefix(emitter.names)}"
+    emitter.emit_line(f"static int {name}(PyObject *self) {{")
+    emitter.emit_line(
+        "{}val = {}{}(self);".format(
+            emitter.ctype_spaced(fn.ret_type), NATIVE_PREFIX, fn.cname(emitter.names)
+        )
+    )
+    emitter.emit_error_check("val", fn.ret_type, "return -1;")
+    # This wouldn't be that hard to fix but it seems unimportant and
+    # getting error handling and unboxing right would be fiddly. (And
+    # way easier to do in IR!)
+    assert is_bool_rprimitive(fn.ret_type), "Only bool return supported for __bool__"
+    emitter.emit_line("return val;")
+    emitter.emit_line("}")
+
+    return name
+
+
+def generate_del_item_wrapper(cl: ClassIR, fn: FuncIR, emitter: Emitter) -> str:
+    """Generates a wrapper for native __delitem__.
+
+    This is only called from a combined __delitem__/__setitem__ wrapper.
+    """
+    name = "{}{}{}".format(DUNDER_PREFIX, "__delitem__", cl.name_prefix(emitter.names))
+    input_args = ", ".join(f"PyObject *obj_{arg.name}" for arg in fn.args)
+    emitter.emit_line(f"static int {name}({input_args}) {{")
+    generate_set_del_item_wrapper_inner(fn, emitter, fn.args)
+    return name
+
+
+def generate_set_del_item_wrapper(cl: ClassIR, fn: FuncIR, emitter: Emitter) -> str:
+    """Generates a wrapper for native __setitem__ method (also works for __delitem__).
+
+    This is used with the mapping protocol slot. Arguments are taken as *PyObjects and we
+    return a negative C int on error.
+
+    Create a separate wrapper function for __delitem__ as needed and have the
+    __setitem__ wrapper call it if the value is NULL. Return the name
+    of the outer (__setitem__) wrapper.
+    """
+    method_cls = cl.get_method_and_class("__delitem__")
+    del_name = None
+    if method_cls and method_cls[1] == cl:
+        # Generate a separate wrapper for __delitem__
+        del_name = generate_del_item_wrapper(cl, method_cls[0], emitter)
+
+    args = fn.args
+    if fn.name == "__delitem__":
+        # Add an extra argument for value that we expect to be NULL.
+        args = list(args) + [RuntimeArg("___value", object_rprimitive, ARG_POS)]
+
+    name = "{}{}{}".format(DUNDER_PREFIX, "__setitem__", cl.name_prefix(emitter.names))
+    input_args = ", ".join(f"PyObject *obj_{arg.name}" for arg in args)
+    emitter.emit_line(f"static int {name}({input_args}) {{")
+
+    # First check if this is __delitem__
+    emitter.emit_line(f"if (obj_{args[2].name} == NULL) {{")
+    if del_name is not None:
+        # We have a native implementation, so call it
+        emitter.emit_line(f"return {del_name}(obj_{args[0].name}, obj_{args[1].name});")
+    else:
+        # Try to call superclass method instead
+        emitter.emit_line(f"PyObject *super = CPy_Super(CPyModule_builtins, obj_{args[0].name});")
+        emitter.emit_line("if (super == NULL) return -1;")
+        emitter.emit_line(
+            'PyObject *result = PyObject_CallMethod(super, "__delitem__", "O", obj_{});'.format(
+                args[1].name
+            )
+        )
+        emitter.emit_line("Py_DECREF(super);")
+        emitter.emit_line("Py_XDECREF(result);")
+        emitter.emit_line("return result == NULL ? -1 : 0;")
+    emitter.emit_line("}")
+
+    method_cls = cl.get_method_and_class("__setitem__")
+    if method_cls and method_cls[1] == cl:
+        generate_set_del_item_wrapper_inner(fn, emitter, args)
+    else:
+        emitter.emit_line(f"PyObject *super = CPy_Super(CPyModule_builtins, obj_{args[0].name});")
+        emitter.emit_line("if (super == NULL) return -1;")
+        emitter.emit_line("PyObject *result;")
+
+        if method_cls is None and cl.builtin_base is None:
+            msg = f"'{cl.name}' object does not support item assignment"
+            emitter.emit_line(f'PyErr_SetString(PyExc_TypeError, "{msg}");')
+            emitter.emit_line("result = NULL;")
+        else:
+            # A base class may have __setitem__
+            emitter.emit_line(
+                'result = PyObject_CallMethod(super, "__setitem__", "OO", obj_{}, obj_{});'.format(
+                    args[1].name, args[2].name
+                )
+            )
+        emitter.emit_line("Py_DECREF(super);")
+        emitter.emit_line("Py_XDECREF(result);")
+        emitter.emit_line("return result == NULL ? -1 : 0;")
+        emitter.emit_line("}")
+    return name
+
+
+def generate_set_del_item_wrapper_inner(
+    fn: FuncIR, emitter: Emitter, args: Sequence[RuntimeArg]
+) -> None:
+    for arg in args:
+        generate_arg_check(arg.name, arg.type, emitter, GotoHandler("fail"))
+    native_args = ", ".join(f"arg_{arg.name}" for arg in args)
+    emitter.emit_line(
+        "{}val = {}{}({});".format(
+            emitter.ctype_spaced(fn.ret_type), NATIVE_PREFIX, fn.cname(emitter.names), native_args
+        )
+    )
+    emitter.emit_error_check("val", fn.ret_type, "goto fail;")
+    emitter.emit_dec_ref("val", fn.ret_type)
+    emitter.emit_line("return 0;")
+    emitter.emit_label("fail")
+    emitter.emit_line("return -1;")
+    emitter.emit_line("}")
+
+
+def generate_contains_wrapper(cl: ClassIR, fn: FuncIR, emitter: Emitter) -> str:
+    """Generates a wrapper for a native __contains__ method."""
+    name = f"{DUNDER_PREFIX}{fn.name}{cl.name_prefix(emitter.names)}"
+    emitter.emit_line(f"static int {name}(PyObject *self, PyObject *obj_item) {{")
+    generate_arg_check("item", fn.args[1].type, emitter, ReturnHandler("-1"))
+    emitter.emit_line(
+        "{}val = {}{}(self, arg_item);".format(
+            emitter.ctype_spaced(fn.ret_type), NATIVE_PREFIX, fn.cname(emitter.names)
+        )
+    )
+    emitter.emit_error_check("val", fn.ret_type, "return -1;")
+    if is_bool_rprimitive(fn.ret_type):
+        emitter.emit_line("return val;")
+    else:
+        emitter.emit_line("int boolval = PyObject_IsTrue(val);")
+        emitter.emit_dec_ref("val", fn.ret_type)
+        emitter.emit_line("return boolval;")
+    emitter.emit_line("}")
+
+    return name
+
+
+# Helpers
+
+
+def generate_wrapper_core(
+    fn: FuncIR,
+    emitter: Emitter,
+    optional_args: list[RuntimeArg] | None = None,
+    arg_names: list[str] | None = None,
+    cleanups: list[str] | None = None,
+    traceback_code: str | None = None,
+) -> None:
+    """Generates the core part of a wrapper function for a native function.
+
+    This expects each argument as a PyObject * named obj_{arg} as a precondition.
+    It converts the PyObject *s to the necessary types, checking and unboxing if necessary,
+    makes the call, then boxes the result if necessary and returns it.
+    """
+    gen = WrapperGenerator(None, emitter)
+    gen.set_target(fn)
+    if arg_names:
+        gen.arg_names = arg_names
+    gen.cleanups = cleanups or []
+    gen.optional_args = optional_args or []
+    gen.traceback_code = traceback_code or ""
+
+    error = ReturnHandler("NULL") if not gen.use_goto() else GotoHandler("fail")
+    gen.emit_arg_processing(error=error)
+    gen.emit_call()
+    gen.emit_error_handling()
+
+
+def generate_arg_check(
+    name: str,
+    typ: RType,
+    emitter: Emitter,
+    error: ErrorHandler | None = None,
+    *,
+    optional: bool = False,
+    raise_exception: bool = True,
+    bitmap_arg_index: int = 0,
+) -> None:
+    """Insert a runtime check for argument and unbox if necessary.
+
+    The object is named PyObject *obj_{}. This is expected to generate
+    a value of name arg_{} (unboxed if necessary). For each primitive a runtime
+    check ensures the correct type.
+    """
+    error = error or AssignHandler()
+    if typ.is_unboxed:
+        if typ.error_overlap and optional:
+            # Update bitmap is value is provided.
+            init = emitter.c_undefined_value(typ)
+            emitter.emit_line(f"{emitter.ctype(typ)} arg_{name} = {init};")
+            emitter.emit_line(f"if (obj_{name} != NULL) {{")
+            bitmap = bitmap_name(bitmap_arg_index // BITMAP_BITS)
+            emitter.emit_line(f"{bitmap} |= 1 << {bitmap_arg_index & (BITMAP_BITS - 1)};")
+            emitter.emit_unbox(
+                f"obj_{name}",
+                f"arg_{name}",
+                typ,
+                declare_dest=False,
+                raise_exception=raise_exception,
+                error=error,
+                borrow=True,
+            )
+            emitter.emit_line("}")
+        else:
+            # Borrow when unboxing to avoid reference count manipulation.
+            emitter.emit_unbox(
+                f"obj_{name}",
+                f"arg_{name}",
+                typ,
+                declare_dest=True,
+                raise_exception=raise_exception,
+                error=error,
+                borrow=True,
+                optional=optional,
+            )
+    elif is_object_rprimitive(typ):
+        # Object is trivial since any object is valid
+        if optional:
+            emitter.emit_line(f"PyObject *arg_{name};")
+            emitter.emit_line(f"if (obj_{name} == NULL) {{")
+            emitter.emit_line(f"arg_{name} = {emitter.c_error_value(typ)};")
+            emitter.emit_lines("} else {", f"arg_{name} = obj_{name}; ", "}")
+        else:
+            emitter.emit_line(f"PyObject *arg_{name} = obj_{name};")
+    else:
+        emitter.emit_cast(
+            f"obj_{name}",
+            f"arg_{name}",
+            typ,
+            declare_dest=True,
+            raise_exception=raise_exception,
+            error=error,
+            optional=optional,
+        )
+
+
+class WrapperGenerator:
+    """Helper that simplifies the generation of wrapper functions."""
+
+    # TODO: Use this for more wrappers
+
+    def __init__(self, cl: ClassIR | None, emitter: Emitter) -> None:
+        self.cl = cl
+        self.emitter = emitter
+        self.cleanups: list[str] = []
+        self.optional_args: list[RuntimeArg] = []
+        self.traceback_code = ""
+
+    def set_target(self, fn: FuncIR) -> None:
+        """Set the wrapped function.
+
+        It's fine to modify the attributes initialized here later to customize
+        the wrapper function.
+        """
+        self.target_name = fn.name
+        self.target_cname = fn.cname(self.emitter.names)
+        self.num_bitmap_args = fn.sig.num_bitmap_args
+        if self.num_bitmap_args:
+            self.args = fn.args[: -self.num_bitmap_args]
+        else:
+            self.args = fn.args
+        self.arg_names = [arg.name for arg in self.args]
+        self.ret_type = fn.ret_type
+
+    def wrapper_name(self) -> str:
+        """Return the name of the wrapper function."""
+        return "{}{}{}".format(
+            DUNDER_PREFIX,
+            self.target_name,
+            self.cl.name_prefix(self.emitter.names) if self.cl else "",
+        )
+
+    def use_goto(self) -> bool:
+        """Do we use a goto for error handling (instead of straight return)?"""
+        return bool(self.cleanups or self.traceback_code)
+
+    def emit_header(self) -> None:
+        """Emit the function header of the wrapper implementation."""
+        input_args = ", ".join(f"PyObject *obj_{arg}" for arg in self.arg_names)
+        self.emitter.emit_line(
+            "static PyObject *{name}({input_args}) {{".format(
+                name=self.wrapper_name(), input_args=input_args
+            )
+        )
+
+    def emit_arg_processing(
+        self, error: ErrorHandler | None = None, raise_exception: bool = True
+    ) -> None:
+        """Emit validation and unboxing of arguments."""
+        error = error or self.error()
+        bitmap_arg_index = 0
+        for arg_name, arg in zip(self.arg_names, self.args):
+            # Suppress the argument check for *args/**kwargs, since we know it must be right.
+            typ = arg.type if arg.kind not in (ARG_STAR, ARG_STAR2) else object_rprimitive
+            optional = arg in self.optional_args
+            generate_arg_check(
+                arg_name,
+                typ,
+                self.emitter,
+                error,
+                raise_exception=raise_exception,
+                optional=optional,
+                bitmap_arg_index=bitmap_arg_index,
+            )
+            if optional and typ.error_overlap:
+                bitmap_arg_index += 1
+
+    def emit_call(self, not_implemented_handler: str = "") -> None:
+        """Emit call to the wrapper function.
+
+        If not_implemented_handler is non-empty, use this C code to handle
+        a NotImplemented return value (if it's possible based on the return type).
+        """
+        native_args = ", ".join(f"arg_{arg}" for arg in self.arg_names)
+        if self.num_bitmap_args:
+            bitmap_args = ", ".join(
+                [bitmap_name(i) for i in reversed(range(self.num_bitmap_args))]
+            )
+            native_args = f"{native_args}, {bitmap_args}"
+
+        ret_type = self.ret_type
+        emitter = self.emitter
+        if ret_type.is_unboxed or self.use_goto():
+            # TODO: The Py_RETURN macros return the correct PyObject * with reference count
+            #       handling. Are they relevant?
+            emitter.emit_line(
+                "{}retval = {}{}({});".format(
+                    emitter.ctype_spaced(ret_type), NATIVE_PREFIX, self.target_cname, native_args
+                )
+            )
+            emitter.emit_lines(*self.cleanups)
+            if ret_type.is_unboxed:
+                emitter.emit_error_check("retval", ret_type, "return NULL;")
+                emitter.emit_box("retval", "retbox", ret_type, declare_dest=True)
+
+            emitter.emit_line("return {};".format("retbox" if ret_type.is_unboxed else "retval"))
+        else:
+            if not_implemented_handler and not isinstance(ret_type, RInstance):
+                # The return value type may overlap with NotImplemented.
+                emitter.emit_line(
+                    "PyObject *retbox = {}{}({});".format(
+                        NATIVE_PREFIX, self.target_cname, native_args
+                    )
+                )
+                emitter.emit_lines(
+                    "if (retbox == Py_NotImplemented) {",
+                    not_implemented_handler,
+                    "}",
+                    "return retbox;",
+                )
+            else:
+                emitter.emit_line(f"return {NATIVE_PREFIX}{self.target_cname}({native_args});")
+            # TODO: Tracebacks?
+
+    def error(self) -> ErrorHandler:
+        """Figure out how to deal with errors in the wrapper."""
+        if self.cleanups or self.traceback_code:
+            # We'll have a label at the end with error handling code.
+            return GotoHandler("fail")
+        else:
+            # Nothing special needs to done to handle errors, so just return.
+            return ReturnHandler("NULL")
+
+    def emit_error_handling(self) -> None:
+        """Emit error handling block at the end of the wrapper, if needed."""
+        emitter = self.emitter
+        if self.use_goto():
+            emitter.emit_label("fail")
+            emitter.emit_lines(*self.cleanups)
+            if self.traceback_code:
+                emitter.emit_line(self.traceback_code)
+            emitter.emit_line("return NULL;")
+
+    def finish(self) -> None:
+        self.emitter.emit_line("}")

venv/lib/python3.12/site-packages/mypyc/codegen/literals.py

@@ -0,0 +1,302 @@
+from __future__ import annotations
+
+from typing import Final, FrozenSet, Tuple, Union
+from typing_extensions import TypeGuard
+
+# Supported Python literal types. All tuple / frozenset items must have supported
+# literal types as well, but we can't represent the type precisely.
+LiteralValue = Union[
+    str, bytes, int, bool, float, complex, Tuple[object, ...], FrozenSet[object], None
+]
+
+
+def _is_literal_value(obj: object) -> TypeGuard[LiteralValue]:
+    return isinstance(obj, (str, bytes, int, float, complex, tuple, frozenset, type(None)))
+
+
+# Some literals are singletons and handled specially (None, False and True)
+NUM_SINGLETONS: Final = 3
+
+
+class Literals:
+    """Collection of literal values used in a compilation group and related helpers."""
+
+    def __init__(self) -> None:
+        # Each dict maps value to literal index (0, 1, ...)
+        self.str_literals: dict[str, int] = {}
+        self.bytes_literals: dict[bytes, int] = {}
+        self.int_literals: dict[int, int] = {}
+        self.float_literals: dict[float, int] = {}
+        self.complex_literals: dict[complex, int] = {}
+        self.tuple_literals: dict[tuple[object, ...], int] = {}
+        self.frozenset_literals: dict[frozenset[object], int] = {}
+
+    def record_literal(self, value: LiteralValue) -> None:
+        """Ensure that the literal value is available in generated code."""
+        if value is None or value is True or value is False:
+            # These are special cased and always present
+            return
+        if isinstance(value, str):
+            str_literals = self.str_literals
+            if value not in str_literals:
+                str_literals[value] = len(str_literals)
+        elif isinstance(value, bytes):
+            bytes_literals = self.bytes_literals
+            if value not in bytes_literals:
+                bytes_literals[value] = len(bytes_literals)
+        elif isinstance(value, int):
+            int_literals = self.int_literals
+            if value not in int_literals:
+                int_literals[value] = len(int_literals)
+        elif isinstance(value, float):
+            float_literals = self.float_literals
+            if value not in float_literals:
+                float_literals[value] = len(float_literals)
+        elif isinstance(value, complex):
+            complex_literals = self.complex_literals
+            if value not in complex_literals:
+                complex_literals[value] = len(complex_literals)
+        elif isinstance(value, tuple):
+            tuple_literals = self.tuple_literals
+            if value not in tuple_literals:
+                for item in value:
+                    assert _is_literal_value(item)
+                    self.record_literal(item)
+                tuple_literals[value] = len(tuple_literals)
+        elif isinstance(value, frozenset):
+            frozenset_literals = self.frozenset_literals
+            if value not in frozenset_literals:
+                for item in value:
+                    assert _is_literal_value(item)
+                    self.record_literal(item)
+                frozenset_literals[value] = len(frozenset_literals)
+        else:
+            assert False, "invalid literal: %r" % value
+
+    def literal_index(self, value: LiteralValue) -> int:
+        """Return the index to the literals array for given value."""
+        # The array contains first None and booleans, followed by all str values,
+        # followed by bytes values, etc.
+        if value is None:
+            return 0
+        elif value is False:
+            return 1
+        elif value is True:
+            return 2
+        n = NUM_SINGLETONS
+        if isinstance(value, str):
+            return n + self.str_literals[value]
+        n += len(self.str_literals)
+        if isinstance(value, bytes):
+            return n + self.bytes_literals[value]
+        n += len(self.bytes_literals)
+        if isinstance(value, int):
+            return n + self.int_literals[value]
+        n += len(self.int_literals)
+        if isinstance(value, float):
+            return n + self.float_literals[value]
+        n += len(self.float_literals)
+        if isinstance(value, complex):
+            return n + self.complex_literals[value]
+        n += len(self.complex_literals)
+        if isinstance(value, tuple):
+            return n + self.tuple_literals[value]
+        n += len(self.tuple_literals)
+        if isinstance(value, frozenset):
+            return n + self.frozenset_literals[value]
+        assert False, "invalid literal: %r" % value
+
+    def num_literals(self) -> int:
+        # The first three are for None, True and False
+        return (
+            NUM_SINGLETONS
+            + len(self.str_literals)
+            + len(self.bytes_literals)
+            + len(self.int_literals)
+            + len(self.float_literals)
+            + len(self.complex_literals)
+            + len(self.tuple_literals)
+            + len(self.frozenset_literals)
+        )
+
+    # The following methods return the C encodings of literal values
+    # of different types
+
+    def encoded_str_values(self) -> list[bytes]:
+        return _encode_str_values(self.str_literals)
+
+    def encoded_int_values(self) -> list[bytes]:
+        return _encode_int_values(self.int_literals)
+
+    def encoded_bytes_values(self) -> list[bytes]:
+        return _encode_bytes_values(self.bytes_literals)
+
+    def encoded_float_values(self) -> list[str]:
+        return _encode_float_values(self.float_literals)
+
+    def encoded_complex_values(self) -> list[str]:
+        return _encode_complex_values(self.complex_literals)
+
+    def encoded_tuple_values(self) -> list[str]:
+        return self._encode_collection_values(self.tuple_literals)
+
+    def encoded_frozenset_values(self) -> list[str]:
+        return self._encode_collection_values(self.frozenset_literals)
+
+    def _encode_collection_values(
+        self, values: dict[tuple[object, ...], int] | dict[frozenset[object], int]
+    ) -> list[str]:
+        """Encode tuple/frozenset values into a C array.
+
+        The format of the result is like this:
+
+           <number of collections>
+           <length of the first collection>
+           <literal index of first item>
+           ...
+           <literal index of last item>
+           <length of the second collection>
+           ...
+        """
+        value_by_index = {index: value for value, index in values.items()}
+        result = []
+        count = len(values)
+        result.append(str(count))
+        for i in range(count):
+            value = value_by_index[i]
+            result.append(str(len(value)))
+            for item in value:
+                assert _is_literal_value(item)
+                index = self.literal_index(item)
+                result.append(str(index))
+        return result
+
+
+def _encode_str_values(values: dict[str, int]) -> list[bytes]:
+    value_by_index = {index: value for value, index in values.items()}
+    result = []
+    line: list[bytes] = []
+    line_len = 0
+    for i in range(len(values)):
+        value = value_by_index[i]
+        c_literal = format_str_literal(value)
+        c_len = len(c_literal)
+        if line_len > 0 and line_len + c_len > 70:
+            result.append(format_int(len(line)) + b"".join(line))
+            line = []
+            line_len = 0
+        line.append(c_literal)
+        line_len += c_len
+    if line:
+        result.append(format_int(len(line)) + b"".join(line))
+    result.append(b"")
+    return result
+
+
+def _encode_bytes_values(values: dict[bytes, int]) -> list[bytes]:
+    value_by_index = {index: value for value, index in values.items()}
+    result = []
+    line: list[bytes] = []
+    line_len = 0
+    for i in range(len(values)):
+        value = value_by_index[i]
+        c_init = format_int(len(value))
+        c_len = len(c_init) + len(value)
+        if line_len > 0 and line_len + c_len > 70:
+            result.append(format_int(len(line)) + b"".join(line))
+            line = []
+            line_len = 0
+        line.append(c_init + value)
+        line_len += c_len
+    if line:
+        result.append(format_int(len(line)) + b"".join(line))
+    result.append(b"")
+    return result
+
+
+def format_int(n: int) -> bytes:
+    """Format an integer using a variable-length binary encoding."""
+    if n < 128:
+        a = [n]
+    else:
+        a = []
+        while n > 0:
+            a.insert(0, n & 0x7F)
+            n >>= 7
+        for i in range(len(a) - 1):
+            # If the highest bit is set, more 7-bit digits follow
+            a[i] |= 0x80
+    return bytes(a)
+
+
+def format_str_literal(s: str) -> bytes:
+    utf8 = s.encode("utf-8")
+    return format_int(len(utf8)) + utf8
+
+
+def _encode_int_values(values: dict[int, int]) -> list[bytes]:
+    """Encode int values into C strings.
+
+    Values are stored in base 10 and separated by 0 bytes.
+    """
+    value_by_index = {index: value for value, index in values.items()}
+    result = []
+    line: list[bytes] = []
+    line_len = 0
+    for i in range(len(values)):
+        value = value_by_index[i]
+        encoded = b"%d" % value
+        if line_len > 0 and line_len + len(encoded) > 70:
+            result.append(format_int(len(line)) + b"\0".join(line))
+            line = []
+            line_len = 0
+        line.append(encoded)
+        line_len += len(encoded)
+    if line:
+        result.append(format_int(len(line)) + b"\0".join(line))
+    result.append(b"")
+    return result
+
+
+def float_to_c(x: float) -> str:
+    """Return C literal representation of a float value."""
+    s = str(x)
+    if s == "inf":
+        return "INFINITY"
+    elif s == "-inf":
+        return "-INFINITY"
+    elif s == "nan":
+        return "NAN"
+    return s
+
+
+def _encode_float_values(values: dict[float, int]) -> list[str]:
+    """Encode float values into a C array values.
+
+    The result contains the number of values followed by individual values.
+    """
+    value_by_index = {index: value for value, index in values.items()}
+    result = []
+    num = len(values)
+    result.append(str(num))
+    for i in range(num):
+        value = value_by_index[i]
+        result.append(float_to_c(value))
+    return result
+
+
+def _encode_complex_values(values: dict[complex, int]) -> list[str]:
+    """Encode float values into a C array values.
+
+    The result contains the number of values followed by pairs of doubles
+    representing complex numbers.
+    """
+    value_by_index = {index: value for value, index in values.items()}
+    result = []
+    num = len(values)
+    result.append(str(num))
+    for i in range(num):
+        value = value_by_index[i]
+        result.append(float_to_c(value.real))
+        result.append(float_to_c(value.imag))
+    return result