ENH: Improve array function overhead by using vectorcall

This moves dispatching for `__array_function__` into a C-wrapper.  This
helps speed for multiple reasons:
* Avoids one additional dispatching function call to C
* Avoids the use of `*args, **kwargs` which is slower.
* For simple NumPy calls we can stay in the faster "vectorcall" world

This speeds up things generally a little, but can speed things up a lot
when keyword arguments are used on lightweight functions, for example::

    np.can_cast(arr, dtype, casting="same_kind")

is more than twice as fast with this.

There is one alternative in principle to get best speed:  We could inline
the "relevant argument"/dispatcher extraction.  That changes behavior in
an acceptable but larger way (passes default arguments).
Unless the C-entry point seems unwanted, this should be a decent step
in the right direction even if we want to do that eventually, though.

Closes gh-20790
Closes gh-18547  (although not quite sure why)

8 files changed, 413 insertions, 346 deletions

diff --git a/numpy/core/_asarray.py b/numpy/core/_asarray.py
index cbaab8c3f..a9abc5a88 100644
--- a/numpy/core/_asarray.py
+++ b/numpy/core/_asarray.py
@@ -24,10 +24,6 @@ POSSIBLE_FLAGS = {
 }
 
 
-def _require_dispatcher(a, dtype=None, requirements=None, *, like=None):
-    return (like,)
-
-
 @set_array_function_like_doc
 @set_module('numpy')
 def require(a, dtype=None, requirements=None, *, like=None):
@@ -100,10 +96,10 @@ def require(a, dtype=None, requirements=None, *, like=None):
     """
     if like is not None:
         return _require_with_like(
+            like,
             a,
             dtype=dtype,
             requirements=requirements,
-            like=like,
         )
 
     if not requirements:
@@ -135,6 +131,4 @@ def require(a, dtype=None, requirements=None, *, like=None):
     return arr
 
 
-_require_with_like = array_function_dispatch(
-    _require_dispatcher
-)(require)
+_require_with_like = array_function_dispatch()(require)
diff --git a/numpy/core/numeric.py b/numpy/core/numeric.py
index 577f8e7cd..91e35c684 100644
--- a/numpy/core/numeric.py
+++ b/numpy/core/numeric.py
@@ -130,10 +130,6 @@ def zeros_like(a, dtype=None, order='K', subok=True, shape=None):
     return res
 
 
-def _ones_dispatcher(shape, dtype=None, order=None, *, like=None):
-    return(like,)
-
-
 @set_array_function_like_doc
 @set_module('numpy')
 def ones(shape, dtype=None, order='C', *, like=None):
@@ -187,16 +183,13 @@ def ones(shape, dtype=None, order='C', *, like=None):
 
     """
     if like is not None:
-        return _ones_with_like(shape, dtype=dtype, order=order, like=like)
+        return _ones_with_like(like, shape, dtype=dtype, order=order)
 
     a = empty(shape, dtype, order)
     multiarray.copyto(a, 1, casting='unsafe')
     return a
 
-
-_ones_with_like = array_function_dispatch(
-    _ones_dispatcher
-)(ones)
+_ones_with_like = array_function_dispatch()(ones)
 
 
 def _ones_like_dispatcher(a, dtype=None, order=None, subok=None, shape=None):
@@ -323,7 +316,7 @@ def full(shape, fill_value, dtype=None, order='C', *, like=None):
 
     """
     if like is not None:
-        return _full_with_like(shape, fill_value, dtype=dtype, order=order, like=like)
+        return _full_with_like(like, shape, fill_value, dtype=dtype, order=order)
 
     if dtype is None:
         fill_value = asarray(fill_value)
@@ -333,9 +326,7 @@ def full(shape, fill_value, dtype=None, order='C', *, like=None):
     return a
 
 
-_full_with_like = array_function_dispatch(
-    _full_dispatcher
-)(full)
+_full_with_like = array_function_dispatch()(full)
 
 
 def _full_like_dispatcher(a, fill_value, dtype=None, order=None, subok=None, shape=None):
@@ -1778,10 +1769,6 @@ def indices(dimensions, dtype=int, sparse=False):
     return res
 
 
-def _fromfunction_dispatcher(function, shape, *, dtype=None, like=None, **kwargs):
-    return (like,)
-
-
 @set_array_function_like_doc
 @set_module('numpy')
 def fromfunction(function, shape, *, dtype=float, like=None, **kwargs):
@@ -1847,15 +1834,13 @@ def fromfunction(function, shape, *, dtype=float, like=None, **kwargs):
 
     """
     if like is not None:
-        return _fromfunction_with_like(function, shape, dtype=dtype, like=like, **kwargs)
+        return _fromfunction_with_like(like, function, shape, dtype=dtype, **kwargs)
 
     args = indices(shape, dtype=dtype)
     return function(*args, **kwargs)
 
 
-_fromfunction_with_like = array_function_dispatch(
-    _fromfunction_dispatcher
-)(fromfunction)
+_fromfunction_with_like = array_function_dispatch()(fromfunction)
 
 
 def _frombuffer(buf, dtype, shape, order):
@@ -2130,10 +2115,6 @@ def _maketup(descr, val):
         return tuple(res)
 
 
-def _identity_dispatcher(n, dtype=None, *, like=None):
-    return (like,)
-
-
 @set_array_function_like_doc
 @set_module('numpy')
 def identity(n, dtype=None, *, like=None):
@@ -2168,15 +2149,13 @@ def identity(n, dtype=None, *, like=None):
 
     """
     if like is not None:
-        return _identity_with_like(n, dtype=dtype, like=like)
+        return _identity_with_like(like, n, dtype=dtype)
 
     from numpy import eye
     return eye(n, dtype=dtype, like=like)
 
 
-_identity_with_like = array_function_dispatch(
-    _identity_dispatcher
-)(identity)
+_identity_with_like = array_function_dispatch()(identity)
 
 
 def _allclose_dispatcher(a, b, rtol=None, atol=None, equal_nan=None):
diff --git a/numpy/core/overrides.py b/numpy/core/overrides.py
index 46e1fbe2c..25892d5de 100644
--- a/numpy/core/overrides.py
+++ b/numpy/core/overrides.py
@@ -6,7 +6,7 @@ import os
 from .._utils import set_module
 from .._utils._inspect import getargspec
 from numpy.core._multiarray_umath import (
-    add_docstring, implement_array_function, _get_implementing_args)
+    add_docstring,  _get_implementing_args, _ArrayFunctionDispatcher)
 
 
 ARRAY_FUNCTIONS = set()
@@ -33,40 +33,33 @@ def set_array_function_like_doc(public_api):
 
 
 add_docstring(
-    implement_array_function,
+    _ArrayFunctionDispatcher,
     """
-    Implement a function with checks for __array_function__ overrides.
+    Class to wrap functions with checks for __array_function__ overrides.
 
     All arguments are required, and can only be passed by position.
 
     Parameters
     ----------
+    dispatcher : function or None
+        The dispatcher function that returns a single sequence-like object
+        of all arguments relevant.  It must have the same signature (except
+        the default values) as the actual implementation.
+        If ``None``, this is a ``like=`` dispatcher and the
+        ``_ArrayFunctionDispatcher`` must be called with ``like`` as the
+        first (additional and positional) argument.
     implementation : function
         Function that implements the operation on NumPy array without
-        overrides when called like ``implementation(*args, **kwargs)``.
-    public_api : function
-        Function exposed by NumPy's public API originally called like
-        ``public_api(*args, **kwargs)`` on which arguments are now being
-        checked.
-    relevant_args : iterable
-        Iterable of arguments to check for __array_function__ methods.
-    args : tuple
-        Arbitrary positional arguments originally passed into ``public_api``.
-    kwargs : dict
-        Arbitrary keyword arguments originally passed into ``public_api``.
+        overrides when called like.
 
-    Returns
-    -------
-    Result from calling ``implementation()`` or an ``__array_function__``
-    method, as appropriate.
-
-    Raises
-    ------
-    TypeError : if no implementation is found.
+    Attributes
+    ----------
+    _implementation : function
+        The original implementation passed in.
     """)
 
 
-# exposed for testing purposes; used internally by implement_array_function
+# exposed for testing purposes; used internally by _ArrayFunctionDispatcher
 add_docstring(
     _get_implementing_args,
     """
@@ -110,7 +103,7 @@ def verify_matching_signatures(implementation, dispatcher):
                                'default argument values')
 
 
-def array_function_dispatch(dispatcher, module=None, verify=True,
+def array_function_dispatch(dispatcher=None, module=None, verify=True,
                             docs_from_dispatcher=False):
     """Decorator for adding dispatch with the __array_function__ protocol.
 
@@ -118,10 +111,14 @@ def array_function_dispatch(dispatcher, module=None, verify=True,
 
     Parameters
     ----------
-    dispatcher : callable
+    dispatcher : callable or None
         Function that when called like ``dispatcher(*args, **kwargs)`` with
         arguments from the NumPy function call returns an iterable of
         array-like arguments to check for ``__array_function__``.
+
+        If `None`, the first argument is used as the single `like=` argument
+        and not passed on.  A function implementing `like=` must call its
+        dispatcher with `like` as the first non-keyword argument.
     module : str, optional
         __module__ attribute to set on new function, e.g., ``module='numpy'``.
         By default, module is copied from the decorated function.
@@ -154,45 +151,28 @@ def array_function_dispatch(dispatcher, module=None, verify=True,
 
     def decorator(implementation):
         if verify:
-            verify_matching_signatures(implementation, dispatcher)
+            if dispatcher is not None:
+                verify_matching_signatures(implementation, dispatcher)
+            else:
+                # Using __code__ directly similar to verify_matching_signature
+                co = implementation.__code__
+                last_arg = co.co_argcount + co.co_kwonlyargcount - 1
+                last_arg = co.co_varnames[last_arg]
+                if last_arg != "like" or co.co_kwonlyargcount == 0:
+                    raise RuntimeError(
+                        "__array_function__ expects `like=` to be the last "
+                        "argument and a keyword-only argument. "
+                        f"{implementation} does not seem to comply.")
 
         if docs_from_dispatcher:
             add_docstring(implementation, dispatcher.__doc__)
 
-        @functools.wraps(implementation)
-        def public_api(*args, **kwargs):
-            try:
-                relevant_args = dispatcher(*args, **kwargs)
-            except TypeError as exc:
-                # Try to clean up a signature related TypeError.  Such an
-                # error will be something like:
-                #     dispatcher.__name__() got an unexpected keyword argument
-                #
-                # So replace the dispatcher name in this case.  In principle
-                # TypeErrors may be raised from _within_ the dispatcher, so
-                # we check that the traceback contains a string that starts
-                # with the name.  (In principle we could also check the
-                # traceback length, as it would be deeper.)
-                msg = exc.args[0]
-                disp_name = dispatcher.__name__
-                if not isinstance(msg, str) or not msg.startswith(disp_name):
-                    raise
-
-                # Replace with the correct name and re-raise:
-                new_msg = msg.replace(disp_name, public_api.__name__)
-                raise TypeError(new_msg) from None
-
-            return implement_array_function(
-                implementation, public_api, relevant_args, args, kwargs)
-
-        public_api.__code__ = public_api.__code__.replace(
-                co_name=implementation.__name__,
-                co_filename='<__array_function__ internals>')
+        public_api = _ArrayFunctionDispatcher(dispatcher, implementation)
+        public_api = functools.wraps(implementation)(public_api)
+
         if module is not None:
             public_api.__module__ = module
 
-        public_api._implementation = implementation
-
         ARRAY_FUNCTIONS.add(public_api)
 
         return public_api
diff --git a/numpy/core/src/multiarray/arrayfunction_override.c b/numpy/core/src/multiarray/arrayfunction_override.c
index 2bb3fbe28..e27bb516e 100644
--- a/numpy/core/src/multiarray/arrayfunction_override.c
+++ b/numpy/core/src/multiarray/arrayfunction_override.c
@@ -1,11 +1,15 @@
 #define NPY_NO_DEPRECATED_API NPY_API_VERSION
 #define _MULTIARRAYMODULE
 
+#include <Python.h>
+#include "structmember.h"
+
 #include "npy_pycompat.h"
 #include "get_attr_string.h"
 #include "npy_import.h"
 #include "multiarraymodule.h"
 
+#include "arrayfunction_override.h"
 
 /* Return the ndarray.__array_function__ method. */
 static PyObject *
@@ -200,183 +204,67 @@ call_array_function(PyObject* argument, PyObject* method,
 }
 
 
-/**
- * Internal handler for the array-function dispatching. The helper returns
- * either the result, or NotImplemented (as a borrowed reference).
- *
- * @param public_api The public API symbol used for dispatching
- * @param relevant_args Arguments which may implement __array_function__
- * @param args Original arguments
- * @param kwargs Original keyword arguments
- *
- * @returns The result of the dispatched version, or a borrowed reference
- *          to NotImplemented to indicate the default implementation should
- *          be used.
+
+/*
+ * Helper to convert from vectorcall convention, since the protocol requires
+ * args and kwargs to be passed as tuple and dict explicitly.
+ * We always pass a dict, so always returns it.
  */
-static PyObject *
-array_implement_array_function_internal(
-    PyObject *public_api, PyObject *relevant_args,
-    PyObject *args, PyObject *kwargs)
+static int
+get_args_and_kwargs(
+        PyObject *const *fast_args, Py_ssize_t len_args, PyObject *kwnames,
+        PyObject **out_args, PyObject **out_kwargs)
 {
-    PyObject *implementing_args[NPY_MAXARGS];
-    PyObject *array_function_methods[NPY_MAXARGS];
-    PyObject *types = NULL;
-
-    PyObject *result = NULL;
-
-    static PyObject *errmsg_formatter = NULL;
+    len_args = PyVectorcall_NARGS(len_args);
+    PyObject *args = PyTuple_New(len_args);
+    PyObject *kwargs = NULL;
 
-    relevant_args = PySequence_Fast(
-        relevant_args,
-        "dispatcher for __array_function__ did not return an iterable");
-    if (relevant_args == NULL) {
-        return NULL;
+    if (args == NULL) {
+        return -1;
     }
-
-    /* Collect __array_function__ implementations */
-    int num_implementing_args = get_implementing_args_and_methods(
-        relevant_args, implementing_args, array_function_methods);
-    if (num_implementing_args == -1) {
-        goto cleanup;
+    for (Py_ssize_t i = 0; i < len_args; i++) {
+        Py_INCREF(fast_args[i]);
+        PyTuple_SET_ITEM(args, i, fast_args[i]);
     }
-
-    /*
-     * Handle the typical case of no overrides. This is merely an optimization
-     * if some arguments are ndarray objects, but is also necessary if no
-     * arguments implement __array_function__ at all (e.g., if they are all
-     * built-in types).
-     */
-    int any_overrides = 0;
-    for (int j = 0; j < num_implementing_args; j++) {
-        if (!is_default_array_function(array_function_methods[j])) {
-            any_overrides = 1;
-            break;
-        }
-    }
-    if (!any_overrides) {
-        /*
-         * When the default implementation should be called, return
-         * `Py_NotImplemented` to indicate this.
-         */
-        result = Py_NotImplemented;
-        goto cleanup;
-    }
-
-    /*
-     * Create a Python object for types.
-     * We use a tuple, because it's the fastest Python collection to create
-     * and has the bonus of being immutable.
-     */
-    types = PyTuple_New(num_implementing_args);
-    if (types == NULL) {
-        goto cleanup;
-    }
-    for (int j = 0; j < num_implementing_args; j++) {
-        PyObject *arg_type = (PyObject *)Py_TYPE(implementing_args[j]);
-        Py_INCREF(arg_type);
-        PyTuple_SET_ITEM(types, j, arg_type);
-    }
-
-    /* Call __array_function__ methods */
-    for (int j = 0; j < num_implementing_args; j++) {
-        PyObject *argument = implementing_args[j];
-        PyObject *method = array_function_methods[j];
-
-        /*
-         * We use `public_api` instead of `implementation` here so
-         * __array_function__ implementations can do equality/identity
-         * comparisons.
-         */
-        result = call_array_function(
-            argument, method, public_api, types, args, kwargs);
-
-        if (result == Py_NotImplemented) {
-            /* Try the next one */
-            Py_DECREF(result);
-            result = NULL;
+    kwargs = PyDict_New();
+    if (kwnames != NULL) {
+        if (kwargs == NULL) {
+            Py_DECREF(args);
+            return -1;
         }
-        else {
-            /* Either a good result, or an exception was raised. */
-            goto cleanup;
+        Py_ssize_t nkwargs = PyTuple_GET_SIZE(kwnames);
+        for (Py_ssize_t i = 0; i < nkwargs; i++) {
+            PyObject *key = PyTuple_GET_ITEM(kwnames, i);
+            PyObject *value = fast_args[i+len_args];
+            if (PyDict_SetItem(kwargs, key, value) < 0) {
+                Py_DECREF(args);
+                Py_DECREF(kwargs);
+                return -1;
+            }
         }
     }
+    *out_args = args;
+    *out_kwargs = kwargs;
+    return 0;
+}
+
 
+static void
+set_no_matching_types_error(PyObject *public_api, PyObject *types)
+{
+    static PyObject *errmsg_formatter = NULL;
     /* No acceptable override found, raise TypeError. */
     npy_cache_import("numpy.core._internal",
                      "array_function_errmsg_formatter",
                      &errmsg_formatter);
     if (errmsg_formatter != NULL) {
         PyObject *errmsg = PyObject_CallFunctionObjArgs(
-            errmsg_formatter, public_api, types, NULL);
+                errmsg_formatter, public_api, types, NULL);
         if (errmsg != NULL) {
             PyErr_SetObject(PyExc_TypeError, errmsg);
             Py_DECREF(errmsg);
         }
     }
-
-cleanup:
-    for (int j = 0; j < num_implementing_args; j++) {
-        Py_DECREF(implementing_args[j]);
-        Py_DECREF(array_function_methods[j]);
-    }
-    Py_XDECREF(types);
-    Py_DECREF(relevant_args);
-    return result;
-}
-
-
-/*
- * Implements the __array_function__ protocol for a Python function, as described in
- * in NEP-18. See numpy.core.overrides for a full docstring.
- */
-NPY_NO_EXPORT PyObject *
-array_implement_array_function(
-    PyObject *NPY_UNUSED(dummy), PyObject *positional_args)
-{
-    PyObject *res, *implementation, *public_api, *relevant_args, *args, *kwargs;
-
-    if (!PyArg_UnpackTuple(
-            positional_args, "implement_array_function", 5, 5,
-            &implementation, &public_api, &relevant_args, &args, &kwargs)) {
-        return NULL;
-    }
-
-    /*
-     * Remove `like=` kwarg, which is NumPy-exclusive and thus not present
-     * in downstream libraries. If `like=` is specified but doesn't
-     * implement `__array_function__`, raise a `TypeError`.
-     */
-    if (kwargs != NULL && PyDict_Contains(kwargs, npy_ma_str_like)) {
-        PyObject *like_arg = PyDict_GetItem(kwargs, npy_ma_str_like);
-        if (like_arg != NULL) {
-            PyObject *tmp_has_override = get_array_function(like_arg);
-            if (tmp_has_override == NULL) {
-                return PyErr_Format(PyExc_TypeError,
-                        "The `like` argument must be an array-like that "
-                        "implements the `__array_function__` protocol.");
-            }
-            Py_DECREF(tmp_has_override);
-            PyDict_DelItem(kwargs, npy_ma_str_like);
-
-            /*
-             * If `like=` kwarg was removed, `implementation` points to the NumPy
-             * public API, as `public_api` is in that case the wrapper dispatcher
-             * function. For example, in the `np.full` case, `implementation` is
-             * `np.full`, whereas `public_api` is `_full_with_like`. This is done
-             * to ensure `__array_function__` implementations can do
-             * equality/identity comparisons when `like=` is present.
-             */
-            public_api = implementation;
-        }
-    }
-
-    res = array_implement_array_function_internal(
-        public_api, relevant_args, args, kwargs);
-
-    if (res == Py_NotImplemented) {
-        return PyObject_Call(implementation, args, kwargs);
-    }
-    return res;
 }
 
 /*
@@ -392,64 +280,48 @@ array_implement_c_array_function_creation(
     PyObject *args, PyObject *kwargs,
     PyObject *const *fast_args, Py_ssize_t len_args, PyObject *kwnames)
 {
-    PyObject *relevant_args = NULL;
+    PyObject *dispatch_types = NULL;
     PyObject *numpy_module = NULL;
     PyObject *public_api = NULL;
     PyObject *result = NULL;
 
     /* If `like` doesn't implement `__array_function__`, raise a `TypeError` */
-    PyObject *tmp_has_override = get_array_function(like);
-    if (tmp_has_override == NULL) {
+    PyObject *method = get_array_function(like);
+    if (method == NULL) {
         return PyErr_Format(PyExc_TypeError,
                 "The `like` argument must be an array-like that "
                 "implements the `__array_function__` protocol.");
     }
-    Py_DECREF(tmp_has_override);
-
-    if (fast_args != NULL) {
+    if (is_default_array_function(method)) {
         /*
-         * Convert from vectorcall convention, since the protocol requires
-         * the normal convention.  We have to do this late to ensure the
-         * normal path where NotImplemented is returned is fast.
+         * Return a borrowed reference of Py_NotImplemented to defer back to
+         * the original function.
          */
+        Py_DECREF(method);
+        return Py_NotImplemented;
+    }
+
+    dispatch_types = PyTuple_Pack(1, Py_TYPE(like));
+    if (dispatch_types == NULL) {
+        goto finish;
+    }
+
+    /* We have to call __array_function__ properly, which needs some prep */
+    if (fast_args != NULL) {
         assert(args == NULL);
         assert(kwargs == NULL);
-        args = PyTuple_New(len_args);
-        if (args == NULL) {
-            return NULL;
-        }
-        for (Py_ssize_t i = 0; i < len_args; i++) {
-            Py_INCREF(fast_args[i]);
-            PyTuple_SET_ITEM(args, i, fast_args[i]);
-        }
-        if (kwnames != NULL) {
-            kwargs = PyDict_New();
-            if (kwargs == NULL) {
-                Py_DECREF(args);
-                return NULL;
-            }
-            Py_ssize_t nkwargs = PyTuple_GET_SIZE(kwnames);
-            for (Py_ssize_t i = 0; i < nkwargs; i++) {
-                PyObject *key = PyTuple_GET_ITEM(kwnames, i);
-                PyObject *value = fast_args[i+len_args];
-                if (PyDict_SetItem(kwargs, key, value) < 0) {
-                    Py_DECREF(args);
-                    Py_DECREF(kwargs);
-                    return NULL;
-                }
-            }
+        if (get_args_and_kwargs(
+                fast_args, len_args, kwnames, &args, &kwargs) < 0) {
+            goto finish;
         }
     }
 
-    relevant_args = PyTuple_Pack(1, like);
-    if (relevant_args == NULL) {
-        goto finish;
-    }
     /* The like argument must be present in the keyword arguments, remove it */
     if (PyDict_DelItem(kwargs, npy_ma_str_like) < 0) {
         goto finish;
     }
 
+    /* Fetch the actual symbol (the long way right now) */
     numpy_module = PyImport_Import(npy_ma_str_numpy);
     if (numpy_module == NULL) {
         goto finish;
@@ -466,16 +338,20 @@ array_implement_c_array_function_creation(
         goto finish;
     }
 
-    result = array_implement_array_function_internal(
-            public_api, relevant_args, args, kwargs);
+    result = call_array_function(like, method,
+            public_api, dispatch_types, args, kwargs);
 
-  finish:
-    if (kwnames != NULL) {
-        /* args and kwargs were converted from vectorcall convention */
-        Py_XDECREF(args);
-        Py_XDECREF(kwargs);
+    if (result == Py_NotImplemented) {
+        Py_DECREF(result);
+        result = NULL;
+        set_no_matching_types_error(public_api, dispatch_types);
     }
-    Py_XDECREF(relevant_args);
+
+  finish:
+    Py_DECREF(method);
+    Py_XDECREF(args);
+    Py_XDECREF(kwargs);
+    Py_XDECREF(dispatch_types);
     Py_XDECREF(public_api);
     return result;
 }
@@ -530,3 +406,275 @@ cleanup:
     Py_DECREF(relevant_args);
     return result;
 }
+
+
+typedef struct {
+    PyObject_HEAD
+    vectorcallfunc vectorcall;
+    PyObject *dict;
+    PyObject *relevant_arg_func;
+    PyObject *default_impl;
+} PyArray_ArrayFunctionDispatcherObject;
+
+
+static void
+dispatcher_dealloc(PyArray_ArrayFunctionDispatcherObject *self)
+{
+    Py_CLEAR(self->relevant_arg_func);
+    Py_CLEAR(self->default_impl);
+    Py_CLEAR(self->dict);
+    PyObject_FREE(self);
+}
+
+
+static PyObject *
+dispatcher_vectorcall(PyArray_ArrayFunctionDispatcherObject *self,
+        PyObject *const *args, Py_ssize_t len_args, PyObject *kwnames)
+{
+    PyObject *result = NULL;
+    PyObject *types = NULL;
+    PyObject *relevant_args = NULL;
+
+    PyObject *public_api;
+
+    /* __array_function__ passes args, kwargs.  These may be filled: */
+    PyObject *packed_args = NULL;
+    PyObject *packed_kwargs = NULL;
+
+    PyObject *implementing_args[NPY_MAXARGS];
+    PyObject *array_function_methods[NPY_MAXARGS];
+
+    int num_implementing_args;
+
+    if (self->relevant_arg_func != NULL) {
+        public_api = (PyObject *)self;
+
+        /* Typical path, need to call the relevant_arg_func and unpack them */
+        relevant_args = PyObject_Vectorcall(
+                self->relevant_arg_func, args, len_args, kwnames);
+        if (relevant_args == NULL) {
+            return NULL;
+        }
+        Py_SETREF(relevant_args, PySequence_Fast(relevant_args,
+                "dispatcher for __array_function__ did not return an iterable"));
+        if (relevant_args == NULL) {
+            return NULL;
+        }
+
+        num_implementing_args = get_implementing_args_and_methods(
+                relevant_args, implementing_args, array_function_methods);
+        if (num_implementing_args < 0) {
+            Py_DECREF(relevant_args);
+            return NULL;
+        }
+    }
+    else {
+        /* For like= dispatching from Python, the public_symbol is the impl */
+        public_api = self->default_impl;
+
+        /*
+         * We are dealing with `like=` from Python.  For simplicity, the
+         * Python code passes it on as the first argument.
+         */
+        if (PyVectorcall_NARGS(len_args) == 0) {
+            PyErr_Format(PyExc_TypeError,
+                    "`like` argument dispatching, but first argument is not "
+                    "positional in call to %S.", self->default_impl);
+            return NULL;
+        }
+
+        array_function_methods[0] = get_array_function(args[0]);
+        if (array_function_methods[0] == NULL) {
+            return PyErr_Format(PyExc_TypeError,
+                    "The `like` argument must be an array-like that "
+                    "implements the `__array_function__` protocol.");
+        }
+        num_implementing_args = 1;
+        implementing_args[0] = args[0];
+        Py_INCREF(implementing_args[0]);
+
+        /* do not pass the like argument */
+        len_args = PyVectorcall_NARGS(len_args) - 1;
+        len_args |= PY_VECTORCALL_ARGUMENTS_OFFSET;
+        args++;
+    }
+
+    /*
+     * Handle the typical case of no overrides. This is merely an optimization
+     * if some arguments are ndarray objects, but is also necessary if no
+     * arguments implement __array_function__ at all (e.g., if they are all
+     * built-in types).
+     */
+    int any_overrides = 0;
+    for (int j = 0; j < num_implementing_args; j++) {
+        if (!is_default_array_function(array_function_methods[j])) {
+            any_overrides = 1;
+            break;
+        }
+    }
+    if (!any_overrides) {
+        /* Directly call the actual implementation. */
+        result = PyObject_Vectorcall(self->default_impl, args, len_args, kwnames);
+        goto cleanup;
+    }
+
+    /* Find args and kwargs as tuple and dict, as we pass them out: */
+    if (get_args_and_kwargs(
+            args, len_args, kwnames, &packed_args, &packed_kwargs) < 0) {
+        goto cleanup;
+    }
+
+    /*
+     * Create a Python object for types.
+     * We use a tuple, because it's the fastest Python collection to create
+     * and has the bonus of being immutable.
+     */
+    types = PyTuple_New(num_implementing_args);
+    if (types == NULL) {
+        goto cleanup;
+    }
+    for (int j = 0; j < num_implementing_args; j++) {
+        PyObject *arg_type = (PyObject *)Py_TYPE(implementing_args[j]);
+        Py_INCREF(arg_type);
+        PyTuple_SET_ITEM(types, j, arg_type);
+    }
+
+    /* Call __array_function__ methods */
+    for (int j = 0; j < num_implementing_args; j++) {
+        PyObject *argument = implementing_args[j];
+        PyObject *method = array_function_methods[j];
+
+        result = call_array_function(
+                argument, method, public_api, types,
+                packed_args, packed_kwargs);
+
+        if (result == Py_NotImplemented) {
+            /* Try the next one */
+            Py_DECREF(result);
+            result = NULL;
+        }
+        else {
+            /* Either a good result, or an exception was raised. */
+            goto cleanup;
+        }
+    }
+
+    set_no_matching_types_error(public_api, types);
+
+cleanup:
+    for (int j = 0; j < num_implementing_args; j++) {
+        Py_DECREF(implementing_args[j]);
+        Py_DECREF(array_function_methods[j]);
+    }
+    Py_XDECREF(packed_args);
+    Py_XDECREF(packed_kwargs);
+    Py_XDECREF(types);
+    Py_XDECREF(relevant_args);
+    return result;
+}
+
+
+static PyObject *
+dispatcher_new(PyTypeObject *NPY_UNUSED(cls), PyObject *args, PyObject *kwargs)
+{
+    PyArray_ArrayFunctionDispatcherObject *self;
+
+    self = PyObject_New(
+            PyArray_ArrayFunctionDispatcherObject,
+            &PyArrayFunctionDispatcher_Type);
+    if (self == NULL) {
+        return PyErr_NoMemory();
+    }
+
+    char *kwlist[] = {"", "", NULL};
+    if (!PyArg_ParseTupleAndKeywords(
+            args, kwargs, "OO:_ArrayFunctionDispatcher", kwlist,
+            &self->relevant_arg_func, &self->default_impl)) {
+        Py_DECREF(self);
+        return NULL;
+    }
+
+    self->vectorcall = (vectorcallfunc)dispatcher_vectorcall;
+    if (self->relevant_arg_func == Py_None) {
+        /* NULL in the relevant arg function means we use `like=` */
+        Py_CLEAR(self->relevant_arg_func);
+    }
+    else {
+        Py_INCREF(self->relevant_arg_func);
+    }
+    Py_INCREF(self->default_impl);
+
+    /* Need to be like a Python function that has arbitrary attributes */
+    self->dict = PyDict_New();
+    if (self->dict == NULL) {
+        Py_DECREF(self);
+        return NULL;
+    }
+    return (PyObject *)self;
+}
+
+
+static PyObject *
+dispatcher_str(PyArray_ArrayFunctionDispatcherObject *self)
+{
+    return PyObject_Str(self->default_impl);
+}
+
+
+static PyObject *
+dispatcher_repr(PyObject *self)
+{
+    PyObject *name = PyObject_GetAttrString(self, "__name__");
+    if (name == NULL) {
+        return NULL;
+    }
+    /* Print like a normal function */
+    return PyUnicode_FromFormat("<function %S at %p>", name, self);
+}
+
+static PyObject *
+dispatcher_get_implementation(
+        PyArray_ArrayFunctionDispatcherObject *self, void *NPY_UNUSED(closure))
+{
+    Py_INCREF(self->default_impl);
+    return self->default_impl;
+}
+
+
+static PyObject *
+dispatcher_reduce(PyObject *self, PyObject *NPY_UNUSED(args))
+{
+    return PyObject_GetAttrString(self, "__qualname__");
+}
+
+
+static struct PyMethodDef func_dispatcher_methods[] = {
+    {"__reduce__",
+        (PyCFunction)dispatcher_reduce, METH_NOARGS, NULL},
+    {NULL, NULL, 0, NULL}
+};
+
+
+static struct PyGetSetDef func_dispatcher_getset[] = {
+    {"__dict__", &PyObject_GenericGetDict, 0, NULL, 0},
+    {"_implementation", (getter)&dispatcher_get_implementation, 0, NULL, 0},
+    {0, 0, 0, 0, 0}
+};
+
+
+NPY_NO_EXPORT PyTypeObject PyArrayFunctionDispatcher_Type = {
+     PyVarObject_HEAD_INIT(NULL, 0)
+     .tp_name = "numpy._ArrayFunctionDispatcher",
+     .tp_basicsize = sizeof(PyArray_ArrayFunctionDispatcherObject),
+     /* We have a dict, so in theory could traverse, but in practice... */
+     .tp_dictoffset = offsetof(PyArray_ArrayFunctionDispatcherObject, dict),
+     .tp_dealloc = (destructor)dispatcher_dealloc,
+     .tp_new = (newfunc)dispatcher_new,
+     .tp_str = (reprfunc)dispatcher_str,
+     .tp_repr = (reprfunc)dispatcher_repr,
+     .tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_HAVE_VECTORCALL,
+     .tp_methods = func_dispatcher_methods,
+     .tp_getset = func_dispatcher_getset,
+     .tp_call = &PyVectorcall_Call,
+     .tp_vectorcall_offset = offsetof(PyArray_ArrayFunctionDispatcherObject, vectorcall),
+};
diff --git a/numpy/core/src/multiarray/arrayfunction_override.h b/numpy/core/src/multiarray/arrayfunction_override.h
index 09f7ee548..3b8b88bac 100644
--- a/numpy/core/src/multiarray/arrayfunction_override.h
+++ b/numpy/core/src/multiarray/arrayfunction_override.h
@@ -1,9 +1,7 @@
 #ifndef NUMPY_CORE_SRC_MULTIARRAY_ARRAYFUNCTION_OVERRIDE_H_
 #define NUMPY_CORE_SRC_MULTIARRAY_ARRAYFUNCTION_OVERRIDE_H_
 
-NPY_NO_EXPORT PyObject *
-array_implement_array_function(
-    PyObject *NPY_UNUSED(dummy), PyObject *positional_args);
+extern NPY_NO_EXPORT PyTypeObject PyArrayFunctionDispatcher_Type;
 
 NPY_NO_EXPORT PyObject *
 array__get_implementing_args(
diff --git a/numpy/core/src/multiarray/multiarraymodule.c b/numpy/core/src/multiarray/multiarraymodule.c
index 94fa2a909..6b1862b18 100644
--- a/numpy/core/src/multiarray/multiarraymodule.c
+++ b/numpy/core/src/multiarray/multiarraymodule.c
@@ -4539,9 +4539,6 @@ static struct PyMethodDef array_module_methods[] = {
         METH_VARARGS | METH_KEYWORDS, NULL},
     {"_monotonicity", (PyCFunction)arr__monotonicity,
         METH_VARARGS | METH_KEYWORDS, NULL},
-    {"implement_array_function",
-        (PyCFunction)array_implement_array_function,
-        METH_VARARGS, NULL},
     {"interp", (PyCFunction)arr_interp,
         METH_VARARGS | METH_KEYWORDS, NULL},
     {"interp_complex", (PyCFunction)arr_interp_complex,
@@ -5112,6 +5109,12 @@ PyMODINIT_FUNC PyInit__multiarray_umath(void) {
     if (set_typeinfo(d) != 0) {
         goto err;
     }
+    if (PyType_Ready(&PyArrayFunctionDispatcher_Type) < 0) {
+        goto err;
+    }
+    PyDict_SetItemString(
+            d, "_ArrayFunctionDispatcher",
+            (PyObject *)&PyArrayFunctionDispatcher_Type);
     if (PyType_Ready(&PyArrayMethod_Type) < 0) {
         goto err;
     }
diff --git a/numpy/lib/npyio.py b/numpy/lib/npyio.py
index 71d600c30..0c1740df1 100644
--- a/numpy/lib/npyio.py
+++ b/numpy/lib/npyio.py
@@ -760,13 +760,6 @@ def _ensure_ndmin_ndarray(a, *, ndmin: int):
 _loadtxt_chunksize = 50000
 
 
-def _loadtxt_dispatcher(
-        fname, dtype=None, comments=None, delimiter=None,
-        converters=None, skiprows=None, usecols=None, unpack=None,
-        ndmin=None, encoding=None, max_rows=None, *, like=None):
-    return (like,)
-
-
 def _check_nonneg_int(value, name="argument"):
     try:
         operator.index(value)
@@ -1331,10 +1324,10 @@ def loadtxt(fname, dtype=float, comments='#', delimiter=None,
 
     if like is not None:
         return _loadtxt_with_like(
-            fname, dtype=dtype, comments=comments, delimiter=delimiter,
+            like, fname, dtype=dtype, comments=comments, delimiter=delimiter,
             converters=converters, skiprows=skiprows, usecols=usecols,
             unpack=unpack, ndmin=ndmin, encoding=encoding,
-            max_rows=max_rows, like=like
+            max_rows=max_rows
         )
 
     if isinstance(delimiter, bytes):
@@ -1361,9 +1354,7 @@ def loadtxt(fname, dtype=float, comments='#', delimiter=None,
     return arr
 
 
-_loadtxt_with_like = array_function_dispatch(
-    _loadtxt_dispatcher
-)(loadtxt)
+_loadtxt_with_like = array_function_dispatch()(loadtxt)
 
 
 def _savetxt_dispatcher(fname, X, fmt=None, delimiter=None, newline=None,
@@ -1724,17 +1715,6 @@ def fromregex(file, regexp, dtype, encoding=None):
 #####--------------------------------------------------------------------------
 
 
-def _genfromtxt_dispatcher(fname, dtype=None, comments=None, delimiter=None,
-                           skip_header=None, skip_footer=None, converters=None,
-                           missing_values=None, filling_values=None, usecols=None,
-                           names=None, excludelist=None, deletechars=None,
-                           replace_space=None, autostrip=None, case_sensitive=None,
-                           defaultfmt=None, unpack=None, usemask=None, loose=None,
-                           invalid_raise=None, max_rows=None, encoding=None,
-                           *, ndmin=None, like=None):
-    return (like,)
-
-
 @set_array_function_like_doc
 @set_module('numpy')
 def genfromtxt(fname, dtype=float, comments='#', delimiter=None,
@@ -1932,7 +1912,7 @@ def genfromtxt(fname, dtype=float, comments='#', delimiter=None,
 
     if like is not None:
         return _genfromtxt_with_like(
-            fname, dtype=dtype, comments=comments, delimiter=delimiter,
+            like, fname, dtype=dtype, comments=comments, delimiter=delimiter,
             skip_header=skip_header, skip_footer=skip_footer,
             converters=converters, missing_values=missing_values,
             filling_values=filling_values, usecols=usecols, names=names,
@@ -1942,7 +1922,6 @@ def genfromtxt(fname, dtype=float, comments='#', delimiter=None,
             unpack=unpack, usemask=usemask, loose=loose,
             invalid_raise=invalid_raise, max_rows=max_rows, encoding=encoding,
             ndmin=ndmin,
-            like=like
         )
 
     _ensure_ndmin_ndarray_check_param(ndmin)
@@ -2471,9 +2450,7 @@ def genfromtxt(fname, dtype=float, comments='#', delimiter=None,
     return output
 
 
-_genfromtxt_with_like = array_function_dispatch(
-    _genfromtxt_dispatcher
-)(genfromtxt)
+_genfromtxt_with_like = array_function_dispatch()(genfromtxt)
 
 
 def recfromtxt(fname, **kwargs):
diff --git a/numpy/lib/twodim_base.py b/numpy/lib/twodim_base.py
index dcb4ed46c..ed4f98704 100644
--- a/numpy/lib/twodim_base.py
+++ b/numpy/lib/twodim_base.py
@@ -155,10 +155,6 @@ def flipud(m):
     return m[::-1, ...]
 
 
-def _eye_dispatcher(N, M=None, k=None, dtype=None, order=None, *, like=None):
-    return (like,)
-
-
 @set_array_function_like_doc
 @set_module('numpy')
 def eye(N, M=None, k=0, dtype=float, order='C', *, like=None):
@@ -209,7 +205,7 @@ def eye(N, M=None, k=0, dtype=float, order='C', *, like=None):
 
     """
     if like is not None:
-        return _eye_with_like(N, M=M, k=k, dtype=dtype, order=order, like=like)
+        return _eye_with_like(like, N, M=M, k=k, dtype=dtype, order=order)
     if M is None:
         M = N
     m = zeros((N, M), dtype=dtype, order=order)
@@ -228,9 +224,7 @@ def eye(N, M=None, k=0, dtype=float, order='C', *, like=None):
     return m
 
 
-_eye_with_like = array_function_dispatch(
-    _eye_dispatcher
-)(eye)
+_eye_with_like = array_function_dispatch()(eye)
 
 
 def _diag_dispatcher(v, k=None):
@@ -369,10 +363,6 @@ def diagflat(v, k=0):
     return wrap(res)
 
 
-def _tri_dispatcher(N, M=None, k=None, dtype=None, *, like=None):
-    return (like,)
-
-
 @set_array_function_like_doc
 @set_module('numpy')
 def tri(N, M=None, k=0, dtype=float, *, like=None):
@@ -416,7 +406,7 @@ def tri(N, M=None, k=0, dtype=float, *, like=None):
 
     """
     if like is not None:
-        return _tri_with_like(N, M=M, k=k, dtype=dtype, like=like)
+        return _tri_with_like(like, N, M=M, k=k, dtype=dtype)
 
     if M is None:
         M = N
@@ -430,9 +420,7 @@ def tri(N, M=None, k=0, dtype=float, *, like=None):
     return m
 
 
-_tri_with_like = array_function_dispatch(
-    _tri_dispatcher
-)(tri)
+_tri_with_like = array_function_dispatch()(tri)
 
 
 def _trilu_dispatcher(m, k=None):