Merge branch 'master' of github.com:numpy/numpy into usimd-compiled

14 files changed, 171 insertions, 151 deletions

diff --git a/numpy/__init__.pyi b/numpy/__init__.pyi
index 83afd2e49..cf9b3e86a 100644
--- a/numpy/__init__.pyi
+++ b/numpy/__init__.pyi
@@ -870,7 +870,7 @@ class dtype(Generic[_DTypeScalar]):
     @property
     def alignment(self) -> int: ...
     @property
-    def base(self) -> dtype: ...
+    def base(self: _DType) -> _DType: ...
     @property
     def byteorder(self) -> str: ...
     @property
@@ -880,7 +880,7 @@ class dtype(Generic[_DTypeScalar]):
     @property
     def fields(
         self,
-    ) -> Optional[Mapping[str, Union[Tuple[dtype, int], Tuple[dtype, int, Any]]]]: ...
+    ) -> Optional[Mapping[str, Union[Tuple[dtype[Any], int], Tuple[dtype[Any], int, Any]]]]: ...
     @property
     def flags(self) -> int: ...
     @property
@@ -906,14 +906,14 @@ class dtype(Generic[_DTypeScalar]):
     @property
     def ndim(self) -> int: ...
     @property
-    def subdtype(self) -> Optional[Tuple[dtype, _Shape]]: ...
-    def newbyteorder(self, __new_order: _ByteOrder = ...) -> dtype: ...
+    def subdtype(self: _DType) -> Optional[Tuple[_DType, _Shape]]: ...
+    def newbyteorder(self: _DType, __new_order: _ByteOrder = ...) -> _DType: ...
     # Leave str and type for end to avoid having to use `builtins.str`
     # everywhere. See https://github.com/python/mypy/issues/3775
     @property
     def str(self) -> builtins.str: ...
     @property
-    def type(self) -> Type[generic]: ...
+    def type(self) -> Type[_DTypeScalar]: ...
 
 class _flagsobj:
     aligned: bool
@@ -955,24 +955,30 @@ _ArrayLikeInt = Union[
 
 _FlatIterSelf = TypeVar("_FlatIterSelf", bound=flatiter)
 
-class flatiter(Generic[_ArraySelf]):
+class flatiter(Generic[_NdArraySubClass]):
     @property
-    def base(self) -> _ArraySelf: ...
+    def base(self) -> _NdArraySubClass: ...
     @property
     def coords(self) -> _Shape: ...
     @property
     def index(self) -> int: ...
-    def copy(self) -> _ArraySelf: ...
+    def copy(self) -> _NdArraySubClass: ...
     def __iter__(self: _FlatIterSelf) -> _FlatIterSelf: ...
-    def __next__(self) -> generic: ...
+    def __next__(self: flatiter[ndarray[Any, dtype[_ScalarType]]]) -> _ScalarType: ...
     def __len__(self) -> int: ...
     @overload
-    def __getitem__(self, key: Union[int, integer]) -> generic: ...
+    def __getitem__(
+        self: flatiter[ndarray[Any, dtype[_ScalarType]]],
+        key: Union[int, integer],
+    ) -> _ScalarType: ...
     @overload
     def __getitem__(
         self, key: Union[_ArrayLikeInt, slice, ellipsis],
-    ) -> _ArraySelf: ...
-    def __array__(self, __dtype: DTypeLike = ...) -> ndarray: ...
+    ) -> _NdArraySubClass: ...
+    @overload
+    def __array__(self: flatiter[ndarray[Any, _DType]], __dtype: None = ...) -> ndarray[Any, _DType]: ...
+    @overload
+    def __array__(self, __dtype: DTypeLike) -> ndarray[Any, dtype[Any]]: ...
 
 _OrderKACF = Optional[Literal["K", "A", "C", "F"]]
 _OrderACF = Optional[Literal["A", "C", "F"]]
@@ -1004,7 +1010,6 @@ class _ArrayOrScalarCommon:
     def itemsize(self) -> int: ...
     @property
     def nbytes(self) -> int: ...
-    def __array__(self, __dtype: DTypeLike = ...) -> ndarray: ...
     def __bool__(self) -> bool: ...
     def __bytes__(self) -> bytes: ...
     def __str__(self) -> str: ...
@@ -1468,6 +1473,10 @@ class ndarray(_ArrayOrScalarCommon, Generic[_ShapeType, _DType]):
         strides: _ShapeLike = ...,
         order: _OrderKACF = ...,
     ) -> _ArraySelf: ...
+    @overload
+    def __array__(self, __dtype: None = ...) -> ndarray[Any, _DType]: ...
+    @overload
+    def __array__(self, __dtype: DTypeLike) -> ndarray[Any, dtype[Any]]: ...
     @property
     def ctypes(self) -> _ctypes: ...
     @property
@@ -1481,7 +1490,7 @@ class ndarray(_ArrayOrScalarCommon, Generic[_ShapeType, _DType]):
     def byteswap(self: _ArraySelf, inplace: bool = ...) -> _ArraySelf: ...
     def fill(self, value: Any) -> None: ...
     @property
-    def flat(self: _ArraySelf) -> flatiter[_ArraySelf]: ...
+    def flat(self: _NdArraySubClass) -> flatiter[_NdArraySubClass]: ...
     @overload
     def item(self, *args: int) -> Any: ...
     @overload
@@ -1646,6 +1655,10 @@ _NBit_co2 = TypeVar("_NBit_co2", covariant=True, bound=NBitBase)
 class generic(_ArrayOrScalarCommon):
     @abstractmethod
     def __init__(self, *args: Any, **kwargs: Any) -> None: ...
+    @overload
+    def __array__(self: _ScalarType, __dtype: None = ...) -> ndarray[Any, dtype[_ScalarType]]: ...
+    @overload
+    def __array__(self, __dtype: DTypeLike) -> ndarray[Any, dtype[Any]]: ...
     @property
     def base(self) -> None: ...
     @property
@@ -1658,7 +1671,7 @@ class generic(_ArrayOrScalarCommon):
     def strides(self) -> Tuple[()]: ...
     def byteswap(self: _ScalarType, inplace: Literal[False] = ...) -> _ScalarType: ...
     @property
-    def flat(self) -> flatiter[ndarray]: ...
+    def flat(self: _ScalarType) -> flatiter[ndarray[Any, dtype[_ScalarType]]]: ...
     def item(
         self: _ScalarType,
         __args: Union[Literal[0], Tuple[()], Tuple[Literal[0]]] = ...,
diff --git a/numpy/_version.py b/numpy/_version.py
index 73f9f5c7f..6605bf4af 100644
--- a/numpy/_version.py
+++ b/numpy/_version.py
@@ -236,7 +236,7 @@ def git_pieces_from_vcs(tag_prefix, root, verbose, run_command=run_command):
 
     # if there is a tag matching tag_prefix, this yields TAG-NUM-gHEX[-dirty]
     # if there isn't one, this yields HEX[-dirty] (no NUM)
-    describe_out, rc = run_command(GITS, ["describe", "--tags", "--dirty",
+    describe_out, rc = run_command(GITS, ["describe", "--tags", "--dirty=",
                                           "--always", "--long",
                                           "--match", "%s*" % tag_prefix],
                                    cwd=root)
diff --git a/numpy/core/function_base.py b/numpy/core/function_base.py
index 8a1fee99b..e940ac230 100644
--- a/numpy/core/function_base.py
+++ b/numpy/core/function_base.py
@@ -371,7 +371,7 @@ def geomspace(start, stop, num=50, endpoint=True, dtype=None, axis=0):
             6.12323400e-17+1.00000000e+00j,  7.07106781e-01+7.07106781e-01j,
             1.00000000e+00+0.00000000e+00j])
 
-    Graphical illustration of ``endpoint`` parameter:
+    Graphical illustration of `endpoint` parameter:
 
     >>> import matplotlib.pyplot as plt
     >>> N = 10
diff --git a/numpy/core/src/multiarray/einsum_sumprod.c.src b/numpy/core/src/multiarray/einsum_sumprod.c.src
index c58e74287..caba0e00a 100644
--- a/numpy/core/src/multiarray/einsum_sumprod.c.src
+++ b/numpy/core/src/multiarray/einsum_sumprod.c.src
@@ -17,7 +17,8 @@
 
 #include "einsum_sumprod.h"
 #include "einsum_debug.h"
-
+#include "simd/simd.h"
+#include "common.h"
 
 #ifdef NPY_HAVE_SSE_INTRINSICS
 #define EINSUM_USE_SSE1 1
@@ -41,6 +42,13 @@
 
 #define EINSUM_IS_SSE_ALIGNED(x) ((((npy_intp)x)&0xf) == 0)
 
+// ARM/Neon don't have instructions for aligned memory access
+#ifdef NPY_HAVE_NEON
+    #define EINSUM_IS_ALIGNED(x) 0
+#else
+    #define EINSUM_IS_ALIGNED(x) npy_is_aligned(x, NPY_SIMD_WIDTH)
+#endif
+
 /**********************************************/
 
 /**begin repeat
@@ -56,6 +64,10 @@
  *             npy_ubyte, npy_ushort, npy_uint, npy_ulong, npy_ulonglong,
  *             npy_float, npy_float, npy_double, npy_longdouble,
  *             npy_float, npy_double, npy_longdouble#
+ * #sfx  = s8, s16, s32, long, s64,
+ *        u8, u16, u32, ulong, u64,
+ *        half, f32, f64, longdouble,
+ *        f32, f64, clongdouble#
  * #to = ,,,,,
  *       ,,,,,
  *       npy_float_to_half,,,,
@@ -76,6 +88,10 @@
  *            0*5,
  *            0,0,1,0,
  *            0*3#
+ * #NPYV_CHK = 0*5,
+ *             0*5,
+ *             0, NPY_SIMD, NPY_SIMD_F64, 0,
+ *             0*3#
  */
 
 /**begin repeat1
@@ -250,115 +266,80 @@ static void
     @type@ *data0 = (@type@ *)dataptr[0];
     @type@ *data1 = (@type@ *)dataptr[1];
     @type@ *data_out = (@type@ *)dataptr[2];
-
-#if EINSUM_USE_SSE1 && @float32@
-    __m128 a, b;
-#elif EINSUM_USE_SSE2 && @float64@
-    __m128d a, b;
-#endif
-
     NPY_EINSUM_DBG_PRINT1("@name@_sum_of_products_contig_two (%d)\n",
                                                             (int)count);
-
-/* This is placed before the main loop to make small counts faster */
-finish_after_unrolled_loop:
-    switch (count) {
-/**begin repeat2
- * #i = 6, 5, 4, 3, 2, 1, 0#
- */
-        case @i@+1:
-            data_out[@i@] = @to@(@from@(data0[@i@]) *
-                                 @from@(data1[@i@]) +
-                                 @from@(data_out[@i@]));
-/**end repeat2**/
-        case 0:
-            return;
-    }
-
-#if EINSUM_USE_SSE1 && @float32@
+    // NPYV check for @type@
+#if @NPYV_CHK@
     /* Use aligned instructions if possible */
-    if (EINSUM_IS_SSE_ALIGNED(data0) && EINSUM_IS_SSE_ALIGNED(data1) &&
-        EINSUM_IS_SSE_ALIGNED(data_out)) {
-        /* Unroll the loop by 8 */
-        while (count >= 8) {
-            count -= 8;
-
-/**begin repeat2
- * #i = 0, 4#
- */
-            a = _mm_mul_ps(_mm_load_ps(data0+@i@), _mm_load_ps(data1+@i@));
-            b = _mm_add_ps(a, _mm_load_ps(data_out+@i@));
-            _mm_store_ps(data_out+@i@, b);
-/**end repeat2**/
-            data0 += 8;
-            data1 += 8;
-            data_out += 8;
+    const int is_aligned = EINSUM_IS_ALIGNED(data0) && EINSUM_IS_ALIGNED(data1) &&
+                        EINSUM_IS_ALIGNED(data_out);
+    const int vstep = npyv_nlanes_@sfx@;
+
+    /**begin repeat2
+     * #cond = if(is_aligned), else#
+     * #ld = loada, load#
+     * #st = storea, store#
+     */
+    @cond@ {
+        const npy_intp vstepx4 = vstep * 4;
+        for (; count >= vstepx4; count -= vstepx4, data0 += vstepx4, data1 += vstepx4, data_out += vstepx4) {
+            /**begin repeat3
+             * #i = 0, 1, 2, 3#
+             */
+            npyv_@sfx@ a@i@ = npyv_@ld@_@sfx@(data0 + vstep * @i@);
+            npyv_@sfx@ b@i@ = npyv_@ld@_@sfx@(data1 + vstep * @i@);
+            npyv_@sfx@ c@i@ = npyv_@ld@_@sfx@(data_out + vstep * @i@);
+            /**end repeat3**/
+            /**begin repeat3
+             * #i = 0, 1, 2, 3#
+             */
+            npyv_@sfx@ abc@i@ = npyv_muladd_@sfx@(a@i@, b@i@, c@i@);
+            /**end repeat3**/
+            /**begin repeat3
+             * #i = 0, 1, 2, 3#
+             */
+            npyv_@st@_@sfx@(data_out + vstep * @i@, abc@i@);
+            /**end repeat3**/
         }
-
-        /* Finish off the loop */
-        goto finish_after_unrolled_loop;
     }
-#elif EINSUM_USE_SSE2 && @float64@
-    /* Use aligned instructions if possible */
-    if (EINSUM_IS_SSE_ALIGNED(data0) && EINSUM_IS_SSE_ALIGNED(data1) &&
-        EINSUM_IS_SSE_ALIGNED(data_out)) {
-        /* Unroll the loop by 8 */
-        while (count >= 8) {
-            count -= 8;
-
-/**begin repeat2
- * #i = 0, 2, 4, 6#
- */
-            a = _mm_mul_pd(_mm_load_pd(data0+@i@), _mm_load_pd(data1+@i@));
-            b = _mm_add_pd(a, _mm_load_pd(data_out+@i@));
-            _mm_store_pd(data_out+@i@, b);
-/**end repeat2**/
-            data0 += 8;
-            data1 += 8;
-            data_out += 8;
-        }
-
-        /* Finish off the loop */
-        goto finish_after_unrolled_loop;
+    /**end repeat2**/
+    for (; count > 0; count -= vstep, data0 += vstep, data1 += vstep, data_out += vstep) {
+        npyv_@sfx@ a = npyv_load_tillz_@sfx@(data0, count);
+        npyv_@sfx@ b = npyv_load_tillz_@sfx@(data1, count);
+        npyv_@sfx@ c = npyv_load_tillz_@sfx@(data_out, count);
+        npyv_store_till_@sfx@(data_out, count, npyv_muladd_@sfx@(a, b, c));
     }
-#endif
-
-    /* Unroll the loop by 8 */
-    while (count >= 8) {
-        count -= 8;
-
-#if EINSUM_USE_SSE1 && @float32@
-/**begin repeat2
- * #i = 0, 4#
- */
-        a = _mm_mul_ps(_mm_loadu_ps(data0+@i@), _mm_loadu_ps(data1+@i@));
-        b = _mm_add_ps(a, _mm_loadu_ps(data_out+@i@));
-        _mm_storeu_ps(data_out+@i@, b);
-/**end repeat2**/
-#elif EINSUM_USE_SSE2 && @float64@
-/**begin repeat2
- * #i = 0, 2, 4, 6#
- */
-        a = _mm_mul_pd(_mm_loadu_pd(data0+@i@), _mm_loadu_pd(data1+@i@));
-        b = _mm_add_pd(a, _mm_loadu_pd(data_out+@i@));
-        _mm_storeu_pd(data_out+@i@, b);
-/**end repeat2**/
+    npyv_cleanup();
 #else
-/**begin repeat2
- * #i = 0, 1, 2, 3, 4, 5, 6, 7#
- */
-        data_out[@i@] = @to@(@from@(data0[@i@]) *
-                             @from@(data1[@i@]) +
-                             @from@(data_out[@i@]));
-/**end repeat2**/
-#endif
-        data0 += 8;
-        data1 += 8;
-        data_out += 8;
+#ifndef NPY_DISABLE_OPTIMIZATION
+    for (; count >= 4; count -= 4, data0 += 4, data1 += 4, data_out += 4) {
+        /**begin repeat2
+         * #i = 0, 1, 2, 3#
+         */
+        const @type@ a@i@ = @from@(data0[@i@]);
+        const @type@ b@i@ = @from@(data1[@i@]);
+        const @type@ c@i@ = @from@(data_out[@i@]);
+        /**end repeat2**/
+        /**begin repeat2
+         * #i = 0, 1, 2, 3#
+         */
+        const @type@ abc@i@ = a@i@ * b@i@ + c@i@;
+        /**end repeat2**/
+        /**begin repeat2
+         * #i = 0, 1, 2, 3#
+         */
+        data_out[@i@] = @to@(abc@i@);
+        /**end repeat2**/
     }
+#endif // !NPY_DISABLE_OPTIMIZATION
+    for (; count > 0; --count, ++data0, ++data1, ++data_out) {
+        const @type@ a = @from@(*data0);
+        const @type@ b = @from@(*data1);
+        const @type@ c = @from@(*data_out);
+        *data_out = @to@(a * b + c);
+    }
+#endif // NPYV check for @type@
 
-    /* Finish off the loop */
-    goto finish_after_unrolled_loop;
 }
 
 /* Some extra specializations for the two operand case */
diff --git a/numpy/distutils/fcompiler/gnu.py b/numpy/distutils/fcompiler/gnu.py
index 0d9d769c2..68d1501ee 100644
--- a/numpy/distutils/fcompiler/gnu.py
+++ b/numpy/distutils/fcompiler/gnu.py
@@ -126,7 +126,7 @@ class GnuFCompiler(FCompiler):
                     target = '10.9'
                     s = f'Env. variable MACOSX_DEPLOYMENT_TARGET set to {target}'
                     warnings.warn(s, stacklevel=2)
-                os.environ['MACOSX_DEPLOYMENT_TARGET'] = target
+                os.environ['MACOSX_DEPLOYMENT_TARGET'] = str(target)
             opt.extend(['-undefined', 'dynamic_lookup', '-bundle'])
         else:
             opt.append("-shared")
diff --git a/numpy/lib/shape_base.py b/numpy/lib/shape_base.py
index cbc4641d8..f0596444e 100644
--- a/numpy/lib/shape_base.py
+++ b/numpy/lib/shape_base.py
@@ -1088,8 +1088,8 @@ def kron(a, b):
     -----
     The function assumes that the number of dimensions of `a` and `b`
     are the same, if necessary prepending the smallest with ones.
-    If `a.shape = (r0,r1,..,rN)` and `b.shape = (s0,s1,...,sN)`,
-    the Kronecker product has shape `(r0*s0, r1*s1, ..., rN*SN)`.
+    If ``a.shape = (r0,r1,..,rN)`` and ``b.shape = (s0,s1,...,sN)``,
+    the Kronecker product has shape ``(r0*s0, r1*s1, ..., rN*SN)``.
     The elements are products of elements from `a` and `b`, organized
     explicitly by::
 
diff --git a/numpy/ma/core.py b/numpy/ma/core.py
index 4fb7d8c28..d6af22337 100644
--- a/numpy/ma/core.py
+++ b/numpy/ma/core.py
@@ -5439,7 +5439,7 @@ class MaskedArray(ndarray):
             When the array contains unmasked values at the same extremes of the
             datatype, the ordering of these values and the masked values is
             undefined.
-        fill_value : {var}, optional
+        fill_value : scalar or None, optional
             Value used internally for the masked values.
             If ``fill_value`` is not None, it supersedes ``endwith``.
 
@@ -5497,7 +5497,7 @@ class MaskedArray(ndarray):
         axis : {None, integer}
             If None, the index is into the flattened array, otherwise along
             the specified axis
-        fill_value : {var}, optional
+        fill_value : scalar or None, optional
             Value used to fill in the masked values.  If None, the output of
             minimum_fill_value(self._data) is used instead.
         out : {None, array}, optional
@@ -5543,7 +5543,7 @@ class MaskedArray(ndarray):
         axis : {None, integer}
             If None, the index is into the flattened array, otherwise along
             the specified axis
-        fill_value : {var}, optional
+        fill_value : scalar or None, optional
             Value used to fill in the masked values.  If None, the output of
             maximum_fill_value(self._data) is used instead.
         out : {None, array}, optional
@@ -5594,7 +5594,7 @@ class MaskedArray(ndarray):
             When the array contains unmasked values sorting at the same extremes of the
             datatype, the ordering of these values and the masked values is
             undefined.
-        fill_value : {var}, optional
+        fill_value : scalar or None, optional
             Value used internally for the masked values.
             If ``fill_value`` is not None, it supersedes ``endwith``.
 
@@ -5665,7 +5665,7 @@ class MaskedArray(ndarray):
         out : array_like, optional
             Alternative output array in which to place the result.  Must be of
             the same shape and buffer length as the expected output.
-        fill_value : {var}, optional
+        fill_value : scalar or None, optional
             Value used to fill in the masked values.
             If None, use the output of `minimum_fill_value`.
         keepdims : bool, optional
@@ -5799,7 +5799,7 @@ class MaskedArray(ndarray):
         out : array_like, optional
             Alternative output array in which to place the result.  Must
             be of the same shape and buffer length as the expected output.
-        fill_value : {var}, optional
+        fill_value : scalar or None, optional
             Value used to fill in the masked values.
             If None, use the output of maximum_fill_value().
         keepdims : bool, optional
@@ -5876,7 +5876,7 @@ class MaskedArray(ndarray):
             Alternative output array in which to place the result. It must
             have the same shape and buffer length as the expected output
             but the type will be cast if necessary.
-        fill_value : {var}, optional
+        fill_value : scalar or None, optional
             Value used to fill in the masked values.
         keepdims : bool, optional
             If this is set to True, the axes which are reduced are left
diff --git a/numpy/typing/_array_like.py b/numpy/typing/_array_like.py
index a1a604239..63b67b33a 100644
--- a/numpy/typing/_array_like.py
+++ b/numpy/typing/_array_like.py
@@ -1,7 +1,9 @@
+from __future__ import annotations
+
 import sys
-from typing import Any, overload, Sequence, TYPE_CHECKING, Union
+from typing import Any, overload, Sequence, TYPE_CHECKING, Union, TypeVar
 
-from numpy import ndarray
+from numpy import ndarray, dtype
 from ._scalars import _ScalarLike
 from ._dtype_like import DTypeLike
 
@@ -16,12 +18,15 @@ else:
     else:
         HAVE_PROTOCOL = True
 
+_DType = TypeVar("_DType", bound="dtype[Any]")
+
 if TYPE_CHECKING or HAVE_PROTOCOL:
-    class _SupportsArray(Protocol):
-        @overload
-        def __array__(self, __dtype: DTypeLike = ...) -> ndarray: ...
-        @overload
-        def __array__(self, dtype: DTypeLike = ...) -> ndarray: ...
+    # The `_SupportsArray` protocol only cares about the default dtype
+    # (i.e. `dtype=None`) of the to-be returned array.
+    # Concrete implementations of the protocol are responsible for adding
+    # any and all remaining overloads
+    class _SupportsArray(Protocol[_DType]):
+        def __array__(self, dtype: None = ...) -> ndarray[Any, _DType]: ...
 else:
     _SupportsArray = Any
 
@@ -36,5 +41,5 @@ ArrayLike = Union[
     _ScalarLike,
     Sequence[_ScalarLike],
     Sequence[Sequence[Any]],  # TODO: Wait for support for recursive types
-    _SupportsArray,
+    "_SupportsArray[Any]",
 ]
diff --git a/numpy/typing/tests/data/fail/array_like.py b/numpy/typing/tests/data/fail/array_like.py
index a97e72dc7..3bbd29061 100644
--- a/numpy/typing/tests/data/fail/array_like.py
+++ b/numpy/typing/tests/data/fail/array_like.py
@@ -11,6 +11,6 @@ x2: ArrayLike = A()  # E: Incompatible types in assignment
 x3: ArrayLike = {1: "foo", 2: "bar"}  # E: Incompatible types in assignment
 
 scalar = np.int64(1)
-scalar.__array__(dtype=np.float64)  # E: Unexpected keyword argument
+scalar.__array__(dtype=np.float64)  # E: No overload variant
 array = np.array([1])
-array.__array__(dtype=np.float64)  # E: Unexpected keyword argument
+array.__array__(dtype=np.float64)  # E: No overload variant
diff --git a/numpy/typing/tests/data/pass/array_like.py b/numpy/typing/tests/data/pass/array_like.py
index f85724267..563fc08c7 100644
--- a/numpy/typing/tests/data/pass/array_like.py
+++ b/numpy/typing/tests/data/pass/array_like.py
@@ -25,13 +25,13 @@ class A:
 x13: ArrayLike = A()
 
 scalar: _SupportsArray = np.int64(1)
-scalar.__array__(np.float64)
+scalar.__array__(None)
 array: _SupportsArray = np.array(1)
-array.__array__(np.float64)
+array.__array__(None)
 
 a: _SupportsArray = A()
-a.__array__(np.int64)
-a.__array__(dtype=np.int64)
+a.__array__(None)
+a.__array__(dtype=None)
 
 # Escape hatch for when you mean to make something like an object
 # array.
diff --git a/numpy/typing/tests/data/pass/dtype.py b/numpy/typing/tests/data/pass/dtype.py
index cbae8c078..a97edc302 100644
--- a/numpy/typing/tests/data/pass/dtype.py
+++ b/numpy/typing/tests/data/pass/dtype.py
@@ -1,5 +1,7 @@
 import numpy as np
 
+dtype_obj = np.dtype(np.str_)
+
 np.dtype(dtype=np.int64)
 np.dtype(int)
 np.dtype("int")
@@ -33,3 +35,9 @@ class Test:
 
 
 np.dtype(Test())
+
+# Methods and attributes
+dtype_obj.base
+dtype_obj.subdtype
+dtype_obj.newbyteorder()
+dtype_obj.type
diff --git a/numpy/typing/tests/data/pass/flatiter.py b/numpy/typing/tests/data/pass/flatiter.py
index c0219eb2b..4fdf25299 100644
--- a/numpy/typing/tests/data/pass/flatiter.py
+++ b/numpy/typing/tests/data/pass/flatiter.py
@@ -12,3 +12,5 @@ a[0]
 a[[0, 1, 2]]
 a[...]
 a[:]
+a.__array__()
+a.__array__(np.float64)
diff --git a/numpy/typing/tests/data/reveal/dtype.py b/numpy/typing/tests/data/reveal/dtype.py
index d414f2c49..626a15270 100644
--- a/numpy/typing/tests/data/reveal/dtype.py
+++ b/numpy/typing/tests/data/reveal/dtype.py
@@ -1,5 +1,7 @@
 import numpy as np
 
+dtype_obj: np.dtype[np.str_]
+
 reveal_type(np.dtype(np.float64))  # E: numpy.dtype[numpy.floating[numpy.typing._64Bit]]
 reveal_type(np.dtype(np.int64))  # E: numpy.dtype[numpy.signedinteger[numpy.typing._64Bit]]
 
@@ -31,3 +33,9 @@ reveal_type(np.dtype("S8"))  # E: numpy.dtype
 
 # Void
 reveal_type(np.dtype(("U", 10)))  # E: numpy.dtype[numpy.void]
+
+# Methods and attributes
+reveal_type(dtype_obj.base)  # E: numpy.dtype[numpy.str_]
+reveal_type(dtype_obj.subdtype)  # E: Union[Tuple[numpy.dtype[numpy.str_], builtins.tuple[builtins.int]], None]
+reveal_type(dtype_obj.newbyteorder())  # E: numpy.dtype[numpy.str_]
+reveal_type(dtype_obj.type)  # E: Type[numpy.str_]
diff --git a/numpy/typing/tests/data/reveal/flatiter.py b/numpy/typing/tests/data/reveal/flatiter.py
index 56cdc7a0e..221101ebb 100644
--- a/numpy/typing/tests/data/reveal/flatiter.py
+++ b/numpy/typing/tests/data/reveal/flatiter.py
@@ -1,14 +1,17 @@
+from typing import Any
 import numpy as np
 
-a: "np.flatiter[np.ndarray]"
+a: np.flatiter[np.ndarray[Any, np.dtype[np.str_]]]
 
-reveal_type(a.base)  # E: numpy.ndarray*
-reveal_type(a.copy())  # E: numpy.ndarray*
+reveal_type(a.base)  # E: numpy.ndarray[Any, numpy.dtype[numpy.str_]]
+reveal_type(a.copy())  # E: numpy.ndarray[Any, numpy.dtype[numpy.str_]]
 reveal_type(a.coords)  # E: tuple[builtins.int]
 reveal_type(a.index)  # E: int
-reveal_type(iter(a))  # E: Iterator[numpy.generic*]
-reveal_type(next(a))  # E: numpy.generic
-reveal_type(a[0])  # E: numpy.generic
-reveal_type(a[[0, 1, 2]])  # E: numpy.ndarray*
-reveal_type(a[...])  # E: numpy.ndarray*
-reveal_type(a[:])  # E: numpy.ndarray*
+reveal_type(iter(a))  # E: Iterator[numpy.str_]
+reveal_type(next(a))  # E: numpy.str_
+reveal_type(a[0])  # E: numpy.str_
+reveal_type(a[[0, 1, 2]])  # E: numpy.ndarray[Any, numpy.dtype[numpy.str_]]
+reveal_type(a[...])  # E: numpy.ndarray[Any, numpy.dtype[numpy.str_]]
+reveal_type(a[:])  # E: numpy.ndarray[Any, numpy.dtype[numpy.str_]]
+reveal_type(a.__array__())  # E: numpy.ndarray[Any, numpy.dtype[numpy.str_]]
+reveal_type(a.__array__(np.float64))  # E: numpy.ndarray[Any, numpy.dtype[Any]]