standalone: add test_ndarray_subscript_assignment

Currently this is used to interop with nalgebra.

nac3core/irrt/irrt.cpp

@@ -1,4 +1,15 @@
 #include <irrt/exception.hpp>
 #include <irrt/int_types.hpp>
+#include <irrt/list.hpp>
 #include <irrt/math_util.hpp>
+#include <irrt/ndarray/array.hpp>
+#include <irrt/ndarray/basic.hpp>
+#include <irrt/ndarray/broadcast.hpp>
+#include <irrt/ndarray/def.hpp>
+#include <irrt/ndarray/indexing.hpp>
+#include <irrt/ndarray/iter.hpp>
+#include <irrt/ndarray/matmul.hpp>
+#include <irrt/ndarray/reshape.hpp>
+#include <irrt/ndarray/transpose.hpp>
 #include <irrt/original.hpp>
+#include <irrt/slice.hpp>

nac3core/irrt/irrt/list.hpp

@@ -0,0 +1,19 @@
+#pragma once
+
+#include <irrt/int_types.hpp>
+#include <irrt/slice.hpp>
+
+namespace
+{
+/**
+ * @brief A list in NAC3.
+ *
+ * The `items` field is opaque. You must rely on external contexts to
+ * know how to interpret it.
+ */
+template <typename SizeT> struct List
+{
+    uint8_t *items;
+    SizeT len;
+};
+} // namespace

nac3core/irrt/irrt/ndarray/array.hpp

@@ -0,0 +1,130 @@
+#pragma once
+
+#include <irrt/debug.hpp>
+#include <irrt/exception.hpp>
+#include <irrt/int_types.hpp>
+#include <irrt/list.hpp>
+#include <irrt/ndarray/basic.hpp>
+#include <irrt/ndarray/def.hpp>
+
+namespace
+{
+namespace ndarray
+{
+namespace array
+{
+template <typename SizeT>
+void set_and_validate_list_shape_helper(SizeT axis, List<SizeT> *list, SizeT ndims, SizeT *shape)
+{
+    if (shape[axis] == -1)
+    {
+        // Dimension is unspecified. Set it.
+        shape[axis] = list->len;
+    }
+    else
+    {
+        // Dimension is specified. Check.
+        if (shape[axis] != list->len)
+        {
+            // Mismatch, throw an error.
+            // NOTE: NumPy's error message is more complex and needs more PARAMS to display.
+            raise_exception(SizeT, EXN_VALUE_ERROR,
+                            "The requested array has an inhomogenous shape "
+                            "after {0} dimension(s).",
+                            axis, shape[axis], list->len);
+        }
+    }
+
+    if (axis + 1 == ndims)
+    {
+        // `list` has type `list[ItemType]`
+        // Do nothing
+    }
+    else
+    {
+        // `list` has type `list[list[...]]`
+        List<SizeT> **lists = (List<SizeT> **)(list->items);
+        for (SizeT i = 0; i < list->len; i++)
+        {
+            set_and_validate_list_shape_helper<SizeT>(axis + 1, lists[i], ndims, shape);
+        }
+    }
+}
+
+// TODO: Document me
+template <typename SizeT> void set_and_validate_list_shape(List<SizeT> *list, SizeT ndims, SizeT *shape)
+{
+    for (SizeT axis = 0; axis < ndims; axis++)
+    {
+        shape[axis] = -1; // Sentinel to say this dimension is unspecified.
+    }
+    set_and_validate_list_shape_helper<SizeT>(0, list, ndims, shape);
+}
+
+template <typename SizeT>
+void write_list_to_array_helper(SizeT axis, SizeT *index, List<SizeT> *list, NDArray<SizeT> *ndarray)
+{
+    debug_assert_eq(SizeT, list->len, ndarray->shape[axis]);
+    if (IRRT_DEBUG_ASSERT_BOOL)
+    {
+        if (!ndarray::basic::is_c_contiguous(ndarray))
+        {
+            raise_debug_assert(SizeT, "ndarray is not C-contiguous", ndarray->strides[0], ndarray->strides[1],
+                               NO_PARAM);
+        }
+    }
+
+    if (axis + 1 == ndarray->ndims)
+    {
+        // `list` has type `list[ItemType]`
+        // `ndarray` is contiguous, so we can do this, and this is fast.
+        uint8_t *dst = ndarray->data + (ndarray->itemsize * (*index));
+        __builtin_memcpy(dst, list->items, ndarray->itemsize * list->len);
+        *index += list->len;
+    }
+    else
+    {
+        // `list` has type `list[list[...]]`
+        List<SizeT> **lists = (List<SizeT> **)(list->items);
+
+        for (SizeT i = 0; i < list->len; i++)
+        {
+            write_list_to_array_helper<SizeT>(axis + 1, index, lists[i], ndarray);
+        }
+    }
+}
+
+// TODO: Document me
+template <typename SizeT> void write_list_to_array(List<SizeT> *list, NDArray<SizeT> *ndarray)
+{
+    SizeT index = 0;
+    write_list_to_array_helper<SizeT>((SizeT)0, &index, list, ndarray);
+}
+} // namespace array
+} // namespace ndarray
+} // namespace
+
+extern "C"
+{
+    using namespace ndarray::array;
+
+    void __nac3_ndarray_array_set_and_validate_list_shape(List<int32_t> *list, int32_t ndims, int32_t *shape)
+    {
+        set_and_validate_list_shape(list, ndims, shape);
+    }
+
+    void __nac3_ndarray_array_set_and_validate_list_shape64(List<int64_t> *list, int64_t ndims, int64_t *shape)
+    {
+        set_and_validate_list_shape(list, ndims, shape);
+    }
+
+    void __nac3_ndarray_array_write_list_to_array(List<int32_t> *list, NDArray<int32_t> *ndarray)
+    {
+        write_list_to_array(list, ndarray);
+    }
+
+    void __nac3_ndarray_array_write_list_to_array64(List<int64_t> *list, NDArray<int64_t> *ndarray)
+    {
+        write_list_to_array(list, ndarray);
+    }
+}

nac3core/irrt/irrt/ndarray/basic.hpp

@@ -0,0 +1,380 @@
+#pragma once
+
+#include <irrt/debug.hpp>
+#include <irrt/exception.hpp>
+#include <irrt/int_types.hpp>
+#include <irrt/ndarray/def.hpp>
+
+namespace
+{
+namespace ndarray
+{
+namespace basic
+{
+/**
+ * @brief Asserts that `shape` does not contain negative dimensions.
+ *
+ * @param ndims Number of dimensions in `shape`
+ * @param shape The shape to check on
+ */
+template <typename SizeT> void assert_shape_no_negative(SizeT ndims, const SizeT *shape)
+{
+    for (SizeT axis = 0; axis < ndims; axis++)
+    {
+        if (shape[axis] < 0)
+        {
+            raise_exception(SizeT, EXN_VALUE_ERROR,
+                            "negative dimensions are not allowed; axis {0} "
+                            "has dimension {1}",
+                            axis, shape[axis], NO_PARAM);
+        }
+    }
+}
+
+/**
+ * @brief Check two shapes are the same in the context of writing outputting to an ndarray.
+ *
+ * This function throws error messages for output shape mismatches.
+ */
+template <typename SizeT>
+void assert_output_shape_same(SizeT ndarray_ndims, const SizeT *ndarray_shape, SizeT output_ndims,
+                              const SizeT *output_shape)
+{
+    if (ndarray_ndims != output_ndims)
+    {
+        // There is no corresponding NumPy error message like this.
+        raise_exception(SizeT, EXN_VALUE_ERROR, "Cannot write output of ndims {0} to an ndarray with ndims {1}",
+                        output_ndims, ndarray_ndims, NO_PARAM);
+    }
+
+    for (SizeT axis = 0; axis < ndarray_ndims; axis++)
+    {
+        if (ndarray_shape[axis] != output_shape[axis])
+        {
+            // There is no corresponding NumPy error message like this.
+            raise_exception(SizeT, EXN_VALUE_ERROR,
+                            "Mismatched dimensions on axis {0}, output has "
+                            "dimension {1}, but destination ndarray has dimension {2}.",
+                            axis, output_shape[axis], ndarray_shape[axis]);
+        }
+    }
+}
+
+/**
+ * @brief Returns the number of elements of an ndarray given its shape.
+ *
+ * @param ndims Number of dimensions in `shape`
+ * @param shape The shape of the ndarray
+ */
+template <typename SizeT> SizeT calc_size_from_shape(SizeT ndims, const SizeT *shape)
+{
+    SizeT size = 1;
+    for (SizeT axis = 0; axis < ndims; axis++)
+        size *= shape[axis];
+    return size;
+}
+
+/**
+ * @brief Compute the array indices of the `nth` (0-based) element of an ndarray given only its shape.
+ *
+ * @param ndims Number of elements in `shape` and `indices`
+ * @param shape The shape of the ndarray
+ * @param indices The returned indices indexing the ndarray with shape `shape`.
+ * @param nth The index of the element of interest.
+ */
+template <typename SizeT> void set_indices_by_nth(SizeT ndims, const SizeT *shape, SizeT *indices, SizeT nth)
+{
+    for (SizeT i = 0; i < ndims; i++)
+    {
+        SizeT axis = ndims - i - 1;
+        SizeT dim = shape[axis];
+
+        indices[axis] = nth % dim;
+        nth /= dim;
+    }
+}
+
+/**
+ * @brief Return the number of elements of an `ndarray`
+ *
+ * This function corresponds to `<an_ndarray>.size`
+ */
+template <typename SizeT> SizeT size(const NDArray<SizeT> *ndarray)
+{
+    return calc_size_from_shape(ndarray->ndims, ndarray->shape);
+}
+
+/**
+ * @brief Return of the number of its content of an `ndarray`.
+ *
+ * This function corresponds to `<an_ndarray>.nbytes`.
+ */
+template <typename SizeT> SizeT nbytes(const NDArray<SizeT> *ndarray)
+{
+    return size(ndarray) * ndarray->itemsize;
+}
+
+/**
+ * @brief Get the `len()` of an ndarray, and asserts that `ndarray` is a sized object.
+ *
+ * This function corresponds to `<an_ndarray>.__len__`.
+ *
+ * @param dst_length The returned result
+ */
+template <typename SizeT> SizeT len(const NDArray<SizeT> *ndarray)
+{
+    // numpy prohibits `__len__` on unsized objects
+    if (ndarray->ndims == 0)
+    {
+        raise_exception(SizeT, EXN_TYPE_ERROR, "len() of unsized object", NO_PARAM, NO_PARAM, NO_PARAM);
+    }
+    else
+    {
+        return ndarray->shape[0];
+    }
+}
+
+/**
+ * @brief Return a boolean indicating if `ndarray` is (C-)contiguous.
+ *
+ * You may want to see: ndarray's rules for C-contiguity: https://github.com/numpy/numpy/blob/df256d0d2f3bc6833699529824781c58f9c6e697/numpy/core/src/multiarray/flagsobject.c#L95C1-L99C45
+ */
+template <typename SizeT> bool is_c_contiguous(const NDArray<SizeT> *ndarray)
+{
+    // Other references:
+    // - tinynumpy's implementation: https://github.com/wadetb/tinynumpy/blob/0d23d22e07062ffab2afa287374c7b366eebdda1/tinynumpy/tinynumpy.py#L102
+    // - ndarray's flags["C_CONTIGUOUS"]: https://numpy.org/doc/stable/reference/generated/numpy.ndarray.flags.html#numpy.ndarray.flags
+    // - ndarray's rules for C-contiguity: https://github.com/numpy/numpy/blob/df256d0d2f3bc6833699529824781c58f9c6e697/numpy/core/src/multiarray/flagsobject.c#L95C1-L99C45
+
+    // From https://github.com/numpy/numpy/blob/df256d0d2f3bc6833699529824781c58f9c6e697/numpy/core/src/multiarray/flagsobject.c#L95C1-L99C45:
+    //
+    // The traditional rule is that for an array to be flagged as C contiguous,
+    // the following must hold:
+    //
+    // strides[-1] == itemsize
+    // strides[i] == shape[i+1] * strides[i + 1]
+    // [...]
+    // According to these rules, a 0- or 1-dimensional array is either both
+    // C- and F-contiguous, or neither; and an array with 2+ dimensions
+    // can be C- or F- contiguous, or neither, but not both. Though there
+    // there are exceptions for arrays with zero or one item, in the first
+    // case the check is relaxed up to and including the first dimension
+    // with shape[i] == 0. In the second case `strides == itemsize` will
+    // can be true for all dimensions and both flags are set.
+
+    if (ndarray->ndims == 0)
+    {
+        return true;
+    }
+
+    if (ndarray->strides[ndarray->ndims - 1] != ndarray->itemsize)
+    {
+        return false;
+    }
+
+    for (SizeT i = 1; i < ndarray->ndims; i++)
+    {
+        SizeT axis_i = ndarray->ndims - i - 1;
+        if (ndarray->strides[axis_i] != ndarray->shape[axis_i + 1] * ndarray->strides[axis_i + 1])
+        {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+/**
+ * @brief Return the pointer to the element indexed by `indices`.
+ */
+template <typename SizeT> uint8_t *get_pelement_by_indices(const NDArray<SizeT> *ndarray, const SizeT *indices)
+{
+    uint8_t *element = ndarray->data;
+    for (SizeT dim_i = 0; dim_i < ndarray->ndims; dim_i++)
+        element += indices[dim_i] * ndarray->strides[dim_i];
+    return element;
+}
+
+/**
+ * @brief Convenience function. Like `get_pelement_by_indices` but
+ * reinterprets the element pointer.
+ */
+template <typename SizeT, typename T> T *get_ptr(const NDArray<SizeT> *ndarray, const SizeT *indices)
+{
+    return (T *)get_pelement_by_indices(ndarray, indices);
+}
+
+/**
+ * @brief Return the pointer to the nth (0-based) element in a flattened view of `ndarray`.
+ *
+ * This function does no bound check.
+ */
+template <typename SizeT> uint8_t *get_nth_pelement(const NDArray<SizeT> *ndarray, SizeT nth)
+{
+    uint8_t *element = ndarray->data;
+    for (SizeT i = 0; i < ndarray->ndims; i++)
+    {
+        SizeT axis = ndarray->ndims - i - 1;
+        SizeT dim = ndarray->shape[axis];
+        element += ndarray->strides[axis] * (nth % dim);
+        nth /= dim;
+    }
+    return element;
+}
+
+/**
+ * @brief Update the strides of an ndarray given an ndarray `shape`
+ * and assuming that the ndarray is fully c-contagious.
+ *
+ * You might want to read https://ajcr.net/stride-guide-part-1/.
+ */
+template <typename SizeT> void set_strides_by_shape(NDArray<SizeT> *ndarray)
+{
+    SizeT stride_product = 1;
+    for (SizeT i = 0; i < ndarray->ndims; i++)
+    {
+        SizeT axis = ndarray->ndims - i - 1;
+        ndarray->strides[axis] = stride_product * ndarray->itemsize;
+        stride_product *= ndarray->shape[axis];
+    }
+}
+
+/**
+ * @brief Set an element in `ndarray`.
+ *
+ * @param pelement Pointer to the element in `ndarray` to be set.
+ * @param pvalue Pointer to the value `pelement` will be set to.
+ */
+template <typename SizeT> void set_pelement_value(NDArray<SizeT> *ndarray, uint8_t *pelement, const uint8_t *pvalue)
+{
+    __builtin_memcpy(pelement, pvalue, ndarray->itemsize);
+}
+
+/**
+ * @brief Copy data from one ndarray to another of the exact same size and itemsize.
+ *
+ * Both ndarrays will be viewed in their flatten views when copying the elements.
+ */
+template <typename SizeT> void copy_data(const NDArray<SizeT> *src_ndarray, NDArray<SizeT> *dst_ndarray)
+{
+    // TODO: Make this faster with memcpy
+
+    debug_assert_eq(SizeT, src_ndarray->itemsize, dst_ndarray->itemsize);
+
+    for (SizeT i = 0; i < size(src_ndarray); i++)
+    {
+        auto src_element = ndarray::basic::get_nth_pelement(src_ndarray, i);
+        auto dst_element = ndarray::basic::get_nth_pelement(dst_ndarray, i);
+        ndarray::basic::set_pelement_value(dst_ndarray, dst_element, src_element);
+    }
+}
+} // namespace basic
+} // namespace ndarray
+} // namespace
+
+extern "C"
+{
+    using namespace ndarray::basic;
+
+    void __nac3_ndarray_util_assert_shape_no_negative(int32_t ndims, int32_t *shape)
+    {
+        assert_shape_no_negative(ndims, shape);
+    }
+
+    void __nac3_ndarray_util_assert_shape_no_negative64(int64_t ndims, int64_t *shape)
+    {
+        assert_shape_no_negative(ndims, shape);
+    }
+
+    void __nac3_ndarray_util_assert_output_shape_same(int32_t ndarray_ndims, const int32_t *ndarray_shape,
+                                                      int32_t output_ndims, const int32_t *output_shape)
+    {
+        assert_output_shape_same(ndarray_ndims, ndarray_shape, output_ndims, output_shape);
+    }
+
+    void __nac3_ndarray_util_assert_output_shape_same64(int64_t ndarray_ndims, const int64_t *ndarray_shape,
+                                                        int64_t output_ndims, const int64_t *output_shape)
+    {
+        assert_output_shape_same(ndarray_ndims, ndarray_shape, output_ndims, output_shape);
+    }
+
+    uint32_t __nac3_ndarray_size(NDArray<int32_t> *ndarray)
+    {
+        return size(ndarray);
+    }
+
+    uint64_t __nac3_ndarray_size64(NDArray<int64_t> *ndarray)
+    {
+        return size(ndarray);
+    }
+
+    uint32_t __nac3_ndarray_nbytes(NDArray<int32_t> *ndarray)
+    {
+        return nbytes(ndarray);
+    }
+
+    uint64_t __nac3_ndarray_nbytes64(NDArray<int64_t> *ndarray)
+    {
+        return nbytes(ndarray);
+    }
+
+    int32_t __nac3_ndarray_len(NDArray<int32_t> *ndarray)
+    {
+        return len(ndarray);
+    }
+
+    int64_t __nac3_ndarray_len64(NDArray<int64_t> *ndarray)
+    {
+        return len(ndarray);
+    }
+
+    bool __nac3_ndarray_is_c_contiguous(NDArray<int32_t> *ndarray)
+    {
+        return is_c_contiguous(ndarray);
+    }
+
+    bool __nac3_ndarray_is_c_contiguous64(NDArray<int64_t> *ndarray)
+    {
+        return is_c_contiguous(ndarray);
+    }
+
+    uint8_t *__nac3_ndarray_get_nth_pelement(const NDArray<int32_t> *ndarray, int32_t nth)
+    {
+        return get_nth_pelement(ndarray, nth);
+    }
+
+    uint8_t *__nac3_ndarray_get_nth_pelement64(const NDArray<int64_t> *ndarray, int64_t nth)
+    {
+        return get_nth_pelement(ndarray, nth);
+    }
+
+    uint8_t *__nac3_ndarray_get_pelement_by_indices(const NDArray<int32_t> *ndarray, int32_t *indices)
+    {
+        return get_pelement_by_indices(ndarray, indices);
+    }
+
+    uint8_t *__nac3_ndarray_get_pelement_by_indices64(const NDArray<int64_t> *ndarray, int64_t *indices)
+    {
+        return get_pelement_by_indices(ndarray, indices);
+    }
+
+    void __nac3_ndarray_set_strides_by_shape(NDArray<int32_t> *ndarray)
+    {
+        set_strides_by_shape(ndarray);
+    }
+
+    void __nac3_ndarray_set_strides_by_shape64(NDArray<int64_t> *ndarray)
+    {
+        set_strides_by_shape(ndarray);
+    }
+
+    void __nac3_ndarray_copy_data(NDArray<int32_t> *src_ndarray, NDArray<int32_t> *dst_ndarray)
+    {
+        copy_data(src_ndarray, dst_ndarray);
+    }
+
+    void __nac3_ndarray_copy_data64(NDArray<int64_t> *src_ndarray, NDArray<int64_t> *dst_ndarray)
+    {
+        copy_data(src_ndarray, dst_ndarray);
+    }
+}

nac3core/irrt/irrt/ndarray/broadcast.hpp

@@ -0,0 +1,188 @@
+#pragma once
+
+#include <irrt/int_types.hpp>
+#include <irrt/ndarray/def.hpp>
+#include <irrt/slice.hpp>
+
+namespace
+{
+template <typename SizeT> struct ShapeEntry
+{
+    SizeT ndims;
+    SizeT *shape;
+};
+} // namespace
+
+namespace
+{
+namespace ndarray
+{
+namespace broadcast
+{
+/**
+ * @brief Return true if `src_shape` can broadcast to `dst_shape`.
+ *
+ * See https://numpy.org/doc/stable/user/basics.broadcasting.html
+ */
+template <typename SizeT>
+bool can_broadcast_shape_to(SizeT target_ndims, const SizeT *target_shape, SizeT src_ndims, const SizeT *src_shape)
+{
+    if (src_ndims > target_ndims)
+    {
+        return false;
+    }
+
+    for (SizeT i = 0; i < src_ndims; i++)
+    {
+        SizeT target_dim = target_shape[target_ndims - i - 1];
+        SizeT src_dim = src_shape[src_ndims - i - 1];
+        if (!(src_dim == 1 || target_dim == src_dim))
+        {
+            return false;
+        }
+    }
+    return true;
+}
+
+/**
+ * @brief Performs `np.broadcast_shapes(<shapes>)`
+ *
+ * @param num_shapes Number of entries in `shapes`
+ * @param shapes The list of shape to do `np.broadcast_shapes` on.
+ * @param dst_ndims The length of `dst_shape`.
+ *   `dst_ndims` must be `max([shape.ndims for shape in shapes])`, but the caller has to calculate it/provide it.
+ *   for this function since they should already know in order to allocate `dst_shape` in the first place.
+ * @param dst_shape The resulting shape. Must be pre-allocated by the caller. This function calculate the result
+ *   of `np.broadcast_shapes` and write it here.
+ */
+template <typename SizeT>
+void broadcast_shapes(SizeT num_shapes, const ShapeEntry<SizeT> *shapes, SizeT dst_ndims, SizeT *dst_shape)
+{
+    for (SizeT dst_axis = 0; dst_axis < dst_ndims; dst_axis++)
+    {
+        dst_shape[dst_axis] = 1;
+    }
+
+#ifdef IRRT_DEBUG_ASSERT
+    SizeT max_ndims_found = 0;
+#endif
+
+    for (SizeT i = 0; i < num_shapes; i++)
+    {
+        ShapeEntry<SizeT> entry = shapes[i];
+
+        // Check pre-condition: `dst_ndims` must be `max([shape.ndims for shape in shapes])`
+        debug_assert(SizeT, entry.ndims <= dst_ndims);
+
+#ifdef IRRT_DEBUG_ASSERT
+        max_ndims_found = max(max_ndims_found, entry.ndims);
+#endif
+
+        for (SizeT j = 0; j < entry.ndims; j++)
+        {
+            SizeT entry_axis = entry.ndims - j - 1;
+            SizeT dst_axis = dst_ndims - j - 1;
+
+            SizeT entry_dim = entry.shape[entry_axis];
+            SizeT dst_dim = dst_shape[dst_axis];
+
+            if (dst_dim == 1)
+            {
+                dst_shape[dst_axis] = entry_dim;
+            }
+            else if (entry_dim == 1 || entry_dim == dst_dim)
+            {
+                // Do nothing
+            }
+            else
+            {
+                raise_exception(SizeT, EXN_VALUE_ERROR,
+                                "shape mismatch: objects cannot be broadcast "
+                                "to a single shape.",
+                                NO_PARAM, NO_PARAM, NO_PARAM);
+            }
+        }
+    }
+
+    // Check pre-condition: `dst_ndims` must be `max([shape.ndims for shape in shapes])`
+    debug_assert_eq(SizeT, max_ndims_found, dst_ndims);
+}
+
+/**
+ * @brief Perform `np.broadcast_to(<ndarray>, <target_shape>)` and appropriate assertions.
+ *
+ * This function attempts to broadcast `src_ndarray` to a new shape defined by `dst_ndarray.shape`,
+ * and return the result by modifying `dst_ndarray`.
+ *
+ * # Notes on `dst_ndarray`
+ * The caller is responsible for allocating space for the resulting ndarray.
+ * Here is what this function expects from `dst_ndarray` when called:
+ *   - `dst_ndarray->data` does not have to be initialized.
+ *   - `dst_ndarray->itemsize` does not have to be initialized.
+ *   - `dst_ndarray->ndims` must be initialized, determining the length of `dst_ndarray->shape`
+ *   - `dst_ndarray->shape` must be allocated, and must contain the desired target broadcast shape.
+ *   - `dst_ndarray->strides` must be allocated, through it can contain uninitialized values.
+ * When this function call ends:
+ *   - `dst_ndarray->data` is set to `src_ndarray->data` (`dst_ndarray` is just a view to `src_ndarray`)
+ *   - `dst_ndarray->itemsize` is set to `src_ndarray->itemsize`
+ *   - `dst_ndarray->ndims` is unchanged.
+ *   - `dst_ndarray->shape` is unchanged.
+ *   - `dst_ndarray->strides` is updated accordingly by how ndarray broadcast_to works.
+ */
+template <typename SizeT> void broadcast_to(const NDArray<SizeT> *src_ndarray, NDArray<SizeT> *dst_ndarray)
+{
+    if (!ndarray::broadcast::can_broadcast_shape_to(dst_ndarray->ndims, dst_ndarray->shape, src_ndarray->ndims,
+                                                    src_ndarray->shape))
+    {
+        raise_exception(SizeT, EXN_VALUE_ERROR, "operands could not be broadcast together", NO_PARAM, NO_PARAM,
+                        NO_PARAM);
+    }
+
+    dst_ndarray->data = src_ndarray->data;
+    dst_ndarray->itemsize = src_ndarray->itemsize;
+
+    for (SizeT i = 0; i < dst_ndarray->ndims; i++)
+    {
+        SizeT src_axis = src_ndarray->ndims - i - 1;
+        SizeT dst_axis = dst_ndarray->ndims - i - 1;
+        if (src_axis < 0 || (src_ndarray->shape[src_axis] == 1 && dst_ndarray->shape[dst_axis] != 1))
+        {
+            // Freeze the steps in-place
+            dst_ndarray->strides[dst_axis] = 0;
+        }
+        else
+        {
+            dst_ndarray->strides[dst_axis] = src_ndarray->strides[src_axis];
+        }
+    }
+}
+} // namespace broadcast
+} // namespace ndarray
+} // namespace
+
+extern "C"
+{
+    using namespace ndarray::broadcast;
+
+    void __nac3_ndarray_broadcast_to(NDArray<int32_t> *src_ndarray, NDArray<int32_t> *dst_ndarray)
+    {
+        broadcast_to(src_ndarray, dst_ndarray);
+    }
+
+    void __nac3_ndarray_broadcast_to64(NDArray<int64_t> *src_ndarray, NDArray<int64_t> *dst_ndarray)
+    {
+        broadcast_to(src_ndarray, dst_ndarray);
+    }
+
+    void __nac3_ndarray_broadcast_shapes(int32_t num_shapes, const ShapeEntry<int32_t> *shapes, int32_t dst_ndims,
+                                         int32_t *dst_shape)
+    {
+        broadcast_shapes(num_shapes, shapes, dst_ndims, dst_shape);
+    }
+
+    void __nac3_ndarray_broadcast_shapes64(int64_t num_shapes, const ShapeEntry<int64_t> *shapes, int64_t dst_ndims,
+                                           int64_t *dst_shape)
+    {
+        broadcast_shapes(num_shapes, shapes, dst_ndims, dst_shape);
+    }
+}

nac3core/irrt/irrt/ndarray/def.hpp

@@ -0,0 +1,47 @@
+#pragma once
+
+#include <irrt/int_types.hpp>
+
+namespace
+{
+/**
+ * @brief The NDArray object
+ *
+ * The official numpy implementations: https://github.com/numpy/numpy/blob/735a477f0bc2b5b84d0e72d92f224bde78d4e069/doc/source/reference/c-api/types-and-structures.rst
+ */
+template <typename SizeT> struct NDArray
+{
+    /**
+     * @brief The underlying data this `ndarray` is pointing to.
+     *
+     * Must be set to `nullptr` to indicate that this NDArray's `data` is uninitialized.
+     */
+    uint8_t *data;
+
+    /**
+     * @brief The number of bytes of a single element in `data`.
+     */
+    SizeT itemsize;
+
+    /**
+     * @brief The number of dimensions of this shape.
+     */
+    SizeT ndims;
+
+    /**
+     * @brief The NDArray shape, with length equal to `ndims`.
+     *
+     * Note that it may contain 0.
+     */
+    SizeT *shape;
+
+    /**
+     * @brief Array strides, with length equal to `ndims`
+     *
+     * The stride values are in units of bytes, not number of elements.
+     *
+     * Note that `strides` can have negative values.
+     */
+    SizeT *strides;
+};
+} // namespace

nac3core/irrt/irrt/ndarray/indexing.hpp

@@ -0,0 +1,246 @@
+#pragma once
+
+#include <irrt/exception.hpp>
+#include <irrt/int_types.hpp>
+#include <irrt/ndarray/basic.hpp>
+#include <irrt/ndarray/def.hpp>
+#include <irrt/slice.hpp>
+
+namespace
+{
+typedef uint8_t NDIndexType;
+
+/**
+ * @brief A single element index
+ *
+ * `data` points to a `SliceIndex`.
+ */
+
+const NDIndexType ND_INDEX_TYPE_SINGLE_ELEMENT = 0;
+/**
+ * @brief A slice index
+ *
+ * `data` points to a `Slice`.
+ */
+const NDIndexType ND_INDEX_TYPE_SLICE = 1;
+
+/**
+ * @brief `np.newaxis` / `None`
+ *
+ * `data` is unused.
+ */
+const NDIndexType ND_INDEX_TYPE_NEWAXIS = 2;
+
+/**
+ * @brief `Ellipsis` / `...`
+ *
+ * `data` is unused.
+ */
+const NDIndexType ND_INDEX_TYPE_ELLIPSIS = 3;
+
+/**
+ * @brief An index used in ndarray indexing
+ */
+struct NDIndex
+{
+    /**
+     * @brief Enum tag to specify the type of index.
+     *
+     * Please see comments of each enum constant.
+     */
+    NDIndexType type;
+
+    /**
+     * @brief The accompanying data associated with `type`.
+     *
+     * Please see comments of each enum constant.
+     */
+    uint8_t *data;
+};
+} // namespace
+
+namespace
+{
+namespace ndarray
+{
+namespace indexing
+{
+/**
+ * @brief Perform ndarray "basic indexing" (https://numpy.org/doc/stable/user/basics.indexing.html#basic-indexing)
+ *
+ * This is function very similar to performing `dst_ndarray = src_ndarray[indices]` in Python (where the variables
+ * can all be found in the parameter of this function).
+ *
+ * In other words, this function takes in an ndarray (`src_ndarray`), index it with `indices`, and return the
+ * indexed array (by writing the result to `dst_ndarray`).
+ *
+ * This function also does proper assertions on `indices`.
+ *
+ * # Notes on `dst_ndarray`
+ * The caller is responsible for allocating space for the resulting ndarray.
+ * Here is what this function expects from `dst_ndarray` when called:
+ *   - `dst_ndarray->data` does not have to be initialized.
+ *   - `dst_ndarray->itemsize` does not have to be initialized.
+ *   - `dst_ndarray->ndims` must be initialized, and it must be equal to the expected `ndims` of the `dst_ndarray` after
+ *       indexing `src_ndarray` with `indices`.
+ *   - `dst_ndarray->shape` must be allocated, through it can contain uninitialized values.
+ *   - `dst_ndarray->strides` must be allocated, through it can contain uninitialized values.
+ * When this function call ends:
+ *   - `dst_ndarray->data` is set to `src_ndarray->data` (`dst_ndarray` is just a view to `src_ndarray`)
+ *   - `dst_ndarray->itemsize` is set to `src_ndarray->itemsize`
+ *   - `dst_ndarray->ndims` is unchanged.
+ *   - `dst_ndarray->shape` is updated according to how `src_ndarray` is indexed.
+ *   - `dst_ndarray->strides` is updated accordingly by how ndarray indexing works.
+ *
+ * @param indices indices to index `src_ndarray`, ordered in the same way you would write them in Python.
+ * @param src_ndarray The NDArray to be indexed.
+ * @param dst_ndarray The resulting NDArray after indexing. Further details in the comments above,
+ */
+template <typename SizeT>
+void index(SizeT num_indices, const NDIndex *indices, const NDArray<SizeT> *src_ndarray, NDArray<SizeT> *dst_ndarray)
+{
+    // Validate `indices`.
+
+    // Expected value of `dst_ndarray->ndims`.
+    SizeT expected_dst_ndims = src_ndarray->ndims;
+    // To check for "too many indices for array: array is ?-dimensional, but ? were indexed"
+    SizeT num_indexed = 0;
+    // There may be ellipsis `...` in `indices`. There can only be 0 or 1 ellipsis.
+    SizeT num_ellipsis = 0;
+
+    for (SizeT i = 0; i < num_indices; i++)
+    {
+        if (indices[i].type == ND_INDEX_TYPE_SINGLE_ELEMENT)
+        {
+            expected_dst_ndims--;
+            num_indexed++;
+        }
+        else if (indices[i].type == ND_INDEX_TYPE_SLICE)
+        {
+            num_indexed++;
+        }
+        else if (indices[i].type == ND_INDEX_TYPE_NEWAXIS)
+        {
+            expected_dst_ndims++;
+        }
+        else if (indices[i].type == ND_INDEX_TYPE_ELLIPSIS)
+        {
+            num_ellipsis++;
+            if (num_ellipsis > 1)
+            {
+                raise_exception(SizeT, EXN_INDEX_ERROR, "an index can only have a single ellipsis ('...')", NO_PARAM,
+                                NO_PARAM, NO_PARAM);
+            }
+        }
+        else
+        {
+            __builtin_unreachable();
+        }
+    }
+
+    debug_assert_eq(SizeT, expected_dst_ndims, dst_ndarray->ndims);
+
+    if (src_ndarray->ndims - num_indexed < 0)
+    {
+        raise_exception(SizeT, EXN_INDEX_ERROR,
+                        "too many indices for array: array is {0}-dimensional, "
+                        "but {1} were indexed",
+                        src_ndarray->ndims, num_indices, NO_PARAM);
+    }
+
+    dst_ndarray->data = src_ndarray->data;
+    dst_ndarray->itemsize = src_ndarray->itemsize;
+
+    // Reference code: https://github.com/wadetb/tinynumpy/blob/0d23d22e07062ffab2afa287374c7b366eebdda1/tinynumpy/tinynumpy.py#L652
+    SizeT src_axis = 0;
+    SizeT dst_axis = 0;
+
+    for (SliceIndex i = 0; i < num_indices; i++)
+    {
+        const NDIndex *index = &indices[i];
+        if (index->type == ND_INDEX_TYPE_SINGLE_ELEMENT)
+        {
+            SliceIndex input = *((SliceIndex *)index->data);
+            SliceIndex k = slice::resolve_index_in_length(src_ndarray->shape[src_axis], input);
+
+            if (k == slice::OUT_OF_BOUNDS)
+            {
+                raise_exception(SizeT, EXN_INDEX_ERROR,
+                                "index {0} is out of bounds for axis {1} "
+                                "with size {2}",
+                                input, src_axis, src_ndarray->shape[src_axis]);
+            }
+
+            dst_ndarray->data += k * src_ndarray->strides[src_axis];
+
+            src_axis++;
+        }
+        else if (index->type == ND_INDEX_TYPE_SLICE)
+        {
+            Slice *slice = (Slice *)index->data;
+
+            Range range = slice->indices_checked<SizeT>(src_ndarray->shape[src_axis]);
+
+            dst_ndarray->data += (SizeT)range.start * src_ndarray->strides[src_axis];
+            dst_ndarray->strides[dst_axis] = ((SizeT)range.step) * src_ndarray->strides[src_axis];
+            dst_ndarray->shape[dst_axis] = (SizeT)range.len<SizeT>();
+
+            dst_axis++;
+            src_axis++;
+        }
+        else if (index->type == ND_INDEX_TYPE_NEWAXIS)
+        {
+            dst_ndarray->strides[dst_axis] = 0;
+            dst_ndarray->shape[dst_axis] = 1;
+
+            dst_axis++;
+        }
+        else if (index->type == ND_INDEX_TYPE_ELLIPSIS)
+        {
+            // The number of ':' entries this '...' implies.
+            SizeT ellipsis_size = src_ndarray->ndims - num_indexed;
+
+            for (SizeT j = 0; j < ellipsis_size; j++)
+            {
+                dst_ndarray->strides[dst_axis] = src_ndarray->strides[src_axis];
+                dst_ndarray->shape[dst_axis] = src_ndarray->shape[src_axis];
+
+                dst_axis++;
+                src_axis++;
+            }
+        }
+        else
+        {
+            __builtin_unreachable();
+        }
+    }
+
+    for (; dst_axis < dst_ndarray->ndims; dst_axis++, src_axis++)
+    {
+        dst_ndarray->shape[dst_axis] = src_ndarray->shape[src_axis];
+        dst_ndarray->strides[dst_axis] = src_ndarray->strides[src_axis];
+    }
+
+    debug_assert_eq(SizeT, src_ndarray->ndims, src_axis);
+    debug_assert_eq(SizeT, dst_ndarray->ndims, dst_axis);
+}
+} // namespace indexing
+} // namespace ndarray
+} // namespace
+
+extern "C"
+{
+    using namespace ndarray::indexing;
+
+    void __nac3_ndarray_index(int32_t num_indices, NDIndex *indices, NDArray<int32_t> *src_ndarray,
+                              NDArray<int32_t> *dst_ndarray)
+    {
+        index(num_indices, indices, src_ndarray, dst_ndarray);
+    }
+
+    void __nac3_ndarray_index64(int64_t num_indices, NDIndex *indices, NDArray<int64_t> *src_ndarray,
+                                NDArray<int64_t> *dst_ndarray)
+    {
+        index(num_indices, indices, src_ndarray, dst_ndarray);
+    }
+}

nac3core/irrt/irrt/ndarray/iter.hpp

@@ -0,0 +1,142 @@
+#pragma once
+
+#include <irrt/int_types.hpp>
+#include <irrt/ndarray/def.hpp>
+
+namespace
+{
+/**
+ * @brief Helper struct to enumerate through all indices under a shape.
+ *
+ * i.e., If `shape` is `[3, 2]`, by repeating `next()`, then you get:
+ * - `[0, 0]`
+ * - `[0, 1]`
+ * - `[1, 0]`
+ * - `[1, 1]`
+ * - `[2, 0]`
+ * - `[2, 1]`
+ * - end.
+ *
+ * Interesting cases:
+ * - If ndims == 0, there is one enumeration.
+ * - If shape contains zeroes, there are no enumerations.
+ */
+template <typename SizeT> struct NDIter
+{
+    SizeT ndims;
+    SizeT *shape;
+    SizeT *strides;
+
+    /**
+     * @brief The current indices.
+     *
+     * Must be allocated by the caller.
+     */
+    SizeT *indices;
+
+    /**
+     * @brief The nth (0-based) index of the current indices.
+     */
+    SizeT nth;
+
+    /**
+     * @brief Pointer to the current element.
+     */
+    uint8_t *element;
+
+    /**
+     * @brief The product of shape.
+     */
+    SizeT size;
+
+    // TODO:: There is something called backstrides to speedup iteration.
+    // See https://ajcr.net/stride-guide-part-1/, and https://docs.scipy.org/doc/numpy-1.13.0/reference/c-api.types-and-structures.html#c.PyArrayIterObject.PyArrayIterObject.backstrides.
+    // Maybe LLVM is clever and knows how to optimize.
+
+    void initialize(SizeT ndims, SizeT *shape, SizeT *strides, uint8_t *element, SizeT *indices)
+    {
+        this->ndims = ndims;
+        this->shape = shape;
+        this->strides = strides;
+
+        this->indices = indices;
+        this->element = element;
+
+        // Compute size and backstrides
+        this->size = 1;
+        for (SizeT i = 0; i < ndims; i++)
+        {
+            this->size *= shape[i];
+        }
+
+        for (SizeT axis = 0; axis < ndims; axis++)
+            indices[axis] = 0;
+        nth = 0;
+    }
+
+    void initialize_by_ndarray(NDArray<SizeT> *ndarray, SizeT *indices)
+    {
+        this->initialize(ndarray->ndims, ndarray->shape, ndarray->strides, ndarray->data, indices);
+    }
+
+    bool has_next()
+    {
+        return nth < size;
+    }
+
+    void next()
+    {
+        for (SizeT i = 0; i < ndims; i++)
+        {
+            SizeT axis = ndims - i - 1;
+            indices[axis]++;
+            if (indices[axis] >= shape[axis])
+            {
+                indices[axis] = 0;
+
+                // TODO: Can be optimized with backstrides.
+                element -= strides[axis] * (shape[axis] - 1);
+            }
+            else
+            {
+                element += strides[axis];
+                break;
+            }
+        }
+        nth++;
+    }
+};
+} // namespace
+
+extern "C"
+{
+    void __nac3_nditer_initialize(NDIter<int32_t> *iter, NDArray<int32_t> *ndarray, int32_t *indices)
+    {
+        iter->initialize_by_ndarray(ndarray, indices);
+    }
+
+    void __nac3_nditer_initialize64(NDIter<int64_t> *iter, NDArray<int64_t> *ndarray, int64_t *indices)
+    {
+        iter->initialize_by_ndarray(ndarray, indices);
+    }
+
+    bool __nac3_nditer_has_next(NDIter<int32_t> *iter)
+    {
+        return iter->has_next();
+    }
+
+    bool __nac3_nditer_has_next64(NDIter<int64_t> *iter)
+    {
+        return iter->has_next();
+    }
+
+    void __nac3_nditer_next(NDIter<int32_t> *iter)
+    {
+        iter->next();
+    }
+
+    void __nac3_nditer_next64(NDIter<int64_t> *iter)
+    {
+        iter->next();
+    }
+}

nac3core/irrt/irrt/ndarray/matmul.hpp

@@ -0,0 +1,92 @@
+#pragma once
+
+#include <irrt/debug.hpp>
+#include <irrt/exception.hpp>
+#include <irrt/int_types.hpp>
+#include <irrt/ndarray/basic.hpp>
+#include <irrt/ndarray/broadcast.hpp>
+#include <irrt/ndarray/iter.hpp>
+
+// NOTE: Everything would be much easier and elegant if einsum is implemented.
+
+namespace
+{
+namespace ndarray
+{
+namespace matmul
+{
+
+/**
+ * @brief Perform the broadcast in `np.einsum("...ij,...jk->...ik", a, b)`.
+ *
+ * Example:
+ *   Suppose      `a_shape ==      [1, 97, 4, 2]`
+ *       and      `b_shape == [99, 98,  1, 2, 5]`,
+ *
+ *   ...then  `new_a_shape == [99, 98, 97, 4, 2]`,
+ *            `new_b_shape == [99, 98, 97, 2, 5]`,
+ *      and     `dst_shape == [99, 98, 97, 4, 5]`.
+ *                             ^^^^^^^^^^  ^^^^
+ *                          (broadcasted)  (4x2 @ 2x5 => 4x5)
+ *
+ * @param a_ndims Length of `a_shape`.
+ * @param a_shape Shape of `a`.
+ * @param b_ndims Length of `b_shape`.
+ * @param b_shape Shape of `b`.
+ * @param final_ndims Should be equal to `max(a_ndims, b_ndims)`. This is the length of `new_a_shape`,
+ * `new_b_shape`, and `dst_shape` - the number of dimensions after broadcasting.
+ */
+template <typename SizeT>
+void calculate_shapes(SizeT a_ndims, SizeT *a_shape, SizeT b_ndims, SizeT *b_shape, SizeT final_ndims,
+                      SizeT *new_a_shape, SizeT *new_b_shape, SizeT *dst_shape)
+{
+    debug_assert(SizeT, a_ndims >= 2);
+    debug_assert(SizeT, b_ndims >= 2);
+    debug_assert_eq(SizeT, max(a_ndims, b_ndims), final_ndims);
+
+    // Check that a and b are compatible for matmul
+    if (a_shape[a_ndims - 1] != b_shape[b_ndims - 2])
+    {
+        // This is a custom error message. Different from NumPy.
+        raise_exception(SizeT, EXN_VALUE_ERROR, "Cannot multiply LHS (shape ?x{0}) with RHS (shape {1}x?})",
+                        a_shape[a_ndims - 1], b_shape[b_ndims - 2], NO_PARAM);
+    }
+
+    const SizeT num_entries = 2;
+    ShapeEntry<SizeT> entries[num_entries] = {{.ndims = a_ndims - 2, .shape = a_shape},
+                                              {.ndims = b_ndims - 2, .shape = b_shape}};
+
+    // TODO: Optimize this
+    ndarray::broadcast::broadcast_shapes<SizeT>(num_entries, entries, final_ndims - 2, new_a_shape);
+    ndarray::broadcast::broadcast_shapes<SizeT>(num_entries, entries, final_ndims - 2, new_b_shape);
+    ndarray::broadcast::broadcast_shapes<SizeT>(num_entries, entries, final_ndims - 2, dst_shape);
+
+    new_a_shape[final_ndims - 2] = a_shape[a_ndims - 2];
+    new_a_shape[final_ndims - 1] = a_shape[a_ndims - 1];
+    new_b_shape[final_ndims - 2] = b_shape[b_ndims - 2];
+    new_b_shape[final_ndims - 1] = b_shape[b_ndims - 1];
+    dst_shape[final_ndims - 2] = a_shape[a_ndims - 2];
+    dst_shape[final_ndims - 1] = b_shape[b_ndims - 1];
+}
+} // namespace matmul
+} // namespace ndarray
+} // namespace
+
+extern "C"
+{
+    using namespace ndarray::matmul;
+
+    void __nac3_ndarray_matmul_calculate_shapes(int32_t a_ndims, int32_t *a_shape, int32_t b_ndims, int32_t *b_shape,
+                                                int32_t final_ndims, int32_t *new_a_shape, int32_t *new_b_shape,
+                                                int32_t *dst_shape)
+    {
+        calculate_shapes(a_ndims, a_shape, b_ndims, b_shape, final_ndims, new_a_shape, new_b_shape, dst_shape);
+    }
+
+    void __nac3_ndarray_matmul_calculate_shapes64(int64_t a_ndims, int64_t *a_shape, int64_t b_ndims, int64_t *b_shape,
+                                                  int64_t final_ndims, int64_t *new_a_shape, int64_t *new_b_shape,
+                                                  int64_t *dst_shape)
+    {
+        calculate_shapes(a_ndims, a_shape, b_ndims, b_shape, final_ndims, new_a_shape, new_b_shape, dst_shape);
+    }
+}

nac3core/irrt/irrt/ndarray/reshape.hpp

@@ -0,0 +1,125 @@
+#pragma once
+
+#include <irrt/int_types.hpp>
+#include <irrt/ndarray/def.hpp>
+
+namespace
+{
+namespace ndarray
+{
+namespace reshape
+{
+/**
+ * @brief Perform assertions on and resolve unknown dimensions in `new_shape` in `np.reshape(<ndarray>, new_shape)`
+ *
+ * If `new_shape` indeed contains unknown dimensions (specified with `-1`, just like numpy), `new_shape` will be
+ * modified to contain the resolved dimension.
+ *
+ * To perform assertions on and resolve unknown dimensions in `new_shape`, we don't need the actual
+ * `<ndarray>` object itself, but only the `.size` of the `<ndarray>`.
+ *
+ * @param size The `.size` of `<ndarray>`
+ * @param new_ndims Number of elements in `new_shape`
+ * @param new_shape Target shape to reshape to
+ */
+template <typename SizeT> void resolve_and_check_new_shape(SizeT size, SizeT new_ndims, SizeT *new_shape)
+{
+    // Is there a -1 in `new_shape`?
+    bool neg1_exists = false;
+    // Location of -1, only initialized if `neg1_exists` is true
+    SizeT neg1_axis_i;
+    // The computed ndarray size of `new_shape`
+    SizeT new_size = 1;
+
+    for (SizeT axis_i = 0; axis_i < new_ndims; axis_i++)
+    {
+        SizeT dim = new_shape[axis_i];
+        if (dim < 0)
+        {
+            if (dim == -1)
+            {
+                if (neg1_exists)
+                {
+                    // Multiple `-1` found. Throw an error.
+                    raise_exception(SizeT, EXN_VALUE_ERROR, "can only specify one unknown dimension", NO_PARAM,
+                                    NO_PARAM, NO_PARAM);
+                }
+                else
+                {
+                    neg1_exists = true;
+                    neg1_axis_i = axis_i;
+                }
+            }
+            else
+            {
+                // TODO: What? In `np.reshape` any negative dimensions is
+                // treated like its `-1`.
+                //
+                // Try running `np.zeros((3, 4)).reshape((-999, 2))`
+                //
+                // It is not documented by numpy.
+                // Throw an error for now...
+
+                raise_exception(SizeT, EXN_VALUE_ERROR, "Found non -1 negative dimension {0} on axis {1}", dim, axis_i,
+                                NO_PARAM);
+            }
+        }
+        else
+        {
+            new_size *= dim;
+        }
+    }
+
+    bool can_reshape;
+    if (neg1_exists)
+    {
+        // Let `x` be the unknown dimension
+        // Solve `x * <new_size> = <size>`
+        if (new_size == 0 && size == 0)
+        {
+            // `x` has infinitely many solutions
+            can_reshape = false;
+        }
+        else if (new_size == 0 && size != 0)
+        {
+            // `x` has no solutions
+            can_reshape = false;
+        }
+        else if (size % new_size != 0)
+        {
+            // `x` has no integer solutions
+            can_reshape = false;
+        }
+        else
+        {
+            can_reshape = true;
+            new_shape[neg1_axis_i] = size / new_size; // Resolve dimension
+        }
+    }
+    else
+    {
+        can_reshape = (new_size == size);
+    }
+
+    if (!can_reshape)
+    {
+        raise_exception(SizeT, EXN_VALUE_ERROR, "cannot reshape array of size {0} into given shape", size, NO_PARAM,
+                        NO_PARAM);
+    }
+}
+} // namespace reshape
+} // namespace ndarray
+} // namespace
+
+extern "C"
+{
+    void __nac3_ndarray_reshape_resolve_and_check_new_shape(int32_t size, int32_t new_ndims, int32_t *new_shape)
+    {
+        ndarray::reshape::resolve_and_check_new_shape(size, new_ndims, new_shape);
+    }
+
+    void __nac3_ndarray_reshape_resolve_and_check_new_shape64(int64_t size, int64_t new_ndims, int64_t *new_shape)
+    {
+        ndarray::reshape::resolve_and_check_new_shape(size, new_ndims, new_shape);
+    }
+}

nac3core/irrt/irrt/ndarray/transpose.hpp

@@ -0,0 +1,155 @@
+#pragma once
+
+#include <irrt/int_types.hpp>
+#include <irrt/ndarray/def.hpp>
+#include <irrt/slice.hpp>
+
+/*
+ * Notes on `np.transpose(<array>, <axes>)`
+ *
+ * TODO: `axes`, if specified, can actually contain negative indices,
+ * but it is not documented in numpy.
+ *
+ * Supporting it for now.
+ */
+
+namespace
+{
+namespace ndarray
+{
+namespace transpose
+{
+/**
+ * @brief Do assertions on `<axes>` in `np.transpose(<array>, <axes>)`.
+ *
+ * Note that `np.transpose`'s `<axe>` argument is optional. If the argument
+ * is specified but the user, use this function to do assertions on it.
+ *
+ * @param ndims The number of dimensions of `<array>`
+ * @param num_axes Number of elements in `<axes>` as specified by the user.
+ *   This should be equal to `ndims`. If not, a "ValueError: axes don't match array" is thrown.
+ * @param axes The user specified `<axes>`.
+ */
+template <typename SizeT> void assert_transpose_axes(SizeT ndims, SizeT num_axes, const SizeT *axes)
+{
+    if (ndims != num_axes)
+    {
+        raise_exception(SizeT, EXN_VALUE_ERROR, "axes don't match array", NO_PARAM, NO_PARAM, NO_PARAM);
+    }
+
+    // TODO: Optimize this
+    bool *axe_specified = (bool *)__builtin_alloca(sizeof(bool) * ndims);
+    for (SizeT i = 0; i < ndims; i++)
+        axe_specified[i] = false;
+
+    for (SizeT i = 0; i < ndims; i++)
+    {
+        SizeT axis = slice::resolve_index_in_length(ndims, axes[i]);
+        if (axis == slice::OUT_OF_BOUNDS)
+        {
+            // TODO: numpy actually throws a `numpy.exceptions.AxisError`
+            raise_exception(SizeT, EXN_VALUE_ERROR, "axis {0} is out of bounds for array of dimension {1}", axis, ndims,
+                            NO_PARAM);
+        }
+
+        if (axe_specified[axis])
+        {
+            raise_exception(SizeT, EXN_VALUE_ERROR, "repeated axis in transpose", NO_PARAM, NO_PARAM, NO_PARAM);
+        }
+
+        axe_specified[axis] = true;
+    }
+}
+
+/**
+ * @brief Create a transpose view of `src_ndarray` and perform proper assertions.
+ *
+ * This function is very similar to doing `dst_ndarray = np.transpose(src_ndarray, <axes>)`.
+ * If `<axes>` is supposed to be `None`, caller can pass in a `nullptr` to `<axes>`.
+ *
+ * The transpose view created is returned by modifying `dst_ndarray`.
+ *
+ * The caller is responsible for setting up `dst_ndarray` before calling this function.
+ * Here is what this function expects from `dst_ndarray` when called:
+ *   - `dst_ndarray->data` does not have to be initialized.
+ *   - `dst_ndarray->itemsize` does not have to be initialized.
+ *   - `dst_ndarray->ndims` must be initialized, must be equal to `src_ndarray->ndims`.
+ *   - `dst_ndarray->shape` must be allocated, through it can contain uninitialized values.
+ *   - `dst_ndarray->strides` must be allocated, through it can contain uninitialized values.
+ * When this function call ends:
+ *   - `dst_ndarray->data` is set to `src_ndarray->data` (`dst_ndarray` is just a view to `src_ndarray`)
+ *   - `dst_ndarray->itemsize` is set to `src_ndarray->itemsize`
+ *   - `dst_ndarray->ndims` is unchanged
+ *   - `dst_ndarray->shape` is updated according to how `np.transpose` works
+ *   - `dst_ndarray->strides` is updated according to how `np.transpose` works
+ *
+ * @param src_ndarray The NDArray to build a transpose view on
+ * @param dst_ndarray The resulting NDArray after transpose. Further details in the comments above,
+ * @param num_axes Number of elements in axes. Unused if `axes` is nullptr.
+ * @param axes Axes permutation. Set it to `nullptr` if `<axes>` is `None`.
+ */
+template <typename SizeT>
+void transpose(const NDArray<SizeT> *src_ndarray, NDArray<SizeT> *dst_ndarray, SizeT num_axes, const SizeT *axes)
+{
+    debug_assert_eq(SizeT, src_ndarray->ndims, dst_ndarray->ndims);
+    const auto ndims = src_ndarray->ndims;
+
+    if (axes != nullptr)
+        assert_transpose_axes(ndims, num_axes, axes);
+
+    dst_ndarray->data = src_ndarray->data;
+    dst_ndarray->itemsize = src_ndarray->itemsize;
+
+    // Check out https://ajcr.net/stride-guide-part-2/ to see how `np.transpose` works behind the scenes.
+    if (axes == nullptr)
+    {
+        // `np.transpose(<array>, axes=None)`
+
+        /*
+         * Minor note: `np.transpose(<array>, axes=None)` is equivalent to
+         * `np.transpose(<array>, axes=[N-1, N-2, ..., 0])` - basically it
+         * is reversing the order of strides and shape.
+         *
+         * This is a fast implementation to handle this special (but very common) case.
+         */
+
+        for (SizeT axis = 0; axis < ndims; axis++)
+        {
+            dst_ndarray->shape[axis] = src_ndarray->shape[ndims - axis - 1];
+            dst_ndarray->strides[axis] = src_ndarray->strides[ndims - axis - 1];
+        }
+    }
+    else
+    {
+        // `np.transpose(<array>, <axes>)`
+
+        // Permute strides and shape according to `axes`, while resolving negative indices in `axes`
+        for (SizeT axis = 0; axis < ndims; axis++)
+        {
+            // `i` cannot be OUT_OF_BOUNDS because of assertions
+            SizeT i = slice::resolve_index_in_length(ndims, axes[axis]);
+
+            dst_ndarray->shape[axis] = src_ndarray->shape[i];
+            dst_ndarray->strides[axis] = src_ndarray->strides[i];
+        }
+    }
+}
+} // namespace transpose
+} // namespace ndarray
+} // namespace
+
+extern "C"
+{
+    using namespace ndarray::transpose;
+    void __nac3_ndarray_transpose(const NDArray<int32_t> *src_ndarray, NDArray<int32_t> *dst_ndarray, int32_t num_axes,
+                                  const int32_t *axes)
+    {
+        transpose(src_ndarray, dst_ndarray, num_axes, axes);
+    }
+
+    void __nac3_ndarray_transpose64(const NDArray<int64_t> *src_ndarray, NDArray<int64_t> *dst_ndarray,
+                                    int64_t num_axes, const int64_t *axes)
+    {
+        transpose(src_ndarray, dst_ndarray, num_axes, axes);
+    }
+}

nac3core/irrt/irrt/original.hpp

@@ -4,7 +4,7 @@
 #include <irrt/math_util.hpp>
 
 // NDArray indices are always `uint32_t`.
-using NDIndex = uint32_t;
+using NDIndexInt = uint32_t;
 // The type of an index or a value describing the length of a range/slice is always `int32_t`.
 using SliceIndex = int32_t;
 
@@ -44,7 +44,7 @@ SizeT __nac3_ndarray_calc_size_impl(const SizeT *list_data, SizeT list_len, Size
 }
 
 template <typename SizeT>
-void __nac3_ndarray_calc_nd_indices_impl(SizeT index, const SizeT *dims, SizeT num_dims, NDIndex *idxs)
+void __nac3_ndarray_calc_nd_indices_impl(SizeT index, const SizeT *dims, SizeT num_dims, NDIndexInt *idxs)
 {
     SizeT stride = 1;
     for (SizeT dim = 0; dim < num_dims; dim++)
@@ -57,7 +57,7 @@ void __nac3_ndarray_calc_nd_indices_impl(SizeT index, const SizeT *dims, SizeT n
 }
 
 template <typename SizeT>
-SizeT __nac3_ndarray_flatten_index_impl(const SizeT *dims, SizeT num_dims, const NDIndex *indices, SizeT num_indices)
+SizeT __nac3_ndarray_flatten_index_impl(const SizeT *dims, SizeT num_dims, const NDIndexInt *indices, SizeT num_indices)
 {
     SizeT idx = 0;
     SizeT stride = 1;
@@ -115,8 +115,8 @@ void __nac3_ndarray_calc_broadcast_impl(const SizeT *lhs_dims, SizeT lhs_ndims,
 }
 
 template <typename SizeT>
-void __nac3_ndarray_calc_broadcast_idx_impl(const SizeT *src_dims, SizeT src_ndims, const NDIndex *in_idx,
-                                            NDIndex *out_idx)
+void __nac3_ndarray_calc_broadcast_idx_impl(const SizeT *src_dims, SizeT src_ndims, const NDIndexInt *in_idx,
+                                            NDIndexInt *out_idx)
 {
     for (SizeT i = 0; i < src_ndims; ++i)
     {
@@ -324,23 +324,23 @@ extern "C"
         return __nac3_ndarray_calc_size_impl(list_data, list_len, begin_idx, end_idx);
     }
 
-    void __nac3_ndarray_calc_nd_indices(uint32_t index, const uint32_t *dims, uint32_t num_dims, NDIndex *idxs)
+    void __nac3_ndarray_calc_nd_indices(uint32_t index, const uint32_t *dims, uint32_t num_dims, NDIndexInt *idxs)
     {
         __nac3_ndarray_calc_nd_indices_impl(index, dims, num_dims, idxs);
     }
 
-    void __nac3_ndarray_calc_nd_indices64(uint64_t index, const uint64_t *dims, uint64_t num_dims, NDIndex *idxs)
+    void __nac3_ndarray_calc_nd_indices64(uint64_t index, const uint64_t *dims, uint64_t num_dims, NDIndexInt *idxs)
     {
         __nac3_ndarray_calc_nd_indices_impl(index, dims, num_dims, idxs);
     }
 
-    uint32_t __nac3_ndarray_flatten_index(const uint32_t *dims, uint32_t num_dims, const NDIndex *indices,
+    uint32_t __nac3_ndarray_flatten_index(const uint32_t *dims, uint32_t num_dims, const NDIndexInt *indices,
                                           uint32_t num_indices)
     {
         return __nac3_ndarray_flatten_index_impl(dims, num_dims, indices, num_indices);
     }
 
-    uint64_t __nac3_ndarray_flatten_index64(const uint64_t *dims, uint64_t num_dims, const NDIndex *indices,
+    uint64_t __nac3_ndarray_flatten_index64(const uint64_t *dims, uint64_t num_dims, const NDIndexInt *indices,
                                             uint64_t num_indices)
     {
         return __nac3_ndarray_flatten_index_impl(dims, num_dims, indices, num_indices);
@@ -358,14 +358,14 @@ extern "C"
         return __nac3_ndarray_calc_broadcast_impl(lhs_dims, lhs_ndims, rhs_dims, rhs_ndims, out_dims);
     }
 
-    void __nac3_ndarray_calc_broadcast_idx(const uint32_t *src_dims, uint32_t src_ndims, const NDIndex *in_idx,
-                                           NDIndex *out_idx)
+    void __nac3_ndarray_calc_broadcast_idx(const uint32_t *src_dims, uint32_t src_ndims, const NDIndexInt *in_idx,
+                                           NDIndexInt *out_idx)
     {
         __nac3_ndarray_calc_broadcast_idx_impl(src_dims, src_ndims, in_idx, out_idx);
     }
 
-    void __nac3_ndarray_calc_broadcast_idx64(const uint64_t *src_dims, uint64_t src_ndims, const NDIndex *in_idx,
-                                             NDIndex *out_idx)
+    void __nac3_ndarray_calc_broadcast_idx64(const uint64_t *src_dims, uint64_t src_ndims, const NDIndexInt *in_idx,
+                                             NDIndexInt *out_idx)
     {
         __nac3_ndarray_calc_broadcast_idx_impl(src_dims, src_ndims, in_idx, out_idx);
     }

nac3core/irrt/irrt/slice.hpp

@@ -0,0 +1,200 @@
+#pragma once
+
+#include <irrt/debug.hpp>
+#include <irrt/exception.hpp>
+#include <irrt/int_types.hpp>
+#include <irrt/math_util.hpp>
+
+// The type of an index or a value describing the length of a
+// range/slice is always `int32_t`.
+using SliceIndex = int32_t;
+
+namespace
+{
+
+/**
+ * @brief A Python range.
+ */
+struct Range
+{
+    SliceIndex start;
+    SliceIndex stop;
+    SliceIndex step;
+
+    /**
+     * @brief Calculate the `len()` of this range.
+     */
+    template <typename SizeT> SliceIndex len()
+    {
+        // Reference: https://github.com/python/cpython/blob/9dbd12375561a393eaec4b21ee4ac568a407cdb0/Objects/rangeobject.c#L933
+        debug_assert(SizeT, step != 0);
+
+        if (step > 0 && start < stop)
+            return 1 + (stop - 1 - start) / step;
+        else if (step < 0 && start > stop)
+            return 1 + (start - 1 - stop) / (-step);
+        else
+            return 0;
+    }
+};
+
+namespace slice
+{
+/**
+ * @brief Resolve a slice index under a given length like Python indexing.
+ *
+ * In Python, if you have a `list` of length 100, `list[-1]` resolves to
+ * `list[99]`, so `resolve_index_in_length_clamped(100, -1)` returns `99`.
+ *
+ * If `length` is 0, 0 is returned for any value of `index`.
+ *
+ * If `index` is out of bounds, clamps the returned value between `0` and
+ * `length - 1` (inclusive).
+ *
+ */
+SliceIndex resolve_index_in_length_clamped(SliceIndex length, SliceIndex index)
+{
+    if (index < 0)
+    {
+        return max<SliceIndex>(length + index, 0);
+    }
+    else
+    {
+        return min<SliceIndex>(length, index);
+    }
+}
+
+const SliceIndex OUT_OF_BOUNDS = -1;
+
+/**
+ * @brief Like `resolve_index_in_length_clamped`, but returns `OUT_OF_BOUNDS`
+ * if `index` is out of bounds.
+ */
+SliceIndex resolve_index_in_length(SliceIndex length, SliceIndex index)
+{
+    SliceIndex resolved = index < 0 ? length + index : index;
+    if (0 <= resolved && resolved < length)
+    {
+        return resolved;
+    }
+    else
+    {
+        return OUT_OF_BOUNDS;
+    }
+}
+} // namespace slice
+
+/**
+ * @brief A Python-like slice with **unresolved** indices.
+ */
+struct Slice
+{
+    bool start_defined;
+    SliceIndex start;
+
+    bool stop_defined;
+    SliceIndex stop;
+
+    bool step_defined;
+    SliceIndex step;
+
+    Slice()
+    {
+        this->reset();
+    }
+
+    void reset()
+    {
+        this->start_defined = false;
+        this->stop_defined = false;
+        this->step_defined = false;
+    }
+
+    void set_start(SliceIndex start)
+    {
+        this->start_defined = true;
+        this->start = start;
+    }
+
+    void set_stop(SliceIndex stop)
+    {
+        this->stop_defined = true;
+        this->stop = stop;
+    }
+
+    void set_step(SliceIndex step)
+    {
+        this->step_defined = true;
+        this->step = step;
+    }
+
+    /**
+     * @brief Resolve this slice as a range.
+     *
+     * In Python, this would be `range(*slice(start, stop, step).indices(length))`.
+     */
+    template <typename SizeT> Range indices(SliceIndex length)
+    {
+        // Reference: https://github.com/python/cpython/blob/main/Objects/sliceobject.c#L388
+        debug_assert(SizeT, length >= 0);
+
+        Range result;
+
+        result.step = step_defined ? step : 1;
+        bool step_is_negative = result.step < 0;
+
+        SliceIndex lower, upper;
+        if (step_is_negative)
+        {
+            lower = -1;
+            upper = length - 1;
+        }
+        else
+        {
+            lower = 0;
+            upper = length;
+        }
+
+        if (start_defined)
+        {
+            result.start = start < 0 ? max(lower, start + length) : min(upper, start);
+        }
+        else
+        {
+            result.start = step_is_negative ? upper : lower;
+        }
+
+        if (stop_defined)
+        {
+            result.stop = stop < 0 ? max(lower, stop + length) : min(upper, stop);
+        }
+        else
+        {
+            result.stop = step_is_negative ? lower : upper;
+        }
+
+        return result;
+    }
+
+    /**
+     * @brief Like `.indices()` but with assertions.
+     */
+    template <typename SizeT> Range indices_checked(SliceIndex length)
+    {
+        // TODO: Switch to `SizeT length`
+
+        if (length < 0)
+        {
+            raise_exception(SizeT, EXN_VALUE_ERROR, "length should not be negative, got {0}", length, NO_PARAM,
+                            NO_PARAM);
+        }
+
+        if (this->step_defined && this->step == 0)
+        {
+            raise_exception(SizeT, EXN_VALUE_ERROR, "slice step cannot be zero", NO_PARAM, NO_PARAM, NO_PARAM);
+        }
+
+        return this->indices<SizeT>(length);
+    }
+};
+} // namespace

nac3core/src/codegen/expr.rs

@@ -1,8 +1,8 @@
 use crate::{
     codegen::{
         classes::{
-            ArrayLikeIndexer, ArrayLikeValue, ListType, ListValue, NDArrayValue, ProxyType,
-            ProxyValue, RangeValue, TypedArrayLikeAccessor, UntypedArrayLikeAccessor,
+            ArrayLikeIndexer, ArrayLikeValue, ListType, ListValue, ProxyType, ProxyValue,
+            RangeValue, UntypedArrayLikeAccessor,
         },
         concrete_type::{ConcreteFuncArg, ConcreteTypeEnum, ConcreteTypeStore},
         gen_in_range_check, get_llvm_abi_type, get_llvm_type, get_va_count_arg_name,
@@ -11,7 +11,8 @@ use crate::{
             call_expect, call_float_floor, call_float_pow, call_float_powi, call_int_smax,
             call_memcpy_generic,
         },
-        need_sret, numpy,
+        need_sret,
+        object::ndarray::{NDArrayOut, ScalarOrNDArray},
         stmt::{
             gen_for_callback_incrementing, gen_if_callback, gen_if_else_expr_callback, gen_raise,
             gen_var,
@@ -19,11 +20,7 @@ use crate::{
         CodeGenContext, CodeGenTask, CodeGenerator,
     },
     symbol_resolver::{SymbolValue, ValueEnum},
-    toplevel::{
-        helper::PrimDef,
-        numpy::{make_ndarray_ty, unpack_ndarray_var_tys},
-        DefinitionId, TopLevelDef,
-    },
+    toplevel::{helper::PrimDef, DefinitionId, TopLevelDef},
     typecheck::{
         magic_methods::{Binop, BinopVariant, HasOpInfo},
         typedef::{FunSignature, FuncArg, Type, TypeEnum, TypeVarId, Unifier, VarMap},
@@ -32,7 +29,10 @@ use crate::{
 use inkwell::{
     attributes::{Attribute, AttributeLoc},
     types::{AnyType, BasicType, BasicTypeEnum},
-    values::{BasicValueEnum, CallSiteValue, FunctionValue, IntValue, PointerValue, StructValue},
+    values::{
+        BasicValue, BasicValueEnum, CallSiteValue, FunctionValue, IntValue, PointerValue,
+        StructValue,
+    },
     AddressSpace, IntPredicate, OptimizationLevel,
 };
 use itertools::{chain, izip, Either, Itertools};
@@ -43,6 +43,14 @@ use nac3parser::ast::{
 use std::iter::{repeat, repeat_with};
 use std::{collections::HashMap, convert::TryInto, iter::once, iter::zip};
 
+use super::{
+    model::*,
+    object::{
+        any::AnyObject,
+        ndarray::{indexing::util::gen_ndarray_subscript_ndindices, NDArrayObject},
+    },
+};
+
 pub fn get_subst_key(
     unifier: &mut Unifier,
     obj: Option<Type>,
@@ -1539,99 +1547,75 @@ pub fn gen_binop_expr_with_values<'ctx, G: CodeGenerator>(
     } else if ty1.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id())
         || ty2.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id())
     {
-        let llvm_usize = generator.get_size_type(ctx.ctx);
+        let left =
+            ScalarOrNDArray::split_object(generator, ctx, AnyObject { ty: ty1, value: left_val });
+        let right =
+            ScalarOrNDArray::split_object(generator, ctx, AnyObject { ty: ty2, value: right_val });
 
-        let is_ndarray1 = ty1.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id());
-        let is_ndarray2 = ty2.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id());
+        // Inhomogeneous binary operations are not supported.
+        assert!(ctx.unifier.unioned(left.get_dtype(), right.get_dtype()));
 
-        if is_ndarray1 && is_ndarray2 {
-            let (ndarray_dtype1, _) = unpack_ndarray_var_tys(&mut ctx.unifier, ty1);
-            let (ndarray_dtype2, _) = unpack_ndarray_var_tys(&mut ctx.unifier, ty2);
+        let common_dtype = left.get_dtype();
 
-            assert!(ctx.unifier.unioned(ndarray_dtype1, ndarray_dtype2));
+        let out = match op.variant {
+            BinopVariant::Normal => NDArrayOut::NewNDArray { dtype: common_dtype },
+            BinopVariant::AugAssign => {
+                // If this is an augmented assignment.
+                // `left` has to be an ndarray. If it were a scalar then NAC3 simply doesn't support it.
+                if let ScalarOrNDArray::NDArray(out_ndarray) = left {
+                    NDArrayOut::WriteToNDArray { ndarray: out_ndarray }
+                } else {
+                    panic!("left must be an ndarray")
+                }
+            }
+        };
 
-            let left_val =
-                NDArrayValue::from_ptr_val(left_val.into_pointer_value(), llvm_usize, None);
-            let right_val =
-                NDArrayValue::from_ptr_val(right_val.into_pointer_value(), llvm_usize, None);
-
-            let res = if op.base == Operator::MatMult {
-                // MatMult is the only binop which is not an elementwise op
-                numpy::ndarray_matmul_2d(
-                    generator,
-                    ctx,
-                    ndarray_dtype1,
-                    match op.variant {
-                        BinopVariant::Normal => None,
-                        BinopVariant::AugAssign => Some(left_val),
-                    },
-                    left_val,
-                    right_val,
-                )?
-            } else {
-                numpy::ndarray_elementwise_binop_impl(
-                    generator,
-                    ctx,
-                    ndarray_dtype1,
-                    match op.variant {
-                        BinopVariant::Normal => None,
-                        BinopVariant::AugAssign => Some(left_val),
-                    },
-                    (left_val.as_base_value().into(), false),
-                    (right_val.as_base_value().into(), false),
-                    |generator, ctx, (lhs, rhs)| {
-                        gen_binop_expr_with_values(
-                            generator,
-                            ctx,
-                            (&Some(ndarray_dtype1), lhs),
-                            op,
-                            (&Some(ndarray_dtype2), rhs),
-                            ctx.current_loc,
-                        )?
-                        .unwrap()
-                        .to_basic_value_enum(
-                            ctx,
-                            generator,
-                            ndarray_dtype1,
-                        )
-                    },
-                )?
-            };
-
-            Ok(Some(res.as_base_value().into()))
+        if op.base == Operator::MatMult {
+            // Handle matrix multiplication.
+            let left = left.to_ndarray(generator, ctx);
+            let right = right.to_ndarray(generator, ctx);
+            let result = NDArrayObject::matmul(generator, ctx, left, right, out)
+                .split_unsized(generator, ctx);
+            Ok(Some(ValueEnum::Dynamic(result.to_basic_value_enum())))
         } else {
-            let (ndarray_dtype, _) =
-                unpack_ndarray_var_tys(&mut ctx.unifier, if is_ndarray1 { ty1 } else { ty2 });
-            let ndarray_val = NDArrayValue::from_ptr_val(
-                if is_ndarray1 { left_val } else { right_val }.into_pointer_value(),
-                llvm_usize,
-                None,
-            );
-            let res = numpy::ndarray_elementwise_binop_impl(
+            // For other operations, they are all elementwise operations.
+
+            // There are only three cases:
+            // - LHS is a scalar, RHS is an ndarray.
+            // - LHS is an ndarray, RHS is a scalar.
+            // - LHS is an ndarray, RHS is an ndarray.
+            //
+            // For all cases, the scalar operand is promoted to an ndarray,
+            // the two are then broadcasted, and starmapped through.
+
+            let left = left.to_ndarray(generator, ctx);
+            let right = right.to_ndarray(generator, ctx);
+
+            let result = NDArrayObject::broadcast_starmap(
                 generator,
                 ctx,
-                ndarray_dtype,
-                match op.variant {
-                    BinopVariant::Normal => None,
-                    BinopVariant::AugAssign => Some(ndarray_val),
-                },
-                (left_val, !is_ndarray1),
-                (right_val, !is_ndarray2),
-                |generator, ctx, (lhs, rhs)| {
-                    gen_binop_expr_with_values(
+                &[left, right],
+                out,
+                |generator, ctx, scalars| {
+                    let left_value = scalars[0];
+                    let right_value = scalars[1];
+
+                    let result = gen_binop_expr_with_values(
                         generator,
                         ctx,
-                        (&Some(ndarray_dtype), lhs),
+                        (&Some(left.dtype), left_value),
                         op,
-                        (&Some(ndarray_dtype), rhs),
+                        (&Some(right.dtype), right_value),
                         ctx.current_loc,
                     )?
                     .unwrap()
-                    .to_basic_value_enum(ctx, generator, ndarray_dtype)
-                },
-            )?;
+                    .to_basic_value_enum(ctx, generator, common_dtype)?;
 
-            Ok(Some(res.as_base_value().into()))
+                    Ok(result)
+                },
+            )
+            .unwrap();
+            Ok(Some(ValueEnum::Dynamic(result.instance.value.as_basic_value_enum())))
         }
     } else {
         let left_ty_enum = ctx.unifier.get_ty_immutable(left_ty.unwrap());
@@ -1789,14 +1773,12 @@ pub fn gen_unaryop_expr_with_values<'ctx, G: CodeGenerator>(
             _ => val.into(),
         }
     } else if ty.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id()) {
-        let llvm_usize = generator.get_size_type(ctx.ctx);
-        let (ndarray_dtype, _) = unpack_ndarray_var_tys(&mut ctx.unifier, ty);
-
-        let val = NDArrayValue::from_ptr_val(val.into_pointer_value(), llvm_usize, None);
+        let ndarray = AnyObject { value: val, ty };
+        let ndarray = NDArrayObject::from_object(generator, ctx, ndarray);
 
         // ndarray uses `~` rather than `not` to perform elementwise inversion, convert it before
         // passing it to the elementwise codegen function
-        let op = if ndarray_dtype.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::Bool.id()) {
+        let op = if ndarray.dtype.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::Bool.id()) {
             if op == ast::Unaryop::Invert {
                 ast::Unaryop::Not
             } else {
@@ -1809,20 +1791,18 @@ pub fn gen_unaryop_expr_with_values<'ctx, G: CodeGenerator>(
             op
         };
 
-        let res = numpy::ndarray_elementwise_unaryop_impl(
+        let mapped_ndarray = ndarray.map(
             generator,
             ctx,
-            ndarray_dtype,
-            None,
-            val,
-            |generator, ctx, val| {
-                gen_unaryop_expr_with_values(generator, ctx, op, (&Some(ndarray_dtype), val))?
+            NDArrayOut::NewNDArray { dtype: ndarray.dtype },
+            |generator, ctx, scalar| {
+                gen_unaryop_expr_with_values(generator, ctx, op, (&Some(ndarray.dtype), scalar))?
                     .unwrap()
-                    .to_basic_value_enum(ctx, generator, ndarray_dtype)
+                    .to_basic_value_enum(ctx, generator, ndarray.dtype)
             },
         )?;
 
-        res.as_base_value().into()
+        ValueEnum::Dynamic(mapped_ndarray.instance.value.as_basic_value_enum())
     } else {
         unimplemented!()
     }))
@@ -1865,85 +1845,46 @@ pub fn gen_cmpop_expr_with_values<'ctx, G: CodeGenerator>(
         if left_ty.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id())
             || right_ty.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id())
         {
-            let llvm_usize = generator.get_size_type(ctx.ctx);
-
-            let (Some(left_ty), lhs) = left else { unreachable!() };
-            let (Some(right_ty), rhs) = comparators[0] else { unreachable!() };
+            let (Some(left_ty), left) = left else { unreachable!() };
+            let (Some(right_ty), right) = comparators[0] else { unreachable!() };
             let op = ops[0];
 
-            let is_ndarray1 =
-                left_ty.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id());
-            let is_ndarray2 =
-                right_ty.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id());
+            let left = AnyObject { value: left, ty: left_ty };
+            let left =
+                ScalarOrNDArray::split_object(generator, ctx, left).to_ndarray(generator, ctx);
 
-            return if is_ndarray1 && is_ndarray2 {
-                let (ndarray_dtype1, _) = unpack_ndarray_var_tys(&mut ctx.unifier, left_ty);
-                let (ndarray_dtype2, _) = unpack_ndarray_var_tys(&mut ctx.unifier, right_ty);
+            let right = AnyObject { value: right, ty: right_ty };
+            let right =
+                ScalarOrNDArray::split_object(generator, ctx, right).to_ndarray(generator, ctx);
 
-                assert!(ctx.unifier.unioned(ndarray_dtype1, ndarray_dtype2));
+            let result_ndarray = NDArrayObject::broadcast_starmap(
+                generator,
+                ctx,
+                &[left, right],
+                NDArrayOut::NewNDArray { dtype: ctx.primitives.bool },
+                |generator, ctx, scalars| {
+                    let left_scalar = scalars[0];
+                    let right_scalar = scalars[1];
 
-                let left_val =
-                    NDArrayValue::from_ptr_val(lhs.into_pointer_value(), llvm_usize, None);
-                let res = numpy::ndarray_elementwise_binop_impl(
-                    generator,
-                    ctx,
-                    ctx.primitives.bool,
-                    None,
-                    (left_val.as_base_value().into(), false),
-                    (rhs, false),
-                    |generator, ctx, (lhs, rhs)| {
-                        let val = gen_cmpop_expr_with_values(
-                            generator,
-                            ctx,
-                            (Some(ndarray_dtype1), lhs),
-                            &[op],
-                            &[(Some(ndarray_dtype2), rhs)],
-                        )?
-                        .unwrap()
-                        .to_basic_value_enum(
-                            ctx,
-                            generator,
-                            ctx.primitives.bool,
-                        )?;
+                    let val = gen_cmpop_expr_with_values(
+                        generator,
+                        ctx,
+                        (Some(left.dtype), left_scalar),
+                        &[op],
+                        &[(Some(right.dtype), right_scalar)],
+                    )?
+                    .unwrap()
+                    .to_basic_value_enum(
+                        ctx,
+                        generator,
+                        ctx.primitives.bool,
+                    )?;
 
-                        Ok(generator.bool_to_i8(ctx, val.into_int_value()).into())
-                    },
-                )?;
+                    Ok(generator.bool_to_i8(ctx, val.into_int_value()).into())
+                },
+            )?;
 
-                Ok(Some(res.as_base_value().into()))
-            } else {
-                let (ndarray_dtype, _) = unpack_ndarray_var_tys(
-                    &mut ctx.unifier,
-                    if is_ndarray1 { left_ty } else { right_ty },
-                );
-                let res = numpy::ndarray_elementwise_binop_impl(
-                    generator,
-                    ctx,
-                    ctx.primitives.bool,
-                    None,
-                    (lhs, !is_ndarray1),
-                    (rhs, !is_ndarray2),
-                    |generator, ctx, (lhs, rhs)| {
-                        let val = gen_cmpop_expr_with_values(
-                            generator,
-                            ctx,
-                            (Some(ndarray_dtype), lhs),
-                            &[op],
-                            &[(Some(ndarray_dtype), rhs)],
-                        )?
-                        .unwrap()
-                        .to_basic_value_enum(
-                            ctx,
-                            generator,
-                            ctx.primitives.bool,
-                        )?;
-
-                        Ok(generator.bool_to_i8(ctx, val.into_int_value()).into())
-                    },
-                )?;
-
-                Ok(Some(res.as_base_value().into()))
-            };
+            return Ok(Some(result_ndarray.instance.value.into()));
         }
     }
 
@@ -2266,338 +2207,6 @@ pub fn gen_cmpop_expr<'ctx, G: CodeGenerator>(
     )
 }
 
-/// Generates code for a subscript expression on an `ndarray`.
-///
-/// * `ty` - The `Type` of the `NDArray` elements.
-/// * `ndims` - The `Type` of the `NDArray` number-of-dimensions `Literal`.
-/// * `v` - The `NDArray` value.
-/// * `slice` - The slice expression used to subscript into the `ndarray`.
-fn gen_ndarray_subscript_expr<'ctx, G: CodeGenerator>(
-    generator: &mut G,
-    ctx: &mut CodeGenContext<'ctx, '_>,
-    ty: Type,
-    ndims: Type,
-    v: NDArrayValue<'ctx>,
-    slice: &Expr<Option<Type>>,
-) -> Result<Option<ValueEnum<'ctx>>, String> {
-    let llvm_i1 = ctx.ctx.bool_type();
-    let llvm_i32 = ctx.ctx.i32_type();
-    let llvm_usize = generator.get_size_type(ctx.ctx);
-
-    let TypeEnum::TLiteral { values, .. } = &*ctx.unifier.get_ty_immutable(ndims) else {
-        unreachable!()
-    };
-
-    let ndims = values
-        .iter()
-        .map(|ndim| u64::try_from(ndim.clone()).map_err(|()| ndim.clone()))
-        .collect::<Result<Vec<_>, _>>()
-        .map_err(|val| {
-            format!(
-                "Expected non-negative literal for ndarray.ndims, got {}",
-                i128::try_from(val).unwrap()
-            )
-        })?;
-
-    assert!(!ndims.is_empty());
-
-    // The number of dimensions subscripted by the index expression.
-    // Slicing a ndarray will yield the same number of dimensions, whereas indexing into a
-    // dimension will remove a dimension.
-    let subscripted_dims = match &slice.node {
-        ExprKind::Tuple { elts, .. } => elts.iter().fold(0, |acc, value_subexpr| {
-            if let ExprKind::Slice { .. } = &value_subexpr.node {
-                acc
-            } else {
-                acc + 1
-            }
-        }),
-
-        ExprKind::Slice { .. } => 0,
-        _ => 1,
-    };
-
-    let ndarray_ndims_ty = ctx.unifier.get_fresh_literal(
-        ndims.iter().map(|v| SymbolValue::U64(v - subscripted_dims)).collect(),
-        None,
-    );
-    let ndarray_ty =
-        make_ndarray_ty(&mut ctx.unifier, &ctx.primitives, Some(ty), Some(ndarray_ndims_ty));
-    let llvm_pndarray_t = ctx.get_llvm_type(generator, ndarray_ty).into_pointer_type();
-    let llvm_ndarray_t = llvm_pndarray_t.get_element_type().into_struct_type();
-    let llvm_ndarray_data_t = ctx.get_llvm_type(generator, ty).as_basic_type_enum();
-    let sizeof_elem = llvm_ndarray_data_t.size_of().unwrap();
-
-    // Check that len is non-zero
-    let len = v.load_ndims(ctx);
-    ctx.make_assert(
-        generator,
-        ctx.builder.build_int_compare(IntPredicate::SGT, len, llvm_usize.const_zero(), "").unwrap(),
-        "0:IndexError",
-        "too many indices for array: array is {0}-dimensional but 1 were indexed",
-        [Some(len), None, None],
-        slice.location,
-    );
-
-    // Normalizes a possibly-negative index to its corresponding positive index
-    let normalize_index = |generator: &mut G,
-                           ctx: &mut CodeGenContext<'ctx, '_>,
-                           index: IntValue<'ctx>,
-                           dim: u64| {
-        gen_if_else_expr_callback(
-            generator,
-            ctx,
-            |_, ctx| {
-                Ok(ctx
-                    .builder
-                    .build_int_compare(IntPredicate::SGE, index, index.get_type().const_zero(), "")
-                    .unwrap())
-            },
-            |_, _| Ok(Some(index)),
-            |generator, ctx| {
-                let llvm_i32 = ctx.ctx.i32_type();
-
-                let len = unsafe {
-                    v.dim_sizes().get_typed_unchecked(
-                        ctx,
-                        generator,
-                        &llvm_usize.const_int(dim, true),
-                        None,
-                    )
-                };
-
-                let index = ctx
-                    .builder
-                    .build_int_add(
-                        len,
-                        ctx.builder.build_int_s_extend(index, llvm_usize, "").unwrap(),
-                        "",
-                    )
-                    .unwrap();
-
-                Ok(Some(ctx.builder.build_int_truncate(index, llvm_i32, "").unwrap()))
-            },
-        )
-        .map(|v| v.map(BasicValueEnum::into_int_value))
-    };
-
-    // Converts a slice expression into a slice-range tuple
-    let expr_to_slice = |generator: &mut G,
-                         ctx: &mut CodeGenContext<'ctx, '_>,
-                         node: &ExprKind<Option<Type>>,
-                         dim: u64| {
-        match node {
-            ExprKind::Constant { value: Constant::Int(v), .. } => {
-                let Some(index) =
-                    normalize_index(generator, ctx, llvm_i32.const_int(*v as u64, true), dim)?
-                else {
-                    return Ok(None);
-                };
-
-                Ok(Some((index, index, llvm_i32.const_int(1, true))))
-            }
-
-            ExprKind::Slice { lower, upper, step } => {
-                let dim_sz = unsafe {
-                    v.dim_sizes().get_typed_unchecked(
-                        ctx,
-                        generator,
-                        &llvm_usize.const_int(dim, false),
-                        None,
-                    )
-                };
-
-                handle_slice_indices(lower, upper, step, ctx, generator, dim_sz)
-            }
-
-            _ => {
-                let Some(index) = generator.gen_expr(ctx, slice)? else { return Ok(None) };
-                let index = index
-                    .to_basic_value_enum(ctx, generator, slice.custom.unwrap())?
-                    .into_int_value();
-                let Some(index) = normalize_index(generator, ctx, index, dim)? else {
-                    return Ok(None);
-                };
-
-                Ok(Some((index, index, llvm_i32.const_int(1, true))))
-            }
-        }
-    };
-
-    let make_indices_arr = |generator: &mut G,
-                            ctx: &mut CodeGenContext<'ctx, '_>|
-     -> Result<_, String> {
-        Ok(if let ExprKind::Tuple { elts, .. } = &slice.node {
-            let llvm_int_ty = ctx.get_llvm_type(generator, elts[0].custom.unwrap());
-            let index_addr = generator.gen_array_var_alloc(
-                ctx,
-                llvm_int_ty,
-                llvm_usize.const_int(elts.len() as u64, false),
-                None,
-            )?;
-
-            for (i, elt) in elts.iter().enumerate() {
-                let Some(index) = generator.gen_expr(ctx, elt)? else {
-                    return Ok(None);
-                };
-
-                let index = index
-                    .to_basic_value_enum(ctx, generator, elt.custom.unwrap())?
-                    .into_int_value();
-                let Some(index) = normalize_index(generator, ctx, index, 0)? else {
-                    return Ok(None);
-                };
-
-                let store_ptr = unsafe {
-                    index_addr.ptr_offset_unchecked(
-                        ctx,
-                        generator,
-                        &llvm_usize.const_int(i as u64, false),
-                        None,
-                    )
-                };
-                ctx.builder.build_store(store_ptr, index).unwrap();
-            }
-
-            Some(index_addr)
-        } else if let Some(index) = generator.gen_expr(ctx, slice)? {
-            let llvm_int_ty = ctx.get_llvm_type(generator, slice.custom.unwrap());
-            let index_addr = generator.gen_array_var_alloc(
-                ctx,
-                llvm_int_ty,
-                llvm_usize.const_int(1u64, false),
-                None,
-            )?;
-
-            let index =
-                index.to_basic_value_enum(ctx, generator, slice.custom.unwrap())?.into_int_value();
-            let Some(index) = normalize_index(generator, ctx, index, 0)? else { return Ok(None) };
-
-            let store_ptr = unsafe {
-                index_addr.ptr_offset_unchecked(ctx, generator, &llvm_usize.const_zero(), None)
-            };
-            ctx.builder.build_store(store_ptr, index).unwrap();
-
-            Some(index_addr)
-        } else {
-            None
-        })
-    };
-
-    Ok(Some(if ndims.len() == 1 && ndims[0] - subscripted_dims == 0 {
-        let Some(index_addr) = make_indices_arr(generator, ctx)? else { return Ok(None) };
-
-        v.data().get(ctx, generator, &index_addr, None).into()
-    } else {
-        match &slice.node {
-            ExprKind::Tuple { elts, .. } => {
-                let slices = elts
-                    .iter()
-                    .enumerate()
-                    .map(|(dim, elt)| expr_to_slice(generator, ctx, &elt.node, dim as u64))
-                    .take_while_inclusive(|slice| slice.as_ref().is_ok_and(Option::is_some))
-                    .collect::<Result<Vec<_>, _>>()?;
-                if slices.len() < elts.len() {
-                    return Ok(None);
-                }
-
-                let slices = slices.into_iter().map(Option::unwrap).collect_vec();
-
-                numpy::ndarray_sliced_copy(generator, ctx, ty, v, &slices)?.as_base_value().into()
-            }
-
-            ExprKind::Slice { .. } => {
-                let Some(slice) = expr_to_slice(generator, ctx, &slice.node, 0)? else {
-                    return Ok(None);
-                };
-
-                numpy::ndarray_sliced_copy(generator, ctx, ty, v, &[slice])?.as_base_value().into()
-            }
-
-            _ => {
-                // Accessing an element from a multi-dimensional `ndarray`
-
-                let Some(index_addr) = make_indices_arr(generator, ctx)? else { return Ok(None) };
-
-                // Create a new array, remove the top dimension from the dimension-size-list, and copy the
-                // elements over
-                let subscripted_ndarray =
-                    generator.gen_var_alloc(ctx, llvm_ndarray_t.into(), None)?;
-                let ndarray = NDArrayValue::from_ptr_val(subscripted_ndarray, llvm_usize, None);
-
-                let num_dims = v.load_ndims(ctx);
-                ndarray.store_ndims(
-                    ctx,
-                    generator,
-                    ctx.builder
-                        .build_int_sub(num_dims, llvm_usize.const_int(1, false), "")
-                        .unwrap(),
-                );
-
-                let ndarray_num_dims = ndarray.load_ndims(ctx);
-                ndarray.create_dim_sizes(ctx, llvm_usize, ndarray_num_dims);
-
-                let ndarray_num_dims = ctx
-                    .builder
-                    .build_int_z_extend_or_bit_cast(
-                        ndarray.load_ndims(ctx),
-                        llvm_usize.size_of().get_type(),
-                        "",
-                    )
-                    .unwrap();
-                let v_dims_src_ptr = unsafe {
-                    v.dim_sizes().ptr_offset_unchecked(
-                        ctx,
-                        generator,
-                        &llvm_usize.const_int(1, false),
-                        None,
-                    )
-                };
-                call_memcpy_generic(
-                    ctx,
-                    ndarray.dim_sizes().base_ptr(ctx, generator),
-                    v_dims_src_ptr,
-                    ctx.builder
-                        .build_int_mul(ndarray_num_dims, llvm_usize.size_of(), "")
-                        .map(Into::into)
-                        .unwrap(),
-                    llvm_i1.const_zero(),
-                );
-
-                let ndarray_num_elems = call_ndarray_calc_size(
-                    generator,
-                    ctx,
-                    &ndarray.dim_sizes().as_slice_value(ctx, generator),
-                    (None, None),
-                );
-                let ndarray_num_elems = ctx
-                    .builder
-                    .build_int_z_extend_or_bit_cast(ndarray_num_elems, sizeof_elem.get_type(), "")
-                    .unwrap();
-                ndarray.create_data(ctx, llvm_ndarray_data_t, ndarray_num_elems);
-
-                let v_data_src_ptr = v.data().ptr_offset(ctx, generator, &index_addr, None);
-                call_memcpy_generic(
-                    ctx,
-                    ndarray.data().base_ptr(ctx, generator),
-                    v_data_src_ptr,
-                    ctx.builder
-                        .build_int_mul(
-                            ndarray_num_elems,
-                            llvm_ndarray_data_t.size_of().unwrap(),
-                            "",
-                        )
-                        .map(Into::into)
-                        .unwrap(),
-                    llvm_i1.const_zero(),
-                );
-
-                ndarray.as_base_value().into()
-            }
-        }
-    }))
-}
-
 /// See [`CodeGenerator::gen_expr`].
 pub fn gen_expr<'ctx, G: CodeGenerator>(
     generator: &mut G,
@@ -3237,18 +2846,26 @@ pub fn gen_expr<'ctx, G: CodeGenerator>(
                         v.data().get(ctx, generator, &index, None).into()
                     }
                 }
-                TypeEnum::TObj { obj_id, params, .. } if *obj_id == PrimDef::NDArray.id() => {
-                    let (ty, ndims) = params.iter().map(|(_, ty)| ty).collect_tuple().unwrap();
-
-                    let v = if let Some(v) = generator.gen_expr(ctx, value)? {
-                        v.to_basic_value_enum(ctx, generator, value.custom.unwrap())?
-                            .into_pointer_value()
-                    } else {
+                TypeEnum::TObj { obj_id, .. } if *obj_id == PrimDef::NDArray.id() => {
+                    let Some(ndarray) = generator.gen_expr(ctx, value)? else {
                         return Ok(None);
                     };
-                    let v = NDArrayValue::from_ptr_val(v, usize, None);
 
-                    return gen_ndarray_subscript_expr(generator, ctx, *ty, *ndims, v, slice);
+                    let ndarray_ty = value.custom.unwrap();
+                    let ndarray = ndarray.to_basic_value_enum(ctx, generator, ndarray_ty)?;
+
+                    let ndarray = NDArrayObject::from_object(
+                        generator,
+                        ctx,
+                        AnyObject { ty: ndarray_ty, value: ndarray },
+                    );
+
+                    let indices = gen_ndarray_subscript_ndindices(generator, ctx, slice)?;
+                    let result = ndarray
+                        .index(generator, ctx, &indices)
+                        .split_unsized(generator, ctx)
+                        .to_basic_value_enum();
+                    return Ok(Some(ValueEnum::Dynamic(result)));
                 }
                 TypeEnum::TTuple { .. } => {
                     let index: u32 =
@@ -3291,3 +2908,42 @@ pub fn gen_expr<'ctx, G: CodeGenerator>(
         _ => unimplemented!(),
     }))
 }
+
+/// Generate LLVM IR for an [`ExprKind::Slice`]
+#[allow(clippy::type_complexity)]
+pub fn gen_slice<'ctx, G: CodeGenerator>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    lower: &Option<Box<Expr<Option<Type>>>>,
+    upper: &Option<Box<Expr<Option<Type>>>>,
+    step: &Option<Box<Expr<Option<Type>>>>,
+) -> Result<
+    (
+        Option<Instance<'ctx, Int<Int32>>>,
+        Option<Instance<'ctx, Int<Int32>>>,
+        Option<Instance<'ctx, Int<Int32>>>,
+    ),
+    String,
+> {
+    let mut help = |value_expr: &Option<Box<Expr<Option<Type>>>>| -> Result<_, String> {
+        Ok(match value_expr {
+            None => None,
+            Some(value_expr) => {
+                let value_expr = generator
+                    .gen_expr(ctx, value_expr)?
+                    .unwrap()
+                    .to_basic_value_enum(ctx, generator, ctx.primitives.int32)?;
+
+                let value_expr = Int(Int32).check_value(generator, ctx.ctx, value_expr).unwrap();
+
+                Some(value_expr)
+            }
+        })
+    };
+
+    let lower = help(lower)?;
+    let upper = help(upper)?;
+    let step = help(step)?;
+
+    Ok((lower, upper, step))
+}

nac3core/src/codegen/irrt/mod.rs

@@ -5,10 +5,17 @@ use super::{
         ArrayLikeIndexer, ArrayLikeValue, ArraySliceValue, ListValue, NDArrayValue,
         TypedArrayLikeAdapter, UntypedArrayLikeAccessor,
     },
-    llvm_intrinsics, CodeGenContext, CodeGenerator,
+    llvm_intrinsics,
+    model::*,
+    object::{
+        list::List,
+        ndarray::{broadcast::ShapeEntry, indexing::NDIndex, nditer::NDIter, NDArray},
+    },
+    CodeGenContext, CodeGenerator,
 };
 use crate::codegen::classes::TypedArrayLikeAccessor;
 use crate::codegen::stmt::gen_for_callback_incrementing;
+use function::CallFunction;
 use inkwell::{
     attributes::{Attribute, AttributeLoc},
     context::Context,
@@ -973,3 +980,280 @@ pub fn setup_irrt_exceptions<'ctx>(
         global.set_initializer(&exn_id);
     }
 }
+
+pub fn call_nac3_ndarray_util_assert_shape_no_negative<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndims: Instance<'ctx, Int<SizeT>>,
+    shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    let name = get_sizet_dependent_function_name(
+        generator,
+        ctx,
+        "__nac3_ndarray_util_assert_shape_no_negative",
+    );
+    CallFunction::begin(generator, ctx, &name).arg(ndims).arg(shape).returning_void();
+}
+
+pub fn call_nac3_ndarray_util_assert_output_shape_same<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray_ndims: Instance<'ctx, Int<SizeT>>,
+    ndarray_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    output_ndims: Instance<'ctx, Int<SizeT>>,
+    output_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    let name = get_sizet_dependent_function_name(
+        generator,
+        ctx,
+        "__nac3_ndarray_util_assert_output_shape_same",
+    );
+    CallFunction::begin(generator, ctx, &name)
+        .arg(ndarray_ndims)
+        .arg(ndarray_shape)
+        .arg(output_ndims)
+        .arg(output_shape)
+        .returning_void();
+}
+
+pub fn call_nac3_ndarray_size<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) -> Instance<'ctx, Int<SizeT>> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_size");
+    CallFunction::begin(generator, ctx, &name).arg(ndarray).returning_auto("size")
+}
+
+pub fn call_nac3_ndarray_nbytes<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) -> Instance<'ctx, Int<SizeT>> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_nbytes");
+    CallFunction::begin(generator, ctx, &name).arg(ndarray).returning_auto("nbytes")
+}
+
+pub fn call_nac3_ndarray_len<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) -> Instance<'ctx, Int<SizeT>> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_len");
+    CallFunction::begin(generator, ctx, &name).arg(ndarray).returning_auto("len")
+}
+
+pub fn call_nac3_ndarray_is_c_contiguous<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) -> Instance<'ctx, Int<Bool>> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_is_c_contiguous");
+    CallFunction::begin(generator, ctx, &name).arg(ndarray).returning_auto("is_c_contiguous")
+}
+
+pub fn call_nac3_ndarray_get_nth_pelement<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    index: Instance<'ctx, Int<SizeT>>,
+) -> Instance<'ctx, Ptr<Int<Byte>>> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_get_nth_pelement");
+    CallFunction::begin(generator, ctx, &name).arg(ndarray).arg(index).returning_auto("pelement")
+}
+
+pub fn call_nac3_ndarray_get_pelement_by_indices<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    indices: Instance<'ctx, Ptr<Int<SizeT>>>,
+) -> Instance<'ctx, Ptr<Int<Byte>>> {
+    let name =
+        get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_get_pelement_by_indices");
+    CallFunction::begin(generator, ctx, &name).arg(ndarray).arg(indices).returning_auto("pelement")
+}
+
+pub fn call_nac3_ndarray_set_strides_by_shape<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) {
+    let name =
+        get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_set_strides_by_shape");
+    CallFunction::begin(generator, ctx, &name).arg(ndarray).returning_void();
+}
+
+pub fn call_nac3_ndarray_copy_data<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    src_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    dst_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_copy_data");
+    CallFunction::begin(generator, ctx, &name).arg(src_ndarray).arg(dst_ndarray).returning_void();
+}
+
+pub fn call_nac3_nditer_initialize<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    iter: Instance<'ctx, Ptr<Struct<NDIter>>>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    indices: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_nditer_initialize");
+    CallFunction::begin(generator, ctx, &name).arg(iter).arg(ndarray).arg(indices).returning_void();
+}
+
+pub fn call_nac3_nditer_has_next<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    iter: Instance<'ctx, Ptr<Struct<NDIter>>>,
+) -> Instance<'ctx, Int<Bool>> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_nditer_has_next");
+    CallFunction::begin(generator, ctx, &name).arg(iter).returning_auto("has_next")
+}
+
+pub fn call_nac3_nditer_next<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    iter: Instance<'ctx, Ptr<Struct<NDIter>>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_nditer_next");
+    CallFunction::begin(generator, ctx, &name).arg(iter).returning_void();
+}
+
+pub fn call_nac3_ndarray_index<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    num_indices: Instance<'ctx, Int<SizeT>>,
+    indices: Instance<'ctx, Ptr<Struct<NDIndex>>>,
+    src_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    dst_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_index");
+    CallFunction::begin(generator, ctx, &name)
+        .arg(num_indices)
+        .arg(indices)
+        .arg(src_ndarray)
+        .arg(dst_ndarray)
+        .returning_void();
+}
+
+pub fn call_nac3_ndarray_array_set_and_validate_list_shape<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    list: Instance<'ctx, Ptr<Struct<List<Int<Byte>>>>>,
+    ndims: Instance<'ctx, Int<SizeT>>,
+    shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    let name = get_sizet_dependent_function_name(
+        generator,
+        ctx,
+        "__nac3_ndarray_array_set_and_validate_list_shape",
+    );
+    CallFunction::begin(generator, ctx, &name).arg(list).arg(ndims).arg(shape).returning_void();
+}
+
+pub fn call_nac3_ndarray_array_write_list_to_array<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    list: Instance<'ctx, Ptr<Struct<List<Int<Byte>>>>>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) {
+    let name = get_sizet_dependent_function_name(
+        generator,
+        ctx,
+        "__nac3_ndarray_array_write_list_to_array",
+    );
+    CallFunction::begin(generator, ctx, &name).arg(list).arg(ndarray).returning_void();
+}
+
+pub fn call_nac3_ndarray_reshape_resolve_and_check_new_shape<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    size: Instance<'ctx, Int<SizeT>>,
+    new_ndims: Instance<'ctx, Int<SizeT>>,
+    new_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    let name = get_sizet_dependent_function_name(
+        generator,
+        ctx,
+        "__nac3_ndarray_reshape_resolve_and_check_new_shape",
+    );
+    CallFunction::begin(generator, ctx, &name)
+        .arg(size)
+        .arg(new_ndims)
+        .arg(new_shape)
+        .returning_void();
+}
+
+pub fn call_nac3_ndarray_broadcast_to<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    src_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    dst_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_broadcast_to");
+    CallFunction::begin(generator, ctx, &name).arg(src_ndarray).arg(dst_ndarray).returning_void();
+}
+
+pub fn call_nac3_ndarray_broadcast_shapes<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    num_shape_entries: Instance<'ctx, Int<SizeT>>,
+    shape_entries: Instance<'ctx, Ptr<Struct<ShapeEntry>>>,
+    dst_ndims: Instance<'ctx, Int<SizeT>>,
+    dst_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_broadcast_shapes");
+    CallFunction::begin(generator, ctx, &name)
+        .arg(num_shape_entries)
+        .arg(shape_entries)
+        .arg(dst_ndims)
+        .arg(dst_shape)
+        .returning_void();
+}
+
+pub fn call_nac3_ndarray_transpose<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    src_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    dst_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    num_axes: Instance<'ctx, Int<SizeT>>,
+    axes: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_transpose");
+    CallFunction::begin(generator, ctx, &name)
+        .arg(src_ndarray)
+        .arg(dst_ndarray)
+        .arg(num_axes)
+        .arg(axes)
+        .returning_void();
+}
+
+#[allow(clippy::too_many_arguments)]
+pub fn call_nac3_ndarray_matmul_calculate_shapes<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    a_ndims: Instance<'ctx, Int<SizeT>>,
+    a_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    b_ndims: Instance<'ctx, Int<SizeT>>,
+    b_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    final_ndims: Instance<'ctx, Int<SizeT>>,
+    new_a_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    new_b_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    dst_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    let name =
+        get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_matmul_calculate_shapes");
+    CallFunction::begin(generator, ctx, &name)
+        .arg(a_ndims)
+        .arg(a_shape)
+        .arg(b_ndims)
+        .arg(b_shape)
+        .arg(final_ndims)
+        .arg(new_a_shape)
+        .arg(new_b_shape)
+        .arg(dst_shape)
+        .returning_void();
+}

nac3core/src/codegen/mod.rs

@@ -1,7 +1,7 @@
 use crate::{
-    codegen::classes::{ListType, NDArrayType, ProxyType, RangeType},
+    codegen::classes::{ListType, ProxyType, RangeType},
     symbol_resolver::{StaticValue, SymbolResolver},
-    toplevel::{helper::PrimDef, numpy::unpack_ndarray_var_tys, TopLevelContext, TopLevelDef},
+    toplevel::{helper::PrimDef, TopLevelContext, TopLevelDef},
     typecheck::{
         type_inferencer::{CodeLocation, PrimitiveStore},
         typedef::{CallId, FuncArg, Type, TypeEnum, Unifier},
@@ -24,7 +24,9 @@ use inkwell::{
     AddressSpace, IntPredicate, OptimizationLevel,
 };
 use itertools::Itertools;
+use model::*;
 use nac3parser::ast::{Location, Stmt, StrRef};
+use object::ndarray::NDArray;
 use parking_lot::{Condvar, Mutex};
 use std::collections::{HashMap, HashSet};
 use std::sync::{
@@ -491,12 +493,7 @@ fn get_llvm_type<'ctx, G: CodeGenerator + ?Sized>(
                         }
 
                         TObj { obj_id, .. } if *obj_id == PrimDef::NDArray.id() => {
-                            let (dtype, _) = unpack_ndarray_var_tys(unifier, ty);
-                            let element_type = get_llvm_type(
-                                ctx, module, generator, unifier, top_level, type_cache, dtype,
-                            );
-
-                            NDArrayType::new(generator, ctx, element_type).as_base_type().into()
+                            Ptr(Struct(NDArray)).get_type(generator, ctx).as_basic_type_enum()
                         }
 
                         _ => unreachable!(

nac3core/src/codegen/model/mod.rs

@@ -6,6 +6,7 @@ pub mod function;
 mod int;
 mod ptr;
 mod structure;
+pub mod util;
 
 pub use any::*;
 pub use array::*;

nac3core/src/codegen/model/util.rs

@@ -0,0 +1,40 @@
+use crate::codegen::{
+    stmt::{gen_for_callback_incrementing, BreakContinueHooks},
+    CodeGenContext, CodeGenerator,
+};
+
+use super::*;
+
+/// Like [`gen_for_callback_incrementing`] with [`Model`] abstractions.
+pub fn gen_for_model<'ctx, 'a, G, F, N>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, 'a>,
+    start: Instance<'ctx, Int<N>>,
+    stop: Instance<'ctx, Int<N>>,
+    step: Instance<'ctx, Int<N>>,
+    body: F,
+) -> Result<(), String>
+where
+    G: CodeGenerator + ?Sized,
+    F: FnOnce(
+        &mut G,
+        &mut CodeGenContext<'ctx, 'a>,
+        BreakContinueHooks<'ctx>,
+        Instance<'ctx, Int<N>>,
+    ) -> Result<(), String>,
+    N: IntKind<'ctx> + Default,
+{
+    let int_model = Int(N::default());
+    gen_for_callback_incrementing(
+        generator,
+        ctx,
+        None,
+        start.value,
+        (stop.value, false),
+        |g, ctx, hooks, i| {
+            let i = int_model.believe_value(i);
+            body(g, ctx, hooks, i)
+        },
+        step.value,
+    )
+}

nac3core/src/codegen/object/list.rs

@@ -0,0 +1,112 @@
+use crate::{
+    codegen::{model::*, CodeGenContext, CodeGenerator},
+    typecheck::typedef::{iter_type_vars, Type, TypeEnum},
+};
+
+use super::any::AnyObject;
+
+/// Fields of [`List`]
+pub struct ListFields<'ctx, F: FieldTraversal<'ctx>, Item: Model<'ctx>> {
+    /// Array pointer to content
+    pub items: F::Out<Ptr<Item>>,
+    /// Number of items in the array
+    pub len: F::Out<Int<SizeT>>,
+}
+
+/// A list in NAC3.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct List<Item> {
+    /// Model of the list items
+    pub item: Item,
+}
+
+impl<'ctx, Item: Model<'ctx>> StructKind<'ctx> for List<Item> {
+    type Fields<F: FieldTraversal<'ctx>> = ListFields<'ctx, F, Item>;
+
+    fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+        Self::Fields {
+            items: traversal.add("data", Ptr(self.item)),
+            len: traversal.add_auto("len"),
+        }
+    }
+}
+
+/// A NAC3 Python List object.
+#[derive(Debug, Clone, Copy)]
+pub struct ListObject<'ctx> {
+    /// Typechecker type of the list items
+    pub item_type: Type,
+    pub instance: Instance<'ctx, Ptr<Struct<List<Any<'ctx>>>>>,
+}
+
+impl<'ctx> ListObject<'ctx> {
+    /// Create a [`ListObject`] from an LLVM value and its typechecker [`Type`].
+    pub fn from_object<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        object: AnyObject<'ctx>,
+    ) -> Self {
+        // Check typechecker type and extract `item_type`
+        let item_type = match &*ctx.unifier.get_ty(object.ty) {
+            TypeEnum::TObj { obj_id, params, .. }
+                if *obj_id == ctx.primitives.list.obj_id(&ctx.unifier).unwrap() =>
+            {
+                iter_type_vars(params).next().unwrap().ty // Extract `item_type`
+            }
+            _ => {
+                panic!("Expecting type to be a list, but got {}", ctx.unifier.stringify(object.ty))
+            }
+        };
+
+        let plist = Ptr(Struct(List { item: Any(ctx.get_llvm_type(generator, item_type)) }));
+
+        // Create object
+        let value = plist.check_value(generator, ctx.ctx, object.value).unwrap();
+        ListObject { item_type, instance: value }
+    }
+
+    /// Check if an object can be converted into an ndarray.
+    pub fn is_instance(ctx: &mut CodeGenContext<'ctx, '_>, object: AnyObject<'ctx>) -> bool {
+        matches!(&*ctx.unifier.get_ty(object.ty), TypeEnum::TObj { obj_id, .. } if *obj_id == ctx.primitives.list.obj_id(&ctx.unifier).unwrap())
+    }
+
+    /// Get the `len()` of this list.
+    pub fn len<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<SizeT>> {
+        self.instance.get(generator, ctx, |f| f.len)
+    }
+
+    /// Get the `items` field as an opaque pointer.
+    pub fn get_opaque_items_ptr<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Ptr<Int<Byte>>> {
+        self.instance.get(generator, ctx, |f| f.items).pointer_cast(generator, ctx, Int(Byte))
+    }
+
+    /// Get the value of this [`ListObject`] as a list with opaque items.
+    ///
+    /// This function allocates on the stack to create the list, but the
+    /// reference to the `items` are preserved.
+    pub fn get_opaque_list_ptr<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Ptr<Struct<List<Int<Byte>>>>> {
+        let opaque_list = Struct(List { item: Int(Byte) }).alloca(generator, ctx);
+
+        // Copy items pointer
+        let items = self.get_opaque_items_ptr(generator, ctx);
+        opaque_list.set(ctx, |f| f.items, items);
+
+        // Copy len
+        let len = self.instance.get(generator, ctx, |f| f.len);
+        opaque_list.set(ctx, |f| f.len, len);
+
+        opaque_list
+    }
+}

nac3core/src/codegen/object/mod.rs

@@ -1 +1,4 @@
 pub mod any;
+pub mod list;
+pub mod ndarray;
+pub mod tuple;

nac3core/src/codegen/object/ndarray/array.rs

@@ -0,0 +1,178 @@
+use super::NDArrayObject;
+use crate::{
+    codegen::{
+        irrt::{
+            call_nac3_ndarray_array_set_and_validate_list_shape,
+            call_nac3_ndarray_array_write_list_to_array,
+        },
+        model::*,
+        object::{any::AnyObject, list::ListObject},
+        stmt::gen_if_else_expr_callback,
+        CodeGenContext, CodeGenerator,
+    },
+    toplevel::helper::{arraylike_flatten_element_type, arraylike_get_ndims},
+    typecheck::typedef::{Type, TypeEnum},
+};
+
+fn get_list_object_dtype_and_ndims<'ctx>(
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    list: ListObject<'ctx>,
+) -> (Type, u64) {
+    let dtype = arraylike_flatten_element_type(&mut ctx.unifier, list.item_type);
+
+    let ndims = arraylike_get_ndims(&mut ctx.unifier, list.item_type);
+    let ndims = ndims + 1; // To count `list` itself.
+
+    (dtype, ndims)
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    fn make_np_array_list_copy_impl<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        list: ListObject<'ctx>,
+    ) -> Self {
+        let (dtype, ndims_int) = get_list_object_dtype_and_ndims(ctx, list);
+        let list_value = list.get_opaque_list_ptr(generator, ctx);
+
+        // Validate `list` has a consistent shape.
+        // Raise an exception if `list` is something abnormal like `[[1, 2], [3]]`.
+        // If `list` has a consistent shape, deduce the shape and write it to `shape`.
+        let ndims = Int(SizeT).const_int(generator, ctx.ctx, ndims_int);
+        let shape = Int(SizeT).array_alloca(generator, ctx, ndims.value);
+        call_nac3_ndarray_array_set_and_validate_list_shape(
+            generator, ctx, list_value, ndims, shape,
+        );
+
+        let ndarray = NDArrayObject::alloca(generator, ctx, dtype, ndims_int);
+        ndarray.copy_shape_from_array(generator, ctx, shape);
+        ndarray.create_data(generator, ctx);
+
+        // Copy all contents from the list.
+        call_nac3_ndarray_array_write_list_to_array(generator, ctx, list_value, ndarray.instance);
+
+        ndarray
+    }
+
+    fn make_np_array_list_try_no_copy_impl<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        list: ListObject<'ctx>,
+    ) -> Self {
+        // np_array without copying is only possible `list` is not nested.
+        //
+        // If `list` is `list[T]`, we can create an ndarray with `data` set
+        // to the array pointer of `list`.
+        //
+        // If `list` is `list[list[T]]` or worse, copy.
+
+        let (dtype, ndims) = get_list_object_dtype_and_ndims(ctx, list);
+        if ndims == 1 {
+            // `list` is not nested
+            let ndarray = NDArrayObject::alloca(generator, ctx, dtype, 1);
+
+            // Set data
+            let data = list.get_opaque_items_ptr(generator, ctx);
+            ndarray.instance.set(ctx, |f| f.data, data);
+
+            // ndarray->shape[0] = list->len;
+            let shape = ndarray.instance.get(generator, ctx, |f| f.shape);
+            let list_len = list.instance.get(generator, ctx, |f| f.len);
+            shape.set_index_const(ctx, 0, list_len);
+
+            // Set strides, the `data` is contiguous
+            ndarray.set_strides_contiguous(generator, ctx);
+
+            ndarray
+        } else {
+            // `list` is nested, copy
+            NDArrayObject::make_np_array_list_copy_impl(generator, ctx, list)
+        }
+    }
+
+    fn make_np_array_list_impl<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        list: ListObject<'ctx>,
+        copy: Instance<'ctx, Int<Bool>>,
+    ) -> Self {
+        let (dtype, ndims) = get_list_object_dtype_and_ndims(ctx, list);
+
+        let ndarray = gen_if_else_expr_callback(
+            generator,
+            ctx,
+            |_generator, _ctx| Ok(copy.value),
+            |generator, ctx| {
+                let ndarray = NDArrayObject::make_np_array_list_copy_impl(generator, ctx, list);
+                Ok(Some(ndarray.instance.value))
+            },
+            |generator, ctx| {
+                let ndarray =
+                    NDArrayObject::make_np_array_list_try_no_copy_impl(generator, ctx, list);
+                Ok(Some(ndarray.instance.value))
+            },
+        )
+        .unwrap()
+        .unwrap();
+
+        NDArrayObject::from_value_and_unpacked_types(generator, ctx, ndarray, dtype, ndims)
+    }
+
+    pub fn make_np_array_ndarray_impl<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        ndarray: NDArrayObject<'ctx>,
+        copy: Instance<'ctx, Int<Bool>>,
+    ) -> Self {
+        let ndarray_val = gen_if_else_expr_callback(
+            generator,
+            ctx,
+            |_generator, _ctx| Ok(copy.value),
+            |generator, ctx| {
+                let ndarray = ndarray.make_copy(generator, ctx); // Force copy
+                Ok(Some(ndarray.instance.value))
+            },
+            |_generator, _ctx| {
+                // No need to copy. Return `ndarray` itself.
+                Ok(Some(ndarray.instance.value))
+            },
+        )
+        .unwrap()
+        .unwrap();
+
+        NDArrayObject::from_value_and_unpacked_types(
+            generator,
+            ctx,
+            ndarray_val,
+            ndarray.dtype,
+            ndarray.ndims,
+        )
+    }
+
+    /// Create a new ndarray like `np.array()`.
+    ///
+    /// NOTE: The `ndmin` argument is not here. You may want to
+    /// do [`NDArrayObject::atleast_nd`] to achieve that.
+    pub fn make_np_array<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        object: AnyObject<'ctx>,
+        copy: Instance<'ctx, Int<Bool>>,
+    ) -> Self {
+        match &*ctx.unifier.get_ty(object.ty) {
+            TypeEnum::TObj { obj_id, .. }
+                if *obj_id == ctx.primitives.list.obj_id(&ctx.unifier).unwrap() =>
+            {
+                let list = ListObject::from_object(generator, ctx, object);
+                NDArrayObject::make_np_array_list_impl(generator, ctx, list, copy)
+            }
+            TypeEnum::TObj { obj_id, .. }
+                if *obj_id == ctx.primitives.ndarray.obj_id(&ctx.unifier).unwrap() =>
+            {
+                let ndarray = NDArrayObject::from_object(generator, ctx, object);
+                NDArrayObject::make_np_array_ndarray_impl(generator, ctx, ndarray, copy)
+            }
+            _ => panic!("Unrecognized object type: {}", ctx.unifier.stringify(object.ty)), // Typechecker ensures this
+        }
+    }
+}

nac3core/src/codegen/object/ndarray/broadcast.rs

@@ -0,0 +1,135 @@
+use itertools::Itertools;
+
+use crate::codegen::{
+    irrt::{call_nac3_ndarray_broadcast_shapes, call_nac3_ndarray_broadcast_to},
+    model::*,
+    CodeGenContext, CodeGenerator,
+};
+
+use super::NDArrayObject;
+
+/// Fields of [`ShapeEntry`]
+pub struct ShapeEntryFields<'ctx, F: FieldTraversal<'ctx>> {
+    pub ndims: F::Out<Int<SizeT>>,
+    pub shape: F::Out<Ptr<Int<SizeT>>>,
+}
+
+/// An IRRT structure used in broadcasting.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct ShapeEntry;
+
+impl<'ctx> StructKind<'ctx> for ShapeEntry {
+    type Fields<F: FieldTraversal<'ctx>> = ShapeEntryFields<'ctx, F>;
+
+    fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+        Self::Fields { ndims: traversal.add_auto("ndims"), shape: traversal.add_auto("shape") }
+    }
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Create a broadcast view on this ndarray with a target shape.
+    ///
+    /// The input shape will be checked to make sure that it contains no negative values.
+    ///
+    /// * `target_ndims` - The ndims type after broadcasting to the given shape.
+    /// The caller has to figure this out for this function.
+    /// * `target_shape` - An array pointer pointing to the target shape.
+    #[must_use]
+    pub fn broadcast_to<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        target_ndims: u64,
+        target_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) -> Self {
+        let broadcast_ndarray = NDArrayObject::alloca(generator, ctx, self.dtype, target_ndims);
+        broadcast_ndarray.copy_shape_from_array(generator, ctx, target_shape);
+
+        call_nac3_ndarray_broadcast_to(generator, ctx, self.instance, broadcast_ndarray.instance);
+        broadcast_ndarray
+    }
+}
+/// A result produced by [`broadcast_all_ndarrays`]
+#[derive(Debug, Clone)]
+pub struct BroadcastAllResult<'ctx> {
+    /// The statically known `ndims` of the broadcast result.
+    pub ndims: u64,
+    /// The broadcasting shape.
+    pub shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    /// Broadcasted views on the inputs.
+    ///
+    /// All of them will have `shape` [`BroadcastAllResult::shape`] and
+    /// `ndims` [`BroadcastAllResult::ndims`]. The length of the vector
+    /// is the same as the input.
+    pub ndarrays: Vec<NDArrayObject<'ctx>>,
+}
+
+/// Helper function to call `call_nac3_ndarray_broadcast_shapes`
+fn broadcast_shapes<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    in_shape_entries: &[(Instance<'ctx, Ptr<Int<SizeT>>>, u64)], // (shape, shape's length/ndims)
+    broadcast_ndims: u64,
+    broadcast_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    // Prepare input shape entries to be passed to `call_nac3_ndarray_broadcast_shapes`.
+    let num_shape_entries =
+        Int(SizeT).const_int(generator, ctx.ctx, u64::try_from(in_shape_entries.len()).unwrap());
+    let shape_entries = Struct(ShapeEntry).array_alloca(generator, ctx, num_shape_entries.value);
+    for (i, (in_shape, in_ndims)) in in_shape_entries.iter().enumerate() {
+        let pshape_entry = shape_entries.offset_const(ctx, i as u64);
+
+        let in_ndims = Int(SizeT).const_int(generator, ctx.ctx, *in_ndims);
+        pshape_entry.set(ctx, |f| f.ndims, in_ndims);
+
+        pshape_entry.set(ctx, |f| f.shape, *in_shape);
+    }
+
+    let broadcast_ndims = Int(SizeT).const_int(generator, ctx.ctx, broadcast_ndims);
+    call_nac3_ndarray_broadcast_shapes(
+        generator,
+        ctx,
+        num_shape_entries,
+        shape_entries,
+        broadcast_ndims,
+        broadcast_shape,
+    );
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Broadcast all ndarrays according to `np.broadcast()` and return a [`BroadcastAllResult`]
+    /// containing all the information of the result of the broadcast operation.
+    pub fn broadcast<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        ndarrays: &[Self],
+    ) -> BroadcastAllResult<'ctx> {
+        assert!(!ndarrays.is_empty());
+
+        // Infer the broadcast output ndims.
+        let broadcast_ndims_int = ndarrays.iter().map(|ndarray| ndarray.ndims).max().unwrap();
+
+        let broadcast_ndims = Int(SizeT).const_int(generator, ctx.ctx, broadcast_ndims_int);
+        let broadcast_shape = Int(SizeT).array_alloca(generator, ctx, broadcast_ndims.value);
+
+        let shape_entries = ndarrays
+            .iter()
+            .map(|ndarray| (ndarray.instance.get(generator, ctx, |f| f.shape), ndarray.ndims))
+            .collect_vec();
+        broadcast_shapes(generator, ctx, &shape_entries, broadcast_ndims_int, broadcast_shape);
+
+        // Broadcast all the inputs to shape `dst_shape`.
+        let broadcast_ndarrays: Vec<_> = ndarrays
+            .iter()
+            .map(|ndarray| {
+                ndarray.broadcast_to(generator, ctx, broadcast_ndims_int, broadcast_shape)
+            })
+            .collect_vec();
+
+        BroadcastAllResult {
+            ndims: broadcast_ndims_int,
+            shape: broadcast_shape,
+            ndarrays: broadcast_ndarrays,
+        }
+    }
+}

nac3core/src/codegen/object/ndarray/contiguous.rs

@@ -0,0 +1,134 @@
+use crate::{
+    codegen::{model::*, CodeGenContext, CodeGenerator},
+    typecheck::typedef::Type,
+};
+
+use super::NDArrayObject;
+
+/// Fields of [`ContiguousNDArray`]
+pub struct ContiguousNDArrayFields<'ctx, F: FieldTraversal<'ctx>, Item: Model<'ctx>> {
+    pub ndims: F::Out<Int<SizeT>>,
+    pub shape: F::Out<Ptr<Int<SizeT>>>,
+    pub data: F::Out<Ptr<Item>>,
+}
+
+/// An ndarray without strides and non-opaque `data` field in NAC3.
+#[derive(Debug, Clone, Copy)]
+pub struct ContiguousNDArray<M> {
+    /// [`Model`] of the items.
+    pub item: M,
+}
+
+impl<'ctx, Item: Model<'ctx>> StructKind<'ctx> for ContiguousNDArray<Item> {
+    type Fields<F: FieldTraversal<'ctx>> = ContiguousNDArrayFields<'ctx, F, Item>;
+
+    fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+        Self::Fields {
+            ndims: traversal.add_auto("ndims"),
+            shape: traversal.add_auto("shape"),
+            data: traversal.add("data", Ptr(self.item)),
+        }
+    }
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Create a [`ContiguousNDArray`] from the contents of this ndarray.
+    ///
+    /// This function may or may not be expensive depending on if this ndarray has contiguous data.
+    ///
+    /// If this ndarray is not C-contiguous, this function will allocate memory on the stack for the `data` field of
+    /// the returned [`ContiguousNDArray`] and copy contents of this ndarray to there.
+    ///
+    /// If this ndarray is C-contiguous, contents of this ndarray will not be copied. The created [`ContiguousNDArray`]
+    /// will share memory with this ndarray.
+    ///
+    /// The `item_model` sets the [`Model`] of the returned [`ContiguousNDArray`]'s `Item` model for type-safety, and
+    /// should match the `ctx.get_llvm_type()` of this ndarray's `dtype`. Otherwise this function panics. Use model [`Any`]
+    /// if you don't care/cannot know the [`Model`] in advance.
+    pub fn make_contiguous_ndarray<G: CodeGenerator + ?Sized, Item: Model<'ctx>>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        item_model: Item,
+    ) -> Instance<'ctx, Ptr<Struct<ContiguousNDArray<Item>>>> {
+        // Sanity check on `self.dtype` and `item_model`.
+        let dtype_llvm = ctx.get_llvm_type(generator, self.dtype);
+        item_model.check_type(generator, ctx.ctx, dtype_llvm).unwrap();
+
+        let cdarray_model = Struct(ContiguousNDArray { item: item_model });
+
+        let current_bb = ctx.builder.get_insert_block().unwrap();
+        let then_bb = ctx.ctx.insert_basic_block_after(current_bb, "then_bb");
+        let else_bb = ctx.ctx.insert_basic_block_after(then_bb, "else_bb");
+        let end_bb = ctx.ctx.insert_basic_block_after(else_bb, "end_bb");
+
+        // Allocate and setup the resulting [`ContiguousNDArray`].
+        let result = cdarray_model.alloca(generator, ctx);
+
+        // Set ndims and shape.
+        let ndims = self.ndims_llvm(generator, ctx.ctx);
+        result.set(ctx, |f| f.ndims, ndims);
+
+        let shape = self.instance.get(generator, ctx, |f| f.shape);
+        result.set(ctx, |f| f.shape, shape);
+
+        let is_contiguous = self.is_c_contiguous(generator, ctx);
+        ctx.builder.build_conditional_branch(is_contiguous.value, then_bb, else_bb).unwrap();
+
+        // Inserting into then_bb; This ndarray is contiguous.
+        ctx.builder.position_at_end(then_bb);
+        let data = self.instance.get(generator, ctx, |f| f.data);
+        let data = data.pointer_cast(generator, ctx, item_model);
+        result.set(ctx, |f| f.data, data);
+        ctx.builder.build_unconditional_branch(end_bb).unwrap();
+
+        // Inserting into else_bb; This ndarray is not contiguous. Do a full-copy on `data`.
+        // `make_copy` produces an ndarray with contiguous `data`.
+        ctx.builder.position_at_end(else_bb);
+        let copied_ndarray = self.make_copy(generator, ctx);
+        let data = copied_ndarray.instance.get(generator, ctx, |f| f.data);
+        let data = data.pointer_cast(generator, ctx, item_model);
+        result.set(ctx, |f| f.data, data);
+        ctx.builder.build_unconditional_branch(end_bb).unwrap();
+
+        // Reposition to end_bb for continuation
+        ctx.builder.position_at_end(end_bb);
+
+        result
+    }
+
+    /// Create an [`NDArrayObject`] from a [`ContiguousNDArray`].
+    ///
+    /// The operation is super cheap. The newly created [`NDArrayObject`] will share the
+    /// same memory as the [`ContiguousNDArray`].
+    ///
+    /// `ndims` has to be provided as [`NDArrayObject`] requires a statically known `ndims` value, despite
+    /// the fact that the information should be contained within the [`ContiguousNDArray`].
+    pub fn from_contiguous_ndarray<G: CodeGenerator + ?Sized, Item: Model<'ctx>>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        carray: Instance<'ctx, Ptr<Struct<ContiguousNDArray<Item>>>>,
+        dtype: Type,
+        ndims: u64,
+    ) -> Self {
+        // Sanity check on `dtype` and `contiguous_array`'s `Item` model.
+        let dtype_llvm = ctx.get_llvm_type(generator, dtype);
+        carray.model.0 .0.item.check_type(generator, ctx.ctx, dtype_llvm).unwrap();
+
+        // TODO: Debug assert `ndims == carray.ndims` to catch bugs.
+
+        // Allocate the resulting ndarray.
+        let ndarray = NDArrayObject::alloca(generator, ctx, dtype, ndims);
+
+        // Copy shape and update strides
+        let shape = carray.get(generator, ctx, |f| f.shape);
+        ndarray.copy_shape_from_array(generator, ctx, shape);
+        ndarray.set_strides_contiguous(generator, ctx);
+
+        // Share data
+        let data = carray.get(generator, ctx, |f| f.data).pointer_cast(generator, ctx, Int(Byte));
+        ndarray.instance.set(ctx, |f| f.data, data);
+
+        ndarray
+    }
+}

nac3core/src/codegen/object/ndarray/factory.rs

@@ -0,0 +1,176 @@
+use inkwell::{values::BasicValueEnum, IntPredicate};
+
+use crate::{
+    codegen::{
+        irrt::call_nac3_ndarray_util_assert_shape_no_negative, model::*, CodeGenContext,
+        CodeGenerator,
+    },
+    typecheck::typedef::Type,
+};
+
+use super::NDArrayObject;
+
+/// Get the zero value in `np.zeros()` of a `dtype`.
+fn ndarray_zero_value<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    dtype: Type,
+) -> BasicValueEnum<'ctx> {
+    if [ctx.primitives.int32, ctx.primitives.uint32]
+        .iter()
+        .any(|ty| ctx.unifier.unioned(dtype, *ty))
+    {
+        ctx.ctx.i32_type().const_zero().into()
+    } else if [ctx.primitives.int64, ctx.primitives.uint64]
+        .iter()
+        .any(|ty| ctx.unifier.unioned(dtype, *ty))
+    {
+        ctx.ctx.i64_type().const_zero().into()
+    } else if ctx.unifier.unioned(dtype, ctx.primitives.float) {
+        ctx.ctx.f64_type().const_zero().into()
+    } else if ctx.unifier.unioned(dtype, ctx.primitives.bool) {
+        ctx.ctx.bool_type().const_zero().into()
+    } else if ctx.unifier.unioned(dtype, ctx.primitives.str) {
+        ctx.gen_string(generator, "").into()
+    } else {
+        panic!("unrecognized dtype: {}", ctx.unifier.stringify(dtype));
+    }
+}
+
+/// Get the one value in `np.ones()` of a `dtype`.
+fn ndarray_one_value<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    dtype: Type,
+) -> BasicValueEnum<'ctx> {
+    if [ctx.primitives.int32, ctx.primitives.uint32]
+        .iter()
+        .any(|ty| ctx.unifier.unioned(dtype, *ty))
+    {
+        let is_signed = ctx.unifier.unioned(dtype, ctx.primitives.int32);
+        ctx.ctx.i32_type().const_int(1, is_signed).into()
+    } else if [ctx.primitives.int64, ctx.primitives.uint64]
+        .iter()
+        .any(|ty| ctx.unifier.unioned(dtype, *ty))
+    {
+        let is_signed = ctx.unifier.unioned(dtype, ctx.primitives.int64);
+        ctx.ctx.i64_type().const_int(1, is_signed).into()
+    } else if ctx.unifier.unioned(dtype, ctx.primitives.float) {
+        ctx.ctx.f64_type().const_float(1.0).into()
+    } else if ctx.unifier.unioned(dtype, ctx.primitives.bool) {
+        ctx.ctx.bool_type().const_int(1, false).into()
+    } else if ctx.unifier.unioned(dtype, ctx.primitives.str) {
+        ctx.gen_string(generator, "1").into()
+    } else {
+        panic!("unrecognized dtype: {}", ctx.unifier.stringify(dtype));
+    }
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Create an ndarray like `np.empty`.
+    pub fn make_np_empty<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        ndims: u64,
+        shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) -> Self {
+        // Validate `shape`
+        let ndims_llvm = Int(SizeT).const_int(generator, ctx.ctx, ndims);
+        call_nac3_ndarray_util_assert_shape_no_negative(generator, ctx, ndims_llvm, shape);
+
+        let ndarray = NDArrayObject::alloca(generator, ctx, dtype, ndims);
+        ndarray.copy_shape_from_array(generator, ctx, shape);
+        ndarray.create_data(generator, ctx);
+
+        ndarray
+    }
+
+    /// Create an ndarray like `np.full`.
+    pub fn make_np_full<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        ndims: u64,
+        shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+        fill_value: BasicValueEnum<'ctx>,
+    ) -> Self {
+        let ndarray = NDArrayObject::make_np_empty(generator, ctx, dtype, ndims, shape);
+        ndarray.fill(generator, ctx, fill_value);
+        ndarray
+    }
+
+    /// Create an ndarray like `np.zero`.
+    pub fn make_np_zeros<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        ndims: u64,
+        shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) -> Self {
+        let fill_value = ndarray_zero_value(generator, ctx, dtype);
+        NDArrayObject::make_np_full(generator, ctx, dtype, ndims, shape, fill_value)
+    }
+
+    /// Create an ndarray like `np.ones`.
+    pub fn make_np_ones<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        ndims: u64,
+        shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) -> Self {
+        let fill_value = ndarray_one_value(generator, ctx, dtype);
+        NDArrayObject::make_np_full(generator, ctx, dtype, ndims, shape, fill_value)
+    }
+
+    /// Create an ndarray like `np.eye`.
+    pub fn make_np_eye<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        nrows: Instance<'ctx, Int<SizeT>>,
+        ncols: Instance<'ctx, Int<SizeT>>,
+        offset: Instance<'ctx, Int<SizeT>>,
+    ) -> Self {
+        let ndzero = ndarray_zero_value(generator, ctx, dtype);
+        let ndone = ndarray_one_value(generator, ctx, dtype);
+
+        let ndarray = NDArrayObject::alloca_dynamic_shape(generator, ctx, dtype, &[nrows, ncols]);
+
+        // Create data and make the matrix like look np.eye()
+        ndarray.create_data(generator, ctx);
+        ndarray
+            .foreach(generator, ctx, |generator, ctx, _hooks, nditer| {
+                // NOTE: rows and cols can never be zero here, since this ndarray's `np.size` would be zero
+                // and this loop would not execute.
+
+                // Load up `row_i` and `col_i` from indices.
+                let row_i = nditer.get_indices().get_index_const(generator, ctx, 0);
+                let col_i = nditer.get_indices().get_index_const(generator, ctx, 1);
+
+                let be_one = row_i.add(ctx, offset).compare(ctx, IntPredicate::EQ, col_i);
+                let value = ctx.builder.build_select(be_one.value, ndone, ndzero, "value").unwrap();
+
+                let p = nditer.get_pointer(generator, ctx);
+                ctx.builder.build_store(p, value).unwrap();
+
+                Ok(())
+            })
+            .unwrap();
+
+        ndarray
+    }
+
+    /// Create an ndarray like `np.identity`.
+    pub fn make_np_identity<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        size: Instance<'ctx, Int<SizeT>>,
+    ) -> Self {
+        // Convenient implementation
+        let offset = Int(SizeT).const_0(generator, ctx.ctx);
+        NDArrayObject::make_np_eye(generator, ctx, dtype, size, size, offset)
+    }
+}

nac3core/src/codegen/object/ndarray/indexing.rs

@@ -0,0 +1,293 @@
+use crate::codegen::{irrt::call_nac3_ndarray_index, model::*, CodeGenContext, CodeGenerator};
+
+use super::NDArrayObject;
+
+pub type NDIndexType = Byte;
+
+/// Fields of [`NDIndex`]
+#[derive(Debug, Clone, Copy)]
+pub struct NDIndexFields<'ctx, F: FieldTraversal<'ctx>> {
+    pub type_: F::Out<Int<NDIndexType>>, // Defined to be uint8_t in IRRT
+    pub data: F::Out<Ptr<Int<Byte>>>,
+}
+
+/// An IRRT representation of an ndarray subscript index.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct NDIndex;
+
+impl<'ctx> StructKind<'ctx> for NDIndex {
+    type Fields<F: FieldTraversal<'ctx>> = NDIndexFields<'ctx, F>;
+
+    fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+        Self::Fields { type_: traversal.add_auto("type"), data: traversal.add_auto("data") }
+    }
+}
+
+/// Fields of [`Slice`]
+#[derive(Debug, Clone)]
+pub struct SliceFields<'ctx, F: FieldTraversal<'ctx>> {
+    pub start_defined: F::Out<Int<Bool>>,
+    pub start: F::Out<Int<Int32>>,
+    pub stop_defined: F::Out<Int<Bool>>,
+    pub stop: F::Out<Int<Int32>>,
+    pub step_defined: F::Out<Int<Bool>>,
+    pub step: F::Out<Int<Int32>>,
+}
+
+/// An IRRT representation of an (unresolved) slice.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct Slice;
+
+impl<'ctx> StructKind<'ctx> for Slice {
+    type Fields<F: FieldTraversal<'ctx>> = SliceFields<'ctx, F>;
+
+    fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+        Self::Fields {
+            start_defined: traversal.add_auto("start_defined"),
+            start: traversal.add_auto("start"),
+            stop_defined: traversal.add_auto("stop_defined"),
+            stop: traversal.add_auto("stop"),
+            step_defined: traversal.add_auto("step_defined"),
+            step: traversal.add_auto("step"),
+        }
+    }
+}
+
+/// A convenience structure to prepare a [`Slice`].
+#[derive(Debug, Clone)]
+pub struct RustSlice<'ctx> {
+    pub start: Option<Instance<'ctx, Int<Int32>>>,
+    pub stop: Option<Instance<'ctx, Int<Int32>>>,
+    pub step: Option<Instance<'ctx, Int<Int32>>>,
+}
+
+impl<'ctx> RustSlice<'ctx> {
+    /// Write the contents to an LLVM [`Slice`].
+    pub fn write_to_slice<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        dst_slice_ptr: Instance<'ctx, Ptr<Struct<Slice>>>,
+    ) {
+        let false_ = Int(Bool).const_false(generator, ctx.ctx);
+        let true_ = Int(Bool).const_true(generator, ctx.ctx);
+
+        match self.start {
+            Some(start) => {
+                dst_slice_ptr.gep(ctx, |f| f.start_defined).store(ctx, true_);
+                dst_slice_ptr.gep(ctx, |f| f.start).store(ctx, start);
+            }
+            None => dst_slice_ptr.gep(ctx, |f| f.start_defined).store(ctx, false_),
+        }
+
+        match self.stop {
+            Some(stop) => {
+                dst_slice_ptr.gep(ctx, |f| f.stop_defined).store(ctx, true_);
+                dst_slice_ptr.gep(ctx, |f| f.stop).store(ctx, stop);
+            }
+            None => dst_slice_ptr.gep(ctx, |f| f.stop_defined).store(ctx, false_),
+        }
+
+        match self.step {
+            Some(step) => {
+                dst_slice_ptr.gep(ctx, |f| f.step_defined).store(ctx, true_);
+                dst_slice_ptr.gep(ctx, |f| f.step).store(ctx, step);
+            }
+            None => dst_slice_ptr.gep(ctx, |f| f.step_defined).store(ctx, false_),
+        }
+    }
+}
+
+// A convenience enum to prepare an [`NDIndex`].
+#[derive(Debug, Clone)]
+pub enum RustNDIndex<'ctx> {
+    SingleElement(Instance<'ctx, Int<Int32>>), // TODO: To be SizeT
+    Slice(RustSlice<'ctx>),
+    NewAxis,
+    Ellipsis,
+}
+
+impl<'ctx> RustNDIndex<'ctx> {
+    /// Get the value to set `NDIndex::type` for this variant.
+    fn get_type_id(&self) -> u64 {
+        // Defined in IRRT, must be in sync
+        match self {
+            RustNDIndex::SingleElement(_) => 0,
+            RustNDIndex::Slice(_) => 1,
+            RustNDIndex::NewAxis => 2,
+            RustNDIndex::Ellipsis => 3,
+        }
+    }
+
+    /// Write the contents to an LLVM [`NDIndex`].
+    fn write_to_ndindex<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        dst_ndindex_ptr: Instance<'ctx, Ptr<Struct<NDIndex>>>,
+    ) {
+        // Set `dst_ndindex_ptr->type`
+        dst_ndindex_ptr.gep(ctx, |f| f.type_).store(
+            ctx,
+            Int(NDIndexType::default()).const_int(generator, ctx.ctx, self.get_type_id()),
+        );
+
+        // Set `dst_ndindex_ptr->data`
+        match self {
+            RustNDIndex::SingleElement(in_index) => {
+                let index_ptr = Int(Int32).alloca(generator, ctx);
+                index_ptr.store(ctx, *in_index);
+
+                dst_ndindex_ptr
+                    .gep(ctx, |f| f.data)
+                    .store(ctx, index_ptr.pointer_cast(generator, ctx, Int(Byte)));
+            }
+            RustNDIndex::Slice(in_rust_slice) => {
+                let user_slice_ptr = Struct(Slice).alloca(generator, ctx);
+                in_rust_slice.write_to_slice(generator, ctx, user_slice_ptr);
+
+                dst_ndindex_ptr
+                    .gep(ctx, |f| f.data)
+                    .store(ctx, user_slice_ptr.pointer_cast(generator, ctx, Int(Byte)));
+            }
+            RustNDIndex::NewAxis | RustNDIndex::Ellipsis => {}
+        }
+    }
+
+    /// Allocate an array of `NDIndex`es on the stack and return its stack pointer.
+    pub fn alloca_ndindices<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        in_ndindices: &[RustNDIndex<'ctx>],
+    ) -> (Instance<'ctx, Int<SizeT>>, Instance<'ctx, Ptr<Struct<NDIndex>>>) {
+        let ndindex_model = Struct(NDIndex);
+
+        let num_ndindices = Int(SizeT).const_int(generator, ctx.ctx, in_ndindices.len() as u64);
+        let ndindices = ndindex_model.array_alloca(generator, ctx, num_ndindices.value);
+        for (i, in_ndindex) in in_ndindices.iter().enumerate() {
+            let pndindex = ndindices.offset_const(ctx, i as u64);
+            in_ndindex.write_to_ndindex(generator, ctx, pndindex);
+        }
+
+        (num_ndindices, ndindices)
+    }
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Get the ndims [`Type`] after indexing with a given slice.
+    #[must_use]
+    pub fn deduce_ndims_after_indexing_with(&self, indices: &[RustNDIndex<'ctx>]) -> u64 {
+        let mut ndims = self.ndims;
+        for index in indices {
+            match index {
+                RustNDIndex::SingleElement(_) => {
+                    ndims -= 1; // Single elements decrements ndims
+                }
+                RustNDIndex::NewAxis => {
+                    ndims += 1; // `np.newaxis` / `none` adds a new axis
+                }
+                RustNDIndex::Ellipsis | RustNDIndex::Slice(_) => {}
+            }
+        }
+        ndims
+    }
+
+    /// Index into the ndarray, and return a newly-allocated view on this ndarray.
+    ///
+    /// This function behaves like NumPy's ndarray indexing, but if the indices index
+    /// into a single element, an unsized ndarray is returned.
+    #[must_use]
+    pub fn index<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        indices: &[RustNDIndex<'ctx>],
+    ) -> Self {
+        let dst_ndims = self.deduce_ndims_after_indexing_with(indices);
+        let dst_ndarray = NDArrayObject::alloca(generator, ctx, self.dtype, dst_ndims);
+
+        let (num_indices, indices) = RustNDIndex::alloca_ndindices(generator, ctx, indices);
+        call_nac3_ndarray_index(
+            generator,
+            ctx,
+            num_indices,
+            indices,
+            self.instance,
+            dst_ndarray.instance,
+        );
+
+        dst_ndarray
+    }
+}
+
+pub mod util {
+    use itertools::Itertools;
+    use nac3parser::ast::{Expr, ExprKind};
+
+    use crate::{
+        codegen::{expr::gen_slice, model::*, CodeGenContext, CodeGenerator},
+        typecheck::typedef::Type,
+    };
+
+    use super::{RustNDIndex, RustSlice};
+
+    /// Generate LLVM code to transform an ndarray subscript expression to
+    /// its list of [`RustNDIndex`]
+    ///
+    /// i.e.,
+    /// ```python
+    /// my_ndarray[::3, 1, :2:]
+    ///            ^^^^^^^^^^^ Then these into a three `RustNDIndex`es
+    /// ```
+    pub fn gen_ndarray_subscript_ndindices<'ctx, G: CodeGenerator>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        subscript: &Expr<Option<Type>>,
+    ) -> Result<Vec<RustNDIndex<'ctx>>, String> {
+        // TODO: Support https://numpy.org/doc/stable/user/basics.indexing.html#dimensional-indexing-tools
+
+        // Annoying notes about `slice`
+        //  - `my_array[5]`
+        //    - slice is a `Constant`
+        //  - `my_array[:5]`
+        //    - slice is a `Slice`
+        //  - `my_array[:]`
+        //    - slice is a `Slice`, but lower upper step would all be `Option::None`
+        //  - `my_array[:, :]`
+        //    - slice is now a `Tuple` of two `Slice`-s
+        //
+        // In summary:
+        //  - when there is a comma "," within [], `slice` will be a `Tuple` of the entries.
+        //  - when there is not comma "," within [] (i.e., just a single entry), `slice` will be that entry itself.
+        //
+        // So we first "flatten" out the slice expression
+        let index_exprs = match &subscript.node {
+            ExprKind::Tuple { elts, .. } => elts.iter().collect_vec(),
+            _ => vec![subscript],
+        };
+
+        // Process all index expressions
+        let mut rust_ndindices: Vec<RustNDIndex> = Vec::with_capacity(index_exprs.len()); // Not using iterators here because `?` is used here.
+        for index_expr in index_exprs {
+            // NOTE: Currently nac3core's slices do not have an object representation,
+            // so the code/implementation looks awkward - we have to do pattern matching on the expression
+            let ndindex = if let ExprKind::Slice { lower, upper, step } = &index_expr.node {
+                // Handle slices
+                let (lower, upper, step) = gen_slice(generator, ctx, lower, upper, step)?;
+                RustNDIndex::Slice(RustSlice { start: lower, stop: upper, step })
+            } else {
+                // Treat and handle everything else as a single element index.
+                let index = generator.gen_expr(ctx, index_expr)?.unwrap().to_basic_value_enum(
+                    ctx,
+                    generator,
+                    ctx.primitives.int32, // Must be int32, this checks for illegal values
+                )?;
+                let index = Int(Int32).check_value(generator, ctx.ctx, index).unwrap();
+
+                RustNDIndex::SingleElement(index)
+            };
+            rust_ndindices.push(ndindex);
+        }
+        Ok(rust_ndindices)
+    }
+}

nac3core/src/codegen/object/ndarray/map.rs

@@ -0,0 +1,220 @@
+use inkwell::values::BasicValueEnum;
+use itertools::Itertools;
+
+use crate::{
+    codegen::{
+        object::ndarray::{AnyObject, NDArrayObject},
+        stmt::gen_for_callback,
+        CodeGenContext, CodeGenerator,
+    },
+    typecheck::typedef::Type,
+};
+
+use super::{nditer::NDIterHandle, NDArrayOut, ScalarOrNDArray};
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Generate LLVM IR to broadcast `ndarray`s together, and starmap through them with `mapping` elementwise.
+    ///
+    /// `mapping` is an LLVM IR generator. The input of `mapping` is the list of elements when iterating through
+    /// the input `ndarrays` after broadcasting. The output of `mapping` is the result of the elementwise operation.
+    ///
+    /// `out` specifies whether the result should be a new ndarray or to be written an existing ndarray.
+    pub fn broadcast_starmap<'a, G, MappingFn>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, 'a>,
+        ndarrays: &[Self],
+        out: NDArrayOut<'ctx>,
+        mapping: MappingFn,
+    ) -> Result<Self, String>
+    where
+        G: CodeGenerator + ?Sized,
+        MappingFn: FnOnce(
+            &mut G,
+            &mut CodeGenContext<'ctx, 'a>,
+            &[BasicValueEnum<'ctx>],
+        ) -> Result<BasicValueEnum<'ctx>, String>,
+    {
+        // Broadcast inputs
+        let broadcast_result = NDArrayObject::broadcast(generator, ctx, ndarrays);
+
+        let out_ndarray = match out {
+            NDArrayOut::NewNDArray { dtype } => {
+                // Create a new ndarray based on the broadcast shape.
+                let result_ndarray =
+                    NDArrayObject::alloca(generator, ctx, dtype, broadcast_result.ndims);
+                result_ndarray.copy_shape_from_array(generator, ctx, broadcast_result.shape);
+                result_ndarray.create_data(generator, ctx);
+                result_ndarray
+            }
+            NDArrayOut::WriteToNDArray { ndarray: result_ndarray } => {
+                // Use an existing ndarray.
+
+                // Check that its shape is compatible with the broadcast shape.
+                result_ndarray.assert_can_be_written_by_out(
+                    generator,
+                    ctx,
+                    broadcast_result.ndims,
+                    broadcast_result.shape,
+                );
+                result_ndarray
+            }
+        };
+
+        // Map element-wise and store results into `mapped_ndarray`.
+        let nditer = NDIterHandle::new(generator, ctx, out_ndarray);
+        gen_for_callback(
+            generator,
+            ctx,
+            Some("broadcast_starmap"),
+            |generator, ctx| {
+                // Create NDIters for all broadcasted input ndarrays.
+                let other_nditers = broadcast_result
+                    .ndarrays
+                    .iter()
+                    .map(|ndarray| NDIterHandle::new(generator, ctx, *ndarray))
+                    .collect_vec();
+                Ok((nditer, other_nditers))
+            },
+            |generator, ctx, (out_nditer, _in_nditers)| {
+                // We can simply use `out_nditer`'s `has_next()`.
+                // `in_nditers`' `has_next()`s should return the same value.
+                Ok(out_nditer.has_next(generator, ctx).value)
+            },
+            |generator, ctx, _hooks, (out_nditer, in_nditers)| {
+                // Get all the scalars from the broadcasted input ndarrays, pass them to `mapping`,
+                // and write to `out_ndarray`.
+
+                let in_scalars = in_nditers
+                    .iter()
+                    .map(|nditer| nditer.get_scalar(generator, ctx).value)
+                    .collect_vec();
+
+                let result = mapping(generator, ctx, &in_scalars)?;
+
+                let p = out_nditer.get_pointer(generator, ctx);
+                ctx.builder.build_store(p, result).unwrap();
+
+                Ok(())
+            },
+            |generator, ctx, (out_nditer, in_nditers)| {
+                // Advance all iterators
+                out_nditer.next(generator, ctx);
+                in_nditers.iter().for_each(|nditer| nditer.next(generator, ctx));
+                Ok(())
+            },
+        )?;
+
+        Ok(out_ndarray)
+    }
+
+    /// Map through this ndarray with an elementwise function.
+    pub fn map<'a, G, Mapping>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, 'a>,
+        out: NDArrayOut<'ctx>,
+        mapping: Mapping,
+    ) -> Result<Self, String>
+    where
+        G: CodeGenerator + ?Sized,
+        Mapping: FnOnce(
+            &mut G,
+            &mut CodeGenContext<'ctx, 'a>,
+            BasicValueEnum<'ctx>,
+        ) -> Result<BasicValueEnum<'ctx>, String>,
+    {
+        NDArrayObject::broadcast_starmap(
+            generator,
+            ctx,
+            &[*self],
+            out,
+            |generator, ctx, scalars| mapping(generator, ctx, scalars[0]),
+        )
+    }
+}
+
+impl<'ctx> ScalarOrNDArray<'ctx> {
+    /// Starmap through a list of inputs using `mapping`, where an input could be an ndarray, a scalar.
+    ///
+    /// This function is very helpful when implementing NumPy functions that takes on either scalars or ndarrays or a mix of them
+    /// as their inputs and produces either an ndarray with broadcast, or a scalar if all its inputs are all scalars.
+    ///
+    /// For example ,this function can be used to implement `np.add`, which has the following behaviors:
+    /// - `np.add(3, 4) = 7` # (scalar, scalar) -> scalar
+    /// - `np.add(3, np.array([4, 5, 6]))` # (scalar, ndarray) -> ndarray; the first `scalar` is converted into an ndarray and broadcasted.
+    /// - `np.add(np.array([[1], [2], [3]]), np.array([[4, 5, 6]]))` # (ndarray, ndarray) -> ndarray; there is broadcasting.
+    ///
+    /// ## Details:
+    ///
+    /// If `inputs` are all [`ScalarOrNDArray::Scalar`], the output will be a [`ScalarOrNDArray::Scalar`] with type `ret_dtype`.
+    ///
+    /// Otherwise (if there are any [`ScalarOrNDArray::NDArray`] in `inputs`), all inputs will be 'as-ndarray'-ed into ndarrays,
+    /// then all inputs (now all ndarrays) will be passed to [`NDArrayObject::broadcasting_starmap`] and **create** a new ndarray
+    /// with dtype `ret_dtype`.
+    pub fn broadcasting_starmap<'a, G, MappingFn>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, 'a>,
+        inputs: &[ScalarOrNDArray<'ctx>],
+        ret_dtype: Type,
+        mapping: MappingFn,
+    ) -> Result<ScalarOrNDArray<'ctx>, String>
+    where
+        G: CodeGenerator + ?Sized,
+        MappingFn: FnOnce(
+            &mut G,
+            &mut CodeGenContext<'ctx, 'a>,
+            &[BasicValueEnum<'ctx>],
+        ) -> Result<BasicValueEnum<'ctx>, String>,
+    {
+        // Check if all inputs are Scalars
+        let all_scalars: Option<Vec<_>> = inputs.iter().map(AnyObject::try_from).try_collect().ok();
+
+        if let Some(scalars) = all_scalars {
+            let scalars = scalars.iter().map(|scalar| scalar.value).collect_vec();
+            let value = mapping(generator, ctx, &scalars)?;
+
+            Ok(ScalarOrNDArray::Scalar(AnyObject { ty: ret_dtype, value }))
+        } else {
+            // Promote all input to ndarrays and map through them.
+            let inputs = inputs.iter().map(|input| input.to_ndarray(generator, ctx)).collect_vec();
+            let ndarray = NDArrayObject::broadcast_starmap(
+                generator,
+                ctx,
+                &inputs,
+                NDArrayOut::NewNDArray { dtype: ret_dtype },
+                mapping,
+            )?;
+            Ok(ScalarOrNDArray::NDArray(ndarray))
+        }
+    }
+
+    /// Map through this [`ScalarOrNDArray`] with an elementwise function.
+    ///
+    /// If this is a scalar, `mapping` will directly act on the scalar. This function will return a [`ScalarOrNDArray::Scalar`] of that result.
+    ///
+    /// If this is an ndarray, `mapping` will be applied to the elements of the ndarray. A new ndarray of the results will be created and
+    /// returned as a [`ScalarOrNDArray::NDArray`].
+    pub fn map<'a, G, Mapping>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, 'a>,
+        ret_dtype: Type,
+        mapping: Mapping,
+    ) -> Result<ScalarOrNDArray<'ctx>, String>
+    where
+        G: CodeGenerator + ?Sized,
+        Mapping: FnOnce(
+            &mut G,
+            &mut CodeGenContext<'ctx, 'a>,
+            BasicValueEnum<'ctx>,
+        ) -> Result<BasicValueEnum<'ctx>, String>,
+    {
+        ScalarOrNDArray::broadcasting_starmap(
+            generator,
+            ctx,
+            &[*self],
+            ret_dtype,
+            |generator, ctx, scalars| mapping(generator, ctx, scalars[0]),
+        )
+    }
+}

nac3core/src/codegen/object/ndarray/matmul.rs

@@ -0,0 +1,207 @@
+use std::cmp::max;
+
+use nac3parser::ast::Operator;
+use util::gen_for_model;
+
+use crate::{
+    codegen::{
+        expr::gen_binop_expr_with_values, irrt::call_nac3_ndarray_matmul_calculate_shapes,
+        model::*, object::ndarray::indexing::RustNDIndex, CodeGenContext, CodeGenerator,
+    },
+    typecheck::{magic_methods::Binop, typedef::Type},
+};
+
+use super::{NDArrayObject, NDArrayOut};
+
+/// Perform `np.einsum("...ij,...jk->...ik", in_a, in_b)`.
+///
+/// `dst_dtype` defines the dtype of the returned ndarray.
+fn matmul_at_least_2d<'ctx, G: CodeGenerator>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    dst_dtype: Type,
+    in_a: NDArrayObject<'ctx>,
+    in_b: NDArrayObject<'ctx>,
+) -> NDArrayObject<'ctx> {
+    assert!(in_a.ndims >= 2);
+    assert!(in_b.ndims >= 2);
+
+    // Deduce ndims of the result of matmul.
+    let ndims_int = max(in_a.ndims, in_b.ndims);
+    let ndims = Int(SizeT).const_int(generator, ctx.ctx, ndims_int);
+
+    let num_0 = Int(SizeT).const_int(generator, ctx.ctx, 0);
+    let num_1 = Int(SizeT).const_int(generator, ctx.ctx, 1);
+
+    // Broadcasts `in_a.shape[:-2]` and `in_b.shape[:-2]` together and allocate the
+    // destination ndarray to store the result of matmul.
+    let (a, b, dst) = {
+        let in_a_ndims = in_a.ndims_llvm(generator, ctx.ctx);
+        let in_a_shape = in_a.instance.get(generator, ctx, |f| f.shape);
+        let in_b_ndims = in_b.ndims_llvm(generator, ctx.ctx);
+        let in_b_shape = in_b.instance.get(generator, ctx, |f| f.shape);
+        let a_shape = Int(SizeT).array_alloca(generator, ctx, ndims.value);
+        let b_shape = Int(SizeT).array_alloca(generator, ctx, ndims.value);
+        let dst_shape = Int(SizeT).array_alloca(generator, ctx, ndims.value);
+
+        // Matmul dimension compatibility is checked here.
+        call_nac3_ndarray_matmul_calculate_shapes(
+            generator, ctx, in_a_ndims, in_a_shape, in_b_ndims, in_b_shape, ndims, a_shape,
+            b_shape, dst_shape,
+        );
+
+        let a = in_a.broadcast_to(generator, ctx, ndims_int, a_shape);
+        let b = in_b.broadcast_to(generator, ctx, ndims_int, b_shape);
+
+        let dst = NDArrayObject::alloca(generator, ctx, dst_dtype, ndims_int);
+        dst.copy_shape_from_array(generator, ctx, dst_shape);
+        dst.create_data(generator, ctx);
+
+        (a, b, dst)
+    };
+
+    let len =
+        a.instance.get(generator, ctx, |f| f.shape).get_index_const(generator, ctx, ndims_int - 1);
+
+    let at_row = ndims_int - 2;
+    let at_col = ndims_int - 1;
+
+    let dst_dtype_llvm = ctx.get_llvm_type(generator, dst_dtype);
+    let dst_zero = dst_dtype_llvm.const_zero();
+
+    dst.foreach(generator, ctx, |generator, ctx, _, hdl| {
+        let pdst_ij = hdl.get_pointer(generator, ctx);
+
+        ctx.builder.build_store(pdst_ij, dst_zero).unwrap();
+
+        let indices = hdl.get_indices();
+        let i = indices.get_index_const(generator, ctx, at_row);
+        let j = indices.get_index_const(generator, ctx, at_col);
+
+        gen_for_model(generator, ctx, num_0, len, num_1, |generator, ctx, _, k| {
+            // `indices` is modified to index into `a` and `b`, and restored.
+            indices.set_index_const(ctx, at_row, i);
+            indices.set_index_const(ctx, at_col, k);
+            let a_ik = a.get_scalar_by_indices(generator, ctx, indices);
+
+            indices.set_index_const(ctx, at_row, k);
+            indices.set_index_const(ctx, at_col, j);
+            let b_kj = b.get_scalar_by_indices(generator, ctx, indices);
+
+            // Restore `indices`.
+            indices.set_index_const(ctx, at_row, i);
+            indices.set_index_const(ctx, at_col, j);
+
+            // x = a_[...]ik * b_[...]kj
+            let x = gen_binop_expr_with_values(
+                generator,
+                ctx,
+                (&Some(a.dtype), a_ik.value),
+                Binop::normal(Operator::Mult),
+                (&Some(b.dtype), b_kj.value),
+                ctx.current_loc,
+            )?
+            .unwrap()
+            .to_basic_value_enum(ctx, generator, dst_dtype)?;
+
+            // dst_[...]ij += x
+            let dst_ij = ctx.builder.build_load(pdst_ij, "").unwrap();
+            let dst_ij = gen_binop_expr_with_values(
+                generator,
+                ctx,
+                (&Some(dst_dtype), dst_ij),
+                Binop::normal(Operator::Add),
+                (&Some(dst_dtype), x),
+                ctx.current_loc,
+            )?
+            .unwrap()
+            .to_basic_value_enum(ctx, generator, dst_dtype)?;
+            ctx.builder.build_store(pdst_ij, dst_ij).unwrap();
+
+            Ok(())
+        })
+    })
+    .unwrap();
+
+    dst
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Perform `np.matmul` according to the rules in
+    /// <https://numpy.org/doc/stable/reference/generated/numpy.matmul.html>.
+    ///
+    /// This function always return an [`NDArrayObject`]. You may want to use [`NDArrayObject::split_unsized`]
+    /// to handle when the output could be a scalar.
+    ///
+    /// `dst_dtype` defines the dtype of the returned ndarray.
+    pub fn matmul<G: CodeGenerator>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        a: Self,
+        b: Self,
+        out: NDArrayOut<'ctx>,
+    ) -> Self {
+        // Sanity check, but type inference should prevent this.
+        assert!(a.ndims > 0 && b.ndims > 0, "np.matmul disallows scalar input");
+
+        /*
+            If both arguments are 2-D they are multiplied like conventional matrices.
+            If either argument is N-D, N > 2, it is treated as a stack of matrices residing in the last two indices and broadcast accordingly.
+            If the first argument is 1-D, it is promoted to a matrix by prepending a 1 to its dimensions. After matrix multiplication the prepended 1 is removed.
+            If the second argument is 1-D, it is promoted to a matrix by appending a 1 to its dimensions. After matrix multiplication the appended 1 is removed.
+        */
+
+        let new_a = if a.ndims == 1 {
+            // Prepend 1 to its dimensions
+            a.index(generator, ctx, &[RustNDIndex::NewAxis, RustNDIndex::Ellipsis])
+        } else {
+            a
+        };
+
+        let new_b = if b.ndims == 1 {
+            // Append 1 to its dimensions
+            b.index(generator, ctx, &[RustNDIndex::Ellipsis, RustNDIndex::NewAxis])
+        } else {
+            b
+        };
+
+        // NOTE: `result` will always be a newly allocated ndarray.
+        // Current implementation cannot do in-place matrix muliplication.
+        let mut result = matmul_at_least_2d(generator, ctx, out.get_dtype(), new_a, new_b);
+
+        // Postprocessing on the result to remove prepended/appended axes.
+        let mut postindices = vec![];
+        let zero = Int(Int32).const_0(generator, ctx.ctx);
+
+        if a.ndims == 1 {
+            // Remove the prepended 1
+            postindices.push(RustNDIndex::SingleElement(zero));
+        }
+
+        if b.ndims == 1 {
+            // Remove the appended 1
+            postindices.push(RustNDIndex::Ellipsis);
+            postindices.push(RustNDIndex::SingleElement(zero));
+        }
+
+        if !postindices.is_empty() {
+            result = result.index(generator, ctx, &postindices);
+        }
+
+        match out {
+            NDArrayOut::NewNDArray { .. } => result,
+            NDArrayOut::WriteToNDArray { ndarray: out_ndarray } => {
+                let result_shape = result.instance.get(generator, ctx, |f| f.shape);
+                out_ndarray.assert_can_be_written_by_out(
+                    generator,
+                    ctx,
+                    result.ndims,
+                    result_shape,
+                );
+
+                out_ndarray.copy_data_from(generator, ctx, result);
+                out_ndarray
+            }
+        }
+    }
+}

nac3core/src/codegen/object/ndarray/mod.rs

@@ -0,0 +1,657 @@
+pub mod array;
+pub mod broadcast;
+pub mod contiguous;
+pub mod factory;
+pub mod indexing;
+pub mod map;
+pub mod matmul;
+pub mod nditer;
+pub mod shape_util;
+pub mod view;
+
+use inkwell::{
+    context::Context,
+    types::BasicType,
+    values::{BasicValue, BasicValueEnum, PointerValue},
+    AddressSpace,
+};
+
+use crate::{
+    codegen::{
+        irrt::{
+            call_nac3_ndarray_copy_data, call_nac3_ndarray_get_nth_pelement,
+            call_nac3_ndarray_get_pelement_by_indices, call_nac3_ndarray_is_c_contiguous,
+            call_nac3_ndarray_len, call_nac3_ndarray_nbytes,
+            call_nac3_ndarray_set_strides_by_shape, call_nac3_ndarray_size,
+            call_nac3_ndarray_util_assert_output_shape_same,
+        },
+        model::*,
+        CodeGenContext, CodeGenerator,
+    },
+    toplevel::{
+        helper::{create_ndims, extract_ndims},
+        numpy::{make_ndarray_ty, unpack_ndarray_var_tys},
+    },
+    typecheck::typedef::{Type, TypeEnum},
+};
+
+use super::{any::AnyObject, tuple::TupleObject};
+
+/// Fields of [`NDArray`]
+pub struct NDArrayFields<'ctx, F: FieldTraversal<'ctx>> {
+    pub data: F::Out<Ptr<Int<Byte>>>,
+    pub itemsize: F::Out<Int<SizeT>>,
+    pub ndims: F::Out<Int<SizeT>>,
+    pub shape: F::Out<Ptr<Int<SizeT>>>,
+    pub strides: F::Out<Ptr<Int<SizeT>>>,
+}
+
+/// A strided ndarray in NAC3.
+///
+/// See IRRT implementation for details about its fields.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct NDArray;
+
+impl<'ctx> StructKind<'ctx> for NDArray {
+    type Fields<F: FieldTraversal<'ctx>> = NDArrayFields<'ctx, F>;
+
+    fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+        Self::Fields {
+            data: traversal.add_auto("data"),
+            itemsize: traversal.add_auto("itemsize"),
+            ndims: traversal.add_auto("ndims"),
+            shape: traversal.add_auto("shape"),
+            strides: traversal.add_auto("strides"),
+        }
+    }
+}
+
+/// A NAC3 Python ndarray object.
+#[derive(Debug, Clone, Copy)]
+pub struct NDArrayObject<'ctx> {
+    pub dtype: Type,
+    pub ndims: u64,
+    pub instance: Instance<'ctx, Ptr<Struct<NDArray>>>,
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Attempt to convert an [`AnyObject`] into an [`NDArrayObject`].
+    pub fn from_object<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        object: AnyObject<'ctx>,
+    ) -> NDArrayObject<'ctx> {
+        let (dtype, ndims) = unpack_ndarray_var_tys(&mut ctx.unifier, object.ty);
+        let ndims = extract_ndims(&ctx.unifier, ndims);
+        Self::from_value_and_unpacked_types(generator, ctx, object.value, dtype, ndims)
+    }
+
+    /// Like [`NDArrayObject::from_object`] but you directly supply the ndarray's
+    /// `dtype` and `ndims`.
+    pub fn from_value_and_unpacked_types<V: BasicValue<'ctx>, G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        value: V,
+        dtype: Type,
+        ndims: u64,
+    ) -> Self {
+        let value = Ptr(Struct(NDArray)).check_value(generator, ctx.ctx, value).unwrap();
+        NDArrayObject { dtype, ndims, instance: value }
+    }
+
+    /// Check if an object can be converted into an ndarray.
+    pub fn is_instance(ctx: &mut CodeGenContext<'ctx, '_>, object: AnyObject<'ctx>) -> bool {
+        matches!(&*ctx.unifier.get_ty(object.ty), TypeEnum::TObj { obj_id, .. } if *obj_id == ctx.primitives.ndarray.obj_id(&ctx.unifier).unwrap())
+    }
+
+    /// Get this ndarray's `ndims` as an LLVM constant.
+    pub fn ndims_llvm<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> Instance<'ctx, Int<SizeT>> {
+        Int(SizeT).const_int(generator, ctx, self.ndims)
+    }
+
+    /// Get the typechecker ndarray type of this [`NDArrayObject`].
+    pub fn get_type(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> Type {
+        let ndims = create_ndims(&mut ctx.unifier, self.ndims);
+        make_ndarray_ty(&mut ctx.unifier, &ctx.primitives, Some(self.dtype), Some(ndims))
+    }
+
+    /// Forget that this is an ndarray and convert into an [`AnyObject`].
+    pub fn to_any(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> AnyObject<'ctx> {
+        let ty = self.get_type(ctx);
+        AnyObject { value: self.instance.value.as_basic_value_enum(), ty }
+    }
+
+    /// Allocate an ndarray on the stack given its `ndims` and `dtype`.
+    ///
+    /// `shape` and `strides` will be automatically allocated on the stack.
+    //e
+    /// The returned ndarray's content will be:
+    /// - `data`: set to `nullptr`.
+    /// - `itemsize`: set to the `sizeof()` of `dtype`.
+    /// - `ndims`: set to the value of  `ndims`.
+    /// - `shape`: allocated with an array of length `ndims` with uninitialized values.
+    /// - `strides`: allocated with an array of length `ndims` with uninitialized values.
+    pub fn alloca<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        ndims: u64,
+    ) -> Self {
+        let ndarray = Struct(NDArray).alloca(generator, ctx);
+
+        let data = Ptr(Int(Byte)).nullptr(generator, ctx.ctx);
+        ndarray.set(ctx, |f| f.data, data);
+
+        let itemsize = ctx.get_llvm_type(generator, dtype).size_of().unwrap();
+        let itemsize = Int(SizeT).s_extend_or_bit_cast(generator, ctx, itemsize);
+        ndarray.set(ctx, |f| f.itemsize, itemsize);
+
+        let ndims_val = Int(SizeT).const_int(generator, ctx.ctx, ndims);
+        ndarray.set(ctx, |f| f.ndims, ndims_val);
+
+        let shape = Int(SizeT).array_alloca(generator, ctx, ndims_val.value);
+        ndarray.set(ctx, |f| f.shape, shape);
+
+        let strides = Int(SizeT).array_alloca(generator, ctx, ndims_val.value);
+        ndarray.set(ctx, |f| f.strides, strides);
+
+        NDArrayObject { dtype, ndims, instance: ndarray }
+    }
+
+    /// Convenience function. Allocate an [`NDArrayObject`] with a statically known shape.
+    ///
+    /// The returned [`NDArrayObject`]'s `data` and `strides` are uninitialized.
+    pub fn alloca_constant_shape<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        shape: &[u64],
+    ) -> Self {
+        let ndarray = NDArrayObject::alloca(generator, ctx, dtype, shape.len() as u64);
+
+        // Write shape
+        let dst_shape = ndarray.instance.get(generator, ctx, |f| f.shape);
+        for (i, dim) in shape.iter().enumerate() {
+            let dim = Int(SizeT).const_int(generator, ctx.ctx, *dim);
+            dst_shape.offset_const(ctx, i as u64).store(ctx, dim);
+        }
+
+        ndarray
+    }
+
+    /// Convenience function. Allocate an [`NDArrayObject`] with a dynamically known shape.
+    ///
+    /// The returned [`NDArrayObject`]'s `data` and `strides` are uninitialized.
+    pub fn alloca_dynamic_shape<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        shape: &[Instance<'ctx, Int<SizeT>>],
+    ) -> Self {
+        let ndarray = NDArrayObject::alloca(generator, ctx, dtype, shape.len() as u64);
+
+        // Write shape
+        let dst_shape = ndarray.instance.get(generator, ctx, |f| f.shape);
+        for (i, dim) in shape.iter().enumerate() {
+            dst_shape.offset_const(ctx, i as u64).store(ctx, *dim);
+        }
+
+        ndarray
+    }
+
+    /// Initialize an ndarray's `data` by allocating a buffer on the stack.
+    /// The allocated data buffer is considered to be *owned* by the ndarray.
+    ///
+    /// `strides` of the ndarray will also be updated with `set_strides_by_shape`.
+    ///
+    /// `shape` and `itemsize` of the ndarray ***must*** be initialized first.
+    pub fn create_data<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) {
+        let nbytes = self.nbytes(generator, ctx);
+
+        let data = Int(Byte).array_alloca(generator, ctx, nbytes.value);
+        self.instance.set(ctx, |f| f.data, data);
+
+        self.set_strides_contiguous(generator, ctx);
+    }
+
+    /// Copy shape dimensions from an array.
+    pub fn copy_shape_from_array<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) {
+        let num_items = self.ndims_llvm(generator, ctx.ctx).value;
+        self.instance.get(generator, ctx, |f| f.shape).copy_from(generator, ctx, shape, num_items);
+    }
+
+    /// Copy shape dimensions from an ndarray.
+    /// Panics if `ndims` mismatches.
+    pub fn copy_shape_from_ndarray<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        src_ndarray: NDArrayObject<'ctx>,
+    ) {
+        assert_eq!(self.ndims, src_ndarray.ndims);
+        let src_shape = src_ndarray.instance.get(generator, ctx, |f| f.shape);
+        self.copy_shape_from_array(generator, ctx, src_shape);
+    }
+
+    /// Copy strides dimensions from an array.
+    pub fn copy_strides_from_array<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        strides: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) {
+        let num_items = self.ndims_llvm(generator, ctx.ctx).value;
+        self.instance
+            .get(generator, ctx, |f| f.strides)
+            .copy_from(generator, ctx, strides, num_items);
+    }
+
+    /// Copy strides dimensions from an ndarray.
+    /// Panics if `ndims` mismatches.
+    pub fn copy_strides_from_ndarray<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        src_ndarray: NDArrayObject<'ctx>,
+    ) {
+        assert_eq!(self.ndims, src_ndarray.ndims);
+        let src_strides = src_ndarray.instance.get(generator, ctx, |f| f.strides);
+        self.copy_strides_from_array(generator, ctx, src_strides);
+    }
+
+    /// Get the `np.size()` of this ndarray.
+    pub fn size<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<SizeT>> {
+        call_nac3_ndarray_size(generator, ctx, self.instance)
+    }
+
+    /// Get the `ndarray.nbytes` of this ndarray.
+    pub fn nbytes<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<SizeT>> {
+        call_nac3_ndarray_nbytes(generator, ctx, self.instance)
+    }
+
+    /// Get the `len()` of this ndarray.
+    pub fn len<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<SizeT>> {
+        call_nac3_ndarray_len(generator, ctx, self.instance)
+    }
+
+    /// Check if this ndarray is C-contiguous.
+    ///
+    /// See NumPy's `flags["C_CONTIGUOUS"]`: <https://numpy.org/doc/stable/reference/generated/numpy.ndarray.flags.html#numpy.ndarray.flags>
+    pub fn is_c_contiguous<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<Bool>> {
+        call_nac3_ndarray_is_c_contiguous(generator, ctx, self.instance)
+    }
+
+    /// Get the pointer to the n-th (0-based) element.
+    ///
+    /// The returned pointer has the element type of the LLVM type of this ndarray's `dtype`.
+    pub fn get_nth_pelement<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        nth: Instance<'ctx, Int<SizeT>>,
+    ) -> PointerValue<'ctx> {
+        let elem_ty = ctx.get_llvm_type(generator, self.dtype);
+
+        let p = call_nac3_ndarray_get_nth_pelement(generator, ctx, self.instance, nth);
+        ctx.builder
+            .build_pointer_cast(p.value, elem_ty.ptr_type(AddressSpace::default()), "")
+            .unwrap()
+    }
+
+    /// Get the n-th (0-based) scalar.
+    pub fn get_nth_scalar<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        nth: Instance<'ctx, Int<SizeT>>,
+    ) -> AnyObject<'ctx> {
+        let ptr = self.get_nth_pelement(generator, ctx, nth);
+        let value = ctx.builder.build_load(ptr, "").unwrap();
+        AnyObject { ty: self.dtype, value }
+    }
+
+    /// Get the pointer to the element indexed by `indices`.
+    ///
+    /// The returned pointer has the element type of the LLVM type of this ndarray's `dtype`.
+    pub fn get_pelement_by_indices<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        indices: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) -> PointerValue<'ctx> {
+        let elem_ty = ctx.get_llvm_type(generator, self.dtype);
+
+        let p = call_nac3_ndarray_get_pelement_by_indices(generator, ctx, self.instance, indices);
+        ctx.builder
+            .build_pointer_cast(p.value, elem_ty.ptr_type(AddressSpace::default()), "")
+            .unwrap()
+    }
+
+    /// Get the scalar indexed by `indices`.
+    pub fn get_scalar_by_indices<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        indices: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) -> AnyObject<'ctx> {
+        let ptr = self.get_pelement_by_indices(generator, ctx, indices);
+        let value = ctx.builder.build_load(ptr, "").unwrap();
+        AnyObject { ty: self.dtype, value }
+    }
+
+    /// Call [`call_nac3_ndarray_set_strides_by_shape`] on this ndarray to update `strides`.
+    ///
+    /// Update the ndarray's strides to make the ndarray contiguous.
+    pub fn set_strides_contiguous<G: CodeGenerator + ?Sized>(
+        self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) {
+        call_nac3_ndarray_set_strides_by_shape(generator, ctx, self.instance);
+    }
+
+    /// Clone/Copy this ndarray - Allocate a new ndarray with the same shape as this ndarray and copy the contents over.
+    ///
+    /// The new ndarray will own its data and will be C-contiguous.
+    #[must_use]
+    pub fn make_copy<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Self {
+        let clone = NDArrayObject::alloca(generator, ctx, self.dtype, self.ndims);
+
+        let shape = self.instance.gep(ctx, |f| f.shape).load(generator, ctx);
+        clone.copy_shape_from_array(generator, ctx, shape);
+        clone.create_data(generator, ctx);
+        clone.copy_data_from(generator, ctx, *self);
+        clone
+    }
+
+    /// Copy data from another ndarray.
+    ///
+    /// This ndarray and `src` is that their `np.size()` should be the same. Their shapes
+    /// do not matter. The copying order is determined by how their flattened views look.
+    ///
+    /// Panics if the `dtype`s of ndarrays are different.
+    pub fn copy_data_from<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        src: NDArrayObject<'ctx>,
+    ) {
+        assert!(ctx.unifier.unioned(self.dtype, src.dtype), "self and src dtype should match");
+        call_nac3_ndarray_copy_data(generator, ctx, src.instance, self.instance);
+    }
+
+    /// Returns true if this ndarray is unsized - `ndims == 0` and only contains a scalar.
+    #[must_use]
+    pub fn is_unsized(&self) -> bool {
+        self.ndims == 0
+    }
+
+    /// If this ndarray is unsized, return its sole value as an [`AnyObject`].
+    /// Otherwise, do nothing and return the ndarray itself.
+    pub fn split_unsized<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> ScalarOrNDArray<'ctx> {
+        if self.is_unsized() {
+            // NOTE: `np.size(self) == 0` here is never possible.
+            let zero = Int(SizeT).const_0(generator, ctx.ctx);
+            let value = self.get_nth_scalar(generator, ctx, zero).value;
+
+            ScalarOrNDArray::Scalar(AnyObject { ty: self.dtype, value })
+        } else {
+            ScalarOrNDArray::NDArray(*self)
+        }
+    }
+
+    /// Fill the ndarray with a scalar.
+    ///
+    /// `fill_value` must have the same LLVM type as the `dtype` of this ndarray.
+    pub fn fill<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        value: BasicValueEnum<'ctx>,
+    ) {
+        self.foreach(generator, ctx, |generator, ctx, _hooks, nditer| {
+            let p = nditer.get_pointer(generator, ctx);
+            ctx.builder.build_store(p, value).unwrap();
+            Ok(())
+        })
+        .unwrap();
+    }
+
+    /// Create the shape tuple of this ndarray like `np.shape(<ndarray>)`.
+    ///
+    /// The returned integers in the tuple are in int32.
+    pub fn make_shape_tuple<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> TupleObject<'ctx> {
+        // TODO: Return a tuple of SizeT
+
+        let mut objects = Vec::with_capacity(self.ndims as usize);
+
+        for i in 0..self.ndims {
+            let dim = self
+                .instance
+                .get(generator, ctx, |f| f.shape)
+                .get_index_const(generator, ctx, i)
+                .truncate(generator, ctx, Int32);
+
+            objects.push(AnyObject {
+                ty: ctx.primitives.int32,
+                value: dim.value.as_basic_value_enum(),
+            });
+        }
+
+        TupleObject::from_objects(generator, ctx, objects)
+    }
+
+    /// Create the strides tuple of this ndarray like `np.strides(<ndarray>)`.
+    ///
+    /// The returned integers in the tuple are in int32.
+    pub fn make_strides_tuple<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> TupleObject<'ctx> {
+        // TODO: Return a tuple of SizeT.
+
+        let mut objects = Vec::with_capacity(self.ndims as usize);
+
+        for i in 0..self.ndims {
+            let dim = self
+                .instance
+                .get(generator, ctx, |f| f.strides)
+                .get_index_const(generator, ctx, i)
+                .truncate(generator, ctx, Int32);
+
+            objects.push(AnyObject {
+                ty: ctx.primitives.int32,
+                value: dim.value.as_basic_value_enum(),
+            });
+        }
+
+        TupleObject::from_objects(generator, ctx, objects)
+    }
+
+    /// Create an unsized ndarray to contain `object`.
+    pub fn make_unsized<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        object: AnyObject<'ctx>,
+    ) -> NDArrayObject<'ctx> {
+        // We have to put the value on the stack to get a data pointer.
+        let data = ctx.builder.build_alloca(object.value.get_type(), "make_unsized").unwrap();
+        ctx.builder.build_store(data, object.value).unwrap();
+        let data = Ptr(Int(Byte)).pointer_cast(generator, ctx, data);
+
+        let ndarray = NDArrayObject::alloca(generator, ctx, object.ty, 0);
+        ndarray.instance.set(ctx, |f| f.data, data);
+        ndarray
+    }
+    /// Check if this `NDArray` can be used as an `out` ndarray for an operation.
+    ///
+    /// Raise an exception if the shapes do not match.
+    pub fn assert_can_be_written_by_out<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        out_ndims: u64,
+        out_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) {
+        let ndarray_ndims = self.ndims_llvm(generator, ctx.ctx);
+        let ndarray_shape = self.instance.get(generator, ctx, |f| f.shape);
+
+        let output_ndims = Int(SizeT).const_int(generator, ctx.ctx, out_ndims);
+        let output_shape = out_shape;
+
+        call_nac3_ndarray_util_assert_output_shape_same(
+            generator,
+            ctx,
+            ndarray_ndims,
+            ndarray_shape,
+            output_ndims,
+            output_shape,
+        );
+    }
+}
+
+/// A convenience enum for implementing functions that acts on scalars or ndarrays or both.
+#[derive(Debug, Clone, Copy)]
+pub enum ScalarOrNDArray<'ctx> {
+    Scalar(AnyObject<'ctx>),
+    NDArray(NDArrayObject<'ctx>),
+}
+
+impl<'ctx> TryFrom<&ScalarOrNDArray<'ctx>> for AnyObject<'ctx> {
+    type Error = ();
+
+    fn try_from(value: &ScalarOrNDArray<'ctx>) -> Result<Self, Self::Error> {
+        match value {
+            ScalarOrNDArray::Scalar(scalar) => Ok(*scalar),
+            ScalarOrNDArray::NDArray(_ndarray) => Err(()),
+        }
+    }
+}
+
+impl<'ctx> TryFrom<&ScalarOrNDArray<'ctx>> for NDArrayObject<'ctx> {
+    type Error = ();
+
+    fn try_from(value: &ScalarOrNDArray<'ctx>) -> Result<Self, Self::Error> {
+        match value {
+            ScalarOrNDArray::Scalar(_scalar) => Err(()),
+            ScalarOrNDArray::NDArray(ndarray) => Ok(*ndarray),
+        }
+    }
+}
+
+impl<'ctx> ScalarOrNDArray<'ctx> {
+    /// Split on `object` either into a scalar or an ndarray.
+    ///
+    /// If `object` is an ndarray, [`ScalarOrNDArray::NDArray`].
+    ///
+    /// For everything else, it is wrapped with [`ScalarOrNDArray::Scalar`].
+    pub fn split_object<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        object: AnyObject<'ctx>,
+    ) -> ScalarOrNDArray<'ctx> {
+        if NDArrayObject::is_instance(ctx, object) {
+            let ndarray = NDArrayObject::from_object(generator, ctx, object);
+            ScalarOrNDArray::NDArray(ndarray)
+        } else {
+            ScalarOrNDArray::Scalar(object)
+        }
+    }
+
+    /// Get the underlying [`BasicValueEnum<'ctx>`] of this [`ScalarOrNDArray`].
+    #[must_use]
+    pub fn to_basic_value_enum(self) -> BasicValueEnum<'ctx> {
+        match self {
+            ScalarOrNDArray::Scalar(scalar) => scalar.value,
+            ScalarOrNDArray::NDArray(ndarray) => ndarray.instance.value.as_basic_value_enum(),
+        }
+    }
+
+    /// Convert this [`ScalarOrNDArray`] to an ndarray - behaves like `np.asarray`.
+    /// - If this is an ndarray, the ndarray is returned.
+    /// - If this is a scalar, this function returns new ndarray created with [`NDArrayObject::make_unsized`].
+    pub fn to_ndarray<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> NDArrayObject<'ctx> {
+        match self {
+            ScalarOrNDArray::NDArray(ndarray) => *ndarray,
+            ScalarOrNDArray::Scalar(scalar) => NDArrayObject::make_unsized(generator, ctx, *scalar),
+        }
+    }
+
+    /// Get the dtype of the ndarray created if this were called with [`ScalarOrNDArray::to_ndarray`].
+    #[must_use]
+    pub fn get_dtype(&self) -> Type {
+        match self {
+            ScalarOrNDArray::NDArray(ndarray) => ndarray.dtype,
+            ScalarOrNDArray::Scalar(scalar) => scalar.ty,
+        }
+    }
+}
+
+/// An helper enum specifying how a function should produce its output.
+///
+/// Many functions in NumPy has an optional `out` parameter (e.g., `matmul`). If `out` is specified
+/// with an ndarray, the result of a function will be written to `out`. If `out` is not specified, a function will
+/// create a new ndarray and store the result in it.
+#[derive(Debug, Clone, Copy)]
+pub enum NDArrayOut<'ctx> {
+    /// Tell a function should create a new ndarray with the expected element type `dtype`.
+    NewNDArray { dtype: Type },
+    /// Tell a function to write the result to `ndarray`.
+    WriteToNDArray { ndarray: NDArrayObject<'ctx> },
+}
+
+impl<'ctx> NDArrayOut<'ctx> {
+    /// Get the dtype of this output.
+    pub fn get_dtype(&self) -> Type {
+        match self {
+            NDArrayOut::NewNDArray { dtype } => *dtype,
+            NDArrayOut::WriteToNDArray { ndarray } => ndarray.dtype,
+        }
+    }
+}

nac3core/src/codegen/object/ndarray/nditer.rs

@@ -0,0 +1,168 @@
+use inkwell::{types::BasicType, values::PointerValue, AddressSpace};
+
+use crate::codegen::{
+    irrt::{call_nac3_nditer_has_next, call_nac3_nditer_initialize, call_nac3_nditer_next},
+    model::*,
+    object::any::AnyObject,
+    stmt::{gen_for_callback, BreakContinueHooks},
+    CodeGenContext, CodeGenerator,
+};
+
+use super::NDArrayObject;
+
+/// Fields of [`NDIter`]
+pub struct NDIterFields<'ctx, F: FieldTraversal<'ctx>> {
+    pub ndims: F::Out<Int<SizeT>>,
+    pub shape: F::Out<Ptr<Int<SizeT>>>,
+    pub strides: F::Out<Ptr<Int<SizeT>>>,
+
+    pub indices: F::Out<Ptr<Int<SizeT>>>,
+    pub nth: F::Out<Int<SizeT>>,
+    pub element: F::Out<Ptr<Int<Byte>>>,
+
+    pub size: F::Out<Int<SizeT>>,
+}
+
+/// An IRRT helper structure used to iterate through an ndarray.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct NDIter;
+
+impl<'ctx> StructKind<'ctx> for NDIter {
+    type Fields<F: FieldTraversal<'ctx>> = NDIterFields<'ctx, F>;
+
+    fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+        Self::Fields {
+            ndims: traversal.add_auto("ndims"),
+            shape: traversal.add_auto("shape"),
+            strides: traversal.add_auto("strides"),
+
+            indices: traversal.add_auto("indices"),
+            nth: traversal.add_auto("nth"),
+            element: traversal.add_auto("element"),
+
+            size: traversal.add_auto("size"),
+        }
+    }
+}
+
+/// A helper structure containing extra details of an [`NDIter`].
+#[derive(Debug, Clone)]
+pub struct NDIterHandle<'ctx> {
+    instance: Instance<'ctx, Ptr<Struct<NDIter>>>,
+    /// The ndarray this [`NDIter`] to iterating over.
+    ndarray: NDArrayObject<'ctx>,
+    /// The current indices of [`NDIter`].
+    indices: Instance<'ctx, Ptr<Int<SizeT>>>,
+}
+
+impl<'ctx> NDIterHandle<'ctx> {
+    /// Allocate an [`NDIter`] that iterates through an ndarray.
+    pub fn new<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        ndarray: NDArrayObject<'ctx>,
+    ) -> Self {
+        let nditer = Struct(NDIter).alloca(generator, ctx);
+        let ndims = ndarray.ndims_llvm(generator, ctx.ctx);
+
+        // The caller has the responsibility to allocate 'indices' for `NDIter`.
+        let indices = Int(SizeT).array_alloca(generator, ctx, ndims.value);
+        call_nac3_nditer_initialize(generator, ctx, nditer, ndarray.instance, indices);
+
+        NDIterHandle { ndarray, instance: nditer, indices }
+    }
+
+    #[must_use]
+    pub fn has_next<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<Bool>> {
+        call_nac3_nditer_has_next(generator, ctx, self.instance)
+    }
+
+    pub fn next<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) {
+        call_nac3_nditer_next(generator, ctx, self.instance);
+    }
+
+    /// Get pointer to the current element.
+    #[must_use]
+    pub fn get_pointer<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> PointerValue<'ctx> {
+        let elem_ty = ctx.get_llvm_type(generator, self.ndarray.dtype);
+
+        let p = self.instance.get(generator, ctx, |f| f.element);
+        ctx.builder
+            .build_pointer_cast(p.value, elem_ty.ptr_type(AddressSpace::default()), "element")
+            .unwrap()
+    }
+
+    /// Get the value of the current element.
+    #[must_use]
+    pub fn get_scalar<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> AnyObject<'ctx> {
+        let p = self.get_pointer(generator, ctx);
+        let value = ctx.builder.build_load(p, "value").unwrap();
+        AnyObject { ty: self.ndarray.dtype, value }
+    }
+
+    /// Get the index of the current element.
+    #[must_use]
+    pub fn get_index<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<SizeT>> {
+        self.instance.get(generator, ctx, |f| f.nth)
+    }
+
+    /// Get the indices of the current element.
+    #[must_use]
+    pub fn get_indices(&self) -> Instance<'ctx, Ptr<Int<SizeT>>> {
+        self.indices
+    }
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Iterate through every element in the ndarray.
+    ///
+    /// `body` also access to [`BreakContinueHooks`] to short-circuit.
+    pub fn foreach<'a, G, F>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, 'a>,
+        body: F,
+    ) -> Result<(), String>
+    where
+        G: CodeGenerator + ?Sized,
+        F: FnOnce(
+            &mut G,
+            &mut CodeGenContext<'ctx, 'a>,
+            BreakContinueHooks<'ctx>,
+            NDIterHandle<'ctx>,
+        ) -> Result<(), String>,
+    {
+        gen_for_callback(
+            generator,
+            ctx,
+            Some("ndarray_foreach"),
+            |generator, ctx| Ok(NDIterHandle::new(generator, ctx, *self)),
+            |generator, ctx, nditer| Ok(nditer.has_next(generator, ctx).value),
+            |generator, ctx, hooks, nditer| body(generator, ctx, hooks, nditer),
+            |generator, ctx, nditer| {
+                nditer.next(generator, ctx);
+                Ok(())
+            },
+        )
+    }
+}

nac3core/src/codegen/object/ndarray/shape_util.rs

@@ -0,0 +1,105 @@
+use util::gen_for_model;
+
+use crate::{
+    codegen::{
+        model::*,
+        object::{any::AnyObject, list::ListObject, tuple::TupleObject},
+        CodeGenContext, CodeGenerator,
+    },
+    typecheck::typedef::TypeEnum,
+};
+
+/// Parse a NumPy-like "int sequence" input and return the int sequence as an array and its length.
+///
+/// * `sequence` - The `sequence` parameter.
+/// * `sequence_ty` - The typechecker type of `sequence`
+///
+/// The `sequence` argument type may only be one of the following:
+///   1. A list of `int32`;   e.g., `np.empty([600, 800, 3])`
+///   2. A tuple of `int32`;  e.g., `np.empty((600, 800, 3))`
+///   3. A scalar `int32`;    e.g., `np.empty(3)`, this is functionally equivalent to `np.empty([3])`
+///
+/// All `int32` values will be sign-extended to `SizeT`.
+pub fn parse_numpy_int_sequence<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    input_sequence: AnyObject<'ctx>,
+) -> (Instance<'ctx, Int<SizeT>>, Instance<'ctx, Ptr<Int<SizeT>>>) {
+    let zero = Int(SizeT).const_0(generator, ctx.ctx);
+    let one = Int(SizeT).const_1(generator, ctx.ctx);
+
+    // The result `list` to return.
+    match &*ctx.unifier.get_ty(input_sequence.ty) {
+        TypeEnum::TObj { obj_id, .. }
+            if *obj_id == ctx.primitives.list.obj_id(&ctx.unifier).unwrap() =>
+        {
+            // 1. A list of `int32`; e.g., `np.empty([600, 800, 3])`
+
+            // Check `input_sequence`
+            let input_sequence = ListObject::from_object(generator, ctx, input_sequence);
+
+            let len = input_sequence.instance.get(generator, ctx, |f| f.len);
+            let result = Int(SizeT).array_alloca(generator, ctx, len.value);
+
+            // Load all the `int32`s from the input_sequence, cast them to `SizeT`, and store them into `result`
+            gen_for_model(generator, ctx, zero, len, one, |generator, ctx, _hooks, i| {
+                // Load the i-th int32 in the input sequence
+                let int = input_sequence
+                    .instance
+                    .get(generator, ctx, |f| f.items)
+                    .get_index(generator, ctx, i.value)
+                    .value
+                    .into_int_value();
+
+                // Cast to SizeT
+                let int = Int(SizeT).s_extend_or_bit_cast(generator, ctx, int);
+
+                // Store
+                result.set_index(ctx, i.value, int);
+
+                Ok(())
+            })
+            .unwrap();
+
+            (len, result)
+        }
+        TypeEnum::TTuple { .. } => {
+            // 2. A tuple of ints; e.g., `np.empty((600, 800, 3))`
+
+            let input_sequence = TupleObject::from_object(ctx, input_sequence);
+
+            let len = input_sequence.len(generator, ctx);
+
+            let result = Int(SizeT).array_alloca(generator, ctx, len.value);
+
+            for i in 0..input_sequence.num_elements() {
+                // Get the i-th element off of the tuple and load it into `result`.
+                let int = input_sequence.index(ctx, i).value.into_int_value();
+                let int = Int(SizeT).s_extend_or_bit_cast(generator, ctx, int);
+
+                result.set_index_const(ctx, i as u64, int);
+            }
+
+            (len, result)
+        }
+        TypeEnum::TObj { obj_id, .. }
+            if *obj_id == ctx.primitives.int32.obj_id(&ctx.unifier).unwrap() =>
+        {
+            // 3. A scalar int; e.g., `np.empty(3)`, this is functionally equivalent to `np.empty([3])`
+            let input_int = input_sequence.value.into_int_value();
+
+            let len = Int(SizeT).const_1(generator, ctx.ctx);
+            let result = Int(SizeT).array_alloca(generator, ctx, len.value);
+            let int = Int(SizeT).s_extend_or_bit_cast(generator, ctx, input_int);
+
+            // Storing into result[0]
+            result.store(ctx, int);
+
+            (len, result)
+        }
+        _ => panic!(
+            "encountered unknown sequence type: {}",
+            ctx.unifier.stringify(input_sequence.ty)
+        ),
+    }
+}

nac3core/src/codegen/object/ndarray/view.rs

@@ -0,0 +1,119 @@
+use crate::codegen::{
+    irrt::{call_nac3_ndarray_reshape_resolve_and_check_new_shape, call_nac3_ndarray_transpose},
+    model::*,
+    CodeGenContext, CodeGenerator,
+};
+
+use super::{indexing::RustNDIndex, NDArrayObject};
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Make sure the ndarray is at least `ndmin`-dimensional.
+    ///
+    /// If this ndarray's `ndims` is less than `ndmin`, a view is created on this with 1s prepended to the shape.
+    /// If this ndarray's `ndims` is not less than `ndmin`, this function does nothing and return this ndarray.
+    #[must_use]
+    pub fn atleast_nd<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        ndmin: u64,
+    ) -> Self {
+        if self.ndims < ndmin {
+            // return this_ndarray[np.newaxis, np.newaxis, and more, ...]
+            let mut indices = vec![];
+            for _ in self.ndims..ndmin {
+                indices.push(RustNDIndex::NewAxis);
+            }
+            indices.push(RustNDIndex::Ellipsis);
+            self.index(generator, ctx, &indices)
+        } else {
+            *self
+        }
+    }
+
+    /// Create a reshaped view on this ndarray like `np.reshape()`.
+    ///
+    /// If there is a `-1` in `new_shape`, it will be resolved; `new_shape` would **NOT** be modified as a result.
+    ///
+    /// If reshape without copying is impossible, this function will allocate a new ndarray and copy contents.
+    ///
+    /// * `new_ndims` - The number of dimensions of `new_shape` as a [`Type`].
+    /// * `new_shape` - The target shape to do `np.reshape()`.
+    #[must_use]
+    pub fn reshape_or_copy<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        new_ndims: u64,
+        new_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) -> Self {
+        // TODO: The current criterion for whether to do a full copy or not is by checking `is_c_contiguous`,
+        // but this is not optimal - there are cases when the ndarray is not contiguous but could be reshaped
+        // without copying data. Look into how numpy does it.
+
+        let current_bb = ctx.builder.get_insert_block().unwrap();
+        let then_bb = ctx.ctx.insert_basic_block_after(current_bb, "then_bb");
+        let else_bb = ctx.ctx.insert_basic_block_after(then_bb, "else_bb");
+        let end_bb = ctx.ctx.insert_basic_block_after(else_bb, "end_bb");
+
+        let dst_ndarray = NDArrayObject::alloca(generator, ctx, self.dtype, new_ndims);
+        dst_ndarray.copy_shape_from_array(generator, ctx, new_shape);
+
+        // Reolsve negative indices
+        let size = self.size(generator, ctx);
+        let dst_ndims = dst_ndarray.ndims_llvm(generator, ctx.ctx);
+        let dst_shape = dst_ndarray.instance.get(generator, ctx, |f| f.shape);
+        call_nac3_ndarray_reshape_resolve_and_check_new_shape(
+            generator, ctx, size, dst_ndims, dst_shape,
+        );
+
+        let is_c_contiguous = self.is_c_contiguous(generator, ctx);
+        ctx.builder.build_conditional_branch(is_c_contiguous.value, then_bb, else_bb).unwrap();
+
+        // Inserting into then_bb: reshape is possible without copying
+        ctx.builder.position_at_end(then_bb);
+        dst_ndarray.set_strides_contiguous(generator, ctx);
+        dst_ndarray.instance.set(ctx, |f| f.data, self.instance.get(generator, ctx, |f| f.data));
+        ctx.builder.build_unconditional_branch(end_bb).unwrap();
+
+        // Inserting into else_bb: reshape is impossible without copying
+        ctx.builder.position_at_end(else_bb);
+        dst_ndarray.create_data(generator, ctx);
+        dst_ndarray.copy_data_from(generator, ctx, *self);
+        ctx.builder.build_unconditional_branch(end_bb).unwrap();
+
+        // Reposition for continuation
+        ctx.builder.position_at_end(end_bb);
+
+        dst_ndarray
+    }
+
+    /// Create a transposed view on this ndarray like `np.transpose(<ndarray>, <axes> = None)`.
+    /// * `axes` - If specified, should be an array of the permutation (negative indices are **allowed**).
+    #[must_use]
+    pub fn transpose<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        axes: Option<Instance<'ctx, Ptr<Int<SizeT>>>>,
+    ) -> Self {
+        // Define models
+        let transposed_ndarray = NDArrayObject::alloca(generator, ctx, self.dtype, self.ndims);
+
+        let num_axes = self.ndims_llvm(generator, ctx.ctx);
+
+        // `axes = nullptr` if `axes` is unspecified.
+        let axes = axes.unwrap_or_else(|| Ptr(Int(SizeT)).nullptr(generator, ctx.ctx));
+
+        call_nac3_ndarray_transpose(
+            generator,
+            ctx,
+            self.instance,
+            transposed_ndarray.instance,
+            num_axes,
+            axes,
+        );
+
+        transposed_ndarray
+    }
+}

nac3core/src/codegen/object/tuple.rs

@@ -0,0 +1,104 @@
+use inkwell::values::StructValue;
+use itertools::Itertools;
+
+use crate::{
+    codegen::{model::*, CodeGenContext, CodeGenerator},
+    typecheck::typedef::{Type, TypeEnum},
+};
+
+use super::any::AnyObject;
+
+/// A NAC3 tuple object.
+///
+/// NOTE: This struct has no copy trait.
+#[derive(Debug, Clone)]
+pub struct TupleObject<'ctx> {
+    /// The type of the tuple.
+    pub tys: Vec<Type>,
+    /// The underlying LLVM struct value of this tuple.
+    pub value: StructValue<'ctx>,
+}
+
+impl<'ctx> TupleObject<'ctx> {
+    pub fn from_object(ctx: &mut CodeGenContext<'ctx, '_>, object: AnyObject<'ctx>) -> Self {
+        // TODO: Keep `is_vararg_ctx` from TTuple?
+
+        // Sanity check on object type.
+        let TypeEnum::TTuple { ty: tys, .. } = &*ctx.unifier.get_ty(object.ty) else {
+            panic!(
+                "Expected type to be a TypeEnum::TTuple, got {}",
+                ctx.unifier.stringify(object.ty)
+            );
+        };
+
+        // Check number of fields
+        let value = object.value.into_struct_value();
+        let value_num_fields = value.get_type().count_fields() as usize;
+        assert!(
+            value_num_fields == tys.len(),
+            "Tuple type has {} item(s), but the LLVM struct value has {} field(s)",
+            tys.len(),
+            value_num_fields
+        );
+
+        TupleObject { tys: tys.clone(), value }
+    }
+
+    /// Check if an object can be converted into a tuple.
+    pub fn is_instance(ctx: &mut CodeGenContext<'ctx, '_>, object: AnyObject<'ctx>) -> bool {
+        matches!(&*ctx.unifier.get_ty(object.ty), TypeEnum::TTuple { .. })
+    }
+
+    /// Convenience function. Create a [`TupleObject`] from an iterator of objects.
+    pub fn from_objects<I, G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        objects: I,
+    ) -> Self
+    where
+        I: IntoIterator<Item = AnyObject<'ctx>>,
+    {
+        let (values, tys): (Vec<_>, Vec<_>) =
+            objects.into_iter().map(|object| (object.value, object.ty)).unzip();
+
+        let llvm_tys = tys.iter().map(|ty| ctx.get_llvm_type(generator, *ty)).collect_vec();
+        let llvm_tuple_ty = ctx.ctx.struct_type(&llvm_tys, false);
+
+        let pllvm_tuple = ctx.builder.build_alloca(llvm_tuple_ty, "tuple").unwrap();
+        for (i, val) in values.into_iter().enumerate() {
+            let pval = ctx.builder.build_struct_gep(pllvm_tuple, i as u32, "value").unwrap();
+            ctx.builder.build_store(pval, val).unwrap();
+        }
+
+        let value = ctx.builder.build_load(pllvm_tuple, "").unwrap().into_struct_value();
+        TupleObject { tys, value }
+    }
+
+    #[must_use]
+    pub fn num_elements(&self) -> usize {
+        self.tys.len()
+    }
+
+    /// Get the `len()` of this tuple.
+    #[must_use]
+    pub fn len<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<SizeT>> {
+        Int(SizeT).const_int(generator, ctx.ctx, self.num_elements() as u64)
+    }
+
+    /// Get the `i`-th (0-based) object in this tuple.
+    pub fn index(&self, ctx: &mut CodeGenContext<'ctx, '_>, i: usize) -> AnyObject<'ctx> {
+        assert!(
+            i < self.num_elements(),
+            "Tuple object with length {} have index {i}",
+            self.num_elements()
+        );
+
+        let value = ctx.builder.build_extract_value(self.value, i as u32, "tuple[{i}]").unwrap();
+        let ty = self.tys[i];
+        AnyObject { ty, value }
+    }
+}

nac3core/src/codegen/stmt.rs

@@ -2,6 +2,12 @@ use super::{
     super::symbol_resolver::ValueEnum,
     expr::destructure_range,
     irrt::{handle_slice_indices, list_slice_assignment},
+    object::{
+        any::AnyObject,
+        ndarray::{
+            indexing::util::gen_ndarray_subscript_ndindices, NDArrayObject, ScalarOrNDArray,
+        },
+    },
     CodeGenContext, CodeGenerator,
 };
 use crate::{
@@ -401,7 +407,47 @@ pub fn gen_setitem<'ctx, G: CodeGenerator>(
             if *obj_id == ctx.primitives.ndarray.obj_id(&ctx.unifier).unwrap() =>
         {
             // Handle NDArray item assignment
-            todo!("ndarray subscript assignment is not yet implemented");
+            // Process target
+            let target = generator
+                .gen_expr(ctx, target)?
+                .unwrap()
+                .to_basic_value_enum(ctx, generator, target_ty)?;
+            let target = AnyObject { value: target, ty: target_ty };
+
+            // Process key
+            let key = gen_ndarray_subscript_ndindices(generator, ctx, key)?;
+
+            // Process value
+            let value = value.to_basic_value_enum(ctx, generator, value_ty)?;
+            let value = AnyObject { value, ty: value_ty };
+
+            /*
+                Reference code:
+                ```python
+                target = target[key]
+                value = np.asarray(value)
+
+                shape = np.broadcast_shape((target, value))
+
+                target = np.broadcast_to(target, shape)
+                value = np.broadcast_to(value, shape)
+
+                ...and finally copy 1-1 from value to target.
+                ```
+            */
+
+            let target = NDArrayObject::from_object(generator, ctx, target);
+            let target = target.index(generator, ctx, &key);
+
+            let value =
+                ScalarOrNDArray::split_object(generator, ctx, value).to_ndarray(generator, ctx);
+
+            let broadcast_result = NDArrayObject::broadcast(generator, ctx, &[target, value]);
+
+            let target = broadcast_result.ndarrays[0];
+            let value = broadcast_result.ndarrays[1];
+
+            target.copy_data_from(generator, ctx, value);
         }
         _ => {
             panic!("encountered unknown target type: {}", ctx.unifier.stringify(target_ty));

nac3core/src/toplevel/builtins.rs

@@ -1,6 +1,6 @@
 use std::iter::once;
 
-use helper::{debug_assert_prim_is_allowed, make_exception_fields, PrimDefDetails};
+use helper::{debug_assert_prim_is_allowed, extract_ndims, make_exception_fields, PrimDefDetails};
 use indexmap::IndexMap;
 use inkwell::{
     attributes::{Attribute, AttributeLoc},
@@ -9,6 +9,7 @@ use inkwell::{
     IntPredicate,
 };
 use itertools::Either;
+use numpy::unpack_ndarray_var_tys;
 use strum::IntoEnumIterator;
 
 use crate::{
@@ -16,6 +17,10 @@ use crate::{
         builtin_fns,
         classes::{ProxyValue, RangeValue},
         numpy::*,
+        object::{
+            any::AnyObject,
+            ndarray::{shape_util::parse_numpy_int_sequence, NDArrayObject},
+        },
         stmt::exn_constructor,
     },
     symbol_resolver::SymbolValue,
@@ -511,6 +516,14 @@ impl<'a> BuiltinBuilder<'a> {
             | PrimDef::FunNpEye
             | PrimDef::FunNpIdentity => self.build_ndarray_other_factory_function(prim),
 
+            PrimDef::FunNpShape | PrimDef::FunNpStrides => {
+                self.build_ndarray_property_getter_function(prim)
+            }
+
+            PrimDef::FunNpBroadcastTo | PrimDef::FunNpTranspose | PrimDef::FunNpReshape => {
+                self.build_ndarray_view_function(prim)
+            }
+
             PrimDef::FunStr => self.build_str_function(),
 
             PrimDef::FunFloor | PrimDef::FunFloor64 | PrimDef::FunCeil | PrimDef::FunCeil64 => {
@@ -576,10 +589,6 @@ impl<'a> BuiltinBuilder<'a> {
             | PrimDef::FunNpHypot
             | PrimDef::FunNpNextAfter => self.build_np_2ary_function(prim),
 
-            PrimDef::FunNpTranspose | PrimDef::FunNpReshape => {
-                self.build_np_sp_ndarray_function(prim)
-            }
-
             PrimDef::FunNpDot
             | PrimDef::FunNpLinalgCholesky
             | PrimDef::FunNpLinalgQr
@@ -1385,6 +1394,146 @@ impl<'a> BuiltinBuilder<'a> {
         }
     }
 
+    fn build_ndarray_property_getter_function(&mut self, prim: PrimDef) -> TopLevelDef {
+        debug_assert_prim_is_allowed(prim, &[PrimDef::FunNpShape, PrimDef::FunNpStrides]);
+
+        let mut var_map = self.num_var_map.clone();
+        var_map.insert(self.ndarray_dtype_tvar.id, self.ndarray_dtype_tvar.ty);
+
+        let in_ndarray_ty = self.unifier.get_fresh_var_with_range(
+            &[self.primitives.ndarray],
+            Some("T".into()),
+            None,
+        );
+
+        match prim {
+            PrimDef::FunNpShape | PrimDef::FunNpStrides => {
+                // The fnuction signatures of `np_shape` an `np_size` are the same.
+
+                // The return type is a tuple of variable length depending on the ndims of the input ndarray.
+                let ret_ty = self.unifier.get_dummy_var().ty; // Handled by special folding
+
+                create_fn_by_codegen(
+                    self.unifier,
+                    &VarMap::new(),
+                    prim.name(),
+                    ret_ty,
+                    &[(in_ndarray_ty.ty, "a")],
+                    Box::new(move |ctx, obj, fun, args, generator| {
+                        assert!(obj.is_none());
+                        assert_eq!(args.len(), 1);
+
+                        let ndarray_ty = fun.0.args[0].ty;
+                        let ndarray =
+                            args[0].1.clone().to_basic_value_enum(ctx, generator, ndarray_ty)?;
+
+                        let ndarray = AnyObject { ty: ndarray_ty, value: ndarray };
+                        let ndarray = NDArrayObject::from_object(generator, ctx, ndarray);
+
+                        let result_tuple = match prim {
+                            PrimDef::FunNpShape => ndarray.make_shape_tuple(generator, ctx),
+                            PrimDef::FunNpStrides => ndarray.make_strides_tuple(generator, ctx),
+                            _ => unreachable!(),
+                        };
+
+                        Ok(Some(result_tuple.value.as_basic_value_enum()))
+                    }),
+                )
+            }
+            _ => unreachable!(),
+        }
+    }
+
+    /// Build np/sp functions that take as input `NDArray` only
+    fn build_ndarray_view_function(&mut self, prim: PrimDef) -> TopLevelDef {
+        debug_assert_prim_is_allowed(
+            prim,
+            &[PrimDef::FunNpBroadcastTo, PrimDef::FunNpTranspose, PrimDef::FunNpReshape],
+        );
+
+        let in_ndarray_ty = self.unifier.get_fresh_var_with_range(
+            &[self.primitives.ndarray],
+            Some("T".into()),
+            None,
+        );
+
+        match prim {
+            PrimDef::FunNpTranspose => {
+                create_fn_by_codegen(
+                    self.unifier,
+                    &VarMap::new(),
+                    prim.name(),
+                    in_ndarray_ty.ty,
+                    &[(in_ndarray_ty.ty, "x")],
+                    Box::new(move |ctx, _, fun, args, generator| {
+                        let arg_ty = fun.0.args[0].ty;
+                        let arg_val =
+                            args[0].1.clone().to_basic_value_enum(ctx, generator, arg_ty)?;
+
+                        let arg = AnyObject { ty: arg_ty, value: arg_val };
+                        let ndarray = NDArrayObject::from_object(generator, ctx, arg);
+
+                        let ndarray = ndarray.transpose(generator, ctx, None); // TODO: Add axes argument
+                        Ok(Some(ndarray.instance.value.as_basic_value_enum()))
+                    }),
+                )
+            }
+
+            // NOTE: on `ndarray_factory_fn_shape_arg_tvar` and
+            // the `param_ty` for `create_fn_by_codegen`.
+            //
+            // Similar to `build_ndarray_from_shape_factory_function` we delegate the responsibility of typechecking
+            // to [`typecheck::type_inferencer::Inferencer::fold_numpy_function_call_shape_argument`],
+            // and use a dummy [`TypeVar`] `ndarray_factory_fn_shape_arg_tvar` as a placeholder for `param_ty`.
+            PrimDef::FunNpBroadcastTo | PrimDef::FunNpReshape => {
+                let ret_ty = self.unifier.get_dummy_var().ty; // Handled by special holding
+
+                create_fn_by_codegen(
+                    self.unifier,
+                    &VarMap::new(),
+                    prim.name(),
+                    ret_ty,
+                    &[
+                        (in_ndarray_ty.ty, "x"),
+                        (self.ndarray_factory_fn_shape_arg_tvar.ty, "shape"), // Handled by special folding
+                    ],
+                    Box::new(move |ctx, _, fun, args, generator| {
+                        let ndarray_ty = fun.0.args[0].ty;
+                        let ndarray_val =
+                            args[0].1.clone().to_basic_value_enum(ctx, generator, ndarray_ty)?;
+
+                        let shape_ty = fun.0.args[1].ty;
+                        let shape_val =
+                            args[1].1.clone().to_basic_value_enum(ctx, generator, shape_ty)?;
+
+                        let ndarray = AnyObject { value: ndarray_val, ty: ndarray_ty };
+                        let ndarray = NDArrayObject::from_object(generator, ctx, ndarray);
+
+                        let shape = AnyObject { value: shape_val, ty: shape_ty };
+                        let (_, shape) = parse_numpy_int_sequence(generator, ctx, shape);
+
+                        // The ndims after reshaping is gotten from the return type of the call.
+                        let (_, ndims) = unpack_ndarray_var_tys(&mut ctx.unifier, fun.0.ret);
+                        let ndims = extract_ndims(&ctx.unifier, ndims);
+
+                        let new_ndarray = match prim {
+                            PrimDef::FunNpBroadcastTo => {
+                                ndarray.broadcast_to(generator, ctx, ndims, shape)
+                            }
+                            PrimDef::FunNpReshape => {
+                                ndarray.reshape_or_copy(generator, ctx, ndims, shape)
+                            }
+                            _ => unreachable!(),
+                        };
+                        Ok(Some(new_ndarray.instance.value.as_basic_value_enum()))
+                    }),
+                )
+            }
+
+            _ => unreachable!(),
+        }
+    }
+
     /// Build the `str()` function.
     fn build_str_function(&mut self) -> TopLevelDef {
         let prim = PrimDef::FunStr;
@@ -1872,57 +2021,6 @@ impl<'a> BuiltinBuilder<'a> {
         }
     }
 
-    /// Build np/sp functions that take as input `NDArray` only
-    fn build_np_sp_ndarray_function(&mut self, prim: PrimDef) -> TopLevelDef {
-        debug_assert_prim_is_allowed(prim, &[PrimDef::FunNpTranspose, PrimDef::FunNpReshape]);
-
-        match prim {
-            PrimDef::FunNpTranspose => {
-                let ndarray_ty = self.unifier.get_fresh_var_with_range(
-                    &[self.ndarray_num_ty],
-                    Some("T".into()),
-                    None,
-                );
-                create_fn_by_codegen(
-                    self.unifier,
-                    &into_var_map([ndarray_ty]),
-                    prim.name(),
-                    ndarray_ty.ty,
-                    &[(ndarray_ty.ty, "x")],
-                    Box::new(move |ctx, _, fun, args, generator| {
-                        let arg_ty = fun.0.args[0].ty;
-                        let arg_val =
-                            args[0].1.clone().to_basic_value_enum(ctx, generator, arg_ty)?;
-                        Ok(Some(ndarray_transpose(generator, ctx, (arg_ty, arg_val))?))
-                    }),
-                )
-            }
-
-            // NOTE: on `ndarray_factory_fn_shape_arg_tvar` and
-            // the `param_ty` for `create_fn_by_codegen`.
-            //
-            // Similar to `build_ndarray_from_shape_factory_function` we delegate the responsibility of typechecking
-            // to [`typecheck::type_inferencer::Inferencer::fold_numpy_function_call_shape_argument`],
-            // and use a dummy [`TypeVar`] `ndarray_factory_fn_shape_arg_tvar` as a placeholder for `param_ty`.
-            PrimDef::FunNpReshape => create_fn_by_codegen(
-                self.unifier,
-                &VarMap::new(),
-                prim.name(),
-                self.ndarray_num_ty,
-                &[(self.ndarray_num_ty, "x"), (self.ndarray_factory_fn_shape_arg_tvar.ty, "shape")],
-                Box::new(move |ctx, _, fun, args, generator| {
-                    let x1_ty = fun.0.args[0].ty;
-                    let x1_val = args[0].1.clone().to_basic_value_enum(ctx, generator, x1_ty)?;
-                    let x2_ty = fun.0.args[1].ty;
-                    let x2_val = args[1].1.clone().to_basic_value_enum(ctx, generator, x2_ty)?;
-                    Ok(Some(ndarray_reshape(generator, ctx, (x1_ty, x1_val), (x2_ty, x2_val))?))
-                }),
-            ),
-
-            _ => unreachable!(),
-        }
-    }
-
     /// Build `np_linalg` and `sp_linalg` functions
     ///
     /// The input to these functions must be floating point `NDArray`
@@ -1954,10 +2052,12 @@ impl<'a> BuiltinBuilder<'a> {
                 Box::new(move |ctx, _, fun, args, generator| {
                     let x1_ty = fun.0.args[0].ty;
                     let x1_val = args[0].1.clone().to_basic_value_enum(ctx, generator, x1_ty)?;
+
                     let x2_ty = fun.0.args[1].ty;
                     let x2_val = args[1].1.clone().to_basic_value_enum(ctx, generator, x2_ty)?;
 
-                    Ok(Some(ndarray_dot(generator, ctx, (x1_ty, x1_val), (x2_ty, x2_val))?))
+                    let result = ndarray_dot(generator, ctx, (x1_ty, x1_val), (x2_ty, x2_val))?;
+                    Ok(Some(result))
                 }),
             ),
 

nac3core/src/toplevel/helper.rs

@@ -52,6 +52,15 @@ pub enum PrimDef {
     FunNpEye,
     FunNpIdentity,
 
+    // NumPy ndarray property getters
+    FunNpShape,
+    FunNpStrides,
+
+    // NumPy ndarray view functions
+    FunNpBroadcastTo,
+    FunNpTranspose,
+    FunNpReshape,
+
     // Miscellaneous NumPy & SciPy functions
     FunNpRound,
     FunNpFloor,
@@ -99,8 +108,6 @@ pub enum PrimDef {
     FunNpLdExp,
     FunNpHypot,
     FunNpNextAfter,
-    FunNpTranspose,
-    FunNpReshape,
 
     // Linalg functions
     FunNpDot,
@@ -238,6 +245,15 @@ impl PrimDef {
             PrimDef::FunNpEye => fun("np_eye", None),
             PrimDef::FunNpIdentity => fun("np_identity", None),
 
+            // NumPy NDArray property getters,
+            PrimDef::FunNpShape => fun("np_shape", None),
+            PrimDef::FunNpStrides => fun("np_strides", None),
+
+            // NumPy NDArray view functions
+            PrimDef::FunNpBroadcastTo => fun("np_broadcast_to", None),
+            PrimDef::FunNpTranspose => fun("np_transpose", None),
+            PrimDef::FunNpReshape => fun("np_reshape", None),
+
             // Miscellaneous NumPy & SciPy functions
             PrimDef::FunNpRound => fun("np_round", None),
             PrimDef::FunNpFloor => fun("np_floor", None),
@@ -285,8 +301,6 @@ impl PrimDef {
             PrimDef::FunNpLdExp => fun("np_ldexp", None),
             PrimDef::FunNpHypot => fun("np_hypot", None),
             PrimDef::FunNpNextAfter => fun("np_nextafter", None),
-            PrimDef::FunNpTranspose => fun("np_transpose", None),
-            PrimDef::FunNpReshape => fun("np_reshape", None),
 
             // Linalg functions
             PrimDef::FunNpDot => fun("np_dot", None),

nac3core/src/typecheck/magic_methods.rs

@@ -1,5 +1,5 @@
 use crate::symbol_resolver::SymbolValue;
-use crate::toplevel::helper::PrimDef;
+use crate::toplevel::helper::{extract_ndims, PrimDef};
 use crate::toplevel::numpy::{make_ndarray_ty, unpack_ndarray_var_tys};
 use crate::typecheck::{
     type_inferencer::*,
@@ -520,36 +520,41 @@ pub fn typeof_binop(
         }
 
         Operator::MatMult => {
-            let (_, lhs_ndims) = unpack_ndarray_var_tys(unifier, lhs);
-            let lhs_ndims = match &*unifier.get_ty_immutable(lhs_ndims) {
-                TypeEnum::TLiteral { values, .. } => {
-                    assert_eq!(values.len(), 1);
-                    u64::try_from(values[0].clone()).unwrap()
+            let (lhs_dtype, lhs_ndims) = unpack_ndarray_var_tys(unifier, lhs);
+            let lhs_ndims = extract_ndims(unifier, lhs_ndims);
+
+            let (rhs_dtype, rhs_ndims) = unpack_ndarray_var_tys(unifier, rhs);
+            let rhs_ndims = extract_ndims(unifier, rhs_ndims);
+
+            if !(unifier.unioned(lhs_dtype, primitives.float)
+                && unifier.unioned(rhs_dtype, primitives.float))
+            {
+                return Err(format!(
+                    "ndarray.__matmul__ only supports float64 operations, but LHS has type {} and RHS has type {}",
+                    unifier.stringify(lhs),
+                    unifier.stringify(rhs)
+                ));
+            }
+
+            let result_ndims = match (lhs_ndims, rhs_ndims) {
+                (0, _) | (_, 0) => {
+                    return Err(
+                        "ndarray.__matmul__ does not allow unsized ndarray input".to_string()
+                    )
                 }
-                _ => unreachable!(),
-            };
-            let (_, rhs_ndims) = unpack_ndarray_var_tys(unifier, rhs);
-            let rhs_ndims = match &*unifier.get_ty_immutable(rhs_ndims) {
-                TypeEnum::TLiteral { values, .. } => {
-                    assert_eq!(values.len(), 1);
-                    u64::try_from(values[0].clone()).unwrap()
-                }
-                _ => unreachable!(),
+                (1, 1) => 0,
+                (1, _) => rhs_ndims - 1,
+                (_, 1) => lhs_ndims - 1,
+                (m, n) => max(m, n),
             };
 
-            match (lhs_ndims, rhs_ndims) {
-                (2, 2) => typeof_ndarray_broadcast(unifier, primitives, lhs, rhs)?,
-                (lhs, rhs) if lhs == 0 || rhs == 0 => {
-                    return Err(format!(
-                    "Input operand {} does not have enough dimensions (has {lhs}, requires {rhs})",
-                    u8::from(rhs == 0)
-                ))
-                }
-                (lhs, rhs) => {
-                    return Err(format!(
-                        "ndarray.__matmul__ on {lhs}D and {rhs}D operands not supported"
-                    ))
-                }
+            if result_ndims == 0 {
+                // If the result is unsized, NumPy returns a scalar.
+                primitives.float
+            } else {
+                let result_ndims_ty =
+                    unifier.get_fresh_literal(vec![SymbolValue::U64(result_ndims)], None);
+                make_ndarray_ty(unifier, primitives, Some(primitives.float), Some(result_ndims_ty))
             }
         }
 
@@ -748,7 +753,7 @@ pub fn set_primitives_magic_methods(store: &PrimitiveStore, unifier: &mut Unifie
     impl_div(unifier, store, ndarray_t, &[ndarray_t, ndarray_dtype_t], None);
     impl_floordiv(unifier, store, ndarray_t, &[ndarray_unsized_t, ndarray_unsized_dtype_t], None);
     impl_mod(unifier, store, ndarray_t, &[ndarray_unsized_t, ndarray_unsized_dtype_t], None);
-    impl_matmul(unifier, store, ndarray_t, &[ndarray_t], Some(ndarray_t));
+    impl_matmul(unifier, store, ndarray_t, &[ndarray_unsized_t], None);
     impl_sign(unifier, store, ndarray_t, Some(ndarray_t));
     impl_invert(unifier, store, ndarray_t, Some(ndarray_t));
     impl_eq(unifier, store, ndarray_t, &[ndarray_unsized_t, ndarray_unsized_dtype_t], None);

nac3core/src/typecheck/type_inferencer/mod.rs

@@ -1,7 +1,7 @@
 use std::cmp::max;
 use std::collections::{HashMap, HashSet};
 use std::convert::{From, TryInto};
-use std::iter::once;
+use std::iter::{self, once};
 use std::{cell::RefCell, sync::Arc};
 
 use super::{
@@ -1181,6 +1181,45 @@ impl<'a> Inferencer<'a> {
             }));
         }
 
+        if ["np_shape".into(), "np_strides".into()].contains(id) && args.len() == 1 {
+            let ndarray = self.fold_expr(args.remove(0))?;
+
+            let ndims = arraylike_get_ndims(self.unifier, ndarray.custom.unwrap());
+
+            // Make a tuple of size `ndims` full of int32 (TODO: Make it usize)
+            let ret_ty = TypeEnum::TTuple {
+                ty: iter::repeat(self.primitives.int32).take(ndims as usize).collect_vec(),
+                is_vararg_ctx: false,
+            };
+            let ret_ty = self.unifier.add_ty(ret_ty);
+
+            let func_ty = TypeEnum::TFunc(FunSignature {
+                args: vec![FuncArg {
+                    name: "a".into(),
+                    default_value: None,
+                    ty: ndarray.custom.unwrap(),
+                    is_vararg: false,
+                }],
+                ret: ret_ty,
+                vars: VarMap::new(),
+            });
+            let func_ty = self.unifier.add_ty(func_ty);
+
+            return Ok(Some(Located {
+                location,
+                custom: Some(ret_ty),
+                node: ExprKind::Call {
+                    func: Box::new(Located {
+                        custom: Some(func_ty),
+                        location: func.location,
+                        node: ExprKind::Name { id: *id, ctx: *ctx },
+                    }),
+                    args: vec![ndarray],
+                    keywords: vec![],
+                },
+            }));
+        }
+
         if id == &"np_dot".into() {
             let arg0 = self.fold_expr(args.remove(0))?;
             let arg1 = self.fold_expr(args.remove(0))?;

nac3standalone/demo/interpret_demo.py

@@ -179,6 +179,15 @@ def patch(module):
     module.np_identity = np.identity
     module.np_array = np.array
 
+    # NumPy NDArray view functions
+    module.np_broadcast_to = np.broadcast_to
+    module.np_transpose = np.transpose
+    module.np_reshape = np.reshape
+
+    # NumPy NDArray property getters
+    module.np_shape = np.shape
+    module.np_strides = lambda ndarray: ndarray.strides
+
     # NumPy Math functions
     module.np_isnan = np.isnan
     module.np_isinf = np.isinf
@@ -218,8 +227,6 @@ def patch(module):
     module.np_ldexp = np.ldexp
     module.np_hypot = np.hypot
     module.np_nextafter = np.nextafter
-    module.np_transpose = np.transpose
-    module.np_reshape = np.reshape
 
     # SciPy Math functions
     module.sp_spec_erf = special.erf

nac3standalone/demo/src/ndarray.py

@@ -68,6 +68,19 @@ def output_ndarray_float_2(n: ndarray[float, Literal[2]]):
         for c in range(len(n[r])):
             output_float64(n[r][c])
 
+def output_ndarray_float_3(n: ndarray[float, Literal[3]]):
+    for d in range(len(n)):
+        for r in range(len(n[d])):
+            for c in range(len(n[d][r])):
+                output_float64(n[d][r][c])
+
+def output_ndarray_float_4(n: ndarray[float, Literal[4]]):
+    for x in range(len(n)):
+        for y in range(len(n[x])):
+            for z in range(len(n[x][y])):
+                for w in range(len(n[x][y][z])):
+                    output_float64(n[x][y][z][w])
+
 def consume_ndarray_1(n: ndarray[float, Literal[1]]):
     pass
 
@@ -186,6 +199,27 @@ def test_ndarray_nd_idx():
     output_float64(x[1, 0])
     output_float64(x[1, 1])
 
+def test_ndarray_subscript_assignment():
+    xs = np_array([[11111.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 99999.0]])
+    output_ndarray_float_2(xs)
+
+    output_int32(np_shape(xs[1:, 1::])[0])
+    output_int32(np_shape(xs[1:, 1::])[1])
+    xs[1:, 1::] = 77.0
+    output_ndarray_float_2(xs)
+
+    xs[0, -1] = 88.0
+    output_ndarray_float_2(xs)
+
+    ys = np_transpose(xs)[1::-1]
+    ys[0] = np_array([44.0, 55.0, 66.0])
+    output_ndarray_float_2(xs)
+    output_ndarray_float_2(ys)
+
+    zs = np_reshape(xs, (-1,))
+    zs[2:5] = np_array([1000.0, 2000.0, 3000.0])
+    output_ndarray_float_2(xs)
+
 def test_ndarray_add():
     x = np_identity(2)
     y = x + np_ones([2, 2])
@@ -530,11 +564,59 @@ def test_ndarray_ipow_broadcast_scalar():
     output_ndarray_float_2(x)
 
 def test_ndarray_matmul():
-    x = np_identity(2)
-    y = x @ np_ones([2, 2])
+    # 2D @ 2D -> 2D
+    a1 = np_array([[2.0, 3.0], [5.0, 7.0]])
+    b1 = np_array([[11.0, 13.0], [17.0, 23.0]])
+    c1 = a1 @ b1
+    output_int32(np_shape(c1)[0])
+    output_int32(np_shape(c1)[1])
+    output_ndarray_float_2(c1)
 
-    output_ndarray_float_2(x)
-    output_ndarray_float_2(y)
+    # 1D @ 1D -> Scalar
+    a2 = np_array([2.0, 3.0, 5.0])
+    b2 = np_array([7.0, 11.0, 13.0])
+    c2 = a2 @ b2
+    output_float64(c2)
+
+    # 2D @ 1D -> 1D
+    a3 = np_array([[1.0, 2.0, 3.0], [7.0, 8.0, 9.0]])
+    b3 = np_array([4.0, 5.0, 6.0])
+    c3 = a3 @ b3
+    output_int32(np_shape(c3)[0])
+    output_ndarray_float_1(c3)
+
+    # 1D @ 2D -> 1D
+    a4 = np_array([1.0, 2.0, 3.0])
+    b4 = np_array([[4.0, 5.0], [6.0, 7.0], [8.0, 9.0]])
+    c4 = a4 @ b4
+    output_int32(np_shape(c4)[0])
+    output_ndarray_float_1(c4)
+
+    # Broadcasting
+    a5 = np_array([
+        [[ 0.0,  1.0,  2.0,  3.0],
+         [ 4.0,  5.0,  6.0,  7.0]],
+        [[ 8.0,  9.0, 10.0, 11.0],
+         [12.0, 13.0, 14.0, 15.0]],
+        [[16.0, 17.0, 18.0, 19.0],
+         [20.0, 21.0, 22.0, 23.0]]
+    ])
+    b5 = np_array([
+        [[[ 0.0,  1.0,  2.0],
+          [ 3.0,  4.0,  5.0],
+          [ 6.0,  7.0,  8.0],
+          [ 9.0, 10.0, 11.0]]],
+        [[[12.0, 13.0, 14.0],
+          [15.0, 16.0, 17.0],
+          [18.0, 19.0, 20.0],
+          [21.0, 22.0, 23.0]]]
+    ])
+    c5 = a5 @ b5
+    output_int32(np_shape(c5)[0])
+    output_int32(np_shape(c5)[1])
+    output_int32(np_shape(c5)[2])
+    output_int32(np_shape(c5)[3])
+    output_ndarray_float_4(c5)
 
 def test_ndarray_imatmul():
     x = np_identity(2)
@@ -1580,6 +1662,7 @@ def run() -> int32:
     test_ndarray_neg_idx()
     test_ndarray_slices()
     test_ndarray_nd_idx()
+    test_ndarray_subscript_assignment()
 
     test_ndarray_add()
     test_ndarray_add_broadcast()