WIP

Including immediate variants of these operators.

nac3artiq/src/symbol_resolver.rs

@@ -5,7 +5,7 @@ use nac3core::{
     toplevel::{
         DefinitionId, 
         helper::PRIMITIVE_DEF_IDS, 
-        numpy::{make_ndarray_ty, unpack_ndarray_tvars},
+        numpy::{make_ndarray_ty, unpack_ndarray_var_tys},
         TopLevelDef,
     },
     typecheck::{
@@ -654,7 +654,7 @@ impl InnerResolver {
                 }
             }
             (TypeEnum::TObj { obj_id, .. }, false) if *obj_id == PRIMITIVE_DEF_IDS.ndarray => {
-                let (ty, ndims) = unpack_ndarray_tvars(unifier, extracted_ty);
+                let (ty, ndims) = unpack_ndarray_var_tys(unifier, extracted_ty);
                 let len: usize = self.helper.len_fn.call1(py, (obj,))?.extract(py)?;
                 if len == 0 {
                     assert!(matches!(

nac3core/src/codegen/classes.rs

@@ -1,17 +1,17 @@
 use inkwell::{
     IntPredicate,
     types::{AnyTypeEnum, BasicTypeEnum, IntType, PointerType},
-    values::{ArrayValue, BasicValueEnum, IntValue, PointerValue},
+    values::{BasicValueEnum, IntValue, PointerValue},
 };
 use crate::codegen::{
     CodeGenContext,
     CodeGenerator,
-    irrt::{call_ndarray_calc_size, call_ndarray_flatten_index, call_ndarray_flatten_index_const},
+    irrt::{call_ndarray_calc_size, call_ndarray_flatten_index},
     llvm_intrinsics::call_int_umin,
     stmt::gen_for_callback_incrementing,
 };
 
-/// An LLVM value that is array-like, i.e. it contains a contiguous, sequenced collection of 
+/// An LLVM value that is array-like, i.e. it contains a contiguous, sequenced collection of
 /// elements.
 pub trait ArrayLikeValue<'ctx> {
     /// Returns the element type of this array-like value.
@@ -1162,98 +1162,6 @@ impl<'ctx> ArrayLikeIndexer<'ctx> for NDArrayDataProxy<'ctx, '_> {
 impl<'ctx> UntypedArrayLikeAccessor<'ctx, IntValue<'ctx>> for NDArrayDataProxy<'ctx, '_> {}
 impl<'ctx> UntypedArrayLikeMutator<'ctx, IntValue<'ctx>> for NDArrayDataProxy<'ctx, '_> {}
 
-impl<'ctx> ArrayLikeIndexer<'ctx, ArrayValue<'ctx>> for NDArrayDataProxy<'ctx, '_> {
-    unsafe fn ptr_offset_unchecked<G: CodeGenerator + ?Sized>(
-        &self,
-        ctx: &mut CodeGenContext<'ctx, '_>,
-        generator: &mut G,
-        indices: ArrayValue<'ctx>,
-        name: Option<&str>,
-    ) -> PointerValue<'ctx> {
-        let index = call_ndarray_flatten_index_const(
-            generator,
-            ctx,
-            *self.0,
-            indices,
-        );
-
-        unsafe {
-            ctx.builder.build_in_bounds_gep(
-                self.base_ptr(ctx, generator),
-                &[index],
-                name.unwrap_or_default(),
-            )
-        }.unwrap()
-    }
-
-    fn ptr_offset<G: CodeGenerator + ?Sized>(
-        &self,
-        ctx: &mut CodeGenContext<'ctx, '_>,
-        generator: &mut G,
-        indices: ArrayValue<'ctx>,
-        name: Option<&str>,
-    ) -> PointerValue<'ctx> {
-        let llvm_usize = generator.get_size_type(ctx.ctx);
-
-        let indices_elem_ty = indices.get_type().get_element_type();
-        let Ok(indices_elem_ty) = IntType::try_from(indices_elem_ty) else {
-            panic!("Expected [int32] but got [{indices_elem_ty}]")
-        };
-        assert_eq!(indices_elem_ty.get_bit_width(), 32, "Expected [int32] but got [{indices_elem_ty}]");
-
-        let nidx_leq_ndims = ctx.builder.build_int_compare(
-            IntPredicate::SLE,
-            llvm_usize.const_int(indices.get_type().len() as u64, false),
-            self.0.load_ndims(ctx),
-            ""
-        ).unwrap();
-        ctx.make_assert(
-            generator,
-            nidx_leq_ndims,
-            "0:IndexError",
-            "invalid index to scalar variable",
-            [None, None, None],
-            ctx.current_loc,
-        );
-
-        for idx in 0..indices.get_type().len() {
-            let i = llvm_usize.const_int(idx as u64, false);
-
-            let dim_idx = ctx.builder
-                .build_extract_value(indices, idx, "")
-                .map(BasicValueEnum::into_int_value)
-                .map(|v| ctx.builder.build_int_z_extend_or_bit_cast(v, llvm_usize, "").unwrap())
-                .unwrap();
-            let dim_sz = unsafe {
-                self.0.dim_sizes().get_typed_unchecked(ctx, generator, i, None)
-            };
-
-            let dim_lt = ctx.builder.build_int_compare(
-                IntPredicate::SLT,
-                dim_idx,
-                dim_sz,
-                ""
-            ).unwrap();
-
-            ctx.make_assert(
-                generator,
-                dim_lt,
-                "0:IndexError",
-                "index {0} is out of bounds for axis 0 with size {1}",
-                [Some(dim_idx), Some(dim_sz), None],
-                ctx.current_loc,
-            );
-        }
-
-        unsafe {
-            self.ptr_offset_unchecked(ctx, generator, indices, name)
-        }
-    }
-}
-
-impl<'ctx> UntypedArrayLikeAccessor<'ctx, ArrayValue<'ctx>> for NDArrayDataProxy<'ctx, '_> {}
-impl<'ctx> UntypedArrayLikeMutator<'ctx, ArrayValue<'ctx>> for NDArrayDataProxy<'ctx, '_> {}
-
 impl<'ctx, Index: UntypedArrayLikeAccessor<'ctx>> ArrayLikeIndexer<'ctx, Index> for NDArrayDataProxy<'ctx, '_> {
     unsafe fn ptr_offset_unchecked<G: CodeGenerator + ?Sized>(
         &self,
@@ -1326,6 +1234,9 @@ impl<'ctx, Index: UntypedArrayLikeAccessor<'ctx>> ArrayLikeIndexer<'ctx, Index>
                         self.0.dim_sizes().get_typed_unchecked(ctx, generator, i, None),
                     )
                 };
+                let dim_idx = ctx.builder
+                    .build_int_z_extend_or_bit_cast(dim_idx, dim_sz.get_type(), "")
+                    .unwrap();
 
                 let dim_lt = ctx.builder.build_int_compare(
                     IntPredicate::SLT,

nac3core/src/codegen/expr.rs

@@ -16,6 +16,7 @@ use crate::{
         get_llvm_abi_type,
         irrt::*,
         llvm_intrinsics::{call_expect, call_float_floor, call_float_pow, call_float_powi},
+        numpy,
         stmt::{gen_raise, gen_var},
         CodeGenContext, CodeGenTask,
     },
@@ -23,7 +24,7 @@ use crate::{
     toplevel::{
         DefinitionId, 
         helper::PRIMITIVE_DEF_IDS,
-        numpy::make_ndarray_ty,
+        numpy::{make_ndarray_ty, unpack_ndarray_var_tys},
         TopLevelDef,
     },
     typecheck::{
@@ -42,6 +43,7 @@ use itertools::{chain, izip, Itertools, Either};
 use nac3parser::ast::{
     self, Boolop, Comprehension, Constant, Expr, ExprKind, Location, Operator, StrRef,
 };
+use crate::codegen::classes::ArraySliceValue;
 
 use super::{CodeGenerator, llvm_intrinsics::call_memcpy_generic, need_sret};
 
@@ -1089,34 +1091,22 @@ pub fn gen_comprehension<'ctx, G: CodeGenerator>(
     Ok(Some(list.as_ptr_value().into()))
 }
 
-/// Generates LLVM IR for a [binary operator expression][expr].
-///
-/// * `left` - The left-hand side of the binary operator.
-/// * `op` - The operator applied on the operands.
-/// * `right` - The right-hand side of the binary operator.
-/// * `loc` - The location of the full expression.
-/// * `is_aug_assign` - Whether the binary operator expression is also an assignment operator.
-pub fn gen_binop_expr<'ctx, G: CodeGenerator>(
+/// Generates LLVM IR for a binary operator expression using the [`Type`] and 
+/// [LLVM value][`BasicValueEnum`] of the operands.
+pub fn gen_binop_expr_with_values<'ctx, G: CodeGenerator>(
     generator: &mut G,
     ctx: &mut CodeGenContext<'ctx, '_>,
-    left: &Expr<Option<Type>>,
+    left: (&Option<Type>, BasicValueEnum<'ctx>),
     op: &Operator,
-    right: &Expr<Option<Type>>,
+    right: (&Option<Type>, BasicValueEnum<'ctx>),
     loc: Location,
     is_aug_assign: bool,
 ) -> Result<Option<ValueEnum<'ctx>>, String> {
-    let ty1 = ctx.unifier.get_representative(left.custom.unwrap());
-    let ty2 = ctx.unifier.get_representative(right.custom.unwrap());
-    let left_val = if let Some(v) = generator.gen_expr(ctx, left)? {
-        v.to_basic_value_enum(ctx, generator, left.custom.unwrap())?
-    } else {
-        return Ok(None)
-    };
-    let right_val = if let Some(v) = generator.gen_expr(ctx, right)? {
-        v.to_basic_value_enum(ctx, generator, right.custom.unwrap())?
-    } else {
-        return Ok(None)
-    };
+    let (left_ty, left_val) = left;
+    let (right_ty, right_val) = right;
+
+    let ty1 = ctx.unifier.get_representative(left_ty.unwrap());
+    let ty2 = ctx.unifier.get_representative(right_ty.unwrap());
 
     // we can directly compare the types, because we've got their representatives
     // which would be unchanged until further unification, which we would never do
@@ -1140,8 +1130,46 @@ pub fn gen_binop_expr<'ctx, G: CodeGenerator>(
             Some("f_pow_i")
         );
         Ok(Some(res.into()))
+    } else if matches!(&*ctx.unifier.get_ty(ty1), TypeEnum::TObj { obj_id, .. } if obj_id == &PRIMITIVE_DEF_IDS.ndarray) && matches!(&*ctx.unifier.get_ty(ty2), TypeEnum::TObj { obj_id, .. } if obj_id == &PRIMITIVE_DEF_IDS.ndarray) {
+        let llvm_usize = generator.get_size_type(ctx.ctx);
+        let (ndarray_dtype1, _) = unpack_ndarray_var_tys(&mut ctx.unifier, ty1);
+        let (ndarray_dtype2, _) = unpack_ndarray_var_tys(&mut ctx.unifier, ty1);
+
+        assert!(ctx.unifier.unioned(ndarray_dtype1, ndarray_dtype2));
+
+        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 = numpy::ndarray_elementwise_binop_impl(
+            generator,
+            ctx,
+            ndarray_dtype1,
+            if is_aug_assign { Some(left_val) } else { None },
+            left_val,
+            right_val,
+            |generator, ctx, elem_ty, (lhs, rhs)| {
+                gen_binop_expr_with_values(
+                    generator,
+                    ctx,
+                    (&Some(elem_ty), lhs),
+                    op,
+                    (&Some(elem_ty), rhs),
+                    ctx.current_loc,
+                    is_aug_assign,
+                )?.unwrap().to_basic_value_enum(ctx, generator, elem_ty)
+            },
+        )?;
+
+        Ok(Some(res.as_ptr_value().into()))
     } else {
-        let left_ty_enum = ctx.unifier.get_ty_immutable(left.custom.unwrap());
+        let left_ty_enum = ctx.unifier.get_ty_immutable(left_ty.unwrap());
         let TypeEnum::TObj { fields, obj_id, .. } = left_ty_enum.as_ref() else {
             unreachable!("must be tobj")
         };
@@ -1161,7 +1189,7 @@ pub fn gen_binop_expr<'ctx, G: CodeGenerator>(
         let signature = if let Some(call) = ctx.calls.get(&loc.into()) {
             ctx.unifier.get_call_signature(*call).unwrap()
         } else {
-            let left_enum_ty = ctx.unifier.get_ty_immutable(left.custom.unwrap());
+            let left_enum_ty = ctx.unifier.get_ty_immutable(left_ty.unwrap());
             let TypeEnum::TObj { fields, .. } = left_enum_ty.as_ref() else {
                 unreachable!("must be tobj")
             };
@@ -1186,13 +1214,51 @@ pub fn gen_binop_expr<'ctx, G: CodeGenerator>(
         generator
             .gen_call(
                 ctx,
-                Some((left.custom.unwrap(), left_val.into())),
+                Some((left_ty.unwrap(), left_val.into())),
                 (&signature, fun_id),
                 vec![(None, right_val.into())],
             ).map(|f| f.map(Into::into))
     }
 }
 
+/// Generates LLVM IR for a [binary operator expression][expr].
+///
+/// * `left` - The left-hand side of the binary operator.
+/// * `op` - The operator applied on the operands.
+/// * `right` - The right-hand side of the binary operator.
+/// * `loc` - The location of the full expression.
+/// * `is_aug_assign` - Whether the binary operator expression is also an assignment operator.
+pub fn gen_binop_expr<'ctx, G: CodeGenerator>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    left: &Expr<Option<Type>>,
+    op: &Operator,
+    right: &Expr<Option<Type>>,
+    loc: Location,
+    is_aug_assign: bool,
+) -> Result<Option<ValueEnum<'ctx>>, String> {
+    let left_val = if let Some(v) = generator.gen_expr(ctx, left)? {
+        v.to_basic_value_enum(ctx, generator, left.custom.unwrap())?
+    } else {
+        return Ok(None)
+    };
+    let right_val = if let Some(v) = generator.gen_expr(ctx, right)? {
+        v.to_basic_value_enum(ctx, generator, right.custom.unwrap())?
+    } else {
+        return Ok(None)
+    };
+
+    gen_binop_expr_with_values(
+        generator,
+        ctx,
+        (&left.custom, left_val),
+        op,
+        (&right.custom, right_val),
+        loc,
+        is_aug_assign,
+    )
+}
+
 /// Generates code for a subscript expression on an `ndarray`.
 ///
 /// * `ty` - The `Type` of the `NDArray` elements.
@@ -1265,12 +1331,14 @@ fn gen_ndarray_subscript_expr<'ctx, G: CodeGenerator>(
         } else {
             return Ok(None)
         };
+        let index_addr = generator.gen_var_alloc(ctx, index.get_type().into(), None)?;
+        ctx.builder.build_store(index_addr, index).unwrap();
 
         Ok(Some(v.data()
             .get(
                 ctx,
                 generator,
-                ctx.ctx.i32_type().const_array(&[index]),
+                ArraySliceValue::from_ptr_val(index_addr, llvm_usize.const_int(1, false), None),
                 None,
             )
             .into()))
@@ -1286,6 +1354,8 @@ fn gen_ndarray_subscript_expr<'ctx, G: CodeGenerator>(
         } else {
             return Ok(None)
         };
+        let index_addr = generator.gen_var_alloc(ctx, index.get_type().into(), None)?;
+        ctx.builder.build_store(index_addr, index).unwrap();
 
         // Create a new array, remove the top dimension from the dimension-size-list, and copy the
         // elements over
@@ -1340,7 +1410,7 @@ fn gen_ndarray_subscript_expr<'ctx, G: CodeGenerator>(
         let v_data_src_ptr = v.data().ptr_offset(
             ctx,
             generator,
-            ctx.ctx.i32_type().const_array(&[index]),
+            ArraySliceValue::from_ptr_val(index_addr, llvm_usize.const_int(1, false), None),
             None
         );
         call_memcpy_generic(

nac3core/src/codegen/irrt/irrt.c

@@ -8,6 +8,8 @@ typedef unsigned _BitInt(64) uint64_t;
 # define MAX(a, b) (a > b ? a : b)
 # define MIN(a, b) (a > b ? b : a)
 
+# define NULL ((void *) 0)
+
 // adapted from GNU Scientific Library: https://git.savannah.gnu.org/cgit/gsl.git/tree/sys/pow_int.c
 // need to make sure `exp >= 0` before calling this function
 #define DEF_INT_EXP(T) T __nac3_int_exp_##T( \
@@ -243,13 +245,13 @@ void __nac3_ndarray_calc_nd_indices64(
     uint64_t index,
     const uint64_t* dims,
     uint64_t num_dims,
-    uint64_t* idxs
+    uint32_t* idxs
 ) {
     uint64_t stride = 1;
     for (uint64_t dim = 0; dim < num_dims; dim++) {
         uint64_t i = num_dims - dim - 1;
         __builtin_assume(dims[i] > 0);
-        idxs[i] = (index / stride) % dims[i];
+        idxs[i] = (uint32_t) ((index / stride) % dims[i]);
         stride *= dims[i];
     }
 }
@@ -293,3 +295,87 @@ uint64_t __nac3_ndarray_flatten_index64(
     }
     return idx;
 }
+
+void __nac3_ndarray_calc_broadcast(
+    const uint32_t *lhs_dims,
+    uint32_t lhs_ndims,
+    const uint32_t *rhs_dims,
+    uint32_t rhs_ndims,
+    uint32_t *out_dims
+) {
+    uint32_t max_ndims = lhs_ndims > rhs_ndims ? lhs_ndims : rhs_ndims;
+
+    for (uint32_t i = 0; i < max_ndims; ++i) {
+        uint32_t *lhs_dim_sz = i < lhs_ndims ? &lhs_dims[lhs_ndims - i - 1] : NULL;
+        uint32_t *rhs_dim_sz = i < rhs_ndims ? &rhs_dims[rhs_ndims - i - 1] : NULL;
+        uint32_t *out_dim = &out_dims[max_ndims - i - 1];
+
+        if (lhs_dim_sz == NULL) {
+            *out_dim = *rhs_dim_sz;
+        } else if (rhs_dim_sz == NULL) {
+            *out_dim = *lhs_dim_sz;
+        } else if (*lhs_dim_sz == 1) {
+            *out_dim = *rhs_dim_sz;
+        } else if (*rhs_dim_sz == 1) {
+            *out_dim = *lhs_dim_sz;
+        } else if (*lhs_dim_sz == *rhs_dim_sz) {
+            *out_dim = *lhs_dim_sz;
+        } else {
+            __builtin_unreachable();
+        }
+    }
+}
+
+void __nac3_ndarray_calc_broadcast64(
+    const uint64_t *lhs_dims,
+    uint64_t lhs_ndims,
+    const uint64_t *rhs_dims,
+    uint64_t rhs_ndims,
+    uint64_t *out_dims
+) {
+    uint64_t max_ndims = lhs_ndims > rhs_ndims ? lhs_ndims : rhs_ndims;
+
+    for (uint64_t i = 0; i < max_ndims; ++i) {
+        uint64_t *lhs_dim_sz = i < lhs_ndims ? &lhs_dims[lhs_ndims - i - 1] : NULL;
+        uint64_t *rhs_dim_sz = i < rhs_ndims ? &rhs_dims[rhs_ndims - i - 1] : NULL;
+        uint64_t *out_dim = &out_dims[max_ndims - i - 1];
+
+        if (lhs_dim_sz == NULL) {
+            *out_dim = *rhs_dim_sz;
+        } else if (rhs_dim_sz == NULL) {
+            *out_dim = *lhs_dim_sz;
+        } else if (*lhs_dim_sz == 1) {
+            *out_dim = *rhs_dim_sz;
+        } else if (*rhs_dim_sz == 1) {
+            *out_dim = *lhs_dim_sz;
+        } else if (*lhs_dim_sz == *rhs_dim_sz) {
+            *out_dim = *lhs_dim_sz;
+        } else {
+            __builtin_unreachable();
+        }
+    }
+}
+
+void __nac3_ndarray_calc_broadcast_idx(
+    const uint32_t *src_dims,
+    uint32_t src_ndims,
+    const uint32_t *in_idx,
+    uint32_t *out_idx
+) {
+    for (uint32_t i = 0; i < src_ndims; ++i) {
+        uint32_t src_i = src_ndims - i - 1;
+        out_idx[src_i] = src_dims[src_i] == 1 ? 0 : in_idx[src_i];
+    }
+}
+
+void __nac3_ndarray_calc_broadcast_idx64(
+    const uint64_t *src_dims,
+    uint64_t src_ndims,
+    const uint32_t *in_idx,
+    uint32_t *out_idx
+) {
+    for (uint64_t i = 0; i < src_ndims; ++i) {
+        uint64_t src_i = src_ndims - i - 1;
+        out_idx[src_i] = src_dims[src_i] == 1 ? 0 : (uint32_t) in_idx[src_i];
+    }
+}

nac3core/src/codegen/irrt/mod.rs

@@ -1,9 +1,18 @@
 use crate::typecheck::typedef::Type;
 
 use super::{
-    classes::{ArrayLikeIndexer, ArrayLikeValue, ListValue, NDArrayValue, UntypedArrayLikeMutator},
+    classes::{
+        ArrayLikeIndexer,
+        ArrayLikeValue,
+        ArraySliceValue,
+        ListValue,
+        NDArrayValue,
+        TypedArrayLikeAdapter,
+        UntypedArrayLikeAccessor,
+    },
     CodeGenContext,
     CodeGenerator,
+    llvm_intrinsics,
 };
 use inkwell::{
     attributes::{Attribute, AttributeLoc},
@@ -11,7 +20,7 @@ use inkwell::{
     memory_buffer::MemoryBuffer,
     module::Module,
     types::{BasicTypeEnum, IntType},
-    values::{ArrayValue, BasicValueEnum, CallSiteValue, FloatValue, IntValue, PointerValue},
+    values::{BasicValueEnum, CallSiteValue, FloatValue, IntValue},
     AddressSpace, IntPredicate,
 };
 use itertools::Either;
@@ -619,7 +628,8 @@ pub fn call_ndarray_calc_size<'ctx, G, Dims>(
         .unwrap()
 }
 
-/// Generates a call to `__nac3_ndarray_calc_nd_indices`.
+/// Generates a call to `__nac3_ndarray_calc_nd_indices`. Returns a [`TypeArrayLikeAdpater`]
+/// containing `i32` indices of the flattened index.
 ///
 /// * `index` - The index to compute the multidimensional index for.
 /// * `ndarray` - LLVM pointer to the `NDArray`. This value must be the LLVM representation of an
@@ -629,10 +639,11 @@ pub fn call_ndarray_calc_nd_indices<'ctx, G: CodeGenerator + ?Sized>(
     ctx: &mut CodeGenContext<'ctx, '_>,
     index: IntValue<'ctx>,
     ndarray: NDArrayValue<'ctx>,
-) -> PointerValue<'ctx> {
+) -> TypedArrayLikeAdapter<'ctx, IntValue<'ctx>> {
     let llvm_void = ctx.ctx.void_type();
+    let llvm_i32 = ctx.ctx.i32_type();
     let llvm_usize = generator.get_size_type(ctx.ctx);
-
+    let llvm_pi32 = llvm_i32.ptr_type(AddressSpace::default());
     let llvm_pusize = llvm_usize.ptr_type(AddressSpace::default());
 
     let ndarray_calc_nd_indices_fn_name = match llvm_usize.get_bit_width() {
@@ -646,7 +657,7 @@ pub fn call_ndarray_calc_nd_indices<'ctx, G: CodeGenerator + ?Sized>(
                 llvm_usize.into(),
                 llvm_pusize.into(),
                 llvm_usize.into(),
-                llvm_pusize.into(),
+                llvm_pi32.into(),
             ],
             false,
         );
@@ -658,7 +669,7 @@ pub fn call_ndarray_calc_nd_indices<'ctx, G: CodeGenerator + ?Sized>(
     let ndarray_dims = ndarray.dim_sizes();
 
     let indices = ctx.builder.build_array_alloca(
-        llvm_usize,
+        llvm_i32,
         ndarray_num_dims,
         "",
     ).unwrap();
@@ -676,7 +687,11 @@ pub fn call_ndarray_calc_nd_indices<'ctx, G: CodeGenerator + ?Sized>(
         )
         .unwrap();
 
-    indices
+    TypedArrayLikeAdapter::from(
+        ArraySliceValue::from_ptr_val(indices, ndarray_num_dims, None),
+        Box::new(|_, v| v.into_int_value()),
+        Box::new(|_, v| v.into()),
+    )
 }
 
 fn call_ndarray_flatten_index_impl<'ctx, G, Indices>(
@@ -771,46 +786,152 @@ pub fn call_ndarray_flatten_index<'ctx, G, Index>(
         indices,
     )
 }
-/// Generates a call to `__nac3_ndarray_flatten_index`. Returns the flattened index for the
-/// multidimensional index.
-///
-/// * `ndarray` - LLVM pointer to the `NDArray`. This value must be the LLVM representation of an
-/// `NDArray`.
-/// * `indices` - The multidimensional index to compute the flattened index for.
-pub fn call_ndarray_flatten_index_const<'ctx, G: CodeGenerator + ?Sized>(
+
+/// Generates a call to `__nac3_ndarray_calc_broadcast`. Returns a tuple containing the number of
+/// dimension and size of each dimension of the resultant `ndarray`.
+pub fn call_ndarray_calc_broadcast<'ctx, G: CodeGenerator + ?Sized>(
     generator: &mut G,
     ctx: &mut CodeGenContext<'ctx, '_>,
-    ndarray: NDArrayValue<'ctx>,
-    indices: ArrayValue<'ctx>,
-) -> IntValue<'ctx> {
+    lhs: NDArrayValue<'ctx>,
+    rhs: NDArrayValue<'ctx>,
+) -> TypedArrayLikeAdapter<'ctx, IntValue<'ctx>> {
     let llvm_usize = generator.get_size_type(ctx.ctx);
+    let llvm_pusize = llvm_usize.ptr_type(AddressSpace::default());
 
-    let indices_size = indices.get_type().len();
-    let indices_alloca = generator.gen_array_var_alloc(
-        ctx,
-        indices.get_type().get_element_type(),
-        llvm_usize.const_int(indices_size as u64, false),
-        None,
-    ).unwrap();
-    for i in 0..indices_size {
-        let v = ctx.builder.build_extract_value(indices, i, "")
-            .unwrap()
-            .into_int_value();
+    let ndarray_calc_broadcast_fn_name = match llvm_usize.get_bit_width() {
+        32 => "__nac3_ndarray_calc_broadcast",
+        64 => "__nac3_ndarray_calc_broadcast64",
+        bw => unreachable!("Unsupported size type bit width: {}", bw)
+    };
+    let ndarray_calc_broadcast_fn = ctx.module.get_function(ndarray_calc_broadcast_fn_name).unwrap_or_else(|| {
+        let fn_type = llvm_usize.fn_type(
+            &[
+                llvm_pusize.into(),
+                llvm_usize.into(),
+                llvm_pusize.into(),
+                llvm_usize.into(),
+                llvm_pusize.into(),
+            ],
+            false,
+        );
 
-        unsafe {
-            indices_alloca.set_unchecked(
-                ctx,
-                generator,
-                ctx.ctx.i32_type().const_int(i as u64, false),
-                v.into(),
-            );
-        }
-    }
+        ctx.module.add_function(ndarray_calc_broadcast_fn_name, fn_type, None)
+    });
 
-    call_ndarray_flatten_index_impl(
-        generator,
-        ctx,
-        ndarray,
-        &indices_alloca,
+    let lhs_ndims = lhs.load_ndims(ctx);
+    let rhs_ndims = rhs.load_ndims(ctx);
+    let max_ndims = llvm_intrinsics::call_int_umax(ctx, lhs_ndims, rhs_ndims, None);
+
+    // TODO: Generate assertion checks for whether each dimension is compatible
+    // gen_for_callback_incrementing(
+    //     generator,
+    //     ctx,
+    //     llvm_usize.const_zero(),
+    //     (max_ndims, false),
+    //     |generator, ctx, idx| {
+    //         let lhs_dim_sz =
+    //
+    //         let lhs_elem = lhs.get_dims().get(ctx, generator, idx, None);
+    //         let rhs_elem = rhs.get_dims().get(ctx, generator, idx, None);
+    //
+    //
+    //     },
+    //     llvm_usize.const_int(1, false),
+    // ).unwrap();
+
+    let lhs_dims = lhs.dim_sizes().base_ptr(ctx, generator);
+    let lhs_ndims = lhs.load_ndims(ctx);
+    let rhs_dims = rhs.dim_sizes().base_ptr(ctx, generator);
+    let rhs_ndims = rhs.load_ndims(ctx);
+    let out_dims = ctx.builder.build_array_alloca(llvm_usize, max_ndims, "").unwrap();
+    let out_dims = ArraySliceValue::from_ptr_val(out_dims, max_ndims, None);
+
+    ctx.builder
+        .build_call(
+            ndarray_calc_broadcast_fn,
+            &[
+                lhs_dims.into(),
+                lhs_ndims.into(),
+                rhs_dims.into(),
+                rhs_ndims.into(),
+                out_dims.base_ptr(ctx, generator).into(),
+            ],
+            "",
+        )
+        .unwrap();
+
+    TypedArrayLikeAdapter::from(
+        out_dims,
+        Box::new(|_, v| v.into_int_value()),
+        Box::new(|_, v| v.into()),
+    )
+}
+
+/// Generates a call to `__nac3_ndarray_calc_broadcast_idx`. Returns an [`ArrayAllocaValue`]
+/// containing the indices used for accessing `array` corresponding to the index of the broadcasted
+/// array `broadcast_idx`.
+pub fn call_ndarray_calc_broadcast_index<'ctx, G: CodeGenerator + ?Sized, BroadcastIdx: UntypedArrayLikeAccessor<'ctx>>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    array: NDArrayValue<'ctx>,
+    broadcast_idx: &BroadcastIdx,
+) -> TypedArrayLikeAdapter<'ctx, IntValue<'ctx>> {
+    let llvm_i32 = ctx.ctx.i32_type();
+    let llvm_usize = generator.get_size_type(ctx.ctx);
+    let llvm_pi32 = llvm_i32.ptr_type(AddressSpace::default());
+    let llvm_pusize = llvm_usize.ptr_type(AddressSpace::default());
+
+    let ndarray_calc_broadcast_fn_name = match llvm_usize.get_bit_width() {
+        32 => "__nac3_ndarray_calc_broadcast_idx",
+        64 => "__nac3_ndarray_calc_broadcast_idx64",
+        bw => unreachable!("Unsupported size type bit width: {}", bw)
+    };
+    let ndarray_calc_broadcast_fn = ctx.module.get_function(ndarray_calc_broadcast_fn_name).unwrap_or_else(|| {
+        let fn_type = llvm_usize.fn_type(
+            &[
+                llvm_pusize.into(),
+                llvm_usize.into(),
+                llvm_pi32.into(),
+                llvm_pi32.into(),
+            ],
+            false,
+        );
+
+        ctx.module.add_function(ndarray_calc_broadcast_fn_name, fn_type, None)
+    });
+
+    // TODO: Assertions
+
+    let broadcast_size = broadcast_idx.size(ctx, generator);
+    let out_idx = ctx.builder.build_array_alloca(llvm_i32, broadcast_size, "").unwrap();
+
+    let array_dims = array.dim_sizes().base_ptr(ctx, generator);
+    let array_ndims = array.load_ndims(ctx);
+    let broadcast_idx_ptr = unsafe {
+        broadcast_idx.ptr_offset_unchecked(
+            ctx,
+            generator,
+            llvm_usize.const_zero(),
+            None
+        )
+    };
+
+    ctx.builder
+        .build_call(
+            ndarray_calc_broadcast_fn,
+            &[
+                array_dims.into(),
+                array_ndims.into(),
+                broadcast_idx_ptr.into(),
+                out_idx.into(),
+            ],
+            "",
+        )
+        .unwrap();
+
+    TypedArrayLikeAdapter::from(
+        ArraySliceValue::from_ptr_val(out_idx, broadcast_size, None),
+        Box::new(|_, v| v.into_int_value()),
+        Box::new(|_, v| v.into()),
     )
 }

nac3core/src/codegen/mod.rs

@@ -2,7 +2,7 @@ use crate::{
     symbol_resolver::{StaticValue, SymbolResolver},
     toplevel::{
         helper::PRIMITIVE_DEF_IDS, 
-        numpy::unpack_ndarray_tvars, 
+        numpy::unpack_ndarray_var_tys, 
         TopLevelContext, 
         TopLevelDef,
     },
@@ -451,7 +451,7 @@ fn get_llvm_type<'ctx, G: CodeGenerator + ?Sized>(
 
                         TObj { obj_id, .. } if *obj_id == PRIMITIVE_DEF_IDS.ndarray => {
                             let llvm_usize = generator.get_size_type(ctx);
-                            let (dtype, _) = unpack_ndarray_tvars(unifier, ty);
+                            let (dtype, _) = unpack_ndarray_var_tys(unifier, ty);
                             let element_type = get_llvm_type(
                                 ctx, 
                                 module, 

nac3core/src/codegen/numpy.rs

@@ -12,11 +12,14 @@ use crate::{
             ListValue,
             NDArrayValue,
             TypedArrayLikeAccessor,
+            TypedArrayLikeAdapter,
             UntypedArrayLikeAccessor,
         },
         CodeGenContext,
         CodeGenerator,
         irrt::{
+            call_ndarray_calc_broadcast,
+            call_ndarray_calc_broadcast_index,
             call_ndarray_calc_nd_indices,
             call_ndarray_calc_size,
         },
@@ -26,7 +29,7 @@ use crate::{
     symbol_resolver::ValueEnum,
     toplevel::{
         DefinitionId,
-        numpy::{make_ndarray_ty, unpack_ndarray_tvars},
+        numpy::{make_ndarray_ty, unpack_ndarray_var_tys},
     },
     typecheck::typedef::{FunSignature, Type},
 };
@@ -324,7 +327,7 @@ fn ndarray_fill_indexed<'ctx, G, ValueFn>(
 ) -> Result<(), String>
     where
         G: CodeGenerator + ?Sized,
-        ValueFn: Fn(&mut G, &mut CodeGenContext<'ctx, '_>, PointerValue<'ctx>) -> Result<BasicValueEnum<'ctx>, String>,
+        ValueFn: Fn(&mut G, &mut CodeGenContext<'ctx, '_>, TypedArrayLikeAdapter<'ctx, IntValue<'ctx>>) -> Result<BasicValueEnum<'ctx>, String>,
 {
     ndarray_fill_flattened(
         generator,
@@ -343,6 +346,45 @@ fn ndarray_fill_indexed<'ctx, G, ValueFn>(
     )
 }
 
+/// Generates the LLVM IR for populating the entire `NDArray` using a lambda with the same-indexed
+/// element from two other `NDArray` as its input.
+fn ndarray_broadcast_fill<'ctx, G, ValueFn>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    elem_ty: Type,
+    res: NDArrayValue<'ctx>,
+    lhs: NDArrayValue<'ctx>,
+    rhs: NDArrayValue<'ctx>,
+    value_fn: ValueFn,
+) -> Result<NDArrayValue<'ctx>, String>
+    where
+        G: CodeGenerator + ?Sized,
+        ValueFn: Fn(&mut G, &mut CodeGenContext<'ctx, '_>, Type, (BasicValueEnum<'ctx>, BasicValueEnum<'ctx>)) -> Result<BasicValueEnum<'ctx>, String>,
+{
+    ndarray_fill_indexed(
+        generator,
+        ctx,
+        res,
+        |generator, ctx, idx| {
+            let lhs_idx = call_ndarray_calc_broadcast_index(generator, ctx, lhs, &idx);
+            let rhs_idx = call_ndarray_calc_broadcast_index(generator, ctx, rhs, &idx);
+
+            let elem = unsafe {
+                (
+                    lhs.data().get_unchecked(ctx, generator, lhs_idx, None),
+                    rhs.data().get_unchecked(ctx, generator, rhs_idx, None),
+                )
+            };
+
+            debug_assert_eq!(elem.0.get_type(), elem.1.get_type());
+
+            value_fn(generator, ctx, elem_ty, elem)
+        },
+    )?;
+
+    Ok(res)
+}
+
 /// LLVM-typed implementation for generating the implementation for `ndarray.zeros`.
 ///
 /// * `elem_ty` - The element type of the `NDArray`.
@@ -470,6 +512,7 @@ fn call_ndarray_eye_impl<'ctx, G: CodeGenerator + ?Sized>(
     ncols: IntValue<'ctx>,
     offset: IntValue<'ctx>,
 ) -> Result<NDArrayValue<'ctx>, String> {
+    let llvm_i32 = ctx.ctx.i32_type();
     let llvm_usize = generator.get_size_type(ctx.ctx);
     let llvm_usize_2 = llvm_usize.array_type(2);
 
@@ -498,21 +541,17 @@ fn call_ndarray_eye_impl<'ctx, G: CodeGenerator + ?Sized>(
         ctx,
         ndarray,
         |generator, ctx, indices| {
-            let row = ctx.build_gep_and_load(
-                indices,
-                &[llvm_usize.const_int(0, false)],
-                None,
-            ).into_int_value();
-            let col = ctx.build_gep_and_load(
-                indices,
-                &[llvm_usize.const_int(1, false)],
-                None,
-            ).into_int_value();
+            let (row, col) = unsafe {
+                (
+                    indices.get_typed_unchecked(ctx, generator, llvm_usize.const_zero(), None),
+                    indices.get_typed_unchecked(ctx, generator, llvm_usize.const_int(1, false), None),
+                )
+            };
 
             let col_with_offset = ctx.builder
                 .build_int_add(
                     col,
-                    ctx.builder.build_int_s_extend_or_bit_cast(offset, llvm_usize, "").unwrap(),
+                    ctx.builder.build_int_s_extend_or_bit_cast(offset, llvm_i32, "").unwrap(),
                     "",
                 )
                 .unwrap();
@@ -581,6 +620,58 @@ fn ndarray_copy_impl<'ctx, G: CodeGenerator + ?Sized>(
     Ok(ndarray)
 }
 
+/// LLVM-typed implementation for computing elementwise binary operations.
+///
+/// * `elem_ty` - The element type of the `NDArray`.
+/// * `res` - The `ndarray` instance to write results into, or [`None`] if the result should be
+/// written to a new `ndarray`.
+/// * `value_fn` - Function mapping the two input elements into the result.
+pub fn ndarray_elementwise_binop_impl<'ctx, G, ValueFn>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    elem_ty: Type,
+    res: Option<NDArrayValue<'ctx>>,
+    this: NDArrayValue<'ctx>,
+    other: NDArrayValue<'ctx>,
+    value_fn: ValueFn,
+) -> Result<NDArrayValue<'ctx>, String>
+    where
+        G: CodeGenerator,
+        ValueFn: Fn(&mut G, &mut CodeGenContext<'ctx, '_>, Type, (BasicValueEnum<'ctx>, BasicValueEnum<'ctx>)) -> Result<BasicValueEnum<'ctx>, String>,
+{
+    let ndarray_dims = call_ndarray_calc_broadcast(generator, ctx, this, other);
+    let ndarray = res.unwrap_or_else(|| {
+        create_ndarray_dyn_shape(
+            generator,
+            ctx,
+            elem_ty,
+            &ndarray_dims,
+            |generator, ctx, v| {
+                Ok(v.size(ctx, generator))
+            },
+            |generator, ctx, v, idx| {
+                unsafe {
+                    Ok(v.get_typed_unchecked(ctx, generator, idx, None))
+                }
+            },
+        ).unwrap()
+    });
+
+    ndarray_broadcast_fill(
+        generator,
+        ctx,
+        elem_ty,
+        ndarray,
+        this,
+        other,
+        |generator, ctx, elem_ty, elems| {
+            value_fn(generator, ctx, elem_ty, elems)
+        },
+    )?;
+
+    Ok(ndarray)
+}
+
 /// Generates LLVM IR for `ndarray.empty`.
 pub fn gen_ndarray_empty<'ctx>(
     context: &mut CodeGenContext<'ctx, '_>,
@@ -767,7 +858,7 @@ pub fn gen_ndarray_copy<'ctx>(
     let llvm_usize = generator.get_size_type(context.ctx);
 
     let this_ty = obj.as_ref().unwrap().0;
-    let (this_elem_ty, _) = unpack_ndarray_tvars(&mut context.unifier, this_ty);
+    let (this_elem_ty, _) = unpack_ndarray_var_tys(&mut context.unifier, this_ty);
     let this_arg = obj
         .as_ref()
         .unwrap()

nac3core/src/codegen/stmt.rs

@@ -13,7 +13,7 @@ use crate::{
     toplevel::{
         DefinitionId,
         helper::PRIMITIVE_DEF_IDS,
-        numpy::unpack_ndarray_tvars,
+        numpy::unpack_ndarray_var_tys,
         TopLevelDef,
     },
     typecheck::typedef::{FunSignature, Type, TypeEnum},
@@ -251,7 +251,7 @@ pub fn gen_assign<'ctx, G: CodeGenerator>(
             let ty = match &*ctx.unifier.get_ty_immutable(target.custom.unwrap()) {
                 TypeEnum::TList { ty } => *ty,
                 TypeEnum::TObj { obj_id, .. } if *obj_id == PRIMITIVE_DEF_IDS.ndarray => {
-                    unpack_ndarray_tvars(&mut ctx.unifier, target.custom.unwrap()).0
+                    unpack_ndarray_var_tys(&mut ctx.unifier, target.custom.unwrap()).0
                 }
                 _ => unreachable!(),
             };
@@ -546,7 +546,7 @@ pub fn gen_for_callback<'ctx, 'a, G, I, InitFn, CondFn, BodyFn, UpdateFn>(
 ///     body(x);
 /// }
 /// ```
-/// 
+///
 /// * `init_val` - The initial value of the loop variable. The type of this value will also be used 
 /// as the type of the loop variable.
 /// * `max_val` - A tuple containing the maximum value of the loop variable, and whether the maximum

nac3core/src/toplevel/builtins.rs

@@ -299,6 +299,8 @@ pub fn get_builtins(primitives: &mut (PrimitiveStore, Unifier)) -> BuiltinInfo {
         Some("N".into()),
         None,
     );
+    let size_t = primitives.0.usize();
+
     let var_map: VarMap = vec![(num_ty.1, num_ty.0)].into_iter().collect();
     let exception_fields = vec![
         ("__name__".into(), int32, true),
@@ -345,8 +347,27 @@ pub fn get_builtins(primitives: &mut (PrimitiveStore, Unifier)) -> BuiltinInfo {
         .nth(1)
         .map(|(var_id, ty)| (*ty, *var_id))
         .unwrap();
+    let ndarray_usized_ndims_tvar = primitives.1.get_fresh_const_generic_var(
+        size_t, 
+        Some("ndarray_ndims".into()), 
+        None,
+    );
     let ndarray_copy_ty = *ndarray_fields.get(&"copy".into()).unwrap();
     let ndarray_fill_ty = *ndarray_fields.get(&"fill".into()).unwrap();
+    let ndarray_add_ty = *ndarray_fields.get(&"__add__".into()).unwrap();
+    let ndarray_sub_ty = *ndarray_fields.get(&"__sub__".into()).unwrap();
+    let ndarray_mul_ty = *ndarray_fields.get(&"__mul__".into()).unwrap();
+    let ndarray_truediv_ty = *ndarray_fields.get(&"__truediv__".into()).unwrap();
+    let ndarray_floordiv_ty = *ndarray_fields.get(&"__floordiv__".into()).unwrap();
+    let ndarray_mod_ty = *ndarray_fields.get(&"__mod__".into()).unwrap();
+    let ndarray_pow_ty = *ndarray_fields.get(&"__pow__".into()).unwrap();
+    let ndarray_iadd_ty = *ndarray_fields.get(&"__iadd__".into()).unwrap();
+    let ndarray_isub_ty = *ndarray_fields.get(&"__isub__".into()).unwrap();
+    let ndarray_imul_ty = *ndarray_fields.get(&"__imul__".into()).unwrap();
+    let ndarray_itruediv_ty = *ndarray_fields.get(&"__itruediv__".into()).unwrap();
+    let ndarray_ifloordiv_ty = *ndarray_fields.get(&"__ifloordiv__".into()).unwrap();
+    let ndarray_imod_ty = *ndarray_fields.get(&"__imod__".into()).unwrap();
+    let ndarray_ipow_ty = *ndarray_fields.get(&"__ipow__".into()).unwrap();
 
     let top_level_def_list = vec![
         Arc::new(RwLock::new(TopLevelComposer::make_top_level_class_def(
@@ -524,6 +545,20 @@ pub fn get_builtins(primitives: &mut (PrimitiveStore, Unifier)) -> BuiltinInfo {
             methods: vec![
                 ("copy".into(), ndarray_copy_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 1)),
                 ("fill".into(), ndarray_fill_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 2)),
+                ("__add__".into(), ndarray_add_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 3)),
+                ("__sub__".into(), ndarray_sub_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 4)),
+                ("__mul__".into(), ndarray_mul_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 5)),
+                ("__truediv__".into(), ndarray_mul_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 6)),
+                ("__floordiv__".into(), ndarray_mul_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 7)),
+                ("__mod__".into(), ndarray_mul_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 8)),
+                ("__pow__".into(), ndarray_mul_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 9)),
+                ("__iadd__".into(), ndarray_iadd_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 10)),
+                ("__isub__".into(), ndarray_isub_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 11)),
+                ("__imul__".into(), ndarray_imul_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 12)),
+                ("__itruediv__".into(), ndarray_mul_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 13)),
+                ("__ifloordiv__".into(), ndarray_mul_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 14)),
+                ("__imod__".into(), ndarray_mul_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 15)),
+                ("__ipow__".into(), ndarray_imul_ty.0, DefinitionId(PRIMITIVE_DEF_IDS.ndarray.0 + 16)),
             ],
             ancestors: Vec::default(),
             constructor: None,
@@ -562,6 +597,216 @@ pub fn get_builtins(primitives: &mut (PrimitiveStore, Unifier)) -> BuiltinInfo {
             )))),
             loc: None,
         })),
+        Arc::new(RwLock::new(TopLevelDef::Function {
+            name: "ndarray.__add__".into(),
+            simple_name: "__add__".into(),
+            signature: ndarray_add_ty.0,
+            var_id: vec![ndarray_dtype_var_id, ndarray_ndims_var_id],
+            instance_to_symbol: HashMap::default(),
+            instance_to_stmt: HashMap::default(),
+            resolver: None,
+            codegen_callback: Some(Arc::new(GenCall::new(Box::new(
+                |_, _, _, _, _| {
+                    unreachable!("handled in gen_expr")
+                },
+            )))),
+            loc: None,
+        })),
+        Arc::new(RwLock::new(TopLevelDef::Function {
+            name: "ndarray.__sub__".into(),
+            simple_name: "__sub__".into(),
+            signature: ndarray_sub_ty.0,
+            var_id: vec![ndarray_dtype_var_id, ndarray_ndims_var_id],
+            instance_to_symbol: HashMap::default(),
+            instance_to_stmt: HashMap::default(),
+            resolver: None,
+            codegen_callback: Some(Arc::new(GenCall::new(Box::new(
+                |_, _, _, _, _| {
+                    unreachable!("handled in gen_expr")
+                },
+            )))),
+            loc: None,
+        })),
+        Arc::new(RwLock::new(TopLevelDef::Function {
+            name: "ndarray.__mul__".into(),
+            simple_name: "__mul__".into(),
+            signature: ndarray_mul_ty.0,
+            var_id: vec![ndarray_dtype_var_id, ndarray_ndims_var_id],
+            instance_to_symbol: HashMap::default(),
+            instance_to_stmt: HashMap::default(),
+            resolver: None,
+            codegen_callback: Some(Arc::new(GenCall::new(Box::new(
+                |_, _, _, _, _| {
+                    unreachable!("handled in gen_expr")
+                },
+            )))),
+            loc: None,
+        })),
+        Arc::new(RwLock::new(TopLevelDef::Function {
+            name: "ndarray.__truediv__".into(),
+            simple_name: "__truediv__".into(),
+            signature: ndarray_truediv_ty.0,
+            var_id: vec![ndarray_dtype_var_id, ndarray_ndims_var_id],
+            instance_to_symbol: HashMap::default(),
+            instance_to_stmt: HashMap::default(),
+            resolver: None,
+            codegen_callback: Some(Arc::new(GenCall::new(Box::new(
+                |_, _, _, _, _| {
+                    unreachable!("handled in gen_expr")
+                },
+            )))),
+            loc: None,
+        })),
+        Arc::new(RwLock::new(TopLevelDef::Function {
+            name: "ndarray.__floordiv__".into(),
+            simple_name: "__floordiv__".into(),
+            signature: ndarray_floordiv_ty.0,
+            var_id: vec![ndarray_dtype_var_id, ndarray_ndims_var_id],
+            instance_to_symbol: HashMap::default(),
+            instance_to_stmt: HashMap::default(),
+            resolver: None,
+            codegen_callback: Some(Arc::new(GenCall::new(Box::new(
+                |_, _, _, _, _| {
+                    unreachable!("handled in gen_expr")
+                },
+            )))),
+            loc: None,
+        })),
+        Arc::new(RwLock::new(TopLevelDef::Function {
+            name: "ndarray.__mod__".into(),
+            simple_name: "__mod__".into(),
+            signature: ndarray_mod_ty.0,
+            var_id: vec![ndarray_dtype_var_id, ndarray_ndims_var_id],
+            instance_to_symbol: HashMap::default(),
+            instance_to_stmt: HashMap::default(),
+            resolver: None,
+            codegen_callback: Some(Arc::new(GenCall::new(Box::new(
+                |_, _, _, _, _| {
+                    unreachable!("handled in gen_expr")
+                },
+            )))),
+            loc: None,
+        })),
+        Arc::new(RwLock::new(TopLevelDef::Function {
+            name: "ndarray.__pow__".into(),
+            simple_name: "__pow__".into(),
+            signature: ndarray_pow_ty.0,
+            var_id: vec![ndarray_dtype_var_id, ndarray_ndims_var_id],
+            instance_to_symbol: HashMap::default(),
+            instance_to_stmt: HashMap::default(),
+            resolver: None,
+            codegen_callback: Some(Arc::new(GenCall::new(Box::new(
+                |_, _, _, _, _| {
+                    unreachable!("handled in gen_expr")
+                },
+            )))),
+            loc: None,
+        })),
+        Arc::new(RwLock::new(TopLevelDef::Function {
+            name: "ndarray.__iadd__".into(),
+            simple_name: "__iadd__".into(),
+            signature: ndarray_iadd_ty.0,
+            var_id: vec![ndarray_dtype_var_id, ndarray_ndims_var_id, ndarray_usized_ndims_tvar.1],
+            instance_to_symbol: HashMap::default(),
+            instance_to_stmt: HashMap::default(),
+            resolver: None,
+            codegen_callback: Some(Arc::new(GenCall::new(Box::new(
+                |_, _, _, _, _| {
+                    unreachable!("handled in gen_expr")
+                },
+            )))),
+            loc: None,
+        })),
+        Arc::new(RwLock::new(TopLevelDef::Function {
+            name: "ndarray.__isub__".into(),
+            simple_name: "__isub__".into(),
+            signature: ndarray_isub_ty.0,
+            var_id: vec![ndarray_dtype_var_id, ndarray_ndims_var_id],
+            instance_to_symbol: HashMap::default(),
+            instance_to_stmt: HashMap::default(),
+            resolver: None,
+            codegen_callback: Some(Arc::new(GenCall::new(Box::new(
+                |_, _, _, _, _| {
+                    unreachable!("handled in gen_expr")
+                },
+            )))),
+            loc: None,
+        })),
+        Arc::new(RwLock::new(TopLevelDef::Function {
+            name: "ndarray.__imul__".into(),
+            simple_name: "__imul__".into(),
+            signature: ndarray_imul_ty.0,
+            var_id: vec![ndarray_dtype_var_id, ndarray_ndims_var_id],
+            instance_to_symbol: HashMap::default(),
+            instance_to_stmt: HashMap::default(),
+            resolver: None,
+            codegen_callback: Some(Arc::new(GenCall::new(Box::new(
+                |_, _, _, _, _| {
+                    unreachable!("handled in gen_expr")
+                },
+            )))),
+            loc: None,
+        })),
+        Arc::new(RwLock::new(TopLevelDef::Function {
+            name: "ndarray.__itruediv__".into(),
+            simple_name: "__itruediv__".into(),
+            signature: ndarray_itruediv_ty.0,
+            var_id: vec![ndarray_dtype_var_id, ndarray_ndims_var_id],
+            instance_to_symbol: HashMap::default(),
+            instance_to_stmt: HashMap::default(),
+            resolver: None,
+            codegen_callback: Some(Arc::new(GenCall::new(Box::new(
+                |_, _, _, _, _| {
+                    unreachable!("handled in gen_expr")
+                },
+            )))),
+            loc: None,
+        })),
+        Arc::new(RwLock::new(TopLevelDef::Function {
+            name: "ndarray.__ifloordiv__".into(),
+            simple_name: "__ifloordiv__".into(),
+            signature: ndarray_ifloordiv_ty.0,
+            var_id: vec![ndarray_dtype_var_id, ndarray_ndims_var_id],
+            instance_to_symbol: HashMap::default(),
+            instance_to_stmt: HashMap::default(),
+            resolver: None,
+            codegen_callback: Some(Arc::new(GenCall::new(Box::new(
+                |_, _, _, _, _| {
+                    unreachable!("handled in gen_expr")
+                },
+            )))),
+            loc: None,
+        })),
+        Arc::new(RwLock::new(TopLevelDef::Function {
+            name: "ndarray.__imod__".into(),
+            simple_name: "__imod__".into(),
+            signature: ndarray_imod_ty.0,
+            var_id: vec![ndarray_dtype_var_id, ndarray_ndims_var_id],
+            instance_to_symbol: HashMap::default(),
+            instance_to_stmt: HashMap::default(),
+            resolver: None,
+            codegen_callback: Some(Arc::new(GenCall::new(Box::new(
+                |_, _, _, _, _| {
+                    unreachable!("handled in gen_expr")
+                },
+            )))),
+            loc: None,
+        })),
+        Arc::new(RwLock::new(TopLevelDef::Function {
+            name: "ndarray.__ipow__".into(),
+            simple_name: "__ipow__".into(),
+            signature: ndarray_ipow_ty.0,
+            var_id: vec![ndarray_dtype_var_id, ndarray_ndims_var_id],
+            instance_to_symbol: HashMap::default(),
+            instance_to_stmt: HashMap::default(),
+            resolver: None,
+            codegen_callback: Some(Arc::new(GenCall::new(Box::new(
+                |_, _, _, _, _| {
+                    unreachable!("handled in gen_expr")
+                },
+            )))),
+            loc: None,
+        })),
         Arc::new(RwLock::new(TopLevelDef::Function {
             name: "int32".into(),
             simple_name: "int32".into(),

nac3core/src/toplevel/helper.rs

@@ -1,6 +1,7 @@
 use std::convert::TryInto;
 
 use crate::symbol_resolver::SymbolValue;
+use crate::toplevel::numpy::subst_ndarray_tvars;
 use crate::typecheck::typedef::{Mapping, VarMap};
 use nac3parser::ast::{Constant, Location};
 
@@ -226,11 +227,57 @@ impl TopLevelComposer {
                 (ndarray_ndims_tvar.1, ndarray_ndims_tvar.0),
             ]),
         }));
+        let ndarray_binop_fun_other_ty = unifier.get_fresh_var(None, None);
+        let ndarray_binop_fun_ret_ty = unifier.get_fresh_var(None, None);
+        let ndarray_binop_fun_ty = unifier.add_ty(TypeEnum::TFunc(FunSignature {
+            args: vec![
+                FuncArg {
+                    name: "other".into(),
+                    ty: ndarray_binop_fun_other_ty.0,
+                    default_value: None,
+                },
+            ],
+            ret: ndarray_binop_fun_ret_ty.0,
+            vars: VarMap::from([
+                (ndarray_dtype_tvar.1, ndarray_dtype_tvar.0),
+                (ndarray_ndims_tvar.1, ndarray_ndims_tvar.0),
+            ]),
+        }));
+        let ndarray_truediv_fun_other_ty = unifier.get_fresh_var(None, None);
+        let ndarray_truediv_fun_ret_ty = unifier.get_fresh_var(None, None);
+        let ndarray_truediv_fun_ty = unifier.add_ty(TypeEnum::TFunc(FunSignature {
+            args: vec![
+                FuncArg {
+                    name: "other".into(),
+                    ty: ndarray_truediv_fun_other_ty.0,
+                    default_value: None,
+                },
+            ],
+            ret: ndarray_truediv_fun_ret_ty.0,
+            vars: VarMap::from([
+                (ndarray_dtype_tvar.1, ndarray_dtype_tvar.0),
+                (ndarray_ndims_tvar.1, ndarray_ndims_tvar.0),
+            ]),
+        }));
         let ndarray = unifier.add_ty(TypeEnum::TObj {
             obj_id: PRIMITIVE_DEF_IDS.ndarray,
             fields: Mapping::from([
                 ("copy".into(), (ndarray_copy_fun_ty, true)),
                 ("fill".into(), (ndarray_fill_fun_ty, true)),
+                ("__add__".into(), (ndarray_binop_fun_ty, true)),
+                ("__sub__".into(), (ndarray_binop_fun_ty, true)),
+                ("__mul__".into(), (ndarray_binop_fun_ty, true)),
+                ("__truediv__".into(), (ndarray_truediv_fun_ty, true)),
+                ("__floordiv__".into(), (ndarray_binop_fun_ty, true)),
+                ("__mod__".into(), (ndarray_binop_fun_ty, true)),
+                ("__pow__".into(), (ndarray_binop_fun_ty, true)),
+                ("__iadd__".into(), (ndarray_binop_fun_ty, true)),
+                ("__isub__".into(), (ndarray_binop_fun_ty, true)),
+                ("__imul__".into(), (ndarray_binop_fun_ty, true)),
+                ("__itruediv__".into(), (ndarray_truediv_fun_ty, true)),
+                ("__ifloordiv__".into(), (ndarray_binop_fun_ty, true)),
+                ("__imod__".into(), (ndarray_binop_fun_ty, true)),
+                ("__ipow__".into(), (ndarray_binop_fun_ty, true)),
             ]),
             params: VarMap::from([
                 (ndarray_dtype_tvar.1, ndarray_dtype_tvar.0),
@@ -238,7 +285,16 @@ impl TopLevelComposer {
             ]),
         });
 
+        let ndarray_usized_ndims_tvar = unifier.get_fresh_const_generic_var(size_t_ty, Some("ndarray_ndims".into()), None);
+        let ndarray_unsized = subst_ndarray_tvars(&mut unifier, ndarray, Some(ndarray_usized_ndims_tvar.0), None);
+
         unifier.unify(ndarray_copy_fun_ret_ty.0, ndarray).unwrap();
+        unifier.unify(ndarray_binop_fun_other_ty.0, ndarray_unsized).unwrap();
+        unifier.unify(ndarray_binop_fun_ret_ty.0, ndarray).unwrap();
+
+        let ndarray_float = subst_ndarray_tvars(&mut unifier, ndarray, Some(float), None);
+        unifier.unify(ndarray_truediv_fun_other_ty.0, ndarray).unwrap();
+        unifier.unify(ndarray_truediv_fun_ret_ty.0, ndarray_float).unwrap();
 
         let primitives = PrimitiveStore {
             int32,

nac3core/src/toplevel/numpy.rs

@@ -19,13 +19,30 @@ pub fn make_ndarray_ty(
     dtype: Option<Type>,
     ndims: Option<Type>,
 ) -> Type {
-    let ndarray = primitives.ndarray;
+    subst_ndarray_tvars(unifier, primitives.ndarray, dtype, ndims)
+}
 
+/// Substitutes type variables in `ndarray`.
+///
+/// * `dtype` - The element type of the `ndarray`, or [`None`] if the type variable is not
+/// specialized.
+/// * `ndims` - The number of dimensions of the `ndarray`, or [`None`] if the type variable is not
+/// specialized.
+pub fn subst_ndarray_tvars(
+    unifier: &mut Unifier,
+    ndarray: Type,
+    dtype: Option<Type>,
+    ndims: Option<Type>,
+) -> Type {
     let TypeEnum::TObj { obj_id, params, .. } = &*unifier.get_ty_immutable(ndarray) else {
         panic!("Expected `ndarray` to be TObj, but got {}", unifier.stringify(ndarray))
     };
     debug_assert_eq!(*obj_id, PRIMITIVE_DEF_IDS.ndarray);
 
+    if dtype.is_none() && ndims.is_none() {
+        return ndarray
+    }
+
     let tvar_ids = params.iter()
         .map(|(obj_id, _)| *obj_id)
         .collect_vec();
@@ -42,12 +59,10 @@ pub fn make_ndarray_ty(
     unifier.subst(ndarray, &tvar_subst).unwrap_or(ndarray)
 }
 
-/// Unpacks the type variables of `ndarray` into a tuple. The elements of the tuple corresponds to
-/// `dtype` (the element type) and `ndims` (the number of dimensions) of the `ndarray` respectively.
-pub fn unpack_ndarray_tvars(
+fn unpack_ndarray_tvars(
     unifier: &mut Unifier,
     ndarray: Type,
-) -> (Type, Type) {
+) -> Vec<(u32, Type)> {
     let TypeEnum::TObj { obj_id, params, .. } = &*unifier.get_ty_immutable(ndarray) else {
         panic!("Expected `ndarray` to be TObj, but got {}", unifier.stringify(ndarray))
     };
@@ -56,7 +71,33 @@ pub fn unpack_ndarray_tvars(
 
     params.iter()
         .sorted_by_key(|(obj_id, _)| *obj_id)
-        .map(|(_, ty)| *ty)
+        .map(|(var_id, ty)| (*var_id, *ty))
+        .collect_vec()
+}
+
+/// Unpacks the type variable IDs of `ndarray` into a tuple. The elements of the tuple corresponds 
+/// to `dtype` (the element type) and `ndims` (the number of dimensions) of the `ndarray` 
+/// respectively.
+pub fn unpack_ndarray_var_ids(
+    unifier: &mut Unifier,
+    ndarray: Type,
+) -> (u32, u32) {
+    unpack_ndarray_tvars(unifier, ndarray)
+        .into_iter()
+        .map(|v| v.0)
+        .collect_tuple()
+        .unwrap()
+}
+
+/// Unpacks the type variables of `ndarray` into a tuple. The elements of the tuple corresponds to
+/// `dtype` (the element type) and `ndims` (the number of dimensions) of the `ndarray` respectively.
+pub fn unpack_ndarray_var_tys(
+    unifier: &mut Unifier,
+    ndarray: Type,
+) -> (Type, Type) {
+    unpack_ndarray_tvars(unifier, ndarray)
+        .into_iter()
+        .map(|v| v.1)
         .collect_tuple()
         .unwrap()
 }

nac3core/src/typecheck/magic_methods.rs

@@ -1,3 +1,4 @@
+use crate::toplevel::numpy::make_ndarray_ty;
 use crate::typecheck::{
     type_inferencer::*,
     typedef::{FunSignature, FuncArg, Type, TypeEnum, Unifier, VarMap},
@@ -234,8 +235,14 @@ pub fn impl_bitwise_shift(unifier: &mut Unifier, store: &PrimitiveStore, ty: Typ
 }
 
 /// `Div`
-pub fn impl_div(unifier: &mut Unifier, store: &PrimitiveStore, ty: Type, other_ty: &[Type]) {
-    impl_binop(unifier, store, ty, other_ty, store.float, &[Operator::Div]);
+pub fn impl_div(
+    unifier: &mut Unifier,
+    store: &PrimitiveStore,
+    ty: Type,
+    other_ty: &[Type],
+    ret_ty: Type,
+) {
+    impl_binop(unifier, store, ty, other_ty, ret_ty, &[Operator::Div]);
 }
 
 /// `FloorDiv`
@@ -299,8 +306,10 @@ pub fn set_primitives_magic_methods(store: &PrimitiveStore, unifier: &mut Unifie
         bool: bool_t,
         uint32: uint32_t,
         uint64: uint64_t,
+        ndarray: ndarray_t,
         ..
     } = *store;
+    let size_t = store.usize();
 
     /* int ======== */
     for t in [int32_t, int64_t, uint32_t, uint64_t] {
@@ -308,7 +317,7 @@ pub fn set_primitives_magic_methods(store: &PrimitiveStore, unifier: &mut Unifie
         impl_pow(unifier, store, t, &[t], t);
         impl_bitwise_arithmetic(unifier, store, t);
         impl_bitwise_shift(unifier, store, t);
-        impl_div(unifier, store, t, &[t]);
+        impl_div(unifier, store, t, &[t], float_t);
         impl_floordiv(unifier, store, t, &[t], t);
         impl_mod(unifier, store, t, &[t], t);
         impl_invert(unifier, store, t);
@@ -323,7 +332,7 @@ pub fn set_primitives_magic_methods(store: &PrimitiveStore, unifier: &mut Unifie
     /* float ======== */
     impl_basic_arithmetic(unifier, store, float_t, &[float_t], float_t);
     impl_pow(unifier, store, float_t, &[int32_t, float_t], float_t);
-    impl_div(unifier, store, float_t, &[float_t]);
+    impl_div(unifier, store, float_t, &[float_t], float_t);
     impl_floordiv(unifier, store, float_t, &[float_t], float_t);
     impl_mod(unifier, store, float_t, &[float_t], float_t);
     impl_sign(unifier, store, float_t);
@@ -334,4 +343,14 @@ pub fn set_primitives_magic_methods(store: &PrimitiveStore, unifier: &mut Unifie
     /* bool ======== */
     impl_not(unifier, store, bool_t);
     impl_eq(unifier, store, bool_t);
+
+    /* ndarray ===== */
+    let ndarray_float_t = make_ndarray_ty(unifier, store, Some(float_t), None);
+    let ndarray_usized_ndims_tvar = unifier.get_fresh_const_generic_var(size_t, Some("ndarray_ndims".into()), None);
+    let ndarray_unsized_t = make_ndarray_ty(unifier, store, None, Some(ndarray_usized_ndims_tvar.0));
+    impl_basic_arithmetic(unifier, store, ndarray_t, &[ndarray_unsized_t], ndarray_t);
+    impl_pow(unifier, store, ndarray_t, &[ndarray_unsized_t], ndarray_t);
+    impl_div(unifier, store, ndarray_t, &[ndarray_t], ndarray_float_t);
+    impl_floordiv(unifier, store, ndarray_t, &[ndarray_unsized_t], ndarray_t);
+    impl_mod(unifier, store, ndarray_t, &[ndarray_unsized_t], ndarray_t);
 }

nac3core/src/typecheck/type_inferencer/mod.rs

@@ -9,7 +9,7 @@ use crate::{
     symbol_resolver::{SymbolResolver, SymbolValue}, 
     toplevel::{
         helper::PRIMITIVE_DEF_IDS,
-        numpy::{make_ndarray_ty, unpack_ndarray_tvars},
+        numpy::{make_ndarray_ty, unpack_ndarray_var_tys},
         TopLevelContext,
     },
 };
@@ -1334,7 +1334,7 @@ impl<'a> Inferencer<'a> {
                 let list_like_ty = match &*self.unifier.get_ty(value.custom.unwrap()) {
                     TypeEnum::TList { .. } => self.unifier.add_ty(TypeEnum::TList { ty }),
                     TypeEnum::TObj { obj_id, .. } if *obj_id == PRIMITIVE_DEF_IDS.ndarray => {
-                        let (_, ndims) = unpack_ndarray_tvars(self.unifier, value.custom.unwrap());
+                        let (_, ndims) = unpack_ndarray_var_tys(self.unifier, value.custom.unwrap());
 
                         make_ndarray_ty(self.unifier, self.primitives, Some(ty), Some(ndims))
                     }
@@ -1347,7 +1347,7 @@ impl<'a> Inferencer<'a> {
             ExprKind::Constant { value: ast::Constant::Int(val), .. } => {
                 match &*self.unifier.get_ty(value.custom.unwrap()) {
                     TypeEnum::TObj { obj_id, .. } if *obj_id == PRIMITIVE_DEF_IDS.ndarray => {
-                        let (_, ndims) = unpack_ndarray_tvars(self.unifier, value.custom.unwrap());
+                        let (_, ndims) = unpack_ndarray_var_tys(self.unifier, value.custom.unwrap());
                         self.infer_subscript_ndarray(value, ty, ndims)
                     }
                     _ => {
@@ -1379,7 +1379,7 @@ impl<'a> Inferencer<'a> {
                         Ok(ty)
                     }
                     TypeEnum::TObj { obj_id, .. } if *obj_id == PRIMITIVE_DEF_IDS.ndarray => {
-                        let (_, ndims) = unpack_ndarray_tvars(self.unifier, value.custom.unwrap());
+                        let (_, ndims) = unpack_ndarray_var_tys(self.unifier, value.custom.unwrap());
 
                         self.constrain(slice.custom.unwrap(), self.primitives.usize(), &slice.location)?;
                         self.infer_subscript_ndarray(value, ty, ndims)

nac3standalone/demo/src/ndarray.py

@@ -67,6 +67,181 @@ def test_ndarray_copy():
     output_float64(y[1][0])
     output_float64(y[1][1])
 
+def test_ndarray_add():
+    x = np_identity(2)
+    y = x + np_ones([2, 2])
+
+    output_float64(x[0][0])
+    output_float64(x[0][1])
+    output_float64(x[1][0])
+    output_float64(x[1][1])
+
+    output_float64(y[0][0])
+    output_float64(y[0][1])
+    output_float64(y[1][0])
+    output_float64(y[1][1])
+
+# def test_ndarray_add_broadcast():
+#     x = np_identity(2)
+#     y: ndarray[float, 2] = x + np_ones([2])
+#
+#     output_float64(x[0][0])
+#     output_float64(x[0][1])
+#     output_float64(x[1][0])
+#     output_float64(x[1][1])
+#
+#     output_float64(y[0][0])
+#     output_float64(y[0][1])
+#     output_float64(y[1][0])
+#     output_float64(y[1][1])
+
+def test_ndarray_iadd():
+    x = np_identity(2)
+    x += np_ones([2, 2])
+
+    output_float64(x[0][0])
+    output_float64(x[0][1])
+    output_float64(x[1][0])
+    output_float64(x[1][1])
+
+def test_ndarray_sub():
+    x = np_ones([2, 2])
+    y = x - np_identity(2)
+
+    output_float64(x[0][0])
+    output_float64(x[0][1])
+    output_float64(x[1][0])
+    output_float64(x[1][1])
+
+    output_float64(y[0][0])
+    output_float64(y[0][1])
+    output_float64(y[1][0])
+    output_float64(y[1][1])
+
+def test_ndarray_isub():
+    x = np_ones([2, 2])
+    x -= np_identity(2)
+
+    output_float64(x[0][0])
+    output_float64(x[0][1])
+    output_float64(x[1][0])
+    output_float64(x[1][1])
+
+def test_ndarray_mul():
+    x = np_ones([2, 2])
+    y = x * np_identity(2)
+
+    output_float64(x[0][0])
+    output_float64(x[0][1])
+    output_float64(x[1][0])
+    output_float64(x[1][1])
+
+    output_float64(y[0][0])
+    output_float64(y[0][1])
+    output_float64(y[1][0])
+    output_float64(y[1][1])
+
+def test_ndarray_imul():
+    x = np_ones([2, 2])
+    x *= np_identity(2)
+
+    output_float64(x[0][0])
+    output_float64(x[0][1])
+    output_float64(x[1][0])
+    output_float64(x[1][1])
+
+def test_ndarray_truediv():
+    x = np_identity(2)
+    y = x / np_ones([2, 2])
+
+    output_float64(x[0][0])
+    output_float64(x[0][1])
+    output_float64(x[1][0])
+    output_float64(x[1][1])
+
+    output_float64(y[0][0])
+    output_float64(y[0][1])
+    output_float64(y[1][0])
+    output_float64(y[1][1])
+
+def test_ndarray_itruediv():
+    x = np_identity(2)
+    x /= np_ones([2, 2])
+
+    output_float64(x[0][0])
+    output_float64(x[0][1])
+    output_float64(x[1][0])
+    output_float64(x[1][1])
+
+def test_ndarray_floordiv():
+    x = np_identity(2)
+    y = x // np_ones([2, 2])
+
+    output_float64(x[0][0])
+    output_float64(x[0][1])
+    output_float64(x[1][0])
+    output_float64(x[1][1])
+
+    output_float64(y[0][0])
+    output_float64(y[0][1])
+    output_float64(y[1][0])
+    output_float64(y[1][1])
+
+def test_ndarray_ifloordiv():
+    x = np_identity(2)
+    x //= np_ones([2, 2])
+
+    output_float64(x[0][0])
+    output_float64(x[0][1])
+    output_float64(x[1][0])
+    output_float64(x[1][1])
+
+def test_ndarray_mod():
+    x = np_identity(2)
+    y = x % np_full([2, 2], 2.0)
+
+    output_float64(x[0][0])
+    output_float64(x[0][1])
+    output_float64(x[1][0])
+    output_float64(x[1][1])
+
+    output_float64(y[0][0])
+    output_float64(y[0][1])
+    output_float64(y[1][0])
+    output_float64(y[1][1])
+
+def test_ndarray_imod():
+    x = np_identity(2)
+    x %= np_full([2, 2], 2.0)
+
+    output_float64(x[0][0])
+    output_float64(x[0][1])
+    output_float64(x[1][0])
+    output_float64(x[1][1])
+
+def test_ndarray_pow():
+    x = np_identity(2)
+    y = x ** np_full([2, 2], 2.0)
+
+    output_float64(x[0][0])
+    output_float64(x[0][1])
+    output_float64(x[1][0])
+    output_float64(x[1][1])
+
+    output_float64(y[0][0])
+    output_float64(y[0][1])
+    output_float64(y[1][0])
+    output_float64(y[1][1])
+
+def test_ndarray_ipow():
+    x = np_identity(2)
+    x **= np_full([2, 2], 2.0)
+
+    output_float64(x[0][0])
+    output_float64(x[0][1])
+    output_float64(x[1][0])
+    output_float64(x[1][1])
+
 def run() -> int32:
     test_ndarray_ctor()
     test_ndarray_empty()
@@ -77,5 +252,17 @@ def run() -> int32:
     test_ndarray_identity()
     test_ndarray_fill()
     test_ndarray_copy()
+    test_ndarray_add()
+    test_ndarray_iadd()
+    test_ndarray_sub()
+    test_ndarray_isub()
+    test_ndarray_mul()
+    test_ndarray_imul()
+    test_ndarray_truediv()
+    test_ndarray_itruediv()
+    test_ndarray_floordiv()
+    test_ndarray_ifloordiv()
+    test_ndarray_mod()
+    test_ndarray_imod()
 
     return 0