WIP: core/ndstrides: hold

`cargo test -F test` would compile `nac3core/irrt/irrt_test.cpp`
targetted to the host machine (it gets to use `std`) and run the
test executable.

.clang-format

@@ -0,0 +1,3 @@
+BasedOnStyle: Google
+IndentWidth: 4
+ReflowComments: false

flake.nix

@@ -14,6 +14,7 @@
           ''
           mkdir -p $out/bin
           ln -s ${pkgs.llvmPackages_14.clang-unwrapped}/bin/clang $out/bin/clang-irrt
+          ln -s ${pkgs.llvmPackages_14.clang}/bin/clang $out/bin/clang-irrt-test
           ln -s ${pkgs.llvmPackages_14.llvm.out}/bin/llvm-as $out/bin/llvm-as-irrt
           '';
         demo-linalg-stub = pkgs.rustPlatform.buildRustPackage {
@@ -40,6 +41,7 @@
             cargoLock = {
               lockFile = ./Cargo.lock;
             };
+            cargoTestFlags = [ "--features" "test" ];
             passthru.cargoLock = cargoLock;
             nativeBuildInputs = [ pkgs.python3 (pkgs.wrapClangMulti pkgs.llvmPackages_14.clang) llvm-tools-irrt pkgs.llvmPackages_14.llvm.out llvm-nac3 ];
             buildInputs = [ pkgs.python3 llvm-nac3 ];

nac3artiq/src/codegen.rs

@@ -798,8 +798,7 @@ fn polymorphic_print<'ctx>(
             ctx.module.add_function(fn_name, fn_t, None)
         });
 
-        let fmt = ctx.gen_string(generator, fmt);
+        let fmt = ctx.gen_string(generator, &fmt).get_field(generator, ctx.ctx, |f| f.base).value;
-        let fmt = unsafe { fmt.get_field_at_index_unchecked(0) }.into_pointer_value();
 
         ctx.builder
             .build_call(
@@ -879,20 +878,24 @@ fn polymorphic_print<'ctx>(
                 fmt.push_str("%.*s");
 
                 let true_str = ctx.gen_string(generator, "True");
-                let true_data =
+
-                    unsafe { true_str.get_field_at_index_unchecked(0) }.into_pointer_value();
+                let true_data = true_str.get_field(generator, ctx.ctx, |f| f.base);
-                let true_len = unsafe { true_str.get_field_at_index_unchecked(1) }.into_int_value();
+                let true_len = true_str.get_field(generator, ctx.ctx, |f| f.len);
+
                 let false_str = ctx.gen_string(generator, "False");
-                let false_data =
+
-                    unsafe { false_str.get_field_at_index_unchecked(0) }.into_pointer_value();
+                let false_data = false_str.get_field(generator, ctx.ctx, |f| f.base);
-                let false_len =
+                let false_len = false_str.get_field(generator, ctx.ctx, |f| f.len);
-                    unsafe { false_str.get_field_at_index_unchecked(1) }.into_int_value();
 
                 let bool_val = generator.bool_to_i1(ctx, value.into_int_value());
 
                 args.extend([
-                    ctx.builder.build_select(bool_val, true_len, false_len, "").unwrap(),
+                    ctx.builder
-                    ctx.builder.build_select(bool_val, true_data, false_data, "").unwrap(),
+                        .build_select(bool_val, true_len.value, false_len.value, "")
+                        .unwrap(),
+                    ctx.builder
+                        .build_select(bool_val, true_data.value, false_data.value, "")
+                        .unwrap(),
                 ]);
             }
 

nac3artiq/src/lib.rs

@@ -33,6 +33,7 @@ use inkwell::{
     OptimizationLevel,
 };
 use itertools::Itertools;
+use nac3core::codegen::irrt::setup_irrt_exceptions;
 use nac3core::codegen::{gen_func_impl, CodeGenLLVMOptions, CodeGenTargetMachineOptions};
 use nac3core::toplevel::builtins::get_exn_constructor;
 use nac3core::typecheck::typedef::{into_var_map, TypeEnum, Unifier, VarMap};
@@ -557,6 +558,11 @@ impl Nac3 {
             .register_top_level(synthesized.pop().unwrap(), Some(resolver.clone()), "", false)
             .unwrap();
 
+        // Process IRRT
+        let context = inkwell::context::Context::create();
+        let irrt = load_irrt(&context);
+        setup_irrt_exceptions(&context, &irrt, resolver.as_ref());
+
         let fun_signature =
             FunSignature { args: vec![], ret: self.primitive.none, vars: VarMap::new() };
         let mut store = ConcreteTypeStore::new();
@@ -727,7 +733,7 @@ impl Nac3 {
             membuffer.lock().push(buffer);
         });
 
-        let context = inkwell::context::Context::create();
+        // Link all modules into `main`.
         let buffers = membuffers.lock();
         let main = context
             .create_module_from_ir(MemoryBuffer::create_from_memory_range(&buffers[0], "main"))
@@ -756,8 +762,7 @@ impl Nac3 {
             )
             .unwrap();
 
-        main.link_in_module(load_irrt(&context))
+        main.link_in_module(irrt).map_err(|err| CompileError::new_err(err.to_string()))?;
-            .map_err(|err| CompileError::new_err(err.to_string()))?;
 
         let mut function_iter = main.get_first_function();
         while let Some(func) = function_iter {

nac3core/Cargo.toml

@@ -1,3 +1,6 @@
+[features]
+test = []
+
 [package]
 name = "nac3core"
 version = "0.1.0"

nac3core/build.rs

@@ -3,43 +3,60 @@ use std::{
     env,
     fs::File,
     io::Write,
-    path::Path,
+    path::{Path, PathBuf},
     process::{Command, Stdio},
 };
 
-fn main() {
+const CMD_IRRT_CLANG: &str = "clang-irrt";
-    const FILE: &str = "src/codegen/irrt/irrt.cpp";
+const CMD_IRRT_CLANG_TEST: &str = "clang-irrt-test";
+const CMD_IRRT_LLVM_AS: &str = "llvm-as-irrt";
+
+fn get_out_dir() -> PathBuf {
+    PathBuf::from(env::var("OUT_DIR").unwrap())
+}
+
+fn get_irrt_dir() -> &'static Path {
+    Path::new("irrt")
+}
+
+/// Compile `irrt.cpp` for use in `src/codegen`
+fn compile_irrt_cpp() {
+    let out_dir = get_out_dir();
+    let irrt_dir = get_irrt_dir();
 
     /*
      * HACK: Sadly, clang doesn't let us emit generic LLVM bitcode.
      * Compiling for WASM32 and filtering the output with regex is the closest we can get.
      */
-    let flags: &[&str] = &[
+    let irrt_cpp_path = irrt_dir.join("irrt.cpp");
-        "--target=wasm32",
+
-        FILE,
+    let mut flags = vec![];
-        "-x",
+    flags.push("--target=wasm32");
-        "c++",
+    flags.extend(&["-x", "c++"]);
-        "-fno-discard-value-names",
+    flags.extend(&["-fno-discard-value-names", "-fno-exceptions", "-fno-rtti"]);
-        "-fno-exceptions",
+    flags.push("-emit-llvm");
-        "-fno-rtti",
+    flags.push("-S");
+    flags.extend(&["-Wall", "-Wextra"]);
+    flags.extend(&["-o", "-"]);
+    flags.extend(&["-I", irrt_dir.to_str().unwrap()]);
+    flags.push(irrt_cpp_path.to_str().unwrap());
+
     match env::var("PROFILE").as_deref() {
-            Ok("debug") => "-O0",
+        Ok("debug") => {
-            Ok("release") => "-O3",
+            flags.push("-O0");
+            flags.push("-DIRRT_DEBUG");
+        }
+        Ok("release") => {
+            flags.push("-O3");
+        }
         flavor => panic!("Unknown or missing build flavor {flavor:?}"),
-        },
+    };
-        "-emit-llvm",
-        "-S",
-        "-Wall",
-        "-Wextra",
-        "-o",
-        "-",
-    ];
 
-    println!("cargo:rerun-if-changed={FILE}");
+    // Tell Cargo to rerun if any file under `irrt_dir` (recursive) changes
-    let out_dir = env::var("OUT_DIR").unwrap();
+    println!("cargo:rerun-if-changed={}", irrt_dir.to_str().unwrap());
-    let out_path = Path::new(&out_dir);
 
-    let output = Command::new("clang-irrt")
+    // Compile IRRT and capture the LLVM IR output
+    let output = Command::new(CMD_IRRT_CLANG)
         .args(flags)
         .output()
         .map(|o| {
@@ -52,7 +69,17 @@ fn main() {
     let output = std::str::from_utf8(&output.stdout).unwrap().replace("\r\n", "\n");
     let mut filtered_output = String::with_capacity(output.len());
 
-    let regex_filter = Regex::new(r"(?ms:^define.*?\}$)|(?m:^declare.*?$)").unwrap();
+    // Filter out irrelevant IR
+    //
+    // Regex:
+    // - `(?ms:^define.*?\}$)` captures LLVM `define` blocks
+    // - `(?m:^declare.*?$)` captures LLVM `declare` lines
+    // - `(?m:^%.+?=\s*type\s*\{.+?\}$)` captures LLVM `type` declarations
+    // - `(?m:^@.+?=.+$)` captures global constants
+    let regex_filter = Regex::new(
+        r"(?ms:^define.*?\}$)|(?m:^declare.*?$)|(?m:^%.+?=\s*type\s*\{.+?\}$)|(?m:^@.+?=.+$)",
+    )
+    .unwrap();
     for f in regex_filter.captures_iter(&output) {
         assert_eq!(f.len(), 1);
         filtered_output.push_str(&f[0]);
@@ -63,20 +90,71 @@ fn main() {
         .unwrap()
         .replace_all(&filtered_output, "");
 
-    println!("cargo:rerun-if-env-changed=DEBUG_DUMP_IRRT");
+    // For debugging
-    if env::var("DEBUG_DUMP_IRRT").is_ok() {
+    // Doing `DEBUG_DUMP_IRRT=1 cargo build -p nac3core` dumps the LLVM IR generated
-        let mut file = File::create(out_path.join("irrt.ll")).unwrap();
+    const DEBUG_DUMP_IRRT: &str = "DEBUG_DUMP_IRRT";
+    println!("cargo:rerun-if-env-changed={DEBUG_DUMP_IRRT}");
+    if env::var(DEBUG_DUMP_IRRT).is_ok() {
+        let mut file = File::create(out_dir.join("irrt.ll")).unwrap();
         file.write_all(output.as_bytes()).unwrap();
-        let mut file = File::create(out_path.join("irrt-filtered.ll")).unwrap();
+
+        let mut file = File::create(out_dir.join("irrt-filtered.ll")).unwrap();
         file.write_all(filtered_output.as_bytes()).unwrap();
     }
 
-    let mut llvm_as = Command::new("llvm-as-irrt")
+    // Assemble the emitted and filtered IR to .bc
+    // That .bc will be integrated into nac3core's codegen
+    let mut llvm_as = Command::new(CMD_IRRT_LLVM_AS)
         .stdin(Stdio::piped())
         .arg("-o")
-        .arg(out_path.join("irrt.bc"))
+        .arg(out_dir.join("irrt.bc"))
         .spawn()
         .unwrap();
     llvm_as.stdin.as_mut().unwrap().write_all(filtered_output.as_bytes()).unwrap();
     assert!(llvm_as.wait().unwrap().success());
 }
+
+/// Compile `irrt_test.cpp` for testing
+fn compile_irrt_test_cpp() {
+    let out_dir = get_out_dir();
+    let irrt_dir = get_irrt_dir();
+
+    let exe_path = out_dir.join("irrt_test.out"); // Output path of the compiled test executable
+    let irrt_test_cpp_path = irrt_dir.join("irrt_test.cpp");
+    let flags: &[&str] = &[
+        irrt_test_cpp_path.to_str().unwrap(),
+        "-x",
+        "c++",
+        "-I",
+        irrt_dir.to_str().unwrap(),
+        "-g",
+        "-fno-discard-value-names",
+        "-O0",
+        "-Wall",
+        "-Wextra",
+        "-Werror=return-type",
+        "-lm", // for `tgamma()`, `lgamma()`
+        "-o",
+        exe_path.to_str().unwrap(),
+    ];
+
+    Command::new(CMD_IRRT_CLANG_TEST)
+        .args(flags)
+        .output()
+        .map(|o| {
+            assert!(o.status.success(), "{}", std::str::from_utf8(&o.stderr).unwrap());
+            o
+        })
+        .unwrap();
+    println!("cargo:rerun-if-changed={}", irrt_dir.to_str().unwrap());
+}
+
+fn main() {
+    compile_irrt_cpp();
+
+    // https://github.com/rust-lang/cargo/issues/2549
+    // `cargo test -F test` to also build `irrt_test.cpp
+    if cfg!(feature = "test") {
+        compile_irrt_test_cpp();
+    }
+}

nac3core/irrt/irrt.cpp

@@ -0,0 +1,10 @@
+#define IRRT_DEFINE_TYPEDEF_INTS
+#include <irrt_everything.hpp>
+
+/*
+ * All IRRT implementations.
+ *
+ * We don't have pre-compiled objects, so we are writing all implementations in
+ * headers and concatenate them with `#include` into one massive source file that
+ * contains all the IRRT stuff.
+ */

nac3core/irrt/irrt/core.hpp

@@ -1,27 +1,17 @@
-using int8_t = _BitInt(8);
+#pragma once
-using uint8_t = unsigned _BitInt(8);
+
-using int32_t = _BitInt(32);
+#include <irrt/int_defs.hpp>
-using uint32_t = unsigned _BitInt(32);
+#include <irrt/util.hpp>
-using int64_t = _BitInt(64);
-using uint64_t = unsigned _BitInt(64);
 
 // 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`.
+// The type of an index or a value describing the length of a
+// range/slice is always `int32_t`.
 using SliceIndex = int32_t;
 
 namespace {
-template <typename T>
+// adapted from GNU Scientific Library:
-const T& max(const T& a, const T& b) {
+// https://git.savannah.gnu.org/cgit/gsl.git/tree/sys/pow_int.c
-    return a > b ? a : b;
-}
-
-template <typename T>
-const T& min(const T& a, const T& b) {
-    return a > b ? b : a;
-}
-
-// 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
 template <typename T>
 T __nac3_int_exp_impl(T base, T exp) {
@@ -38,12 +28,8 @@ T __nac3_int_exp_impl(T base, T exp) {
 }
 
 template <typename SizeT>
-SizeT __nac3_ndarray_calc_size_impl(
+SizeT __nac3_ndarray_calc_size_impl(const SizeT* list_data, SizeT list_len,
-    const SizeT* list_data,
+                                    SizeT begin_idx, SizeT end_idx) {
-    SizeT list_len,
-    SizeT begin_idx,
-    SizeT end_idx
-) {
     __builtin_assume(end_idx <= list_len);
 
     SizeT num_elems = 1;
@@ -56,12 +42,8 @@ SizeT __nac3_ndarray_calc_size_impl(
 }
 
 template <typename SizeT>
-void __nac3_ndarray_calc_nd_indices_impl(
+void __nac3_ndarray_calc_nd_indices_impl(SizeT index, const SizeT* dims,
-    SizeT index,
+                                         SizeT num_dims, NDIndexInt* idxs) {
-    const SizeT* dims,
-    SizeT num_dims,
-    NDIndex* idxs
-) {
     SizeT stride = 1;
     for (SizeT dim = 0; dim < num_dims; dim++) {
         SizeT i = num_dims - dim - 1;
@@ -72,12 +54,9 @@ void __nac3_ndarray_calc_nd_indices_impl(
 }
 
 template <typename SizeT>
-SizeT __nac3_ndarray_flatten_index_impl(
+SizeT __nac3_ndarray_flatten_index_impl(const SizeT* dims, SizeT num_dims,
-    const SizeT* dims,
+                                        const NDIndexInt* indices,
-    SizeT num_dims,
+                                        SizeT num_indices) {
-    const NDIndex* indices,
-    SizeT num_indices
-) {
     SizeT idx = 0;
     SizeT stride = 1;
     for (SizeT i = 0; i < num_dims; ++i) {
@@ -93,18 +72,17 @@ SizeT __nac3_ndarray_flatten_index_impl(
 }
 
 template <typename SizeT>
-void __nac3_ndarray_calc_broadcast_impl(
+void __nac3_ndarray_calc_broadcast_impl(const SizeT* lhs_dims, SizeT lhs_ndims,
-    const SizeT* lhs_dims,
+                                        const SizeT* rhs_dims, SizeT rhs_ndims,
-    SizeT lhs_ndims,
+                                        SizeT* out_dims) {
-    const SizeT* rhs_dims,
-    SizeT rhs_ndims,
-    SizeT* out_dims
-) {
     SizeT max_ndims = lhs_ndims > rhs_ndims ? lhs_ndims : rhs_ndims;
 
     for (SizeT i = 0; i < max_ndims; ++i) {
-        const SizeT* lhs_dim_sz = i < lhs_ndims ? &lhs_dims[lhs_ndims - i - 1] : nullptr;
+        const SizeT* lhs_dim_sz =
-        const SizeT* rhs_dim_sz = i < rhs_ndims ? &rhs_dims[rhs_ndims - i - 1] : nullptr;
+            i < lhs_ndims ? &lhs_dims[lhs_ndims - i - 1] : nullptr;
+        const SizeT* rhs_dim_sz =
+            i < rhs_ndims ? &rhs_dims[rhs_ndims - i - 1] : nullptr;
+
         SizeT* out_dim = &out_dims[max_ndims - i - 1];
 
         if (lhs_dim_sz == nullptr) {
@@ -124,12 +102,10 @@ void __nac3_ndarray_calc_broadcast_impl(
 }
 
 template <typename SizeT>
-void __nac3_ndarray_calc_broadcast_idx_impl(
+void __nac3_ndarray_calc_broadcast_idx_impl(const SizeT* src_dims,
-    const SizeT* src_dims,
                                             SizeT src_ndims,
-    const NDIndex* in_idx,
+                                            const NDIndexInt* in_idx,
-    NDIndex* out_idx
+                                            NDIndexInt* out_idx) {
-) {
     for (SizeT i = 0; i < src_ndims; ++i) {
         SizeT src_i = src_ndims - i - 1;
         out_idx[src_i] = src_dims[src_i] == 1 ? 0 : in_idx[src_i];
@@ -143,10 +119,10 @@ extern "C" {
         return __nac3_int_exp_impl(base, exp); \
     }
 
-DEF_nac3_int_exp_(int32_t)
+DEF_nac3_int_exp_(int32_t);
-DEF_nac3_int_exp_(int64_t)
+DEF_nac3_int_exp_(int64_t);
-DEF_nac3_int_exp_(uint32_t)
+DEF_nac3_int_exp_(uint32_t);
-DEF_nac3_int_exp_(uint64_t)
+DEF_nac3_int_exp_(uint64_t);
 
 SliceIndex __nac3_slice_index_bound(SliceIndex i, const SliceIndex len) {
     if (i < 0) {
@@ -160,11 +136,8 @@ SliceIndex __nac3_slice_index_bound(SliceIndex i, const SliceIndex len) {
     return i;
 }
 
-SliceIndex __nac3_range_slice_len(
+SliceIndex __nac3_range_slice_len(const SliceIndex start, const SliceIndex end,
-    const SliceIndex start,
+                                  const SliceIndex step) {
-    const SliceIndex end,
-    const SliceIndex step
-) {
     SliceIndex diff = end - start;
     if (diff > 0 && step > 0) {
         return ((diff - 1) / step) + 1;
@@ -180,62 +153,52 @@ SliceIndex __nac3_range_slice_len(
 // - All the index must *not* be out-of-bound or negative,
 // - The end index is *inclusive*,
 // - The length of src and dest slice size should already
-// be checked: if dest.step == 1 then len(src) <= len(dest) else len(src) == len(dest)
+// be checked: if dest.step == 1 then len(src) <= len(dest) else
+// len(src) == len(dest)
 SliceIndex __nac3_list_slice_assign_var_size(
-    SliceIndex dest_start,
+    SliceIndex dest_start, SliceIndex dest_end, SliceIndex dest_step,
-    SliceIndex dest_end,
+    uint8_t* dest_arr, SliceIndex dest_arr_len, SliceIndex src_start,
-    SliceIndex dest_step,
+    SliceIndex src_end, SliceIndex src_step, uint8_t* src_arr,
-    uint8_t* dest_arr,
+    SliceIndex src_arr_len, const SliceIndex size) {
-    SliceIndex dest_arr_len,
+    /* if dest_arr_len == 0, do nothing since we do not support
-    SliceIndex src_start,
+     * extending list
-    SliceIndex src_end,
+     */
-    SliceIndex src_step,
-    uint8_t* src_arr,
-    SliceIndex src_arr_len,
-    const SliceIndex size
-) {
-    /* if dest_arr_len == 0, do nothing since we do not support extending list */
     if (dest_arr_len == 0) return dest_arr_len;
-    /* if both step is 1, memmove directly, handle the dropping of the list, and shrink size */
+    /* if both step is 1, memmove directly, handle the dropping of
+     * the list, and shrink size */
     if (src_step == dest_step && dest_step == 1) {
-        const SliceIndex src_len = (src_end >= src_start) ? (src_end - src_start + 1) : 0;
+        const SliceIndex src_len =
-        const SliceIndex dest_len = (dest_end >= dest_start) ? (dest_end - dest_start + 1) : 0;
+            (src_end >= src_start) ? (src_end - src_start + 1) : 0;
+        const SliceIndex dest_len =
+            (dest_end >= dest_start) ? (dest_end - dest_start + 1) : 0;
         if (src_len > 0) {
-            __builtin_memmove(
+            __builtin_memmove(dest_arr + dest_start * size,
-                dest_arr + dest_start * size,
+                              src_arr + src_start * size, src_len * size);
-                src_arr + src_start * size,
-                src_len * size
-            );
         }
         if (dest_len > 0) {
             /* dropping */
-            __builtin_memmove(
+            __builtin_memmove(dest_arr + (dest_start + src_len) * size,
-                dest_arr + (dest_start + src_len) * size,
                               dest_arr + (dest_end + 1) * size,
-                (dest_arr_len - dest_end - 1) * size
+                              (dest_arr_len - dest_end - 1) * size);
-            );
         }
         /* shrink size */
         return dest_arr_len - (dest_len - src_len);
     }
     /* if two range overlaps, need alloca */
     uint8_t need_alloca =
-        (dest_arr == src_arr)
+        (dest_arr == src_arr) &&
-        && !(
+        !(max(dest_start, dest_end) < min(src_start, src_end) ||
-            max(dest_start, dest_end) < min(src_start, src_end)
+          max(src_start, src_end) < min(dest_start, dest_end));
-            || max(src_start, src_end) < min(dest_start, dest_end)
-        );
     if (need_alloca) {
-        uint8_t* tmp = reinterpret_cast<uint8_t *>(__builtin_alloca(src_arr_len * size));
+        uint8_t* tmp =
+            reinterpret_cast<uint8_t*>(__builtin_alloca(src_arr_len * size));
         __builtin_memcpy(tmp, src_arr, src_arr_len * size);
         src_arr = tmp;
     }
     SliceIndex src_ind = src_start;
     SliceIndex dest_ind = dest_start;
-    for (;
+    for (; (src_step > 0) ? (src_ind <= src_end) : (src_ind >= src_end);
-        (src_step > 0) ? (src_ind <= src_end) : (src_ind >= src_end);
+         src_ind += src_step, dest_ind += dest_step) {
-        src_ind += src_step, dest_ind += dest_step
-    ) {
         /* for constant optimization */
         if (size == 1) {
             __builtin_memcpy(dest_arr + dest_ind, src_arr + src_ind, 1);
@@ -244,30 +207,26 @@ SliceIndex __nac3_list_slice_assign_var_size(
         } else if (size == 8) {
             __builtin_memcpy(dest_arr + dest_ind * 8, src_arr + src_ind * 8, 8);
         } else {
-            /* memcpy for var size, cannot overlap after previous alloca */
+            /* memcpy for var size, cannot overlap after previous
-            __builtin_memcpy(dest_arr + dest_ind * size, src_arr + src_ind * size, size);
+             * alloca */
+            __builtin_memcpy(dest_arr + dest_ind * size,
+                             src_arr + src_ind * size, size);
         }
     }
     /* only dest_step == 1 can we shrink the dest list. */
     /* size should be ensured prior to calling this function */
     if (dest_step == 1 && dest_end >= dest_start) {
         __builtin_memmove(
-            dest_arr + dest_ind * size,
+            dest_arr + dest_ind * size, dest_arr + (dest_end + 1) * size,
-            dest_arr + (dest_end + 1) * size,
+            (dest_arr_len - dest_end - 1) * size + size + size + size);
-            (dest_arr_len - dest_end - 1) * size
-        );
         return dest_arr_len - (dest_end - dest_ind) - 1;
     }
     return dest_arr_len;
 }
 
-int32_t __nac3_isinf(double x) {
+int32_t __nac3_isinf(double x) { return __builtin_isinf(x); }
-    return __builtin_isinf(x);
-}
 
-int32_t __nac3_isnan(double x) {
+int32_t __nac3_isnan(double x) { return __builtin_isnan(x); }
-    return __builtin_isnan(x);
-}
 
 double tgamma(double arg);
 
@@ -320,95 +279,71 @@ double __nac3_j0(double x) {
     return j0(x);
 }
 
-uint32_t __nac3_ndarray_calc_size(
+uint32_t __nac3_ndarray_calc_size(const uint32_t* list_data, uint32_t list_len,
-    const uint32_t* list_data,
+                                  uint32_t begin_idx, uint32_t end_idx) {
-    uint32_t list_len,
+    return __nac3_ndarray_calc_size_impl(list_data, list_len, begin_idx,
-    uint32_t begin_idx,
+                                         end_idx);
-    uint32_t end_idx
-) {
-    return __nac3_ndarray_calc_size_impl(list_data, list_len, begin_idx, end_idx);
 }
 
-uint64_t __nac3_ndarray_calc_size64(
+uint64_t __nac3_ndarray_calc_size64(const uint64_t* list_data,
-    const uint64_t* list_data,
+                                    uint64_t list_len, uint64_t begin_idx,
-    uint64_t list_len,
+                                    uint64_t end_idx) {
-    uint64_t begin_idx,
+    return __nac3_ndarray_calc_size_impl(list_data, list_len, begin_idx,
-    uint64_t end_idx
+                                         end_idx);
-) {
-    return __nac3_ndarray_calc_size_impl(list_data, list_len, begin_idx, end_idx);
 }
 
-void __nac3_ndarray_calc_nd_indices(
+void __nac3_ndarray_calc_nd_indices(uint32_t index, const uint32_t* dims,
-    uint32_t index,
+                                    uint32_t num_dims, NDIndexInt* idxs) {
-    const uint32_t* dims,
-    uint32_t num_dims,
-    NDIndex* idxs
-) {
     __nac3_ndarray_calc_nd_indices_impl(index, dims, num_dims, idxs);
 }
 
-void __nac3_ndarray_calc_nd_indices64(
+void __nac3_ndarray_calc_nd_indices64(uint64_t index, const uint64_t* dims,
-    uint64_t index,
+                                      uint64_t num_dims, NDIndexInt* idxs) {
-    const uint64_t* dims,
-    uint64_t num_dims,
-    NDIndex* idxs
-) {
     __nac3_ndarray_calc_nd_indices_impl(index, dims, num_dims, idxs);
 }
 
-uint32_t __nac3_ndarray_flatten_index(
+uint32_t __nac3_ndarray_flatten_index(const uint32_t* dims, uint32_t num_dims,
-    const uint32_t* dims,
+                                      const NDIndexInt* indices,
-    uint32_t num_dims,
+                                      uint32_t num_indices) {
-    const NDIndex* indices,
+    return __nac3_ndarray_flatten_index_impl(dims, num_dims, indices,
-    uint32_t num_indices
+                                             num_indices);
-) {
-    return __nac3_ndarray_flatten_index_impl(dims, num_dims, indices, num_indices);
 }
 
-uint64_t __nac3_ndarray_flatten_index64(
+uint64_t __nac3_ndarray_flatten_index64(const uint64_t* dims, uint64_t num_dims,
-    const uint64_t* dims,
+                                        const NDIndexInt* indices,
-    uint64_t num_dims,
+                                        uint64_t num_indices) {
-    const NDIndex* indices,
+    return __nac3_ndarray_flatten_index_impl(dims, num_dims, indices,
-    uint64_t num_indices
+                                             num_indices);
-) {
-    return __nac3_ndarray_flatten_index_impl(dims, num_dims, indices, num_indices);
 }
 
-void __nac3_ndarray_calc_broadcast(
+void __nac3_ndarray_calc_broadcast(const uint32_t* lhs_dims, uint32_t lhs_ndims,
-    const uint32_t* lhs_dims,
+                                   const uint32_t* rhs_dims, uint32_t rhs_ndims,
-    uint32_t lhs_ndims,
+                                   uint32_t* out_dims) {
-    const uint32_t* rhs_dims,
+    return __nac3_ndarray_calc_broadcast_impl(lhs_dims, lhs_ndims, rhs_dims,
-    uint32_t rhs_ndims,
+                                              rhs_ndims, out_dims);
-    uint32_t* out_dims
-) {
-    return __nac3_ndarray_calc_broadcast_impl(lhs_dims, lhs_ndims, rhs_dims, rhs_ndims, out_dims);
 }
 
-void __nac3_ndarray_calc_broadcast64(
+void __nac3_ndarray_calc_broadcast64(const uint64_t* lhs_dims,
-    const uint64_t* lhs_dims,
                                      uint64_t lhs_ndims,
                                      const uint64_t* rhs_dims,
-    uint64_t rhs_ndims,
+                                     uint64_t rhs_ndims, uint64_t* out_dims) {
-    uint64_t* out_dims
+    return __nac3_ndarray_calc_broadcast_impl(lhs_dims, lhs_ndims, rhs_dims,
-) {
+                                              rhs_ndims, out_dims);
-    return __nac3_ndarray_calc_broadcast_impl(lhs_dims, lhs_ndims, rhs_dims, rhs_ndims, out_dims);
 }
 
-void __nac3_ndarray_calc_broadcast_idx(
+void __nac3_ndarray_calc_broadcast_idx(const uint32_t* src_dims,
-    const uint32_t* src_dims,
                                        uint32_t src_ndims,
-    const NDIndex* in_idx,
+                                       const NDIndexInt* in_idx,
-    NDIndex* out_idx
+                                       NDIndexInt* out_idx) {
-) {
+    __nac3_ndarray_calc_broadcast_idx_impl(src_dims, src_ndims, in_idx,
-    __nac3_ndarray_calc_broadcast_idx_impl(src_dims, src_ndims, in_idx, out_idx);
+                                           out_idx);
 }
 
-void __nac3_ndarray_calc_broadcast_idx64(
+void __nac3_ndarray_calc_broadcast_idx64(const uint64_t* src_dims,
-    const uint64_t* src_dims,
                                          uint64_t src_ndims,
-    const NDIndex* in_idx,
+                                         const NDIndexInt* in_idx,
-    NDIndex* out_idx
+                                         NDIndexInt* out_idx) {
-) {
+    __nac3_ndarray_calc_broadcast_idx_impl(src_dims, src_ndims, in_idx,
-    __nac3_ndarray_calc_broadcast_idx_impl(src_dims, src_ndims, in_idx, out_idx);
+                                           out_idx);
 }
 }  // extern "C"

nac3core/irrt/irrt/cslice.hpp

@@ -0,0 +1,9 @@
+#pragma once
+
+#include <irrt/int_defs.hpp>
+
+template <typename SizeT>
+struct CSlice {
+    uint8_t* base;
+    SizeT len;
+};

nac3core/irrt/irrt/debug.hpp

@@ -0,0 +1,15 @@
+#pragma once
+
+#define raise_debug_assert(SizeT, msg, param1, param2, param3) \
+    raise_exception(SizeT, EXN_ASSERTION_ERROR,                \
+                    "IRRT debug assert failed: " msg, param1, param2, param3);
+
+#define debug_assert_eq(SizeT, lhs, rhs)                                       \
+    if (IRRT_DEBUG_ASSERT_BOOL && (lhs) != (rhs)) {                            \
+        raise_debug_assert(SizeT, "LHS = {0}. RHS = {1}", lhs, rhs, NO_PARAM); \
+    }
+
+#define debug_assert(SizeT, expr)                                              \
+    if (IRRT_DEBUG_ASSERT_BOOL && !(expr)) {                                   \
+        raise_debug_assert(SizeT, "Got false.", NO_PARAM, NO_PARAM, NO_PARAM); \
+    }

nac3core/irrt/irrt/exception.hpp

@@ -0,0 +1,123 @@
+#pragma once
+
+#include <irrt/cslice.hpp>
+#include <irrt/int_defs.hpp>
+#include <irrt/util.hpp>
+
+/**
+ * @brief The int type of ARTIQ exception IDs.
+ *
+ * It is always `int32_t`
+ */
+typedef int32_t ExceptionId;
+
+/*
+ * A set of exceptions IRRT can use.
+ * Must be synchronized with `setup_irrt_exceptions` in `nac3core/src/codegen/irrt/mod.rs`.
+ * All exception IDs are initialized by `setup_irrt_exceptions`.
+ */
+#ifdef IRRT_TESTING
+// If we are doing IRRT tests (i.e., running `cargo test -F test`), define them with a fake set of IDs.
+ExceptionId EXN_INDEX_ERROR = 0;
+ExceptionId EXN_VALUE_ERROR = 1;
+ExceptionId EXN_ASSERTION_ERROR = 2;
+ExceptionId EXN_RUNTIME_ERROR = 3;
+ExceptionId EXN_TYPE_ERROR = 4;
+#else
+extern "C" {
+ExceptionId EXN_INDEX_ERROR;
+ExceptionId EXN_VALUE_ERROR;
+ExceptionId EXN_ASSERTION_ERROR;
+ExceptionId EXN_RUNTIME_ERROR;
+ExceptionId EXN_TYPE_ERROR;
+}
+#endif
+
+namespace {
+/**
+ * @brief NAC3's Exception struct
+ */
+template <typename SizeT>
+struct Exception {
+    ExceptionId id;
+    CSlice<SizeT> filename;
+    int32_t line;
+    int32_t column;
+    CSlice<SizeT> function;
+    CSlice<SizeT> msg;
+    int64_t params[3];
+};
+}  // namespace
+
+// Declare/Define `__nac3_raise`
+#ifdef IRRT_TESTING
+#include <cstdio>
+void __nac3_raise(void* err) {
+    // TODO: Print the error content?
+    printf("__nac3_raise called. Exiting...\n");
+    exit(1);
+}
+#else
+/**
+ * @brief Extern function to `__nac3_raise`
+ *
+ * The parameter `err` could be `Exception<int32_t>` or `Exception<int64_t>`. The caller
+ * must make sure to pass `Exception`s with the correct `SizeT` depending on the `size_t` of the runtime.
+ */
+extern "C" void __nac3_raise(void* err);
+#endif
+
+namespace {
+const int64_t NO_PARAM = 0;
+
+// Helper function to raise an exception with `__nac3_raise`
+// Do not use this function directly. See `raise_exception`.
+template <typename SizeT>
+void _raise_exception_helper(ExceptionId id, const char* filename, int32_t line,
+                             const char* function, const char* msg,
+                             int64_t param0, int64_t param1, int64_t param2) {
+    Exception<SizeT> e = {
+        .id = id,
+        .filename = {.base = (uint8_t*)filename,
+                     .len = (int32_t)cstr_utils::length(filename)},
+        .line = line,
+        .column = 0,
+        .function = {.base = (uint8_t*)function,
+                     .len = (int32_t)cstr_utils::length(function)},
+        .msg = {.base = (uint8_t*)msg, .len = (int32_t)cstr_utils::length(msg)},
+    };
+    e.params[0] = param0;
+    e.params[1] = param1;
+    e.params[2] = param2;
+    __nac3_raise((void*)&e);
+    __builtin_unreachable();
+}
+
+/**
+ * @brief Raise an exception with location details (location in the IRRT source files).
+ * @param SizeT The runtime `size_t` type.
+ * @param id The ID of the exception to raise.
+ * @param msg A global constant C-string of the error message.
+ *
+ * `param0` and `param2` are optional format arguments of `msg`. They should be set to
+ * `NO_PARAM` to indicate they are unused.
+ */
+#define raise_exception(SizeT, id, msg, param0, param1, param2)               \
+    _raise_exception_helper<SizeT>(id, __FILE__, __LINE__, __FUNCTION__, msg, \
+                                   param0, param1, param2)
+
+/**
+ * @brief Throw a dummy error for testing.
+ */
+template <typename SizeT>
+void throw_dummy_error() {
+    raise_exception(SizeT, EXN_RUNTIME_ERROR, "dummy error", NO_PARAM, NO_PARAM,
+                    NO_PARAM);
+}
+}  // namespace
+
+extern "C" {
+void __nac3_throw_dummy_error() { throw_dummy_error<int32_t>(); }
+
+void __nac3_throw_dummy_error64() { throw_dummy_error<int64_t>(); }
+}

nac3core/irrt/irrt/int_defs.hpp

@@ -0,0 +1,12 @@
+#pragma once
+
+// This is made toggleable since `irrt_test.cpp` itself would include
+// headers that define these typedefs
+#ifdef IRRT_DEFINE_TYPEDEF_INTS
+using int8_t = _BitInt(8);
+using uint8_t = unsigned _BitInt(8);
+using int32_t = _BitInt(32);
+using uint32_t = unsigned _BitInt(32);
+using int64_t = _BitInt(64);
+using uint64_t = unsigned _BitInt(64);
+#endif

nac3core/irrt/irrt/list.hpp

@@ -0,0 +1,56 @@
+#pragma once
+
+#include <irrt/int_defs.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 list {
+template <typename SizeT>
+void slice_assign(List<SizeT>* dst, List<SizeT>* src, SizeT itemsize,
+                  UserSlice* user_slice) {
+    Slice slice = user_slice->indices_checked<SizeT>(dst->len);
+
+    // NOTE: Python does not have this restriction.
+    if (slice.len() != src->len) {
+        raise_exception(SizeT, EXN_VALUE_ERROR,
+                        "List destination has {} item(s), but source has {} "
+                        "item(s). The lengths must match.",
+                        slice.len(), src->len, NO_PARAM);
+    }
+
+    // TODO: Look into how the original implementation was implemented and optimized.
+    SizeT dst_i = slice.start;
+    SizeT src_i = 0;
+    while (src_i < slice.len()) {
+        __builtin_memcpy(dst->items + dst_i, src->items + src_i, itemsize);
+
+        src_i += 1;
+        dst_i += slice.step;
+    }
+}
+}  // namespace list
+}  // namespace
+
+extern "C" {
+void __nac3_list_slice_assign(List<int32_t>* dst, List<int32_t>* src,
+                              int32_t itemsize, UserSlice* user_slice) {
+    list::slice_assign(dst, src, itemsize, user_slice);
+}
+
+void __nac3_list_slice_assign64(List<int64_t>* dst, List<int64_t>* src,
+                                int64_t itemsize, UserSlice* user_slice) {
+    list::slice_assign(dst, src, itemsize, user_slice);
+}
+}

nac3core/irrt/irrt/ndarray/array.hpp

@@ -0,0 +1,119 @@
+#pragma once
+
+#include <irrt/int_defs.hpp>
+#include <irrt/list.hpp>
+#include <irrt/ndarray/basic.hpp>
+#include <irrt/ndarray/def.hpp>
+
+namespace {
+namespace ndarray {
+namespace array {
+// TODO: Document me
+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);
+}
+
+// TODO: Document me
+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) {
+    // done after set_and_validate(list, ndims, shape), list is well-formed
+    // ndarray->data is allocated and owned
+    // ndarray->itemsize is set
+    // ndarray->ndims is set
+    // ndarray->shape is set
+    // ndarray->strides is ???
+    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_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_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_array_write_list_to_array(List<int32_t>* list,
+                                      NDArray<int32_t>* ndarray) {
+    write_list_to_array(list, ndarray);
+}
+
+void __nac3_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 +1,345 @@
+#pragma once
+
+#include <irrt/exception.hpp>
+#include <irrt/int_defs.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);
+}
+
+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 +1,171 @@
+#pragma once
+
+#include <irrt/int_defs.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 +1,44 @@
+#pragma once
+
+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 +1,221 @@
+#pragma once
+
+#include <irrt/exception.hpp>
+#include <irrt/int_defs.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 `UserRange`.
+ */
+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[indexes]` 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 `indexes`, and return the
+ * indexed array (by writing the result to `dst_ndarray`).
+ *
+ * This function also does proper assertions on `indexes`.
+ *
+ * # 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 `indexes`.
+ *   - `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 indexes Indexes 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_indexes, const NDIndex* indexes,
+           const NDArray<SizeT>* src_ndarray, NDArray<SizeT>* dst_ndarray) {
+    // First, validate `indexes`.
+
+    // 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 `indexes`. There can only be 0 or 1 ellipsis.
+    SizeT num_ellipsis = 0;
+
+    for (SizeT i = 0; i < num_indexes; i++) {
+        if (indexes[i].type == ND_INDEX_TYPE_SINGLE_ELEMENT) {
+            expected_dst_ndims--;
+            num_indexed++;
+        } else if (indexes[i].type == ND_INDEX_TYPE_SLICE) {
+            num_indexed++;
+        } else if (indexes[i].type == ND_INDEX_TYPE_NEWAXIS) {
+            expected_dst_ndims++;
+        } else if (indexes[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_indexes, 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_indexes; i++) {
+        const NDIndex* index = &indexes[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) {
+            UserSlice* input = (UserSlice*)index->data;
+
+            Slice slice =
+                input->indices_checked<SizeT>(src_ndarray->shape[src_axis]);
+
+            dst_ndarray->data +=
+                (SizeT)slice.start * src_ndarray->strides[src_axis];
+            dst_ndarray->strides[dst_axis] =
+                ((SizeT)slice.step) * src_ndarray->strides[src_axis];
+            dst_ndarray->shape[dst_axis] = (SizeT)slice.len();
+
+            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_indexes, NDIndex* indexes,
+                          NDArray<int32_t>* src_ndarray,
+                          NDArray<int32_t>* dst_ndarray) {
+    index(num_indexes, indexes, src_ndarray, dst_ndarray);
+}
+
+void __nac3_ndarray_index64(int64_t num_indexes, NDIndex* indexes,
+                            NDArray<int64_t>* src_ndarray,
+                            NDArray<int64_t>* dst_ndarray) {
+    index(num_indexes, indexes, src_ndarray, dst_ndarray);
+}
+}

nac3core/irrt/irrt/ndarray/iter.hpp

@@ -0,0 +1,118 @@
+#pragma once
+
+#include <irrt/int_defs.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/product.hpp

@@ -0,0 +1,194 @@
+#pragma once
+
+#include <irrt/int_defs.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 {
+
+/*
+ * In einsum notation, the output is the broadcasts performed by `np.einsum("...ij,...jk->...ik", a, b)`.
+ *
+ * Example:
+ *   Suppose      `a_shape == [99,  1, 97, 4, 2]`
+ *       and      `b_shape == [ 1, 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]`.
+ *                             ^^^^^^^^^^  ^^^^
+ *                         (by broadcast)  (4x2 @ 2x5 => 4x5)
+ */
+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);
+
+    const SizeT num_entries = 2;
+    ShapeEntry<SizeT> entries[num_entries] = {
+        {.ndims = a_ndims - 2, .shape = a_shape},
+        {.ndims = b_ndims - 2, .shape = b_shape}};
+
+    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];
+}
+
+/**
+ * @brief Perform `np.matmul(a, b)` but the inputs are both rank >=2 matrices and `a.shape[:-2] == b.shape[:-2]`.
+ *
+ * The compatibility of `a` and `b` (for their `.shape[-2:]`) are asserted.
+ *
+ * Also see https://numpy.org/doc/stable/reference/generated/numpy.matmul.html#numpy-matmul.
+ *
+ * This function expects `dst_ndarray` to contain the following content when called:
+ * - `dst_ndarray->data` is allocated. Can be uninitialized.
+ * - `dst_ndarray->itemsize` is set to `sizeof(T)`.
+ * - `dst_ndarray->ndims` is set appropriately.
+ * - `dst_ndarray->shape` is set appropriately.
+ * - `dst_ndarray->strides` is ignored.
+ *
+ * Moreover, the shapes of `a_ndarray`, `b_ndarray`, and `dst_ndarray` **must be the same**. This implies
+ */
+template <typename SizeT, typename T>
+void matmul_at_least_2d(NDArray<SizeT>* a_ndarray, NDArray<SizeT>* b_ndarray,
+                        NDArray<SizeT>* dst_ndarray) {
+    // All inputs' ndims should be >= 2 and be the same.
+    debug_assert_eq(SizeT, a_ndarray->ndims, b_ndarray->ndims);
+    debug_assert_eq(SizeT, a_ndarray->ndims, dst_ndarray->ndims);
+    debug_assert(SizeT, a_ndarray->ndims >= 2);
+
+    debug_assert_eq(SizeT, a_ndarray->itemsize, sizeof(T));
+    debug_assert_eq(SizeT, b_ndarray->itemsize, sizeof(T));
+    debug_assert_eq(SizeT, dst_ndarray->itemsize, sizeof(T));
+
+    if (IRRT_DEBUG_ASSERT_BOOL) {
+        // Check that the shapes are the same.
+        for (SizeT i = 0; i < a_ndarray->ndims - 2; i++) {
+            if (dst_ndarray->shape[0] != a_ndarray->shape[0]) {
+                raise_debug_assert(
+                    SizeT, "Bad shape. At axis {0}, a has {1}, dst has {2}", i,
+                    a_ndarray->shape[i], dst_ndarray->shape[i]);
+            }
+            if (dst_ndarray->shape[0] != b_ndarray->shape[0]) {
+                raise_debug_assert(
+                    SizeT, "Bad shape. At axis {0}, b has {1}, dst has {2}", i,
+                    b_ndarray->shape[i], dst_ndarray->shape[i]);
+            }
+        }
+    }
+
+    // Number of dimensions dedicated to stacking
+    // e.g., [4, 6, 1, 2, 3]
+    //        ^^^^^^^ count these
+    const SizeT u = a_ndarray->ndims - 2;  // Alias
+
+    SizeT* a_mat_shape = a_ndarray->shape + u;
+    SizeT* b_mat_shape = b_ndarray->shape + u;
+    SizeT* dst_mat_shape = dst_ndarray->shape + u;
+
+    // Assert that dst_ndarray has the correct shape
+    debug_assert_eq(SizeT, dst_mat_shape[0], a_mat_shape[0]);
+    debug_assert_eq(SizeT, dst_mat_shape[1], b_mat_shape[1]);
+
+    // Check that a and b are compatible for matmul
+    if (a_mat_shape[1] != b_mat_shape[0]) {
+        // 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_mat_shape[1], b_mat_shape[0], NO_PARAM);
+    }
+
+    // Iterate through shape[:-2]. i.e,
+    // Given a = [5, 4, 3, m, p] and b = [5, 4, 3, p, n]. We iterate through [5, 4, 3].
+    SizeT* indices =
+        (SizeT*)__builtin_alloca(sizeof(SizeT) * dst_ndarray->ndims);
+    SizeT* mat_indices = indices + u;
+    NDIter<SizeT> iter;
+    iter.initialize(u, dst_ndarray->shape, dst_ndarray->strides,
+                    dst_ndarray->data, indices);
+
+    for (; iter.has_next(); iter.next()) {
+        for (SizeT i = 0; i < dst_mat_shape[0]; i++) {
+            for (SizeT j = 0; j < dst_mat_shape[1]; j++) {
+                // `indices` is being reused to index into different ndarrays.
+                mat_indices[0] = i;
+                mat_indices[1] = j;
+                T* d = ndarray::basic::get_ptr<SizeT, T>(dst_ndarray, indices);
+                *d = 0;
+
+                for (SizeT k = 0; k < a_ndarray->shape[1]; k++) {
+                    mat_indices[0] = i;
+                    mat_indices[1] = k;
+                    T* a =
+                        ndarray::basic::get_ptr<SizeT, T>(a_ndarray, indices);
+
+                    mat_indices[0] = k;
+                    mat_indices[1] = j;
+                    T* b =
+                        ndarray::basic::get_ptr<SizeT, T>(b_ndarray, indices);
+
+                    *d += (*a) * (*b);
+                }
+            }
+        }
+    }
+}
+}  // 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);
+}
+
+void __nac3_ndarray_float64_matmul_at_least_2d(NDArray<int32_t>* a_ndarray,
+                                               NDArray<int32_t>* b_ndarray,
+                                               NDArray<int32_t>* dst_ndarray) {
+    matmul_at_least_2d<int32_t, double>(a_ndarray, b_ndarray, dst_ndarray);
+}
+
+void __nac3_ndarray_float64_matmul_at_least_2d64(
+    NDArray<int64_t>* a_ndarray, NDArray<int64_t>* b_ndarray,
+    NDArray<int64_t>* dst_ndarray) {
+    matmul_at_least_2d<int64_t, double>(a_ndarray, b_ndarray, dst_ndarray);
+}
+}

nac3core/irrt/irrt/ndarray/reshape.hpp

@@ -0,0 +1,106 @@
+#pragma once
+
+#include <irrt/int_defs.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_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_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 +1,145 @@
+#pragma once
+
+#include <irrt/int_defs.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/slice.hpp

@@ -0,0 +1,167 @@
+#pragma once
+
+#include <irrt/int_defs.hpp>
+#include <irrt/slice.hpp>
+#include <irrt/util.hpp>
+
+#include "exception.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-like slice with resolved indices.
+ *
+ * "Resolved indices" means that `start` and `stop` must be positive and are
+ * bound to a known length.
+ */
+struct Slice {
+    SliceIndex start;
+    SliceIndex stop;
+    SliceIndex step;
+
+    /**
+     * @brief Calculate and return the length / the number of the slice.
+     * 
+     * If this were a Python range, this function would be `len(range(start, stop, step))`.
+     */
+    SliceIndex len() {
+        SliceIndex diff = stop - start;
+        if (diff > 0 && step > 0) {
+            return ((diff - 1) / step) + 1;
+        } else if (diff < 0 && step < 0) {
+            return ((diff + 1) / step) + 1;
+        } 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 UserSlice {
+    bool start_defined;
+    SliceIndex start;
+
+    bool stop_defined;
+    SliceIndex stop;
+
+    bool step_defined;
+    SliceIndex step;
+
+    UserSlice() { 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.
+     *
+     * In Python, this would be `slice(start, stop, step).indices(length)`.
+     *
+     * @return A `Slice` with the resolved indices.
+     */
+    Slice indices(SliceIndex length) {
+        Slice result;
+
+        result.step = step_defined ? step : 1;
+        bool step_is_negative = result.step < 0;
+
+        if (start_defined) {
+            result.start =
+                slice::resolve_index_in_length_clamped(length, start);
+        } else {
+            result.start = step_is_negative ? length - 1 : 0;
+        }
+
+        if (stop_defined) {
+            result.stop = slice::resolve_index_in_length_clamped(length, stop);
+        } else {
+            result.stop = step_is_negative ? -1 : length;
+        }
+
+        return result;
+    }
+
+    /**
+     * @brief Like `.indices()` but with assertions.
+     */
+    template <typename SizeT>
+    Slice 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(length);
+    }
+};
+}  // namespace

nac3core/irrt/irrt/util.hpp

@@ -0,0 +1,101 @@
+#pragma once
+
+namespace {
+template <typename T>
+const T& max(const T& a, const T& b) {
+    return a > b ? a : b;
+}
+
+template <typename T>
+const T& min(const T& a, const T& b) {
+    return a > b ? b : a;
+}
+
+template <typename T>
+bool arrays_match(int len, T* as, T* bs) {
+    for (int i = 0; i < len; i++) {
+        if (as[i] != bs[i]) return false;
+    }
+    return true;
+}
+
+namespace cstr_utils {
+/**
+ * @brief Return true if `str` is empty.
+ */
+bool is_empty(const char* str) { return str[0] == '\0'; }
+
+/**
+ * @brief Implementation of `strcmp()`
+ */
+int8_t compare(const char* a, const char* b) {
+    uint32_t i = 0;
+    while (true) {
+        if (a[i] < b[i]) {
+            return -1;
+        } else if (a[i] > b[i]) {
+            return 1;
+        } else {
+            if (a[i] == '\0') {
+                return 0;
+            } else {
+                i++;
+            }
+        }
+    }
+}
+
+/**
+ * @brief Return true two strings have the same content.
+ */
+int8_t equal(const char* a, const char* b) { return compare(a, b) == 0; }
+
+/**
+ * @brief Implementation of `strlen()`.
+ */
+uint32_t length(const char* str) {
+    uint32_t length = 0;
+    while (*str != '\0') {
+        length++;
+        str++;
+    }
+    return length;
+}
+
+/**
+ * @brief Copy a null-terminated string to a buffer with limited size and guaranteed null-termination.
+ *
+ * `dst_max_size` must be greater than 0, otherwise this function has undefined behavior.
+ *
+ * This function attempts to copy everything from `src` from `dst`, and *always* null-terminates `dst`.
+ *
+ * If the size of `dst` is too small, the final byte (`dst[dst_max_size - 1]`) of `dst` will be set to
+ * the null terminator.
+ *
+ * @param src String to copy from.
+ * @param dst Buffer to copy string to.
+ * @param dst_max_size
+ *   Number of bytes of this buffer, including the space needed for the null terminator.
+ *   Must be greater than 0.
+ * @return If `dst` is too small to contain everything in `src`.
+ */
+bool copy(const char* src, char* dst, uint32_t dst_max_size) {
+    for (uint32_t i = 0; i < dst_max_size; i++) {
+        bool is_last = i + 1 == dst_max_size;
+        if (is_last && src[i] != '\0') {
+            dst[i] = '\0';
+            return false;
+        }
+
+        if (src[i] == '\0') {
+            dst[i] = '\0';
+            return true;
+        }
+
+        dst[i] = src[i];
+    }
+
+    __builtin_unreachable();
+}
+}  // namespace cstr_utils
+}  // namespace

nac3core/irrt/irrt_everything.hpp

@@ -0,0 +1,24 @@
+#pragma once
+
+#ifdef IRRT_DEBUG
+#define IRRT_DEBUG_ASSERT
+#define IRRT_DEBUG_ASSERT_BOOL true
+#else
+#define IRRT_DEBUG_ASSERT_BOOL false
+#endif
+
+#include <irrt/core.hpp>
+#include <irrt/debug.hpp>
+#include <irrt/exception.hpp>
+#include <irrt/int_defs.hpp>
+#include <irrt/list.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/product.hpp>
+#include <irrt/ndarray/reshape.hpp>
+#include <irrt/ndarray/transpose.hpp>
+#include <irrt/util.hpp>

nac3core/irrt/irrt_test.cpp

@@ -0,0 +1,25 @@
+// This file will be compiled like a real C++ program,
+// and we do have the luxury to use the standard libraries.
+// That is if the nix flakes do not have issues... especially on msys2...
+
+#include <cstdint>
+#include <cstdio>
+#include <cstdlib>
+
+// Special macro to inform `#include <irrt/*>` that we are testing.
+#define IRRT_TESTING
+
+// Note that failure unit tests are not supported.
+
+#include <test/test_core.hpp>
+#include <test/test_ndarray_basic.hpp>
+#include <test/test_ndarray_broadcast.hpp>
+#include <test/test_ndarray_indexing.hpp>
+
+int main() {
+    test::core::run();
+    test::ndarray_basic::run();
+    test::ndarray_indexing::run();
+    test::ndarray_broadcast::run();
+    return 0;
+}

nac3core/irrt/test/includes.hpp

@@ -0,0 +1,11 @@
+#pragma once
+
+#include <cstdint>
+#include <cstdio>
+#include <cstdlib>
+#include <irrt_everything.hpp>
+#include <test/util.hpp>
+
+/*
+    Include this header for every test_*.cpp
+*/

nac3core/irrt/test/test_core.hpp

@@ -0,0 +1,16 @@
+#pragma once
+
+#include <test/includes.hpp>
+
+namespace test {
+namespace core {
+void test_int_exp() {
+    BEGIN_TEST();
+
+    assert_values_match(125L, __nac3_int_exp_impl<int64_t>(5, 3));
+    assert_values_match(3125L, __nac3_int_exp_impl<int64_t>(5, 5));
+}
+
+void run() { test_int_exp(); }
+}  // namespace core
+}  // namespace test

nac3core/irrt/test/test_ndarray_basic.hpp

@@ -0,0 +1,30 @@
+#pragma once
+
+#include <test/includes.hpp>
+
+namespace test {
+namespace ndarray_basic {
+void test_calc_size_from_shape_normal() {
+    // Test shapes with normal values
+    BEGIN_TEST();
+
+    int64_t shape[4] = {2, 3, 5, 7};
+    assert_values_match(
+        210L, ndarray::basic::util::calc_size_from_shape<int64_t>(4, shape));
+}
+
+void test_calc_size_from_shape_has_zero() {
+    // Test shapes with 0 in them
+    BEGIN_TEST();
+
+    int64_t shape[4] = {2, 0, 5, 7};
+    assert_values_match(
+        0L, ndarray::basic::util::calc_size_from_shape<int64_t>(4, shape));
+}
+
+void run() {
+    test_calc_size_from_shape_normal();
+    test_calc_size_from_shape_has_zero();
+}
+}  // namespace ndarray_basic
+}  // namespace test

nac3core/irrt/test/test_ndarray_broadcast.hpp

@@ -0,0 +1,127 @@
+#pragma once
+
+#include <test/includes.hpp>
+
+namespace test {
+namespace ndarray_broadcast {
+void test_can_broadcast_shape() {
+    BEGIN_TEST();
+
+    assert_values_match(true,
+                        ndarray::broadcast::util::can_broadcast_shape_to(
+                            1, (int32_t[]){3}, 5, (int32_t[]){1, 1, 1, 1, 3}));
+    assert_values_match(false, ndarray::broadcast::util::can_broadcast_shape_to(
+                                   1, (int32_t[]){3}, 2, (int32_t[]){3, 1}));
+    assert_values_match(true, ndarray::broadcast::util::can_broadcast_shape_to(
+                                  1, (int32_t[]){3}, 1, (int32_t[]){3}));
+    assert_values_match(false, ndarray::broadcast::util::can_broadcast_shape_to(
+                                   1, (int32_t[]){1}, 1, (int32_t[]){3}));
+    assert_values_match(true, ndarray::broadcast::util::can_broadcast_shape_to(
+                                  1, (int32_t[]){1}, 1, (int32_t[]){1}));
+    assert_values_match(
+        true, ndarray::broadcast::util::can_broadcast_shape_to(
+                  3, (int32_t[]){256, 256, 3}, 3, (int32_t[]){256, 1, 3}));
+    assert_values_match(true,
+                        ndarray::broadcast::util::can_broadcast_shape_to(
+                            3, (int32_t[]){256, 256, 3}, 1, (int32_t[]){3}));
+    assert_values_match(false,
+                        ndarray::broadcast::util::can_broadcast_shape_to(
+                            3, (int32_t[]){256, 256, 3}, 1, (int32_t[]){2}));
+    assert_values_match(true,
+                        ndarray::broadcast::util::can_broadcast_shape_to(
+                            3, (int32_t[]){256, 256, 3}, 1, (int32_t[]){1}));
+
+    // In cases when the shapes contain zero(es)
+    assert_values_match(true, ndarray::broadcast::util::can_broadcast_shape_to(
+                                  1, (int32_t[]){0}, 1, (int32_t[]){1}));
+    assert_values_match(false, ndarray::broadcast::util::can_broadcast_shape_to(
+                                   1, (int32_t[]){0}, 1, (int32_t[]){2}));
+    assert_values_match(true,
+                        ndarray::broadcast::util::can_broadcast_shape_to(
+                            4, (int32_t[]){0, 4, 0, 0}, 1, (int32_t[]){1}));
+    assert_values_match(
+        true, ndarray::broadcast::util::can_broadcast_shape_to(
+                  4, (int32_t[]){0, 4, 0, 0}, 4, (int32_t[]){1, 1, 1, 1}));
+    assert_values_match(
+        true, ndarray::broadcast::util::can_broadcast_shape_to(
+                  4, (int32_t[]){0, 4, 0, 0}, 4, (int32_t[]){1, 4, 1, 1}));
+    assert_values_match(false, ndarray::broadcast::util::can_broadcast_shape_to(
+                                   2, (int32_t[]){4, 3}, 2, (int32_t[]){0, 3}));
+    assert_values_match(false, ndarray::broadcast::util::can_broadcast_shape_to(
+                                   2, (int32_t[]){4, 3}, 2, (int32_t[]){0, 0}));
+}
+
+void test_ndarray_broadcast() {
+    /*
+    # array = np.array([[19.9, 29.9, 39.9, 49.9]], dtype=np.float64)
+    # >>> [[19.9 29.9 39.9 49.9]]
+    #
+    # array = np.broadcast_to(array, (2, 3, 4))
+    # >>> [[[19.9 29.9 39.9 49.9]
+    # >>>   [19.9 29.9 39.9 49.9]
+    # >>>   [19.9 29.9 39.9 49.9]]
+    # >>>  [[19.9 29.9 39.9 49.9]
+    # >>>   [19.9 29.9 39.9 49.9]
+    # >>>   [19.9 29.9 39.9 49.9]]]
+    #
+    # assery array.strides == (0, 0, 8)
+
+    */
+    BEGIN_TEST();
+
+    double in_data[4] = {19.9, 29.9, 39.9, 49.9};
+    const int32_t in_ndims = 2;
+    int32_t in_shape[in_ndims] = {1, 4};
+    int32_t in_strides[in_ndims] = {};
+    NDArray<int32_t> ndarray = {.data = (uint8_t*)in_data,
+                                .itemsize = sizeof(double),
+                                .ndims = in_ndims,
+                                .shape = in_shape,
+                                .strides = in_strides};
+    ndarray::basic::set_strides_by_shape(&ndarray);
+
+    const int32_t dst_ndims = 3;
+    int32_t dst_shape[dst_ndims] = {2, 3, 4};
+    int32_t dst_strides[dst_ndims] = {};
+    NDArray<int32_t> dst_ndarray = {
+        .ndims = dst_ndims, .shape = dst_shape, .strides = dst_strides};
+
+    ndarray::broadcast::broadcast_to(&ndarray, &dst_ndarray);
+
+    assert_arrays_match(dst_ndims, ((int32_t[]){0, 0, 8}), dst_ndarray.strides);
+
+    assert_values_match(19.9,
+                        *((double*)ndarray::basic::get_pelement_by_indices(
+                            &dst_ndarray, ((int32_t[]){0, 0, 0}))));
+    assert_values_match(29.9,
+                        *((double*)ndarray::basic::get_pelement_by_indices(
+                            &dst_ndarray, ((int32_t[]){0, 0, 1}))));
+    assert_values_match(39.9,
+                        *((double*)ndarray::basic::get_pelement_by_indices(
+                            &dst_ndarray, ((int32_t[]){0, 0, 2}))));
+    assert_values_match(49.9,
+                        *((double*)ndarray::basic::get_pelement_by_indices(
+                            &dst_ndarray, ((int32_t[]){0, 0, 3}))));
+    assert_values_match(19.9,
+                        *((double*)ndarray::basic::get_pelement_by_indices(
+                            &dst_ndarray, ((int32_t[]){0, 1, 0}))));
+    assert_values_match(29.9,
+                        *((double*)ndarray::basic::get_pelement_by_indices(
+                            &dst_ndarray, ((int32_t[]){0, 1, 1}))));
+    assert_values_match(39.9,
+                        *((double*)ndarray::basic::get_pelement_by_indices(
+                            &dst_ndarray, ((int32_t[]){0, 1, 2}))));
+    assert_values_match(49.9,
+                        *((double*)ndarray::basic::get_pelement_by_indices(
+                            &dst_ndarray, ((int32_t[]){0, 1, 3}))));
+    assert_values_match(49.9,
+                        *((double*)ndarray::basic::get_pelement_by_indices(
+                            &dst_ndarray, ((int32_t[]){1, 2, 3}))));
+}
+
+void run() {
+    test_can_broadcast_shape();
+    test_ndarray_broadcast();
+}
+}  // namespace ndarray_broadcast
+}  // namespace test

nac3core/irrt/test/test_ndarray_indexing.hpp

@@ -0,0 +1,165 @@
+#pragma once
+
+#include <test/includes.hpp>
+
+namespace test {
+namespace ndarray_indexing {
+void test_normal_1() {
+    /*
+        Reference Python code:
+        ```python
+        ndarray = np.arange(12, dtype=np.float64).reshape((3, 4));
+        # array([[ 0.,  1.,  2.,  3.],
+        #        [ 4.,  5.,  6.,  7.],
+        #        [ 8.,  9., 10., 11.]])
+
+        dst_ndarray = ndarray[-2:, 1::2]
+        # array([[ 5.,  7.],
+        #        [ 9., 11.]])
+
+        assert dst_ndarray.shape == (2, 2)
+        assert dst_ndarray.strides == (32, 16)
+        assert dst_ndarray[0, 0] == 5.0
+        assert dst_ndarray[0, 1] == 7.0
+        assert dst_ndarray[1, 0] == 9.0
+        assert dst_ndarray[1, 1] == 11.0
+        ```
+    */
+    BEGIN_TEST();
+
+    // Prepare src_ndarray
+    double src_data[12] = {0.0, 1.0, 2.0, 3.0, 4.0,  5.0,
+                           6.0, 7.0, 8.0, 9.0, 10.0, 11.0};
+    int64_t src_itemsize = sizeof(double);
+    const int64_t src_ndims = 2;
+    int64_t src_shape[src_ndims] = {3, 4};
+    int64_t src_strides[src_ndims] = {};
+    NDArray<int64_t> src_ndarray = {.data = (uint8_t *)src_data,
+                                    .itemsize = src_itemsize,
+                                    .ndims = src_ndims,
+                                    .shape = src_shape,
+                                    .strides = src_strides};
+    ndarray::basic::set_strides_by_shape(&src_ndarray);
+
+    // Prepare dst_ndarray
+    const int64_t dst_ndims = 2;
+    int64_t dst_shape[dst_ndims] = {999, 999};    // Empty values
+    int64_t dst_strides[dst_ndims] = {999, 999};  // Empty values
+    NDArray<int64_t> dst_ndarray = {.data = nullptr,
+                                    .ndims = dst_ndims,
+                                    .shape = dst_shape,
+                                    .strides = dst_strides};
+
+    // Create the subscripts in `ndarray[-2::, 1::2]`
+    UserSlice subscript_1;
+    subscript_1.set_start(-2);
+
+    UserSlice subscript_2;
+    subscript_2.set_start(1);
+    subscript_2.set_step(2);
+
+    const int64_t num_indexes = 2;
+    NDIndex indexes[num_indexes] = {
+        {.type = ND_INDEX_TYPE_SLICE, .data = (uint8_t *)&subscript_1},
+        {.type = ND_INDEX_TYPE_SLICE, .data = (uint8_t *)&subscript_2}};
+
+    ndarray::indexing::index(num_indexes, indexes, &src_ndarray, &dst_ndarray);
+
+    int64_t expected_shape[dst_ndims] = {2, 2};
+    int64_t expected_strides[dst_ndims] = {32, 16};
+
+    assert_arrays_match(dst_ndims, expected_shape, dst_ndarray.shape);
+    assert_arrays_match(dst_ndims, expected_strides, dst_ndarray.strides);
+
+    // dst_ndarray[0, 0]
+    assert_values_match(5.0,
+                        *((double *)ndarray::basic::get_pelement_by_indices(
+                            &dst_ndarray, (int64_t[dst_ndims]){0, 0})));
+    // dst_ndarray[0, 1]
+    assert_values_match(7.0,
+                        *((double *)ndarray::basic::get_pelement_by_indices(
+                            &dst_ndarray, (int64_t[dst_ndims]){0, 1})));
+    // dst_ndarray[1, 0]
+    assert_values_match(9.0,
+                        *((double *)ndarray::basic::get_pelement_by_indices(
+                            &dst_ndarray, (int64_t[dst_ndims]){1, 0})));
+    // dst_ndarray[1, 1]
+    assert_values_match(11.0,
+                        *((double *)ndarray::basic::get_pelement_by_indices(
+                            &dst_ndarray, (int64_t[dst_ndims]){1, 1})));
+}
+
+void test_normal_2() {
+    /*
+        ```python
+        ndarray = np.arange(12, dtype=np.float64).reshape((3, 4))
+        # array([[ 0.,  1.,  2.,  3.],
+        #        [ 4.,  5.,  6.,  7.],
+        #        [ 8.,  9., 10., 11.]])
+
+        dst_ndarray = ndarray[2, ::-2]
+        # array([11.,  9.])
+
+        assert dst_ndarray.shape == (2,)
+        assert dst_ndarray.strides == (-16,)
+        assert dst_ndarray[0] == 11.0
+        assert dst_ndarray[1] == 9.0
+        ```
+    */
+    BEGIN_TEST();
+
+    // Prepare src_ndarray
+    double src_data[12] = {0.0, 1.0, 2.0, 3.0, 4.0,  5.0,
+                           6.0, 7.0, 8.0, 9.0, 10.0, 11.0};
+    int64_t src_itemsize = sizeof(double);
+    const int64_t src_ndims = 2;
+    int64_t src_shape[src_ndims] = {3, 4};
+    int64_t src_strides[src_ndims] = {};
+    NDArray<int64_t> src_ndarray = {.data = (uint8_t *)src_data,
+                                    .itemsize = src_itemsize,
+                                    .ndims = src_ndims,
+                                    .shape = src_shape,
+                                    .strides = src_strides};
+    ndarray::basic::set_strides_by_shape(&src_ndarray);
+
+    // Prepare dst_ndarray
+    const int64_t dst_ndims = 1;
+    int64_t dst_shape[dst_ndims] = {999};    // Empty values
+    int64_t dst_strides[dst_ndims] = {999};  // Empty values
+    NDArray<int64_t> dst_ndarray = {.data = nullptr,
+                                    .ndims = dst_ndims,
+                                    .shape = dst_shape,
+                                    .strides = dst_strides};
+
+    // Create the subscripts in `ndarray[2, ::-2]`
+    int64_t subscript_1 = 2;
+
+    UserSlice subscript_2;
+    subscript_2.set_step(-2);
+
+    const int64_t num_indexes = 2;
+    NDIndex indexes[num_indexes] = {
+        {.type = ND_INDEX_TYPE_SINGLE_ELEMENT, .data = (uint8_t *)&subscript_1},
+        {.type = ND_INDEX_TYPE_SLICE, .data = (uint8_t *)&subscript_2}};
+
+    ndarray::indexing::index(num_indexes, indexes, &src_ndarray, &dst_ndarray);
+
+    int64_t expected_shape[dst_ndims] = {2};
+    int64_t expected_strides[dst_ndims] = {-16};
+    assert_arrays_match(dst_ndims, expected_shape, dst_ndarray.shape);
+    assert_arrays_match(dst_ndims, expected_strides, dst_ndarray.strides);
+
+    assert_values_match(11.0,
+                        *((double *)ndarray::basic::get_pelement_by_indices(
+                            &dst_ndarray, (int64_t[dst_ndims]){0})));
+    assert_values_match(9.0,
+                        *((double *)ndarray::basic::get_pelement_by_indices(
+                            &dst_ndarray, (int64_t[dst_ndims]){1})));
+}
+
+void run() {
+    test_normal_1();
+    test_normal_2();
+}
+}  // namespace ndarray_indexing
+}  // namespace test

nac3core/irrt/test/util.hpp

@@ -0,0 +1,131 @@
+#pragma once
+
+#include <cstdio>
+#include <cstdlib>
+
+template <class T>
+void print_value(const T& value);
+
+template <>
+void print_value(const bool& value) {
+    printf("%s", value ? "true" : "false");
+}
+
+template <>
+void print_value(const int8_t& value) {
+    printf("%d", value);
+}
+
+template <>
+void print_value(const int32_t& value) {
+    printf("%d", value);
+}
+
+template <>
+void print_value(const int64_t& value) {
+    printf("%d", value);
+}
+
+template <>
+void print_value(const uint8_t& value) {
+    printf("%u", value);
+}
+
+template <>
+void print_value(const uint32_t& value) {
+    printf("%u", value);
+}
+
+template <>
+void print_value(const uint64_t& value) {
+    printf("%d", value);
+}
+
+template <>
+void print_value(const float& value) {
+    printf("%f", value);
+}
+
+template <>
+void print_value(const double& value) {
+    printf("%f", value);
+}
+
+void __begin_test(const char* function_name, const char* file, int line) {
+    printf("######### Running %s @ %s:%d\n", function_name, file, line);
+}
+
+#define BEGIN_TEST() __begin_test(__FUNCTION__, __FILE__, __LINE__)
+
+void test_fail() {
+    printf("[!] Test failed. Exiting with status code 1.\n");
+    exit(1);
+}
+
+template <typename T>
+void debug_print_array(int len, const T* as) {
+    printf("[");
+    for (int i = 0; i < len; i++) {
+        if (i != 0) printf(", ");
+        print_value(as[i]);
+    }
+    printf("]");
+}
+
+void print_assertion_passed(const char* file, int line) {
+    printf("[*] Assertion passed on %s:%d\n", file, line);
+}
+
+void print_assertion_failed(const char* file, int line) {
+    printf("[!] Assertion failed on %s:%d\n", file, line);
+}
+
+void __assert_true(const char* file, int line, bool cond) {
+    if (cond) {
+        print_assertion_passed(file, line);
+    } else {
+        print_assertion_failed(file, line);
+        test_fail();
+    }
+}
+
+#define assert_true(cond) __assert_true(__FILE__, __LINE__, cond)
+
+template <typename T>
+void __assert_arrays_match(const char* file, int line, int len,
+                           const T* expected, const T* got) {
+    if (arrays_match(len, expected, got)) {
+        print_assertion_passed(file, line);
+    } else {
+        print_assertion_failed(file, line);
+        printf("Expect = ");
+        debug_print_array(len, expected);
+        printf("\n");
+        printf("   Got = ");
+        debug_print_array(len, got);
+        printf("\n");
+        test_fail();
+    }
+}
+
+#define assert_arrays_match(len, expected, got) \
+    __assert_arrays_match(__FILE__, __LINE__, len, expected, got)
+
+template <typename T>
+void __assert_values_match(const char* file, int line, T expected, T got) {
+    if (expected == got) {
+        print_assertion_passed(file, line);
+    } else {
+        print_assertion_failed(file, line);
+        printf("Expect = ");
+        print_value(expected);
+        printf("\n");
+        printf("   Got = ");
+        print_value(got);
+        printf("\n");
+        test_fail();
+    }
+}
+
+#define assert_values_match(expected, got) \
+    __assert_values_match(__FILE__, __LINE__, expected, got)

nac3core/src/codegen/expr.rs

@@ -2,7 +2,7 @@ use crate::{
     codegen::{
         classes::{
             ArrayLikeIndexer, ArrayLikeValue, ListType, ListValue, NDArrayValue, ProxyType,
-            ProxyValue, RangeValue, TypedArrayLikeAccessor, UntypedArrayLikeAccessor,
+            ProxyValue, RangeValue, UntypedArrayLikeAccessor,
         },
         concrete_type::{ConcreteFuncArg, ConcreteTypeEnum, ConcreteTypeStore},
         gen_in_range_check, get_llvm_abi_type, get_llvm_type, get_va_count_arg_name,
@@ -12,6 +12,10 @@ use crate::{
             call_memcpy_generic,
         },
         need_sret, numpy,
+        object::{
+            ndarray::{scalar::split_scalar_or_ndarray, NDArrayObject, NDArrayOut},
+            AnyObject,
+        },
         stmt::{
             gen_for_callback_incrementing, gen_if_callback, gen_if_else_expr_callback, gen_raise,
             gen_var,
@@ -19,11 +23,7 @@ use crate::{
         CodeGenContext, CodeGenTask, CodeGenerator,
     },
     symbol_resolver::{SymbolValue, ValueEnum},
-    toplevel::{
+    toplevel::{helper::PrimDef, numpy::unpack_ndarray_var_tys, DefinitionId, TopLevelDef},
-        helper::PrimDef,
-        numpy::{make_ndarray_ty, unpack_ndarray_var_tys},
-        DefinitionId, TopLevelDef,
-    },
     typecheck::{
         magic_methods::{Binop, BinopVariant, HasOpInfo},
         typedef::{FunSignature, FuncArg, Type, TypeEnum, TypeVarId, Unifier, VarMap},
@@ -32,7 +32,7 @@ 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},
     AddressSpace, IntPredicate, OptimizationLevel,
 };
 use itertools::{chain, izip, Either, Itertools};
@@ -43,6 +43,12 @@ use nac3parser::ast::{
 use std::iter::{repeat, repeat_with};
 use std::{collections::HashMap, convert::TryInto, iter::once, iter::zip};
 
+use super::{
+    model::*,
+    object::ndarray::indexing::util::gen_ndarray_subscript_ndindexes,
+    structure::{CSlice, Exception, ExceptionId},
+};
+
 pub fn get_subst_key(
     unifier: &mut Unifier,
     obj: Option<Type>,
@@ -298,31 +304,14 @@ impl<'ctx, 'a> CodeGenContext<'ctx, 'a> {
                     None
                 }
             }
-            Constant::Str(v) => {
+            Constant::Str(s) => Some(self.gen_string(generator, s).value.into()),
-                assert!(self.unifier.unioned(ty, self.primitives.str));
-                if let Some(v) = self.const_strings.get(v) {
-                    Some(*v)
-                } else {
-                    let str_ptr = self
-                        .builder
-                        .build_global_string_ptr(v, "const")
-                        .map(|v| v.as_pointer_value().into())
-                        .unwrap();
-                    let size = generator.get_size_type(self.ctx).const_int(v.len() as u64, false);
-                    let ty = self.get_llvm_type(generator, self.primitives.str);
-                    let val =
-                        ty.into_struct_type().const_named_struct(&[str_ptr, size.into()]).into();
-                    self.const_strings.insert(v.to_string(), val);
-                    Some(val)
-                }
-            }
             Constant::Ellipsis => {
                 let msg = self.gen_string(generator, "NotImplementedError");
 
                 self.raise_exn(
                     generator,
                     "0:NotImplementedError",
-                    msg.into(),
+                    msg,
                     [None, None, None],
                     self.current_loc,
                 );
@@ -582,98 +571,107 @@ impl<'ctx, 'a> CodeGenContext<'ctx, 'a> {
     }
 
     /// Helper function for generating a LLVM variable storing a [String].
-    pub fn gen_string<G, S>(&mut self, generator: &mut G, s: S) -> StructValue<'ctx>
+    pub fn gen_string<G>(&mut self, generator: &mut G, string: &str) -> Struct<'ctx, CSlice>
     where
         G: CodeGenerator + ?Sized,
-        S: Into<String>,
     {
-        self.gen_const(generator, &Constant::Str(s.into()), self.primitives.str)
+        self.const_strings.get(string).copied().unwrap_or_else(|| {
-            .map(BasicValueEnum::into_struct_value)
+            let sizet_model = IntModel(SizeT);
-            .unwrap()
+            let pbyte_model = PtrModel(IntModel(Byte));
+            let cslice_model = StructModel(CSlice);
+
+            let base = self.builder.build_global_string_ptr(string, "constant_string").unwrap();
+            let base = pbyte_model.believe_value(base.as_pointer_value());
+
+            let len = sizet_model.constant(generator, self.ctx, string.len() as u64);
+
+            let cslice = cslice_model.create_const(generator, self.ctx, base, len);
+
+            self.const_strings.insert(string.to_owned(), cslice);
+
+            cslice
+        })
     }
 
     pub fn raise_exn<G: CodeGenerator + ?Sized>(
         &mut self,
         generator: &mut G,
         name: &str,
-        msg: BasicValueEnum<'ctx>,
+        msg: Struct<'ctx, CSlice>,
-        params: [Option<IntValue<'ctx>>; 3],
+        params: [Option<Int<'ctx, Int64>>; 3],
         loc: Location,
     ) {
-        let zelf = if let Some(exception_val) = self.exception_val {
+        let exn_model = StructModel(Exception);
-            exception_val
+        let exn_id_model = IntModel(ExceptionId::default());
-        } else {
+
-            let ty = self.get_llvm_type(generator, self.primitives.exception).into_pointer_type();
+        let exn_id =
-            let zelf_ty: BasicTypeEnum = ty.get_element_type().into_struct_type().into();
+            exn_id_model.constant(generator, self.ctx, self.resolver.get_string_id(name) as u64);
-            let zelf = generator.gen_var_alloc(self, zelf_ty, Some("exn")).unwrap();
+        let exn = self.exception_val.unwrap_or_else(|| {
-            *self.exception_val.insert(zelf)
+            let exn = exn_model.var_alloca(generator, self, Some("exn")).unwrap();
-        };
+            *self.exception_val.insert(exn)
-        let int32 = self.ctx.i32_type();
-        let zero = int32.const_zero();
-        unsafe {
-            let id_ptr = self.builder.build_in_bounds_gep(zelf, &[zero, zero], "exn.id").unwrap();
-            let id = self.resolver.get_string_id(name);
-            self.builder.build_store(id_ptr, int32.const_int(id as u64, false)).unwrap();
-            let ptr = self
-                .builder
-                .build_in_bounds_gep(zelf, &[zero, int32.const_int(5, false)], "exn.msg")
-                .unwrap();
-            self.builder.build_store(ptr, msg).unwrap();
-            let i64_zero = self.ctx.i64_type().const_zero();
-            for (i, attr_ind) in [6, 7, 8].iter().enumerate() {
-                let ptr = self
-                    .builder
-                    .build_in_bounds_gep(
-                        zelf,
-                        &[zero, int32.const_int(*attr_ind, false)],
-                        "exn.param",
-                    )
-                    .unwrap();
-                let val = params[i].map_or(i64_zero, |v| {
-                    self.builder.build_int_s_extend(v, self.ctx.i64_type(), "sext").unwrap()
         });
-                self.builder.build_store(ptr, val).unwrap();
+
+        exn.set(self, |f| f.id, exn_id);
+        exn.set(self, |f| f.msg, msg);
+        for (i, param) in params.iter().enumerate() {
+            if let Some(param) = param {
+                exn.set(self, |f| f.params[i], *param);
             }
         }
-        gen_raise(generator, self, Some(&zelf.into()), loc);
+
+        gen_raise(generator, self, Some(exn), loc);
     }
 
     pub fn make_assert<G: CodeGenerator + ?Sized>(
         &mut self,
         generator: &mut G,
-        cond: IntValue<'ctx>,
+        cond: IntValue<'ctx>, // IntType can have arbitrary bit width
         err_name: &str,
         err_msg: &str,
         params: [Option<IntValue<'ctx>>; 3],
         loc: Location,
     ) {
+        let param_model = IntModel(Int64);
+        let params = params
+            .map(|p| p.map(|p| param_model.s_extend_or_bit_cast(generator, self, p, "param")));
+
         let err_msg = self.gen_string(generator, err_msg);
-        self.make_assert_impl(generator, cond, err_name, err_msg.into(), params, loc);
+        self.make_assert_impl(generator, cond, err_name, err_msg, params, loc);
     }
 
     pub fn make_assert_impl<G: CodeGenerator + ?Sized>(
         &mut self,
         generator: &mut G,
-        cond: IntValue<'ctx>,
+        cond: IntValue<'ctx>, // IntType can have arbitrary bit width
         err_name: &str,
-        err_msg: BasicValueEnum<'ctx>,
+        err_msg: Struct<'ctx, CSlice>,
-        params: [Option<IntValue<'ctx>>; 3],
+        params: [Option<Int<'ctx, Int64>>; 3],
         loc: Location,
     ) {
-        let i1 = self.ctx.bool_type();
+        let bool_model = IntModel(Bool);
-        let i1_true = i1.const_all_ones();
+
-        // we assume that the condition is most probably true, so the normal path is the most
+        // We assume that the condition is most probably true, so the normal path is the most
-        // probable path
+        // probable path even if this assumption is violated, it does not matter as exception unwinding is.
-        // even if this assumption is violated, it does not matter as exception unwinding is
+        let cond = call_expect(
-        // slow anyway...
+            self,
-        let cond = call_expect(self, cond, i1_true, Some("expect"));
+            generator.bool_to_i1(self, cond),
+            bool_model.const_true(generator, self.ctx).value,
+            Some("expect"),
+        );
+
         let current_bb = self.builder.get_insert_block().unwrap();
         let current_fun = current_bb.get_parent().unwrap();
+
         let then_block = self.ctx.insert_basic_block_after(current_bb, "succ");
         let exn_block = self.ctx.append_basic_block(current_fun, "fail");
+
         self.builder.build_conditional_branch(cond, then_block, exn_block).unwrap();
+
+        // Inserting into `exn_block`
         self.builder.position_at_end(exn_block);
         self.raise_exn(generator, err_name, err_msg, params, loc);
+
+        // Continuation
         self.builder.position_at_end(then_block);
     }
 }
@@ -1539,99 +1537,52 @@ 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 = AnyObject { ty: ty1, value: left_val };
+        let right = AnyObject { ty: ty1, value: right_val };
 
-        let is_ndarray1 = ty1.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id());
+        let left = split_scalar_or_ndarray(generator, ctx, left).as_ndarray(generator, ctx);
-        let is_ndarray2 = ty2.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id());
+        let right = split_scalar_or_ndarray(generator, ctx, right).as_ndarray(generator, ctx);
 
-        if is_ndarray1 && is_ndarray2 {
+        debug_assert!(ctx.unifier.unioned(left.dtype, right.dtype)); // Typechecker ensures this.
-            let (ndarray_dtype1, _) = unpack_ndarray_var_tys(&mut ctx.unifier, ty1);
-            let (ndarray_dtype2, _) = unpack_ndarray_var_tys(&mut ctx.unifier, ty2);
 
-            assert!(ctx.unifier.unioned(ndarray_dtype1, ndarray_dtype2));
+        let common_dtype = left.dtype;
 
-            let left_val =
+        let out = match op.variant {
-                NDArrayValue::from_ptr_val(left_val.into_pointer_value(), llvm_usize, None);
+            BinopVariant::Normal => NDArrayOut::NewNDArray { dtype: common_dtype },
-            let right_val =
+            BinopVariant::AugAssign => NDArrayOut::WriteToNDArray { ndarray: left },
-                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 `left @ right`
+            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, _) =
+            // For other operators like +, -, etc...; do them element-wise-ly
-                unpack_ndarray_var_tys(&mut ctx.unifier, if is_ndarray1 { ty1 } else { ty2 });
+            let result = NDArrayObject::broadcasting_starmap(
-            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(
                 generator,
                 ctx,
-                ndarray_dtype,
+                &[left, right],
-                match op.variant {
+                out,
-                    BinopVariant::Normal => None,
+                |generator, ctx, scalars| {
-                    BinopVariant::AugAssign => Some(ndarray_val),
+                    let left_scalar = scalars[0];
-                },
+                    let right_scalar = scalars[1];
-                (left_val, !is_ndarray1),
+
-                (right_val, !is_ndarray2),
+                    let result = gen_binop_expr_with_values(
-                |generator, ctx, (lhs, rhs)| {
-                    gen_binop_expr_with_values(
                         generator,
                         ctx,
-                        (&Some(ndarray_dtype), lhs),
+                        (&Some(left_scalar.ty), left_scalar.value),
                         op,
-                        (&Some(ndarray_dtype), rhs),
+                        (&Some(right_scalar.ty), right_scalar.value),
                         ctx.current_loc,
                     )?
                     .unwrap()
-                    .to_basic_value_enum(ctx, generator, ndarray_dtype)
+                    .to_basic_value_enum(ctx, generator, common_dtype)?;
+
+                    Ok(AnyObject { value: result, ty: common_dtype })
                 },
             )?;
-
+            Ok(Some(ValueEnum::Dynamic(result.instance.value.as_basic_value_enum())))
-            Ok(Some(res.as_base_value().into()))
         }
     } else {
         let left_ty_enum = ctx.unifier.get_ty_immutable(left_ty.unwrap());
@@ -1722,6 +1673,7 @@ pub fn gen_binop_expr<'ctx, G: CodeGenerator>(
 
 /// Generates LLVM IR for a unary operator expression using the [`Type`] and
 /// [LLVM value][`BasicValueEnum`] of the operands.
+#[allow(clippy::only_used_in_recursion)]
 pub fn gen_unaryop_expr_with_values<'ctx, G: CodeGenerator>(
     generator: &mut G,
     ctx: &mut CodeGenContext<'ctx, '_>,
@@ -1789,40 +1741,42 @@ 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);
+        todo!()
-        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 llvm_usize = generator.get_size_type(ctx.ctx);
+        // let (ndarray_dtype, _) = unpack_ndarray_var_tys(&mut ctx.unifier, ty);
 
-        // ndarray uses `~` rather than `not` to perform elementwise inversion, convert it before
+        // let val = NDArrayValue::from_ptr_val(val.into_pointer_value(), llvm_usize, None);
-        // passing it to the elementwise codegen function
-        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 {
-                unreachable!(
-                    "ufunc {} not supported for ndarray[bool, N]",
-                    op.op_info().method_name,
-                )
-            }
-        } else {
-            op
-        };
 
-        let res = numpy::ndarray_elementwise_unaryop_impl(
+        // // ndarray uses `~` rather than `not` to perform elementwise inversion, convert it before
-            generator,
+        // // passing it to the elementwise codegen function
-            ctx,
+        // let op = if ndarray_dtype.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::Bool.id()) {
-            ndarray_dtype,
+        //     if op == ast::Unaryop::Invert {
-            None,
+        //         ast::Unaryop::Not
-            val,
+        //     } else {
-            |generator, ctx, val| {
+        //         unreachable!(
-                gen_unaryop_expr_with_values(generator, ctx, op, (&Some(ndarray_dtype), val))?
+        //             "ufunc {} not supported for ndarray[bool, N]",
-                    .unwrap()
+        //             op.op_info().method_name,
-                    .to_basic_value_enum(ctx, generator, ndarray_dtype)
+        //         )
-            },
+        //     }
-        )?;
+        // } else {
+        //     op
+        // };
 
-        res.as_base_value().into()
+        // let res = numpy::ndarray_elementwise_unaryop_impl(
+        //     generator,
+        //     ctx,
+        //     ndarray_dtype,
+        //     None,
+        //     val,
+        //     |generator, ctx, val| {
+        //         gen_unaryop_expr_with_values(generator, ctx, op, (&Some(ndarray_dtype), val))?
+        //             .unwrap()
+        //             .to_basic_value_enum(ctx, generator, ndarray_dtype)
+        //     },
+        // )?;
+
+        // res.as_base_value().into()
     } else {
         unimplemented!()
     }))
@@ -2266,338 +2220,6 @@ 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,
@@ -3069,7 +2691,7 @@ pub fn gen_expr<'ctx, G: CodeGenerator>(
                                         ctx.raise_exn(
                                             generator,
                                             "0:UnwrapNoneError",
-                                            err_msg.into(),
+                                            err_msg,
                                             [None, None, None],
                                             ctx.current_loc,
                                         );
@@ -3237,18 +2859,23 @@ 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() => {
+                TypeEnum::TObj { obj_id, .. } if *obj_id == PrimDef::NDArray.id() => {
-                    let (ty, ndims) = params.iter().map(|(_, ty)| ty).collect_tuple().unwrap();
+                    let Some(ndarray) = generator.gen_expr(ctx, value)? else {
-
-                    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 {
                         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 = AnyObject { ty: ndarray_ty, value: ndarray };
+
+                    let ndarray = NDArrayObject::from_object(generator, ctx, ndarray);
+
+                    let indexes = gen_ndarray_subscript_ndindexes(generator, ctx, slice)?;
+                    let result = ndarray
+                        .index(generator, ctx, &indexes, "index_result")
+                        .split_unsized(generator, ctx)
+                        .to_basic_value_enum();
+                    return Ok(Some(ValueEnum::Dynamic(result)));
                 }
                 TypeEnum::TTuple { .. } => {
                     let index: u32 =
@@ -3291,3 +2918,44 @@ 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, IntModel<Int32>>>,
+        Option<Instance<'ctx, IntModel<Int32>>>,
+        Option<Instance<'ctx, IntModel<Int32>>>,
+    ),
+    String,
+> {
+    let i32_model = IntModel(Int32); // TODO: Switch to usize
+
+    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 = i32_model.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

@@ -1,5 +1,12 @@
+use crate::symbol_resolver::SymbolResolver;
 use crate::typecheck::typedef::Type;
 
+mod test;
+
+use super::model::*;
+use super::object::ndarray::broadcast::ShapeEntry;
+use super::object::ndarray::indexing::{NDIndex, UserSlice};
+use super::structure::{List, NDArray, NDIter};
 use super::{
     classes::{
         ArrayLikeIndexer, ArrayLikeValue, ArraySliceValue, ListValue, NDArrayValue,
@@ -9,6 +16,8 @@ use super::{
 };
 use crate::codegen::classes::TypedArrayLikeAccessor;
 use crate::codegen::stmt::gen_for_callback_incrementing;
+use function::{get_sizet_dependent_function_name, CallFunction};
+use inkwell::values::BasicValue;
 use inkwell::{
     attributes::{Attribute, AttributeLoc},
     context::Context,
@@ -414,14 +423,27 @@ pub fn list_slice_assignment<'ctx, G: CodeGenerator + ?Sized>(
         .unwrap();
     let cond_1 = ctx.builder.build_and(dest_step_eq_one, src_slt_dest, "slice_cond_1").unwrap();
     let cond = ctx.builder.build_or(src_eq_dest, cond_1, "slice_cond").unwrap();
+
+    // TODO: Temporary fix. Rewrite `list_slice_assignment` later
+    // Exception params should have been i64
+    {
+        let param_model = IntModel(Int64);
+
+        let src_slice_len =
+            param_model.s_extend_or_bit_cast(generator, ctx, src_slice_len, "src_slice_len");
+        let dest_slice_len =
+            param_model.s_extend_or_bit_cast(generator, ctx, dest_slice_len, "dest_slice_len");
+        let dest_idx_2 = param_model.s_extend_or_bit_cast(generator, ctx, dest_idx.2, "dest_idx_2");
+
         ctx.make_assert(
             generator,
             cond,
             "0:ValueError",
             "attempt to assign sequence of size {0} to slice of size {1} with step size {2}",
-        [Some(src_slice_len), Some(dest_slice_len), Some(dest_idx.2)],
+            [Some(src_slice_len.value), Some(dest_slice_len.value), Some(dest_idx_2.value)],
             ctx.current_loc,
         );
+    }
 
     let new_len = {
         let args = vec![
@@ -873,7 +895,7 @@ pub fn call_ndarray_calc_broadcast<'ctx, G: CodeGenerator + ?Sized>(
 }
 
 /// 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
+/// containing the indices used for accessing `array` corresponding to the index of the broadcast
 /// array `broadcast_idx`.
 pub fn call_ndarray_calc_broadcast_index<
     'ctx,
@@ -928,3 +950,337 @@ pub fn call_ndarray_calc_broadcast_index<
         Box::new(|_, v| v.into()),
     )
 }
+
+pub fn call_nac3_throw_dummy_error<G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &CodeGenContext<'_, '_>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_throw_dummy_error");
+    CallFunction::begin(generator, ctx, &name).returning_void();
+}
+
+/// Initialize all global `EXN_*` exception IDs in IRRT with the [`SymbolResolver`].
+pub fn setup_irrt_exceptions<'ctx>(
+    ctx: &'ctx Context,
+    module: &Module<'ctx>,
+    symbol_resolver: &dyn SymbolResolver,
+) {
+    let exn_id_type = ctx.i32_type();
+
+    let errors = &[
+        ("EXN_INDEX_ERROR", "0:IndexError"),
+        ("EXN_VALUE_ERROR", "0:ValueError"),
+        ("EXN_ASSERTION_ERROR", "0:AssertionError"),
+        ("EXN_RUNTIME_ERROR", "0:RuntimeError"),
+        ("EXN_TYPE_ERROR", "0:TypeError"),
+    ];
+
+    for (irrt_name, symbol_name) in errors {
+        let exn_id = symbol_resolver.get_string_id(symbol_name);
+        let exn_id = exn_id_type.const_int(exn_id as u64, false).as_basic_value_enum();
+
+        let global = module.get_global(irrt_name).unwrap_or_else(|| {
+            panic!("Exception symbol name '{irrt_name}' should exist in the IRRT LLVM module")
+        });
+        global.set_initializer(&exn_id);
+    }
+}
+
+pub fn call_nac3_list_slice_assign<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    dst: Ptr<'ctx, StructModel<List<IntModel<Byte>>>>,
+    src: Ptr<'ctx, StructModel<List<IntModel<Byte>>>>,
+    itemsize: Int<'ctx, SizeT>,
+    user_slice: Ptr<'ctx, StructModel<UserSlice>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_list_slice_assign");
+    CallFunction::begin(generator, ctx, &name)
+        .arg(dst)
+        .arg(src)
+        .arg(itemsize)
+        .arg(user_slice)
+        .returning_void();
+}
+
+pub fn call_nac3_ndarray_util_assert_shape_no_negative<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndims: Int<'ctx, SizeT>,
+    shape: Ptr<'ctx, IntModel<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: Int<'ctx, SizeT>,
+    ndarray_shape: Ptr<'ctx, IntModel<SizeT>>,
+    output_ndims: Int<'ctx, SizeT>,
+    output_shape: Ptr<'ctx, IntModel<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, '_>,
+    pndarray: Ptr<'ctx, StructModel<NDArray>>,
+) -> Int<'ctx, SizeT> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_size");
+    CallFunction::begin(generator, ctx, &name).arg(pndarray).returning_auto("size")
+}
+
+pub fn call_nac3_ndarray_nbytes<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    pndarray: Ptr<'ctx, StructModel<NDArray>>,
+) -> Int<'ctx, SizeT> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_nbytes");
+    CallFunction::begin(generator, ctx, &name).arg(pndarray).returning_auto("nbytes")
+}
+
+pub fn call_nac3_ndarray_len<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    pndarray: Ptr<'ctx, StructModel<NDArray>>,
+) -> Int<'ctx, SizeT> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_len");
+    CallFunction::begin(generator, ctx, &name).arg(pndarray).returning_auto("len")
+}
+
+pub fn call_nac3_ndarray_is_c_contiguous<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray_ptr: Ptr<'ctx, StructModel<NDArray>>,
+) -> Int<'ctx, Bool> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_is_c_contiguous");
+    CallFunction::begin(generator, ctx, &name).arg(ndarray_ptr).returning_auto("is_c_contiguous")
+}
+
+pub fn call_nac3_ndarray_get_nth_pelement<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    pndarray: Ptr<'ctx, StructModel<NDArray>>,
+    index: Int<'ctx, SizeT>,
+) -> Ptr<'ctx, IntModel<Byte>> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_get_nth_pelement");
+    CallFunction::begin(generator, ctx, &name).arg(pndarray).arg(index).returning_auto("pelement")
+}
+
+pub fn call_nac3_ndarray_set_strides_by_shape<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    pdnarray: Ptr<'ctx, StructModel<NDArray>>,
+) {
+    let name =
+        get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_set_strides_by_shape");
+    CallFunction::begin(generator, ctx, &name).arg(pdnarray).returning_void();
+}
+
+pub fn call_nac3_ndarray_copy_data<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    src_ndarray: Ptr<'ctx, StructModel<NDArray>>,
+    dst_ndarray: Ptr<'ctx, StructModel<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_ndarray_index<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    num_indexes: Int<'ctx, SizeT>,
+    indexes: Ptr<'ctx, StructModel<NDIndex>>,
+    src_ndarray: Ptr<'ctx, StructModel<NDArray>>,
+    dst_ndarray: Ptr<'ctx, StructModel<NDArray>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_index");
+    CallFunction::begin(generator, ctx, &name)
+        .arg(num_indexes)
+        .arg(indexes)
+        .arg(src_ndarray)
+        .arg(dst_ndarray)
+        .returning_void();
+}
+
+pub fn call_nac3_ndarray_broadcast_to<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    src_ndarray: Ptr<'ctx, StructModel<NDArray>>,
+    dst_ndarray: Ptr<'ctx, StructModel<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: Int<'ctx, SizeT>,
+    shape_entries: Ptr<'ctx, StructModel<ShapeEntry>>,
+    dst_ndims: Int<'ctx, SizeT>,
+    dst_shape: Ptr<'ctx, IntModel<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_resolve_and_check_new_shape<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    size: Int<'ctx, SizeT>,
+    new_ndims: Int<'ctx, SizeT>,
+    new_shape: Ptr<'ctx, IntModel<SizeT>>,
+) {
+    let name = get_sizet_dependent_function_name(
+        generator,
+        ctx,
+        "__nac3_ndarray_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_transpose<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    src_ndarray: Ptr<'ctx, StructModel<NDArray>>,
+    dst_ndarray: Ptr<'ctx, StructModel<NDArray>>,
+    num_axes: Int<'ctx, SizeT>,
+    axes: Ptr<'ctx, IntModel<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: Int<'ctx, SizeT>,
+    a_shape: Ptr<'ctx, IntModel<SizeT>>,
+    b_ndims: Int<'ctx, SizeT>,
+    b_shape: Ptr<'ctx, IntModel<SizeT>>,
+    final_ndims: Int<'ctx, SizeT>,
+    new_a_shape: Ptr<'ctx, IntModel<SizeT>>,
+    new_b_shape: Ptr<'ctx, IntModel<SizeT>>,
+    dst_shape: Ptr<'ctx, IntModel<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();
+}
+
+pub fn call_nac3_ndarray_float64_matmul_at_least_2d<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    a_ndarray: Ptr<'ctx, StructModel<NDArray>>,
+    b_ndarray: Ptr<'ctx, StructModel<NDArray>>,
+    dst_ndarray: Ptr<'ctx, StructModel<NDArray>>,
+) {
+    let name = get_sizet_dependent_function_name(
+        generator,
+        ctx,
+        "__nac3_ndarray_float64_matmul_at_least_2d",
+    );
+    CallFunction::begin(generator, ctx, &name)
+        .arg(a_ndarray)
+        .arg(b_ndarray)
+        .arg(dst_ndarray)
+        .returning_void();
+}
+
+pub fn call_nac3_array_set_and_validate_list_shape<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    list: Ptr<'ctx, StructModel<List<IntModel<Byte>>>>,
+    ndims: Int<'ctx, SizeT>,
+    shape: Ptr<'ctx, IntModel<SizeT>>,
+) {
+    let name = get_sizet_dependent_function_name(
+        generator,
+        ctx,
+        "__nac3_array_set_and_validate_list_shape",
+    );
+    CallFunction::begin(generator, ctx, &name).arg(list).arg(ndims).arg(shape).returning_void();
+}
+
+pub fn call_nac3_array_write_list_to_array<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    list: Ptr<'ctx, StructModel<List<IntModel<Byte>>>>,
+    ndarray: Ptr<'ctx, StructModel<NDArray>>,
+) {
+    let name =
+        get_sizet_dependent_function_name(generator, ctx, "__nac3_array_write_list_to_array");
+    CallFunction::begin(generator, ctx, &name).arg(list).arg(ndarray).returning_void();
+}
+
+pub fn call_nac3_nditer_initialize<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    iter: Ptr<'ctx, StructModel<NDIter>>,
+    ndarray: Ptr<'ctx, StructModel<NDArray>>,
+    indices: Ptr<'ctx, IntModel<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: Ptr<'ctx, StructModel<NDIter>>,
+) -> Int<'ctx, 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: Ptr<'ctx, StructModel<NDIter>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_nditer_next");
+    CallFunction::begin(generator, ctx, &name).arg(iter).returning_void();
+}

nac3core/src/codegen/irrt/test.rs

@@ -0,0 +1,26 @@
+#[cfg(test)]
+mod tests {
+    use std::{path::Path, process::Command};
+
+    #[test]
+    fn run_irrt_test() {
+        assert!(
+            cfg!(feature = "test"),
+            "Please do `cargo test -F test` to compile `irrt_test.out` and run test"
+        );
+
+        let irrt_test_out_path = Path::new(concat!(env!("OUT_DIR"), "/irrt_test.out"));
+        let output = Command::new(irrt_test_out_path.to_str().unwrap()).output().unwrap();
+
+        if !output.status.success() {
+            eprintln!("irrt_test failed with status {}:", output.status);
+            eprintln!("====== stdout ======");
+            eprintln!("{}", String::from_utf8(output.stdout).unwrap());
+            eprintln!("====== stderr ======");
+            eprintln!("{}", String::from_utf8(output.stderr).unwrap());
+            eprintln!("====================");
+
+            panic!("irrt_test failed");
+        }
+    }
+}

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},
@@ -24,6 +24,7 @@ use inkwell::{
     AddressSpace, IntPredicate, OptimizationLevel,
 };
 use itertools::Itertools;
+use model::*;
 use nac3parser::ast::{Location, Stmt, StrRef};
 use parking_lot::{Condvar, Mutex};
 use std::collections::{HashMap, HashSet};
@@ -32,8 +33,8 @@ use std::sync::{
     Arc,
 };
 use std::thread;
+use structure::{CSlice, Exception, NDArray};
 
-pub mod builtin_fns;
 pub mod classes;
 pub mod concrete_type;
 pub mod expr;
@@ -41,8 +42,12 @@ pub mod extern_fns;
 mod generator;
 pub mod irrt;
 pub mod llvm_intrinsics;
+pub mod model;
 pub mod numpy;
+pub mod numpy_new;
+pub mod object;
 pub mod stmt;
+pub mod structure;
 
 #[cfg(test)]
 mod test;
@@ -168,11 +173,11 @@ pub struct CodeGenContext<'ctx, 'a> {
     pub registry: &'a WorkerRegistry,
 
     /// Cache for constant strings.
-    pub const_strings: HashMap<String, BasicValueEnum<'ctx>>,
+    pub const_strings: HashMap<String, Struct<'ctx, CSlice>>,
 
     /// [`BasicBlock`] containing all `alloca` statements for the current function.
     pub init_bb: BasicBlock<'ctx>,
-    pub exception_val: Option<PointerValue<'ctx>>,
+    pub exception_val: Option<Ptr<'ctx, StructModel<Exception>>>,
 
     /// The header and exit basic blocks of a loop in this context. See
     /// <https://llvm.org/docs/LoopTerminology.html> for explanation of these terminology.
@@ -489,12 +494,8 @@ 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 pndarray_model = PtrModel(StructModel(NDArray));
-                            let element_type = get_llvm_type(
+                            pndarray_model.get_type(generator, ctx).as_basic_type_enum()
-                                ctx, module, generator, unifier, top_level, type_cache, dtype,
-                            );
-
-                            NDArrayType::new(generator, ctx, element_type).as_base_type().into()
                         }
 
                         _ => unreachable!(
@@ -700,43 +701,19 @@ pub fn gen_func_impl<
         ..primitives
     };
 
-    let mut type_cache: HashMap<_, _> = [
+    let cslice_model = StructModel(CSlice);
+    let pexn_model = PtrModel(StructModel(Exception));
+
+    let mut type_cache: HashMap<_, BasicTypeEnum<'ctx>> = [
         (primitives.int32, context.i32_type().into()),
         (primitives.int64, context.i64_type().into()),
         (primitives.uint32, context.i32_type().into()),
         (primitives.uint64, context.i64_type().into()),
         (primitives.float, context.f64_type().into()),
         (primitives.bool, context.i8_type().into()),
-        (primitives.str, {
+        (primitives.str, cslice_model.get_type(generator, context).into()),
-            let name = "str";
-            match module.get_struct_type(name) {
-                None => {
-                    let str_type = context.opaque_struct_type("str");
-                    let fields = [
-                        context.i8_type().ptr_type(AddressSpace::default()).into(),
-                        generator.get_size_type(context).into(),
-                    ];
-                    str_type.set_body(&fields, false);
-                    str_type.into()
-                }
-                Some(t) => t.as_basic_type_enum(),
-            }
-        }),
         (primitives.range, RangeType::new(context).as_base_type().into()),
-        (primitives.exception, {
+        (primitives.exception, pexn_model.get_type(generator, context).into()),
-            let name = "Exception";
-            if let Some(t) = module.get_struct_type(name) {
-                t.ptr_type(AddressSpace::default()).as_basic_type_enum()
-            } else {
-                let exception = context.opaque_struct_type("Exception");
-                let int32 = context.i32_type().into();
-                let int64 = context.i64_type().into();
-                let str_ty = module.get_struct_type("str").unwrap().as_basic_type_enum();
-                let fields = [int32, str_ty, int32, int32, str_ty, str_ty, int64, int64, int64];
-                exception.set_body(&fields, false);
-                exception.ptr_type(AddressSpace::default()).as_basic_type_enum()
-            }
-        }),
     ]
     .iter()
     .copied()

nac3core/src/codegen/model/any.rs

@@ -0,0 +1,40 @@
+use inkwell::{
+    context::Context,
+    types::{BasicType, BasicTypeEnum},
+    values::BasicValueEnum,
+};
+
+use crate::codegen::CodeGenerator;
+
+use super::*;
+
+#[derive(Debug, Clone, Copy)]
+pub struct AnyModel<'ctx>(pub BasicTypeEnum<'ctx>);
+pub type Anything<'ctx> = Instance<'ctx, AnyModel<'ctx>>;
+
+impl<'ctx> Model<'ctx> for AnyModel<'ctx> {
+    type Value = BasicValueEnum<'ctx>;
+    type Type = BasicTypeEnum<'ctx>;
+
+    fn get_type<G: CodeGenerator + ?Sized>(
+        &self,
+        _generator: &G,
+        _ctx: &'ctx Context,
+    ) -> Self::Type {
+        self.0
+    }
+
+    fn check_type<T: BasicType<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        _generator: &mut G,
+        _ctx: &'ctx Context,
+        ty: T,
+    ) -> Result<(), ModelError> {
+        let ty = ty.as_basic_type_enum();
+        if ty == self.0 {
+            Ok(())
+        } else {
+            Err(ModelError(format!("Expecting {}, but got {}", self.0, ty)))
+        }
+    }
+}

nac3core/src/codegen/model/array.rs

@@ -0,0 +1,122 @@
+use inkwell::{
+    context::Context,
+    types::{ArrayType, BasicType, BasicTypeEnum},
+    values::ArrayValue,
+};
+
+use crate::codegen::{CodeGenContext, CodeGenerator};
+
+use super::*;
+
+/// A Model for an [`ArrayType`].
+#[derive(Debug, Clone, Copy)]
+pub struct ArrayModel<Element> {
+    pub len: u32,
+    pub element: Element,
+}
+pub type Array<'ctx, Element> = Instance<'ctx, ArrayModel<Element>>;
+
+impl<'ctx, Element: Model<'ctx>> Model<'ctx> for ArrayModel<Element> {
+    type Value = ArrayValue<'ctx>;
+    type Type = ArrayType<'ctx>;
+
+    fn get_type<G: CodeGenerator + ?Sized>(&self, generator: &G, ctx: &'ctx Context) -> Self::Type {
+        self.element.get_type(generator, ctx).array_type(self.len)
+    }
+
+    fn check_type<T: BasicType<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+        ty: T,
+    ) -> Result<(), ModelError> {
+        let ty = ty.as_basic_type_enum();
+        let BasicTypeEnum::ArrayType(ty) = ty else {
+            return Err(ModelError(format!("Expecting ArrayType, but got {ty:?}")));
+        };
+
+        if ty.len() != self.len {
+            return Err(ModelError(format!(
+                "Expecting ArrayType with size {}, but got an ArrayType with size {}",
+                ty.len(),
+                self.len
+            )));
+        }
+
+        self.element
+            .check_type(generator, ctx, ty.get_element_type())
+            .map_err(|err| err.under_context("an ArrayType"))?;
+
+        Ok(())
+    }
+}
+
+impl<'ctx, Element: Model<'ctx>> Ptr<'ctx, ArrayModel<Element>> {
+    /// Get the pointer to the `i`-th (0-based) array element.
+    pub fn at<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        i: u32,
+        name: &str,
+    ) -> Ptr<'ctx, Element> {
+        assert!(i < self.model.0.len);
+
+        let zero = ctx.ctx.i32_type().const_zero();
+        let i = ctx.ctx.i32_type().const_int(u64::from(i), false);
+        let ptr = unsafe { ctx.builder.build_in_bounds_gep(self.value, &[zero, i], name).unwrap() };
+
+        PtrModel(self.model.0.element).check_value(generator, ctx.ctx, ptr).unwrap()
+    }
+}
+
+/// Like [`ArrayModel`] but length is strongly-typed.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct NArrayModel<const LEN: u32, Element>(pub Element);
+pub type NArray<'ctx, const LEN: u32, Element> = Instance<'ctx, NArrayModel<LEN, Element>>;
+
+impl<'ctx, const LEN: u32, Element: Model<'ctx>> NArrayModel<LEN, Element> {
+    /// Forget the `LEN` constant generic and get an [`ArrayModel`] with the same length.
+    pub fn forget_len(&self) -> ArrayModel<Element> {
+        ArrayModel { element: self.0, len: LEN }
+    }
+}
+
+impl<'ctx, const LEN: u32, Element: Model<'ctx>> Model<'ctx> for NArrayModel<LEN, Element> {
+    type Value = ArrayValue<'ctx>;
+    type Type = ArrayType<'ctx>;
+
+    fn get_type<G: CodeGenerator + ?Sized>(&self, generator: &G, ctx: &'ctx Context) -> Self::Type {
+        // Convenient implementation
+        self.forget_len().get_type(generator, ctx)
+    }
+
+    fn check_type<T: BasicType<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+        ty: T,
+    ) -> Result<(), ModelError> {
+        // Convenient implementation
+        self.forget_len().check_type(generator, ctx, ty)
+    }
+}
+
+impl<'ctx, const LEN: u32, Element: Model<'ctx>> Ptr<'ctx, NArrayModel<LEN, Element>> {
+    /// Get the pointer to the `i`-th (0-based) array element.
+    pub fn at_const<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        i: u32,
+        name: &str,
+    ) -> Ptr<'ctx, Element> {
+        assert!(i < LEN);
+
+        let zero = ctx.ctx.i32_type().const_zero();
+        let i = ctx.ctx.i32_type().const_int(u64::from(i), false);
+        let ptr = unsafe { ctx.builder.build_in_bounds_gep(self.value, &[zero, i], name).unwrap() };
+
+        PtrModel(self.model.0 .0).check_value(generator, ctx.ctx, ptr).unwrap()
+    }
+}

nac3core/src/codegen/model/core.rs

@@ -0,0 +1,123 @@
+use std::fmt;
+
+use inkwell::{context::Context, types::*, values::*};
+
+use super::*;
+use crate::codegen::{CodeGenContext, CodeGenerator};
+
+#[derive(Debug, Clone)]
+pub struct ModelError(pub String);
+
+impl ModelError {
+    pub(super) fn under_context(mut self, context: &str) -> Self {
+        self.0.push_str(" ... in ");
+        self.0.push_str(context);
+        self
+    }
+}
+
+pub trait Model<'ctx>: fmt::Debug + Clone + Copy {
+    type Value: BasicValue<'ctx> + TryFrom<BasicValueEnum<'ctx>>;
+    type Type: BasicType<'ctx>;
+
+    /// Return the [`BasicType`] of this model.
+    #[must_use]
+    fn get_type<G: CodeGenerator + ?Sized>(&self, generator: &G, ctx: &'ctx Context) -> Self::Type;
+
+    /// Check if a [`BasicType`] is the same type of this model.
+    fn check_type<T: BasicType<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+        ty: T,
+    ) -> Result<(), ModelError>;
+
+    /// Create an instance from a value with [`Instance::model`] being this model.
+    ///
+    /// Caller must make sure the type of `value` and the type of this `model` are equivalent.
+    #[must_use]
+    fn believe_value(&self, value: Self::Value) -> Instance<'ctx, Self> {
+        Instance { model: *self, value }
+    }
+
+    /// Check if a [`BasicValue`]'s type is equivalent to the type of this model.
+    /// Wrap it into an [`Instance`] if it is.
+    fn check_value<V: BasicValue<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+        value: V,
+    ) -> Result<Instance<'ctx, Self>, ModelError> {
+        let value = value.as_basic_value_enum();
+        self.check_type(generator, ctx, value.get_type())
+            .map_err(|err| err.under_context(format!("the value {value:?}").as_str()))?;
+
+        let Ok(value) = Self::Value::try_from(value) else {
+            unreachable!("check_type() has bad implementation")
+        };
+        Ok(self.believe_value(value))
+    }
+
+    // Allocate a value on the stack and return its pointer.
+    fn alloca<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        name: &str,
+    ) -> Ptr<'ctx, Self> {
+        let pmodel = PtrModel(*self);
+        let p = ctx.builder.build_alloca(self.get_type(generator, ctx.ctx), name).unwrap();
+        pmodel.believe_value(p)
+    }
+
+    // Allocate an array on the stack and return its pointer.
+    fn array_alloca<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        len: IntValue<'ctx>,
+        name: &str,
+    ) -> Ptr<'ctx, Self> {
+        let pmodel = PtrModel(*self);
+        let p =
+            ctx.builder.build_array_alloca(self.get_type(generator, ctx.ctx), len, name).unwrap();
+        pmodel.believe_value(p)
+    }
+
+    fn var_alloca<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        name: Option<&str>,
+    ) -> Result<Ptr<'ctx, Self>, String> {
+        let pmodel = PtrModel(*self);
+        let ty = self.get_type(generator, ctx.ctx).as_basic_type_enum();
+        let p = generator.gen_var_alloc(ctx, ty, name)?;
+        Ok(pmodel.believe_value(p))
+    }
+
+    fn array_var_alloca<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        len: IntValue<'ctx>,
+        name: Option<&'ctx str>,
+    ) -> Result<Ptr<'ctx, Self>, String> {
+        // TODO: Remove ArraySliceValue
+        let pmodel = PtrModel(*self);
+        let ty = self.get_type(generator, ctx.ctx).as_basic_type_enum();
+        let p = generator.gen_array_var_alloc(ctx, ty, len, name)?;
+        Ok(pmodel.believe_value(PointerValue::from(p)))
+    }
+}
+
+#[derive(Debug, Clone, Copy)]
+pub struct Instance<'ctx, M: Model<'ctx>> {
+    /// The model of this instance.
+    pub model: M,
+    /// The value of this instance.
+    ///
+    /// Caller must make sure the type of `value` and the type of this `model` are equivalent,
+    /// down to having the same [`IntType::get_bit_width`] in case of [`IntType`] for example.
+    pub value: M::Value,
+}

nac3core/src/codegen/model/float.rs

@@ -0,0 +1,88 @@
+use std::fmt;
+
+use inkwell::{context::Context, types::FloatType, values::FloatValue};
+
+use crate::codegen::CodeGenerator;
+
+use super::*;
+
+pub trait FloatKind<'ctx>: fmt::Debug + Clone + Copy {
+    fn get_float_type<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &G,
+        ctx: &'ctx Context,
+    ) -> FloatType<'ctx>;
+}
+
+#[derive(Debug, Clone, Copy, Default)]
+pub struct Float32;
+#[derive(Debug, Clone, Copy, Default)]
+pub struct Float64;
+
+impl<'ctx> FloatKind<'ctx> for Float32 {
+    fn get_float_type<G: CodeGenerator + ?Sized>(
+        &self,
+        _generator: &G,
+        ctx: &'ctx Context,
+    ) -> FloatType<'ctx> {
+        ctx.f32_type()
+    }
+}
+
+impl<'ctx> FloatKind<'ctx> for Float64 {
+    fn get_float_type<G: CodeGenerator + ?Sized>(
+        &self,
+        _generator: &G,
+        ctx: &'ctx Context,
+    ) -> FloatType<'ctx> {
+        ctx.f64_type()
+    }
+}
+
+#[derive(Debug, Clone, Copy)]
+pub struct AnyFloat<'ctx>(FloatType<'ctx>);
+
+impl<'ctx> FloatKind<'ctx> for AnyFloat<'ctx> {
+    fn get_float_type<G: CodeGenerator + ?Sized>(
+        &self,
+        _generator: &G,
+        _ctx: &'ctx Context,
+    ) -> FloatType<'ctx> {
+        self.0
+    }
+}
+
+#[derive(Debug, Clone, Copy, Default)]
+pub struct FloatModel<N>(pub N);
+pub type Float<'ctx, N> = Instance<'ctx, FloatModel<N>>;
+
+impl<'ctx, N: FloatKind<'ctx>> Model<'ctx> for FloatModel<N> {
+    type Value = FloatValue<'ctx>;
+    type Type = FloatType<'ctx>;
+
+    fn get_type<G: CodeGenerator + ?Sized>(&self, generator: &G, ctx: &'ctx Context) -> Self::Type {
+        self.0.get_float_type(generator, ctx)
+    }
+
+    fn check_type<T: inkwell::types::BasicType<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+        ty: T,
+    ) -> Result<(), ModelError> {
+        let ty = ty.as_basic_type_enum();
+        let Ok(ty) = FloatType::try_from(ty) else {
+            return Err(ModelError(format!("Expecting FloatType, but got {ty:?}")));
+        };
+
+        let exp_ty = self.0.get_float_type(generator, ctx);
+
+        // TODO: Inkwell does not have get_bit_width for FloatType?
+        // TODO: Quick hack for now, but does this actually work?
+        if ty != exp_ty {
+            return Err(ModelError(format!("Expecting {exp_ty:?}, but got {ty:?}")));
+        }
+
+        Ok(())
+    }
+}

nac3core/src/codegen/model/function.rs

@@ -0,0 +1,125 @@
+use inkwell::{
+    attributes::{Attribute, AttributeLoc},
+    types::{BasicMetadataTypeEnum, BasicType, FunctionType},
+    values::{AnyValue, BasicMetadataValueEnum, BasicValue, BasicValueEnum, CallSiteValue},
+};
+use itertools::Itertools;
+
+use crate::codegen::{CodeGenContext, CodeGenerator};
+
+use super::*;
+
+// When [`TypeContext::size_type`] is 32-bits, the function name is "{fn_name}".
+// When [`TypeContext::size_type`] is 64-bits, the function name is "{fn_name}64".
+#[must_use]
+pub fn get_sizet_dependent_function_name<G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &CodeGenContext<'_, '_>,
+    name: &str,
+) -> String {
+    let mut name = name.to_owned();
+    match generator.get_size_type(ctx.ctx).get_bit_width() {
+        32 => {}
+        64 => name.push_str("64"),
+        bit_width => {
+            panic!("Unsupported int type bit width {bit_width}, must be either 32-bits or 64-bits")
+        }
+    }
+    name
+}
+
+#[derive(Debug, Clone, Copy)]
+struct Arg<'ctx> {
+    ty: BasicMetadataTypeEnum<'ctx>,
+    val: BasicMetadataValueEnum<'ctx>,
+}
+
+/// A structure to construct & call an LLVM function.
+///
+/// This is a helper to reduce IRRT Inkwell function call boilerplate
+// TODO: Remove the lifetimes somehow? There is 4 of them.
+pub struct CallFunction<'ctx, 'a, 'b, 'c, 'd, G: CodeGenerator + ?Sized> {
+    generator: &'d mut G,
+    ctx: &'b CodeGenContext<'ctx, 'a>,
+    /// Function name
+    name: &'c str,
+    /// Call arguments
+    args: Vec<Arg<'ctx>>,
+    /// LLVM function Attributes
+    attrs: Vec<&'static str>,
+}
+
+impl<'ctx, 'a, 'b, 'c, 'd, G: CodeGenerator + ?Sized> CallFunction<'ctx, 'a, 'b, 'c, 'd, G> {
+    pub fn begin(generator: &'d mut G, ctx: &'b CodeGenContext<'ctx, 'a>, name: &'c str) -> Self {
+        CallFunction { generator, ctx, name, args: Vec::new(), attrs: Vec::new() }
+    }
+
+    /// Push a list of LLVM function attributes to the function declaration.
+    #[must_use]
+    pub fn attrs(mut self, attrs: Vec<&'static str>) -> Self {
+        self.attrs = attrs;
+        self
+    }
+
+    /// Push a call argument to the function call.
+    #[allow(clippy::needless_pass_by_value)]
+    #[must_use]
+    pub fn arg<M: Model<'ctx>>(mut self, arg: Instance<'ctx, M>) -> Self {
+        let arg = Arg {
+            ty: arg.model.get_type(self.generator, self.ctx.ctx).as_basic_type_enum().into(),
+            val: arg.value.as_basic_value_enum().into(),
+        };
+        self.args.push(arg);
+        self
+    }
+
+    /// Call the function and expect the function to return a value of type of `return_model`.
+    #[must_use]
+    pub fn returning<M: Model<'ctx>>(self, name: &str, return_model: M) -> Instance<'ctx, M> {
+        let ret_ty = return_model.get_type(self.generator, self.ctx.ctx);
+
+        let ret = self.get_function(|tys| ret_ty.fn_type(tys, false), name);
+        let ret = BasicValueEnum::try_from(ret.as_any_value_enum()).unwrap(); // Must work
+        let ret = return_model.check_value(self.generator, self.ctx.ctx, ret).unwrap(); // Must work
+        ret
+    }
+
+    /// Like [`CallFunction::returning_`] but `return_model` is automatically inferred.
+    #[must_use]
+    pub fn returning_auto<M: Model<'ctx> + Default>(self, name: &str) -> Instance<'ctx, M> {
+        self.returning(name, M::default())
+    }
+
+    /// Call the function and expect the function to return a void-type.
+    pub fn returning_void(self) {
+        let ret_ty = self.ctx.ctx.void_type();
+
+        let _ = self.get_function(|tys| ret_ty.fn_type(tys, false), "");
+    }
+
+    fn get_function<F>(&self, make_fn_type: F, return_value_name: &str) -> CallSiteValue<'ctx>
+    where
+        F: FnOnce(&[BasicMetadataTypeEnum<'ctx>]) -> FunctionType<'ctx>,
+    {
+        // Get the LLVM function.
+        let func = self.ctx.module.get_function(self.name).unwrap_or_else(|| {
+            // Declare the function if it doesn't exist.
+            let tys = self.args.iter().map(|arg| arg.ty).collect_vec();
+
+            let func_type = make_fn_type(&tys);
+            let func = self.ctx.module.add_function(self.name, func_type, None);
+
+            for attr in &self.attrs {
+                func.add_attribute(
+                    AttributeLoc::Function,
+                    self.ctx.ctx.create_enum_attribute(Attribute::get_named_enum_kind_id(attr), 0),
+                );
+            }
+
+            func
+        });
+
+        let vals = self.args.iter().map(|arg| arg.val).collect_vec();
+        self.ctx.builder.build_call(func, &vals, return_value_name).unwrap()
+    }
+}

nac3core/src/codegen/model/int.rs

@@ -0,0 +1,275 @@
+use std::fmt;
+
+use inkwell::{context::Context, types::IntType, values::IntValue, IntPredicate};
+
+use crate::codegen::{CodeGenContext, CodeGenerator};
+
+use super::*;
+
+pub trait IntKind<'ctx>: fmt::Debug + Clone + Copy {
+    fn get_int_type<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &G,
+        ctx: &'ctx Context,
+    ) -> IntType<'ctx>;
+}
+
+#[derive(Debug, Clone, Copy, Default)]
+pub struct Bool;
+#[derive(Debug, Clone, Copy, Default)]
+pub struct Byte;
+#[derive(Debug, Clone, Copy, Default)]
+pub struct Int32;
+#[derive(Debug, Clone, Copy, Default)]
+pub struct Int64;
+#[derive(Debug, Clone, Copy, Default)]
+pub struct SizeT;
+
+impl<'ctx> IntKind<'ctx> for Bool {
+    fn get_int_type<G: CodeGenerator + ?Sized>(
+        &self,
+        _generator: &G,
+        ctx: &'ctx Context,
+    ) -> IntType<'ctx> {
+        ctx.bool_type()
+    }
+}
+
+impl<'ctx> IntKind<'ctx> for Byte {
+    fn get_int_type<G: CodeGenerator + ?Sized>(
+        &self,
+        _generator: &G,
+        ctx: &'ctx Context,
+    ) -> IntType<'ctx> {
+        ctx.i8_type()
+    }
+}
+
+impl<'ctx> IntKind<'ctx> for Int32 {
+    fn get_int_type<G: CodeGenerator + ?Sized>(
+        &self,
+        _generator: &G,
+        ctx: &'ctx Context,
+    ) -> IntType<'ctx> {
+        ctx.i32_type()
+    }
+}
+
+impl<'ctx> IntKind<'ctx> for Int64 {
+    fn get_int_type<G: CodeGenerator + ?Sized>(
+        &self,
+        _generator: &G,
+        ctx: &'ctx Context,
+    ) -> IntType<'ctx> {
+        ctx.i64_type()
+    }
+}
+
+impl<'ctx> IntKind<'ctx> for SizeT {
+    fn get_int_type<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &G,
+        ctx: &'ctx Context,
+    ) -> IntType<'ctx> {
+        generator.get_size_type(ctx)
+    }
+}
+
+#[derive(Debug, Clone, Copy)]
+pub struct AnyInt<'ctx>(pub IntType<'ctx>);
+
+impl<'ctx> IntKind<'ctx> for AnyInt<'ctx> {
+    fn get_int_type<G: CodeGenerator + ?Sized>(
+        &self,
+        _generator: &G,
+        _ctx: &'ctx Context,
+    ) -> IntType<'ctx> {
+        self.0
+    }
+}
+
+#[derive(Debug, Clone, Copy, Default)]
+pub struct IntModel<N>(pub N);
+pub type Int<'ctx, N> = Instance<'ctx, IntModel<N>>;
+
+impl<'ctx, N: IntKind<'ctx>> Model<'ctx> for IntModel<N> {
+    type Value = IntValue<'ctx>;
+    type Type = IntType<'ctx>;
+
+    fn get_type<G: CodeGenerator + ?Sized>(&self, generator: &G, ctx: &'ctx Context) -> Self::Type {
+        self.0.get_int_type(generator, ctx)
+    }
+
+    fn check_type<T: inkwell::types::BasicType<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+        ty: T,
+    ) -> Result<(), ModelError> {
+        let ty = ty.as_basic_type_enum();
+        let Ok(ty) = IntType::try_from(ty) else {
+            return Err(ModelError(format!("Expecting IntType, but got {ty:?}")));
+        };
+
+        let exp_ty = self.0.get_int_type(generator, ctx);
+        if ty.get_bit_width() != exp_ty.get_bit_width() {
+            return Err(ModelError(format!(
+                "Expecting IntType to have {} bit(s), but got {} bit(s)",
+                exp_ty.get_bit_width(),
+                ty.get_bit_width()
+            )));
+        }
+
+        Ok(())
+    }
+}
+
+impl<'ctx, N: IntKind<'ctx>> IntModel<N> {
+    pub fn constant<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+        value: u64,
+    ) -> Int<'ctx, N> {
+        let value = self.get_type(generator, ctx).const_int(value, false);
+        self.believe_value(value)
+    }
+
+    pub fn const_0<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> Int<'ctx, N> {
+        self.constant(generator, ctx, 0)
+    }
+
+    pub fn const_1<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> Int<'ctx, N> {
+        self.constant(generator, ctx, 1)
+    }
+
+    pub fn const_all_1s<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> Int<'ctx, N> {
+        let value = self.get_type(generator, ctx).const_all_ones();
+        self.believe_value(value)
+    }
+
+    pub fn s_extend_or_bit_cast<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        value: IntValue<'ctx>,
+        name: &str,
+    ) -> Int<'ctx, N> {
+        let value = ctx
+            .builder
+            .build_int_s_extend_or_bit_cast(value, self.get_type(generator, ctx.ctx), name)
+            .unwrap();
+        self.believe_value(value)
+    }
+
+    pub fn truncate<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        value: IntValue<'ctx>,
+        name: &str,
+    ) -> Int<'ctx, N> {
+        let value =
+            ctx.builder.build_int_truncate(value, self.get_type(generator, ctx.ctx), name).unwrap();
+        self.believe_value(value)
+    }
+}
+
+impl IntModel<Bool> {
+    #[must_use]
+    pub fn const_false<'ctx, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> Int<'ctx, Bool> {
+        self.constant(generator, ctx, 0)
+    }
+
+    #[must_use]
+    pub fn const_true<'ctx, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> Int<'ctx, Bool> {
+        self.constant(generator, ctx, 1)
+    }
+}
+
+impl<'ctx, N: IntKind<'ctx>> Int<'ctx, N> {
+    pub fn s_extend_or_bit_cast<NewN: IntKind<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        to_int_kind: NewN,
+        name: &str,
+    ) -> Int<'ctx, NewN> {
+        IntModel(to_int_kind).s_extend_or_bit_cast(generator, ctx, self.value, name)
+    }
+
+    pub fn truncate<NewN: IntKind<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        to_int_kind: NewN,
+        name: &str,
+    ) -> Int<'ctx, NewN> {
+        IntModel(to_int_kind).truncate(generator, ctx, self.value, name)
+    }
+
+    #[must_use]
+    pub fn add(
+        &self,
+        ctx: &CodeGenContext<'ctx, '_>,
+        other: Int<'ctx, N>,
+        name: &str,
+    ) -> Int<'ctx, N> {
+        let value = ctx.builder.build_int_add(self.value, other.value, name).unwrap();
+        self.model.believe_value(value)
+    }
+
+    #[must_use]
+    pub fn sub(
+        &self,
+        ctx: &CodeGenContext<'ctx, '_>,
+        other: Int<'ctx, N>,
+        name: &str,
+    ) -> Int<'ctx, N> {
+        let value = ctx.builder.build_int_sub(self.value, other.value, name).unwrap();
+        self.model.believe_value(value)
+    }
+
+    #[must_use]
+    pub fn mul(
+        &self,
+        ctx: &CodeGenContext<'ctx, '_>,
+        other: Int<'ctx, N>,
+        name: &str,
+    ) -> Int<'ctx, N> {
+        let value = ctx.builder.build_int_mul(self.value, other.value, name).unwrap();
+        self.model.believe_value(value)
+    }
+
+    pub fn compare(
+        &self,
+        ctx: &CodeGenContext<'ctx, '_>,
+        op: IntPredicate,
+        other: Int<'ctx, N>,
+        name: &str,
+    ) -> Int<'ctx, Bool> {
+        let bool_model = IntModel(Bool);
+        let value = ctx.builder.build_int_compare(op, self.value, other.value, name).unwrap();
+        bool_model.believe_value(value)
+    }
+}

nac3core/src/codegen/model/mod.rs

@@ -0,0 +1,17 @@
+mod any;
+mod array;
+mod core;
+mod float;
+pub mod function;
+mod int;
+mod ptr;
+mod structure;
+pub mod util;
+
+pub use any::*;
+pub use array::*;
+pub use core::*;
+pub use float::*;
+pub use int::*;
+pub use ptr::*;
+pub use structure::*;

nac3core/src/codegen/model/ptr.rs

@@ -0,0 +1,145 @@
+use inkwell::{
+    context::Context,
+    types::{BasicType, BasicTypeEnum, PointerType},
+    values::{IntValue, PointerValue},
+    AddressSpace,
+};
+
+use crate::codegen::{CodeGenContext, CodeGenerator};
+
+use super::*;
+
+#[derive(Debug, Clone, Copy, Default)]
+pub struct PtrModel<Element>(pub Element);
+pub type Ptr<'ctx, Element> = Instance<'ctx, PtrModel<Element>>;
+
+impl<'ctx, Element: Model<'ctx>> Model<'ctx> for PtrModel<Element> {
+    type Value = PointerValue<'ctx>;
+    type Type = PointerType<'ctx>;
+
+    fn get_type<G: CodeGenerator + ?Sized>(&self, generator: &G, ctx: &'ctx Context) -> Self::Type {
+        self.0.get_type(generator, ctx).ptr_type(AddressSpace::default())
+    }
+
+    fn check_type<T: BasicType<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+        ty: T,
+    ) -> Result<(), ModelError> {
+        let ty = ty.as_basic_type_enum();
+        let Ok(ty) = PointerType::try_from(ty) else {
+            return Err(ModelError(format!("Expecting PointerType, but got {ty:?}")));
+        };
+
+        let elem_ty = ty.get_element_type();
+        let Ok(elem_ty) = BasicTypeEnum::try_from(elem_ty) else {
+            return Err(ModelError(format!(
+                "Expecting pointer element type to be a BasicTypeEnum, but got {elem_ty:?}"
+            )));
+        };
+
+        // TODO: inkwell `get_element_type()` will be deprecated.
+        // Remove the check for `get_element_type()` when the time comes.
+        self.0
+            .check_type(generator, ctx, elem_ty)
+            .map_err(|err| err.under_context("a PointerType"))?;
+
+        Ok(())
+    }
+}
+
+impl<'ctx, Element: Model<'ctx>> PtrModel<Element> {
+    /// Return a ***constant*** nullptr.
+    pub fn nullptr<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> Ptr<'ctx, Element> {
+        let ptr = self.get_type(generator, ctx).const_null();
+        self.believe_value(ptr)
+    }
+
+    /// Cast a pointer into this model with [`inkwell::builder::Builder::build_pointer_cast`]
+    pub fn pointer_cast<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        ptr: PointerValue<'ctx>,
+        name: &str,
+    ) -> Ptr<'ctx, Element> {
+        let ptr =
+            ctx.builder.build_pointer_cast(ptr, self.get_type(generator, ctx.ctx), name).unwrap();
+        self.believe_value(ptr)
+    }
+}
+
+impl<'ctx, Element: Model<'ctx>> Ptr<'ctx, Element> {
+    /// Offset the pointer by [`inkwell::builder::Builder::build_in_bounds_gep`].
+    #[must_use]
+    pub fn offset<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        offset: IntValue<'ctx>,
+        name: &str,
+    ) -> Ptr<'ctx, Element> {
+        let new_ptr =
+            unsafe { ctx.builder.build_in_bounds_gep(self.value, &[offset], name).unwrap() };
+        self.model.check_value(generator, ctx.ctx, new_ptr).unwrap()
+    }
+
+    /// Offset the pointer by [`inkwell::builder::Builder::build_in_bounds_gep`] by a constant offset.
+    #[must_use]
+    pub fn offset_const<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        offset: u64,
+        name: &str,
+    ) -> Ptr<'ctx, Element> {
+        let offset = ctx.ctx.i32_type().const_int(offset, false);
+        self.offset(generator, ctx, offset, name)
+    }
+
+    /// Load the value with [`inkwell::builder::Builder::build_load`].
+    pub fn load<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        name: &str,
+    ) -> Instance<'ctx, Element> {
+        let value = ctx.builder.build_load(self.value, name).unwrap();
+        self.model.0.check_value(generator, ctx.ctx, value).unwrap() // If unwrap() panics, there is a logic error.
+    }
+
+    /// Store a value with [`inkwell::builder::Builder::build_store`].
+    pub fn store(&self, ctx: &CodeGenContext<'ctx, '_>, value: Instance<'ctx, Element>) {
+        ctx.builder.build_store(self.value, value.value).unwrap();
+    }
+
+    /// Return a casted pointer of element type `NewElement` with [`inkwell::builder::Builder::build_pointer_cast`].
+    pub fn pointer_cast<NewElement: Model<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        new_model: NewElement,
+        name: &str,
+    ) -> Ptr<'ctx, NewElement> {
+        PtrModel(new_model).pointer_cast(generator, ctx, self.value, name)
+    }
+
+    /// Check if the pointer is null with [`inkwell::builder::Builder::build_is_null`].
+    pub fn is_null(&self, ctx: &CodeGenContext<'ctx, '_>, name: &str) -> Int<'ctx, Bool> {
+        let bool_model = IntModel(Bool);
+        let value = ctx.builder.build_is_null(self.value, name).unwrap();
+        bool_model.believe_value(value)
+    }
+
+    /// Check if the pointer is not null with [`inkwell::builder::Builder::build_is_not_null`].
+    pub fn is_not_null(&self, ctx: &CodeGenContext<'ctx, '_>, name: &str) -> Int<'ctx, Bool> {
+        let bool_model = IntModel(Bool);
+        let value = ctx.builder.build_is_not_null(self.value, name).unwrap();
+        bool_model.believe_value(value)
+    }
+}

nac3core/src/codegen/model/structure.rs

@@ -0,0 +1,222 @@
+use std::fmt;
+
+use inkwell::{
+    context::Context,
+    types::{BasicType, BasicTypeEnum, StructType},
+    values::StructValue,
+};
+
+use crate::codegen::{CodeGenContext, CodeGenerator};
+
+use super::*;
+
+#[derive(Debug, Clone, Copy)]
+pub struct GepField<M> {
+    pub gep_index: u64,
+    pub name: &'static str,
+    pub model: M,
+}
+
+pub trait FieldTraversal<'ctx> {
+    type Out<M>;
+
+    fn add<M: Model<'ctx>>(&mut self, name: &'static str, model: M) -> Self::Out<M>;
+
+    /// Like [`FieldTraversal::visit`] but [`Model`] is automatically inferred from [`Default`] trait.
+    fn add_auto<M: Model<'ctx> + Default>(&mut self, name: &'static str) -> Self::Out<M> {
+        self.add(name, M::default())
+    }
+}
+
+pub struct GepFieldTraversal {
+    gep_index_counter: u64,
+}
+
+impl<'ctx> FieldTraversal<'ctx> for GepFieldTraversal {
+    type Out<M> = GepField<M>;
+
+    fn add<M: Model<'ctx>>(&mut self, name: &'static str, model: M) -> Self::Out<M> {
+        let gep_index = self.gep_index_counter;
+        self.gep_index_counter += 1;
+        Self::Out { gep_index, name, model }
+    }
+}
+
+struct TypeFieldTraversal<'ctx, 'a, G: CodeGenerator + ?Sized> {
+    generator: &'a G,
+    ctx: &'ctx Context,
+    field_types: Vec<BasicTypeEnum<'ctx>>,
+}
+
+impl<'ctx, 'a, G: CodeGenerator + ?Sized> FieldTraversal<'ctx> for TypeFieldTraversal<'ctx, 'a, G> {
+    type Out<M> = ();
+
+    fn add<M: Model<'ctx>>(&mut self, _name: &'static str, model: M) -> Self::Out<M> {
+        let t = model.get_type(self.generator, self.ctx).as_basic_type_enum();
+        self.field_types.push(t);
+    }
+}
+
+struct CheckTypeFieldTraversal<'ctx, 'a, G: CodeGenerator + ?Sized> {
+    generator: &'a mut G,
+    ctx: &'ctx Context,
+    index: u32,
+    scrutinee: StructType<'ctx>,
+    errors: Vec<ModelError>,
+}
+
+impl<'ctx, 'a, G: CodeGenerator + ?Sized> FieldTraversal<'ctx>
+    for CheckTypeFieldTraversal<'ctx, 'a, G>
+{
+    type Out<M> = ();
+
+    fn add<M: Model<'ctx>>(&mut self, name: &'static str, model: M) -> Self::Out<M> {
+        let i = self.index;
+        self.index += 1;
+
+        if let Some(t) = self.scrutinee.get_field_type_at_index(i) {
+            if let Err(err) = model.check_type(self.generator, self.ctx, t) {
+                self.errors.push(err.under_context(format!("field #{i} '{name}'").as_str()));
+            }
+        } // Otherwise, it will be caught
+    }
+}
+
+pub trait StructKind<'ctx>: fmt::Debug + Clone + Copy {
+    type Fields<F: FieldTraversal<'ctx>>;
+
+    fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F>;
+
+    fn fields(&self) -> Self::Fields<GepFieldTraversal> {
+        self.traverse_fields(&mut GepFieldTraversal { gep_index_counter: 0 })
+    }
+
+    fn get_struct_type<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &G,
+        ctx: &'ctx Context,
+    ) -> StructType<'ctx> {
+        let mut traversal = TypeFieldTraversal { generator, ctx, field_types: Vec::new() };
+        self.traverse_fields(&mut traversal);
+
+        ctx.struct_type(&traversal.field_types, false)
+    }
+}
+
+#[derive(Debug, Clone, Copy, Default)]
+pub struct StructModel<S>(pub S);
+pub type Struct<'ctx, S> = Instance<'ctx, StructModel<S>>;
+
+impl<'ctx, S: StructKind<'ctx>> Model<'ctx> for StructModel<S> {
+    type Value = StructValue<'ctx>;
+    type Type = StructType<'ctx>;
+
+    fn get_type<G: CodeGenerator + ?Sized>(&self, generator: &G, ctx: &'ctx Context) -> Self::Type {
+        self.0.get_struct_type(generator, ctx)
+    }
+
+    fn check_type<T: BasicType<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+        ty: T,
+    ) -> Result<(), ModelError> {
+        let ty = ty.as_basic_type_enum();
+        let Ok(ty) = StructType::try_from(ty) else {
+            return Err(ModelError(format!("Expecting StructType, but got {ty:?}")));
+        };
+
+        let mut traversal =
+            CheckTypeFieldTraversal { generator, ctx, index: 0, errors: Vec::new(), scrutinee: ty };
+        self.0.traverse_fields(&mut traversal);
+
+        let exp_num_fields = traversal.index;
+        let got_num_fields = u32::try_from(ty.get_field_types().len()).unwrap();
+        if exp_num_fields != got_num_fields {
+            return Err(ModelError(format!(
+                "Expecting StructType with {exp_num_fields} field(s), but got {got_num_fields}"
+            )));
+        }
+
+        if !traversal.errors.is_empty() {
+            return Err(traversal.errors[0].clone()); // TODO: Return other errors as well
+        }
+
+        Ok(())
+    }
+}
+
+impl<'ctx, S: StructKind<'ctx>> Struct<'ctx, S> {
+    pub fn get_field<G: CodeGenerator + ?Sized, M, GetField>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+        get_field: GetField,
+    ) -> Instance<'ctx, M>
+    where
+        M: Model<'ctx>,
+        GetField: FnOnce(S::Fields<GepFieldTraversal>) -> GepField<M>,
+    {
+        let field = get_field(self.model.0.fields());
+        let val = self.value.get_field_at_index(field.gep_index as u32).unwrap();
+        field.model.check_value(generator, ctx, val).unwrap()
+    }
+}
+
+impl<'ctx, S: StructKind<'ctx>> Ptr<'ctx, StructModel<S>> {
+    pub fn gep<M, GetField>(
+        &self,
+        ctx: &CodeGenContext<'ctx, '_>,
+        get_field: GetField,
+    ) -> Ptr<'ctx, M>
+    where
+        M: Model<'ctx>,
+        GetField: FnOnce(S::Fields<GepFieldTraversal>) -> GepField<M>,
+    {
+        let field = get_field(self.model.0 .0.fields());
+        let llvm_i32 = ctx.ctx.i32_type(); // i64 would segfault
+
+        let ptr = unsafe {
+            ctx.builder
+                .build_in_bounds_gep(
+                    self.value,
+                    &[llvm_i32.const_zero(), llvm_i32.const_int(field.gep_index, false)],
+                    field.name,
+                )
+                .unwrap()
+        };
+
+        let ptr_model = PtrModel(field.model);
+        ptr_model.believe_value(ptr)
+    }
+
+    /// Convenience function equivalent to `.gep(...).load(...)`.
+    pub fn get<M, GetField, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        get_field: GetField,
+        name: &str,
+    ) -> Instance<'ctx, M>
+    where
+        M: Model<'ctx>,
+        GetField: FnOnce(S::Fields<GepFieldTraversal>) -> GepField<M>,
+    {
+        self.gep(ctx, get_field).load(generator, ctx, name)
+    }
+
+    /// Convenience function equivalent to `.gep(...).store(...)`.
+    pub fn set<M, GetField>(
+        &self,
+        ctx: &CodeGenContext<'ctx, '_>,
+        get_field: GetField,
+        value: Instance<'ctx, M>,
+    ) where
+        M: Model<'ctx>,
+        GetField: FnOnce(S::Fields<GepFieldTraversal>) -> GepField<M>,
+    {
+        self.gep(ctx, get_field).store(ctx, value);
+    }
+}
+
+// TODO: Add an opaque struct type?

nac3core/src/codegen/model/util.rs

@@ -0,0 +1,62 @@
+use inkwell::{types::BasicType, values::IntValue};
+
+/// `llvm.memcpy` but under the [`Model`] abstraction
+use crate::codegen::{
+    llvm_intrinsics::call_memcpy_generic,
+    stmt::{gen_for_callback_incrementing, BreakContinueHooks},
+    CodeGenContext, CodeGenerator,
+};
+
+use super::*;
+
+/// Convenience function.
+///
+/// Like [`call_memcpy_generic`] but with model abstractions and `is_volatile` set to `false`.
+pub fn call_memcpy_model<'ctx, Item: Model<'ctx> + Default, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &CodeGenContext<'ctx, '_>,
+    dst_array: Ptr<'ctx, Item>,
+    src_array: Ptr<'ctx, Item>,
+    num_items: IntValue<'ctx>,
+) {
+    let itemsize = Item::default().get_type(generator, ctx.ctx).size_of().unwrap();
+    let totalsize = ctx.builder.build_int_mul(itemsize, num_items, "totalsize").unwrap(); // TODO: Int types may not match.
+    let is_volatile = ctx.ctx.bool_type().const_zero();
+    call_memcpy_generic(ctx, dst_array.value, src_array.value, totalsize, is_volatile);
+}
+
+/// Like [`gen_for_callback_incrementing`] with [`Model`] abstractions.
+/// The [`IntKind`] is automatically inferred.
+pub fn gen_for_model_auto<'ctx, 'a, G, F, I>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, 'a>,
+    start: Int<'ctx, I>,
+    stop: Int<'ctx, I>,
+    step: Int<'ctx, I>,
+    body: F,
+) -> Result<(), String>
+where
+    G: CodeGenerator + ?Sized,
+    F: FnOnce(
+        &mut G,
+        &mut CodeGenContext<'ctx, 'a>,
+        BreakContinueHooks<'ctx>,
+        Int<'ctx, I>,
+    ) -> Result<(), String>,
+    I: IntKind<'ctx> + Default,
+{
+    let int_model = IntModel(I::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/numpy.rs

@@ -257,7 +257,7 @@ fn ndarray_zero_value<'ctx, G: CodeGenerator + ?Sized>(
     } else if ctx.unifier.unioned(elem_ty, ctx.primitives.bool) {
         ctx.ctx.bool_type().const_zero().into()
     } else if ctx.unifier.unioned(elem_ty, ctx.primitives.str) {
-        ctx.gen_string(generator, "").into()
+        ctx.gen_string(generator, "").value.into()
     } else {
         unreachable!()
     }
@@ -285,7 +285,7 @@ fn ndarray_one_value<'ctx, G: CodeGenerator + ?Sized>(
     } else if ctx.unifier.unioned(elem_ty, ctx.primitives.bool) {
         ctx.ctx.bool_type().const_int(1, false).into()
     } else if ctx.unifier.unioned(elem_ty, ctx.primitives.str) {
-        ctx.gen_string(generator, "1").into()
+        ctx.gen_string(generator, "1").value.into()
     } else {
         unreachable!()
     }

nac3core/src/codegen/numpy_new.rs

@@ -0,0 +1,210 @@
+// TODO: Replace numpy.rs
+
+use inkwell::values::{BasicValue, BasicValueEnum};
+use nac3parser::ast::StrRef;
+
+use crate::{
+    codegen::object::{ndarray::scalar::split_scalar_or_ndarray, tuple::TupleObject},
+    symbol_resolver::ValueEnum,
+    toplevel::{helper::extract_ndims, numpy::unpack_ndarray_var_tys, DefinitionId},
+    typecheck::typedef::{FunSignature, Type},
+};
+
+use super::{
+    irrt::call_nac3_ndarray_util_assert_shape_no_negative,
+    model::*,
+    object::{
+        ndarray::{shape_util::parse_numpy_int_sequence, NDArrayObject},
+        AnyObject,
+    },
+    CodeGenContext, CodeGenerator,
+};
+
+/// Generates LLVM IR for `np.broadcast_to`.
+pub fn gen_ndarray_broadcast_to<'ctx>(
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    obj: &Option<(Type, ValueEnum<'ctx>)>,
+    fun: (&FunSignature, DefinitionId),
+    args: &[(Option<StrRef>, ValueEnum<'ctx>)],
+    generator: &mut dyn CodeGenerator,
+) -> Result<BasicValueEnum<'ctx>, String> {
+    assert!(obj.is_none());
+    assert_eq!(args.len(), 2);
+
+    // Parse argument #1 input
+    let input_ty = fun.0.args[0].ty;
+    let input = args[0].1.clone().to_basic_value_enum(ctx, generator, input_ty)?;
+    let input = AnyObject { ty: input_ty, value: input };
+
+    // Parse argument #2 shape
+    let shape_ty = fun.0.args[1].ty;
+    let shape = args[1].1.clone().to_basic_value_enum(ctx, generator, shape_ty)?;
+    let shape = AnyObject { ty: shape_ty, value: shape };
+
+    // Define models
+    let sizet_model = IntModel(SizeT);
+
+    // Extract broadcast_ndims, this is the only way to get the
+    // ndims of the ndarray result statically.
+    let (_, broadcast_ndims_ty) = unpack_ndarray_var_tys(&mut ctx.unifier, fun.0.ret);
+    let broadcast_ndims = extract_ndims(&ctx.unifier, broadcast_ndims_ty);
+
+    // Process `input`
+    let in_ndarray = split_scalar_or_ndarray(generator, ctx, input).as_ndarray(generator, ctx);
+
+    // Process `shape`
+    let (_, broadcast_shape) = parse_numpy_int_sequence(generator, ctx, shape);
+    // NOTE: shape.size should equal to `broadcasted_ndims`.
+    let broadcast_ndims_llvm = sizet_model.constant(generator, ctx.ctx, broadcast_ndims);
+    call_nac3_ndarray_util_assert_shape_no_negative(
+        generator,
+        ctx,
+        broadcast_ndims_llvm,
+        broadcast_shape,
+    );
+
+    // Create broadcast view
+    let broadcast_ndarray =
+        in_ndarray.broadcast_to(generator, ctx, broadcast_ndims, broadcast_shape);
+
+    Ok(broadcast_ndarray.instance.value.as_basic_value_enum())
+}
+
+/// Generates LLVM IR for `np.reshape`.
+pub fn gen_ndarray_reshape<'ctx>(
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    obj: &Option<(Type, ValueEnum<'ctx>)>,
+    fun: (&FunSignature, DefinitionId),
+    args: &[(Option<StrRef>, ValueEnum<'ctx>)],
+    generator: &mut dyn CodeGenerator,
+) -> Result<BasicValueEnum<'ctx>, String> {
+    assert!(obj.is_none());
+    assert_eq!(args.len(), 2);
+
+    // Parse argument #1 input
+    let input_ty = fun.0.args[0].ty;
+    let input = args[0].1.clone().to_basic_value_enum(ctx, generator, input_ty)?;
+    let input = AnyObject { ty: input_ty, value: input };
+
+    // Parse argument #2 shape
+    let shape_ty = fun.0.args[1].ty;
+    let shape = args[1].1.clone().to_basic_value_enum(ctx, generator, shape_ty)?;
+    let shape = AnyObject { ty: shape_ty, value: shape };
+
+    // Extract reshaped_ndims
+    let (_, reshaped_ndims_ty) = unpack_ndarray_var_tys(&mut ctx.unifier, fun.0.ret);
+    let reshaped_ndims = extract_ndims(&ctx.unifier, reshaped_ndims_ty);
+
+    // Process `input`
+    let in_ndarray = split_scalar_or_ndarray(generator, ctx, input).as_ndarray(generator, ctx);
+
+    // Process the shape input from user and resolve negative indices.
+    // The resulting `new_shape`'s size should be equal to reshaped_ndims.
+    // This is ensured by the typechecker.
+    let (_, new_shape) = parse_numpy_int_sequence(generator, ctx, shape);
+    let reshaped_ndarray = in_ndarray.reshape_or_copy(generator, ctx, reshaped_ndims, new_shape);
+
+    Ok(reshaped_ndarray.instance.value.as_basic_value_enum())
+}
+
+/// Generates LLVM IR for `np.arange`.
+pub fn gen_ndarray_arange<'ctx>(
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    obj: &Option<(Type, ValueEnum<'ctx>)>,
+    fun: (&FunSignature, DefinitionId),
+    args: &[(Option<StrRef>, ValueEnum<'ctx>)],
+    generator: &mut dyn CodeGenerator,
+) -> Result<BasicValueEnum<'ctx>, String> {
+    assert!(obj.is_none());
+    assert_eq!(args.len(), 1);
+
+    // Parse argument #1 len
+    let input_ty = fun.0.args[0].ty;
+    let input = args[0].1.clone().to_basic_value_enum(ctx, generator, input_ty)?.into_int_value();
+
+    // Implementation
+    let input_dim = IntModel(SizeT).s_extend_or_bit_cast(generator, ctx, input, "input_dim");
+    let ndarray = NDArrayObject::from_np_arange(generator, ctx, input_dim);
+
+    Ok(ndarray.instance.value.as_basic_value_enum())
+}
+
+/// Generates LLVM IR for `np.size`.
+pub fn gen_ndarray_size<'ctx>(
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    obj: &Option<(Type, ValueEnum<'ctx>)>,
+    fun: (&FunSignature, DefinitionId),
+    args: &[(Option<StrRef>, ValueEnum<'ctx>)],
+    generator: &mut dyn CodeGenerator,
+) -> Result<BasicValueEnum<'ctx>, String> {
+    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 size = ndarray.size(generator, ctx).truncate(generator, ctx, Int32, "size");
+    Ok(size.value.as_basic_value_enum())
+}
+
+/// Generates LLVM IR for `np.shape`.
+pub fn gen_ndarray_shape<'ctx>(
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    obj: &Option<(Type, ValueEnum<'ctx>)>,
+    fun: (&FunSignature, DefinitionId),
+    args: &[(Option<StrRef>, ValueEnum<'ctx>)],
+    generator: &mut dyn CodeGenerator,
+) -> Result<BasicValueEnum<'ctx>, String> {
+    assert!(obj.is_none());
+    assert_eq!(args.len(), 1);
+
+    // Parse argument #1 ndarray
+    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 };
+
+    // Process ndarray
+    let ndarray = NDArrayObject::from_object(generator, ctx, ndarray);
+    Ok(ndarray.make_shape_tuple(generator, ctx).value.as_basic_value_enum())
+}
+
+/// Generates LLVM IR for `<ndarray>.strides`.
+pub fn gen_ndarray_strides<'ctx>(
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    obj: &Option<(Type, ValueEnum<'ctx>)>,
+    fun: (&FunSignature, DefinitionId),
+    args: &[(Option<StrRef>, ValueEnum<'ctx>)],
+    generator: &mut dyn CodeGenerator,
+) -> Result<BasicValueEnum<'ctx>, String> {
+    // TODO: Code duplication: This function looks exactly like `gen_ndarray_shapes`.
+
+    assert!(obj.is_none());
+    assert_eq!(args.len(), 1);
+
+    // Parse argument #1 ndarray
+    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 };
+
+    // Process ndarray
+    let ndarray = NDArrayObject::from_object(generator, ctx, ndarray);
+
+    let mut objects = Vec::with_capacity(ndarray.ndims as usize);
+
+    for i in 0..ndarray.ndims {
+        let dim = ndarray
+            .instance
+            .get(generator, ctx, |f| f.strides, "")
+            .offset_const(generator, ctx, i, "")
+            .load(generator, ctx, "dim");
+        let dim = dim.truncate(generator, ctx, Int32, "dim"); // TODO: keep using SizeT
+
+        objects
+            .push(AnyObject { ty: ctx.primitives.int32, value: dim.value.as_basic_value_enum() });
+    }
+
+    let strides = TupleObject::create(generator, ctx, objects, "strides");
+    Ok(strides.value.as_basic_value_enum())
+}

nac3core/src/codegen/object/list.rs

@@ -0,0 +1,121 @@
+use crate::{
+    codegen::{
+        irrt::{call_nac3_list_slice_assign, list_slice_assignment},
+        model::*,
+        object::ndarray::indexing::UserSlice,
+        structure::List,
+        CodeGenContext, CodeGenerator,
+    },
+    typecheck::typedef::{iter_type_vars, Type, TypeEnum},
+};
+
+use super::{ndarray::indexing::RustUserSlice, AnyObject};
+
+/// 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: Ptr<'ctx, StructModel<List<AnyModel<'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 item_model = AnyModel(ctx.get_llvm_type(generator, item_type));
+        let plist_model = PtrModel(StructModel(List { item: item_model }));
+
+        // Create object
+        let value = plist_model.check_value(generator, ctx.ctx, object.value).unwrap();
+        ListObject { item_type, instance: value }
+    }
+
+    /// 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, '_>,
+    ) -> Ptr<'ctx, IntModel<Byte>> {
+        self.instance.get(generator, ctx, |f| f.items, "items").pointer_cast(
+            generator,
+            ctx,
+            IntModel(Byte),
+            "items_opaque",
+        )
+    }
+
+    /// 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, '_>,
+    ) -> Ptr<'ctx, StructModel<List<IntModel<Byte>>>> {
+        let opaque_list_model = StructModel(List { item: IntModel(Byte) });
+        let opaque_list_ptr = opaque_list_model.alloca(generator, ctx, "opaque_list_ptr");
+
+        // Copy items pointer
+        let items = self.get_opaque_items_ptr(generator, ctx);
+        opaque_list_ptr.set(ctx, |f| f.items, items);
+
+        // Copy len
+        let len = self.instance.get(generator, ctx, |f| f.len, "len");
+        opaque_list_ptr.set(ctx, |f| f.len, len);
+
+        opaque_list_ptr
+    }
+
+    /// Get the `len()` of this list.
+    pub fn len<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Int<'ctx, SizeT> {
+        self.instance.get(generator, ctx, |f| f.len, "list_len")
+    }
+
+    pub fn slice_assign_from<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        user_slice: &RustUserSlice<'ctx>,
+        source: ListObject<'ctx>,
+    ) {
+        // Sanity check
+        assert!(ctx.unifier.unioned(self.item_type, source.item_type));
+
+        let user_slice_model = StructModel(UserSlice);
+        let puser_slice = user_slice_model.alloca(generator, ctx, "user_slice");
+        user_slice.write_to_user_slice(generator, ctx, puser_slice);
+
+        let itemsize = self.instance.model.get_type(generator, ctx.ctx).size_of();
+
+        call_nac3_list_slice_assign(
+            generator,
+            ctx,
+            self.get_opaque_list_ptr(generator, ctx),
+            source.instance.value,
+            itemsize,
+            user_slice,
+        );
+        todo!()
+    }
+}

nac3core/src/codegen/object/mod.rs

@@ -0,0 +1,608 @@
+use inkwell::{
+    values::{BasicValue, BasicValueEnum, FloatValue, IntValue},
+    FloatPredicate, IntPredicate,
+};
+use itertools::Itertools;
+use list::ListObject;
+use ndarray::{NDArrayObject, NDArrayOut};
+use range::RangeObject;
+use tuple::TupleObject;
+
+use crate::{
+    toplevel::helper::PrimDef,
+    typecheck::typedef::{Type, TypeEnum},
+};
+
+use super::{llvm_intrinsics, model::*, CodeGenContext, CodeGenerator};
+
+pub mod list;
+pub mod ndarray;
+pub mod range;
+pub mod tuple;
+
+/// Convenience function to crash the program when types of arguments are not supported.
+/// Used to be debugged with a stacktrace.
+fn unsupported_type<I>(ctx: &CodeGenContext<'_, '_>, tys: I) -> !
+where
+    I: IntoIterator<Item = Type>,
+{
+    unreachable!(
+        "unsupported types found '{}'",
+        tys.into_iter().map(|ty| format!("'{}'", ctx.unifier.stringify(ty))).join(", "),
+    )
+}
+
+#[derive(Debug, Clone, Copy)]
+pub enum FloorOrCeil {
+    Floor,
+    Ceil,
+}
+
+#[derive(Debug, Clone, Copy)]
+pub enum MinOrMax {
+    Min,
+    Max,
+}
+
+fn signed_ints(ctx: &CodeGenContext<'_, '_>) -> Vec<Type> {
+    vec![ctx.primitives.int32, ctx.primitives.int64]
+}
+
+fn unsigned_ints(ctx: &CodeGenContext<'_, '_>) -> Vec<Type> {
+    vec![ctx.primitives.uint32, ctx.primitives.uint64]
+}
+
+fn ints(ctx: &CodeGenContext<'_, '_>) -> Vec<Type> {
+    vec![ctx.primitives.int32, ctx.primitives.int64, ctx.primitives.uint32, ctx.primitives.uint64]
+}
+
+fn int_like(ctx: &CodeGenContext<'_, '_>) -> Vec<Type> {
+    vec![
+        ctx.primitives.bool,
+        ctx.primitives.int32,
+        ctx.primitives.int64,
+        ctx.primitives.uint32,
+        ctx.primitives.uint64,
+    ]
+}
+
+#[derive(Debug, Clone, Copy)]
+pub struct AnyObject<'ctx> {
+    pub ty: Type,
+    pub value: BasicValueEnum<'ctx>,
+}
+
+impl<'ctx> AnyObject<'ctx> {
+    /// Returns true if this object's type is a [`TypeEnum::TObj`] and has the object ID as `prim`.
+    pub fn is_obj(&self, ctx: &mut CodeGenContext<'ctx, '_>, prim: PrimDef) -> bool {
+        match &*ctx.unifier.get_ty(self.ty) {
+            TypeEnum::TObj { obj_id, .. } => *obj_id == prim.id(),
+            _ => false,
+        }
+    }
+
+    /// Returns true if this object's type is a [`TypeEnum::TTuple`]
+    pub fn is_tuple(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> bool {
+        matches!(&*ctx.unifier.get_ty(self.ty), TypeEnum::TTuple { .. })
+    }
+
+    pub fn into_tuple() {}
+
+    pub fn is_int32(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> bool {
+        ctx.unifier.unioned(self.ty, ctx.primitives.int32)
+    }
+
+    pub fn into_int32(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> Int<'ctx, Int32> {
+        assert!(self.is_int32(ctx));
+        IntModel(Int32).believe_value(self.value.into_int_value())
+    }
+
+    pub fn is_uint32(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> bool {
+        ctx.unifier.unioned(self.ty, ctx.primitives.uint32)
+    }
+
+    pub fn is_int64(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> bool {
+        ctx.unifier.unioned(self.ty, ctx.primitives.int64)
+    }
+
+    pub fn is_uint64(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> bool {
+        ctx.unifier.unioned(self.ty, ctx.primitives.uint64)
+    }
+
+    pub fn is_bool(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> bool {
+        ctx.unifier.unioned(self.ty, ctx.primitives.bool)
+    }
+
+    /// Returns true if the object type is `bool`, `int32`, `int64`, `uint32`, or `uint64`.
+    pub fn is_int_like(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> bool {
+        ctx.unifier.unioned_any(self.ty, int_like(ctx))
+    }
+
+    /// Returns true if the object type is `int32`, `int64`.
+    pub fn is_signed_int(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> bool {
+        ctx.unifier.unioned_any(self.ty, signed_ints(ctx))
+    }
+
+    /// Returns true if the object type is `uint32`, `uint64`.
+    pub fn is_unsigned_int(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> bool {
+        ctx.unifier.unioned_any(self.ty, unsigned_ints(ctx))
+    }
+
+    pub fn into_int(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> IntValue<'ctx> {
+        assert!(self.is_int_like(ctx));
+        self.value.into_int_value()
+    }
+
+    pub fn is_float(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> bool {
+        self.is_obj(ctx, PrimDef::Float)
+    }
+
+    pub fn into_float(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> Float<'ctx, Float64> {
+        assert!(self.is_float(ctx));
+        FloatModel(Float64).believe_value(self.value.into_float_value())
+    }
+
+    pub fn is_ndarray(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> bool {
+        self.is_obj(ctx, PrimDef::NDArray)
+    }
+
+    pub fn into_ndarray<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> NDArrayObject<'ctx> {
+        NDArrayObject::from_object(generator, ctx, *self)
+    }
+
+    /// Create an object from a boolean from an i1.
+    ///
+    /// NOTE: In NAC3, booleans are i8. This function does converts the input i1 to an i8.
+    pub fn from_bool(ctx: &mut CodeGenContext<'ctx, '_>, n: Int<'ctx, Bool>) -> AnyObject<'ctx> {
+        let llvm_i8 = ctx.ctx.i8_type();
+        let value = ctx.builder.build_int_z_extend(n.value, llvm_i8, "bool").unwrap();
+        AnyObject { value: value.as_basic_value_enum(), ty: ctx.primitives.bool }
+    }
+
+    /// Helper function to compare two scalars.
+    ///
+    /// Only int-to-int and float-to-float comparisons are allowed.
+    ///
+    /// Panic otherwise.
+    pub fn compare_int_or_float_by_predicate<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        lhs: AnyObject<'ctx>,
+        rhs: AnyObject<'ctx>,
+        int_predicate: IntPredicate,
+        float_predicate: FloatPredicate,
+        name: &str,
+    ) -> Int<'ctx, Bool> {
+        assert!(ctx.unifier.unioned(lhs.ty, rhs.ty), "lhs and rhs type should be the same");
+
+        let bool_model = IntModel(Bool);
+
+        let common_ty = lhs.ty;
+        let result = if lhs.is_float(ctx) {
+            let lhs = lhs.into_float(ctx);
+            let rhs = rhs.into_float(ctx);
+            ctx.builder.build_float_compare(float_predicate, lhs.value, rhs.value, name).unwrap()
+        } else if ctx.unifier.unioned_any(common_ty, int_like(ctx)) {
+            let lhs = lhs.into_int(ctx);
+            let rhs = rhs.into_int(ctx);
+            ctx.builder.build_int_compare(int_predicate, lhs, rhs, name).unwrap()
+        } else {
+            unsupported_type(ctx, [lhs.ty, rhs.ty]);
+        };
+
+        bool_model.check_value(generator, ctx.ctx, result).unwrap()
+    }
+
+    /// Helper function for `int32()`, `int64()`, `uint32()`, and `uint64()`.
+    ///
+    /// TODO: Document me
+    fn cast_to_int_conversion<'a, G, HandleFloatFn>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, 'a>,
+        ret_int_ty: Type,
+        handle_float: HandleFloatFn,
+    ) -> AnyObject<'ctx>
+    where
+        G: CodeGenerator + ?Sized,
+        HandleFloatFn:
+            FnOnce(&mut G, &mut CodeGenContext<'ctx, 'a>, FloatValue<'ctx>) -> IntValue<'ctx>,
+    {
+        let ret_int_ty_llvm = ctx.get_llvm_type(generator, ret_int_ty).into_int_type();
+
+        let result = if self.is_float(ctx) {
+            // Handle float to int
+            let n = self.into_float(ctx);
+            handle_float(generator, ctx, n.value)
+        } else if self.is_int_like(ctx) {
+            // Handle int to a new int type
+            let n = self.into_int(ctx);
+            if n.get_type().get_bit_width() <= ret_int_ty_llvm.get_bit_width() {
+                ctx.builder.build_int_z_extend(n, ret_int_ty_llvm, "zext").unwrap()
+            } else {
+                ctx.builder.build_int_truncate(n, ret_int_ty_llvm, "trunc").unwrap()
+            }
+        } else {
+            unsupported_type(ctx, [self.ty]);
+        };
+
+        assert_eq!(ret_int_ty_llvm.get_bit_width(), result.get_type().get_bit_width()); // Sanity check
+        AnyObject { value: result.into(), ty: ret_int_ty }
+    }
+
+    /// Call `int32()` on this object.
+    #[must_use]
+    pub fn call_int32<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> AnyObject<'ctx> {
+        self.cast_to_int_conversion(
+            generator,
+            ctx,
+            ctx.primitives.int32,
+            |_generator, ctx, input| {
+                let n =
+                    ctx.builder.build_float_to_signed_int(input, ctx.ctx.i64_type(), "").unwrap();
+                ctx.builder.build_int_truncate(n, ctx.ctx.i32_type(), "conv").unwrap()
+            },
+        )
+    }
+
+    /// Call `int64()` on this object.
+    #[must_use]
+    pub fn call_int64<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> AnyObject<'ctx> {
+        self.cast_to_int_conversion(
+            generator,
+            ctx,
+            ctx.primitives.int64,
+            |_generator, ctx, input| {
+                ctx.builder.build_float_to_signed_int(input, ctx.ctx.i64_type(), "").unwrap()
+            },
+        )
+    }
+
+    /// Call `uint32()` on this object.
+    #[must_use]
+    pub fn call_uint32<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> AnyObject<'ctx> {
+        self.cast_to_int_conversion(generator, ctx, ctx.primitives.uint32, |_generator, ctx, n| {
+            let n_gez = ctx
+                .builder
+                .build_float_compare(FloatPredicate::OGE, n, n.get_type().const_zero(), "")
+                .unwrap();
+
+            let to_int32 =
+                ctx.builder.build_float_to_signed_int(n, ctx.ctx.i32_type(), "").unwrap();
+            let to_uint64 =
+                ctx.builder.build_float_to_unsigned_int(n, ctx.ctx.i64_type(), "").unwrap();
+
+            ctx.builder
+                .build_select(
+                    n_gez,
+                    ctx.builder.build_int_truncate(to_uint64, ctx.ctx.i32_type(), "").unwrap(),
+                    to_int32,
+                    "conv",
+                )
+                .unwrap()
+                .into_int_value()
+        })
+    }
+
+    /// Call `uint64()` on this object.
+    #[must_use]
+    pub fn call_uint64<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> AnyObject<'ctx> {
+        self.cast_to_int_conversion(generator, ctx, ctx.primitives.uint64, |_generator, ctx, n| {
+            let val_gez = ctx
+                .builder
+                .build_float_compare(FloatPredicate::OGE, n, n.get_type().const_zero(), "")
+                .unwrap();
+
+            let to_int64 =
+                ctx.builder.build_float_to_signed_int(n, ctx.ctx.i64_type(), "").unwrap();
+            let to_uint64 =
+                ctx.builder.build_float_to_unsigned_int(n, ctx.ctx.i64_type(), "").unwrap();
+
+            ctx.builder.build_select(val_gez, to_uint64, to_int64, "conv").unwrap().into_int_value()
+        })
+    }
+
+    // Get the `len()` of this object.
+    #[must_use]
+    pub fn call_len<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> AnyObject<'ctx> {
+        // TODO: Switch to returning SizeT
+        let result = match &*ctx.unifier.get_ty_immutable(self.ty) {
+            TypeEnum::TTuple { .. } => {
+                let tuple = TupleObject::from_object(ctx, *self);
+                tuple.len(generator, ctx).truncate(generator, ctx, Int32, "tuple_len_32")
+            }
+            TypeEnum::TObj { obj_id, .. }
+                if *obj_id == ctx.primitives.range.obj_id(&ctx.unifier).unwrap() =>
+            {
+                let range = RangeObject::from_object(generator, ctx, *self);
+                range.len(generator, ctx)
+            }
+            TypeEnum::TObj { obj_id, .. }
+                if *obj_id == ctx.primitives.list.obj_id(&ctx.unifier).unwrap() =>
+            {
+                let list = ListObject::from_object(generator, ctx, *self);
+                list.len(generator, ctx).truncate(generator, ctx, Int32, "list_len_i32")
+            }
+            TypeEnum::TObj { obj_id, .. }
+                if *obj_id == ctx.primitives.ndarray.obj_id(&ctx.unifier).unwrap() =>
+            {
+                let ndarray = NDArrayObject::from_object(generator, ctx, *self);
+                ndarray.len(generator, ctx).truncate(generator, ctx, Int32, "ndarray_len_i32")
+            }
+            _ => unreachable!(),
+        };
+
+        AnyObject { ty: ctx.primitives.int32, value: result.value.as_basic_value_enum() }
+    }
+
+    /// Like [`AnyObject::call_bool`] but this returns an `Int<'ctx, Bool>` instead of an object.
+    pub fn bool(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> Int<'ctx, Bool> {
+        let bool_model = IntModel(Bool);
+        if self.is_int_like(ctx) {
+            let n = self.into_int(ctx);
+            let n = ctx
+                .builder
+                .build_int_compare(inkwell::IntPredicate::NE, n, n.get_type().const_zero(), "bool")
+                .unwrap();
+            bool_model.believe_value(n)
+        } else if self.is_float(ctx) {
+            let n = self.value.into_float_value();
+            let n = ctx
+                .builder
+                .build_float_compare(FloatPredicate::UNE, n, n.get_type().const_zero(), "bool")
+                .unwrap();
+            bool_model.believe_value(n)
+        } else {
+            unsupported_type(ctx, [self.ty])
+        }
+    }
+
+    /// Call `bool()` on this object.
+    #[must_use]
+    pub fn call_bool(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> AnyObject<'ctx> {
+        let n = self.bool(ctx);
+        AnyObject::from_bool(ctx, n)
+    }
+
+    /// Call `float()` on this object.
+    #[must_use]
+    pub fn call_float(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> AnyObject<'ctx> {
+        let f64_model = FloatModel(Float64);
+        let llvm_f64 = ctx.ctx.f64_type();
+
+        let result = if self.is_float(ctx) {
+            self.into_float(ctx)
+        } else if self.is_signed_int(ctx) || self.is_bool(ctx) {
+            let n = self.into_int(ctx);
+            let n = ctx.builder.build_signed_int_to_float(n, llvm_f64, "sitofp").unwrap();
+            f64_model.believe_value(n)
+        } else if self.is_unsigned_int(ctx) {
+            let n = self.into_int(ctx);
+            let n = ctx.builder.build_unsigned_int_to_float(n, llvm_f64, "uitofp").unwrap();
+            f64_model.believe_value(n)
+        } else {
+            unsupported_type(ctx, [self.ty]);
+        };
+
+        AnyObject { ty: ctx.primitives.float, value: result.value.as_basic_value_enum() }
+    }
+
+    // Call `abs()` on this object.
+    #[must_use]
+    pub fn call_abs<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Self {
+        if self.is_float(ctx) {
+            let n = self.value.into_float_value();
+            let n = llvm_intrinsics::call_float_fabs(ctx, n, Some("abs"));
+            AnyObject { value: n.into(), ty: ctx.primitives.float }
+        } else if self.is_unsigned_int(ctx) || self.is_signed_int(ctx) {
+            let is_poisoned = ctx.ctx.bool_type().const_zero(); // is_poisoned = false
+
+            let n = self.value.into_int_value();
+            let n = llvm_intrinsics::call_int_abs(ctx, n, is_poisoned, Some("abs"));
+            AnyObject { value: n.into(), ty: self.ty }
+        } else if self.is_ndarray(ctx) {
+            let ndarray = self.into_ndarray(generator, ctx);
+            ndarray
+                .map(
+                    generator,
+                    ctx,
+                    NDArrayOut::NewNDArray { dtype: ndarray.dtype },
+                    |generator, ctx, scalar| Ok(scalar.call_abs(generator, ctx)),
+                )
+                .unwrap()
+                .to_any_object(ctx)
+        } else {
+            unsupported_type(ctx, [self.ty])
+        }
+    }
+
+    // Call `round()` on this object.
+    //
+    // It is possible to specify which kind of int type to return.
+    #[must_use]
+    pub fn call_round<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        ret_int_ty: Type,
+    ) -> AnyObject<'ctx> {
+        let ret_int_ty_llvm = ctx.get_llvm_type(generator, ret_int_ty).into_int_type();
+
+        let result = if ctx.unifier.unioned(self.ty, ctx.primitives.float) {
+            let n = self.value.into_float_value();
+            let n = llvm_intrinsics::call_float_round(ctx, n, None);
+            ctx.builder.build_float_to_signed_int(n, ret_int_ty_llvm, "round").unwrap()
+        } else {
+            unsupported_type(ctx, [self.ty])
+        };
+        AnyObject { ty: ret_int_ty, value: result.as_basic_value_enum() }
+    }
+
+    /// Call `np_round()` on this object.
+    ///
+    /// NOTE: `np.round()` has different behaviors than `round()` when in comes to "tie" cases and return type.
+    #[must_use]
+    pub fn call_np_round<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> AnyObject<'ctx> {
+        if self.is_float(ctx) {
+            let n = self.into_float(ctx);
+            let n = llvm_intrinsics::call_float_rint(ctx, n.value, None);
+            AnyObject { ty: ctx.primitives.float, value: n.as_basic_value_enum() }
+        } else if self.is_ndarray(ctx) {
+            let ndarray = self.into_ndarray(generator, ctx);
+            ndarray
+                .map(
+                    generator,
+                    ctx,
+                    NDArrayOut::NewNDArray { dtype: ndarray.dtype },
+                    |generator, ctx, scalar| Ok(scalar.call_np_round(generator, ctx)),
+                )
+                .unwrap()
+                .to_any_object(ctx)
+        } else {
+            unsupported_type(ctx, [self.ty])
+        }
+    }
+
+    /// Call `min()` or `max()` on two objects.
+    #[must_use]
+    pub fn call_min_or_max(
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        kind: MinOrMax,
+        a: AnyObject<'ctx>,
+        b: AnyObject<'ctx>,
+    ) -> AnyObject<'ctx> {
+        if !ctx.unifier.unioned(a.ty, b.ty) {
+            unsupported_type(ctx, [a.ty, b.ty])
+        }
+
+        let common_ty = a.ty;
+
+        if a.is_float(ctx) {
+            let function = match kind {
+                MinOrMax::Min => llvm_intrinsics::call_float_minnum,
+                MinOrMax::Max => llvm_intrinsics::call_float_maxnum,
+            };
+
+            let a = a.into_float(ctx).value;
+            let b = b.into_float(ctx).value;
+            let result = function(ctx, a, b, None);
+            AnyObject { value: result.as_basic_value_enum(), ty: ctx.primitives.float }
+        } else if a.is_unsigned_int(ctx) || a.is_bool(ctx) {
+            // Treating bool has an unsigned int since that is convenient
+            let function = match kind {
+                MinOrMax::Min => llvm_intrinsics::call_int_umin,
+                MinOrMax::Max => llvm_intrinsics::call_int_umax,
+            };
+
+            let a = a.into_int(ctx);
+            let b = b.into_int(ctx);
+            let result = function(ctx, a, b, None);
+            AnyObject { value: result.as_basic_value_enum(), ty: common_ty }
+        } else if a.is_signed_int(ctx) {
+            let function = match kind {
+                MinOrMax::Min => llvm_intrinsics::call_int_smin,
+                MinOrMax::Max => llvm_intrinsics::call_int_smax,
+            };
+
+            let a = a.into_int(ctx);
+            let b = b.into_int(ctx);
+            let result = function(ctx, a, b, None);
+            AnyObject { value: result.as_basic_value_enum(), ty: common_ty }
+        } else {
+            unsupported_type(ctx, [common_ty])
+        }
+    }
+
+    /// Call `floor()` or `ceil()` on this object.
+    ///
+    /// It is possible to specify which kind of int type to return.
+    #[must_use]
+    pub fn call_floor_or_ceil<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        kind: FloorOrCeil,
+        ret_int_ty: Type,
+    ) -> Self {
+        let ret_int_ty_llvm = ctx.get_llvm_type(generator, ret_int_ty).into_int_type();
+
+        if self.is_float(ctx) {
+            let function = match kind {
+                FloorOrCeil::Floor => llvm_intrinsics::call_float_floor,
+                FloorOrCeil::Ceil => llvm_intrinsics::call_float_ceil,
+            };
+
+            let n = self.into_float(ctx).value;
+            let n = function(ctx, n, None);
+            let n = ctx.builder.build_float_to_signed_int(n, ret_int_ty_llvm, "").unwrap();
+            AnyObject { ty: ret_int_ty, value: n.as_basic_value_enum() }
+        } else {
+            unsupported_type(ctx, [self.ty])
+        }
+    }
+
+    /// Call `np_floor()` or `np_ceil()` on this object.
+    #[must_use]
+    pub fn call_np_floor_or_ceil<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        kind: FloorOrCeil,
+    ) -> Self {
+        // TODO:
+        if self.is_float(ctx) {
+            let function = match kind {
+                FloorOrCeil::Floor => llvm_intrinsics::call_float_floor,
+                FloorOrCeil::Ceil => llvm_intrinsics::call_float_ceil,
+            };
+            let n = self.into_float(ctx).value;
+            let n = function(ctx, n, None);
+            AnyObject { ty: ctx.primitives.float, value: n.as_basic_value_enum() }
+        } else if self.is_ndarray(ctx) {
+            let ndarray = self.into_ndarray(generator, ctx);
+            ndarray
+                .map(
+                    generator,
+                    ctx,
+                    NDArrayOut::NewNDArray { dtype: ctx.primitives.float },
+                    |generator, ctx, scalar| Ok(scalar.call_np_floor_or_ceil(generator, ctx, kind)),
+                )
+                .unwrap()
+                .to_any_object(ctx)
+        } else {
+            unsupported_type(ctx, [self.ty])
+        }
+    }
+}

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

@@ -0,0 +1,176 @@
+use super::NDArrayObject;
+use crate::{
+    codegen::{
+        irrt::{call_nac3_array_set_and_validate_list_shape, call_nac3_array_write_list_to_array},
+        model::*,
+        object::{list::ListObject, AnyObject},
+        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 from_np_array_list_copy_impl<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        list: ListObject<'ctx>,
+    ) -> Self {
+        let sizet_model = IntModel(SizeT);
+
+        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 = sizet_model.constant(generator, ctx.ctx, ndims_int);
+        let shape = sizet_model.array_alloca(generator, ctx, ndims.value, "shape");
+        call_nac3_array_set_and_validate_list_shape(generator, ctx, list_value, ndims, shape);
+
+        let ndarray = NDArrayObject::alloca(generator, ctx, dtype, ndims_int, "ndarray_from_list");
+        ndarray.copy_shape_from_array(generator, ctx, shape);
+        ndarray.create_data(generator, ctx);
+
+        // Copy all contents from the list.
+        call_nac3_array_write_list_to_array(generator, ctx, list_value, ndarray.instance);
+
+        ndarray
+    }
+
+    fn from_np_array_list_try_no_copy_impl<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        list: ListObject<'ctx>,
+    ) -> Self {
+        // np_array without copy 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`.
+
+        let sizet_model = IntModel(SizeT);
+        let zero = sizet_model.const_0(generator, ctx.ctx);
+
+        let (dtype, ndims) = get_list_object_dtype_and_ndims(ctx, list);
+        if ndims == 1 {
+            // `list` is not nested, does not need to copy.
+            let ndarray =
+                NDArrayObject::alloca(generator, ctx, dtype, 1, "ndarray_from_list_no_copy");
+
+            // Set data
+            let data = list.get_opaque_items_ptr(generator, ctx);
+            ndarray.instance.set(ctx, |f| f.data, data);
+
+            // Set shape
+            //   dim = list->len;
+            //   shape[0] = dim;
+            let shape = ndarray.instance.get(generator, ctx, |f| f.shape, "shape");
+            let dim = list.instance.get(generator, ctx, |f| f.len, "dim");
+            shape.offset(generator, ctx, zero.value, "pdim").store(ctx, dim);
+
+            // Set strides, the `data` is contiguous
+            ndarray.update_strides_by_shape(generator, ctx);
+
+            // Done
+            ndarray
+        } else {
+            // `list` is nested, it is impossible to not copy.
+            NDArrayObject::from_np_array_list_copy_impl(generator, ctx, list)
+        }
+    }
+
+    fn from_np_array_list_impl<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        list: ListObject<'ctx>,
+        copy: Int<'ctx, 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::from_np_array_list_copy_impl(generator, ctx, list);
+                Ok(Some(ndarray.instance.value))
+            },
+            |generator, ctx| {
+                let ndarray =
+                    NDArrayObject::from_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 from_np_array_ndarray_impl<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        ndarray: NDArrayObject<'ctx>,
+        copy: Int<'ctx, 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, "np_array_copied_ndarray"); // 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,
+        )
+    }
+
+    pub fn from_np_array<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        object: AnyObject<'ctx>,
+        copy: Int<'ctx, 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::from_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::from_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 +1,161 @@
+use itertools::Itertools;
+
+use crate::codegen::{
+    irrt::{
+        call_nac3_ndarray_broadcast_shapes, call_nac3_ndarray_broadcast_to,
+        call_nac3_ndarray_util_assert_shape_no_negative,
+    },
+    model::*,
+    CodeGenContext, CodeGenerator,
+};
+
+use super::NDArrayObject;
+
+/// Fields of [`ShapeEntry`]
+pub struct ShapeEntryFields<'ctx, F: FieldTraversal<'ctx>> {
+    pub ndims: F::Out<IntModel<SizeT>>,
+    pub shape: F::Out<PtrModel<IntModel<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: Ptr<'ctx, IntModel<SizeT>>,
+    ) -> Self {
+        let target_ndims_llvm = IntModel(SizeT).constant(generator, ctx.ctx, target_ndims);
+        call_nac3_ndarray_util_assert_shape_no_negative(
+            generator,
+            ctx,
+            target_ndims_llvm,
+            target_shape,
+        );
+
+        let broadcast_ndarray = NDArrayObject::alloca(
+            generator,
+            ctx,
+            self.dtype,
+            target_ndims,
+            "broadcast_ndarray_to_dst",
+        );
+        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: Ptr<'ctx, IntModel<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>>,
+}
+
+pub fn broadcast_shapes<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    in_entries: &[(Ptr<'ctx, IntModel<SizeT>>, u64)],
+    broadcast_ndims: u64,
+    broadcast_shape: Ptr<'ctx, IntModel<SizeT>>,
+) {
+    let sizet_model = IntModel(SizeT);
+    let shape_model = StructModel(ShapeEntry);
+
+    // Prepare input shape entries
+    let num_shape_entries =
+        sizet_model.constant(generator, ctx.ctx, u64::try_from(in_entries.len()).unwrap());
+    let shape_entries =
+        shape_model.array_alloca(generator, ctx, num_shape_entries.value, "shape_entries");
+    for (i, (in_shape, in_ndims)) in in_entries.iter().enumerate() {
+        let i = sizet_model.constant(generator, ctx.ctx, i as u64).value;
+        let pshape_entry = shape_entries.offset(generator, ctx, i, "shape_entry");
+
+        let in_ndims = sizet_model.constant(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 = sizet_model.constant(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> {
+    // TODO: DOCUMENT: Behaves like `np.broadcast()`, except returns results differently.
+    pub fn broadcast<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        ndarrays: &[Self],
+    ) -> BroadcastAllResult<'ctx> {
+        assert!(!ndarrays.is_empty());
+
+        let sizet_model = IntModel(SizeT);
+
+        // Infer the broadcast output ndims.
+        let broadcast_ndims_int = ndarrays.iter().map(|ndarray| ndarray.ndims).max().unwrap();
+
+        let broadcast_ndims = sizet_model.constant(generator, ctx.ctx, broadcast_ndims_int);
+        let broadcast_shape =
+            sizet_model.array_alloca(generator, ctx, broadcast_ndims.value, "broadcast_shape");
+
+        let shape_entries = ndarrays
+            .iter()
+            .map(|ndarray| {
+                (ndarray.instance.get(generator, ctx, |f| f.shape, "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/factory.rs

@@ -0,0 +1,238 @@
+use inkwell::{values::BasicValueEnum, IntPredicate};
+
+use super::NDArrayObject;
+use crate::{
+    codegen::{
+        irrt::call_nac3_ndarray_util_assert_shape_no_negative, model::*, object::AnyObject,
+        CodeGenContext, CodeGenerator,
+    },
+    typecheck::typedef::Type,
+};
+
+/// 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, "").value.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").value.into()
+    } else {
+        panic!("unrecognized dtype: {}", ctx.unifier.stringify(dtype));
+    }
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Create an ndarray like `np.empty`.
+    pub fn from_np_empty<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        ndims: u64,
+        shape: Ptr<'ctx, IntModel<SizeT>>,
+    ) -> Self {
+        // Validate `shape`
+        // TODO: Should the caller be responsible for this instead?
+        let ndims_llvm = IntModel(SizeT).constant(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, "full_ndarray");
+        ndarray.copy_shape_from_array(generator, ctx, shape);
+        ndarray.create_data(generator, ctx);
+
+        ndarray
+    }
+
+    /// Create an ndarray like `np.full`.
+    pub fn from_np_full<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        ndims: u64,
+        shape: Ptr<'ctx, IntModel<SizeT>>,
+        fill_value: AnyObject<'ctx>,
+    ) -> Self {
+        // Sanity check on `fill_value`'s dtype.
+        assert!(ctx.unifier.unioned(dtype, fill_value.ty));
+
+        let ndarray = NDArrayObject::from_np_empty(generator, ctx, dtype, ndims, shape);
+        ndarray.fill(generator, ctx, fill_value);
+        ndarray
+    }
+
+    /// Create an ndarray like `np.zero`.
+    pub fn from_np_zero<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        ndims: u64,
+        shape: Ptr<'ctx, IntModel<SizeT>>,
+    ) -> Self {
+        let fill_value = ndarray_zero_value(generator, ctx, dtype);
+        let fill_value = AnyObject { value: fill_value, ty: dtype };
+        NDArrayObject::from_np_full(generator, ctx, dtype, ndims, shape, fill_value)
+    }
+
+    /// Create an ndarray like `np.ones`.
+    pub fn from_np_ones<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        ndims: u64,
+        shape: Ptr<'ctx, IntModel<SizeT>>,
+    ) -> Self {
+        let fill_value = ndarray_one_value(generator, ctx, dtype);
+        let fill_value = AnyObject { value: fill_value, ty: dtype };
+        NDArrayObject::from_np_full(generator, ctx, dtype, ndims, shape, fill_value)
+    }
+
+    /// Create an ndarray like `np.arange`.
+    /// The returned ndarray's `dtype` is always `float`
+    pub fn from_np_arange<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        length: Int<'ctx, SizeT>,
+    ) -> Self {
+        let ndarray = NDArrayObject::alloca(
+            generator,
+            ctx,
+            ctx.primitives.float,
+            1, // ndims = 1
+            "arange_ndarray",
+        );
+
+        // `ndarray.shape[0] = length`
+        ndarray
+            .instance
+            .get(generator, ctx, |f| f.shape, "shape")
+            .offset_const(generator, ctx, 0, "dim")
+            .store(ctx, length);
+
+        // Create data and set elements
+        ndarray.create_data(generator, ctx);
+        ndarray
+            .foreach(generator, ctx, |generator, ctx, _hooks, nditer| {
+                // Get the index of the current element, convert that index to float, and write it.
+                // This is how we get [0.0, 1.0, 2.0, ...].
+                let index = nditer.get_index(generator, ctx);
+                let pelement = nditer.get_pointer(generator, ctx);
+
+                let val = ctx
+                    .builder
+                    .build_unsigned_int_to_float(index.value, ctx.ctx.f64_type(), "val")
+                    .unwrap();
+                ctx.builder.build_store(pelement, val).unwrap();
+                Ok(())
+            })
+            .unwrap();
+
+        ndarray
+    }
+
+    /// Create an ndarray like `np.eye`.
+    pub fn from_np_eye<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        nrows: Int<'ctx, SizeT>,
+        ncols: Int<'ctx, SizeT>,
+        offset: Int<'ctx, 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],
+            "eye_ndarray",
+        );
+
+        // 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()
+                    .offset_const(generator, ctx, 0, "")
+                    .load(generator, ctx, "row_i");
+                let col_i = nditer
+                    .get_indices()
+                    .offset_const(generator, ctx, 1, "")
+                    .load(generator, ctx, "col_i");
+
+                // Write to element
+                let be_one =
+                    row_i.add(ctx, offset, "").compare(ctx, IntPredicate::EQ, col_i, "write_one");
+                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 from_np_identity<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        size: Int<'ctx, SizeT>,
+    ) -> Self {
+        // Convenient implementation
+        let offset = IntModel(SizeT).const_0(generator, ctx.ctx);
+        NDArrayObject::from_np_eye(generator, ctx, dtype, size, size, offset)
+    }
+}

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

@@ -0,0 +1,128 @@
+use inkwell::{FloatPredicate, IntPredicate};
+
+use crate::codegen::{
+    model::*,
+    object::{AnyObject, MinOrMax},
+    stmt::gen_if_callback,
+    CodeGenContext, CodeGenerator,
+};
+
+use super::NDArrayObject;
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Helper function to implement NAC3's builtin `np_min()`, `np_max()`, `np_argmin()`, and `np_argmax()`.
+    ///
+    /// Generate LLVM IR to find the extremum and index of the **first** extremum value.
+    ///
+    /// Care has also been taken to make the error messages match that of NumPy.
+    fn min_max_argmin_argmax_helper<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        kind: MinOrMax,
+        on_empty_err_msg: &str,
+    ) -> (AnyObject<'ctx>, Int<'ctx, SizeT>) {
+        let sizet_model = IntModel(SizeT);
+        let dtype_llvm = ctx.get_llvm_type(generator, self.dtype);
+
+        // If the ndarray is empty, throw an error.
+        let is_empty = self.is_empty(generator, ctx);
+        ctx.make_assert(
+            generator,
+            is_empty.value,
+            "0:ValueError",
+            on_empty_err_msg,
+            [None, None, None],
+            ctx.current_loc,
+        );
+
+        // Setup and initialize the extremum to be the first element in the ndarray
+        let pextremum_index = sizet_model.alloca(generator, ctx, "extremum_index");
+        let pextremum = ctx.builder.build_alloca(dtype_llvm, "extremum").unwrap();
+
+        let zero = sizet_model.const_0(generator, ctx.ctx);
+        pextremum_index.store(ctx, zero);
+
+        let first_scalar = self.get_nth_scalar(generator, ctx, zero);
+        ctx.builder.build_store(pextremum, first_scalar.value).unwrap();
+
+        self.foreach(generator, ctx, |generator, ctx, _hooks, nditer| {
+            let old_extremum = ctx.builder.build_load(pextremum, "current_extremum").unwrap();
+            let old_extremum = AnyObject { ty: self.dtype, value: old_extremum };
+
+            let scalar = nditer.get_scalar(generator, ctx);
+            let new_extremum = AnyObject::call_min_or_max(ctx, kind, old_extremum, scalar);
+
+            gen_if_callback(
+                generator,
+                ctx,
+                |generator, ctx| {
+                    // Is new_extremum is more extreme than old_extremum?
+                    let cmp = AnyObject::compare_int_or_float_by_predicate(
+                        generator,
+                        ctx,
+                        new_extremum,
+                        old_extremum,
+                        IntPredicate::NE,
+                        FloatPredicate::ONE,
+                        "",
+                    );
+                    Ok(cmp.value)
+                },
+                |generator, ctx| {
+                    // Yes, update the extremum index
+                    let index = nditer.get_index(generator, ctx);
+                    pextremum_index.store(ctx, index);
+                    Ok(())
+                },
+                |_generator, _ctx| {
+                    // No, do nothing
+                    Ok(())
+                },
+            )
+        })
+        .unwrap();
+
+        // Finally return the extremum and extremum index.
+        let extremum_index = pextremum_index.load(generator, ctx, "extremum_index");
+
+        let extremum = ctx.builder.build_load(pextremum, "extremum_value").unwrap();
+        let extremum = AnyObject { ty: self.dtype, value: extremum };
+
+        (extremum, extremum_index)
+    }
+
+    /// Invoke NAC3's builtin `np_min()` or `np_max()`.
+    pub fn min_or_max<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        kind: MinOrMax,
+    ) -> AnyObject<'ctx> {
+        let on_empty_err_msg = format!(
+            "zero-size array to reduction operation {} which has no identity",
+            match kind {
+                MinOrMax::Min => "minimum",
+                MinOrMax::Max => "maximum",
+            }
+        );
+        self.min_max_argmin_argmax_helper(generator, ctx, kind, &on_empty_err_msg).0
+    }
+
+    /// Invoke NAC3's builtin `np_argmin()` or `np_argmax()`.
+    pub fn argmin_or_argmax<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        kind: MinOrMax,
+    ) -> Int<'ctx, SizeT> {
+        let on_empty_err_msg = format!(
+            "attempt to get {} of an empty sequence",
+            match kind {
+                MinOrMax::Min => "argmin",
+                MinOrMax::Max => "argmax",
+            }
+        );
+        self.min_max_argmin_argmax_helper(generator, ctx, kind, &on_empty_err_msg).1
+    }
+}

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

@@ -0,0 +1,321 @@
+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<IntModel<NDIndexType>>, // Defined to be uint8_t in IRRT
+    pub data: F::Out<PtrModel<IntModel<Byte>>>,
+}
+
+/// An IRRT representation fo 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 [`UserSlice`]
+#[derive(Debug, Clone)]
+pub struct UserSliceFields<'ctx, F: FieldTraversal<'ctx>> {
+    pub start_defined: F::Out<IntModel<Bool>>,
+    pub start: F::Out<IntModel<Int32>>,
+    pub stop_defined: F::Out<IntModel<Bool>>,
+    pub stop: F::Out<IntModel<Int32>>,
+    pub step_defined: F::Out<IntModel<Bool>>,
+    pub step: F::Out<IntModel<Int32>>,
+}
+
+/// An IRRT representation of a user slice.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct UserSlice;
+
+impl<'ctx> StructKind<'ctx> for UserSlice {
+    type Fields<F: FieldTraversal<'ctx>> = UserSliceFields<'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 [`UserSlice`].
+#[derive(Debug, Clone)]
+pub struct RustUserSlice<'ctx> {
+    pub start: Option<Int<'ctx, Int32>>,
+    pub stop: Option<Int<'ctx, Int32>>,
+    pub step: Option<Int<'ctx, Int32>>,
+}
+
+impl<'ctx> RustUserSlice<'ctx> {
+    /// Write the contents to an LLVM [`UserSlice`].
+    pub fn write_to_user_slice<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        dst_slice_ptr: Ptr<'ctx, StructModel<UserSlice>>,
+    ) {
+        let bool_model = IntModel(Bool);
+
+        let false_ = bool_model.constant(generator, ctx.ctx, 0);
+        let true_ = bool_model.constant(generator, ctx.ctx, 1);
+
+        // TODO: Code duplication. Probably okay...?
+
+        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 variant to store the content and type of an NDIndex in high level.
+#[derive(Debug, Clone)]
+pub enum RustNDIndex<'ctx> {
+    SingleElement(Int<'ctx, Int32>), // TODO: To be SizeT
+    Slice(RustUserSlice<'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: Ptr<'ctx, StructModel<NDIndex>>,
+    ) {
+        let ndindex_type_model = IntModel(NDIndexType::default());
+        let i32_model = IntModel(Int32);
+        let user_slice_model = StructModel(UserSlice);
+
+        // Set `dst_ndindex_ptr->type`
+        dst_ndindex_ptr
+            .gep(ctx, |f| f.type_)
+            .store(ctx, ndindex_type_model.constant(generator, ctx.ctx, self.get_type_id()));
+
+        // Set `dst_ndindex_ptr->data`
+        match self {
+            RustNDIndex::SingleElement(in_index) => {
+                let index_ptr = i32_model.alloca(generator, ctx, "index");
+                index_ptr.store(ctx, *in_index);
+
+                dst_ndindex_ptr
+                    .gep(ctx, |f| f.data)
+                    .store(ctx, index_ptr.pointer_cast(generator, ctx, IntModel(Byte), ""));
+            }
+            RustNDIndex::Slice(in_rust_slice) => {
+                let user_slice_ptr = user_slice_model.alloca(generator, ctx, "user_slice");
+                in_rust_slice.write_to_user_slice(generator, ctx, user_slice_ptr);
+
+                dst_ndindex_ptr
+                    .gep(ctx, |f| f.data)
+                    .store(ctx, user_slice_ptr.pointer_cast(generator, ctx, IntModel(Byte), ""));
+            }
+            RustNDIndex::NewAxis | RustNDIndex::Ellipsis => {}
+        }
+    }
+
+    /// Allocate an array of `NDIndex`es on the stack and return its stack pointer.
+    pub fn alloca_ndindexes<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        in_ndindexes: &[RustNDIndex<'ctx>],
+    ) -> (Int<'ctx, SizeT>, Ptr<'ctx, StructModel<NDIndex>>) {
+        let sizet_model = IntModel(SizeT);
+        let ndindex_model = StructModel(NDIndex);
+
+        let num_ndindexes = sizet_model.constant(generator, ctx.ctx, in_ndindexes.len() as u64);
+        let ndindexes =
+            ndindex_model.array_alloca(generator, ctx, num_ndindexes.value, "ndindexes");
+        for (i, in_ndindex) in in_ndindexes.iter().enumerate() {
+            let i = sizet_model.constant(generator, ctx.ctx, i as u64);
+            let pndindex = ndindexes.offset(generator, ctx, i.value, "");
+            in_ndindex.write_to_ndindex(generator, ctx, pndindex);
+        }
+
+        (num_ndindexes, ndindexes)
+    }
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Get the ndims [`Type`] after indexing with a given slice.
+    #[must_use]
+    pub fn deduce_ndims_after_indexing_with(&self, indexes: &[RustNDIndex<'ctx>]) -> u64 {
+        let mut ndims = self.ndims;
+        for index in indexes {
+            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 indexes 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, '_>,
+        indexes: &[RustNDIndex<'ctx>],
+        name: &str,
+    ) -> Self {
+        let dst_ndims = self.deduce_ndims_after_indexing_with(indexes);
+        let dst_ndarray = NDArrayObject::alloca(generator, ctx, self.dtype, dst_ndims, name);
+
+        let (num_indexes, indexes) = RustNDIndex::alloca_ndindexes(generator, ctx, indexes);
+        call_nac3_ndarray_index(
+            generator,
+            ctx,
+            num_indexes,
+            indexes,
+            self.instance,
+            dst_ndarray.instance,
+        );
+
+        dst_ndarray
+    }
+}
+
+pub mod util {
+    use itertools::Itertools;
+    use nac3parser::ast::{Constant, Expr, ExprKind};
+
+    use crate::{
+        codegen::{expr::gen_slice, model::*, CodeGenContext, CodeGenerator},
+        typecheck::typedef::{Type, TypeEnum},
+    };
+
+    use super::{RustNDIndex, RustUserSlice};
+
+    /// 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_ndindexes<'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
+        let i32_model = IntModel(Int32);
+
+        // 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_ndindexes: 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
+
+                // Helper function here to deduce code duplication
+                let (lower, upper, step) = gen_slice(generator, ctx, lower, upper, step)?;
+                RustNDIndex::Slice(RustUserSlice { start: lower, stop: upper, step })
+            } else if let ExprKind::Constant { value: Constant::Ellipsis, .. } = &index_expr.node {
+                // Handle '...'
+                RustNDIndex::Ellipsis
+            } else {
+                match &*ctx.unifier.get_ty(index_expr.custom.unwrap()) {
+                    TypeEnum::TObj { obj_id, .. }
+                        if *obj_id == ctx.primitives.option.obj_id(&ctx.unifier).unwrap() =>
+                    {
+                        // Handle `np.newaxis` / `None`
+                        RustNDIndex::NewAxis
+                    }
+                    _ => {
+                        // 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 = i32_model.check_value(generator, ctx.ctx, index).unwrap();
+
+                        RustNDIndex::SingleElement(index)
+                    }
+                }
+            };
+            rust_ndindexes.push(ndindex);
+        }
+        Ok(rust_ndindexes)
+    }
+}

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

@@ -0,0 +1,200 @@
+use itertools::Itertools;
+
+use crate::{
+    codegen::{
+        object::ndarray::{AnyObject, NDArrayObject},
+        stmt::gen_for_callback,
+        CodeGenContext, CodeGenerator,
+    },
+    typecheck::typedef::Type,
+};
+
+use super::{nditer::NDIterHandle, scalar::ScalarOrNDArray, NDArrayOut};
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// TODO: Document me. Has complex behavior.
+    /// and explain why `ret_dtype` has to be specified beforehand.
+    pub fn broadcasting_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>,
+            &[AnyObject<'ctx>],
+        ) -> Result<AnyObject<'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,
+                    "mapped_ndarray",
+                );
+                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.check_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)).collect_vec();
+
+                let result = mapping(generator, ctx, &in_scalars)?;
+                // Sanity check on result's ty
+                assert!(ctx.unifier.unioned(result.ty, out_ndarray.dtype));
+
+                let p = out_nditer.get_pointer(generator, ctx);
+                ctx.builder.build_store(p, result.value).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)
+    }
+
+    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>,
+            AnyObject<'ctx>,
+        ) -> Result<AnyObject<'ctx>, String>,
+    {
+        NDArrayObject::broadcasting_starmap(
+            generator,
+            ctx,
+            &[*self],
+            out,
+            |generator, ctx, scalars| mapping(generator, ctx, scalars[0]),
+        )
+    }
+}
+
+impl<'ctx> ScalarOrNDArray<'ctx> {
+    /// TODO: Document me. Has complex behavior.
+    /// and explain why `ret_dtype` has to be specified beforehand.
+    pub fn broadcasting_starmap<'a, G, MappingFn>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, 'a>,
+        inputs: &[Self],
+        ret_dtype: Type,
+        mapping: MappingFn,
+    ) -> Result<Self, String>
+    where
+        G: CodeGenerator + ?Sized,
+        MappingFn: FnOnce(
+            &mut G,
+            &mut CodeGenContext<'ctx, 'a>,
+            &[AnyObject<'ctx>],
+        ) -> Result<AnyObject<'ctx>, String>,
+    {
+        // Check if all inputs are AnyObjects
+        let all_scalars: Option<Vec<_>> = inputs.iter().map(AnyObject::try_from).try_collect().ok();
+
+        if let Some(scalars) = all_scalars {
+            let scalar = mapping(generator, ctx, &scalars)?;
+
+            // Sanity check on scalar's type
+            assert!(ctx.unifier.unioned(scalar.ty, ret_dtype));
+
+            Ok(ScalarOrNDArray::Scalar(scalar))
+        } else {
+            // Promote all input to ndarrays and map through them.
+            let inputs = inputs.iter().map(|input| input.as_ndarray(generator, ctx)).collect_vec();
+            let ndarray = NDArrayObject::broadcasting_starmap(
+                generator,
+                ctx,
+                &inputs,
+                NDArrayOut::NewNDArray { dtype: ret_dtype },
+                mapping,
+            )?;
+            Ok(ScalarOrNDArray::NDArray(ndarray))
+        }
+    }
+
+    pub fn map<'a, G, Mapping>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, 'a>,
+        ret_dtype: Type,
+        mapping: Mapping,
+    ) -> Result<Self, String>
+    where
+        G: CodeGenerator + ?Sized,
+        Mapping: FnOnce(
+            &mut G,
+            &mut CodeGenContext<'ctx, 'a>,
+            AnyObject<'ctx>,
+        ) -> Result<AnyObject<'ctx>, String>,
+    {
+        ScalarOrNDArray::broadcasting_starmap(
+            generator,
+            ctx,
+            &[*self],
+            ret_dtype,
+            |generator, ctx, scalars| mapping(generator, ctx, scalars[0]),
+        )
+    }
+}

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

@@ -0,0 +1,831 @@
+pub mod array;
+pub mod broadcast;
+pub mod factory;
+pub mod functions;
+pub mod indexing;
+pub mod mapping;
+pub mod nalgebra;
+pub mod nditer;
+pub mod product;
+pub mod scalar;
+pub mod shape_util;
+
+use crate::{
+    codegen::{
+        irrt::{
+            call_nac3_ndarray_copy_data, call_nac3_ndarray_get_nth_pelement,
+            call_nac3_ndarray_is_c_contiguous, call_nac3_ndarray_len, call_nac3_ndarray_nbytes,
+            call_nac3_ndarray_resolve_and_check_new_shape, call_nac3_ndarray_set_strides_by_shape,
+            call_nac3_ndarray_size, call_nac3_ndarray_transpose,
+            call_nac3_ndarray_util_assert_output_shape_same,
+        },
+        model::*,
+        stmt::{gen_for_callback, BreakContinueHooks},
+        structure::{NDArray, SimpleNDArray},
+        CodeGenContext, CodeGenerator,
+    },
+    toplevel::{
+        helper::{create_ndims, extract_ndims},
+        numpy::{make_ndarray_ty, unpack_ndarray_var_tys},
+    },
+    typecheck::typedef::Type,
+};
+use indexing::RustNDIndex;
+use inkwell::{
+    context::Context,
+    types::BasicType,
+    values::{BasicValue, PointerValue},
+    AddressSpace, IntPredicate,
+};
+use nditer::NDIterHandle;
+use scalar::ScalarOrNDArray;
+use util::call_memcpy_model;
+
+use super::{tuple::TupleObject, AnyObject};
+
+/// A NAC3 Python ndarray object.
+#[derive(Debug, Clone, Copy)]
+pub struct NDArrayObject<'ctx> {
+    pub dtype: Type,
+    pub ndims: u64,
+    pub instance: Ptr<'ctx, StructModel<NDArray>>,
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Create an [`NDArrayObject`] 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 {
+        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 pndarray_model = PtrModel(StructModel(NDArray));
+        let value = pndarray_model.check_value(generator, ctx.ctx, value).unwrap();
+        NDArrayObject { dtype, ndims, instance: value }
+    }
+
+    /// Forget that this is an ndarray and convert to an [`AnyObject`].
+    pub fn to_any_object(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> AnyObject<'ctx> {
+        let ty = self.get_ndarray_type(ctx);
+        AnyObject { value: self.instance.value.as_basic_value_enum(), ty }
+    }
+
+    /// Create a [`SimpleNDArray`] 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 [`SimpleNDArray`] and copy contents of this ndarray to there.
+    ///
+    /// If this ndarray is C-contiguous, contents of this ndarray will not be copied. The created [`SimpleNDArray`]
+    /// will have the same `data` field as this ndarray.
+    ///
+    /// The `item_model` sets the [`Model`] of the returned [`SimpleNDArray`]'s `Item` model, and should match the
+    /// `ctx.get_llvm_type()` of this ndarray's `dtype`. Otherwise this function panics.
+    pub fn make_simple_ndarray<G: CodeGenerator + ?Sized, Item: Model<'ctx>>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        item_model: Item,
+        name: &str,
+    ) -> Ptr<'ctx, StructModel<SimpleNDArray<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 simple_ndarray_model = StructModel(SimpleNDArray { 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 [`SimpleNDArray`].
+        let result = simple_ndarray_model.alloca(generator, ctx, name);
+
+        // Set ndims and shape.
+        let ndims = self.get_ndims(generator, ctx.ctx);
+        result.set(ctx, |f| f.ndims, ndims);
+
+        let shape = self.instance.get(generator, ctx, |f| f.shape, "shape");
+        result.set(ctx, |f| f.shape, shape);
+
+        // Set data, but we do things differently if this ndarray is contiguous.
+        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.
+        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`.
+        // TODO: Reimplement this? This method does give us the contiguous `data`, but
+        // this creates a few extra bytes of useless information because an entire NDArray
+        // is allocated. Though this is super convenient.
+        let data = self.make_copy(generator, ctx, "").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 [`SimpleNDArray`].
+    ///
+    /// This operation is super cheap. The newly created [`NDArray`] will not copy contents
+    /// from `simple_ndarray`, but only having its `data` and `shape` pointing to `simple_array`.
+    pub fn from_simple_ndarray<G: CodeGenerator + ?Sized, Item: Model<'ctx>>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        simple_ndarray: Ptr<'ctx, StructModel<SimpleNDArray<Item>>>,
+        dtype: Type,
+        ndims: u64,
+    ) -> Self {
+        // Sanity check on `dtype` and `simple_array`'s `Item` model.
+        let dtype_llvm = ctx.get_llvm_type(generator, dtype);
+        simple_ndarray.model.0 .0.item.check_type(generator, ctx.ctx, dtype_llvm).unwrap();
+
+        let byte_model = IntModel(Byte);
+
+        // TODO: Check if `ndims` is consistent with that in `simple_array`?
+
+        // Allocate the resulting ndarray.
+        let ndarray = NDArrayObject::alloca(generator, ctx, dtype, ndims, "from_simple_ndarray");
+
+        // Set data, shape by simply copying addresses.
+        let data = simple_ndarray
+            .get(generator, ctx, |f| f.data, "")
+            .pointer_cast(generator, ctx, byte_model, "data");
+        ndarray.instance.set(ctx, |f| f.data, data);
+
+        let shape = simple_ndarray.get(generator, ctx, |f| f.shape, "shape");
+        ndarray.instance.set(ctx, |f| f.shape, shape);
+
+        // Set strides. We know `data` is contiguous.
+        ndarray.update_strides_by_shape(generator, ctx);
+
+        ndarray
+    }
+
+    /// Get the typechecker ndarray type of this [`NDArrayObject`].
+    pub fn get_ndarray_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))
+    }
+
+    /// Get the `np.size()` of this ndarray.
+    pub fn size<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Int<'ctx, 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, '_>,
+    ) -> Int<'ctx, 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, '_>,
+    ) -> Int<'ctx, 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, '_>,
+    ) -> Int<'ctx, 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`.
+    ///
+    /// There is no out-of-bounds check.
+    pub fn get_nth_pointer<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        nth: Int<'ctx, SizeT>,
+        name: &str,
+    ) -> 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()), name)
+            .unwrap()
+    }
+
+    /// Get the n-th (0-based) scalar.
+    ///
+    /// There is no out-of-bounds check.
+    pub fn get_nth_scalar<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        nth: Int<'ctx, SizeT>,
+    ) -> AnyObject<'ctx> {
+        let p = self.get_nth_pointer(generator, ctx, nth, "value");
+        let value = ctx.builder.build_load(p, "value").unwrap();
+        AnyObject { ty: self.dtype, value }
+    }
+
+    /// Set the n-th (0-based) scalar.
+    ///
+    /// There is no out-of-bounds check.
+    pub fn set_nth_scalar<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        nth: Int<'ctx, SizeT>,
+        scalar: AnyObject<'ctx>,
+    ) {
+        // Sanity check on scalar's `dtype`
+        assert!(ctx.unifier.unioned(scalar.ty, self.dtype));
+
+        let pscalar = self.get_nth_pointer(generator, ctx, nth, "pscalar");
+        ctx.builder.build_store(pscalar, scalar.value).unwrap();
+    }
+
+    /// Call [`call_nac3_ndarray_set_strides_by_shape`] on this ndarray to update `strides`.
+    ///
+    /// Please refer to the IRRT implementation to see its purpose.
+    pub fn update_strides_by_shape<G: CodeGenerator + ?Sized>(
+        self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) {
+        call_nac3_ndarray_set_strides_by_shape(generator, ctx, self.instance);
+    }
+
+    /// 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);
+    }
+
+    /// 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,
+        name: &str,
+    ) -> Self {
+        let sizet_model = IntModel(SizeT);
+        let ndarray_model = StructModel(NDArray);
+        let ndarray_data_model = PtrModel(IntModel(Byte));
+
+        let pndarray = ndarray_model.alloca(generator, ctx, name);
+
+        let data = ndarray_data_model.nullptr(generator, ctx.ctx);
+        pndarray.set(ctx, |f| f.data, data);
+
+        let itemsize = ctx.get_llvm_type(generator, dtype).size_of().unwrap();
+        let itemsize =
+            sizet_model.s_extend_or_bit_cast(generator, ctx, itemsize, "alloca_itemsize");
+        pndarray.set(ctx, |f| f.itemsize, itemsize);
+
+        let ndims_val = sizet_model.constant(generator, ctx.ctx, ndims);
+        pndarray.set(ctx, |f| f.ndims, ndims_val);
+
+        let shape = sizet_model.array_alloca(generator, ctx, ndims_val.value, "alloca_shape");
+        pndarray.set(ctx, |f| f.shape, shape);
+
+        let strides = sizet_model.array_alloca(generator, ctx, ndims_val.value, "alloca_strides");
+        pndarray.set(ctx, |f| f.strides, strides);
+
+        NDArrayObject { dtype, ndims, instance: pndarray }
+    }
+
+    /// Convenience function.
+    /// Like [`NDArrayObject::alloca_uninitialized`] but directly takes the typechecker type of the ndarray.
+    pub fn alloca_ndarray_type<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        ndarray_ty: Type,
+        name: &str,
+    ) -> Self {
+        let (dtype, ndims) = unpack_ndarray_var_tys(&mut ctx.unifier, ndarray_ty);
+        let ndims = extract_ndims(&ctx.unifier, ndims);
+        Self::alloca(generator, ctx, dtype, ndims, name)
+    }
+
+    /// 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],
+        name: &str,
+    ) -> Self {
+        let sizet_model = IntModel(SizeT);
+
+        let ndarray = NDArrayObject::alloca(generator, ctx, dtype, shape.len() as u64, name);
+
+        // Write shape
+        let dst_shape = ndarray.instance.get(generator, ctx, |f| f.shape, "shape");
+        for (i, dim) in shape.iter().enumerate() {
+            let dim = sizet_model.constant(generator, ctx.ctx, *dim);
+            dst_shape.offset_const(generator, 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: &[Int<'ctx, SizeT>],
+        name: &str,
+    ) -> Self {
+        let ndarray = NDArrayObject::alloca(generator, ctx, dtype, shape.len() as u64, name);
+
+        // Write shape
+        let dst_shape = ndarray.instance.get(generator, ctx, |f| f.shape, "shape");
+        for (i, dim) in shape.iter().enumerate() {
+            dst_shape.offset_const(generator, ctx, i as u64, "").store(ctx, *dim);
+        }
+
+        ndarray
+    }
+
+    /// 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, '_>,
+        name: &str,
+    ) -> Self {
+        let clone = NDArrayObject::alloca(generator, ctx, self.dtype, self.ndims, name);
+
+        let shape = self.instance.gep(ctx, |f| f.shape).load(generator, ctx, "shape");
+        clone.copy_shape_from_array(generator, ctx, shape);
+        clone.create_data(generator, ctx);
+        clone.copy_data_from(generator, ctx, *self);
+        clone
+    }
+
+    /// Get this ndarray's `ndims` as an LLVM constant.
+    pub fn get_ndims<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> Int<'ctx, SizeT> {
+        let sizet_model = IntModel(SizeT);
+        sizet_model.constant(generator, ctx, self.ndims)
+    }
+
+    /// Get if this ndarray's `np.size` is `0` - containing no content.
+    pub fn is_empty<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Int<'ctx, Bool> {
+        let sizet_model = IntModel(SizeT);
+
+        let size = self.size(generator, ctx);
+        size.compare(ctx, IntPredicate::EQ, sizet_model.const_0(generator, ctx.ctx), "is_empty")
+    }
+
+    /// Return true if this ndarray is unsized - `ndims == 0` and only contains a scalar.
+    ///
+    /// This is a staticially known property of ndarrays. This is why it is returning
+    /// a Rust boolean instead of a [`BasicValue`].
+    #[must_use]
+    pub fn is_unsized(&self) -> bool {
+        self.ndims == 0
+    }
+
+    /// If this ndarray is unsized, return its sole value as a [`AnyObject`]. Otherwise, do nothing.
+    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 sizet_model = IntModel(SizeT);
+            let zero = sizet_model.const_0(generator, ctx.ctx);
+            ScalarOrNDArray::Scalar(self.get_nth_scalar(generator, ctx, zero))
+        } else {
+            ScalarOrNDArray::NDArray(*self)
+        }
+    }
+
+    /// 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 byte_model = IntModel(Byte);
+
+        let nbytes = self.nbytes(generator, ctx);
+
+        let data = byte_model.array_alloca(generator, ctx, nbytes.value, "data");
+        self.instance.set(ctx, |f| f.data, data);
+
+        self.update_strides_by_shape(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, '_>,
+        src_shape: Ptr<'ctx, IntModel<SizeT>>,
+    ) {
+        let dst_shape = self.instance.get(generator, ctx, |f| f.shape, "dst_shape");
+        let num_items = self.get_ndims(generator, ctx.ctx).value;
+        call_memcpy_model(generator, ctx, dst_shape, src_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, "src_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, '_>,
+        src_strides: Ptr<'ctx, IntModel<SizeT>>,
+    ) {
+        let dst_strides = self.instance.get(generator, ctx, |f| f.strides, "dst_strides");
+        let num_items = self.get_ndims(generator, ctx.ctx).value;
+        call_memcpy_model(generator, ctx, dst_strides, src_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, "src_strides");
+        self.copy_strides_from_array(generator, ctx, src_strides);
+    }
+
+    /// 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(())
+            },
+        )
+    }
+
+    /// 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 {
+            let mut indices = vec![];
+            for _ in self.ndims..ndmin {
+                indices.push(RustNDIndex::NewAxis);
+            }
+            indices.push(RustNDIndex::Ellipsis);
+            self.index(generator, ctx, &indices, "atleast_nd_ndarray")
+        } else {
+            *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, '_>,
+        scalar: AnyObject<'ctx>,
+    ) {
+        // Sanity check on scalar's type.
+        assert!(ctx.unifier.unioned(self.dtype, scalar.ty));
+
+        self.foreach(generator, ctx, |generator, ctx, _hooks, nditer| {
+            let p = nditer.get_pointer(generator, ctx);
+            ctx.builder.build_store(p, scalar.value).unwrap();
+            Ok(())
+        })
+        .unwrap();
+    }
+
+    /// 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: Ptr<'ctx, IntModel<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, "reshaped_ndarray");
+        dst_ndarray.copy_shape_from_array(generator, ctx, new_shape);
+
+        // Reolsve negative indices
+        let size = self.size(generator, ctx);
+        let dst_ndims = dst_ndarray.get_ndims(generator, ctx.ctx);
+        let dst_shape =
+            dst_ndarray.instance.get(generator, ctx, |f| f.shape, "reshaped_ndarray_shape");
+        call_nac3_ndarray_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.update_strides_by_shape(generator, ctx);
+        dst_ndarray.instance.set(
+            ctx,
+            |f| f.data,
+            self.instance.get(generator, ctx, |f| f.data, "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 flattened view of this ndarray, like `np.ravel()`.
+    ///
+    /// Uses [`NDArrayObject::reshape_or_copy`] under-the-hood so ndarray may or may not be copied.
+    #[must_use]
+    pub fn ravel_or_copy<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Self {
+        // Define models
+        let sizet_model = IntModel(SizeT);
+
+        let num0 = sizet_model.const_0(generator, ctx.ctx);
+        let num1 = sizet_model.const_1(generator, ctx.ctx);
+        let num_neg1 = sizet_model.const_all_1s(generator, ctx.ctx);
+
+        // Create `[-1]` and pass to `reshape_or_copy`.
+        let new_shape = sizet_model.array_alloca(generator, ctx, num1.value, "new_shape");
+        new_shape.offset(generator, ctx, num0.value, "").store(ctx, num_neg1);
+
+        self.reshape_or_copy(generator, ctx, 1, new_shape)
+    }
+
+    /// 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<Ptr<'ctx, IntModel<SizeT>>>,
+    ) -> Self {
+        // Define models
+        let sizet_model = IntModel(SizeT);
+
+        let transposed_ndarray =
+            NDArrayObject::alloca(generator, ctx, self.dtype, self.ndims, "transposed_ndarray");
+
+        let num_axes = self.get_ndims(generator, ctx.ctx);
+
+        // `axes = nullptr` if `axes` is unspecified.
+        let axes = axes.unwrap_or_else(|| PtrModel(sizet_model).nullptr(generator, ctx.ctx));
+
+        call_nac3_ndarray_transpose(
+            generator,
+            ctx,
+            self.instance,
+            transposed_ndarray.instance,
+            num_axes,
+            axes,
+        );
+
+        transposed_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 check_can_be_written_by_out<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        out_ndims: u64,
+        out_shape: Ptr<'ctx, IntModel<SizeT>>,
+    ) {
+        let sizet_model = IntModel(SizeT);
+
+        let ndarray_ndims = self.get_ndims(generator, ctx.ctx);
+        let ndarray_shape = self.instance.get(generator, ctx, |f| f.shape, "shape");
+
+        let output_ndims = sizet_model.constant(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,
+        );
+    }
+
+    /// 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: Don't return a tuple of int32s.
+
+        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, "")
+                .offset_const(generator, ctx, i, "")
+                .load(generator, ctx, "dim");
+            let dim = dim.truncate(generator, ctx, Int32, "dim"); // TODO: keep using SizeT
+
+            objects.push(AnyObject {
+                ty: ctx.primitives.int32,
+                value: dim.value.as_basic_value_enum(),
+            });
+        }
+
+        TupleObject::create(generator, ctx, objects, "shape")
+    }
+
+    /// 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: Don't return a tuple of int32s.
+
+        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, "")
+                .offset_const(generator, ctx, i, "")
+                .load(generator, ctx, "dim");
+            let dim = dim.truncate(generator, ctx, Int32, "dim"); // TODO: keep using SizeT
+
+            objects.push(AnyObject {
+                ty: ctx.primitives.int32,
+                value: dim.value.as_basic_value_enum(),
+            });
+        }
+
+        TupleObject::create(generator, ctx, objects, "strides")
+    }
+}
+
+/// TODO: Document me
+#[derive(Debug, Clone, Copy)]
+pub enum NDArrayOut<'ctx> {
+    NewNDArray { dtype: Type },
+    WriteToNDArray { ndarray: NDArrayObject<'ctx> },
+}

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

@@ -0,0 +1,53 @@
+use inkwell::values::{BasicValue, BasicValueEnum};
+
+use crate::codegen::{model::*, structure::SimpleNDArray, CodeGenContext, CodeGenerator};
+
+use super::NDArrayObject;
+
+pub fn perform_nalgebra_call<'ctx, 'a, const NUM_INPUTS: usize, const NUM_OUTPUTS: usize, G, F>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, 'a>,
+    inputs: [NDArrayObject<'ctx>; NUM_INPUTS],
+    output_ndims: [u64; NUM_OUTPUTS],
+    invoke_function: F,
+) -> [NDArrayObject<'ctx>; NUM_OUTPUTS]
+where
+    G: CodeGenerator + ?Sized,
+    F: FnOnce(
+        &mut CodeGenContext<'ctx, 'a>,
+        [BasicValueEnum<'ctx>; NUM_INPUTS],
+        [BasicValueEnum<'ctx>; NUM_OUTPUTS],
+    ),
+{
+    // TODO: Allow stacked inputs. See NumPy docs.
+
+    let f64_model = FloatModel(Float64);
+    let simple_ndarray_model = StructModel(SimpleNDArray { item: f64_model });
+
+    // Prepare inputs & outputs and invoke
+    let inputs = inputs.map(|input| {
+        // Sanity check. Typechecker ensures this.
+        assert!(ctx.unifier.unioned(input.dtype, ctx.primitives.float));
+
+        input
+            .make_simple_ndarray(generator, ctx, FloatModel(Float64), "nalgebra_input")
+            .value
+            .as_basic_value_enum()
+    });
+    let outputs = [simple_ndarray_model.alloca(generator, ctx, "nalgebra_output"); NUM_OUTPUTS];
+    invoke_function(ctx, inputs, outputs.map(|output| output.value.as_basic_value_enum()));
+
+    // Turn the outputs into strided NDArrays
+    let mut output_i = 0;
+    outputs.map(|output| {
+        let out = NDArrayObject::from_simple_ndarray(
+            generator,
+            ctx,
+            output,
+            ctx.primitives.float,
+            output_ndims[output_i],
+        );
+        output_i += 1;
+        out
+    })
+}

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

@@ -0,0 +1,88 @@
+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::AnyObject,
+    structure::NDIter,
+    CodeGenContext, CodeGenerator,
+};
+
+use super::NDArrayObject;
+
+#[derive(Debug, Clone)]
+pub struct NDIterHandle<'ctx> {
+    ndarray: NDArrayObject<'ctx>,
+    instance: Ptr<'ctx, StructModel<NDIter>>,
+    indices: Ptr<'ctx, IntModel<SizeT>>,
+}
+
+impl<'ctx> NDIterHandle<'ctx> {
+    pub fn new<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        ndarray: NDArrayObject<'ctx>,
+    ) -> Self {
+        let nditer = StructModel(NDIter).alloca(generator, ctx, "nditer");
+        let ndims = ndarray.get_ndims(generator, ctx.ctx);
+        let indices = IntModel(SizeT).array_alloca(generator, ctx, ndims.value, "indices");
+        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, '_>,
+    ) -> Int<'ctx, 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);
+    }
+
+    #[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, "element");
+        ctx.builder
+            .build_pointer_cast(p.value, elem_ty.ptr_type(AddressSpace::default()), "element")
+            .unwrap()
+    }
+
+    #[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 }
+    }
+
+    #[must_use]
+    pub fn get_index<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Int<'ctx, SizeT> {
+        self.instance.get(generator, ctx, |f| f.nth, "index")
+    }
+
+    #[must_use]
+    pub fn get_indices(&self) -> Ptr<'ctx, IntModel<SizeT>> {
+        self.indices
+    }
+}

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

@@ -0,0 +1,159 @@
+use std::cmp::max;
+
+use crate::codegen::{
+    irrt::{
+        call_nac3_ndarray_float64_matmul_at_least_2d, call_nac3_ndarray_matmul_calculate_shapes,
+    },
+    model::*,
+    object::ndarray::indexing::RustNDIndex,
+    CodeGenContext, CodeGenerator,
+};
+
+use super::{NDArrayObject, NDArrayOut};
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// TODO: Document me
+    fn matmul_helper<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        a: Self,
+        b: Self,
+    ) -> Self {
+        assert!(a.ndims >= 2);
+        assert!(b.ndims >= 2);
+
+        assert!(ctx.unifier.unioned(ctx.primitives.float, a.dtype));
+        assert!(ctx.unifier.unioned(ctx.primitives.float, b.dtype));
+
+        let sizet_model = IntModel(SizeT);
+
+        let final_ndims_int = max(a.ndims, b.ndims);
+
+        let a_ndims = a.get_ndims(generator, ctx.ctx);
+        let a_shape = a.instance.get(generator, ctx, |f| f.shape, "a_shape");
+        let b_ndims = b.get_ndims(generator, ctx.ctx);
+        let b_shape = b.instance.get(generator, ctx, |f| f.shape, "b_shape");
+        let final_ndims = sizet_model.constant(generator, ctx.ctx, final_ndims_int);
+        let new_a_shape =
+            sizet_model.array_alloca(generator, ctx, final_ndims.value, "new_a_shape");
+        let new_b_shape =
+            sizet_model.array_alloca(generator, ctx, final_ndims.value, "new_b_shape");
+        let dst_shape = sizet_model.array_alloca(generator, ctx, final_ndims.value, "dst_shape");
+
+        call_nac3_ndarray_matmul_calculate_shapes(
+            generator,
+            ctx,
+            a_ndims,
+            a_shape,
+            b_ndims,
+            b_shape,
+            final_ndims,
+            new_a_shape,
+            new_b_shape,
+            dst_shape,
+        );
+
+        let new_a = a.broadcast_to(generator, ctx, final_ndims_int, new_a_shape);
+        let new_b = b.broadcast_to(generator, ctx, final_ndims_int, new_b_shape);
+        let dst = NDArrayObject::alloca(
+            generator,
+            ctx,
+            ctx.primitives.float,
+            final_ndims_int,
+            "matmul_result",
+        );
+        dst.copy_shape_from_array(generator, ctx, dst_shape);
+        dst.create_data(generator, ctx);
+
+        call_nac3_ndarray_float64_matmul_at_least_2d(
+            generator,
+            ctx,
+            new_a.instance,
+            new_b.instance,
+            dst.instance,
+        );
+
+        dst
+    }
+
+    /// 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`].
+    pub fn matmul<G: CodeGenerator + ?Sized>(
+        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 indexes 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], "new_a")
+        } else {
+            a
+        };
+
+        let new_b = if b.ndims == 1 {
+            // Append 1 to its dimensions
+            b.index(generator, ctx, &[RustNDIndex::Ellipsis, RustNDIndex::NewAxis], "new_a")
+        } else {
+            b
+        };
+
+        // NOTE: `result` will always be a newly allocated ndarray.
+        // Current implementation cannot do in-place matrix muliplication.
+        let mut result = NDArrayObject::matmul_helper(generator, ctx, new_a, new_b);
+
+        let i32_model = IntModel(Int32); // TODO: Upgrade to SizeT
+        let zero = i32_model.const_0(generator, ctx.ctx);
+
+        if a.ndims == 1 {
+            // Remove the prepended 1
+            result = result.index(
+                generator,
+                ctx,
+                &[RustNDIndex::SingleElement(zero), RustNDIndex::Ellipsis],
+                "result_no_prepend_1",
+            );
+        }
+
+        if b.ndims == 1 {
+            // Remove the appended 1
+            result = result.index(
+                generator,
+                ctx,
+                &[RustNDIndex::Ellipsis, RustNDIndex::SingleElement(zero)],
+                "result_no_append_1",
+            );
+        }
+
+        match out {
+            NDArrayOut::NewNDArray { dtype } => {
+                // We don't support auto-casting right now, nor anything other than float64.
+                // Force the output dtype to be float64.
+                assert!(ctx.unifier.unioned(ctx.primitives.float, dtype));
+                result
+            }
+            NDArrayOut::WriteToNDArray { ndarray: out_ndarray } => {
+                // TODO: It is possible to check the shapes before computing the matmul to save resources.
+                let result_shape = result.instance.get(generator, ctx, |f| f.shape, "result_shape");
+                out_ndarray.check_can_be_written_by_out(generator, ctx, result.ndims, result_shape);
+
+                // TODO: We can just set `out_ndarray.data` to `result.data`. Should we?
+                out_ndarray.copy_data_from(generator, ctx, result);
+                out_ndarray
+            }
+        }
+    }
+}

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

@@ -0,0 +1,131 @@
+use inkwell::values::{BasicValue, BasicValueEnum};
+
+use crate::{
+    codegen::{model::*, object::AnyObject, CodeGenContext, CodeGenerator},
+    typecheck::typedef::{Type, TypeEnum},
+};
+
+use super::NDArrayObject;
+
+impl<'ctx> AnyObject<'ctx> {
+    /// Promote this scalar to an unsized ndarray (like doing `np.asarray`).
+    ///
+    /// The scalar value is allocated onto the stack, and the ndarray's `data` will point to that
+    /// allocated value.
+    pub fn as_ndarray<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> NDArrayObject<'ctx> {
+        let pbyte_model = PtrModel(IntModel(Byte));
+
+        // We have to put the value on the stack to get a data pointer.
+        let data = ctx.builder.build_alloca(self.value.get_type(), "as_ndarray_scalar").unwrap();
+        ctx.builder.build_store(data, self.value).unwrap();
+        let data = pbyte_model.pointer_cast(generator, ctx, data, "data");
+
+        let ndarray = NDArrayObject::alloca(generator, ctx, self.ty, 0, "scalar_ndarray");
+        ndarray.instance.set(ctx, |f| f.data, data);
+        ndarray
+    }
+}
+
+/// A convenience enum for implementing scalar/ndarray agnostic utilities.
+#[derive(Debug, Clone, Copy)]
+pub enum ScalarOrNDArray<'ctx> {
+    Scalar(AnyObject<'ctx>),
+    NDArray(NDArrayObject<'ctx>),
+}
+
+impl<'ctx> ScalarOrNDArray<'ctx> {
+    /// 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(),
+        }
+    }
+
+    #[must_use]
+    pub fn into_scalar(&self) -> AnyObject<'ctx> {
+        match self {
+            ScalarOrNDArray::NDArray(_ndarray) => panic!("Got NDArray"),
+            ScalarOrNDArray::Scalar(scalar) => *scalar,
+        }
+    }
+
+    #[must_use]
+    pub fn into_ndarray(&self) -> NDArrayObject<'ctx> {
+        match self {
+            ScalarOrNDArray::NDArray(ndarray) => *ndarray,
+            ScalarOrNDArray::Scalar(_scalar) => panic!("Got Scalar"),
+        }
+    }
+
+    /// Convert this [`ScalarOrNDArray`] to an ndarray - behaves like `np.asarray`.
+    /// - If this is an ndarray, the ndarray is returned.
+    /// - If this is a scalar, an unsized ndarray view is created on it.
+    pub fn as_ndarray<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> NDArrayObject<'ctx> {
+        match self {
+            ScalarOrNDArray::NDArray(ndarray) => *ndarray,
+            ScalarOrNDArray::Scalar(scalar) => scalar.as_ndarray(generator, ctx),
+        }
+    }
+
+    #[must_use]
+    pub fn dtype(&self) -> Type {
+        match self {
+            ScalarOrNDArray::Scalar(scalar) => scalar.ty,
+            ScalarOrNDArray::NDArray(ndarray) => ndarray.dtype,
+        }
+    }
+}
+
+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),
+        }
+    }
+}
+
+/// Split an [`AnyObject`] into a [`ScalarOrNDArray`] depending on its [`Type`].
+pub fn split_scalar_or_ndarray<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    object: AnyObject<'ctx>,
+) -> ScalarOrNDArray<'ctx> {
+    // TODO: Automatically convert a list into an ndarray?
+
+    match &*ctx.unifier.get_ty(object.ty) {
+        TypeEnum::TObj { obj_id, .. }
+            if *obj_id == ctx.primitives.ndarray.obj_id(&ctx.unifier).unwrap() =>
+        {
+            let ndarray = NDArrayObject::from_object(generator, ctx, object);
+            ScalarOrNDArray::NDArray(ndarray)
+        }
+        _ => {
+            let scalar = AnyObject { ty: object.ty, value: object.value };
+            ScalarOrNDArray::Scalar(scalar)
+        }
+    }
+}

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

@@ -0,0 +1,111 @@
+use util::gen_for_model_auto;
+
+use crate::{
+    codegen::{
+        model::*,
+        object::{list::ListObject, tuple::TupleObject, AnyObject},
+        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>,
+) -> (Int<'ctx, SizeT>, Ptr<'ctx, IntModel<SizeT>>) {
+    let sizet_model = IntModel(SizeT);
+
+    let zero = sizet_model.const_0(generator, ctx.ctx);
+    let one = sizet_model.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.gep(ctx, |f| f.len).load(generator, ctx, "len");
+            let result = sizet_model.array_alloca(generator, ctx, len.value, "int_sequence");
+
+            // Load all the `int32`s from the input_sequence, cast them to `SizeT`, and store them into `result`
+            gen_for_model_auto(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, "")
+                    .offset(generator, ctx, i.value, "")
+                    .load(generator, ctx, "")
+                    .value
+                    .into_int_value();
+
+                // Cast to SizeT
+                let int = sizet_model.s_extend_or_bit_cast(generator, ctx, int, "int");
+
+                // Store
+                result.offset(generator, ctx, i.value, "int").store(ctx, 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_int = input_sequence.len_static();
+            let len = sizet_model.constant(generator, ctx.ctx, len_int as u64);
+
+            let result = sizet_model.array_alloca(generator, ctx, len.value, "int_sequence");
+
+            for i in 0..len_int {
+                // Get the i-th element off of the tuple and load it into `result`.
+                let int = input_sequence.get(ctx, i, "dim").value.into_int_value();
+                let int = sizet_model.s_extend_or_bit_cast(generator, ctx, int, "int");
+
+                let offset = sizet_model.constant(generator, ctx.ctx, i as u64);
+                result.offset(generator, ctx, offset.value, "int").store(ctx, 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 = sizet_model.const_1(generator, ctx.ctx);
+            let result = sizet_model.array_alloca(generator, ctx, len.value, "int_sequence");
+
+            let int = sizet_model.s_extend_or_bit_cast(generator, ctx, input_int, "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/range.rs

@@ -0,0 +1,41 @@
+use crate::codegen::{
+    irrt::calculate_len_for_slice_range, model::*, structure::RangeModel, CodeGenContext,
+    CodeGenerator,
+};
+
+use super::AnyObject;
+
+/// A `range` in NAC3
+pub struct RangeObject<'ctx> {
+    pub instance: Ptr<'ctx, RangeModel>,
+}
+
+impl<'ctx> RangeObject<'ctx> {
+    pub fn from_object<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        object: AnyObject<'ctx>,
+    ) -> Self {
+        assert!(ctx.unifier.unioned(ctx.primitives.range, object.ty)); // Sanity check on type.
+
+        let model = PtrModel(RangeModel::default());
+        let instance = model.check_value(generator, ctx.ctx, object.value).unwrap();
+        RangeObject { instance }
+    }
+
+    /// Get the `len()` of this range.
+    pub fn len<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Int<'ctx, Int32> {
+        let start = self.instance.gep_start(generator, ctx, "").load(generator, ctx, "start");
+        let stop = self.instance.gep_stop(generator, ctx, "").load(generator, ctx, "stop");
+        let step = self.instance.gep_step(generator, ctx, "").load(generator, ctx, "step");
+
+        // TODO: Refactor this
+        let len =
+            calculate_len_for_slice_range(generator, ctx, start.value, stop.value, step.value);
+        IntModel(Int32).check_value(generator, ctx.ctx, len).unwrap()
+    }
+}

nac3core/src/codegen/object/tuple.rs

@@ -0,0 +1,113 @@
+use core::panic;
+
+use inkwell::values::StructValue;
+use itertools::Itertools;
+
+use crate::{
+    codegen::{model::*, CodeGenContext, CodeGenerator},
+    typecheck::typedef::{Type, TypeEnum},
+};
+
+use super::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 value of this tuple.
+    pub value: StructValue<'ctx>,
+}
+
+impl<'ctx> TupleObject<'ctx> {
+    // NOTE: There is no Model abstraction for Tuples with arbitrary lengths.
+    // Everything has to be done raw with Inkwell.
+
+    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)
+            );
+        };
+
+        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 }
+    }
+
+    /// Convenience function. Create a [`TupleObject`] from an iterator of objects.
+    pub fn create<I, G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        objects: I,
+        name: &str,
+    ) -> 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, name).unwrap().into_struct_value();
+        TupleObject { tys, value }
+    }
+
+    /// Get the `len()` of this tuple statically.
+    ///
+    /// We statically know the lengths of tuples in NAC3 when compiling.
+    #[must_use]
+    pub fn len_static(&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, '_>,
+    ) -> Int<'ctx, SizeT> {
+        IntModel(SizeT).constant(generator, ctx.ctx, self.len_static() as u64)
+    }
+
+    /// Check if this tuple is an empty/unit tuple.
+    #[must_use]
+    pub fn is_empty(&self) -> bool {
+        self.len_static() == 0
+    }
+
+    /// Get the `i`-th (0-based) object in this tuple.
+    pub fn get(&self, ctx: &mut CodeGenContext<'ctx, '_>, i: usize, name: &str) -> AnyObject<'ctx> {
+        assert!(
+            i < self.len_static(),
+            "Tuple object with length {} have index {i}",
+            self.len_static()
+        );
+
+        let value = ctx.builder.build_extract_value(self.value, i as u32, name).unwrap();
+        let ty = self.tys[i];
+        AnyObject { ty, value }
+    }
+}

nac3core/src/codegen/stmt.rs

@@ -1,8 +1,14 @@
+use super::model::*;
+use super::object::ndarray::indexing::util::gen_ndarray_subscript_ndindexes;
+use super::object::ndarray::scalar::split_scalar_or_ndarray;
+use super::object::ndarray::NDArrayObject;
+use super::object::AnyObject;
 use super::{
     super::symbol_resolver::ValueEnum,
     expr::destructure_range,
     irrt::{handle_slice_indices, list_slice_assignment},
-    CodeGenContext, CodeGenerator,
+    structure::{CSlice, Exception},
+    CodeGenContext, CodeGenerator, Int32, IntModel, Ptr, StructModel,
 };
 use crate::{
     codegen::{
@@ -401,7 +407,45 @@ 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_ndindexes(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, "assign_target_ndarray");
+
+            let value = split_scalar_or_ndarray(generator, ctx, value).as_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));
@@ -638,8 +682,12 @@ where
     I: Clone,
     InitFn: FnOnce(&mut G, &mut CodeGenContext<'ctx, 'a>) -> Result<I, String>,
     CondFn: FnOnce(&mut G, &mut CodeGenContext<'ctx, 'a>, I) -> Result<IntValue<'ctx>, String>,
-    BodyFn:
+    BodyFn: FnOnce(
-        FnOnce(&mut G, &mut CodeGenContext<'ctx, 'a>, BreakContinueHooks, I) -> Result<(), String>,
+        &mut G,
+        &mut CodeGenContext<'ctx, 'a>,
+        BreakContinueHooks<'ctx>,
+        I,
+    ) -> Result<(), String>,
     UpdateFn: FnOnce(&mut G, &mut CodeGenContext<'ctx, 'a>, I) -> Result<(), String>,
 {
     let label = label.unwrap_or("for");
@@ -719,7 +767,7 @@ where
     BodyFn: FnOnce(
         &mut G,
         &mut CodeGenContext<'ctx, 'a>,
-        BreakContinueHooks,
+        BreakContinueHooks<'ctx>,
         IntValue<'ctx>,
     ) -> Result<(), String>,
 {
@@ -1259,47 +1307,36 @@ pub fn exn_constructor<'ctx>(
 pub fn gen_raise<'ctx, G: CodeGenerator + ?Sized>(
     generator: &mut G,
     ctx: &mut CodeGenContext<'ctx, '_>,
-    exception: Option<&BasicValueEnum<'ctx>>,
+    exception: Option<Ptr<'ctx, StructModel<Exception>>>,
     loc: Location,
 ) {
-    if let Some(exception) = exception {
+    if let Some(pexn) = exception {
-        unsafe {
+        let i32_model = IntModel(Int32);
-            let int32 = ctx.ctx.i32_type();
+        let cslice_model = StructModel(CSlice);
-            let zero = int32.const_zero();
-            let exception = exception.into_pointer_value();
-            let file_ptr = ctx
-                .builder
-                .build_in_bounds_gep(exception, &[zero, int32.const_int(1, false)], "file_ptr")
-                .unwrap();
-            let filename = ctx.gen_string(generator, loc.file.0);
-            ctx.builder.build_store(file_ptr, filename).unwrap();
-            let row_ptr = ctx
-                .builder
-                .build_in_bounds_gep(exception, &[zero, int32.const_int(2, false)], "row_ptr")
-                .unwrap();
-            ctx.builder.build_store(row_ptr, int32.const_int(loc.row as u64, false)).unwrap();
-            let col_ptr = ctx
-                .builder
-                .build_in_bounds_gep(exception, &[zero, int32.const_int(3, false)], "col_ptr")
-                .unwrap();
-            ctx.builder.build_store(col_ptr, int32.const_int(loc.column as u64, false)).unwrap();
 
-            let current_fun = ctx.builder.get_insert_block().unwrap().get_parent().unwrap();
+        // Get and store filename
-            let fun_name = ctx.gen_string(generator, current_fun.get_name().to_str().unwrap());
+        let filename = loc.file.0;
-            let name_ptr = ctx
+        let filename = ctx.gen_string(generator, &String::from(filename)).value;
-                .builder
+        let filename = cslice_model.check_value(generator, ctx.ctx, filename).unwrap();
-                .build_in_bounds_gep(exception, &[zero, int32.const_int(4, false)], "name_ptr")
+        pexn.set(ctx, |f| f.filename, filename);
-                .unwrap();
+
-            ctx.builder.build_store(name_ptr, fun_name).unwrap();
+        let row = i32_model.constant(generator, ctx.ctx, loc.row as u64);
-        }
+        pexn.set(ctx, |f| f.line, row);
+
+        let column = i32_model.constant(generator, ctx.ctx, loc.column as u64);
+        pexn.set(ctx, |f| f.column, column);
+
+        let current_fn = ctx.builder.get_insert_block().unwrap().get_parent().unwrap();
+        let fn_name = ctx.gen_string(generator, current_fn.get_name().to_str().unwrap());
+        pexn.set(ctx, |f| f.function, fn_name);
 
         let raise = get_builtins(generator, ctx, "__nac3_raise");
-        let exception = *exception;
+        ctx.build_call_or_invoke(raise, &[pexn.value.into()], "raise");
-        ctx.build_call_or_invoke(raise, &[exception], "raise");
     } else {
         let resume = get_builtins(generator, ctx, "__nac3_resume");
         ctx.build_call_or_invoke(resume, &[], "resume");
     }
+
     ctx.builder.build_unreachable().unwrap();
 }
 
@@ -1761,30 +1798,41 @@ pub fn gen_stmt<G: CodeGenerator>(
                 } else {
                     return Ok(());
                 };
-                gen_raise(generator, ctx, Some(&exc), stmt.location);
+
+                let pexn_model = PtrModel(StructModel(Exception));
+                let exn = pexn_model.check_value(generator, ctx.ctx, exc).unwrap();
+
+                gen_raise(generator, ctx, Some(exn), stmt.location);
             } else {
                 gen_raise(generator, ctx, None, stmt.location);
             }
         }
         StmtKind::Assert { test, msg, .. } => {
-            let test = if let Some(v) = generator.gen_expr(ctx, test)? {
+            let byte_model = IntModel(Byte);
-                v.to_basic_value_enum(ctx, generator, test.custom.unwrap())?
+            let cslice_model = StructModel(CSlice);
-            } else {
+
+            let Some(test) = generator.gen_expr(ctx, test)? else {
                 return Ok(());
             };
+            let test = test.to_basic_value_enum(ctx, generator, ctx.primitives.bool)?;
+            let test = byte_model.check_value(generator, ctx.ctx, test).unwrap(); // Python `bool` is represented as `i8` in nac3core
+
+            // Check `msg`
             let err_msg = match msg {
                 Some(msg) => {
-                    if let Some(v) = generator.gen_expr(ctx, msg)? {
+                    let Some(msg) = generator.gen_expr(ctx, msg)? else {
-                        v.to_basic_value_enum(ctx, generator, msg.custom.unwrap())?
-                    } else {
                         return Ok(());
-                    }
-                }
-                None => ctx.gen_string(generator, "").into(),
                     };
+
+                    let msg = msg.to_basic_value_enum(ctx, generator, ctx.primitives.str)?;
+                    cslice_model.check_value(generator, ctx.ctx, msg).unwrap()
+                }
+                None => ctx.gen_string(generator, ""),
+            };
+
             ctx.make_assert_impl(
                 generator,
-                generator.bool_to_i1(ctx, test.into_int_value()),
+                test.value,
                 "0:AssertionError",
                 err_msg,
                 [None, None, None],

nac3core/src/codegen/structure.rs

@@ -0,0 +1,256 @@
+use inkwell::context::Context;
+
+use crate::codegen::model::*;
+
+use super::{CodeGenContext, CodeGenerator};
+
+/// Fields of [`CSlice`]
+pub struct CSliceFields<'ctx, F: FieldTraversal<'ctx>> {
+    /// Pointer to data.
+    pub base: F::Out<PtrModel<IntModel<Byte>>>,
+    /// Number of bytes of data.
+    pub len: F::Out<IntModel<SizeT>>,
+}
+
+/// See <https://crates.io/crates/cslice>.
+///
+/// Additionally, see <https://github.com/m-labs/artiq/blob/b0d2705c385f64b6e6711c1726cd9178f40b598e/artiq/firmware/libeh/eh_artiq.rs>)
+/// for ARTIQ-specific notes.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct CSlice;
+
+impl<'ctx> StructKind<'ctx> for CSlice {
+    type Fields<F: FieldTraversal<'ctx>> = CSliceFields<'ctx, F>;
+
+    fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+        Self::Fields { base: traversal.add_auto("base"), len: traversal.add_auto("len") }
+    }
+}
+
+impl StructModel<CSlice> {
+    /// Create a [`CSlice`].
+    ///
+    /// `base` and `len` must be LLVM global constants.
+    pub fn create_const<'ctx, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+        base: Ptr<'ctx, IntModel<Byte>>,
+        len: Int<'ctx, SizeT>,
+    ) -> Struct<'ctx, CSlice> {
+        let value = self
+            .0
+            .get_struct_type(generator, ctx)
+            .const_named_struct(&[base.value.into(), len.value.into()]);
+        self.believe_value(value)
+    }
+}
+
+/// The LLVM int type of an Exception ID.
+pub type ExceptionId = Int32;
+
+/// Fields of [`Exception<'ctx>`]
+///
+/// The definition came from `pub struct Exception<'a>` in
+/// <https://github.com/m-labs/artiq/blob/master/artiq/firmware/libeh/eh_artiq.rs>.
+pub struct ExceptionFields<'ctx, F: FieldTraversal<'ctx>> {
+    /// nac3core's ID of the exception
+    pub id: F::Out<IntModel<ExceptionId>>,
+    /// The name of the file this `Exception` was raised in.
+    pub filename: F::Out<StructModel<CSlice>>,
+    /// The line number in the file this `Exception` was raised in.
+    pub line: F::Out<IntModel<Int32>>,
+    /// The column number in the file this `Exception` was raised in.
+    pub column: F::Out<IntModel<Int32>>,
+    /// The name of the Python function this `Exception` was raised in.
+    pub function: F::Out<StructModel<CSlice>>,
+    /// The message of this Exception.
+    ///
+    /// The message can optionally contain integer parameters `{0}`, `{1}`, and `{2}` in its string,
+    /// where they will be substituted by `params[0]`, `params[1]`, and `params[2]` respectively (as `int64_t`s).
+    /// Here is an example:
+    ///
+    /// ```ignore
+    /// "Index {0} is out of bounds! List only has {1} element(s)."
+    /// ```
+    ///
+    /// In this case, `params[0]` and `params[1]` must be specified, and `params[2]` is ***unused***.
+    /// Having only 3 parameters is a constraint in ARTIQ.
+    pub msg: F::Out<StructModel<CSlice>>,
+    pub params: [F::Out<IntModel<Int64>>; 3],
+}
+
+/// nac3core & ARTIQ's Exception
+#[derive(Debug, Clone, Copy, Default)]
+pub struct Exception;
+
+impl<'ctx> StructKind<'ctx> for Exception {
+    type Fields<F: FieldTraversal<'ctx>> = ExceptionFields<'ctx, F>;
+
+    fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+        Self::Fields {
+            id: traversal.add_auto("id"),
+            filename: traversal.add_auto("filename"),
+            line: traversal.add_auto("line"),
+            column: traversal.add_auto("column"),
+            function: traversal.add_auto("function"),
+            msg: traversal.add_auto("msg"),
+            params: [
+                traversal.add_auto("params[0]"),
+                traversal.add_auto("params[1]"),
+                traversal.add_auto("params[2]"),
+            ],
+        }
+    }
+}
+
+/// Fields of [`List`]
+pub struct ListFields<'ctx, F: FieldTraversal<'ctx>, Item: Model<'ctx>> {
+    /// Array pointer to content
+    pub items: F::Out<PtrModel<Item>>,
+    /// Number of items in the array
+    pub len: F::Out<IntModel<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", PtrModel(self.item)),
+            len: traversal.add_auto("len"),
+        }
+    }
+}
+
+/// Fields of [`NDArray`]
+pub struct NDArrayFields<'ctx, F: FieldTraversal<'ctx>> {
+    pub data: F::Out<PtrModel<IntModel<Byte>>>,
+    pub itemsize: F::Out<IntModel<SizeT>>,
+    pub ndims: F::Out<IntModel<SizeT>>,
+    pub shape: F::Out<PtrModel<IntModel<SizeT>>>,
+    pub strides: F::Out<PtrModel<IntModel<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"),
+        }
+    }
+}
+
+/// Fields of [`SimpleNDArray`]
+#[derive(Debug, Clone, Copy)]
+pub struct SimpleNDArrayFields<'ctx, F: FieldTraversal<'ctx>, Item: Model<'ctx>> {
+    pub ndims: F::Out<IntModel<SizeT>>,
+    pub shape: F::Out<PtrModel<IntModel<SizeT>>>,
+    pub data: F::Out<PtrModel<Item>>,
+}
+
+/// An ndarray without strides and non-opaque `data` field in NAC3.
+#[derive(Debug, Clone, Copy)]
+pub struct SimpleNDArray<M> {
+    /// [`Model`] of the items.
+    pub item: M,
+}
+
+impl<'ctx, Item: Model<'ctx>> StructKind<'ctx> for SimpleNDArray<Item> {
+    type Fields<F: FieldTraversal<'ctx>> = SimpleNDArrayFields<'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", PtrModel(self.item)),
+        }
+    }
+}
+
+/// Fields of [`NDIter`]
+pub struct NDIterFields<'ctx, F: FieldTraversal<'ctx>> {
+    pub ndims: F::Out<IntModel<SizeT>>,
+    pub shape: F::Out<PtrModel<IntModel<SizeT>>>,
+    pub strides: F::Out<PtrModel<IntModel<SizeT>>>,
+
+    pub indices: F::Out<PtrModel<IntModel<SizeT>>>,
+    pub nth: F::Out<IntModel<SizeT>>,
+    pub element: F::Out<PtrModel<IntModel<Byte>>>,
+
+    pub size: F::Out<IntModel<SizeT>>,
+}
+
+/// An IRRT helper structure used when iterating 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 NAC3 `range`. It is an array of 3 int32s.
+// TODO: Use `pub type RangeModel<N> = NArrayModel<3, IntModel<N>>` in the future when
+// `range` type is type dependent.
+pub type RangeModel = NArrayModel<3, IntModel<Int32>>;
+
+impl<'ctx> Ptr<'ctx, RangeModel> {
+    pub fn gep_start<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        name: &str,
+    ) -> Ptr<'ctx, IntModel<Int32>> {
+        self.at_const(generator, ctx, 0, name)
+    }
+
+    pub fn gep_stop<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        name: &str,
+    ) -> Ptr<'ctx, IntModel<Int32>> {
+        self.at_const(generator, ctx, 1, name)
+    }
+
+    pub fn gep_step<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        name: &str,
+    ) -> Ptr<'ctx, IntModel<Int32>> {
+        self.at_const(generator, ctx, 2, name)
+    }
+}

nac3core/src/toplevel/helper.rs

@@ -51,6 +51,17 @@ pub enum PrimDef {
     FunNpArray,
     FunNpEye,
     FunNpIdentity,
+    FunNpArange,
+
+    // NumPy view functions
+    FunNpBroadcastTo,
+    FunNpReshape,
+    FunNpTranspose,
+
+    // NumPy NDArray property getters
+    FunNpSize,
+    FunNpShape,
+    FunNpStrides,
 
     // Miscellaneous NumPy & SciPy functions
     FunNpRound,
@@ -99,8 +110,6 @@ pub enum PrimDef {
     FunNpLdExp,
     FunNpHypot,
     FunNpNextAfter,
-    FunNpTranspose,
-    FunNpReshape,
 
     // Linalg functions
     FunNpDot,
@@ -237,6 +246,17 @@ impl PrimDef {
             PrimDef::FunNpArray => fun("np_array", None),
             PrimDef::FunNpEye => fun("np_eye", None),
             PrimDef::FunNpIdentity => fun("np_identity", None),
+            PrimDef::FunNpArange => fun("np_arange", None),
+
+            // NumPy view functions
+            PrimDef::FunNpBroadcastTo => fun("np_broadcast_to", None),
+            PrimDef::FunNpReshape => fun("np_reshape", None),
+            PrimDef::FunNpTranspose => fun("np_transpose", None),
+
+            // NumPy NDArray property getters
+            PrimDef::FunNpSize => fun("np_size", None),
+            PrimDef::FunNpShape => fun("np_shape", None),
+            PrimDef::FunNpStrides => fun("np_strides", None),
 
             // Miscellaneous NumPy & SciPy functions
             PrimDef::FunNpRound => fun("np_round", None),
@@ -285,8 +305,6 @@ 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),
@@ -1000,3 +1018,23 @@ pub fn arraylike_get_ndims(unifier: &mut Unifier, ty: Type) -> u64 {
         _ => 0,
     }
 }
+
+/// Extract an ndarray's `ndims` [type][`Type`] in `u64`. Panic if not possible.
+/// The `ndims` must only contain 1 value.
+#[must_use]
+pub fn extract_ndims(unifier: &Unifier, ndims_ty: Type) -> u64 {
+    let ndims_ty_enum = unifier.get_ty_immutable(ndims_ty);
+    let TypeEnum::TLiteral { values, .. } = &*ndims_ty_enum else {
+        panic!("ndims_ty should be a TLiteral");
+    };
+
+    assert_eq!(values.len(), 1, "ndims_ty TLiteral should only contain 1 value");
+
+    let ndims = values[0].clone();
+    u64::try_from(ndims).unwrap()
+}
+
+/// Return an ndarray's `ndims` as a typechecker [`Type`] from its `u64` value.
+pub fn create_ndims(unifier: &mut Unifier, ndims: u64) -> Type {
+    unifier.get_fresh_literal(vec![SymbolValue::U64(ndims)], None)
+}

nac3core/src/toplevel/mod.rs

@@ -31,6 +31,7 @@ pub mod builtins;
 pub mod composer;
 pub mod helper;
 pub mod numpy;
+pub mod option;
 pub mod type_annotation;
 use composer::*;
 use type_annotation::*;

nac3core/src/toplevel/option.rs

@@ -0,0 +1,46 @@
+use itertools::Itertools;
+
+use crate::{
+    toplevel::helper::PrimDef,
+    typecheck::{
+        type_inferencer::PrimitiveStore,
+        typedef::{Type, TypeEnum, Unifier, VarMap},
+    },
+};
+
+// TODO: This entire module is duplicated code (numpy.rs also has these kinds of things)
+
+/// Creates a `option` [`Type`] with the given type arguments.
+///
+/// * `dtype` - The element type of the `option`, or [`None`] if the type variable is not
+/// specialized.
+/// * `ndims` - The number of dimensions of the `option`, or [`None`] if the type variable is not
+/// specialized.
+pub fn make_option_ty(
+    unifier: &mut Unifier,
+    primitives: &PrimitiveStore,
+    dtype: Option<Type>,
+) -> Type {
+    subst_option_tvars(unifier, primitives.option, dtype)
+}
+
+/// Substitutes type variables in `option`.
+///
+/// * `dtype` - The element type of the `option`, or [`None`] if the type variable is not
+/// specialized.
+pub fn subst_option_tvars(unifier: &mut Unifier, option: Type, dtype: Option<Type>) -> Type {
+    let TypeEnum::TObj { obj_id, params, .. } = &*unifier.get_ty_immutable(option) else {
+        panic!("Expected `option` to be TObj, but got {}", unifier.stringify(option))
+    };
+    debug_assert_eq!(*obj_id, PrimDef::Option.id());
+
+    let tvar_ids = params.iter().map(|(obj_id, _)| *obj_id).collect_vec();
+    debug_assert_eq!(tvar_ids.len(), 1);
+
+    let mut tvar_subst = VarMap::new();
+    if let Some(dtype) = dtype {
+        tvar_subst.insert(tvar_ids[0], dtype);
+    }
+
+    unifier.subst(option, &tvar_subst).unwrap_or(option)
+}

nac3core/src/typecheck/function_check.rs

@@ -80,7 +80,7 @@ impl<'a> Inferencer<'a> {
                 return Err(HashSet::from([format!(
                     "expected concrete type at {} but got {}",
                     expr.location,
-                    self.unifier.get_ty(*ty).get_type_name()
+                    self.unifier.stringify(*ty)
                 )]));
             }
         }

nac3core/src/typecheck/magic_methods.rs

@@ -1,5 +1,5 @@
 use crate::symbol_resolver::SymbolValue;
-use crate::toplevel::helper::PrimDef;
+use crate::toplevel::helper::{arraylike_flatten_element_type, arraylike_get_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_dtype = arraylike_flatten_element_type(unifier, lhs);
-            let lhs_ndims = match &*unifier.get_ty_immutable(lhs_ndims) {
+            let rhs_dtype = arraylike_flatten_element_type(unifier, rhs);
-                TypeEnum::TLiteral { values, .. } => {
+
-                    assert_eq!(values.len(), 1);
+            let lhs_ndims = arraylike_get_ndims(unifier, lhs);
-                    u64::try_from(values[0].clone()).unwrap()
+            let rhs_ndims = arraylike_get_ndims(unifier, rhs);
+
+            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)
+                ));
             }
-                _ => unreachable!(),
+
-            };
+            let result_ndims = match (lhs_ndims, rhs_ndims) {
-            let (_, rhs_ndims) = unpack_ndarray_var_tys(unifier, rhs);
+                (0, _) | (_, 0) => {
-            let rhs_ndims = match &*unifier.get_ty_immutable(rhs_ndims) {
+                    return Err(
-                TypeEnum::TLiteral { values, .. } => {
+                        "ndarray.__matmul__ does not allow unsized ndarray input".to_string()
-                    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) {
+            if result_ndims == 0 {
-                (2, 2) => typeof_ndarray_broadcast(unifier, primitives, lhs, rhs)?,
+                // If the result is unsized, NumPy returns a scalar.
-                (lhs, rhs) if lhs == 0 || rhs == 0 => {
+                primitives.float
-                    return Err(format!(
+            } else {
-                    "Input operand {} does not have enough dimensions (has {lhs}, requires {rhs})",
+                let result_ndims_ty =
-                    u8::from(rhs == 0)
+                    unifier.get_fresh_literal(vec![SymbolValue::U64(result_ndims)], None);
-                ))
+                make_ndarray_ty(unifier, primitives, Some(primitives.float), Some(result_ndims_ty))
-                }
-                (lhs, rhs) => {
-                    return Err(format!(
-                        "ndarray.__matmul__ on {lhs}D and {rhs}D operands not supported"
-                    ))
-                }
             }
         }
 
@@ -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], Some(ndarray_t));
     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::{
@@ -12,6 +12,7 @@ use super::{
         RecordField, RecordKey, Type, TypeEnum, TypeVar, Unifier, VarMap,
     },
 };
+use crate::toplevel::option::make_option_ty;
 use crate::{
     symbol_resolver::{SymbolResolver, SymbolValue},
     toplevel::{
@@ -22,6 +23,7 @@ use crate::{
     typecheck::typedef::Mapping,
 };
 use itertools::{izip, Itertools};
+use nac3parser::ast::Constant;
 use nac3parser::ast::{
     self,
     fold::{self, Fold},
@@ -1325,7 +1327,7 @@ impl<'a> Inferencer<'a> {
                 arg_ty.obj_id(self.unifier).is_some_and(|id| id == PrimDef::NDArray.id())
             }) {
                 // typeof_ndarray_broadcast requires both dtypes to be the same, but ldexp accepts
-                // (float, int32), so convert it to align with the dtype of the first arg
+                // (float, int32), so convert it to align with t#he dtype of the first arg
                 let arg1_ty = if id == &"np_ldexp".into() {
                     if arg1_ty.obj_id(self.unifier).is_some_and(|id| id == PrimDef::NDArray.id()) {
                         let (_, ndims) = unpack_ndarray_var_tys(self.unifier, arg1_ty);
@@ -1501,7 +1503,7 @@ impl<'a> Inferencer<'a> {
                 },
             }));
         }
-        // 2-argument ndarray n-dimensional factory functions
+
         if id == &"np_reshape".into() && args.len() == 2 {
             let arg0 = self.fold_expr(args.remove(0))?;
 
@@ -1546,6 +1548,50 @@ impl<'a> Inferencer<'a> {
                 },
             }));
         }
+
+        if ["np_shape".into(), "np_strides".into()].contains(id) && args.len() == 1 {
+            // Output tuple size depends on input ndarray's ndims.
+
+            let ndarray = self.fold_expr(args.remove(0))?;
+
+            let ndims = arraylike_get_ndims(self.unifier, ndarray.custom.unwrap());
+
+            // Create 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![],
+                },
+            }));
+        }
+
         // 2-argument ndarray n-dimensional creation functions
         if id == &"np_full".into() && args.len() == 2 {
             let ExprKind::List { elts, .. } = &args[0].node else {
@@ -2264,14 +2310,39 @@ impl<'a> Inferencer<'a> {
         // We will also take the opportunity to deduce `dims_to_subtract` as well
         let mut dims_to_subtract: i128 = 0;
         for index in indices {
-            if let ExprKind::Slice { lower, upper, step } = &index.node {
+            match &index.node {
+                ExprKind::Slice { lower, upper, step } => {
+                    // Handle slices
                     for v in [lower.as_ref(), upper.as_ref(), step.as_ref()].iter().flatten() {
                         self.constrain(v.custom.unwrap(), self.primitives.int32, &v.location)?;
                     }
-            } else {
+                }
+                ExprKind::Constant { value: Constant::Ellipsis, .. } => {
+                    // Handle `...`.
+
+                    // See https://git.m-labs.hk/M-Labs/nac3/issues/486
+                    // Force `...` to have `()` (completely bogus) to make it concrete.
+                    let empty_tuple = TypeEnum::TTuple { ty: vec![], is_vararg_ctx: false };
+                    let empty_tuple = self.unifier.add_ty(empty_tuple);
+                    self.unify(index.custom.unwrap(), empty_tuple, &index.location)?;
+                }
+                ExprKind::Name { id, .. } if id == &"none".into() => {
+                    // Handle `np.newaxis` / `None`.
+                    dims_to_subtract -= 1;
+
+                    // "none" itself has type `Option[T]`, and since we have a stray `T` (non-concrete type).
+                    // We will force the type to be `Option[()]` to make it concrete. (TODO: is there a void type?)
+                    let empty_tuple = TypeEnum::TTuple { ty: vec![], is_vararg_ctx: false };
+                    let empty_tuple = self.unifier.add_ty(empty_tuple);
+                    let expected_type =
+                        make_option_ty(self.unifier, self.primitives, Some(empty_tuple));
+                    self.unify(index.custom.unwrap(), expected_type, &index.location)?;
+                }
+                _ => {
                     // Treat anything else as an integer index, and force unify their type to int32.
-                self.unify(index.custom.unwrap(), self.primitives.int32, &index.location)?;
                     dims_to_subtract += 1;
+                    self.unify(index.custom.unwrap(), self.primitives.int32, &index.location)?;
+                }
             }
         }
 

nac3core/src/typecheck/typedef/mod.rs

@@ -342,6 +342,14 @@ impl Unifier {
         self.unification_table.unioned(a, b)
     }
 
+    /// Determine if a type unions with a type in `tys`.
+    pub fn unioned_any<I>(&mut self, a: Type, tys: I) -> bool
+    where
+        I: IntoIterator<Item = Type>,
+    {
+        tys.into_iter().any(|ty| self.unioned(a, ty))
+    }
+
     pub fn from_shared_unifier(unifier: &SharedUnifier) -> Unifier {
         let lock = unifier.lock().unwrap();
         Unifier {

nac3standalone/demo/interpret_demo.py

@@ -218,8 +218,16 @@ def patch(module):
     module.np_ldexp = np.ldexp
     module.np_hypot = np.hypot
     module.np_nextafter = np.nextafter
-    module.np_transpose = np.transpose
+
+    # NumPy view functions
+    module.np_broadcast_to = np.broadcast_to
     module.np_reshape = np.reshape
+    module.np_transpose = np.transpose
+
+    # NumPy NDArray property getter functions
+    module.np_size = np.size
+    module.np_shape = np.shape
+    module.np_strides = lambda ndarray: ndarray.strides
 
     # SciPy Math functions
     module.sp_spec_erf = special.erf

nac3standalone/src/main.rs

@@ -14,6 +14,7 @@ use inkwell::{
     memory_buffer::MemoryBuffer, passes::PassBuilderOptions, support::is_multithreaded, targets::*,
     OptimizationLevel,
 };
+use nac3core::codegen::irrt::setup_irrt_exceptions;
 use nac3core::{
     codegen::{
         concrete_type::ConcreteTypeStore, irrt::load_irrt, CodeGenLLVMOptions,
@@ -314,6 +315,16 @@ fn main() {
     let resolver =
         Arc::new(Resolver(internal_resolver.clone())) as Arc<dyn SymbolResolver + Send + Sync>;
 
+    let context = inkwell::context::Context::create();
+
+    // Process IRRT
+    let irrt = load_irrt(&context);
+    setup_irrt_exceptions(&context, &irrt, resolver.as_ref());
+    if emit_llvm {
+        irrt.write_bitcode_to_path(Path::new("irrt.bc"));
+    }
+
+    // Process the Python script
     let parser_result = parser::parse_program(&program, file_name.into()).unwrap();
 
     for stmt in parser_result {
@@ -418,8 +429,8 @@ fn main() {
     registry.add_task(task);
     registry.wait_tasks_complete(handles);
 
+    // Link all modules together into `main`
     let buffers = membuffers.lock();
-    let context = inkwell::context::Context::create();
     let main = context
         .create_module_from_ir(MemoryBuffer::create_from_memory_range(&buffers[0], "main"))
         .unwrap();
@@ -439,12 +450,9 @@ fn main() {
         main.link_in_module(other).unwrap();
     }
 
-    let irrt = load_irrt(&context);
-    if emit_llvm {
-        irrt.write_bitcode_to_path(Path::new("irrt.bc"));
-    }
     main.link_in_module(irrt).unwrap();
 
+    // Private all functions except "run"
     let mut function_iter = main.get_first_function();
     while let Some(func) = function_iter {
         if func.count_basic_blocks() > 0 && func.get_name().to_str().unwrap() != "run" {
@@ -453,6 +461,7 @@ fn main() {
         function_iter = func.get_next_function();
     }
 
+    // Optimize `main`
     let target_machine = llvm_options
         .target
         .create_target_machine(llvm_options.opt_level)
@@ -466,6 +475,7 @@ fn main() {
         panic!("Failed to run optimization for module `main`: {}", err.to_string());
     }
 
+    // Write output
     target_machine
         .write_to_file(&main, FileType::Object, Path::new("module.o"))
         .expect("couldn't write module to file");