core/ndstrides: implement cmpop

Overlapping is not handled. Currently it has undefined behavior.

.clang-format

@@ -0,0 +1,32 @@
+BasedOnStyle: LLVM
+
+Language: Cpp
+Standard: Cpp11
+
+AccessModifierOffset: -1
+AlignEscapedNewlines: Left
+AlwaysBreakAfterReturnType: None
+AlwaysBreakTemplateDeclarations: Yes
+AllowAllParametersOfDeclarationOnNextLine: false
+AllowShortFunctionsOnASingleLine: Inline
+BinPackParameters: false
+BreakBeforeBinaryOperators: NonAssignment
+BreakBeforeTernaryOperators: true
+BreakConstructorInitializers: AfterColon
+BreakInheritanceList: AfterColon
+ColumnLimit: 120
+ConstructorInitializerAllOnOneLineOrOnePerLine: true
+ContinuationIndentWidth: 4
+DerivePointerAlignment: false
+IndentCaseLabels: true
+IndentPPDirectives: None
+IndentWidth: 4
+MaxEmptyLinesToKeep: 1
+PointerAlignment: Left
+ReflowComments: true
+SortIncludes: false
+SortUsingDeclarations: true
+SpaceAfterTemplateKeyword: false
+SpacesBeforeTrailingComments: 2
+TabWidth: 4
+UseTab: Never

Cargo.lock

@@ -117,9 +117,9 @@ checksum = "1fd0f2584146f6f2ef48085050886acf353beff7305ebd1ae69500e27c67f64b"
 
 [[package]]
 name = "cc"
-version = "1.1.13"
+version = "1.1.15"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "72db2f7947ecee9b03b510377e8bb9077afa27176fdbff55c51027e976fdcc48"
+checksum = "57b6a275aa2903740dc87da01c62040406b8812552e97129a63ea8850a17c6e6"
 dependencies = [
  "shlex",
 ]
@@ -161,7 +161,7 @@ dependencies = [
  "heck 0.5.0",
  "proc-macro2",
  "quote",
- "syn 2.0.75",
+ "syn 2.0.76",
 ]
 
 [[package]]
@@ -310,9 +310,9 @@ dependencies = [
 
 [[package]]
 name = "fastrand"
-version = "2.1.0"
+version = "2.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9fc0510504f03c51ada170672ac806f1f105a88aa97a5281117e1ddc3368e51a"
+checksum = "e8c02a5121d4ea3eb16a80748c74f5549a5665e4c21333c6098f283870fbdea6"
 
 [[package]]
 name = "fixedbitset"
@@ -424,7 +424,7 @@ checksum = "4fa4d8d74483041a882adaa9a29f633253a66dde85055f0495c121620ac484b2"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.75",
+ "syn 2.0.76",
 ]
 
 [[package]]
@@ -510,9 +510,9 @@ checksum = "bbd2bcb4c963f2ddae06a2efc7e9f3591312473c50c6685e1f298068316e66fe"
 
 [[package]]
 name = "libc"
-version = "0.2.157"
+version = "0.2.158"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "374af5f94e54fa97cf75e945cce8a6b201e88a1a07e688b47dfd2a59c66dbd86"
+checksum = "d8adc4bb1803a324070e64a98ae98f38934d91957a99cfb3a43dcbc01bc56439"
 
 [[package]]
 name = "libloading"
@@ -752,7 +752,7 @@ dependencies = [
  "phf_shared 0.11.2",
  "proc-macro2",
  "quote",
- "syn 2.0.75",
+ "syn 2.0.76",
 ]
 
 [[package]]
@@ -856,7 +856,7 @@ dependencies = [
  "proc-macro2",
  "pyo3-macros-backend",
  "quote",
- "syn 2.0.75",
+ "syn 2.0.76",
 ]
 
 [[package]]
@@ -869,14 +869,14 @@ dependencies = [
  "proc-macro2",
  "pyo3-build-config",
  "quote",
- "syn 2.0.75",
+ "syn 2.0.76",
 ]
 
 [[package]]
 name = "quote"
-version = "1.0.36"
+version = "1.0.37"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0fa76aaf39101c457836aec0ce2316dbdc3ab723cdda1c6bd4e6ad4208acaca7"
+checksum = "b5b9d34b8991d19d98081b46eacdd8eb58c6f2b201139f7c5f643cc155a633af"
 dependencies = [
  "proc-macro2",
 ]
@@ -942,9 +942,9 @@ dependencies = [
 
 [[package]]
 name = "redox_users"
-version = "0.4.5"
+version = "0.4.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bd283d9651eeda4b2a83a43c1c91b266c40fd76ecd39a50a8c630ae69dc72891"
+checksum = "ba009ff324d1fc1b900bd1fdb31564febe58a8ccc8a6fdbb93b543d33b13ca43"
 dependencies = [
  "getrandom",
  "libredox",
@@ -989,9 +989,9 @@ dependencies = [
 
 [[package]]
 name = "rustix"
-version = "0.38.34"
+version = "0.38.35"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "70dc5ec042f7a43c4a73241207cecc9873a06d45debb38b329f8541d85c2730f"
+checksum = "a85d50532239da68e9addb745ba38ff4612a242c1c7ceea689c4bc7c2f43c36f"
 dependencies = [
  "bitflags",
  "errno",
@@ -1035,29 +1035,29 @@ checksum = "61697e0a1c7e512e84a621326239844a24d8207b4669b41bc18b32ea5cbf988b"
 
 [[package]]
 name = "serde"
-version = "1.0.208"
+version = "1.0.209"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cff085d2cb684faa248efb494c39b68e522822ac0de72ccf08109abde717cfb2"
+checksum = "99fce0ffe7310761ca6bf9faf5115afbc19688edd00171d81b1bb1b116c63e09"
 dependencies = [
  "serde_derive",
 ]
 
 [[package]]
 name = "serde_derive"
-version = "1.0.208"
+version = "1.0.209"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "24008e81ff7613ed8e5ba0cfaf24e2c2f1e5b8a0495711e44fcd4882fca62bcf"
+checksum = "a5831b979fd7b5439637af1752d535ff49f4860c0f341d1baeb6faf0f4242170"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.75",
+ "syn 2.0.76",
 ]
 
 [[package]]
 name = "serde_json"
-version = "1.0.125"
+version = "1.0.127"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "83c8e735a073ccf5be70aa8066aa984eaf2fa000db6c8d0100ae605b366d31ed"
+checksum = "8043c06d9f82bd7271361ed64f415fe5e12a77fdb52e573e7f06a516dea329ad"
 dependencies = [
  "itoa",
  "memchr",
@@ -1147,7 +1147,7 @@ dependencies = [
  "proc-macro2",
  "quote",
  "rustversion",
- "syn 2.0.75",
+ "syn 2.0.76",
 ]
 
 [[package]]
@@ -1163,9 +1163,9 @@ dependencies = [
 
 [[package]]
 name = "syn"
-version = "2.0.75"
+version = "2.0.76"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f6af063034fc1935ede7be0122941bafa9bacb949334d090b77ca98b5817c7d9"
+checksum = "578e081a14e0cefc3279b0472138c513f37b41a08d5a3cca9b6e4e8ceb6cd525"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -1232,7 +1232,7 @@ checksum = "a4558b58466b9ad7ca0f102865eccc95938dca1a74a856f2b57b6629050da261"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.75",
+ "syn 2.0.76",
 ]
 
 [[package]]
@@ -1310,9 +1310,9 @@ checksum = "0336d538f7abc86d282a4189614dfaa90810dfc2c6f6427eaf88e16311dd225d"
 
 [[package]]
 name = "unicode-xid"
-version = "0.2.4"
+version = "0.2.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f962df74c8c05a667b5ee8bcf162993134c104e96440b663c8daa176dc772d8c"
+checksum = "229730647fbc343e3a80e463c1db7f78f3855d3f3739bee0dda773c9a037c90a"
 
 [[package]]
 name = "unicode_names2"
@@ -1510,5 +1510,5 @@ checksum = "fa4f8080344d4671fb4e831a13ad1e68092748387dfc4f55e356242fae12ce3e"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.75",
+ "syn 2.0.76",
 ]

nac3artiq/src/codegen.rs

@@ -2,7 +2,7 @@ use nac3core::{
     codegen::{
         classes::{
             ArrayLikeIndexer, ArrayLikeValue, ArraySliceValue, ListValue, NDArrayType,
-            NDArrayValue, RangeValue, UntypedArrayLikeAccessor,
+            NDArrayValue, ProxyType, ProxyValue, RangeValue, UntypedArrayLikeAccessor,
         },
         expr::{destructure_range, gen_call},
         irrt::call_ndarray_calc_size,
@@ -22,7 +22,7 @@ use inkwell::{
     module::Linkage,
     types::{BasicType, IntType},
     values::{BasicValueEnum, PointerValue, StructValue},
-    AddressSpace, IntPredicate,
+    AddressSpace, IntPredicate, OptimizationLevel,
 };
 
 use pyo3::{
@@ -32,6 +32,7 @@ use pyo3::{
 
 use crate::{symbol_resolver::InnerResolver, timeline::TimeFns};
 
+use inkwell::values::IntValue;
 use itertools::Itertools;
 use std::{
     collections::{hash_map::DefaultHasher, HashMap},
@@ -486,13 +487,10 @@ fn format_rpc_arg<'ctx>(
             let buffer = ctx.builder.build_array_alloca(llvm_i8, buffer_size, "rpc.arg").unwrap();
             let buffer = ArraySliceValue::from_ptr_val(buffer, buffer_size, Some("rpc.arg"));
 
-            let ppdata = generator.gen_var_alloc(ctx, llvm_arg_ty.element_type(), None).unwrap();
-            ctx.builder.build_store(ppdata, llvm_arg.data().base_ptr(ctx, generator)).unwrap();
-
             call_memcpy_generic(
                 ctx,
                 buffer.base_ptr(ctx, generator),
-                ppdata,
+                llvm_arg.ptr_to_data(ctx),
                 llvm_pdata_sizeof,
                 llvm_i1.const_zero(),
             );
@@ -528,6 +526,298 @@ fn format_rpc_arg<'ctx>(
     arg_slot
 }
 
+/// Formats an RPC return value to conform to the expected format required by NAC3.
+fn format_rpc_ret<'ctx>(
+    generator: &mut dyn CodeGenerator,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ret_ty: Type,
+) -> Option<BasicValueEnum<'ctx>> {
+    // -- receive value:
+    // T result = {
+    //   void *ret_ptr = alloca(sizeof(T));
+    //   void *ptr = ret_ptr;
+    //   loop: int size = rpc_recv(ptr);
+    //   // Non-zero: Provide `size` bytes of extra storage for variable-length data.
+    //   if(size) { ptr = alloca(size); goto loop; }
+    //   else *(T*)ret_ptr
+    // }
+
+    let llvm_i8 = ctx.ctx.i8_type();
+    let llvm_i32 = ctx.ctx.i32_type();
+    let llvm_i8_8 = ctx.ctx.struct_type(&[llvm_i8.array_type(8).into()], false);
+    let llvm_pi8 = llvm_i8.ptr_type(AddressSpace::default());
+
+    let rpc_recv = ctx.module.get_function("rpc_recv").unwrap_or_else(|| {
+        ctx.module.add_function("rpc_recv", llvm_i32.fn_type(&[llvm_pi8.into()], false), None)
+    });
+
+    if ctx.unifier.unioned(ret_ty, ctx.primitives.none) {
+        ctx.build_call_or_invoke(rpc_recv, &[llvm_pi8.const_null().into()], "rpc_recv");
+        return None;
+    }
+
+    let prehead_bb = ctx.builder.get_insert_block().unwrap();
+    let current_function = prehead_bb.get_parent().unwrap();
+    let head_bb = ctx.ctx.append_basic_block(current_function, "rpc.head");
+    let alloc_bb = ctx.ctx.append_basic_block(current_function, "rpc.continue");
+    let tail_bb = ctx.ctx.append_basic_block(current_function, "rpc.tail");
+
+    let llvm_ret_ty = ctx.get_llvm_abi_type(generator, ret_ty);
+
+    let result = match &*ctx.unifier.get_ty_immutable(ret_ty) {
+        TypeEnum::TObj { obj_id, .. } if *obj_id == PrimDef::NDArray.id() => {
+            let llvm_i1 = ctx.ctx.bool_type();
+            let llvm_usize = generator.get_size_type(ctx.ctx);
+
+            // Round `val` up to its modulo `power_of_two`
+            let round_up = |ctx: &mut CodeGenContext<'ctx, '_>,
+                            val: IntValue<'ctx>,
+                            power_of_two: IntValue<'ctx>| {
+                debug_assert_eq!(
+                    val.get_type().get_bit_width(),
+                    power_of_two.get_type().get_bit_width()
+                );
+
+                let llvm_val_t = val.get_type();
+
+                let max_rem = ctx
+                    .builder
+                    .build_int_sub(power_of_two, llvm_val_t.const_int(1, false), "")
+                    .unwrap();
+                ctx.builder
+                    .build_and(
+                        ctx.builder.build_int_add(val, max_rem, "").unwrap(),
+                        ctx.builder.build_not(max_rem, "").unwrap(),
+                        "",
+                    )
+                    .unwrap()
+            };
+
+            // Setup types
+            let (elem_ty, ndims) = unpack_ndarray_var_tys(&mut ctx.unifier, ret_ty);
+            let llvm_elem_ty = ctx.get_llvm_type(generator, elem_ty);
+            let llvm_ret_ty = NDArrayType::new(generator, ctx.ctx, llvm_elem_ty);
+
+            // Allocate the resulting ndarray
+            // A condition after format_rpc_ret ensures this will not be popped this off.
+            let ndarray = llvm_ret_ty.new_value(generator, ctx, Some("rpc.result"));
+
+            // Setup ndims
+            let ndims =
+                if let TypeEnum::TLiteral { values, .. } = &*ctx.unifier.get_ty_immutable(ndims) {
+                    assert_eq!(values.len(), 1);
+
+                    u64::try_from(values[0].clone()).unwrap()
+                } else {
+                    unreachable!();
+                };
+            // Set `ndarray.ndims`
+            ndarray.store_ndims(ctx, generator, llvm_usize.const_int(ndims, false));
+            // Allocate `ndarray.shape` [size_t; ndims]
+            ndarray.create_dim_sizes(ctx, llvm_usize, ndarray.load_ndims(ctx));
+
+            /*
+                ndarray now:
+                  - .ndims: initialized
+                  - .shape: allocated but uninitialized .shape
+                  - .data: uninitialized
+            */
+
+            let llvm_usize_sizeof = ctx
+                .builder
+                .build_int_truncate_or_bit_cast(llvm_usize.size_of(), llvm_usize, "")
+                .unwrap();
+            let llvm_pdata_sizeof = ctx
+                .builder
+                .build_int_truncate_or_bit_cast(
+                    llvm_ret_ty.element_type().size_of().unwrap(),
+                    llvm_usize,
+                    "",
+                )
+                .unwrap();
+            let llvm_elem_sizeof = ctx
+                .builder
+                .build_int_truncate_or_bit_cast(llvm_elem_ty.size_of().unwrap(), llvm_usize, "")
+                .unwrap();
+
+            // Allocates a buffer for the initial RPC'ed object, which is guaranteed to be
+            // (4 + 4 * ndims) bytes with 8-byte alignment
+            let sizeof_dims =
+                ctx.builder.build_int_mul(ndarray.load_ndims(ctx), llvm_usize_sizeof, "").unwrap();
+            let unaligned_buffer_size =
+                ctx.builder.build_int_add(sizeof_dims, llvm_pdata_sizeof, "").unwrap();
+            let buffer_size = round_up(ctx, unaligned_buffer_size, llvm_usize.const_int(8, false));
+
+            let stackptr = call_stacksave(ctx, None);
+            // Just to be absolutely sure, alloca in [i8 x 8] slices to force 8-byte alignment
+            let buffer = ctx
+                .builder
+                .build_array_alloca(
+                    llvm_i8_8,
+                    ctx.builder
+                        .build_int_unsigned_div(buffer_size, llvm_usize.const_int(8, false), "")
+                        .unwrap(),
+                    "rpc.buffer",
+                )
+                .unwrap();
+            let buffer = ctx
+                .builder
+                .build_bitcast(buffer, llvm_pi8, "")
+                .map(BasicValueEnum::into_pointer_value)
+                .unwrap();
+            let buffer = ArraySliceValue::from_ptr_val(buffer, buffer_size, None);
+
+            // The first call to `rpc_recv` reads the top-level ndarray object: [pdata, shape]
+            //
+            // The returned value is the number of bytes for `ndarray.data`.
+            let ndarray_nbytes = ctx
+                .build_call_or_invoke(
+                    rpc_recv,
+                    &[buffer.base_ptr(ctx, generator).into()], // Reads [usize; ndims]. NOTE: We are allocated [size_t; ndims].
+                    "rpc.size.next",
+                )
+                .map(BasicValueEnum::into_int_value)
+                .unwrap();
+
+            // debug_assert(ndarray_nbytes > 0)
+            if ctx.registry.llvm_options.opt_level == OptimizationLevel::None {
+                ctx.make_assert(
+                    generator,
+                    ctx.builder
+                        .build_int_compare(
+                            IntPredicate::UGT,
+                            ndarray_nbytes,
+                            ndarray_nbytes.get_type().const_zero(),
+                            "",
+                        )
+                        .unwrap(),
+                    "0:AssertionError",
+                    "Unexpected RPC termination for ndarray - Expected data buffer next",
+                    [None, None, None],
+                    ctx.current_loc,
+                );
+            }
+
+            // Copy shape from the buffer to `ndarray.shape`.
+            let pbuffer_dims =
+                unsafe { buffer.ptr_offset_unchecked(ctx, generator, &llvm_pdata_sizeof, None) };
+
+            call_memcpy_generic(
+                ctx,
+                ndarray.dim_sizes().base_ptr(ctx, generator),
+                pbuffer_dims,
+                sizeof_dims,
+                llvm_i1.const_zero(),
+            );
+            // Restore stack from before allocation of buffer
+            call_stackrestore(ctx, stackptr);
+
+            // Allocate `ndarray.data`.
+            // `ndarray.shape` must be initialized beforehand in this implementation
+            //   (for ndarray.create_data() to know how many elements to allocate)
+            let num_elements =
+                call_ndarray_calc_size(generator, ctx, &ndarray.dim_sizes(), (None, None));
+
+            // debug_assert(nelems * sizeof(T) >= ndarray_nbytes)
+            if ctx.registry.llvm_options.opt_level == OptimizationLevel::None {
+                let sizeof_data =
+                    ctx.builder.build_int_mul(num_elements, llvm_elem_sizeof, "").unwrap();
+
+                ctx.make_assert(
+                    generator,
+                    ctx.builder.build_int_compare(IntPredicate::UGE,
+                        sizeof_data,
+                        ndarray_nbytes,
+                        "",
+                    ).unwrap(),
+                    "0:AssertionError",
+                    "Unexpected allocation size request for ndarray data - Expected up to {0} bytes, got {1} bytes",
+                    [Some(sizeof_data), Some(ndarray_nbytes), None],
+                    ctx.current_loc,
+                );
+            }
+
+            ndarray.create_data(ctx, llvm_elem_ty, num_elements);
+
+            let ndarray_data = ndarray.data().base_ptr(ctx, generator);
+            let ndarray_data_i8 =
+                ctx.builder.build_pointer_cast(ndarray_data, llvm_pi8, "").unwrap();
+
+            // NOTE: Currently on `prehead_bb`
+            ctx.builder.build_unconditional_branch(head_bb).unwrap();
+
+            // Inserting into `head_bb`. Do `rpc_recv` for `data` recursively.
+            ctx.builder.position_at_end(head_bb);
+
+            let phi = ctx.builder.build_phi(llvm_pi8, "rpc.ptr").unwrap();
+            phi.add_incoming(&[(&ndarray_data_i8, prehead_bb)]);
+
+            let alloc_size = ctx
+                .build_call_or_invoke(rpc_recv, &[phi.as_basic_value()], "rpc.size.next")
+                .map(BasicValueEnum::into_int_value)
+                .unwrap();
+
+            let is_done = ctx
+                .builder
+                .build_int_compare(IntPredicate::EQ, llvm_i32.const_zero(), alloc_size, "rpc.done")
+                .unwrap();
+            ctx.builder.build_conditional_branch(is_done, tail_bb, alloc_bb).unwrap();
+
+            ctx.builder.position_at_end(alloc_bb);
+            // Align the allocation to sizeof(T)
+            let alloc_size = round_up(ctx, alloc_size, llvm_elem_sizeof);
+            let alloc_ptr = ctx
+                .builder
+                .build_array_alloca(
+                    llvm_elem_ty,
+                    ctx.builder.build_int_unsigned_div(alloc_size, llvm_elem_sizeof, "").unwrap(),
+                    "rpc.alloc",
+                )
+                .unwrap();
+            let alloc_ptr =
+                ctx.builder.build_pointer_cast(alloc_ptr, llvm_pi8, "rpc.alloc.ptr").unwrap();
+            phi.add_incoming(&[(&alloc_ptr, alloc_bb)]);
+            ctx.builder.build_unconditional_branch(head_bb).unwrap();
+
+            ctx.builder.position_at_end(tail_bb);
+            ndarray.as_base_value().into()
+        }
+
+        _ => {
+            let slot = ctx.builder.build_alloca(llvm_ret_ty, "rpc.ret.slot").unwrap();
+            let slotgen = ctx.builder.build_bitcast(slot, llvm_pi8, "rpc.ret.ptr").unwrap();
+            ctx.builder.build_unconditional_branch(head_bb).unwrap();
+            ctx.builder.position_at_end(head_bb);
+
+            let phi = ctx.builder.build_phi(llvm_pi8, "rpc.ptr").unwrap();
+            phi.add_incoming(&[(&slotgen, prehead_bb)]);
+            let alloc_size = ctx
+                .build_call_or_invoke(rpc_recv, &[phi.as_basic_value()], "rpc.size.next")
+                .unwrap()
+                .into_int_value();
+            let is_done = ctx
+                .builder
+                .build_int_compare(IntPredicate::EQ, llvm_i32.const_zero(), alloc_size, "rpc.done")
+                .unwrap();
+
+            ctx.builder.build_conditional_branch(is_done, tail_bb, alloc_bb).unwrap();
+            ctx.builder.position_at_end(alloc_bb);
+
+            let alloc_ptr =
+                ctx.builder.build_array_alloca(llvm_pi8, alloc_size, "rpc.alloc").unwrap();
+            let alloc_ptr =
+                ctx.builder.build_bitcast(alloc_ptr, llvm_pi8, "rpc.alloc.ptr").unwrap();
+            phi.add_incoming(&[(&alloc_ptr, alloc_bb)]);
+            ctx.builder.build_unconditional_branch(head_bb).unwrap();
+
+            ctx.builder.position_at_end(tail_bb);
+            ctx.builder.build_load(slot, "rpc.result").unwrap()
+        }
+    };
+
+    Some(result)
+}
+
 fn rpc_codegen_callback_fn<'ctx>(
     ctx: &mut CodeGenContext<'ctx, '_>,
     obj: Option<(Type, ValueEnum<'ctx>)>,
@@ -663,63 +953,14 @@ fn rpc_codegen_callback_fn<'ctx>(
     // reclaim stack space used by arguments
     call_stackrestore(ctx, stackptr);
 
-    // -- receive value:
-    // T result = {
-    //   void *ret_ptr = alloca(sizeof(T));
-    //   void *ptr = ret_ptr;
-    //   loop: int size = rpc_recv(ptr);
-    //   // Non-zero: Provide `size` bytes of extra storage for variable-length data.
-    //   if(size) { ptr = alloca(size); goto loop; }
-    //   else *(T*)ret_ptr
-    // }
-    let rpc_recv = ctx.module.get_function("rpc_recv").unwrap_or_else(|| {
-        ctx.module.add_function("rpc_recv", int32.fn_type(&[ptr_type.into()], false), None)
-    });
+    let result = format_rpc_ret(generator, ctx, fun.0.ret);
 
-    if ctx.unifier.unioned(fun.0.ret, ctx.primitives.none) {
-        ctx.build_call_or_invoke(rpc_recv, &[ptr_type.const_null().into()], "rpc_recv");
-        return Ok(None);
-    }
-
-    let prehead_bb = ctx.builder.get_insert_block().unwrap();
-    let current_function = prehead_bb.get_parent().unwrap();
-    let head_bb = ctx.ctx.append_basic_block(current_function, "rpc.head");
-    let alloc_bb = ctx.ctx.append_basic_block(current_function, "rpc.continue");
-    let tail_bb = ctx.ctx.append_basic_block(current_function, "rpc.tail");
-
-    let ret_ty = ctx.get_llvm_abi_type(generator, fun.0.ret);
-    let need_load = !ret_ty.is_pointer_type();
-    let slot = ctx.builder.build_alloca(ret_ty, "rpc.ret.slot").unwrap();
-    let slotgen = ctx.builder.build_bitcast(slot, ptr_type, "rpc.ret.ptr").unwrap();
-    ctx.builder.build_unconditional_branch(head_bb).unwrap();
-    ctx.builder.position_at_end(head_bb);
-
-    let phi = ctx.builder.build_phi(ptr_type, "rpc.ptr").unwrap();
-    phi.add_incoming(&[(&slotgen, prehead_bb)]);
-    let alloc_size = ctx
-        .build_call_or_invoke(rpc_recv, &[phi.as_basic_value()], "rpc.size.next")
-        .unwrap()
-        .into_int_value();
-    let is_done = ctx
-        .builder
-        .build_int_compare(inkwell::IntPredicate::EQ, int32.const_zero(), alloc_size, "rpc.done")
-        .unwrap();
-
-    ctx.builder.build_conditional_branch(is_done, tail_bb, alloc_bb).unwrap();
-    ctx.builder.position_at_end(alloc_bb);
-
-    let alloc_ptr = ctx.builder.build_array_alloca(ptr_type, alloc_size, "rpc.alloc").unwrap();
-    let alloc_ptr = ctx.builder.build_bitcast(alloc_ptr, ptr_type, "rpc.alloc.ptr").unwrap();
-    phi.add_incoming(&[(&alloc_ptr, alloc_bb)]);
-    ctx.builder.build_unconditional_branch(head_bb).unwrap();
-
-    ctx.builder.position_at_end(tail_bb);
-
-    let result = ctx.builder.build_load(slot, "rpc.result").unwrap();
-    if need_load {
+    if !result.is_some_and(|res| res.get_type().is_pointer_type()) {
+        // An RPC returning an NDArray would not touch here.
         call_stackrestore(ctx, stackptr);
     }
-    Ok(Some(result))
+
+    Ok(result)
 }
 
 pub fn attributes_writeback(

nac3artiq/src/lib.rs

@@ -448,7 +448,6 @@ impl Nac3 {
                         pyid_to_type: pyid_to_type.clone(),
                         primitive_ids: self.primitive_ids.clone(),
                         global_value_ids: global_value_ids.clone(),
-                        class_names: Mutex::default(),
                         name_to_pyid: name_to_pyid.clone(),
                         module: module.clone(),
                         id_to_pyval: RwLock::default(),
@@ -540,7 +539,6 @@ impl Nac3 {
             pyid_to_type: pyid_to_type.clone(),
             primitive_ids: self.primitive_ids.clone(),
             global_value_ids: global_value_ids.clone(),
-            class_names: Mutex::default(),
             id_to_pyval: RwLock::default(),
             id_to_primitive: RwLock::default(),
             field_to_val: RwLock::default(),
@@ -557,6 +555,10 @@ 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, resolver.as_ref());
+
         let fun_signature =
             FunSignature { args: vec![], ret: self.primitive.none, vars: VarMap::new() };
         let mut store = ConcreteTypeStore::new();
@@ -727,7 +729,7 @@ 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 +758,7 @@ impl Nac3 {
             )
             .unwrap();
 
-        main.link_in_module(load_irrt(&context))
-            .map_err(|err| CompileError::new_err(err.to_string()))?;
+        main.link_in_module(irrt).map_err(|err| CompileError::new_err(err.to_string()))?;
 
         let mut function_iter = main.get_first_function();
         while let Some(func) = function_iter {

nac3artiq/src/symbol_resolver.rs

@@ -23,7 +23,7 @@ use nac3core::{
     },
 };
 use nac3parser::ast::{self, StrRef};
-use parking_lot::{Mutex, RwLock};
+use parking_lot::RwLock;
 use pyo3::{
     types::{PyDict, PyTuple},
     PyAny, PyObject, PyResult, Python,
@@ -79,7 +79,6 @@ pub struct InnerResolver {
     pub id_to_primitive: RwLock<HashMap<u64, PrimitiveValue>>,
     pub field_to_val: RwLock<HashMap<ResolverField, Option<PyFieldHandle>>>,
     pub global_value_ids: Arc<RwLock<HashMap<u64, PyObject>>>,
-    pub class_names: Mutex<HashMap<StrRef, Type>>,
     pub pyid_to_def: Arc<RwLock<HashMap<u64, DefinitionId>>>,
     pub pyid_to_type: Arc<RwLock<HashMap<u64, Type>>>,
     pub primitive_ids: PrimitivePythonId,

nac3core/build.rs

@@ -8,37 +8,50 @@ use std::{
 };
 
 fn main() {
-    const FILE: &str = "src/codegen/irrt/irrt.cpp";
+    let out_dir = env::var("OUT_DIR").unwrap();
+    let out_dir = Path::new(&out_dir);
+    let irrt_dir = Path::new("irrt");
+
+    let irrt_cpp_path = irrt_dir.join("irrt.cpp");
 
     /*
      * 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 mut flags: Vec<&str> = vec![
         "--target=wasm32",
-        FILE,
         "-x",
         "c++",
+        "-std=c++20",
         "-fno-discard-value-names",
         "-fno-exceptions",
         "-fno-rtti",
-        match env::var("PROFILE").as_deref() {
-            Ok("debug") => "-O0",
-            Ok("release") => "-O3",
-            flavor => panic!("Unknown or missing build flavor {flavor:?}"),
-        },
         "-emit-llvm",
         "-S",
         "-Wall",
         "-Wextra",
         "-o",
         "-",
+        "-I",
+        irrt_dir.to_str().unwrap(),
+        irrt_cpp_path.to_str().unwrap(),
     ];
 
-    println!("cargo:rerun-if-changed={FILE}");
-    let out_dir = env::var("OUT_DIR").unwrap();
-    let out_path = Path::new(&out_dir);
+    match env::var("PROFILE").as_deref() {
+        Ok("debug") => {
+            flags.push("-O0");
+            flags.push("-DIRRT_DEBUG_ASSERT");
+        }
+        Ok("release") => {
+            flags.push("-O3");
+        }
+        flavor => panic!("Unknown or missing build flavor {flavor:?}"),
+    }
 
+    // Tell Cargo to rerun if any file under `irrt_dir` (recursive) changes
+    println!("cargo:rerun-if-changed={}", irrt_dir.to_str().unwrap());
+
+    // Compile IRRT and capture the LLVM IR output
     let output = Command::new("clang-irrt")
         .args(flags)
         .output()
@@ -52,7 +65,17 @@ 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]);
@@ -63,18 +86,22 @@ fn main() {
         .unwrap()
         .replace_all(&filtered_output, "");
 
-    println!("cargo:rerun-if-env-changed=DEBUG_DUMP_IRRT");
-    if env::var("DEBUG_DUMP_IRRT").is_ok() {
-        let mut file = File::create(out_path.join("irrt.ll")).unwrap();
+    // For debugging
+    // Doing `DEBUG_DUMP_IRRT=1 cargo build -p nac3core` dumps the LLVM IR generated
+    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")
         .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();

nac3core/irrt/irrt.cpp

@@ -0,0 +1,15 @@
+#include "irrt/exception.hpp"
+#include "irrt/int_types.hpp"
+#include "irrt/list.hpp"
+#include "irrt/math.hpp"
+#include "irrt/ndarray.hpp"
+#include "irrt/range.hpp"
+#include "irrt/slice.hpp"
+#include "irrt/ndarray/basic.hpp"
+#include "irrt/ndarray/def.hpp"
+#include "irrt/ndarray/iter.hpp"
+#include "irrt/ndarray/indexing.hpp"
+#include "irrt/ndarray/array.hpp"
+#include "irrt/ndarray/reshape.hpp"
+#include "irrt/ndarray/broadcast.hpp"
+#include "irrt/ndarray/transpose.hpp"

nac3core/irrt/irrt/cslice.hpp

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

nac3core/irrt/irrt/debug.hpp

@@ -0,0 +1,25 @@
+#pragma once
+
+// Set in nac3core/build.rs
+#ifdef IRRT_DEBUG_ASSERT
+#define IRRT_DEBUG_ASSERT_BOOL true
+#else
+#define IRRT_DEBUG_ASSERT_BOOL false
+#endif
+
+#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 constexpr (IRRT_DEBUG_ASSERT_BOOL) {                                        \
+        if ((lhs) != (rhs)) {                                                      \
+            raise_debug_assert(SizeT, "LHS = {0}. RHS = {1}", lhs, rhs, NO_PARAM); \
+        }                                                                          \
+    }
+
+#define debug_assert(SizeT, expr)                                                  \
+    if constexpr (IRRT_DEBUG_ASSERT_BOOL) {                                        \
+        if (!(expr)) {                                                             \
+            raise_debug_assert(SizeT, "Got false.", NO_PARAM, NO_PARAM, NO_PARAM); \
+        }                                                                          \
+    }

nac3core/irrt/irrt/exception.hpp

@@ -0,0 +1,85 @@
+#pragma once
+
+#include "irrt/cslice.hpp"
+#include "irrt/int_types.hpp"
+
+/**
+ * @brief The int type of ARTIQ exception IDs.
+ */
+typedef int32_t ExceptionId;
+
+/*
+ * Set of exceptions C++ IRRT can use.
+ * Must be synchronized with `setup_irrt_exceptions` in `nac3core/src/codegen/irrt/mod.rs`.
+ */
+extern "C" {
+ExceptionId EXN_INDEX_ERROR;
+ExceptionId EXN_VALUE_ERROR;
+ExceptionId EXN_ASSERTION_ERROR;
+ExceptionId EXN_TYPE_ERROR;
+}
+
+/**
+ * @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);
+
+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];
+};
+
+constexpr int64_t NO_PARAM = 0;
+
+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 = reinterpret_cast<uint8_t*>(const_cast<char*>(filename)),
+                     .len = static_cast<int32_t>(__builtin_strlen(filename))},
+        .line = line,
+        .column = 0,
+        .function = {.base = reinterpret_cast<uint8_t*>(const_cast<char*>(function)),
+                     .len = static_cast<int32_t>(__builtin_strlen(function))},
+        .msg = {.base = reinterpret_cast<uint8_t*>(const_cast<char*>(msg)),
+                .len = static_cast<int32_t>(__builtin_strlen(msg))},
+    };
+    e.params[0] = param0;
+    e.params[1] = param1;
+    e.params[2] = param2;
+    __nac3_raise(reinterpret_cast<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` to `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)
+}  // namespace

nac3core/irrt/irrt/int_types.hpp

@@ -0,0 +1,13 @@
+#pragma once
+
+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);
+
+// NDArray indices are always `uint32_t`.
+using NDIndexInt = uint32_t;
+// The type of an index or a value describing the length of a range/slice is always `int32_t`.
+using SliceIndex = int32_t;

nac3core/irrt/irrt/list.hpp

@@ -0,0 +1,90 @@
+#pragma once
+
+#include "irrt/int_types.hpp"
+#include "irrt/math_util.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
+
+extern "C" {
+// Handle list assignment and dropping part of the list when
+// both dest_step and src_step are +1.
+// - 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)
+SliceIndex __nac3_list_slice_assign_var_size(SliceIndex dest_start,
+                                             SliceIndex dest_end,
+                                             SliceIndex dest_step,
+                                             uint8_t* dest_arr,
+                                             SliceIndex dest_arr_len,
+                                             SliceIndex src_start,
+                                             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 (src_step == dest_step && dest_step == 1) {
+        const SliceIndex src_len = (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(dest_arr + dest_start * size, src_arr + src_start * size, src_len * size);
+        }
+        if (dest_len > 0) {
+            /* dropping */
+            __builtin_memmove(dest_arr + (dest_start + src_len) * size, dest_arr + (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)
+                          && !(max(dest_start, dest_end) < min(src_start, src_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));
+        __builtin_memcpy(tmp, src_arr, src_arr_len * size);
+        src_arr = tmp;
+    }
+    SliceIndex src_ind = src_start;
+    SliceIndex dest_ind = dest_start;
+    for (; (src_step > 0) ? (src_ind <= src_end) : (src_ind >= src_end); 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);
+        } else if (size == 4) {
+            __builtin_memcpy(dest_arr + dest_ind * 4, src_arr + src_ind * 4, 4);
+        } 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 */
+            __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_end + 1) * size,
+                          (dest_arr_len - dest_end - 1) * size);
+        return dest_arr_len - (dest_end - dest_ind) - 1;
+    }
+    return dest_arr_len;
+}
+}  // extern "C"

nac3core/irrt/irrt/math.hpp

@@ -0,0 +1,93 @@
+#pragma once
+
+namespace {
+// 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) {
+    T res = 1;
+    /* repeated squaring method */
+    do {
+        if (exp & 1) {
+            res *= base; /* for n odd */
+        }
+        exp >>= 1;
+        base *= base;
+    } while (exp);
+    return res;
+}
+}  // namespace
+
+#define DEF_nac3_int_exp_(T)                   \
+    T __nac3_int_exp_##T(T base, T exp) {      \
+        return __nac3_int_exp_impl(base, exp); \
+    }
+
+extern "C" {
+
+// Putting semicolons here to make clang-format not reformat this into
+// a stair shape.
+DEF_nac3_int_exp_(int32_t);
+DEF_nac3_int_exp_(int64_t);
+DEF_nac3_int_exp_(uint32_t);
+DEF_nac3_int_exp_(uint64_t);
+
+int32_t __nac3_isinf(double x) {
+    return __builtin_isinf(x);
+}
+
+int32_t __nac3_isnan(double x) {
+    return __builtin_isnan(x);
+}
+
+double tgamma(double arg);
+
+double __nac3_gamma(double z) {
+    // Handling for denormals
+    //     | x                 | Python gamma(x) | C tgamma(x) |
+    // --- | ----------------- | --------------- | ----------- |
+    // (1) | nan               | nan             | nan         |
+    // (2) | -inf              | -inf            | inf         |
+    // (3) | inf               | inf             | inf         |
+    // (4) | 0.0               | inf             | inf         |
+    // (5) | {-1.0, -2.0, ...} | inf             | nan         |
+
+    // (1)-(3)
+    if (__builtin_isinf(z) || __builtin_isnan(z)) {
+        return z;
+    }
+
+    double v = tgamma(z);
+
+    // (4)-(5)
+    return __builtin_isinf(v) || __builtin_isnan(v) ? __builtin_inf() : v;
+}
+
+double lgamma(double arg);
+
+double __nac3_gammaln(double x) {
+    // libm's handling of value overflows differs from scipy:
+    // - scipy: gammaln(-inf) -> -inf
+    // - libm : lgamma(-inf) -> inf
+
+    if (__builtin_isinf(x)) {
+        return x;
+    }
+
+    return lgamma(x);
+}
+
+double j0(double x);
+
+double __nac3_j0(double x) {
+    // libm's handling of value overflows differs from scipy:
+    // - scipy: j0(inf) -> nan
+    // - libm : j0(inf) -> 0.0
+
+    if (__builtin_isinf(x)) {
+        return __builtin_nan("");
+    }
+
+    return j0(x);
+}
+}

nac3core/irrt/irrt/math_util.hpp

@@ -0,0 +1,13 @@
+#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;
+}
+}  // namespace

nac3core/irrt/irrt/ndarray.hpp

@@ -0,0 +1,151 @@
+#pragma once
+
+#include "irrt/int_types.hpp"
+
+// TODO: To be deleted since NDArray with strides is done.
+
+namespace {
+template<typename SizeT>
+SizeT __nac3_ndarray_calc_size_impl(const SizeT* list_data, SizeT list_len, SizeT begin_idx, SizeT end_idx) {
+    __builtin_assume(end_idx <= list_len);
+
+    SizeT num_elems = 1;
+    for (SizeT i = begin_idx; i < end_idx; ++i) {
+        SizeT val = list_data[i];
+        __builtin_assume(val > 0);
+        num_elems *= val;
+    }
+    return num_elems;
+}
+
+template<typename SizeT>
+void __nac3_ndarray_calc_nd_indices_impl(SizeT index, const SizeT* dims, SizeT num_dims, NDIndexInt* idxs) {
+    SizeT stride = 1;
+    for (SizeT dim = 0; dim < num_dims; dim++) {
+        SizeT i = num_dims - dim - 1;
+        __builtin_assume(dims[i] > 0);
+        idxs[i] = (index / stride) % dims[i];
+        stride *= dims[i];
+    }
+}
+
+template<typename SizeT>
+SizeT __nac3_ndarray_flatten_index_impl(const SizeT* dims,
+                                        SizeT num_dims,
+                                        const NDIndexInt* indices,
+                                        SizeT num_indices) {
+    SizeT idx = 0;
+    SizeT stride = 1;
+    for (SizeT i = 0; i < num_dims; ++i) {
+        SizeT ri = num_dims - i - 1;
+        if (ri < num_indices) {
+            idx += stride * indices[ri];
+        }
+
+        __builtin_assume(dims[i] > 0);
+        stride *= dims[ri];
+    }
+    return idx;
+}
+
+template<typename SizeT>
+void __nac3_ndarray_calc_broadcast_impl(const SizeT* lhs_dims,
+                                        SizeT lhs_ndims,
+                                        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* 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) {
+            *out_dim = *rhs_dim_sz;
+        } else if (rhs_dim_sz == nullptr) {
+            *out_dim = *lhs_dim_sz;
+        } else if (*lhs_dim_sz == 1) {
+            *out_dim = *rhs_dim_sz;
+        } else if (*rhs_dim_sz == 1) {
+            *out_dim = *lhs_dim_sz;
+        } else if (*lhs_dim_sz == *rhs_dim_sz) {
+            *out_dim = *lhs_dim_sz;
+        } else {
+            __builtin_unreachable();
+        }
+    }
+}
+
+template<typename SizeT>
+void __nac3_ndarray_calc_broadcast_idx_impl(const SizeT* src_dims,
+                                            SizeT src_ndims,
+                                            const NDIndexInt* in_idx,
+                                            NDIndexInt* out_idx) {
+    for (SizeT i = 0; i < src_ndims; ++i) {
+        SizeT src_i = src_ndims - i - 1;
+        out_idx[src_i] = src_dims[src_i] == 1 ? 0 : in_idx[src_i];
+    }
+}
+}  // namespace
+
+extern "C" {
+uint32_t __nac3_ndarray_calc_size(const uint32_t* list_data, uint32_t list_len, uint32_t begin_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(const uint64_t* list_data, uint64_t list_len, uint64_t begin_idx, uint64_t end_idx) {
+    return __nac3_ndarray_calc_size_impl(list_data, list_len, begin_idx, end_idx);
+}
+
+void __nac3_ndarray_calc_nd_indices(uint32_t index, const uint32_t* dims, uint32_t num_dims, NDIndexInt* idxs) {
+    __nac3_ndarray_calc_nd_indices_impl(index, dims, num_dims, idxs);
+}
+
+void __nac3_ndarray_calc_nd_indices64(uint64_t index, const uint64_t* dims, uint64_t num_dims, NDIndexInt* idxs) {
+    __nac3_ndarray_calc_nd_indices_impl(index, dims, num_dims, idxs);
+}
+
+uint32_t
+__nac3_ndarray_flatten_index(const uint32_t* dims, uint32_t num_dims, const NDIndexInt* indices, uint32_t num_indices) {
+    return __nac3_ndarray_flatten_index_impl(dims, num_dims, indices, num_indices);
+}
+
+uint64_t __nac3_ndarray_flatten_index64(const uint64_t* dims,
+                                        uint64_t num_dims,
+                                        const NDIndexInt* indices,
+                                        uint64_t num_indices) {
+    return __nac3_ndarray_flatten_index_impl(dims, num_dims, indices, num_indices);
+}
+
+void __nac3_ndarray_calc_broadcast(const uint32_t* lhs_dims,
+                                   uint32_t lhs_ndims,
+                                   const uint32_t* rhs_dims,
+                                   uint32_t rhs_ndims,
+                                   uint32_t* out_dims) {
+    return __nac3_ndarray_calc_broadcast_impl(lhs_dims, lhs_ndims, rhs_dims, rhs_ndims, out_dims);
+}
+
+void __nac3_ndarray_calc_broadcast64(const uint64_t* lhs_dims,
+                                     uint64_t lhs_ndims,
+                                     const uint64_t* rhs_dims,
+                                     uint64_t rhs_ndims,
+                                     uint64_t* out_dims) {
+    return __nac3_ndarray_calc_broadcast_impl(lhs_dims, lhs_ndims, rhs_dims, rhs_ndims, out_dims);
+}
+
+void __nac3_ndarray_calc_broadcast_idx(const uint32_t* src_dims,
+                                       uint32_t src_ndims,
+                                       const NDIndexInt* in_idx,
+                                       NDIndexInt* out_idx) {
+    __nac3_ndarray_calc_broadcast_idx_impl(src_dims, src_ndims, in_idx, out_idx);
+}
+
+void __nac3_ndarray_calc_broadcast_idx64(const uint64_t* src_dims,
+                                         uint64_t src_ndims,
+                                         const NDIndexInt* in_idx,
+                                         NDIndexInt* out_idx) {
+    __nac3_ndarray_calc_broadcast_idx_impl(src_dims, src_ndims, in_idx, out_idx);
+}
+}

nac3core/irrt/irrt/ndarray/array.hpp

@@ -0,0 +1,134 @@
+#pragma once
+
+#include "irrt/debug.hpp"
+#include "irrt/exception.hpp"
+#include "irrt/int_types.hpp"
+#include "irrt/list.hpp"
+#include "irrt/ndarray/basic.hpp"
+#include "irrt/ndarray/def.hpp"
+
+namespace {
+namespace ndarray {
+namespace array {
+/**
+ * @brief In the context of `np.array(<list>)`, deduce the ndarray's shape produced by `<list>` and raise
+ * an exception if there is anything wrong with `<shape>` (e.g., inconsistent dimensions `np.array([[1.0, 2.0],
+ * [3.0]])`)
+ *
+ * If this function finds no issues with `<list>`, the deduced shape is written to `shape`. The caller has the
+ * responsibility to allocate `[SizeT; ndims]` for `shape`. The caller must also initialize `shape` with `-1`s because
+ * of implementation details.
+ */
+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);
+        }
+    }
+}
+
+/**
+ * @brief See `set_and_validate_list_shape_helper`.
+ */
+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);
+}
+
+/**
+ * @brief In the context of `np.array(<list>)`, copied the contents stored in `list` to `ndarray`.
+ *
+ * `list` is assumed to be "legal". (i.e., no inconsistent dimensions)
+ *
+ * # Notes on `ndarray`
+ * The caller is responsible for allocating space for `ndarray`.
+ * Here is what this function expects from `ndarray` when called:
+ *   - `ndarray->data` has to be allocated, contiguous, and may contain uninitialized values.
+ *   - `ndarray->itemsize` has to be initialized.
+ *   - `ndarray->ndims` has to be initialized.
+ *   - `ndarray->shape` has to be initialized.
+ *   - `ndarray->strides` is ignored, but note that `ndarray->data` is contiguous.
+ * When this function call ends:
+ *   - `ndarray->data` is written with contents from `<list>`.
+ */
+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[scalar]`
+        // `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);
+        }
+    }
+}
+
+/**
+ * @brief See `write_list_to_array_helper`.
+ */
+template<typename SizeT>
+void write_list_to_array(List<SizeT>* list, NDArray<SizeT>* ndarray) {
+    SizeT index = 0;
+    write_list_to_array_helper<SizeT>((SizeT)0, &index, list, ndarray);
+}
+}  // namespace array
+}  // namespace ndarray
+}  // namespace
+
+extern "C" {
+using namespace ndarray::array;
+
+void __nac3_ndarray_array_set_and_validate_list_shape(List<int32_t>* list, int32_t ndims, int32_t* shape) {
+    set_and_validate_list_shape(list, ndims, shape);
+}
+
+void __nac3_ndarray_array_set_and_validate_list_shape64(List<int64_t>* list, int64_t ndims, int64_t* shape) {
+    set_and_validate_list_shape(list, ndims, shape);
+}
+
+void __nac3_ndarray_array_write_list_to_array(List<int32_t>* list, NDArray<int32_t>* ndarray) {
+    write_list_to_array(list, ndarray);
+}
+
+void __nac3_ndarray_array_write_list_to_array64(List<int64_t>* list, NDArray<int64_t>* ndarray) {
+    write_list_to_array(list, ndarray);
+}
+}

nac3core/irrt/irrt/ndarray/basic.hpp

@@ -0,0 +1,341 @@
+#pragma once
+
+#include "irrt/debug.hpp"
+#include "irrt/exception.hpp"
+#include "irrt/int_types.hpp"
+#include "irrt/ndarray/def.hpp"
+
+namespace {
+namespace ndarray {
+namespace basic {
+/**
+ * @brief Assert 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 Assert that two shapes are the same in the context of writing output to an ndarray.
+ */
+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 Return 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 length.
+ */
+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) {
+    // 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` along the ndarray's axes.
+ *
+ * This function does no bound check.
+ */
+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 Return the pointer to the nth (0-based) element of `ndarray` in flattened view.
+ *
+ * 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` to be contiguous.
+ *
+ * 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 when we see a contiguous segment.
+    // TODO: Handle overlapping.
+
+    debug_assert_eq(SizeT, src_ndarray->itemsize, dst_ndarray->itemsize);
+
+    for (SizeT i = 0; i < size(src_ndarray); i++) {
+        auto src_element = ndarray::basic::get_nth_pelement(src_ndarray, i);
+        auto dst_element = ndarray::basic::get_nth_pelement(dst_ndarray, i);
+        ndarray::basic::set_pelement_value(dst_ndarray, dst_element, src_element);
+    }
+}
+}  // namespace basic
+}  // namespace ndarray
+}  // namespace
+
+extern "C" {
+using namespace ndarray::basic;
+
+void __nac3_ndarray_util_assert_shape_no_negative(int32_t ndims, int32_t* shape) {
+    assert_shape_no_negative(ndims, shape);
+}
+
+void __nac3_ndarray_util_assert_shape_no_negative64(int64_t ndims, int64_t* shape) {
+    assert_shape_no_negative(ndims, shape);
+}
+
+void __nac3_ndarray_util_assert_output_shape_same(int32_t ndarray_ndims,
+                                                  const int32_t* ndarray_shape,
+                                                  int32_t output_ndims,
+                                                  const int32_t* output_shape) {
+    assert_output_shape_same(ndarray_ndims, ndarray_shape, output_ndims, output_shape);
+}
+
+void __nac3_ndarray_util_assert_output_shape_same64(int64_t ndarray_ndims,
+                                                    const int64_t* ndarray_shape,
+                                                    int64_t output_ndims,
+                                                    const int64_t* output_shape) {
+    assert_output_shape_same(ndarray_ndims, ndarray_shape, output_ndims, output_shape);
+}
+
+uint32_t __nac3_ndarray_size(NDArray<int32_t>* ndarray) {
+    return size(ndarray);
+}
+
+uint64_t __nac3_ndarray_size64(NDArray<int64_t>* ndarray) {
+    return size(ndarray);
+}
+
+uint32_t __nac3_ndarray_nbytes(NDArray<int32_t>* ndarray) {
+    return nbytes(ndarray);
+}
+
+uint64_t __nac3_ndarray_nbytes64(NDArray<int64_t>* ndarray) {
+    return nbytes(ndarray);
+}
+
+int32_t __nac3_ndarray_len(NDArray<int32_t>* ndarray) {
+    return len(ndarray);
+}
+
+int64_t __nac3_ndarray_len64(NDArray<int64_t>* ndarray) {
+    return len(ndarray);
+}
+
+bool __nac3_ndarray_is_c_contiguous(NDArray<int32_t>* ndarray) {
+    return is_c_contiguous(ndarray);
+}
+
+bool __nac3_ndarray_is_c_contiguous64(NDArray<int64_t>* ndarray) {
+    return is_c_contiguous(ndarray);
+}
+
+uint8_t* __nac3_ndarray_get_nth_pelement(const NDArray<int32_t>* ndarray, int32_t nth) {
+    return get_nth_pelement(ndarray, nth);
+}
+
+uint8_t* __nac3_ndarray_get_nth_pelement64(const NDArray<int64_t>* ndarray, int64_t nth) {
+    return get_nth_pelement(ndarray, nth);
+}
+
+uint8_t* __nac3_ndarray_get_pelement_by_indices(const NDArray<int32_t>* ndarray, int32_t* indices) {
+    return get_pelement_by_indices(ndarray, indices);
+}
+
+uint8_t* __nac3_ndarray_get_pelement_by_indices64(const NDArray<int64_t>* ndarray, int64_t* indices) {
+    return get_pelement_by_indices(ndarray, indices);
+}
+
+void __nac3_ndarray_set_strides_by_shape(NDArray<int32_t>* ndarray) {
+    set_strides_by_shape(ndarray);
+}
+
+void __nac3_ndarray_set_strides_by_shape64(NDArray<int64_t>* ndarray) {
+    set_strides_by_shape(ndarray);
+}
+
+void __nac3_ndarray_copy_data(NDArray<int32_t>* src_ndarray, NDArray<int32_t>* dst_ndarray) {
+    copy_data(src_ndarray, dst_ndarray);
+}
+
+void __nac3_ndarray_copy_data64(NDArray<int64_t>* src_ndarray, NDArray<int64_t>* dst_ndarray) {
+    copy_data(src_ndarray, dst_ndarray);
+}
+}

nac3core/irrt/irrt/ndarray/broadcast.hpp

@@ -0,0 +1,165 @@
+#pragma once
+
+#include "irrt/int_types.hpp"
+#include "irrt/ndarray/def.hpp"
+#include "irrt/slice.hpp"
+
+namespace {
+template<typename SizeT>
+struct ShapeEntry {
+    SizeT ndims;
+    SizeT* shape;
+};
+}  // namespace
+
+namespace {
+namespace ndarray {
+namespace broadcast {
+/**
+ * @brief Return true if `src_shape` can broadcast to `dst_shape`.
+ *
+ * See https://numpy.org/doc/stable/user/basics.broadcasting.html
+ */
+template<typename SizeT>
+bool can_broadcast_shape_to(SizeT target_ndims, const SizeT* target_shape, SizeT src_ndims, const SizeT* src_shape) {
+    if (src_ndims > target_ndims) {
+        return false;
+    }
+
+    for (SizeT i = 0; i < src_ndims; i++) {
+        SizeT target_dim = target_shape[target_ndims - i - 1];
+        SizeT src_dim = src_shape[src_ndims - i - 1];
+        if (!(src_dim == 1 || target_dim == src_dim)) {
+            return false;
+        }
+    }
+    return true;
+}
+
+/**
+ * @brief Performs `np.broadcast_shapes(<shapes>)`
+ *
+ * @param num_shapes Number of entries in `shapes`
+ * @param shapes The list of shape to do `np.broadcast_shapes` on.
+ * @param dst_ndims The length of `dst_shape`.
+ *   `dst_ndims` must be `max([shape.ndims for shape in shapes])`, but the caller has to calculate it/provide it.
+ *   for this function since they should already know in order to allocate `dst_shape` in the first place.
+ * @param dst_shape The resulting shape. Must be pre-allocated by the caller. This function calculate the result
+ *   of `np.broadcast_shapes` and write it here.
+ */
+template<typename SizeT>
+void broadcast_shapes(SizeT num_shapes, const ShapeEntry<SizeT>* shapes, SizeT dst_ndims, SizeT* dst_shape) {
+    for (SizeT dst_axis = 0; dst_axis < dst_ndims; dst_axis++) {
+        dst_shape[dst_axis] = 1;
+    }
+
+#ifdef IRRT_DEBUG_ASSERT
+    SizeT max_ndims_found = 0;
+#endif
+
+    for (SizeT i = 0; i < num_shapes; i++) {
+        ShapeEntry<SizeT> entry = shapes[i];
+
+        // Check pre-condition: `dst_ndims` must be `max([shape.ndims for shape in shapes])`
+        debug_assert(SizeT, entry.ndims <= dst_ndims);
+
+#ifdef IRRT_DEBUG_ASSERT
+        max_ndims_found = max(max_ndims_found, entry.ndims);
+#endif
+
+        for (SizeT j = 0; j < entry.ndims; j++) {
+            SizeT entry_axis = entry.ndims - j - 1;
+            SizeT dst_axis = dst_ndims - j - 1;
+
+            SizeT entry_dim = entry.shape[entry_axis];
+            SizeT dst_dim = dst_shape[dst_axis];
+
+            if (dst_dim == 1) {
+                dst_shape[dst_axis] = entry_dim;
+            } else if (entry_dim == 1 || entry_dim == dst_dim) {
+                // Do nothing
+            } else {
+                raise_exception(SizeT, EXN_VALUE_ERROR,
+                                "shape mismatch: objects cannot be broadcast "
+                                "to a single shape.",
+                                NO_PARAM, NO_PARAM, NO_PARAM);
+            }
+        }
+    }
+
+    // Check pre-condition: `dst_ndims` must be `max([shape.ndims for shape in shapes])`
+    debug_assert_eq(SizeT, max_ndims_found, dst_ndims);
+}
+
+/**
+ * @brief Perform `np.broadcast_to(<ndarray>, <target_shape>)` and appropriate assertions.
+ *
+ * This function attempts to broadcast `src_ndarray` to a new shape defined by `dst_ndarray.shape`,
+ * and return the result by modifying `dst_ndarray`.
+ *
+ * # Notes on `dst_ndarray`
+ * The caller is responsible for allocating space for the resulting ndarray.
+ * Here is what this function expects from `dst_ndarray` when called:
+ *   - `dst_ndarray->data` does not have to be initialized.
+ *   - `dst_ndarray->itemsize` does not have to be initialized.
+ *   - `dst_ndarray->ndims` must be initialized, determining the length of `dst_ndarray->shape`
+ *   - `dst_ndarray->shape` must be allocated, and must contain the desired target broadcast shape.
+ *   - `dst_ndarray->strides` must be allocated, through it can contain uninitialized values.
+ * When this function call ends:
+ *   - `dst_ndarray->data` is set to `src_ndarray->data` (`dst_ndarray` is just a view to `src_ndarray`)
+ *   - `dst_ndarray->itemsize` is set to `src_ndarray->itemsize`
+ *   - `dst_ndarray->ndims` is unchanged.
+ *   - `dst_ndarray->shape` is unchanged.
+ *   - `dst_ndarray->strides` is updated accordingly by how ndarray broadcast_to works.
+ */
+template<typename SizeT>
+void broadcast_to(const NDArray<SizeT>* src_ndarray, NDArray<SizeT>* dst_ndarray) {
+    if (!ndarray::broadcast::can_broadcast_shape_to(dst_ndarray->ndims, dst_ndarray->shape, src_ndarray->ndims,
+                                                    src_ndarray->shape)) {
+        raise_exception(SizeT, EXN_VALUE_ERROR, "operands could not be broadcast together", NO_PARAM, NO_PARAM,
+                        NO_PARAM);
+    }
+
+    dst_ndarray->data = src_ndarray->data;
+    dst_ndarray->itemsize = src_ndarray->itemsize;
+
+    for (SizeT i = 0; i < dst_ndarray->ndims; i++) {
+        SizeT src_axis = src_ndarray->ndims - i - 1;
+        SizeT dst_axis = dst_ndarray->ndims - i - 1;
+        if (src_axis < 0 || (src_ndarray->shape[src_axis] == 1 && dst_ndarray->shape[dst_axis] != 1)) {
+            // Freeze the steps in-place
+            dst_ndarray->strides[dst_axis] = 0;
+        } else {
+            dst_ndarray->strides[dst_axis] = src_ndarray->strides[src_axis];
+        }
+    }
+}
+}  // namespace broadcast
+}  // namespace ndarray
+}  // namespace
+
+extern "C" {
+using namespace ndarray::broadcast;
+
+void __nac3_ndarray_broadcast_to(NDArray<int32_t>* src_ndarray, NDArray<int32_t>* dst_ndarray) {
+    broadcast_to(src_ndarray, dst_ndarray);
+}
+
+void __nac3_ndarray_broadcast_to64(NDArray<int64_t>* src_ndarray, NDArray<int64_t>* dst_ndarray) {
+    broadcast_to(src_ndarray, dst_ndarray);
+}
+
+void __nac3_ndarray_broadcast_shapes(int32_t num_shapes,
+                                     const ShapeEntry<int32_t>* shapes,
+                                     int32_t dst_ndims,
+                                     int32_t* dst_shape) {
+    broadcast_shapes(num_shapes, shapes, dst_ndims, dst_shape);
+}
+
+void __nac3_ndarray_broadcast_shapes64(int64_t num_shapes,
+                                       const ShapeEntry<int64_t>* shapes,
+                                       int64_t dst_ndims,
+                                       int64_t* dst_shape) {
+    broadcast_shapes(num_shapes, shapes, dst_ndims, dst_shape);
+}
+}

nac3core/irrt/irrt/ndarray/def.hpp

@@ -0,0 +1,45 @@
+#pragma once
+
+#include "irrt/int_types.hpp"
+
+namespace {
+/**
+ * @brief The NDArray object
+ *
+ * Official numpy implementation:
+ * 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.
+     */
+    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 or contain 0.
+     */
+    SizeT* strides;
+};
+}  // namespace

nac3core/irrt/irrt/ndarray/indexing.hpp

@@ -0,0 +1,220 @@
+#pragma once
+
+#include "irrt/exception.hpp"
+#include "irrt/int_types.hpp"
+#include "irrt/ndarray/basic.hpp"
+#include "irrt/ndarray/def.hpp"
+#include "irrt/range.hpp"
+#include "irrt/slice.hpp"
+
+namespace {
+typedef uint8_t NDIndexType;
+
+/**
+ * @brief A single element index
+ *
+ * `data` points to a `int32_t`.
+ */
+const NDIndexType ND_INDEX_TYPE_SINGLE_ELEMENT = 0;
+
+/**
+ * @brief A slice index
+ *
+ * `data` points to a `Slice<int32_t>`.
+ */
+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
+ *
+ * That is:
+ * ```
+ * my_ndarray[::-1, 3, ..., np.newaxis]
+ *            ^^^^  ^  ^^^  ^^^^^^^^^^ each of these is represented by an NDIndex.
+ * ```
+ */
+struct NDIndex {
+    /**
+     * @brief Enum tag to specify the type of index.
+     *
+     * Please see the comment of each enum constant.
+     */
+    NDIndexType type;
+
+    /**
+     * @brief The accompanying data associated with `type`.
+     *
+     * Please see the comment 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 function is very similar to performing `dst_ndarray = src_ndarray[indices]` in Python.
+ *
+ * This function also does proper assertions on `indices` to check for out of bounds access and more.
+ *
+ * # Notes on `dst_ndarray`
+ * The caller is responsible for allocating space for the resulting ndarray.
+ * Here is what this function expects from `dst_ndarray` when called:
+ *   - `dst_ndarray->data` does not have to be initialized.
+ *   - `dst_ndarray->itemsize` does not have to be initialized.
+ *   - `dst_ndarray->ndims` must be initialized, and it must be equal to the expected `ndims` of the `dst_ndarray` after
+ *       indexing `src_ndarray` with `indices`.
+ *   - `dst_ndarray->shape` must be allocated, through it can contain uninitialized values.
+ *   - `dst_ndarray->strides` must be allocated, through it can contain uninitialized values.
+ * When this function call ends:
+ *   - `dst_ndarray->data` is set to `src_ndarray->data`.
+ *   - `dst_ndarray->itemsize` is set to `src_ndarray->itemsize`.
+ *   - `dst_ndarray->ndims` is unchanged.
+ *   - `dst_ndarray->shape` is updated according to how `src_ndarray` is indexed.
+ *   - `dst_ndarray->strides` is updated accordingly by how ndarray indexing works.
+ *
+ * @param indices indices to index `src_ndarray`, ordered in the same way you would write them in Python.
+ * @param src_ndarray The NDArray to be indexed.
+ * @param dst_ndarray The resulting NDArray after indexing. Further details in the comments above,
+ */
+template<typename SizeT>
+void index(SizeT num_indices, const NDIndex* indices, const NDArray<SizeT>* src_ndarray, NDArray<SizeT>* dst_ndarray) {
+    // Validate `indices`.
+
+    // Expected value of `dst_ndarray->ndims`.
+    SizeT expected_dst_ndims = src_ndarray->ndims;
+    // To check for "too many indices for array: array is ?-dimensional, but ? were indexed"
+    SizeT num_indexed = 0;
+    // There may be ellipsis `...` in `indices`. There can only be 0 or 1 ellipsis.
+    SizeT num_ellipsis = 0;
+
+    for (SizeT i = 0; i < num_indices; i++) {
+        if (indices[i].type == ND_INDEX_TYPE_SINGLE_ELEMENT) {
+            expected_dst_ndims--;
+            num_indexed++;
+        } else if (indices[i].type == ND_INDEX_TYPE_SLICE) {
+            num_indexed++;
+        } else if (indices[i].type == ND_INDEX_TYPE_NEWAXIS) {
+            expected_dst_ndims++;
+        } else if (indices[i].type == ND_INDEX_TYPE_ELLIPSIS) {
+            num_ellipsis++;
+            if (num_ellipsis > 1) {
+                raise_exception(SizeT, EXN_INDEX_ERROR, "an index can only have a single ellipsis ('...')", NO_PARAM,
+                                NO_PARAM, NO_PARAM);
+            }
+        } else {
+            __builtin_unreachable();
+        }
+    }
+
+    debug_assert_eq(SizeT, expected_dst_ndims, dst_ndarray->ndims);
+
+    if (src_ndarray->ndims - num_indexed < 0) {
+        raise_exception(SizeT, EXN_INDEX_ERROR,
+                        "too many indices for array: array is {0}-dimensional, "
+                        "but {1} were indexed",
+                        src_ndarray->ndims, num_indices, NO_PARAM);
+    }
+
+    dst_ndarray->data = src_ndarray->data;
+    dst_ndarray->itemsize = src_ndarray->itemsize;
+
+    // Reference code:
+    // https://github.com/wadetb/tinynumpy/blob/0d23d22e07062ffab2afa287374c7b366eebdda1/tinynumpy/tinynumpy.py#L652
+    SizeT src_axis = 0;
+    SizeT dst_axis = 0;
+
+    for (int32_t i = 0; i < num_indices; i++) {
+        const NDIndex* index = &indices[i];
+        if (index->type == ND_INDEX_TYPE_SINGLE_ELEMENT) {
+            SizeT input = (SizeT) * ((int32_t*)index->data);
+
+            SizeT k = slice::resolve_index_in_length(src_ndarray->shape[src_axis], input);
+            if (k == -1) {
+                raise_exception(SizeT, EXN_INDEX_ERROR,
+                                "index {0} is out of bounds for axis {1} "
+                                "with size {2}",
+                                input, src_axis, src_ndarray->shape[src_axis]);
+            }
+
+            dst_ndarray->data += k * src_ndarray->strides[src_axis];
+
+            src_axis++;
+        } else if (index->type == ND_INDEX_TYPE_SLICE) {
+            Slice<int32_t>* slice = (Slice<int32_t>*)index->data;
+
+            Range<int32_t> range = slice->indices_checked<SizeT>(src_ndarray->shape[src_axis]);
+
+            dst_ndarray->data += (SizeT)range.start * src_ndarray->strides[src_axis];
+            dst_ndarray->strides[dst_axis] = ((SizeT)range.step) * src_ndarray->strides[src_axis];
+            dst_ndarray->shape[dst_axis] = (SizeT)range.len<SizeT>();
+
+            dst_axis++;
+            src_axis++;
+        } else if (index->type == ND_INDEX_TYPE_NEWAXIS) {
+            dst_ndarray->strides[dst_axis] = 0;
+            dst_ndarray->shape[dst_axis] = 1;
+
+            dst_axis++;
+        } else if (index->type == ND_INDEX_TYPE_ELLIPSIS) {
+            // The number of ':' entries this '...' implies.
+            SizeT ellipsis_size = src_ndarray->ndims - num_indexed;
+
+            for (SizeT j = 0; j < ellipsis_size; j++) {
+                dst_ndarray->strides[dst_axis] = src_ndarray->strides[src_axis];
+                dst_ndarray->shape[dst_axis] = src_ndarray->shape[src_axis];
+
+                dst_axis++;
+                src_axis++;
+            }
+        } else {
+            __builtin_unreachable();
+        }
+    }
+
+    for (; dst_axis < dst_ndarray->ndims; dst_axis++, src_axis++) {
+        dst_ndarray->shape[dst_axis] = src_ndarray->shape[src_axis];
+        dst_ndarray->strides[dst_axis] = src_ndarray->strides[src_axis];
+    }
+
+    debug_assert_eq(SizeT, src_ndarray->ndims, src_axis);
+    debug_assert_eq(SizeT, dst_ndarray->ndims, dst_axis);
+}
+}  // namespace indexing
+}  // namespace ndarray
+}  // namespace
+
+extern "C" {
+using namespace ndarray::indexing;
+
+void __nac3_ndarray_index(int32_t num_indices,
+                          NDIndex* indices,
+                          NDArray<int32_t>* src_ndarray,
+                          NDArray<int32_t>* dst_ndarray) {
+    index(num_indices, indices, src_ndarray, dst_ndarray);
+}
+
+void __nac3_ndarray_index64(int64_t num_indices,
+                            NDIndex* indices,
+                            NDArray<int64_t>* src_ndarray,
+                            NDArray<int64_t>* dst_ndarray) {
+    index(num_indices, indices, src_ndarray, dst_ndarray);
+}
+}

nac3core/irrt/irrt/ndarray/iter.hpp

@@ -0,0 +1,146 @@
+#pragma once
+
+#include "irrt/int_types.hpp"
+#include "irrt/ndarray/def.hpp"
+
+namespace {
+/**
+ * @brief Helper struct to enumerate through an ndarray *efficiently*.
+ *
+ * Example usage (in pseudo-code):
+ * ```
+ * // Suppose my_ndarray has been initialized, with shape [2, 3] and dtype `double`
+ * NDIter nditer;
+ * nditer.initialize(my_ndarray);
+ * while (nditer.has_element()) {
+ *    // This body is run 6 (= my_ndarray.size) times.
+ *
+ *    // [0, 0] -> [0, 1] -> [0, 2] -> [1, 0] -> [1, 1] -> [1, 2] -> end
+ *    print(nditer.indices);
+ *
+ *    // 0 -> 1 -> 2 -> 3 -> 4 -> 5
+ *    print(nditer.nth);
+ *
+ *    // <1st element> -> <2nd element> -> ... -> <6th element> -> end
+ *    print(*((double *) nditer.element))
+ *
+ *    nditer.next(); // Go to next element.
+ * }
+ * ```
+ *
+ * Interesting cases:
+ * - If `my_ndarray.ndims` == 0, there is one iteration.
+ * - If `my_ndarray.shape` contains zeroes, there are no iterations.
+ */
+template<typename SizeT>
+struct NDIter {
+    // Information about the ndarray being iterated over.
+    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.
+     *
+     * Initially this is 0.
+     */
+    SizeT nth;
+
+    /**
+     * @brief Pointer to the current element.
+     *
+     * Initially this points to first element of the ndarray.
+     */
+    uint8_t* element;
+
+    /**
+     * @brief Cache for the product of shape.
+     *
+     * Could be 0 if `shape` has 0s in it.
+     */
+    SizeT size;
+
+    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
+        this->size = 1;
+        for (SizeT i = 0; i < ndims; i++) {
+            this->size *= shape[i];
+        }
+
+        // `indices` starts on all 0s.
+        for (SizeT axis = 0; axis < ndims; axis++)
+            indices[axis] = 0;
+        nth = 0;
+    }
+
+    void initialize_by_ndarray(NDArray<SizeT>* ndarray, SizeT* indices) {
+        // NOTE: ndarray->data is pointing to the first element, and `NDIter`'s `element` should also point to the first
+        // element as well.
+        this->initialize(ndarray->ndims, ndarray->shape, ndarray->strides, ndarray->data, indices);
+    }
+
+    // Is the current iteration valid?
+    // If true, then `element`, `indices` and `nth` contain details about the current element.
+    bool has_element() { return nth < size; }
+
+    // Go to the next element.
+    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: 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.
+                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_element(NDIter<int32_t>* iter) {
+    return iter->has_element();
+}
+
+bool __nac3_nditer_has_element64(NDIter<int64_t>* iter) {
+    return iter->has_element();
+}
+
+void __nac3_nditer_next(NDIter<int32_t>* iter) {
+    iter->next();
+}
+
+void __nac3_nditer_next64(NDIter<int64_t>* iter) {
+    iter->next();
+}
+}

nac3core/irrt/irrt/ndarray/reshape.hpp

@@ -0,0 +1,99 @@
+#pragma once
+
+#include "irrt/exception.hpp"
+#include "irrt/int_types.hpp"
+#include "irrt/ndarray/def.hpp"
+
+namespace {
+namespace ndarray {
+namespace reshape {
+/**
+ * @brief Perform assertions on and resolve unknown dimensions in `new_shape` in `np.reshape(<ndarray>, new_shape)`
+ *
+ * If `new_shape` indeed contains unknown dimensions (specified with `-1`, just like numpy), `new_shape` will be
+ * modified to contain the resolved dimension.
+ *
+ * To perform assertions on and resolve unknown dimensions in `new_shape`, we don't need the actual
+ * `<ndarray>` object itself, but only the `.size` of the `<ndarray>`.
+ *
+ * @param size The `.size` of `<ndarray>`
+ * @param new_ndims Number of elements in `new_shape`
+ * @param new_shape Target shape to reshape to
+ */
+template<typename SizeT>
+void resolve_and_check_new_shape(SizeT size, SizeT new_ndims, SizeT* new_shape) {
+    // Is there a -1 in `new_shape`?
+    bool neg1_exists = false;
+    // Location of -1, only initialized if `neg1_exists` is true
+    SizeT neg1_axis_i;
+    // The computed ndarray size of `new_shape`
+    SizeT new_size = 1;
+
+    for (SizeT axis_i = 0; axis_i < new_ndims; axis_i++) {
+        SizeT dim = new_shape[axis_i];
+        if (dim < 0) {
+            if (dim == -1) {
+                if (neg1_exists) {
+                    // Multiple `-1` found. Throw an error.
+                    raise_exception(SizeT, EXN_VALUE_ERROR, "can only specify one unknown dimension", NO_PARAM,
+                                    NO_PARAM, NO_PARAM);
+                } else {
+                    neg1_exists = true;
+                    neg1_axis_i = axis_i;
+                }
+            } else {
+                // TODO: What? In `np.reshape` any negative dimensions is
+                // treated like its `-1`.
+                //
+                // Try running `np.zeros((3, 4)).reshape((-999, 2))`
+                //
+                // It is not documented by numpy.
+                // Throw an error for now...
+
+                raise_exception(SizeT, EXN_VALUE_ERROR, "Found non -1 negative dimension {0} on axis {1}", dim, axis_i,
+                                NO_PARAM);
+            }
+        } else {
+            new_size *= dim;
+        }
+    }
+
+    bool can_reshape;
+    if (neg1_exists) {
+        // Let `x` be the unknown dimension
+        // Solve `x * <new_size> = <size>`
+        if (new_size == 0 && size == 0) {
+            // `x` has infinitely many solutions
+            can_reshape = false;
+        } else if (new_size == 0 && size != 0) {
+            // `x` has no solutions
+            can_reshape = false;
+        } else if (size % new_size != 0) {
+            // `x` has no integer solutions
+            can_reshape = false;
+        } else {
+            can_reshape = true;
+            new_shape[neg1_axis_i] = size / new_size;  // Resolve dimension
+        }
+    } else {
+        can_reshape = (new_size == size);
+    }
+
+    if (!can_reshape) {
+        raise_exception(SizeT, EXN_VALUE_ERROR, "cannot reshape array of size {0} into given shape", size, NO_PARAM,
+                        NO_PARAM);
+    }
+}
+}  // namespace reshape
+}  // namespace ndarray
+}  // namespace
+
+extern "C" {
+void __nac3_ndarray_reshape_resolve_and_check_new_shape(int32_t size, int32_t new_ndims, int32_t* new_shape) {
+    ndarray::reshape::resolve_and_check_new_shape(size, new_ndims, new_shape);
+}
+
+void __nac3_ndarray_reshape_resolve_and_check_new_shape64(int64_t size, int64_t new_ndims, int64_t* new_shape) {
+    ndarray::reshape::resolve_and_check_new_shape(size, new_ndims, new_shape);
+}
+}

nac3core/irrt/irrt/ndarray/transpose.hpp

@@ -0,0 +1,145 @@
+#pragma once
+
+#include "irrt/debug.hpp"
+#include "irrt/exception.hpp"
+#include "irrt/int_types.hpp"
+#include "irrt/ndarray/def.hpp"
+#include "irrt/slice.hpp"
+
+/*
+ * Notes on `np.transpose(<array>, <axes>)`
+ *
+ * TODO: `axes`, if specified, can actually contain negative indices,
+ * but it is not documented in numpy.
+ *
+ * Supporting it for now.
+ */
+
+namespace {
+namespace ndarray {
+namespace transpose {
+/**
+ * @brief Do assertions on `<axes>` in `np.transpose(<array>, <axes>)`.
+ *
+ * Note that `np.transpose`'s `<axe>` argument is optional. If the argument
+ * is specified but the user, use this function to do assertions on it.
+ *
+ * @param ndims The number of dimensions of `<array>`
+ * @param num_axes Number of elements in `<axes>` as specified by the user.
+ *   This should be equal to `ndims`. If not, a "ValueError: axes don't match array" is thrown.
+ * @param axes The user specified `<axes>`.
+ */
+template<typename SizeT>
+void assert_transpose_axes(SizeT ndims, SizeT num_axes, const SizeT* axes) {
+    if (ndims != num_axes) {
+        raise_exception(SizeT, EXN_VALUE_ERROR, "axes don't match array", NO_PARAM, NO_PARAM, NO_PARAM);
+    }
+
+    // TODO: Optimize this
+    bool* axe_specified = (bool*)__builtin_alloca(sizeof(bool) * ndims);
+    for (SizeT i = 0; i < ndims; i++)
+        axe_specified[i] = false;
+
+    for (SizeT i = 0; i < ndims; i++) {
+        SizeT axis = slice::resolve_index_in_length(ndims, axes[i]);
+        if (axis == -1) {
+            // 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/range.hpp

@@ -0,0 +1,47 @@
+#pragma once
+
+#include "irrt/debug.hpp"
+#include "irrt/int_types.hpp"
+
+namespace {
+namespace range {
+template<typename T>
+T len(T start, T stop, T step) {
+    // Reference:
+    // https://github.com/python/cpython/blob/9dbd12375561a393eaec4b21ee4ac568a407cdb0/Objects/rangeobject.c#L933
+    if (step > 0 && start < stop)
+        return 1 + (stop - 1 - start) / step;
+    else if (step < 0 && start > stop)
+        return 1 + (start - 1 - stop) / (-step);
+    else
+        return 0;
+}
+}  // namespace range
+
+/**
+ * @brief A Python range.
+ */
+template<typename T>
+struct Range {
+    T start;
+    T stop;
+    T step;
+
+    /**
+     * @brief Calculate the `len()` of this range.
+     */
+    template<typename SizeT>
+    T len() {
+        debug_assert(SizeT, step != 0);
+        return range::len(start, stop, step);
+    }
+};
+}  // namespace
+
+extern "C" {
+using namespace range;
+
+SliceIndex __nac3_range_slice_len(const SliceIndex start, const SliceIndex end, const SliceIndex step) {
+    return len(start, end, step);
+}
+}

nac3core/irrt/irrt/slice.hpp

@@ -0,0 +1,156 @@
+#pragma once
+
+#include "irrt/debug.hpp"
+#include "irrt/exception.hpp"
+#include "irrt/int_types.hpp"
+#include "irrt/math_util.hpp"
+#include "irrt/range.hpp"
+
+namespace {
+namespace slice {
+/**
+ * @brief Resolve a possibly negative index in a list of a known length.
+ *
+ * Returns -1 if the resolved index is out of the list's bounds.
+ */
+template<typename T>
+T resolve_index_in_length(T length, T index) {
+    T resolved = index < 0 ? length + index : index;
+    if (0 <= resolved && resolved < length) {
+        return resolved;
+    } else {
+        return -1;
+    }
+}
+
+/**
+ * @brief Resolve a slice as a range.
+ *
+ * This is equivalent to `range(*slice(start, stop, step).indices(length))` in Python.
+ */
+template<typename T>
+void indices(bool start_defined,
+             T start,
+             bool stop_defined,
+             T stop,
+             bool step_defined,
+             T step,
+             T length,
+             T* range_start,
+             T* range_stop,
+             T* range_step) {
+    // Reference: https://github.com/python/cpython/blob/main/Objects/sliceobject.c#L388
+    *range_step = step_defined ? step : 1;
+    bool step_is_negative = *range_step < 0;
+
+    T lower, upper;
+    if (step_is_negative) {
+        lower = -1;
+        upper = length - 1;
+    } else {
+        lower = 0;
+        upper = length;
+    }
+
+    if (start_defined) {
+        *range_start = start < 0 ? max(lower, start + length) : min(upper, start);
+    } else {
+        *range_start = step_is_negative ? upper : lower;
+    }
+
+    if (stop_defined) {
+        *range_stop = stop < 0 ? max(lower, stop + length) : min(upper, stop);
+    } else {
+        *range_stop = step_is_negative ? lower : upper;
+    }
+}
+}  // namespace slice
+
+/**
+ * @brief A Python-like slice with **unresolved** indices.
+ */
+template<typename T>
+struct Slice {
+    bool start_defined;
+    T start;
+
+    bool stop_defined;
+    T stop;
+
+    bool step_defined;
+    T step;
+
+    Slice() { this->reset(); }
+
+    void reset() {
+        this->start_defined = false;
+        this->stop_defined = false;
+        this->step_defined = false;
+    }
+
+    void set_start(T start) {
+        this->start_defined = true;
+        this->start = start;
+    }
+
+    void set_stop(T stop) {
+        this->stop_defined = true;
+        this->stop = stop;
+    }
+
+    void set_step(T step) {
+        this->step_defined = true;
+        this->step = step;
+    }
+
+    /**
+     * @brief Resolve this slice as a range.
+     *
+     * In Python, this would be `range(*slice(start, stop, step).indices(length))`.
+     */
+    template<typename SizeT>
+    Range<T> indices(T length) {
+        // Reference:
+        // https://github.com/python/cpython/blob/main/Objects/sliceobject.c#L388
+        debug_assert(SizeT, length >= 0);
+
+        Range<T> result;
+        slice::indices(start_defined, start, stop_defined, stop, step_defined, step, length, &result.start,
+                       &result.stop, &result.step);
+        return result;
+    }
+
+    /**
+     * @brief Like `.indices()` but with assertions.
+     */
+    template<typename SizeT>
+    Range<T> indices_checked(T length) {
+        // TODO: Switch to `SizeT length`
+
+        if (length < 0) {
+            raise_exception(SizeT, EXN_VALUE_ERROR, "length should not be negative, got {0}", length, NO_PARAM,
+                            NO_PARAM);
+        }
+
+        if (this->step_defined && this->step == 0) {
+            raise_exception(SizeT, EXN_VALUE_ERROR, "slice step cannot be zero", NO_PARAM, NO_PARAM, NO_PARAM);
+        }
+
+        return this->indices<SizeT>(length);
+    }
+};
+}  // namespace
+
+extern "C" {
+SliceIndex __nac3_slice_index_bound(SliceIndex i, const SliceIndex len) {
+    if (i < 0) {
+        i = len + i;
+    }
+    if (i < 0) {
+        return 0;
+    } else if (i > len) {
+        return len;
+    }
+    return i;
+}
+}

nac3core/src/codegen/builtin_fns.rs

@@ -4,11 +4,11 @@ use inkwell::{FloatPredicate, IntPredicate, OptimizationLevel};
 use itertools::Itertools;
 
 use crate::codegen::classes::{
-    ArrayLikeValue, NDArrayValue, ProxyValue, RangeValue, TypedArrayLikeAccessor,
-    UntypedArrayLikeAccessor, UntypedArrayLikeMutator,
+    NDArrayValue, ProxyValue, RangeValue, UntypedArrayLikeAccessor, UntypedArrayLikeMutator,
 };
 use crate::codegen::expr::destructure_range;
 use crate::codegen::irrt::calculate_len_for_slice_range;
+use crate::codegen::macros::codegen_unreachable;
 use crate::codegen::numpy::ndarray_elementwise_unaryop_impl;
 use crate::codegen::stmt::gen_for_callback_incrementing;
 use crate::codegen::{extern_fns, irrt, llvm_intrinsics, numpy, CodeGenContext, CodeGenerator};
@@ -16,11 +16,18 @@ use crate::toplevel::helper::PrimDef;
 use crate::toplevel::numpy::unpack_ndarray_var_tys;
 use crate::typecheck::typedef::{Type, TypeEnum};
 
+use super::model::*;
+use super::object::any::AnyObject;
+use super::object::list::ListObject;
+use super::object::ndarray::NDArrayObject;
+use super::object::tuple::TupleObject;
+
 /// Shorthand for [`unreachable!()`] when a type of argument is not supported.
 ///
 /// The generated message will contain the function name and the name of the unsupported type.
 fn unsupported_type(ctx: &CodeGenContext<'_, '_>, fn_name: &str, tys: &[Type]) -> ! {
-    unreachable!(
+    codegen_unreachable!(
+        ctx,
         "{fn_name}() not supported for '{}'",
         tys.iter().map(|ty| format!("'{}'", ctx.unifier.stringify(*ty))).join(", "),
     )
@@ -32,58 +39,33 @@ pub fn call_len<'ctx, G: CodeGenerator + ?Sized>(
     ctx: &mut CodeGenContext<'ctx, '_>,
     n: (Type, BasicValueEnum<'ctx>),
 ) -> Result<IntValue<'ctx>, String> {
-    let llvm_i32 = ctx.ctx.i32_type();
-    let range_ty = ctx.primitives.range;
     let (arg_ty, arg) = n;
-
-    Ok(if ctx.unifier.unioned(arg_ty, range_ty) {
+    Ok(if ctx.unifier.unioned(arg_ty, ctx.primitives.range) {
         let arg = RangeValue::from_ptr_val(arg.into_pointer_value(), Some("range"));
         let (start, end, step) = destructure_range(ctx, arg);
         calculate_len_for_slice_range(generator, ctx, start, end, step)
     } else {
-        match &*ctx.unifier.get_ty_immutable(arg_ty) {
-            TypeEnum::TTuple { ty, .. } => llvm_i32.const_int(ty.len() as u64, false),
-            TypeEnum::TObj { obj_id, .. } if *obj_id == PrimDef::List.id() => {
-                let zero = llvm_i32.const_zero();
-                let len = ctx
-                    .build_gep_and_load(
-                        arg.into_pointer_value(),
-                        &[zero, llvm_i32.const_int(1, false)],
-                        None,
-                    )
-                    .into_int_value();
-                ctx.builder.build_int_truncate_or_bit_cast(len, llvm_i32, "len").unwrap()
+        let arg = AnyObject { ty: arg_ty, value: arg };
+        let len: Instance<'ctx, Int<Int32>> = match &*ctx.unifier.get_ty(arg_ty) {
+            TypeEnum::TTuple { .. } => {
+                let tuple = TupleObject::from_object(ctx, arg);
+                tuple.len(generator, ctx).truncate_or_bit_cast(generator, ctx, Int32)
             }
-            TypeEnum::TObj { obj_id, .. } if *obj_id == PrimDef::NDArray.id() => {
-                let llvm_usize = generator.get_size_type(ctx.ctx);
-
-                let arg = NDArrayValue::from_ptr_val(arg.into_pointer_value(), llvm_usize, None);
-
-                let ndims = arg.dim_sizes().size(ctx, generator);
-                ctx.make_assert(
-                    generator,
-                    ctx.builder
-                        .build_int_compare(IntPredicate::NE, ndims, llvm_usize.const_zero(), "")
-                        .unwrap(),
-                    "0:TypeError",
-                    "len() of unsized object",
-                    [None, None, None],
-                    ctx.current_loc,
-                );
-
-                let len = unsafe {
-                    arg.dim_sizes().get_typed_unchecked(
-                        ctx,
-                        generator,
-                        &llvm_usize.const_zero(),
-                        None,
-                    )
-                };
-
-                ctx.builder.build_int_truncate_or_bit_cast(len, llvm_i32, "len").unwrap()
+            TypeEnum::TObj { obj_id, .. }
+                if *obj_id == ctx.primitives.ndarray.obj_id(&ctx.unifier).unwrap() =>
+            {
+                let ndarray = NDArrayObject::from_object(generator, ctx, arg);
+                ndarray.len(generator, ctx).truncate_or_bit_cast(generator, ctx, Int32)
             }
-            _ => unreachable!(),
-        }
+            TypeEnum::TObj { obj_id, .. }
+                if *obj_id == ctx.primitives.list.obj_id(&ctx.unifier).unwrap() =>
+            {
+                let list = ListObject::from_object(generator, ctx, arg);
+                list.len(generator, ctx).truncate_or_bit_cast(generator, ctx, Int32)
+            }
+            _ => unsupported_type(ctx, "len", &[arg_ty]),
+        };
+        len.value
     })
 }
 
@@ -784,7 +766,7 @@ pub fn call_numpy_minimum<'ctx, G: CodeGenerator + ?Sized>(
             } else if is_ndarray2 {
                 unpack_ndarray_var_tys(&mut ctx.unifier, x2_ty).0
             } else {
-                unreachable!()
+                codegen_unreachable!(ctx)
             };
 
             let x1_scalar_ty = if is_ndarray1 { dtype } else { x1_ty };
@@ -888,7 +870,7 @@ pub fn call_numpy_max_min<'ctx, G: CodeGenerator + ?Sized>(
             match fn_name {
                 "np_argmin" | "np_argmax" => llvm_int64.const_zero().into(),
                 "np_max" | "np_min" => a,
-                _ => unreachable!(),
+                _ => codegen_unreachable!(ctx),
             }
         }
         BasicValueEnum::PointerValue(n)
@@ -943,7 +925,7 @@ pub fn call_numpy_max_min<'ctx, G: CodeGenerator + ?Sized>(
                         "np_argmax" | "np_max" => {
                             call_max(ctx, (elem_ty, accumulator), (elem_ty, elem))
                         }
-                        _ => unreachable!(),
+                        _ => codegen_unreachable!(ctx),
                     };
 
                     let updated_idx = match (accumulator, result) {
@@ -980,7 +962,7 @@ pub fn call_numpy_max_min<'ctx, G: CodeGenerator + ?Sized>(
             match fn_name {
                 "np_argmin" | "np_argmax" => ctx.builder.build_load(res_idx, "").unwrap(),
                 "np_max" | "np_min" => ctx.builder.build_load(accumulator_addr, "").unwrap(),
-                _ => unreachable!(),
+                _ => codegen_unreachable!(ctx),
             }
         }
 
@@ -1046,7 +1028,7 @@ pub fn call_numpy_maximum<'ctx, G: CodeGenerator + ?Sized>(
             } else if is_ndarray2 {
                 unpack_ndarray_var_tys(&mut ctx.unifier, x2_ty).0
             } else {
-                unreachable!()
+                codegen_unreachable!(ctx)
             };
 
             let x1_scalar_ty = if is_ndarray1 { dtype } else { x1_ty };
@@ -1486,7 +1468,7 @@ pub fn call_numpy_arctan2<'ctx, G: CodeGenerator + ?Sized>(
             } else if is_ndarray2 {
                 unpack_ndarray_var_tys(&mut ctx.unifier, x2_ty).0
             } else {
-                unreachable!()
+                codegen_unreachable!(ctx)
             };
 
             let x1_scalar_ty = if is_ndarray1 { dtype } else { x1_ty };
@@ -1553,7 +1535,7 @@ pub fn call_numpy_copysign<'ctx, G: CodeGenerator + ?Sized>(
             } else if is_ndarray2 {
                 unpack_ndarray_var_tys(&mut ctx.unifier, x2_ty).0
             } else {
-                unreachable!()
+                codegen_unreachable!(ctx)
             };
 
             let x1_scalar_ty = if is_ndarray1 { dtype } else { x1_ty };
@@ -1620,7 +1602,7 @@ pub fn call_numpy_fmax<'ctx, G: CodeGenerator + ?Sized>(
             } else if is_ndarray2 {
                 unpack_ndarray_var_tys(&mut ctx.unifier, x2_ty).0
             } else {
-                unreachable!()
+                codegen_unreachable!(ctx)
             };
 
             let x1_scalar_ty = if is_ndarray1 { dtype } else { x1_ty };
@@ -1687,7 +1669,7 @@ pub fn call_numpy_fmin<'ctx, G: CodeGenerator + ?Sized>(
             } else if is_ndarray2 {
                 unpack_ndarray_var_tys(&mut ctx.unifier, x2_ty).0
             } else {
-                unreachable!()
+                codegen_unreachable!(ctx)
             };
 
             let x1_scalar_ty = if is_ndarray1 { dtype } else { x1_ty };
@@ -1810,7 +1792,7 @@ pub fn call_numpy_hypot<'ctx, G: CodeGenerator + ?Sized>(
             } else if is_ndarray2 {
                 unpack_ndarray_var_tys(&mut ctx.unifier, x2_ty).0
             } else {
-                unreachable!()
+                codegen_unreachable!(ctx)
             };
 
             let x1_scalar_ty = if is_ndarray1 { dtype } else { x1_ty };
@@ -1877,7 +1859,7 @@ pub fn call_numpy_nextafter<'ctx, G: CodeGenerator + ?Sized>(
             } else if is_ndarray2 {
                 unpack_ndarray_var_tys(&mut ctx.unifier, x2_ty).0
             } else {
-                unreachable!()
+                codegen_unreachable!(ctx)
             };
 
             let x1_scalar_ty = if is_ndarray1 { dtype } else { x1_ty };

nac3core/src/codegen/classes.rs

@@ -1404,7 +1404,7 @@ impl<'ctx> NDArrayValue<'ctx> {
 
     /// Returns the double-indirection pointer to the `data` array, as if by calling `getelementptr`
     /// on the field.
-    fn ptr_to_data(&self, ctx: &CodeGenContext<'ctx, '_>) -> PointerValue<'ctx> {
+    pub fn ptr_to_data(&self, ctx: &CodeGenContext<'ctx, '_>) -> PointerValue<'ctx> {
         let llvm_i32 = ctx.ctx.i32_type();
         let var_name = self.name.map(|v| format!("{v}.data.addr")).unwrap_or_default();
 

nac3core/src/codegen/irrt/irrt.cpp

@@ -1,414 +0,0 @@
-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);
-
-// NDArray indices are always `uint32_t`.
-using NDIndex = uint32_t;
-// The type of an index or a value describing the length of a range/slice is always `int32_t`.
-using SliceIndex = int32_t;
-
-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;
-}
-
-// 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) {
-    T res = 1;
-    /* repeated squaring method */
-    do {
-        if (exp & 1) {
-            res *= base; /* for n odd */
-        }
-        exp >>= 1;
-        base *= base;
-    } while (exp);
-    return res;
-}
-
-template <typename SizeT>
-SizeT __nac3_ndarray_calc_size_impl(
-    const SizeT* list_data,
-    SizeT list_len,
-    SizeT begin_idx,
-    SizeT end_idx
-) {
-    __builtin_assume(end_idx <= list_len);
-
-    SizeT num_elems = 1;
-    for (SizeT i = begin_idx; i < end_idx; ++i) {
-        SizeT val = list_data[i];
-        __builtin_assume(val > 0);
-        num_elems *= val;
-    }
-    return num_elems;
-}
-
-template <typename SizeT>
-void __nac3_ndarray_calc_nd_indices_impl(
-    SizeT index,
-    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;
-        __builtin_assume(dims[i] > 0);
-        idxs[i] = (index / stride) % dims[i];
-        stride *= dims[i];
-    }
-}
-
-template <typename SizeT>
-SizeT __nac3_ndarray_flatten_index_impl(
-    const SizeT* dims,
-    SizeT num_dims,
-    const NDIndex* indices,
-    SizeT num_indices
-) {
-    SizeT idx = 0;
-    SizeT stride = 1;
-    for (SizeT i = 0; i < num_dims; ++i) {
-        SizeT ri = num_dims - i - 1;
-        if (ri < num_indices) {
-            idx += stride * indices[ri];
-        }
-
-        __builtin_assume(dims[i] > 0);
-        stride *= dims[ri];
-    }
-    return idx;
-}
-
-template <typename SizeT>
-void __nac3_ndarray_calc_broadcast_impl(
-    const SizeT* lhs_dims,
-    SizeT lhs_ndims,
-    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* 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) {
-            *out_dim = *rhs_dim_sz;
-        } else if (rhs_dim_sz == nullptr) {
-            *out_dim = *lhs_dim_sz;
-        } else if (*lhs_dim_sz == 1) {
-            *out_dim = *rhs_dim_sz;
-        } else if (*rhs_dim_sz == 1) {
-            *out_dim = *lhs_dim_sz;
-        } else if (*lhs_dim_sz == *rhs_dim_sz) {
-            *out_dim = *lhs_dim_sz;
-        } else {
-            __builtin_unreachable();
-        }
-    }
-}
-
-template <typename SizeT>
-void __nac3_ndarray_calc_broadcast_idx_impl(
-    const SizeT* src_dims,
-    SizeT src_ndims,
-    const NDIndex* in_idx,
-    NDIndex* 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];
-    }
-}
-}  // namespace
-
-extern "C" {
-#define DEF_nac3_int_exp_(T) \
-    T __nac3_int_exp_##T(T base, T exp) {\
-        return __nac3_int_exp_impl(base, exp);\
-    }
-
-DEF_nac3_int_exp_(int32_t)
-DEF_nac3_int_exp_(int64_t)
-DEF_nac3_int_exp_(uint32_t)
-DEF_nac3_int_exp_(uint64_t)
-
-SliceIndex __nac3_slice_index_bound(SliceIndex i, const SliceIndex len) {
-    if (i < 0) {
-        i = len + i;
-    }
-    if (i < 0) {
-        return 0;
-    } else if (i > len) {
-        return len;
-    }
-    return i;
-}
-
-SliceIndex __nac3_range_slice_len(
-    const SliceIndex start,
-    const SliceIndex end,
-    const SliceIndex step
-) {
-    SliceIndex diff = end - 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;
-    }
-}
-
-// Handle list assignment and dropping part of the list when
-// both dest_step and src_step are +1.
-// - 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)
-SliceIndex __nac3_list_slice_assign_var_size(
-    SliceIndex dest_start,
-    SliceIndex dest_end,
-    SliceIndex dest_step,
-    uint8_t* dest_arr,
-    SliceIndex dest_arr_len,
-    SliceIndex src_start,
-    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 (src_step == dest_step && dest_step == 1) {
-        const SliceIndex src_len = (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(
-                dest_arr + dest_start * size,
-                src_arr + src_start * size,
-                src_len * size
-            );
-        }
-        if (dest_len > 0) {
-            /* dropping */
-            __builtin_memmove(
-                dest_arr + (dest_start + src_len) * size,
-                dest_arr + (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)
-        && !(
-            max(dest_start, dest_end) < min(src_start, src_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));
-        __builtin_memcpy(tmp, src_arr, src_arr_len * size);
-        src_arr = tmp;
-    }
-    SliceIndex src_ind = src_start;
-    SliceIndex dest_ind = dest_start;
-    for (;
-        (src_step > 0) ? (src_ind <= src_end) : (src_ind >= src_end);
-        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);
-        } else if (size == 4) {
-            __builtin_memcpy(dest_arr + dest_ind * 4, src_arr + src_ind * 4, 4);
-        } 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 */
-            __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_end + 1) * 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) {
-    return __builtin_isinf(x);
-}
-
-int32_t __nac3_isnan(double x) {
-    return __builtin_isnan(x);
-}
-
-double tgamma(double arg);
-
-double __nac3_gamma(double z) {
-    // Handling for denormals
-    //     | x                 | Python gamma(x) | C tgamma(x) |
-    // --- | ----------------- | --------------- | ----------- |
-    // (1) | nan               | nan             | nan         |
-    // (2) | -inf              | -inf            | inf         |
-    // (3) | inf               | inf             | inf         |
-    // (4) | 0.0               | inf             | inf         |
-    // (5) | {-1.0, -2.0, ...} | inf             | nan         |
-
-    // (1)-(3)
-    if (__builtin_isinf(z) || __builtin_isnan(z)) {
-        return z;
-    }
-
-    double v = tgamma(z);
-
-    // (4)-(5)
-    return __builtin_isinf(v) || __builtin_isnan(v) ? __builtin_inf() : v;
-}
-
-double lgamma(double arg);
-
-double __nac3_gammaln(double x) {
-    // libm's handling of value overflows differs from scipy:
-    // - scipy: gammaln(-inf) -> -inf
-    // - libm : lgamma(-inf) -> inf
-
-    if (__builtin_isinf(x)) {
-        return x;
-    }
-
-    return lgamma(x);
-}
-
-double j0(double x);
-
-double __nac3_j0(double x) {
-    // libm's handling of value overflows differs from scipy:
-    // - scipy: j0(inf) -> nan
-    // - libm : j0(inf) -> 0.0
-
-    if (__builtin_isinf(x)) {
-        return __builtin_nan("");
-    }
-
-    return j0(x);
-}
-
-uint32_t __nac3_ndarray_calc_size(
-    const uint32_t* list_data,
-    uint32_t list_len,
-    uint32_t begin_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(
-    const uint64_t* list_data,
-    uint64_t list_len,
-    uint64_t begin_idx,
-    uint64_t end_idx
-) {
-    return __nac3_ndarray_calc_size_impl(list_data, list_len, begin_idx, end_idx);
-}
-
-void __nac3_ndarray_calc_nd_indices(
-    uint32_t index,
-    const uint32_t* dims,
-    uint32_t num_dims,
-    NDIndex* idxs
-) {
-    __nac3_ndarray_calc_nd_indices_impl(index, dims, num_dims, idxs);
-}
-
-void __nac3_ndarray_calc_nd_indices64(
-    uint64_t index,
-    const uint64_t* dims,
-    uint64_t num_dims,
-    NDIndex* idxs
-) {
-    __nac3_ndarray_calc_nd_indices_impl(index, dims, num_dims, idxs);
-}
-
-uint32_t __nac3_ndarray_flatten_index(
-    const uint32_t* dims,
-    uint32_t num_dims,
-    const NDIndex* indices,
-    uint32_t num_indices
-) {
-    return __nac3_ndarray_flatten_index_impl(dims, num_dims, indices, num_indices);
-}
-
-uint64_t __nac3_ndarray_flatten_index64(
-    const uint64_t* dims,
-    uint64_t num_dims,
-    const NDIndex* indices,
-    uint64_t num_indices
-) {
-    return __nac3_ndarray_flatten_index_impl(dims, num_dims, indices, num_indices);
-}
-
-void __nac3_ndarray_calc_broadcast(
-    const uint32_t* lhs_dims,
-    uint32_t lhs_ndims,
-    const uint32_t* rhs_dims,
-    uint32_t rhs_ndims,
-    uint32_t* out_dims
-) {
-    return __nac3_ndarray_calc_broadcast_impl(lhs_dims, lhs_ndims, rhs_dims, rhs_ndims, out_dims);
-}
-
-void __nac3_ndarray_calc_broadcast64(
-    const uint64_t* lhs_dims,
-    uint64_t lhs_ndims,
-    const uint64_t* rhs_dims,
-    uint64_t rhs_ndims,
-    uint64_t* out_dims
-) {
-    return __nac3_ndarray_calc_broadcast_impl(lhs_dims, lhs_ndims, rhs_dims, rhs_ndims, out_dims);
-}
-
-void __nac3_ndarray_calc_broadcast_idx(
-    const uint32_t* src_dims,
-    uint32_t src_ndims,
-    const NDIndex* in_idx,
-    NDIndex* out_idx
-) {
-    __nac3_ndarray_calc_broadcast_idx_impl(src_dims, src_ndims, in_idx, out_idx);
-}
-
-void __nac3_ndarray_calc_broadcast_idx64(
-    const uint64_t* src_dims,
-    uint64_t src_ndims,
-    const NDIndex* in_idx,
-    NDIndex* out_idx
-) {
-    __nac3_ndarray_calc_broadcast_idx_impl(src_dims, src_ndims, in_idx, out_idx);
-}
-}  // extern "C"

nac3core/src/codegen/irrt/mod.rs

@@ -1,28 +1,35 @@
-use crate::typecheck::typedef::Type;
+use crate::{symbol_resolver::SymbolResolver, typecheck::typedef::Type};
 
 use super::{
     classes::{
         ArrayLikeIndexer, ArrayLikeValue, ArraySliceValue, ListValue, NDArrayValue,
-        TypedArrayLikeAdapter, UntypedArrayLikeAccessor,
+        TypedArrayLikeAccessor, TypedArrayLikeAdapter, UntypedArrayLikeAccessor,
     },
-    llvm_intrinsics, CodeGenContext, CodeGenerator,
+    llvm_intrinsics,
+    macros::codegen_unreachable,
+    model::*,
+    object::{
+        list::List,
+        ndarray::{broadcast::ShapeEntry, indexing::NDIndex, nditer::NDIter, NDArray},
+    },
+    stmt::gen_for_callback_incrementing,
+    CodeGenContext, CodeGenerator,
 };
-use crate::codegen::classes::TypedArrayLikeAccessor;
-use crate::codegen::stmt::gen_for_callback_incrementing;
+use function::FnCall;
 use inkwell::{
     attributes::{Attribute, AttributeLoc},
     context::Context,
     memory_buffer::MemoryBuffer,
     module::Module,
     types::{BasicTypeEnum, IntType},
-    values::{BasicValueEnum, CallSiteValue, FloatValue, IntValue},
+    values::{BasicValue, BasicValueEnum, CallSiteValue, FloatValue, IntValue},
     AddressSpace, IntPredicate,
 };
 use itertools::Either;
 use nac3parser::ast::Expr;
 
 #[must_use]
-pub fn load_irrt(ctx: &Context) -> Module {
+pub fn load_irrt<'ctx>(ctx: &'ctx Context, symbol_resolver: &dyn SymbolResolver) -> Module<'ctx> {
     let bitcode_buf = MemoryBuffer::create_from_memory_range(
         include_bytes!(concat!(env!("OUT_DIR"), "/irrt.bc")),
         "irrt_bitcode_buffer",
@@ -38,6 +45,25 @@ pub fn load_irrt(ctx: &Context) -> Module {
         let function = irrt_mod.get_function(symbol).unwrap();
         function.add_attribute(AttributeLoc::Function, ctx.create_enum_attribute(inline_attr, 0));
     }
+
+    // Initialize all global `EXN_*` exception IDs in IRRT with the [`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_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 = irrt_mod.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);
+    }
+
     irrt_mod
 }
 
@@ -55,7 +81,7 @@ pub fn integer_power<'ctx, G: CodeGenerator + ?Sized>(
         (64, 64, true) => "__nac3_int_exp_int64_t",
         (32, 32, false) => "__nac3_int_exp_uint32_t",
         (64, 64, false) => "__nac3_int_exp_uint64_t",
-        _ => unreachable!(),
+        _ => codegen_unreachable!(ctx),
     };
     let base_type = base.get_type();
     let pow_fun = ctx.module.get_function(symbol).unwrap_or_else(|| {
@@ -441,7 +467,7 @@ pub fn list_slice_assignment<'ctx, G: CodeGenerator + ?Sized>(
                     BasicTypeEnum::IntType(t) => t.size_of(),
                     BasicTypeEnum::PointerType(t) => t.size_of(),
                     BasicTypeEnum::StructType(t) => t.size_of().unwrap(),
-                    _ => unreachable!(),
+                    _ => codegen_unreachable!(ctx),
                 };
                 ctx.builder.build_int_truncate_or_bit_cast(s, int32, "size").unwrap()
             }
@@ -586,7 +612,7 @@ where
     let ndarray_calc_size_fn_name = match llvm_usize.get_bit_width() {
         32 => "__nac3_ndarray_calc_size",
         64 => "__nac3_ndarray_calc_size64",
-        bw => unreachable!("Unsupported size type bit width: {}", bw),
+        bw => codegen_unreachable!(ctx, "Unsupported size type bit width: {}", bw),
     };
     let ndarray_calc_size_fn_t = llvm_usize.fn_type(
         &[llvm_pusize.into(), llvm_usize.into(), llvm_usize.into(), llvm_usize.into()],
@@ -637,7 +663,7 @@ pub fn call_ndarray_calc_nd_indices<'ctx, G: CodeGenerator + ?Sized>(
     let ndarray_calc_nd_indices_fn_name = match llvm_usize.get_bit_width() {
         32 => "__nac3_ndarray_calc_nd_indices",
         64 => "__nac3_ndarray_calc_nd_indices64",
-        bw => unreachable!("Unsupported size type bit width: {}", bw),
+        bw => codegen_unreachable!(ctx, "Unsupported size type bit width: {}", bw),
     };
     let ndarray_calc_nd_indices_fn =
         ctx.module.get_function(ndarray_calc_nd_indices_fn_name).unwrap_or_else(|| {
@@ -706,7 +732,7 @@ where
     let ndarray_flatten_index_fn_name = match llvm_usize.get_bit_width() {
         32 => "__nac3_ndarray_flatten_index",
         64 => "__nac3_ndarray_flatten_index64",
-        bw => unreachable!("Unsupported size type bit width: {}", bw),
+        bw => codegen_unreachable!(ctx, "Unsupported size type bit width: {}", bw),
     };
     let ndarray_flatten_index_fn =
         ctx.module.get_function(ndarray_flatten_index_fn_name).unwrap_or_else(|| {
@@ -774,7 +800,7 @@ pub fn call_ndarray_calc_broadcast<'ctx, G: CodeGenerator + ?Sized>(
     let ndarray_calc_broadcast_fn_name = match llvm_usize.get_bit_width() {
         32 => "__nac3_ndarray_calc_broadcast",
         64 => "__nac3_ndarray_calc_broadcast64",
-        bw => unreachable!("Unsupported size type bit width: {}", bw),
+        bw => codegen_unreachable!(ctx, "Unsupported size type bit width: {}", bw),
     };
     let ndarray_calc_broadcast_fn =
         ctx.module.get_function(ndarray_calc_broadcast_fn_name).unwrap_or_else(|| {
@@ -894,7 +920,7 @@ pub fn call_ndarray_calc_broadcast_index<
     let ndarray_calc_broadcast_fn_name = match llvm_usize.get_bit_width() {
         32 => "__nac3_ndarray_calc_broadcast_idx",
         64 => "__nac3_ndarray_calc_broadcast_idx64",
-        bw => unreachable!("Unsupported size type bit width: {}", bw),
+        bw => codegen_unreachable!(ctx, "Unsupported size type bit width: {}", bw),
     };
     let ndarray_calc_broadcast_fn =
         ctx.module.get_function(ndarray_calc_broadcast_fn_name).unwrap_or_else(|| {
@@ -929,3 +955,268 @@ pub fn call_ndarray_calc_broadcast_index<
         Box::new(|_, v| v.into()),
     )
 }
+
+// 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
+}
+
+pub fn call_nac3_ndarray_util_assert_shape_no_negative<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndims: Instance<'ctx, Int<SizeT>>,
+    shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    let name = get_sizet_dependent_function_name(
+        generator,
+        ctx,
+        "__nac3_ndarray_util_assert_shape_no_negative",
+    );
+    FnCall::builder(generator, ctx, &name).arg(ndims).arg(shape).returning_void();
+}
+
+pub fn call_nac3_ndarray_util_assert_output_shape_same<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray_ndims: Instance<'ctx, Int<SizeT>>,
+    ndarray_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    output_ndims: Instance<'ctx, Int<SizeT>>,
+    output_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    let name = get_sizet_dependent_function_name(
+        generator,
+        ctx,
+        "__nac3_ndarray_util_assert_output_shape_same",
+    );
+    FnCall::builder(generator, ctx, &name)
+        .arg(ndarray_ndims)
+        .arg(ndarray_shape)
+        .arg(output_ndims)
+        .arg(output_shape)
+        .returning_void();
+}
+
+pub fn call_nac3_ndarray_size<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) -> Instance<'ctx, Int<SizeT>> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_size");
+    FnCall::builder(generator, ctx, &name).arg(ndarray).returning_auto("size")
+}
+
+pub fn call_nac3_ndarray_nbytes<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) -> Instance<'ctx, Int<SizeT>> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_nbytes");
+    FnCall::builder(generator, ctx, &name).arg(ndarray).returning_auto("nbytes")
+}
+
+pub fn call_nac3_ndarray_len<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) -> Instance<'ctx, Int<SizeT>> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_len");
+    FnCall::builder(generator, ctx, &name).arg(ndarray).returning_auto("len")
+}
+
+pub fn call_nac3_ndarray_is_c_contiguous<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) -> Instance<'ctx, Int<Bool>> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_is_c_contiguous");
+    FnCall::builder(generator, ctx, &name).arg(ndarray).returning_auto("is_c_contiguous")
+}
+
+pub fn call_nac3_ndarray_get_nth_pelement<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    index: Instance<'ctx, Int<SizeT>>,
+) -> Instance<'ctx, Ptr<Int<Byte>>> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_get_nth_pelement");
+    FnCall::builder(generator, ctx, &name).arg(ndarray).arg(index).returning_auto("pelement")
+}
+
+pub fn call_nac3_ndarray_get_pelement_by_indices<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    indices: Instance<'ctx, Ptr<Int<SizeT>>>,
+) -> Instance<'ctx, Ptr<Int<Byte>>> {
+    let name =
+        get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_get_pelement_by_indices");
+    FnCall::builder(generator, ctx, &name).arg(ndarray).arg(indices).returning_auto("pelement")
+}
+
+pub fn call_nac3_ndarray_set_strides_by_shape<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) {
+    let name =
+        get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_set_strides_by_shape");
+    FnCall::builder(generator, ctx, &name).arg(ndarray).returning_void();
+}
+
+pub fn call_nac3_ndarray_copy_data<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    src_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    dst_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_copy_data");
+    FnCall::builder(generator, ctx, &name).arg(src_ndarray).arg(dst_ndarray).returning_void();
+}
+
+pub fn call_nac3_nditer_initialize<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    iter: Instance<'ctx, Ptr<Struct<NDIter>>>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    indices: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_nditer_initialize");
+    FnCall::builder(generator, ctx, &name).arg(iter).arg(ndarray).arg(indices).returning_void();
+}
+
+pub fn call_nac3_nditer_has_element<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    iter: Instance<'ctx, Ptr<Struct<NDIter>>>,
+) -> Instance<'ctx, Int<Bool>> {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_nditer_has_element");
+    FnCall::builder(generator, ctx, &name).arg(iter).returning_auto("has_element")
+}
+
+pub fn call_nac3_nditer_next<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    iter: Instance<'ctx, Ptr<Struct<NDIter>>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_nditer_next");
+    FnCall::builder(generator, ctx, &name).arg(iter).returning_void();
+}
+
+pub fn call_nac3_ndarray_index<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    num_indices: Instance<'ctx, Int<SizeT>>,
+    indices: Instance<'ctx, Ptr<Struct<NDIndex>>>,
+    src_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    dst_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_index");
+    FnCall::builder(generator, ctx, &name)
+        .arg(num_indices)
+        .arg(indices)
+        .arg(src_ndarray)
+        .arg(dst_ndarray)
+        .returning_void();
+}
+
+pub fn call_nac3_ndarray_array_set_and_validate_list_shape<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    list: Instance<'ctx, Ptr<Struct<List<Int<Byte>>>>>,
+    ndims: Instance<'ctx, Int<SizeT>>,
+    shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    let name = get_sizet_dependent_function_name(
+        generator,
+        ctx,
+        "__nac3_ndarray_array_set_and_validate_list_shape",
+    );
+    FnCall::builder(generator, ctx, &name).arg(list).arg(ndims).arg(shape).returning_void();
+}
+
+pub fn call_nac3_ndarray_array_write_list_to_array<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    list: Instance<'ctx, Ptr<Struct<List<Int<Byte>>>>>,
+    ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) {
+    let name = get_sizet_dependent_function_name(
+        generator,
+        ctx,
+        "__nac3_ndarray_array_write_list_to_array",
+    );
+    FnCall::builder(generator, ctx, &name).arg(list).arg(ndarray).returning_void();
+}
+
+pub fn call_nac3_ndarray_reshape_resolve_and_check_new_shape<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    size: Instance<'ctx, Int<SizeT>>,
+    new_ndims: Instance<'ctx, Int<SizeT>>,
+    new_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    let name = get_sizet_dependent_function_name(
+        generator,
+        ctx,
+        "__nac3_ndarray_reshape_resolve_and_check_new_shape",
+    );
+    FnCall::builder(generator, ctx, &name).arg(size).arg(new_ndims).arg(new_shape).returning_void();
+}
+
+pub fn call_nac3_ndarray_broadcast_to<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    src_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    dst_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_broadcast_to");
+    FnCall::builder(generator, ctx, &name).arg(src_ndarray).arg(dst_ndarray).returning_void();
+}
+
+pub fn call_nac3_ndarray_broadcast_shapes<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    num_shape_entries: Instance<'ctx, Int<SizeT>>,
+    shape_entries: Instance<'ctx, Ptr<Struct<ShapeEntry>>>,
+    dst_ndims: Instance<'ctx, Int<SizeT>>,
+    dst_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_broadcast_shapes");
+    FnCall::builder(generator, ctx, &name)
+        .arg(num_shape_entries)
+        .arg(shape_entries)
+        .arg(dst_ndims)
+        .arg(dst_shape)
+        .returning_void();
+}
+
+pub fn call_nac3_ndarray_transpose<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    src_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    dst_ndarray: Instance<'ctx, Ptr<Struct<NDArray>>>,
+    num_axes: Instance<'ctx, Int<SizeT>>,
+    axes: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_transpose");
+    FnCall::builder(generator, ctx, &name)
+        .arg(src_ndarray)
+        .arg(dst_ndarray)
+        .arg(num_axes)
+        .arg(axes)
+        .returning_void();
+}

nac3core/src/codegen/mod.rs

@@ -1,7 +1,7 @@
 use crate::{
-    codegen::classes::{ListType, NDArrayType, ProxyType, RangeType},
+    codegen::classes::{ListType, ProxyType, RangeType},
     symbol_resolver::{StaticValue, SymbolResolver},
-    toplevel::{helper::PrimDef, numpy::unpack_ndarray_var_tys, TopLevelContext, TopLevelDef},
+    toplevel::{helper::PrimDef, TopLevelContext, TopLevelDef},
     typecheck::{
         type_inferencer::{CodeLocation, PrimitiveStore},
         typedef::{CallId, FuncArg, Type, TypeEnum, Unifier},
@@ -24,7 +24,9 @@ use inkwell::{
     AddressSpace, IntPredicate, OptimizationLevel,
 };
 use itertools::Itertools;
+use model::*;
 use nac3parser::ast::{Location, Stmt, StrRef};
+use object::ndarray::NDArray;
 use parking_lot::{Condvar, Mutex};
 use std::collections::{HashMap, HashSet};
 use std::sync::{
@@ -41,7 +43,9 @@ pub mod extern_fns;
 mod generator;
 pub mod irrt;
 pub mod llvm_intrinsics;
+pub mod model;
 pub mod numpy;
+pub mod object;
 pub mod stmt;
 
 #[cfg(test)]
@@ -50,6 +54,22 @@ mod test;
 use concrete_type::{ConcreteType, ConcreteTypeEnum, ConcreteTypeStore};
 pub use generator::{CodeGenerator, DefaultCodeGenerator};
 
+mod macros {
+    /// Codegen-variant of [`std::unreachable`] which accepts an instance of [`CodeGenContext`] as
+    /// its first argument to provide Python source information to indicate the codegen location
+    /// causing the assertion.
+    macro_rules! codegen_unreachable {
+        ($ctx:expr $(,)?) => {
+            std::unreachable!("unreachable code while processing {}", &$ctx.current_loc)
+        };
+        ($ctx:expr, $($arg:tt)*) => {
+            std::unreachable!("unreachable code while processing {}: {}", &$ctx.current_loc, std::format!("{}", std::format_args!($($arg)+)))
+        };
+    }
+
+    pub(crate) use codegen_unreachable;
+}
+
 #[derive(Default)]
 pub struct StaticValueStore {
     pub lookup: HashMap<Vec<(usize, u64)>, usize>,
@@ -489,12 +509,7 @@ fn get_llvm_type<'ctx, G: CodeGenerator + ?Sized>(
                         }
 
                         TObj { obj_id, .. } if *obj_id == PrimDef::NDArray.id() => {
-                            let (dtype, _) = unpack_ndarray_var_tys(unifier, ty);
-                            let element_type = get_llvm_type(
-                                ctx, module, generator, unifier, top_level, type_cache, dtype,
-                            );
-
-                            NDArrayType::new(generator, ctx, element_type).as_base_type().into()
+                            Ptr(Struct(NDArray)).llvm_type(generator, ctx).as_basic_type_enum()
                         }
 
                         _ => unreachable!(

nac3core/src/codegen/model/any.rs

@@ -0,0 +1,42 @@
+use inkwell::{
+    context::Context,
+    types::{BasicType, BasicTypeEnum},
+    values::BasicValueEnum,
+};
+
+use crate::codegen::CodeGenerator;
+
+use super::*;
+
+/// A [`Model`] of any [`BasicTypeEnum`].
+///
+/// Use this when it is infeasible to use model abstractions.
+#[derive(Debug, Clone, Copy)]
+pub struct Any<'ctx>(pub BasicTypeEnum<'ctx>);
+
+impl<'ctx> Model<'ctx> for Any<'ctx> {
+    type Value = BasicValueEnum<'ctx>;
+    type Type = BasicTypeEnum<'ctx>;
+
+    fn llvm_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 +1,147 @@
+use std::fmt;
+
+use inkwell::{
+    context::Context,
+    types::{ArrayType, BasicType, BasicTypeEnum},
+    values::{ArrayValue, IntValue},
+};
+
+use crate::codegen::{CodeGenContext, CodeGenerator};
+
+use super::*;
+
+/// Trait for Rust structs identifying length values for [`Array`].
+pub trait ArrayLen: fmt::Debug + Clone + Copy {
+    fn length(&self) -> u32;
+}
+
+/// A statically known length.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct Len<const N: u32>;
+
+/// A dynamically known length.
+#[derive(Debug, Clone, Copy)]
+pub struct AnyLen(pub u32);
+
+impl<const N: u32> ArrayLen for Len<N> {
+    fn length(&self) -> u32 {
+        N
+    }
+}
+
+impl ArrayLen for AnyLen {
+    fn length(&self) -> u32 {
+        self.0
+    }
+}
+
+/// A Model for an [`ArrayType`].
+///
+/// `Len` should be of a [`LenKind`] and `Item` should be a of [`Model`].
+#[derive(Debug, Clone, Copy, Default)]
+pub struct Array<Len, Item> {
+    /// Length of this array.
+    pub len: Len,
+    /// [`Model`] of the array items.
+    pub item: Item,
+}
+
+impl<'ctx, Len: ArrayLen, Item: Model<'ctx>> Model<'ctx> for Array<Len, Item> {
+    type Value = ArrayValue<'ctx>;
+    type Type = ArrayType<'ctx>;
+
+    fn llvm_type<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &G,
+        ctx: &'ctx Context,
+    ) -> Self::Type {
+        self.item.llvm_type(generator, ctx).array_type(self.len.length())
+    }
+
+    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.length() {
+            return Err(ModelError(format!(
+                "Expecting ArrayType with size {}, but got an ArrayType with size {}",
+                ty.len(),
+                self.len.length()
+            )));
+        }
+
+        self.item
+            .check_type(generator, ctx, ty.get_element_type())
+            .map_err(|err| err.under_context("an ArrayType"))?;
+
+        Ok(())
+    }
+}
+
+impl<'ctx, Len: ArrayLen, Item: Model<'ctx>> Instance<'ctx, Ptr<Array<Len, Item>>> {
+    /// Get the pointer to the `i`-th (0-based) array element.
+    pub fn gep(
+        &self,
+        ctx: &CodeGenContext<'ctx, '_>,
+        i: IntValue<'ctx>,
+    ) -> Instance<'ctx, Ptr<Item>> {
+        let zero = ctx.ctx.i32_type().const_zero();
+        let ptr = unsafe { ctx.builder.build_in_bounds_gep(self.value, &[zero, i], "").unwrap() };
+
+        unsafe { Ptr(self.model.0.item).believe_value(ptr) }
+    }
+
+    /// Like `gep` but `i` is a constant.
+    pub fn gep_const(&self, ctx: &CodeGenContext<'ctx, '_>, i: u64) -> Instance<'ctx, Ptr<Item>> {
+        assert!(
+            i < u64::from(self.model.0.len.length()),
+            "Index {i} is out of bounds. Array length = {}",
+            self.model.0.len.length()
+        );
+
+        let i = ctx.ctx.i32_type().const_int(i, false);
+        self.gep(ctx, i)
+    }
+
+    /// Convenience function equivalent to `.gep(...).load(...)`.
+    pub fn get<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        i: IntValue<'ctx>,
+    ) -> Instance<'ctx, Item> {
+        self.gep(ctx, i).load(generator, ctx)
+    }
+
+    /// Like `get` but `i` is a constant.
+    pub fn get_const<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        i: u64,
+    ) -> Instance<'ctx, Item> {
+        self.gep_const(ctx, i).load(generator, ctx)
+    }
+
+    /// Convenience function equivalent to `.gep(...).store(...)`.
+    pub fn set(
+        &self,
+        ctx: &CodeGenContext<'ctx, '_>,
+        i: IntValue<'ctx>,
+        value: Instance<'ctx, Item>,
+    ) {
+        self.gep(ctx, i).store(ctx, value);
+    }
+
+    /// Like `set` but `i` is a constant.
+    pub fn set_const(&self, ctx: &CodeGenContext<'ctx, '_>, i: u64, value: Instance<'ctx, Item>) {
+        self.gep_const(ctx, i).store(ctx, value);
+    }
+}

nac3core/src/codegen/model/core.rs

@@ -0,0 +1,207 @@
+use std::fmt;
+
+use inkwell::{context::Context, types::*, values::*};
+use itertools::Itertools;
+
+use super::*;
+use crate::codegen::{CodeGenContext, CodeGenerator};
+
+/// A error type for reporting any [`Model`]-related error (e.g., a [`BasicType`] mismatch).
+#[derive(Debug, Clone)]
+pub struct ModelError(pub String);
+
+impl ModelError {
+    /// Append a context message to the error.
+    pub(super) fn under_context(mut self, context: &str) -> Self {
+        self.0.push_str(" ... in ");
+        self.0.push_str(context);
+        self
+    }
+}
+
+/// Trait for Rust structs identifying [`BasicType`]s in the context of a known [`CodeGenerator`] and [`CodeGenContext`].
+///
+/// For instance,
+/// - [`Int<Int32>`] identifies an [`IntType`] with 32-bits.
+/// - [`Int<SizeT>`] identifies an [`IntType`] with bit-width [`CodeGenerator::get_size_type`].
+/// - [`Ptr<Int<SizeT>>`] identifies a [`PointerType`] that points to an [`IntType`] with bit-width [`CodeGenerator::get_size_type`].
+/// - [`Int<AnyInt>`] identifies an [`IntType`] with bit-width of whatever is set in the [`AnyInt`] object.
+/// - [`Any`] identifies a [`BasicType`] set in the [`Any`] object itself.
+///
+/// You can get the [`BasicType`] out of a model with [`Model::get_type`].
+///
+/// Furthermore, [`Instance<'ctx, M>`] is a simple structure that carries a [`BasicValue`] with [`BasicType`] identified by model `M`.
+///
+/// The main purpose of this abstraction is to have a more Rust type-safe way to use Inkwell and give type-hints for programmers.
+///
+/// ### Notes on `Default` trait
+///
+/// For some models like [`Int<Int32>`] or [`Int<SizeT>`], they have a [`Default`] trait since just by looking at their types, it is possible
+/// to tell the [`BasicType`]s they are identifying.
+///
+/// This can be used to create strongly-typed interfaces accepting only values of a specific [`BasicType`] without having to worry about
+/// writing debug assertions to check, for example, if the programmer has passed in an [`IntValue`] with the wrong bit-width.
+/// ```ignore
+/// fn give_me_i32_and_get_a_size_t_back<'ctx>(i32: Instance<'ctx, Int<Int32>>) -> Instance<'ctx, Int<SizeT>> {
+///     // code...
+/// }
+/// ```
+///
+/// ### Notes on converting between Inkwell and model/ge.
+///
+/// Suppose you have an [`IntValue`], and you want to pass it into a function that takes a [`Instance<'ctx, Int<Int32>>`]. You can do use
+/// [`Model::check_value`] or [`Model::believe_value`].
+/// ```ignore
+/// let my_value: IntValue<'ctx>;
+///
+/// let my_value = Int(Int32).check_value(my_value).unwrap(); // Panics if `my_value` is not 32-bit with a descriptive error message.
+///
+/// // or, if you are absolutely certain that `my_value` is 32-bit and doing extra checks is a waste of time:
+/// let my_value = Int(Int32).believe_value(my_value);
+/// ```
+pub trait Model<'ctx>: fmt::Debug + Clone + Copy {
+    /// The [`BasicType`] *variant* this model is identifying.
+    type Type: BasicType<'ctx>;
+
+    /// The [`BasicValue`] type of the [`BasicType`] of this model.
+    type Value: BasicValue<'ctx> + TryFrom<BasicValueEnum<'ctx>>;
+
+    /// Return the [`BasicType`] of this model.
+    #[must_use]
+    fn llvm_type<G: CodeGenerator + ?Sized>(&self, generator: &G, ctx: &'ctx Context)
+        -> Self::Type;
+
+    /// Get the number of bytes of the [`BasicType`] of this model.
+    fn size_of<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> IntValue<'ctx> {
+        self.llvm_type(generator, ctx).size_of().unwrap()
+    }
+
+    /// Check if a [`BasicType`] matches the [`BasicType`] 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.
+    ///
+    /// # Safety
+    ///
+    /// Caller must make sure the type of `value` and the type of this `model` are equivalent.
+    #[must_use]
+    unsafe 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 the [`BasicValue`] 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")
+        };
+        unsafe { 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, '_>,
+    ) -> Instance<'ctx, Ptr<Self>> {
+        let p = ctx.builder.build_alloca(self.llvm_type(generator, ctx.ctx), "").unwrap();
+        unsafe { Ptr(*self).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>,
+    ) -> Instance<'ctx, Ptr<Self>> {
+        let p =
+            ctx.builder.build_array_alloca(self.llvm_type(generator, ctx.ctx), len, "").unwrap();
+        unsafe { Ptr(*self).believe_value(p) }
+    }
+
+    fn var_alloca<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        name: Option<&str>,
+    ) -> Result<Instance<'ctx, Ptr<Self>>, String> {
+        let ty = self.llvm_type(generator, ctx.ctx).as_basic_type_enum();
+        let p = generator.gen_var_alloc(ctx, ty, name)?;
+        unsafe { Ok(Ptr(*self).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<Instance<'ctx, Ptr<Self>>, String> {
+        // TODO: Remove ArraySliceValue
+        let ty = self.llvm_type(generator, ctx.ctx).as_basic_type_enum();
+        let p = generator.gen_array_var_alloc(ctx, ty, len, name)?;
+        unsafe { Ok(Ptr(*self).believe_value(PointerValue::from(p))) }
+    }
+
+    /// Allocate a constant array.
+    fn const_array<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+        values: &[Instance<'ctx, Self>],
+    ) -> Instance<'ctx, Array<AnyLen, Self>> {
+        macro_rules! make {
+            ($t:expr, $into_value:expr) => {
+                $t.const_array(
+                    &values
+                        .iter()
+                        .map(|x| $into_value(x.value.as_basic_value_enum()))
+                        .collect_vec(),
+                )
+            };
+        }
+
+        let value = match self.llvm_type(generator, ctx).as_basic_type_enum() {
+            BasicTypeEnum::ArrayType(t) => make!(t, BasicValueEnum::into_array_value),
+            BasicTypeEnum::IntType(t) => make!(t, BasicValueEnum::into_int_value),
+            BasicTypeEnum::FloatType(t) => make!(t, BasicValueEnum::into_float_value),
+            BasicTypeEnum::PointerType(t) => make!(t, BasicValueEnum::into_pointer_value),
+            BasicTypeEnum::StructType(t) => make!(t, BasicValueEnum::into_struct_value),
+            BasicTypeEnum::VectorType(t) => make!(t, BasicValueEnum::into_vector_value),
+        };
+
+        Array { len: AnyLen(values.len() as u32), item: *self }
+            .check_value(generator, ctx, value)
+            .unwrap()
+    }
+}
+
+#[derive(Debug, Clone, Copy)]
+pub struct Instance<'ctx, M: Model<'ctx>> {
+    /// The model of this instance.
+    pub model: M,
+
+    /// The value of this instance.
+    ///
+    /// It is guaranteed the [`BasicType`] of `value` is consistent with that of `model`.
+    pub value: M::Value,
+}

nac3core/src/codegen/model/float.rs

@@ -0,0 +1,94 @@
+use std::fmt;
+
+use inkwell::{
+    context::Context,
+    types::{BasicType, 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 Float<N>(pub N);
+
+impl<'ctx, N: FloatKind<'ctx>> Model<'ctx> for Float<N> {
+    type Value = FloatValue<'ctx>;
+    type Type = FloatType<'ctx>;
+
+    fn llvm_type<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &G,
+        ctx: &'ctx Context,
+    ) -> Self::Type {
+        self.0.get_float_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) = 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?
+        if ty != exp_ty {
+            return Err(ModelError(format!("Expecting {exp_ty:?}, but got {ty:?}")));
+        }
+
+        Ok(())
+    }
+}

nac3core/src/codegen/model/function.rs

@@ -0,0 +1,122 @@
+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::*;
+
+#[derive(Debug, Clone, Copy)]
+struct Arg<'ctx> {
+    ty: BasicMetadataTypeEnum<'ctx>,
+    val: BasicMetadataValueEnum<'ctx>,
+}
+
+/// A convenience structure to construct & call an LLVM function.
+///
+/// ### Usage
+///
+/// The syntax is like this:
+/// ```ignore
+/// let result = CallFunction::begin("my_function_name")
+///   .attrs(...)
+///   .arg(arg1)
+///   .arg(arg2)
+///   .arg(arg3)
+///   .returning("my_function_result", Int32);
+/// ```
+///
+/// The function `my_function_name` is called when `.returning()` (or its variants) is called, returning
+/// the result as an `Instance<'ctx, Int<Int32>>`.
+///
+/// If `my_function_name` has not been declared in `ctx.module`, once `.returning()` is called, a function
+/// declaration of `my_function_name` is added to `ctx.module`, where the [`FunctionType`] is deduced from
+/// the argument types and returning type.
+pub struct FnCall<'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> FnCall<'ctx, 'a, 'b, 'c, 'd, G> {
+    pub fn builder(generator: &'d mut G, ctx: &'b CodeGenContext<'ctx, 'a>, name: &'c str) -> Self {
+        FnCall { 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.llvm_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.llvm_type(self.generator, self.ctx.ctx);
+
+        let ret = self.call(|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.call(|tys| ret_ty.fn_type(tys, false), "");
+    }
+
+    fn call<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 +1,422 @@
+use std::{cmp::Ordering, fmt};
+
+use inkwell::{
+    context::Context,
+    types::{BasicType, 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 Int<N>(pub N);
+
+impl<'ctx, N: IntKind<'ctx>> Model<'ctx> for Int<N> {
+    type Value = IntValue<'ctx>;
+    type Type = IntType<'ctx>;
+
+    fn llvm_type<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &G,
+        ctx: &'ctx Context,
+    ) -> Self::Type {
+        self.0.get_int_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) = 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>> Int<N> {
+    pub fn const_int<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+        value: u64,
+        sign_extend: bool,
+    ) -> Instance<'ctx, Self> {
+        let value = self.llvm_type(generator, ctx).const_int(value, sign_extend);
+        unsafe { self.believe_value(value) }
+    }
+
+    pub fn const_0<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> Instance<'ctx, Self> {
+        let value = self.llvm_type(generator, ctx).const_zero();
+        unsafe { self.believe_value(value) }
+    }
+
+    pub fn const_1<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> Instance<'ctx, Self> {
+        self.const_int(generator, ctx, 1, false)
+    }
+
+    pub fn const_all_ones<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> Instance<'ctx, Self> {
+        let value = self.llvm_type(generator, ctx).const_all_ones();
+        unsafe { self.believe_value(value) }
+    }
+
+    pub fn s_extend_or_bit_cast<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        value: IntValue<'ctx>,
+    ) -> Instance<'ctx, Self> {
+        assert!(
+            value.get_type().get_bit_width()
+                <= self.0.get_int_type(generator, ctx.ctx).get_bit_width()
+        );
+        let value = ctx
+            .builder
+            .build_int_s_extend_or_bit_cast(value, self.llvm_type(generator, ctx.ctx), "")
+            .unwrap();
+        unsafe { self.believe_value(value) }
+    }
+
+    pub fn s_extend<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        value: IntValue<'ctx>,
+    ) -> Instance<'ctx, Self> {
+        assert!(
+            value.get_type().get_bit_width()
+                < self.0.get_int_type(generator, ctx.ctx).get_bit_width()
+        );
+        let value =
+            ctx.builder.build_int_s_extend(value, self.llvm_type(generator, ctx.ctx), "").unwrap();
+        unsafe { self.believe_value(value) }
+    }
+
+    pub fn z_extend_or_bit_cast<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        value: IntValue<'ctx>,
+    ) -> Instance<'ctx, Self> {
+        assert!(
+            value.get_type().get_bit_width()
+                <= self.0.get_int_type(generator, ctx.ctx).get_bit_width()
+        );
+        let value = ctx
+            .builder
+            .build_int_z_extend_or_bit_cast(value, self.llvm_type(generator, ctx.ctx), "")
+            .unwrap();
+        unsafe { self.believe_value(value) }
+    }
+
+    pub fn z_extend<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        value: IntValue<'ctx>,
+    ) -> Instance<'ctx, Self> {
+        assert!(
+            value.get_type().get_bit_width()
+                < self.0.get_int_type(generator, ctx.ctx).get_bit_width()
+        );
+        let value =
+            ctx.builder.build_int_z_extend(value, self.llvm_type(generator, ctx.ctx), "").unwrap();
+        unsafe { self.believe_value(value) }
+    }
+
+    pub fn truncate_or_bit_cast<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        value: IntValue<'ctx>,
+    ) -> Instance<'ctx, Self> {
+        assert!(
+            value.get_type().get_bit_width()
+                >= self.0.get_int_type(generator, ctx.ctx).get_bit_width()
+        );
+        let value = ctx
+            .builder
+            .build_int_truncate_or_bit_cast(value, self.llvm_type(generator, ctx.ctx), "")
+            .unwrap();
+        unsafe { self.believe_value(value) }
+    }
+
+    pub fn truncate<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        value: IntValue<'ctx>,
+    ) -> Instance<'ctx, Self> {
+        assert!(
+            value.get_type().get_bit_width()
+                > self.0.get_int_type(generator, ctx.ctx).get_bit_width()
+        );
+        let value =
+            ctx.builder.build_int_truncate(value, self.llvm_type(generator, ctx.ctx), "").unwrap();
+        unsafe { self.believe_value(value) }
+    }
+
+    /// `sext` or `trunc` an int to this model's int type. Does nothing if equal bit-widths.
+    pub fn s_extend_or_truncate<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        value: IntValue<'ctx>,
+    ) -> Instance<'ctx, Self> {
+        let their_width = value.get_type().get_bit_width();
+        let our_width = self.0.get_int_type(generator, ctx.ctx).get_bit_width();
+        match their_width.cmp(&our_width) {
+            Ordering::Less => self.s_extend(generator, ctx, value),
+            Ordering::Equal => unsafe { self.believe_value(value) },
+            Ordering::Greater => self.truncate(generator, ctx, value),
+        }
+    }
+
+    /// `zext` or `trunc` an int to this model's int type. Does nothing if equal bit-widths.
+    pub fn z_extend_or_truncate<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        value: IntValue<'ctx>,
+    ) -> Instance<'ctx, Self> {
+        let their_width = value.get_type().get_bit_width();
+        let our_width = self.0.get_int_type(generator, ctx.ctx).get_bit_width();
+        match their_width.cmp(&our_width) {
+            Ordering::Less => self.z_extend(generator, ctx, value),
+            Ordering::Equal => unsafe { self.believe_value(value) },
+            Ordering::Greater => self.truncate(generator, ctx, value),
+        }
+    }
+}
+
+impl Int<Bool> {
+    #[must_use]
+    pub fn const_false<'ctx, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> Instance<'ctx, Self> {
+        self.const_int(generator, ctx, 0, false)
+    }
+
+    #[must_use]
+    pub fn const_true<'ctx, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> Instance<'ctx, Self> {
+        self.const_int(generator, ctx, 1, false)
+    }
+}
+
+impl<'ctx, N: IntKind<'ctx>> Instance<'ctx, Int<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,
+    ) -> Instance<'ctx, Int<NewN>> {
+        Int(to_int_kind).s_extend_or_bit_cast(generator, ctx, self.value)
+    }
+
+    pub fn s_extend<NewN: IntKind<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        to_int_kind: NewN,
+    ) -> Instance<'ctx, Int<NewN>> {
+        Int(to_int_kind).s_extend(generator, ctx, self.value)
+    }
+
+    pub fn z_extend_or_bit_cast<NewN: IntKind<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        to_int_kind: NewN,
+    ) -> Instance<'ctx, Int<NewN>> {
+        Int(to_int_kind).z_extend_or_bit_cast(generator, ctx, self.value)
+    }
+
+    pub fn z_extend<NewN: IntKind<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        to_int_kind: NewN,
+    ) -> Instance<'ctx, Int<NewN>> {
+        Int(to_int_kind).z_extend(generator, ctx, self.value)
+    }
+
+    pub fn truncate_or_bit_cast<NewN: IntKind<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        to_int_kind: NewN,
+    ) -> Instance<'ctx, Int<NewN>> {
+        Int(to_int_kind).truncate_or_bit_cast(generator, ctx, self.value)
+    }
+
+    pub fn truncate<NewN: IntKind<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        to_int_kind: NewN,
+    ) -> Instance<'ctx, Int<NewN>> {
+        Int(to_int_kind).truncate(generator, ctx, self.value)
+    }
+
+    pub fn s_extend_or_truncate<NewN: IntKind<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        to_int_kind: NewN,
+    ) -> Instance<'ctx, Int<NewN>> {
+        Int(to_int_kind).s_extend_or_truncate(generator, ctx, self.value)
+    }
+
+    pub fn z_extend_or_truncate<NewN: IntKind<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        to_int_kind: NewN,
+    ) -> Instance<'ctx, Int<NewN>> {
+        Int(to_int_kind).z_extend_or_truncate(generator, ctx, self.value)
+    }
+
+    #[must_use]
+    pub fn add(&self, ctx: &CodeGenContext<'ctx, '_>, other: Self) -> Self {
+        let value = ctx.builder.build_int_add(self.value, other.value, "").unwrap();
+        unsafe { self.model.believe_value(value) }
+    }
+
+    #[must_use]
+    pub fn sub(&self, ctx: &CodeGenContext<'ctx, '_>, other: Self) -> Self {
+        let value = ctx.builder.build_int_sub(self.value, other.value, "").unwrap();
+        unsafe { self.model.believe_value(value) }
+    }
+
+    #[must_use]
+    pub fn mul(&self, ctx: &CodeGenContext<'ctx, '_>, other: Self) -> Self {
+        let value = ctx.builder.build_int_mul(self.value, other.value, "").unwrap();
+        unsafe { self.model.believe_value(value) }
+    }
+
+    pub fn compare(
+        &self,
+        ctx: &CodeGenContext<'ctx, '_>,
+        op: IntPredicate,
+        other: Self,
+    ) -> Instance<'ctx, Int<Bool>> {
+        let value = ctx.builder.build_int_compare(op, self.value, other.value, "").unwrap();
+        unsafe { Int(Bool).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 +1,223 @@
+use inkwell::{
+    context::Context,
+    types::{BasicType, BasicTypeEnum, PointerType},
+    values::{IntValue, PointerValue},
+    AddressSpace,
+};
+
+use crate::codegen::{llvm_intrinsics::call_memcpy_generic, CodeGenContext, CodeGenerator};
+
+use super::*;
+
+/// A model for [`PointerType`].
+///
+/// `Item` is the element type this pointer is pointing to, and should be of a [`Model`].
+///
+// TODO: LLVM 15: `Item` is a Rust type-hint for the LLVM type of value the `.store()/.load()` family
+// of functions return. If a truly opaque pointer is needed, tell the programmer to use `OpaquePtr`.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct Ptr<Item>(pub Item);
+
+/// An opaque pointer. Like [`Ptr`] but without any Rust type-hints about its element type.
+///
+/// `.load()/.store()` is not available for [`Instance`]s of opaque pointers.
+pub type OpaquePtr = Ptr<()>;
+
+// TODO: LLVM 15: `Item: Model<'ctx>` don't even need to be a model anymore. It will only be
+// a type-hint for the `.load()/.store()` functions for the `pointee_ty`.
+//
+// See https://thedan64.github.io/inkwell/inkwell/builder/struct.Builder.html#method.build_load.
+impl<'ctx, Item: Model<'ctx>> Model<'ctx> for Ptr<Item> {
+    type Value = PointerValue<'ctx>;
+    type Type = PointerType<'ctx>;
+
+    fn llvm_type<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &G,
+        ctx: &'ctx Context,
+    ) -> Self::Type {
+        // TODO: LLVM 15: ctx.ptr_type(AddressSpace::default())
+        self.0.llvm_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, Item: Model<'ctx>> Ptr<Item> {
+    /// Return a ***constant*** nullptr.
+    pub fn nullptr<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> Instance<'ctx, Ptr<Item>> {
+        let ptr = self.llvm_type(generator, ctx).const_null();
+        unsafe { 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>,
+    ) -> Instance<'ctx, Ptr<Item>> {
+        // TODO: LLVM 15: Write in an impl where `Item` does not have to be `Model<'ctx>`.
+        // TODO: LLVM 15: This function will only have to be:
+        // ```
+        // return self.believe_value(ptr);
+        // ```
+        let t = self.llvm_type(generator, ctx.ctx);
+        let ptr = ctx.builder.build_pointer_cast(ptr, t, "").unwrap();
+        unsafe { self.believe_value(ptr) }
+    }
+}
+
+impl<'ctx, Item: Model<'ctx>> Instance<'ctx, Ptr<Item>> {
+    /// Offset the pointer by [`inkwell::builder::Builder::build_in_bounds_gep`].
+    #[must_use]
+    pub fn offset(
+        &self,
+        ctx: &CodeGenContext<'ctx, '_>,
+        offset: IntValue<'ctx>,
+    ) -> Instance<'ctx, Ptr<Item>> {
+        let p = unsafe { ctx.builder.build_in_bounds_gep(self.value, &[offset], "").unwrap() };
+        unsafe { self.model.believe_value(p) }
+    }
+
+    /// Offset the pointer by [`inkwell::builder::Builder::build_in_bounds_gep`] by a constant offset.
+    #[must_use]
+    pub fn offset_const(
+        &self,
+        ctx: &CodeGenContext<'ctx, '_>,
+        offset: i64,
+    ) -> Instance<'ctx, Ptr<Item>> {
+        let offset = ctx.ctx.i32_type().const_int(offset as u64, true);
+        self.offset(ctx, offset)
+    }
+
+    pub fn set_index(
+        &self,
+        ctx: &CodeGenContext<'ctx, '_>,
+        index: IntValue<'ctx>,
+        value: Instance<'ctx, Item>,
+    ) {
+        self.offset(ctx, index).store(ctx, value);
+    }
+
+    pub fn set_index_const(
+        &self,
+        ctx: &CodeGenContext<'ctx, '_>,
+        index: i64,
+        value: Instance<'ctx, Item>,
+    ) {
+        self.offset_const(ctx, index).store(ctx, value);
+    }
+
+    pub fn get_index<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        index: IntValue<'ctx>,
+    ) -> Instance<'ctx, Item> {
+        self.offset(ctx, index).load(generator, ctx)
+    }
+
+    pub fn get_index_const<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        index: i64,
+    ) -> Instance<'ctx, Item> {
+        self.offset_const(ctx, index).load(generator, ctx)
+    }
+
+    /// Load the value with [`inkwell::builder::Builder::build_load`].
+    pub fn load<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Item> {
+        let value = ctx.builder.build_load(self.value, "").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, Item>) {
+        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<NewItem: Model<'ctx>, G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        new_item: NewItem,
+    ) -> Instance<'ctx, Ptr<NewItem>> {
+        // TODO: LLVM 15: Write in an impl where `Item` does not have to be `Model<'ctx>`.
+        Ptr(new_item).pointer_cast(generator, ctx, self.value)
+    }
+
+    /// Cast this pointer to `uint8_t*`
+    pub fn cast_to_pi8<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Ptr<Int<Byte>>> {
+        Ptr(Int(Byte)).pointer_cast(generator, ctx, self.value)
+    }
+
+    /// Check if the pointer is null with [`inkwell::builder::Builder::build_is_null`].
+    pub fn is_null(&self, ctx: &CodeGenContext<'ctx, '_>) -> Instance<'ctx, Int<Bool>> {
+        let value = ctx.builder.build_is_null(self.value, "").unwrap();
+        unsafe { Int(Bool).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, '_>) -> Instance<'ctx, Int<Bool>> {
+        let value = ctx.builder.build_is_not_null(self.value, "").unwrap();
+        unsafe { Int(Bool).believe_value(value) }
+    }
+
+    /// `memcpy` from another pointer.
+    pub fn copy_from<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        source: Self,
+        num_items: IntValue<'ctx>,
+    ) {
+        // Force extend `num_items` and `itemsize` to `i64` so their types would match.
+        let itemsize = self.model.size_of(generator, ctx.ctx);
+        let itemsize = Int(SizeT).z_extend_or_truncate(generator, ctx, itemsize);
+        let num_items = Int(SizeT).z_extend_or_truncate(generator, ctx, num_items);
+        let totalsize = itemsize.mul(ctx, num_items);
+
+        let is_volatile = ctx.ctx.bool_type().const_zero(); // is_volatile = false
+        call_memcpy_generic(ctx, self.value, source.value, totalsize.value, is_volatile);
+    }
+}

nac3core/src/codegen/model/structure.rs

@@ -0,0 +1,364 @@
+use std::fmt;
+
+use inkwell::{
+    context::Context,
+    types::{BasicType, BasicTypeEnum, StructType},
+    values::{BasicValueEnum, StructValue},
+};
+
+use crate::codegen::{CodeGenContext, CodeGenerator};
+
+use super::*;
+
+/// A traveral that traverses a Rust `struct` that is used to declare an LLVM's struct's field types.
+pub trait FieldTraversal<'ctx> {
+    /// Output type of [`FieldTraversal::add`].
+    type Output<M>;
+
+    /// Traverse through the type of a declared field and do something with it.
+    ///
+    /// * `name` - The cosmetic name of the LLVM field. Used for debugging.
+    /// * `model` - The [`Model`] representing the LLVM type of this field.
+    fn add<M: Model<'ctx>>(&mut self, name: &'static str, model: M) -> Self::Output<M>;
+
+    /// Like [`FieldTraversal::add`] but [`Model`] is automatically inferred from its [`Default`] trait.
+    fn add_auto<M: Model<'ctx> + Default>(&mut self, name: &'static str) -> Self::Output<M> {
+        self.add(name, M::default())
+    }
+}
+
+/// Descriptor of an LLVM struct field.
+#[derive(Debug, Clone, Copy)]
+pub struct GepField<M> {
+    /// The GEP index of this field. This is the index to use with `build_gep`.
+    pub gep_index: u32,
+    /// The cosmetic name of this field.
+    pub name: &'static str,
+    /// The [`Model`] of this field's type.
+    pub model: M,
+}
+
+/// A traversal to calculate the GEP index of fields.
+pub struct GepFieldTraversal {
+    /// The current GEP index.
+    gep_index_counter: u32,
+}
+
+impl<'ctx> FieldTraversal<'ctx> for GepFieldTraversal {
+    type Output<M> = GepField<M>;
+
+    fn add<M: Model<'ctx>>(&mut self, name: &'static str, model: M) -> Self::Output<M> {
+        let gep_index = self.gep_index_counter;
+        self.gep_index_counter += 1;
+        Self::Output { gep_index, name, model }
+    }
+}
+
+/// A traversal to collect the field types of a struct.
+///
+/// This is used to collect field types and construct the LLVM struct type with [`Context::struct_type`].
+struct TypeFieldTraversal<'ctx, 'a, G: CodeGenerator + ?Sized> {
+    generator: &'a G,
+    ctx: &'ctx Context,
+    /// The collected field types so far in exact order.
+    field_types: Vec<BasicTypeEnum<'ctx>>,
+}
+
+impl<'ctx, 'a, G: CodeGenerator + ?Sized> FieldTraversal<'ctx> for TypeFieldTraversal<'ctx, 'a, G> {
+    type Output<M> = (); // Checking types return nothing.
+
+    fn add<M: Model<'ctx>>(&mut self, _name: &'static str, model: M) -> Self::Output<M> {
+        let t = model.llvm_type(self.generator, self.ctx).as_basic_type_enum();
+        self.field_types.push(t);
+    }
+}
+
+/// A traversal to check the types of fields.
+struct CheckTypeFieldTraversal<'ctx, 'a, G: CodeGenerator + ?Sized> {
+    generator: &'a mut G,
+    ctx: &'ctx Context,
+    /// The current GEP index, so we can tell the index of the field we are checking
+    /// and report the GEP index.
+    gep_index_counter: u32,
+    /// The [`StructType`] to check.
+    scrutinee: StructType<'ctx>,
+    /// The list of collected errors so far.
+    errors: Vec<ModelError>,
+}
+
+impl<'ctx, 'a, G: CodeGenerator + ?Sized> FieldTraversal<'ctx>
+    for CheckTypeFieldTraversal<'ctx, 'a, G>
+{
+    type Output<M> = (); // Checking types return nothing.
+
+    fn add<M: Model<'ctx>>(&mut self, name: &'static str, model: M) -> Self::Output<M> {
+        let gep_index = self.gep_index_counter;
+        self.gep_index_counter += 1;
+
+        if let Some(t) = self.scrutinee.get_field_type_at_index(gep_index) {
+            if let Err(err) = model.check_type(self.generator, self.ctx, t) {
+                self.errors
+                    .push(err.under_context(format!("field #{gep_index} '{name}'").as_str()));
+            }
+        }
+        // Otherwise, it will be caught by Struct's `check_type`.
+    }
+}
+
+/// A trait for Rust structs identifying LLVM structures.
+///
+/// ### Example
+///
+/// Suppose you want to define this structure:
+/// ```c
+/// template <typename T>
+/// struct ContiguousNDArray {
+///     size_t ndims;
+///     size_t* shape;
+///     T* data;
+/// }
+/// ```
+///
+/// This is how it should be done:
+/// ```ignore
+/// pub struct ContiguousNDArrayFields<'ctx, F: FieldTraversal<'ctx>, Item: Model<'ctx>> {
+///     pub ndims: F::Out<Int<SizeT>>,
+///     pub shape: F::Out<Ptr<Int<SizeT>>>,
+///     pub data: F::Out<Ptr<Item>>,
+/// }
+///
+/// /// An ndarray without strides and non-opaque `data` field in NAC3.
+/// #[derive(Debug, Clone, Copy)]
+/// pub struct ContiguousNDArray<M> {
+///     /// [`Model`] of the items.
+///     pub item: M,
+/// }
+///
+/// impl<'ctx, Item: Model<'ctx>> StructKind<'ctx> for ContiguousNDArray<Item> {
+///     type Fields<F: FieldTraversal<'ctx>> = ContiguousNDArrayFields<'ctx, F, Item>;
+///
+///     fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+///         // The order of `traversal.add*` is important
+///         Self::Fields {
+///             ndims: traversal.add_auto("ndims"),
+///             shape: traversal.add_auto("shape"),
+///             data: traversal.add("data", Ptr(self.item)),
+///         }
+///     }
+/// }
+/// ```
+///
+/// The [`FieldTraversal`] here is a mechanism to allow the fields of `ContiguousNDArrayFields` to be
+/// traversed to do useful work such as:
+///
+/// - To create the [`StructType`] of `ContiguousNDArray` by collecting [`BasicType`]s of the fields.
+/// - To enable the `.gep(ctx, |f| f.ndims).store(ctx, ...)` syntax.
+///
+/// Suppose now that you have defined `ContiguousNDArray` and you want to allocate a `ContiguousNDArray`
+/// with dtype `float64` in LLVM, this is how you do it:
+/// ```ignore
+/// type F64NDArray = Struct<ContiguousNDArray<Float<Float64>>>; // Type alias for leaner documentation
+/// let model: F64NDArray = Struct(ContigousNDArray { item: Float(Float64) });
+/// let ndarray: Instance<'ctx, Ptr<F64NDArray>> = model.alloca(generator, ctx);
+/// ```
+///
+/// ...and here is how you may manipulate/access `ndarray`:
+///
+/// (NOTE: some arguments have been omitted)
+///
+/// ```ignore
+/// // Get `&ndarray->data`
+/// ndarray.gep(|f| f.data); // type: Instance<'ctx, Ptr<Float<Float64>>>
+///
+/// // Get `ndarray->ndims`
+/// ndarray.get(|f| f.ndims); // type: Instance<'ctx, Int<SizeT>>
+///
+/// // Get `&ndarray->ndims`
+/// ndarray.gep(|f| f.ndims); // type: Instance<'ctx, Ptr<Int<SizeT>>>
+///
+/// // Get `ndarray->shape[0]`
+/// ndarray.get(|f| f.shape).get_index_const(0); // Instance<'ctx, Int<SizeT>>
+///
+/// // Get `&ndarray->shape[2]`
+/// ndarray.get(|f| f.shape).offset_const(2); // Instance<'ctx, Ptr<Int<SizeT>>>
+///
+/// // Do `ndarray->ndims = 3;`
+/// let num_3 = Int(SizeT).const_int(3);
+/// ndarray.set(|f| f.ndims, num_3);
+/// ```
+pub trait StructKind<'ctx>: fmt::Debug + Clone + Copy {
+    /// The associated fields of this struct.
+    type Fields<F: FieldTraversal<'ctx>>;
+
+    /// Traverse through all fields of this [`StructKind`].
+    ///
+    /// Only used internally in this module for implementing other components.
+    fn iter_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F>;
+
+    /// Get a convenience structure to get a struct field's GEP index through its corresponding Rust field.
+    ///
+    /// Only used internally in this module for implementing other components.
+    fn fields(&self) -> Self::Fields<GepFieldTraversal> {
+        self.iter_fields(&mut GepFieldTraversal { gep_index_counter: 0 })
+    }
+
+    /// Get the LLVM [`StructType`] of this [`StructKind`].
+    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.iter_fields(&mut traversal);
+
+        ctx.struct_type(&traversal.field_types, false)
+    }
+}
+
+/// A model for LLVM struct.
+///
+/// `S` should be of a [`StructKind`].
+#[derive(Debug, Clone, Copy, Default)]
+pub struct Struct<S>(pub S);
+
+impl<'ctx, S: StructKind<'ctx>> Struct<S> {
+    /// Create a constant struct value from its fields.
+    ///
+    /// This function also validates `fields` and panic when there is something wrong.
+    pub fn const_struct<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+        fields: &[BasicValueEnum<'ctx>],
+    ) -> Instance<'ctx, Self> {
+        // NOTE: There *could* have been a functor `F<M> = Instance<'ctx, M>` for `S::Fields<F>`
+        // to create a more user-friendly interface, but Rust's type system is not sophisticated enough
+        // and if you try doing that Rust would force you put lifetimes everywhere.
+        let val = ctx.const_struct(fields, false);
+        self.check_value(generator, ctx, val).unwrap()
+    }
+}
+
+impl<'ctx, S: StructKind<'ctx>> Model<'ctx> for Struct<S> {
+    type Value = StructValue<'ctx>;
+    type Type = StructType<'ctx>;
+
+    fn llvm_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:?}")));
+        };
+
+        // Check each field individually.
+        let mut traversal = CheckTypeFieldTraversal {
+            generator,
+            ctx,
+            gep_index_counter: 0,
+            errors: Vec::new(),
+            scrutinee: ty,
+        };
+        self.0.iter_fields(&mut traversal);
+
+        // Check the number of fields.
+        let exp_num_fields = traversal.gep_index_counter;
+        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() {
+            // Currently, only the first error is reported.
+            return Err(traversal.errors[0].clone());
+        }
+
+        Ok(())
+    }
+}
+
+impl<'ctx, S: StructKind<'ctx>> Instance<'ctx, Struct<S>> {
+    /// Get a field with [`StructValue::get_field_at_index`].
+    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).unwrap();
+        field.model.check_value(generator, ctx, val).unwrap()
+    }
+}
+
+impl<'ctx, S: StructKind<'ctx>> Instance<'ctx, Ptr<Struct<S>>> {
+    /// Get a pointer to a field with [`Builder::build_in_bounds_gep`].
+    pub fn gep<M, GetField>(
+        &self,
+        ctx: &CodeGenContext<'ctx, '_>,
+        get_field: GetField,
+    ) -> Instance<'ctx, Ptr<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();
+
+        let ptr = unsafe {
+            ctx.builder
+                .build_in_bounds_gep(
+                    self.value,
+                    &[llvm_i32.const_zero(), llvm_i32.const_int(u64::from(field.gep_index), false)],
+                    field.name,
+                )
+                .unwrap()
+        };
+
+        unsafe { Ptr(field.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,
+    ) -> Instance<'ctx, M>
+    where
+        M: Model<'ctx>,
+        GetField: FnOnce(S::Fields<GepFieldTraversal>) -> GepField<M>,
+    {
+        self.gep(ctx, get_field).load(generator, ctx)
+    }
+
+    /// 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);
+    }
+}

nac3core/src/codegen/model/util.rs

@@ -0,0 +1,42 @@
+use crate::codegen::{
+    stmt::{gen_for_callback_incrementing, BreakContinueHooks},
+    CodeGenContext, CodeGenerator,
+};
+
+use super::*;
+
+/// Like [`gen_for_callback_incrementing`] with [`Model`] abstractions.
+///
+/// `stop` is not included.
+pub fn gen_for_model<'ctx, 'a, G, F, N>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, 'a>,
+    start: Instance<'ctx, Int<N>>,
+    stop: Instance<'ctx, Int<N>>,
+    step: Instance<'ctx, Int<N>>,
+    body: F,
+) -> Result<(), String>
+where
+    G: CodeGenerator + ?Sized,
+    F: FnOnce(
+        &mut G,
+        &mut CodeGenContext<'ctx, 'a>,
+        BreakContinueHooks<'ctx>,
+        Instance<'ctx, Int<N>>,
+    ) -> Result<(), String>,
+    N: IntKind<'ctx> + Default,
+{
+    let int_model = Int(N::default());
+    gen_for_callback_incrementing(
+        generator,
+        ctx,
+        None,
+        start.value,
+        (stop.value, false),
+        |g, ctx, hooks, i| {
+            let i = unsafe { int_model.believe_value(i) };
+            body(g, ctx, hooks, i)
+        },
+        step.value,
+    )
+}

nac3core/src/codegen/numpy.rs

@@ -12,18 +12,24 @@ use crate::{
             call_ndarray_calc_size,
         },
         llvm_intrinsics::{self, call_memcpy_generic},
+        macros::codegen_unreachable,
+        model::*,
+        object::{
+            any::AnyObject,
+            ndarray::{shape_util::parse_numpy_int_sequence, NDArrayObject},
+        },
         stmt::{gen_for_callback_incrementing, gen_for_range_callback, gen_if_else_expr_callback},
         CodeGenContext, CodeGenerator,
     },
     symbol_resolver::ValueEnum,
     toplevel::{
-        helper::PrimDef,
+        helper::extract_ndims,
         numpy::{make_ndarray_ty, unpack_ndarray_var_tys},
         DefinitionId,
     },
     typecheck::{
         magic_methods::Binop,
-        typedef::{FunSignature, Type, TypeEnum},
+        typedef::{FunSignature, Type},
     },
 };
 use inkwell::{
@@ -259,7 +265,7 @@ fn ndarray_zero_value<'ctx, G: CodeGenerator + ?Sized>(
     } else if ctx.unifier.unioned(elem_ty, ctx.primitives.str) {
         ctx.gen_string(generator, "").into()
     } else {
-        unreachable!()
+        codegen_unreachable!(ctx)
     }
 }
 
@@ -287,7 +293,7 @@ fn ndarray_one_value<'ctx, G: CodeGenerator + ?Sized>(
     } else if ctx.unifier.unioned(elem_ty, ctx.primitives.str) {
         ctx.gen_string(generator, "1").into()
     } else {
-        unreachable!()
+        codegen_unreachable!(ctx)
     }
 }
 
@@ -355,7 +361,7 @@ fn call_ndarray_empty_impl<'ctx, G: CodeGenerator + ?Sized>(
 
             create_ndarray_const_shape(generator, ctx, elem_ty, &[shape_int])
         }
-        _ => unreachable!(),
+        _ => codegen_unreachable!(ctx),
     }
 }
 
@@ -626,7 +632,7 @@ fn call_ndarray_full_impl<'ctx, G: CodeGenerator + ?Sized>(
         } else if fill_value.is_int_value() || fill_value.is_float_value() {
             fill_value
         } else {
-            unreachable!()
+            codegen_unreachable!(ctx)
         };
 
         Ok(value)
@@ -1742,8 +1748,13 @@ pub fn gen_ndarray_empty<'ctx>(
     let shape_ty = fun.0.args[0].ty;
     let shape_arg = args[0].1.clone().to_basic_value_enum(context, generator, shape_ty)?;
 
-    call_ndarray_empty_impl(generator, context, context.primitives.float, shape_arg)
-        .map(NDArrayValue::into)
+    let (dtype, ndims) = unpack_ndarray_var_tys(&mut context.unifier, fun.0.ret);
+    let ndims = extract_ndims(&context.unifier, ndims);
+
+    let shape = AnyObject { value: shape_arg, ty: shape_ty };
+    let (_, shape) = parse_numpy_int_sequence(generator, context, shape);
+    let ndarray = NDArrayObject::make_np_empty(generator, context, dtype, ndims, shape);
+    Ok(ndarray.instance.value)
 }
 
 /// Generates LLVM IR for `ndarray.zeros`.
@@ -1760,8 +1771,13 @@ pub fn gen_ndarray_zeros<'ctx>(
     let shape_ty = fun.0.args[0].ty;
     let shape_arg = args[0].1.clone().to_basic_value_enum(context, generator, shape_ty)?;
 
-    call_ndarray_zeros_impl(generator, context, context.primitives.float, shape_arg)
-        .map(NDArrayValue::into)
+    let (dtype, ndims) = unpack_ndarray_var_tys(&mut context.unifier, fun.0.ret);
+    let ndims = extract_ndims(&context.unifier, ndims);
+
+    let shape = AnyObject { value: shape_arg, ty: shape_ty };
+    let (_, shape) = parse_numpy_int_sequence(generator, context, shape);
+    let ndarray = NDArrayObject::make_np_zeros(generator, context, dtype, ndims, shape);
+    Ok(ndarray.instance.value)
 }
 
 /// Generates LLVM IR for `ndarray.ones`.
@@ -1778,8 +1794,13 @@ pub fn gen_ndarray_ones<'ctx>(
     let shape_ty = fun.0.args[0].ty;
     let shape_arg = args[0].1.clone().to_basic_value_enum(context, generator, shape_ty)?;
 
-    call_ndarray_ones_impl(generator, context, context.primitives.float, shape_arg)
-        .map(NDArrayValue::into)
+    let (dtype, ndims) = unpack_ndarray_var_tys(&mut context.unifier, fun.0.ret);
+    let ndims = extract_ndims(&context.unifier, ndims);
+
+    let shape = AnyObject { value: shape_arg, ty: shape_ty };
+    let (_, shape) = parse_numpy_int_sequence(generator, context, shape);
+    let ndarray = NDArrayObject::make_np_ones(generator, context, dtype, ndims, shape);
+    Ok(ndarray.instance.value)
 }
 
 /// Generates LLVM IR for `ndarray.full`.
@@ -1799,8 +1820,14 @@ pub fn gen_ndarray_full<'ctx>(
     let fill_value_arg =
         args[1].1.clone().to_basic_value_enum(context, generator, fill_value_ty)?;
 
-    call_ndarray_full_impl(generator, context, fill_value_ty, shape_arg, fill_value_arg)
-        .map(NDArrayValue::into)
+    let (dtype, ndims) = unpack_ndarray_var_tys(&mut context.unifier, fun.0.ret);
+    let ndims = extract_ndims(&context.unifier, ndims);
+
+    let shape = AnyObject { value: shape_arg, ty: shape_ty };
+    let (_, shape) = parse_numpy_int_sequence(generator, context, shape);
+    let ndarray =
+        NDArrayObject::make_np_full(generator, context, dtype, ndims, shape, fill_value_arg);
+    Ok(ndarray.instance.value)
 }
 
 pub fn gen_ndarray_array<'ctx>(
@@ -1814,26 +1841,6 @@ pub fn gen_ndarray_array<'ctx>(
     assert!(matches!(args.len(), 1..=3));
 
     let obj_ty = fun.0.args[0].ty;
-    let obj_elem_ty = match &*context.unifier.get_ty(obj_ty) {
-        TypeEnum::TObj { obj_id, .. } if *obj_id == PrimDef::NDArray.id() => {
-            unpack_ndarray_var_tys(&mut context.unifier, obj_ty).0
-        }
-
-        TypeEnum::TObj { obj_id, params, .. } if *obj_id == PrimDef::List.id() => {
-            let mut ty = *params.iter().next().unwrap().1;
-            while let TypeEnum::TObj { obj_id, params, .. } = &*context.unifier.get_ty_immutable(ty)
-            {
-                if *obj_id != PrimDef::List.id() {
-                    break;
-                }
-
-                ty = *params.iter().next().unwrap().1;
-            }
-            ty
-        }
-
-        _ => obj_ty,
-    };
     let obj_arg = args[0].1.clone().to_basic_value_enum(context, generator, obj_ty)?;
 
     let copy_arg = if let Some(arg) =
@@ -1849,28 +1856,18 @@ pub fn gen_ndarray_array<'ctx>(
         )
     };
 
-    let ndmin_arg = if let Some(arg) =
-        args.iter().find(|arg| arg.0.is_some_and(|name| name == fun.0.args[2].name))
-    {
-        let ndmin_ty = fun.0.args[2].ty;
-        arg.1.clone().to_basic_value_enum(context, generator, ndmin_ty)?
-    } else {
-        context.gen_symbol_val(
-            generator,
-            fun.0.args[2].default_value.as_ref().unwrap(),
-            fun.0.args[2].ty,
-        )
-    };
+    // The ndmin argument is ignored. We can simply force the ndarray's number of dimensions to be
+    // the `ndims` of the function return type.
+    let (_, ndims) = unpack_ndarray_var_tys(&mut context.unifier, fun.0.ret);
+    let ndims = extract_ndims(&context.unifier, ndims);
 
-    call_ndarray_array_impl(
-        generator,
-        context,
-        obj_elem_ty,
-        obj_arg,
-        copy_arg.into_int_value(),
-        ndmin_arg.into_int_value(),
-    )
-    .map(NDArrayValue::into)
+    let object = AnyObject { value: obj_arg, ty: obj_ty };
+    // NAC3 booleans are i8.
+    let copy = Int(Bool).truncate(generator, context, copy_arg.into_int_value());
+    let ndarray = NDArrayObject::make_np_array(generator, context, object, copy)
+        .atleast_nd(generator, context, ndims);
+
+    Ok(ndarray.instance.value)
 }
 
 /// Generates LLVM IR for `ndarray.eye`.
@@ -1909,15 +1906,23 @@ pub fn gen_ndarray_eye<'ctx>(
         ))
     }?;
 
-    call_ndarray_eye_impl(
-        generator,
-        context,
-        context.primitives.float,
-        nrows_arg.into_int_value(),
-        ncols_arg.into_int_value(),
-        offset_arg.into_int_value(),
-    )
-    .map(NDArrayValue::into)
+    let (dtype, _) = unpack_ndarray_var_tys(&mut context.unifier, fun.0.ret);
+
+    let nrows = Int(Int32)
+        .check_value(generator, context.ctx, nrows_arg)
+        .unwrap()
+        .s_extend_or_bit_cast(generator, context, SizeT);
+    let ncols = Int(Int32)
+        .check_value(generator, context.ctx, ncols_arg)
+        .unwrap()
+        .s_extend_or_bit_cast(generator, context, SizeT);
+    let offset = Int(Int32)
+        .check_value(generator, context.ctx, offset_arg)
+        .unwrap()
+        .s_extend_or_bit_cast(generator, context, SizeT);
+
+    let ndarray = NDArrayObject::make_np_eye(generator, context, dtype, nrows, ncols, offset);
+    Ok(ndarray.instance.value)
 }
 
 /// Generates LLVM IR for `ndarray.identity`.
@@ -1931,20 +1936,15 @@ pub fn gen_ndarray_identity<'ctx>(
     assert!(obj.is_none());
     assert_eq!(args.len(), 1);
 
-    let llvm_usize = generator.get_size_type(context.ctx);
+    let (dtype, _) = unpack_ndarray_var_tys(&mut context.unifier, fun.0.ret);
 
     let n_ty = fun.0.args[0].ty;
     let n_arg = args[0].1.clone().to_basic_value_enum(context, generator, n_ty)?;
 
-    call_ndarray_eye_impl(
-        generator,
-        context,
-        context.primitives.float,
-        n_arg.into_int_value(),
-        n_arg.into_int_value(),
-        llvm_usize.const_zero(),
-    )
-    .map(NDArrayValue::into)
+    let n = Int(Int32).check_value(generator, context.ctx, n_arg).unwrap();
+    let n = n.s_extend_or_bit_cast(generator, context, SizeT);
+    let ndarray = NDArrayObject::make_np_identity(generator, context, dtype, n);
+    Ok(ndarray.instance.value)
 }
 
 /// Generates LLVM IR for `ndarray.copy`.
@@ -1958,20 +1958,14 @@ pub fn gen_ndarray_copy<'ctx>(
     assert!(obj.is_some());
     assert!(args.is_empty());
 
-    let llvm_usize = generator.get_size_type(context.ctx);
-
     let this_ty = obj.as_ref().unwrap().0;
-    let (this_elem_ty, _) = unpack_ndarray_var_tys(&mut context.unifier, this_ty);
     let this_arg =
         obj.as_ref().unwrap().1.clone().to_basic_value_enum(context, generator, this_ty)?;
 
-    ndarray_copy_impl(
-        generator,
-        context,
-        this_elem_ty,
-        NDArrayValue::from_ptr_val(this_arg.into_pointer_value(), llvm_usize, None),
-    )
-    .map(NDArrayValue::into)
+    let this = AnyObject { value: this_arg, ty: this_ty };
+    let this = NDArrayObject::from_object(generator, context, this);
+    let ndarray = this.make_copy(generator, context);
+    Ok(ndarray.instance.value)
 }
 
 /// Generates LLVM IR for `ndarray.fill`.
@@ -1985,443 +1979,18 @@ pub fn gen_ndarray_fill<'ctx>(
     assert!(obj.is_some());
     assert_eq!(args.len(), 1);
 
-    let llvm_usize = generator.get_size_type(context.ctx);
-
     let this_ty = obj.as_ref().unwrap().0;
-    let this_arg = obj
-        .as_ref()
-        .unwrap()
-        .1
-        .clone()
-        .to_basic_value_enum(context, generator, this_ty)?
-        .into_pointer_value();
+    let this_arg =
+        obj.as_ref().unwrap().1.clone().to_basic_value_enum(context, generator, this_ty)?;
     let value_ty = fun.0.args[0].ty;
     let value_arg = args[0].1.clone().to_basic_value_enum(context, generator, value_ty)?;
 
-    ndarray_fill_flattened(
-        generator,
-        context,
-        NDArrayValue::from_ptr_val(this_arg, llvm_usize, None),
-        |generator, ctx, _| {
-            let value = if value_arg.is_pointer_value() {
-                let llvm_i1 = ctx.ctx.bool_type();
-
-                let copy = generator.gen_var_alloc(ctx, value_arg.get_type(), None)?;
-
-                call_memcpy_generic(
-                    ctx,
-                    copy,
-                    value_arg.into_pointer_value(),
-                    value_arg.get_type().size_of().map(Into::into).unwrap(),
-                    llvm_i1.const_zero(),
-                );
-
-                copy.into()
-            } else if value_arg.is_int_value() || value_arg.is_float_value() {
-                value_arg
-            } else {
-                unreachable!()
-            };
-
-            Ok(value)
-        },
-    )?;
-
+    let this = AnyObject { value: this_arg, ty: this_ty };
+    let this = NDArrayObject::from_object(generator, context, this);
+    this.fill(generator, context, value_arg);
     Ok(())
 }
 
-/// Generates LLVM IR for `ndarray.transpose`.
-pub fn ndarray_transpose<'ctx, G: CodeGenerator + ?Sized>(
-    generator: &mut G,
-    ctx: &mut CodeGenContext<'ctx, '_>,
-    x1: (Type, BasicValueEnum<'ctx>),
-) -> Result<BasicValueEnum<'ctx>, String> {
-    const FN_NAME: &str = "ndarray_transpose";
-    let (x1_ty, x1) = x1;
-    let llvm_usize = generator.get_size_type(ctx.ctx);
-
-    if let BasicValueEnum::PointerValue(n1) = x1 {
-        let (elem_ty, _) = unpack_ndarray_var_tys(&mut ctx.unifier, x1_ty);
-        let n1 = NDArrayValue::from_ptr_val(n1, llvm_usize, None);
-        let n_sz = call_ndarray_calc_size(generator, ctx, &n1.dim_sizes(), (None, None));
-
-        // Dimensions are reversed in the transposed array
-        let out = create_ndarray_dyn_shape(
-            generator,
-            ctx,
-            elem_ty,
-            &n1,
-            |_, ctx, n| Ok(n.load_ndims(ctx)),
-            |generator, ctx, n, idx| {
-                let new_idx = ctx.builder.build_int_sub(n.load_ndims(ctx), idx, "").unwrap();
-                let new_idx = ctx
-                    .builder
-                    .build_int_sub(new_idx, new_idx.get_type().const_int(1, false), "")
-                    .unwrap();
-                unsafe { Ok(n.dim_sizes().get_typed_unchecked(ctx, generator, &new_idx, None)) }
-            },
-        )
-        .unwrap();
-
-        gen_for_callback_incrementing(
-            generator,
-            ctx,
-            None,
-            llvm_usize.const_zero(),
-            (n_sz, false),
-            |generator, ctx, _, idx| {
-                let elem = unsafe { n1.data().get_unchecked(ctx, generator, &idx, None) };
-
-                let new_idx = generator.gen_var_alloc(ctx, llvm_usize.into(), None)?;
-                let rem_idx = generator.gen_var_alloc(ctx, llvm_usize.into(), None)?;
-                ctx.builder.build_store(new_idx, llvm_usize.const_zero()).unwrap();
-                ctx.builder.build_store(rem_idx, idx).unwrap();
-
-                // Incrementally calculate the new index in the transposed array
-                // For each index, we first decompose it into the n-dims and use those to reconstruct the new index
-                // The formula used for indexing is:
-                // idx = dim_n * ( ... (dim2 * (dim0 * dim1) + dim1) + dim2 ... ) + dim_n
-                gen_for_callback_incrementing(
-                    generator,
-                    ctx,
-                    None,
-                    llvm_usize.const_zero(),
-                    (n1.load_ndims(ctx), false),
-                    |generator, ctx, _, ndim| {
-                        let ndim_rev =
-                            ctx.builder.build_int_sub(n1.load_ndims(ctx), ndim, "").unwrap();
-                        let ndim_rev = ctx
-                            .builder
-                            .build_int_sub(ndim_rev, llvm_usize.const_int(1, false), "")
-                            .unwrap();
-                        let dim = unsafe {
-                            n1.dim_sizes().get_typed_unchecked(ctx, generator, &ndim_rev, None)
-                        };
-
-                        let rem_idx_val =
-                            ctx.builder.build_load(rem_idx, "").unwrap().into_int_value();
-                        let new_idx_val =
-                            ctx.builder.build_load(new_idx, "").unwrap().into_int_value();
-
-                        let add_component =
-                            ctx.builder.build_int_unsigned_rem(rem_idx_val, dim, "").unwrap();
-                        let rem_idx_val =
-                            ctx.builder.build_int_unsigned_div(rem_idx_val, dim, "").unwrap();
-
-                        let new_idx_val = ctx.builder.build_int_mul(new_idx_val, dim, "").unwrap();
-                        let new_idx_val =
-                            ctx.builder.build_int_add(new_idx_val, add_component, "").unwrap();
-
-                        ctx.builder.build_store(rem_idx, rem_idx_val).unwrap();
-                        ctx.builder.build_store(new_idx, new_idx_val).unwrap();
-
-                        Ok(())
-                    },
-                    llvm_usize.const_int(1, false),
-                )?;
-
-                let new_idx_val = ctx.builder.build_load(new_idx, "").unwrap().into_int_value();
-                unsafe { out.data().set_unchecked(ctx, generator, &new_idx_val, elem) };
-                Ok(())
-            },
-            llvm_usize.const_int(1, false),
-        )?;
-
-        Ok(out.as_base_value().into())
-    } else {
-        unreachable!(
-            "{FN_NAME}() not supported for '{}'",
-            format!("'{}'", ctx.unifier.stringify(x1_ty))
-        )
-    }
-}
-
-/// LLVM-typed implementation for generating the implementation for `ndarray.reshape`.
-///
-/// * `x1` - `NDArray` to reshape.
-/// * `shape` - The `shape` parameter used to construct the new `NDArray`.
-///   Just like numpy, the `shape` argument can be:
-///   1. A list of `int32`;   e.g., `np.reshape(arr, [600, -1, 3])`
-///   2. A tuple of `int32`;  e.g., `np.reshape(arr, (-1, 800, 3))`
-///   3. A scalar `int32`;     e.g., `np.reshape(arr, 3)`
-///
-/// Note that unlike other generating functions, one of the dimensions in the shape can be negative.
-pub fn ndarray_reshape<'ctx, G: CodeGenerator + ?Sized>(
-    generator: &mut G,
-    ctx: &mut CodeGenContext<'ctx, '_>,
-    x1: (Type, BasicValueEnum<'ctx>),
-    shape: (Type, BasicValueEnum<'ctx>),
-) -> Result<BasicValueEnum<'ctx>, String> {
-    const FN_NAME: &str = "ndarray_reshape";
-    let (x1_ty, x1) = x1;
-    let (_, shape) = shape;
-
-    let llvm_usize = generator.get_size_type(ctx.ctx);
-
-    if let BasicValueEnum::PointerValue(n1) = x1 {
-        let (elem_ty, _) = unpack_ndarray_var_tys(&mut ctx.unifier, x1_ty);
-        let n1 = NDArrayValue::from_ptr_val(n1, llvm_usize, None);
-        let n_sz = call_ndarray_calc_size(generator, ctx, &n1.dim_sizes(), (None, None));
-
-        let acc = generator.gen_var_alloc(ctx, llvm_usize.into(), None)?;
-        let num_neg = generator.gen_var_alloc(ctx, llvm_usize.into(), None)?;
-        ctx.builder.build_store(acc, llvm_usize.const_int(1, false)).unwrap();
-        ctx.builder.build_store(num_neg, llvm_usize.const_zero()).unwrap();
-
-        let out = match shape {
-            BasicValueEnum::PointerValue(shape_list_ptr)
-                if ListValue::is_instance(shape_list_ptr, llvm_usize).is_ok() =>
-            {
-                // 1. A list of ints; e.g., `np.reshape(arr, [int64(600), int64(800, -1])`
-
-                let shape_list = ListValue::from_ptr_val(shape_list_ptr, llvm_usize, None);
-                // Check for -1 in dimensions
-                gen_for_callback_incrementing(
-                    generator,
-                    ctx,
-                    None,
-                    llvm_usize.const_zero(),
-                    (shape_list.load_size(ctx, None), false),
-                    |generator, ctx, _, idx| {
-                        let ele =
-                            shape_list.data().get(ctx, generator, &idx, None).into_int_value();
-                        let ele = ctx.builder.build_int_s_extend(ele, llvm_usize, "").unwrap();
-
-                        gen_if_else_expr_callback(
-                            generator,
-                            ctx,
-                            |_, ctx| {
-                                Ok(ctx
-                                    .builder
-                                    .build_int_compare(
-                                        IntPredicate::SLT,
-                                        ele,
-                                        llvm_usize.const_zero(),
-                                        "",
-                                    )
-                                    .unwrap())
-                            },
-                            |_, ctx| -> Result<Option<IntValue>, String> {
-                                let num_neg_value =
-                                    ctx.builder.build_load(num_neg, "").unwrap().into_int_value();
-                                let num_neg_value = ctx
-                                    .builder
-                                    .build_int_add(
-                                        num_neg_value,
-                                        llvm_usize.const_int(1, false),
-                                        "",
-                                    )
-                                    .unwrap();
-                                ctx.builder.build_store(num_neg, num_neg_value).unwrap();
-                                Ok(None)
-                            },
-                            |_, ctx| {
-                                let acc_value =
-                                    ctx.builder.build_load(acc, "").unwrap().into_int_value();
-                                let acc_value =
-                                    ctx.builder.build_int_mul(acc_value, ele, "").unwrap();
-                                ctx.builder.build_store(acc, acc_value).unwrap();
-                                Ok(None)
-                            },
-                        )?;
-                        Ok(())
-                    },
-                    llvm_usize.const_int(1, false),
-                )?;
-                let acc_val = ctx.builder.build_load(acc, "").unwrap().into_int_value();
-                let rem = ctx.builder.build_int_unsigned_div(n_sz, acc_val, "").unwrap();
-                // Generate the output shape by filling -1 with `rem`
-                create_ndarray_dyn_shape(
-                    generator,
-                    ctx,
-                    elem_ty,
-                    &shape_list,
-                    |_, ctx, _| Ok(shape_list.load_size(ctx, None)),
-                    |generator, ctx, shape_list, idx| {
-                        let dim =
-                            shape_list.data().get(ctx, generator, &idx, None).into_int_value();
-                        let dim = ctx.builder.build_int_s_extend(dim, llvm_usize, "").unwrap();
-
-                        Ok(gen_if_else_expr_callback(
-                            generator,
-                            ctx,
-                            |_, ctx| {
-                                Ok(ctx
-                                    .builder
-                                    .build_int_compare(
-                                        IntPredicate::SLT,
-                                        dim,
-                                        llvm_usize.const_zero(),
-                                        "",
-                                    )
-                                    .unwrap())
-                            },
-                            |_, _| Ok(Some(rem)),
-                            |_, _| Ok(Some(dim)),
-                        )?
-                        .unwrap()
-                        .into_int_value())
-                    },
-                )
-            }
-            BasicValueEnum::StructValue(shape_tuple) => {
-                // 2. A tuple of `int32`;  e.g., `np.reshape(arr, (-1, 800, 3))`
-
-                let ndims = shape_tuple.get_type().count_fields();
-                // Check for -1 in dims
-                for dim_i in 0..ndims {
-                    let dim = ctx
-                        .builder
-                        .build_extract_value(shape_tuple, dim_i, "")
-                        .unwrap()
-                        .into_int_value();
-                    let dim = ctx.builder.build_int_s_extend(dim, llvm_usize, "").unwrap();
-
-                    gen_if_else_expr_callback(
-                        generator,
-                        ctx,
-                        |_, ctx| {
-                            Ok(ctx
-                                .builder
-                                .build_int_compare(
-                                    IntPredicate::SLT,
-                                    dim,
-                                    llvm_usize.const_zero(),
-                                    "",
-                                )
-                                .unwrap())
-                        },
-                        |_, ctx| -> Result<Option<IntValue>, String> {
-                            let num_negs =
-                                ctx.builder.build_load(num_neg, "").unwrap().into_int_value();
-                            let num_negs = ctx
-                                .builder
-                                .build_int_add(num_negs, llvm_usize.const_int(1, false), "")
-                                .unwrap();
-                            ctx.builder.build_store(num_neg, num_negs).unwrap();
-                            Ok(None)
-                        },
-                        |_, ctx| {
-                            let acc_val = ctx.builder.build_load(acc, "").unwrap().into_int_value();
-                            let acc_val = ctx.builder.build_int_mul(acc_val, dim, "").unwrap();
-                            ctx.builder.build_store(acc, acc_val).unwrap();
-                            Ok(None)
-                        },
-                    )?;
-                }
-
-                let acc_val = ctx.builder.build_load(acc, "").unwrap().into_int_value();
-                let rem = ctx.builder.build_int_unsigned_div(n_sz, acc_val, "").unwrap();
-                let mut shape = Vec::with_capacity(ndims as usize);
-
-                // Reconstruct shape filling negatives with rem
-                for dim_i in 0..ndims {
-                    let dim = ctx
-                        .builder
-                        .build_extract_value(shape_tuple, dim_i, "")
-                        .unwrap()
-                        .into_int_value();
-                    let dim = ctx.builder.build_int_s_extend(dim, llvm_usize, "").unwrap();
-
-                    let dim = gen_if_else_expr_callback(
-                        generator,
-                        ctx,
-                        |_, ctx| {
-                            Ok(ctx
-                                .builder
-                                .build_int_compare(
-                                    IntPredicate::SLT,
-                                    dim,
-                                    llvm_usize.const_zero(),
-                                    "",
-                                )
-                                .unwrap())
-                        },
-                        |_, _| Ok(Some(rem)),
-                        |_, _| Ok(Some(dim)),
-                    )?
-                    .unwrap()
-                    .into_int_value();
-                    shape.push(dim);
-                }
-                create_ndarray_const_shape(generator, ctx, elem_ty, shape.as_slice())
-            }
-            BasicValueEnum::IntValue(shape_int) => {
-                // 3. A scalar `int32`;     e.g., `np.reshape(arr, 3)`
-                let shape_int = gen_if_else_expr_callback(
-                    generator,
-                    ctx,
-                    |_, ctx| {
-                        Ok(ctx
-                            .builder
-                            .build_int_compare(
-                                IntPredicate::SLT,
-                                shape_int,
-                                llvm_usize.const_zero(),
-                                "",
-                            )
-                            .unwrap())
-                    },
-                    |_, _| Ok(Some(n_sz)),
-                    |_, ctx| {
-                        Ok(Some(ctx.builder.build_int_s_extend(shape_int, llvm_usize, "").unwrap()))
-                    },
-                )?
-                .unwrap()
-                .into_int_value();
-                create_ndarray_const_shape(generator, ctx, elem_ty, &[shape_int])
-            }
-            _ => unreachable!(),
-        }
-        .unwrap();
-
-        // Only allow one dimension to be negative
-        let num_negs = ctx.builder.build_load(num_neg, "").unwrap().into_int_value();
-        ctx.make_assert(
-            generator,
-            ctx.builder
-                .build_int_compare(IntPredicate::ULT, num_negs, llvm_usize.const_int(2, false), "")
-                .unwrap(),
-            "0:ValueError",
-            "can only specify one unknown dimension",
-            [None, None, None],
-            ctx.current_loc,
-        );
-
-        // The new shape must be compatible with the old shape
-        let out_sz = call_ndarray_calc_size(generator, ctx, &out.dim_sizes(), (None, None));
-        ctx.make_assert(
-            generator,
-            ctx.builder.build_int_compare(IntPredicate::EQ, out_sz, n_sz, "").unwrap(),
-            "0:ValueError",
-            "cannot reshape array of size {0} into provided shape of size {1}",
-            [Some(n_sz), Some(out_sz), None],
-            ctx.current_loc,
-        );
-
-        gen_for_callback_incrementing(
-            generator,
-            ctx,
-            None,
-            llvm_usize.const_zero(),
-            (n_sz, false),
-            |generator, ctx, _, idx| {
-                let elem = unsafe { n1.data().get_unchecked(ctx, generator, &idx, None) };
-                unsafe { out.data().set_unchecked(ctx, generator, &idx, elem) };
-                Ok(())
-            },
-            llvm_usize.const_int(1, false),
-        )?;
-
-        Ok(out.as_base_value().into())
-    } else {
-        unreachable!(
-            "{FN_NAME}() not supported for '{}'",
-            format!("'{}'", ctx.unifier.stringify(x1_ty))
-        )
-    }
-}
-
 /// Generates LLVM IR for `ndarray.dot`.
 /// Calculate inner product of two vectors or literals
 /// For matrix multiplication use `np_matmul`
@@ -2483,7 +2052,7 @@ pub fn ndarray_dot<'ctx, G: CodeGenerator + ?Sized>(
                             .build_float_mul(e1, elem2.into_float_value(), "")
                             .unwrap()
                             .as_basic_value_enum(),
-                        _ => unreachable!(),
+                        _ => codegen_unreachable!(ctx),
                     };
                     let acc_val = ctx.builder.build_load(acc, "").unwrap();
                     let acc_val = match acc_val {
@@ -2497,7 +2066,7 @@ pub fn ndarray_dot<'ctx, G: CodeGenerator + ?Sized>(
                             .build_float_add(e1, product.into_float_value(), "")
                             .unwrap()
                             .as_basic_value_enum(),
-                        _ => unreachable!(),
+                        _ => codegen_unreachable!(ctx),
                     };
                     ctx.builder.build_store(acc, acc_val).unwrap();
 
@@ -2514,7 +2083,8 @@ pub fn ndarray_dot<'ctx, G: CodeGenerator + ?Sized>(
         (BasicValueEnum::FloatValue(e1), BasicValueEnum::FloatValue(e2)) => {
             Ok(ctx.builder.build_float_mul(e1, e2, "").unwrap().as_basic_value_enum())
         }
-        _ => unreachable!(
+        _ => codegen_unreachable!(
+            ctx,
             "{FN_NAME}() not supported for '{}'",
             format!("'{}'", ctx.unifier.stringify(x1_ty))
         ),

nac3core/src/codegen/object/any.rs

@@ -0,0 +1,12 @@
+use inkwell::values::BasicValueEnum;
+
+use crate::typecheck::typedef::Type;
+
+/// A NAC3 LLVM Python object of any type.
+#[derive(Debug, Clone, Copy)]
+pub struct AnyObject<'ctx> {
+    /// Typechecker type of the object.
+    pub ty: Type,
+    /// LLVM value of the object.
+    pub value: BasicValueEnum<'ctx>,
+}

nac3core/src/codegen/object/list.rs

@@ -0,0 +1,87 @@
+use crate::{
+    codegen::{model::*, CodeGenContext, CodeGenerator},
+    typecheck::typedef::{iter_type_vars, Type, TypeEnum},
+};
+
+use super::any::AnyObject;
+
+/// Fields of [`List`]
+pub struct ListFields<'ctx, F: FieldTraversal<'ctx>, Item: Model<'ctx>> {
+    /// Array pointer to content
+    pub items: F::Output<Ptr<Item>>,
+    /// Number of items in the array
+    pub len: F::Output<Int<SizeT>>,
+}
+
+/// A list in NAC3.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct List<Item> {
+    /// Model of the list items
+    pub item: Item,
+}
+
+impl<'ctx, Item: Model<'ctx>> StructKind<'ctx> for List<Item> {
+    type Fields<F: FieldTraversal<'ctx>> = ListFields<'ctx, F, Item>;
+
+    fn iter_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+        Self::Fields {
+            items: traversal.add("items", Ptr(self.item)),
+            len: traversal.add_auto("len"),
+        }
+    }
+}
+
+impl<'ctx, Item: Model<'ctx>> Instance<'ctx, Ptr<Struct<List<Item>>>> {
+    /// Cast the items pointer to `uint8_t*`.
+    pub fn with_pi8_items<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Ptr<Struct<List<Int<Byte>>>>> {
+        self.pointer_cast(generator, ctx, Struct(List { item: Int(Byte) }))
+    }
+}
+
+/// A NAC3 Python List object.
+#[derive(Debug, Clone, Copy)]
+pub struct ListObject<'ctx> {
+    /// Typechecker type of the list items
+    pub item_type: Type,
+    pub instance: Instance<'ctx, Ptr<Struct<List<Any<'ctx>>>>>,
+}
+
+impl<'ctx> ListObject<'ctx> {
+    /// Create a [`ListObject`] from an LLVM value and its typechecker [`Type`].
+    pub fn from_object<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        object: AnyObject<'ctx>,
+    ) -> Self {
+        // Check typechecker type and extract `item_type`
+        let item_type = match &*ctx.unifier.get_ty(object.ty) {
+            TypeEnum::TObj { obj_id, params, .. }
+                if *obj_id == ctx.primitives.list.obj_id(&ctx.unifier).unwrap() =>
+            {
+                iter_type_vars(params).next().unwrap().ty // Extract `item_type`
+            }
+            _ => {
+                panic!("Expecting type to be a list, but got {}", ctx.unifier.stringify(object.ty))
+            }
+        };
+
+        let plist = Ptr(Struct(List { item: Any(ctx.get_llvm_type(generator, item_type)) }));
+
+        // Create object
+        let value = plist.check_value(generator, ctx.ctx, object.value).unwrap();
+        ListObject { item_type, instance: value }
+    }
+
+    /// Get the `len()` of this list.
+    pub fn len<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<SizeT>> {
+        self.instance.get(generator, ctx, |f| f.len)
+    }
+}

nac3core/src/codegen/object/mod.rs

@@ -0,0 +1,5 @@
+pub mod any;
+pub mod list;
+pub mod ndarray;
+pub mod tuple;
+pub mod utils;

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

@@ -0,0 +1,184 @@
+use super::NDArrayObject;
+use crate::{
+    codegen::{
+        irrt::{
+            call_nac3_ndarray_array_set_and_validate_list_shape,
+            call_nac3_ndarray_array_write_list_to_array,
+        },
+        model::*,
+        object::{any::AnyObject, list::ListObject},
+        stmt::gen_if_else_expr_callback,
+        CodeGenContext, CodeGenerator,
+    },
+    toplevel::helper::{arraylike_flatten_element_type, arraylike_get_ndims},
+    typecheck::typedef::{Type, TypeEnum},
+};
+
+/// Get the expected `dtype` and `ndims` of the ndarray returned by `np_array(list)`.
+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> {
+    /// Implementation of `np_array(<list>, copy=True)`
+    fn make_np_array_list_copy_true_impl<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        list: ListObject<'ctx>,
+    ) -> Self {
+        let (dtype, ndims_int) = get_list_object_dtype_and_ndims(ctx, list);
+        let list_value = list.instance.with_pi8_items(generator, ctx);
+
+        // Validate `list` has a consistent shape.
+        // Raise an exception if `list` is something abnormal like `[[1, 2], [3]]`.
+        // If `list` has a consistent shape, deduce the shape and write it to `shape`.
+        let ndims = Int(SizeT).const_int(generator, ctx.ctx, ndims_int, false);
+        let shape = Int(SizeT).array_alloca(generator, ctx, ndims.value);
+        call_nac3_ndarray_array_set_and_validate_list_shape(
+            generator, ctx, list_value, ndims, shape,
+        );
+
+        let ndarray = NDArrayObject::alloca(generator, ctx, dtype, ndims_int);
+        ndarray.copy_shape_from_array(generator, ctx, shape);
+        ndarray.create_data(generator, ctx);
+
+        // Copy all contents from the list.
+        call_nac3_ndarray_array_write_list_to_array(generator, ctx, list_value, ndarray.instance);
+
+        ndarray
+    }
+
+    /// Implementation of `np_array(<list>, copy=None)`
+    fn make_np_array_list_copy_none_impl<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        list: ListObject<'ctx>,
+    ) -> Self {
+        // np_array without copying is only possible `list` is not nested.
+        //
+        // If `list` is `list[T]`, we can create an ndarray with `data` set
+        // to the array pointer of `list`.
+        //
+        // If `list` is `list[list[T]]` or worse, copy.
+
+        let (dtype, ndims) = get_list_object_dtype_and_ndims(ctx, list);
+        if ndims == 1 {
+            // `list` is not nested
+            let ndarray = NDArrayObject::alloca(generator, ctx, dtype, 1);
+
+            // Set data
+            let data = list.instance.get(generator, ctx, |f| f.items).cast_to_pi8(generator, ctx);
+            ndarray.instance.set(ctx, |f| f.data, data);
+
+            // ndarray->shape[0] = list->len;
+            let shape = ndarray.instance.get(generator, ctx, |f| f.shape);
+            let list_len = list.instance.get(generator, ctx, |f| f.len);
+            shape.set_index_const(ctx, 0, list_len);
+
+            // Set strides, the `data` is contiguous
+            ndarray.set_strides_contiguous(generator, ctx);
+
+            ndarray
+        } else {
+            // `list` is nested, copy
+            NDArrayObject::make_np_array_list_copy_true_impl(generator, ctx, list)
+        }
+    }
+
+    /// Implementation of `np_array(<list>, copy=copy)`
+    fn make_np_array_list_impl<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        list: ListObject<'ctx>,
+        copy: Instance<'ctx, Int<Bool>>,
+    ) -> Self {
+        let (dtype, ndims) = get_list_object_dtype_and_ndims(ctx, list);
+
+        let ndarray = gen_if_else_expr_callback(
+            generator,
+            ctx,
+            |_generator, _ctx| Ok(copy.value),
+            |generator, ctx| {
+                let ndarray =
+                    NDArrayObject::make_np_array_list_copy_true_impl(generator, ctx, list);
+                Ok(Some(ndarray.instance.value))
+            },
+            |generator, ctx| {
+                let ndarray =
+                    NDArrayObject::make_np_array_list_copy_none_impl(generator, ctx, list);
+                Ok(Some(ndarray.instance.value))
+            },
+        )
+        .unwrap()
+        .unwrap();
+
+        NDArrayObject::from_value_and_unpacked_types(generator, ctx, ndarray, dtype, ndims)
+    }
+
+    /// Implementation of `np_array(<ndarray>, copy=copy)`.
+    pub fn make_np_array_ndarray_impl<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        ndarray: NDArrayObject<'ctx>,
+        copy: Instance<'ctx, Int<Bool>>,
+    ) -> Self {
+        let ndarray_val = gen_if_else_expr_callback(
+            generator,
+            ctx,
+            |_generator, _ctx| Ok(copy.value),
+            |generator, ctx| {
+                let ndarray = ndarray.make_copy(generator, ctx); // Force copy
+                Ok(Some(ndarray.instance.value))
+            },
+            |_generator, _ctx| {
+                // No need to copy. Return `ndarray` itself.
+                Ok(Some(ndarray.instance.value))
+            },
+        )
+        .unwrap()
+        .unwrap();
+
+        NDArrayObject::from_value_and_unpacked_types(
+            generator,
+            ctx,
+            ndarray_val,
+            ndarray.dtype,
+            ndarray.ndims,
+        )
+    }
+
+    /// Create a new ndarray like `np.array()`.
+    ///
+    /// NOTE: The `ndmin` argument is not here. You may want to
+    /// do [`NDArrayObject::atleast_nd`] to achieve that.
+    pub fn make_np_array<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        object: AnyObject<'ctx>,
+        copy: Instance<'ctx, Int<Bool>>,
+    ) -> Self {
+        match &*ctx.unifier.get_ty(object.ty) {
+            TypeEnum::TObj { obj_id, .. }
+                if *obj_id == ctx.primitives.list.obj_id(&ctx.unifier).unwrap() =>
+            {
+                let list = ListObject::from_object(generator, ctx, object);
+                NDArrayObject::make_np_array_list_impl(generator, ctx, list, copy)
+            }
+            TypeEnum::TObj { obj_id, .. }
+                if *obj_id == ctx.primitives.ndarray.obj_id(&ctx.unifier).unwrap() =>
+            {
+                let ndarray = NDArrayObject::from_object(generator, ctx, object);
+                NDArrayObject::make_np_array_ndarray_impl(generator, ctx, ndarray, copy)
+            }
+            _ => panic!("Unrecognized object type: {}", ctx.unifier.stringify(object.ty)), // Typechecker ensures this
+        }
+    }
+}

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

@@ -0,0 +1,139 @@
+use itertools::Itertools;
+
+use crate::codegen::{
+    irrt::{call_nac3_ndarray_broadcast_shapes, call_nac3_ndarray_broadcast_to},
+    model::*,
+    CodeGenContext, CodeGenerator,
+};
+
+use super::NDArrayObject;
+
+/// Fields of [`ShapeEntry`]
+pub struct ShapeEntryFields<'ctx, F: FieldTraversal<'ctx>> {
+    pub ndims: F::Output<Int<SizeT>>,
+    pub shape: F::Output<Ptr<Int<SizeT>>>,
+}
+
+/// An IRRT structure used in broadcasting.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct ShapeEntry;
+
+impl<'ctx> StructKind<'ctx> for ShapeEntry {
+    type Fields<F: FieldTraversal<'ctx>> = ShapeEntryFields<'ctx, F>;
+
+    fn iter_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+        Self::Fields { ndims: traversal.add_auto("ndims"), shape: traversal.add_auto("shape") }
+    }
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Create a broadcast view on this ndarray with a target shape.
+    ///
+    /// The input shape will be checked to make sure that it contains no negative values.
+    ///
+    /// * `target_ndims` - The ndims type after broadcasting to the given shape.
+    ///   The caller has to figure this out for this function.
+    /// * `target_shape` - An array pointer pointing to the target shape.
+    #[must_use]
+    pub fn broadcast_to<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        target_ndims: u64,
+        target_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) -> Self {
+        let broadcast_ndarray = NDArrayObject::alloca(generator, ctx, self.dtype, target_ndims);
+        broadcast_ndarray.copy_shape_from_array(generator, ctx, target_shape);
+
+        call_nac3_ndarray_broadcast_to(generator, ctx, self.instance, broadcast_ndarray.instance);
+        broadcast_ndarray
+    }
+}
+/// A result produced by [`broadcast_all_ndarrays`]
+#[derive(Debug, Clone)]
+pub struct BroadcastAllResult<'ctx> {
+    /// The statically known `ndims` of the broadcast result.
+    pub ndims: u64,
+    /// The broadcasting shape.
+    pub shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    /// Broadcasted views on the inputs.
+    ///
+    /// All of them will have `shape` [`BroadcastAllResult::shape`] and
+    /// `ndims` [`BroadcastAllResult::ndims`]. The length of the vector
+    /// is the same as the input.
+    pub ndarrays: Vec<NDArrayObject<'ctx>>,
+}
+
+/// Helper function to call `call_nac3_ndarray_broadcast_shapes`
+fn broadcast_shapes<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    in_shape_entries: &[(Instance<'ctx, Ptr<Int<SizeT>>>, u64)], // (shape, shape's length/ndims)
+    broadcast_ndims: u64,
+    broadcast_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+) {
+    // Prepare input shape entries to be passed to `call_nac3_ndarray_broadcast_shapes`.
+    let num_shape_entries = Int(SizeT).const_int(
+        generator,
+        ctx.ctx,
+        u64::try_from(in_shape_entries.len()).unwrap(),
+        false,
+    );
+    let shape_entries = Struct(ShapeEntry).array_alloca(generator, ctx, num_shape_entries.value);
+    for (i, (in_shape, in_ndims)) in in_shape_entries.iter().enumerate() {
+        let pshape_entry = shape_entries.offset_const(ctx, i64::try_from(i).unwrap());
+
+        let in_ndims = Int(SizeT).const_int(generator, ctx.ctx, *in_ndims, false);
+        pshape_entry.set(ctx, |f| f.ndims, in_ndims);
+
+        pshape_entry.set(ctx, |f| f.shape, *in_shape);
+    }
+
+    let broadcast_ndims = Int(SizeT).const_int(generator, ctx.ctx, broadcast_ndims, false);
+    call_nac3_ndarray_broadcast_shapes(
+        generator,
+        ctx,
+        num_shape_entries,
+        shape_entries,
+        broadcast_ndims,
+        broadcast_shape,
+    );
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Broadcast all ndarrays according to `np.broadcast()` and return a [`BroadcastAllResult`]
+    /// containing all the information of the result of the broadcast operation.
+    pub fn broadcast<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        ndarrays: &[Self],
+    ) -> BroadcastAllResult<'ctx> {
+        assert!(!ndarrays.is_empty());
+
+        // Infer the broadcast output ndims.
+        let broadcast_ndims_int = ndarrays.iter().map(|ndarray| ndarray.ndims).max().unwrap();
+
+        let broadcast_ndims = Int(SizeT).const_int(generator, ctx.ctx, broadcast_ndims_int, false);
+        let broadcast_shape = Int(SizeT).array_alloca(generator, ctx, broadcast_ndims.value);
+
+        let shape_entries = ndarrays
+            .iter()
+            .map(|ndarray| (ndarray.instance.get(generator, ctx, |f| f.shape), ndarray.ndims))
+            .collect_vec();
+        broadcast_shapes(generator, ctx, &shape_entries, broadcast_ndims_int, broadcast_shape);
+
+        // Broadcast all the inputs to shape `dst_shape`.
+        let broadcast_ndarrays: Vec<_> = ndarrays
+            .iter()
+            .map(|ndarray| {
+                ndarray.broadcast_to(generator, ctx, broadcast_ndims_int, broadcast_shape)
+            })
+            .collect_vec();
+
+        BroadcastAllResult {
+            ndims: broadcast_ndims_int,
+            shape: broadcast_shape,
+            ndarrays: broadcast_ndarrays,
+        }
+    }
+}

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

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

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

@@ -0,0 +1,227 @@
+use crate::codegen::{
+    irrt::call_nac3_ndarray_index,
+    model::*,
+    object::utils::slice::{RustSlice, Slice},
+    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::Output<Int<NDIndexType>>,
+    pub data: F::Output<Ptr<Int<Byte>>>,
+}
+
+/// An IRRT representation of an ndarray subscript index.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct NDIndex;
+
+impl<'ctx> StructKind<'ctx> for NDIndex {
+    type Fields<F: FieldTraversal<'ctx>> = NDIndexFields<'ctx, F>;
+
+    fn iter_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+        Self::Fields { type_: traversal.add_auto("type"), data: traversal.add_auto("data") }
+    }
+}
+
+// A convenience enum representing a [`NDIndex`].
+#[derive(Debug, Clone)]
+pub enum RustNDIndex<'ctx> {
+    SingleElement(Instance<'ctx, Int<Int32>>),
+    Slice(RustSlice<'ctx, Int32>),
+    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,
+        }
+    }
+
+    /// Serialize this [`RustNDIndex`] by writing it into an LLVM [`NDIndex`].
+    fn write_to_ndindex<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        dst_ndindex_ptr: Instance<'ctx, Ptr<Struct<NDIndex>>>,
+    ) {
+        // Set `dst_ndindex_ptr->type`
+        dst_ndindex_ptr.gep(ctx, |f| f.type_).store(
+            ctx,
+            Int(NDIndexType::default()).const_int(generator, ctx.ctx, self.get_type_id(), false),
+        );
+
+        // Set `dst_ndindex_ptr->data`
+        match self {
+            RustNDIndex::SingleElement(in_index) => {
+                let index_ptr = Int(Int32).alloca(generator, ctx);
+                index_ptr.store(ctx, *in_index);
+
+                dst_ndindex_ptr
+                    .gep(ctx, |f| f.data)
+                    .store(ctx, index_ptr.pointer_cast(generator, ctx, Int(Byte)));
+            }
+            RustNDIndex::Slice(in_rust_slice) => {
+                let user_slice_ptr = Struct(Slice(Int32)).alloca(generator, ctx);
+                in_rust_slice.write_to_slice(generator, ctx, user_slice_ptr);
+
+                dst_ndindex_ptr
+                    .gep(ctx, |f| f.data)
+                    .store(ctx, user_slice_ptr.pointer_cast(generator, ctx, Int(Byte)));
+            }
+            RustNDIndex::NewAxis | RustNDIndex::Ellipsis => {}
+        }
+    }
+
+    /// Serialize a list of `RustNDIndex` as a newly allocated LLVM array of `NDIndex`.
+    pub fn make_ndindices<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        in_ndindices: &[RustNDIndex<'ctx>],
+    ) -> (Instance<'ctx, Int<SizeT>>, Instance<'ctx, Ptr<Struct<NDIndex>>>) {
+        let ndindex_model = Struct(NDIndex);
+
+        // Allocate the LLVM ndindices.
+        let num_ndindices =
+            Int(SizeT).const_int(generator, ctx.ctx, in_ndindices.len() as u64, false);
+        let ndindices = ndindex_model.array_alloca(generator, ctx, num_ndindices.value);
+
+        // Initialize all of them.
+        for (i, in_ndindex) in in_ndindices.iter().enumerate() {
+            let pndindex = ndindices.offset_const(ctx, i64::try_from(i).unwrap());
+            in_ndindex.write_to_ndindex(generator, ctx, pndindex);
+        }
+
+        (num_ndindices, ndindices)
+    }
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Get the expected `ndims` after indexing with `indices`.
+    #[must_use]
+    fn deduce_ndims_after_indexing_with(&self, indices: &[RustNDIndex<'ctx>]) -> u64 {
+        let mut ndims = self.ndims;
+        for index in indices {
+            match index {
+                RustNDIndex::SingleElement(_) => {
+                    ndims -= 1; // Single elements decrements ndims
+                }
+                RustNDIndex::NewAxis => {
+                    ndims += 1; // `np.newaxis` / `none` adds a new axis
+                }
+                RustNDIndex::Ellipsis | RustNDIndex::Slice(_) => {}
+            }
+        }
+        ndims
+    }
+
+    /// Index into the ndarray, and return a newly-allocated view on this ndarray.
+    ///
+    /// This function behaves like NumPy's ndarray indexing, but if the indices index
+    /// into a single element, an unsized ndarray is returned.
+    #[must_use]
+    pub fn index<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        indices: &[RustNDIndex<'ctx>],
+    ) -> Self {
+        let dst_ndims = self.deduce_ndims_after_indexing_with(indices);
+        let dst_ndarray = NDArrayObject::alloca(generator, ctx, self.dtype, dst_ndims);
+
+        let (num_indices, indices) = RustNDIndex::make_ndindices(generator, ctx, indices);
+        call_nac3_ndarray_index(
+            generator,
+            ctx,
+            num_indices,
+            indices,
+            self.instance,
+            dst_ndarray.instance,
+        );
+
+        dst_ndarray
+    }
+}
+
+pub mod util {
+    use itertools::Itertools;
+    use nac3parser::ast::{Expr, ExprKind};
+
+    use crate::{
+        codegen::{model::*, object::utils::slice::util::gen_slice, CodeGenContext, CodeGenerator},
+        typecheck::typedef::Type,
+    };
+
+    use super::RustNDIndex;
+
+    /// Generate LLVM code to transform an ndarray subscript expression to
+    /// its list of [`RustNDIndex`]
+    ///
+    /// i.e.,
+    /// ```python
+    /// my_ndarray[::3, 1, :2:]
+    ///            ^^^^^^^^^^^ Then these into a three `RustNDIndex`es
+    /// ```
+    pub fn gen_ndarray_subscript_ndindices<'ctx, G: CodeGenerator>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        subscript: &Expr<Option<Type>>,
+    ) -> Result<Vec<RustNDIndex<'ctx>>, String> {
+        // TODO: Support https://numpy.org/doc/stable/user/basics.indexing.html#dimensional-indexing-tools
+
+        // Annoying notes about `slice`
+        //  - `my_array[5]`
+        //    - slice is a `Constant`
+        //  - `my_array[:5]`
+        //    - slice is a `Slice`
+        //  - `my_array[:]`
+        //    - slice is a `Slice`, but lower upper step would all be `Option::None`
+        //  - `my_array[:, :]`
+        //    - slice is now a `Tuple` of two `Slice`-s
+        //
+        // In summary:
+        //  - when there is a comma "," within [], `slice` will be a `Tuple` of the entries.
+        //  - when there is not comma "," within [] (i.e., just a single entry), `slice` will be that entry itself.
+        //
+        // So we first "flatten" out the slice expression
+        let index_exprs = match &subscript.node {
+            ExprKind::Tuple { elts, .. } => elts.iter().collect_vec(),
+            _ => vec![subscript],
+        };
+
+        // Process all index expressions
+        let mut rust_ndindices: Vec<RustNDIndex> = Vec::with_capacity(index_exprs.len()); // Not using iterators here because `?` is used here.
+        for index_expr in index_exprs {
+            // NOTE: Currently nac3core's slices do not have an object representation,
+            // so the code/implementation looks awkward - we have to do pattern matching on the expression
+            let ndindex = if let ExprKind::Slice { lower, upper, step } = &index_expr.node {
+                // Handle slices
+                let slice = gen_slice(generator, ctx, lower, upper, step)?;
+                RustNDIndex::Slice(slice)
+            } else {
+                // Treat and handle everything else as a single element index.
+                let index = generator.gen_expr(ctx, index_expr)?.unwrap().to_basic_value_enum(
+                    ctx,
+                    generator,
+                    ctx.primitives.int32, // Must be int32, this checks for illegal values
+                )?;
+                let index = Int(Int32).check_value(generator, ctx.ctx, index).unwrap();
+
+                RustNDIndex::SingleElement(index)
+            };
+            rust_ndindices.push(ndindex);
+        }
+        Ok(rust_ndindices)
+    }
+}

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

@@ -0,0 +1,219 @@
+use inkwell::values::BasicValueEnum;
+use itertools::Itertools;
+
+use crate::{
+    codegen::{
+        object::ndarray::{AnyObject, NDArrayObject},
+        stmt::gen_for_callback,
+        CodeGenContext, CodeGenerator,
+    },
+    typecheck::typedef::Type,
+};
+
+use super::{nditer::NDIterHandle, NDArrayOut, ScalarOrNDArray};
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Generate LLVM IR to broadcast `ndarray`s together, and starmap through them with `mapping` elementwise.
+    ///
+    /// `mapping` is an LLVM IR generator. The input of `mapping` is the list of elements when iterating through
+    /// the input `ndarrays` after broadcasting. The output of `mapping` is the result of the elementwise operation.
+    ///
+    /// `out` specifies whether the result should be a new ndarray or to be written an existing ndarray.
+    pub fn broadcast_starmap<'a, G, MappingFn>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, 'a>,
+        ndarrays: &[Self],
+        out: NDArrayOut<'ctx>,
+        mapping: MappingFn,
+    ) -> Result<Self, String>
+    where
+        G: CodeGenerator + ?Sized,
+        MappingFn: FnOnce(
+            &mut G,
+            &mut CodeGenContext<'ctx, 'a>,
+            &[BasicValueEnum<'ctx>],
+        ) -> Result<BasicValueEnum<'ctx>, String>,
+    {
+        // Broadcast inputs
+        let broadcast_result = NDArrayObject::broadcast(generator, ctx, ndarrays);
+
+        let out_ndarray = match out {
+            NDArrayOut::NewNDArray { dtype } => {
+                // Create a new ndarray based on the broadcast shape.
+                let result_ndarray =
+                    NDArrayObject::alloca(generator, ctx, dtype, broadcast_result.ndims);
+                result_ndarray.copy_shape_from_array(generator, ctx, broadcast_result.shape);
+                result_ndarray.create_data(generator, ctx);
+                result_ndarray
+            }
+            NDArrayOut::WriteToNDArray { ndarray: result_ndarray } => {
+                // Use an existing ndarray.
+
+                // Check that its shape is compatible with the broadcast shape.
+                result_ndarray.assert_can_be_written_by_out(
+                    generator,
+                    ctx,
+                    broadcast_result.ndims,
+                    broadcast_result.shape,
+                );
+                result_ndarray
+            }
+        };
+
+        // Map element-wise and store results into `mapped_ndarray`.
+        let nditer = NDIterHandle::new(generator, ctx, out_ndarray);
+        gen_for_callback(
+            generator,
+            ctx,
+            Some("broadcast_starmap"),
+            |generator, ctx| {
+                // Create NDIters for all broadcasted input ndarrays.
+                let other_nditers = broadcast_result
+                    .ndarrays
+                    .iter()
+                    .map(|ndarray| NDIterHandle::new(generator, ctx, *ndarray))
+                    .collect_vec();
+                Ok((nditer, other_nditers))
+            },
+            |generator, ctx, (out_nditer, _in_nditers)| {
+                // We can simply use `out_nditer`'s `has_element()`.
+                // `in_nditers`' `has_element()`s should return the same value.
+                Ok(out_nditer.has_element(generator, ctx).value)
+            },
+            |generator, ctx, _hooks, (out_nditer, in_nditers)| {
+                // Get all the scalars from the broadcasted input ndarrays, pass them to `mapping`,
+                // and write to `out_ndarray`.
+                let in_scalars = in_nditers
+                    .iter()
+                    .map(|nditer| nditer.get_scalar(generator, ctx).value)
+                    .collect_vec();
+
+                let result = mapping(generator, ctx, &in_scalars)?;
+
+                let p = out_nditer.get_pointer(generator, ctx);
+                ctx.builder.build_store(p, result).unwrap();
+
+                Ok(())
+            },
+            |generator, ctx, (out_nditer, in_nditers)| {
+                // Advance all iterators
+                out_nditer.next(generator, ctx);
+                in_nditers.iter().for_each(|nditer| nditer.next(generator, ctx));
+                Ok(())
+            },
+        )?;
+
+        Ok(out_ndarray)
+    }
+
+    /// Map through this ndarray with an elementwise function.
+    pub fn map<'a, G, Mapping>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, 'a>,
+        out: NDArrayOut<'ctx>,
+        mapping: Mapping,
+    ) -> Result<Self, String>
+    where
+        G: CodeGenerator + ?Sized,
+        Mapping: FnOnce(
+            &mut G,
+            &mut CodeGenContext<'ctx, 'a>,
+            BasicValueEnum<'ctx>,
+        ) -> Result<BasicValueEnum<'ctx>, String>,
+    {
+        NDArrayObject::broadcast_starmap(
+            generator,
+            ctx,
+            &[*self],
+            out,
+            |generator, ctx, scalars| mapping(generator, ctx, scalars[0]),
+        )
+    }
+}
+
+impl<'ctx> ScalarOrNDArray<'ctx> {
+    /// Starmap through a list of inputs using `mapping`, where an input could be an ndarray, a scalar.
+    ///
+    /// This function is very helpful when implementing NumPy functions that takes on either scalars or ndarrays or a mix of them
+    /// as their inputs and produces either an ndarray with broadcast, or a scalar if all its inputs are all scalars.
+    ///
+    /// For example ,this function can be used to implement `np.add`, which has the following behaviors:
+    /// - `np.add(3, 4) = 7` # (scalar, scalar) -> scalar
+    /// - `np.add(3, np.array([4, 5, 6]))` # (scalar, ndarray) -> ndarray; the first `scalar` is converted into an ndarray and broadcasted.
+    /// - `np.add(np.array([[1], [2], [3]]), np.array([[4, 5, 6]]))` # (ndarray, ndarray) -> ndarray; there is broadcasting.
+    ///
+    /// ## Details:
+    ///
+    /// If `inputs` are all [`ScalarOrNDArray::Scalar`], the output will be a [`ScalarOrNDArray::Scalar`] with type `ret_dtype`.
+    ///
+    /// Otherwise (if there are any [`ScalarOrNDArray::NDArray`] in `inputs`), all inputs will be 'as-ndarray'-ed into ndarrays,
+    /// then all inputs (now all ndarrays) will be passed to [`NDArrayObject::broadcasting_starmap`] and **create** a new ndarray
+    /// with dtype `ret_dtype`.
+    pub fn broadcasting_starmap<'a, G, MappingFn>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, 'a>,
+        inputs: &[ScalarOrNDArray<'ctx>],
+        ret_dtype: Type,
+        mapping: MappingFn,
+    ) -> Result<ScalarOrNDArray<'ctx>, String>
+    where
+        G: CodeGenerator + ?Sized,
+        MappingFn: FnOnce(
+            &mut G,
+            &mut CodeGenContext<'ctx, 'a>,
+            &[BasicValueEnum<'ctx>],
+        ) -> Result<BasicValueEnum<'ctx>, String>,
+    {
+        // Check if all inputs are Scalars
+        let all_scalars: Option<Vec<_>> = inputs.iter().map(AnyObject::try_from).try_collect().ok();
+
+        if let Some(scalars) = all_scalars {
+            let scalars = scalars.iter().map(|scalar| scalar.value).collect_vec();
+            let value = mapping(generator, ctx, &scalars)?;
+
+            Ok(ScalarOrNDArray::Scalar(AnyObject { ty: ret_dtype, value }))
+        } else {
+            // Promote all input to ndarrays and map through them.
+            let inputs = inputs.iter().map(|input| input.to_ndarray(generator, ctx)).collect_vec();
+            let ndarray = NDArrayObject::broadcast_starmap(
+                generator,
+                ctx,
+                &inputs,
+                NDArrayOut::NewNDArray { dtype: ret_dtype },
+                mapping,
+            )?;
+            Ok(ScalarOrNDArray::NDArray(ndarray))
+        }
+    }
+
+    /// Map through this [`ScalarOrNDArray`] with an elementwise function.
+    ///
+    /// If this is a scalar, `mapping` will directly act on the scalar. This function will return a [`ScalarOrNDArray::Scalar`] of that result.
+    ///
+    /// If this is an ndarray, `mapping` will be applied to the elements of the ndarray. A new ndarray of the results will be created and
+    /// returned as a [`ScalarOrNDArray::NDArray`].
+    pub fn map<'a, G, Mapping>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, 'a>,
+        ret_dtype: Type,
+        mapping: Mapping,
+    ) -> Result<ScalarOrNDArray<'ctx>, String>
+    where
+        G: CodeGenerator + ?Sized,
+        Mapping: FnOnce(
+            &mut G,
+            &mut CodeGenContext<'ctx, 'a>,
+            BasicValueEnum<'ctx>,
+        ) -> Result<BasicValueEnum<'ctx>, String>,
+    {
+        ScalarOrNDArray::broadcasting_starmap(
+            generator,
+            ctx,
+            &[*self],
+            ret_dtype,
+            |generator, ctx, scalars| mapping(generator, ctx, scalars[0]),
+        )
+    }
+}

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

@@ -0,0 +1,638 @@
+pub mod array;
+pub mod broadcast;
+pub mod factory;
+pub mod indexing;
+pub mod map;
+pub mod nditer;
+pub mod shape_util;
+pub mod view;
+
+use inkwell::{
+    context::Context,
+    types::BasicType,
+    values::{BasicValue, BasicValueEnum, PointerValue},
+    AddressSpace,
+};
+
+use crate::{
+    codegen::{
+        irrt::{
+            call_nac3_ndarray_copy_data, call_nac3_ndarray_get_nth_pelement,
+            call_nac3_ndarray_get_pelement_by_indices, call_nac3_ndarray_is_c_contiguous,
+            call_nac3_ndarray_len, call_nac3_ndarray_nbytes,
+            call_nac3_ndarray_set_strides_by_shape, call_nac3_ndarray_size,
+            call_nac3_ndarray_util_assert_output_shape_same,
+        },
+        model::*,
+        CodeGenContext, CodeGenerator,
+    },
+    toplevel::{helper::extract_ndims, numpy::unpack_ndarray_var_tys},
+    typecheck::typedef::{Type, TypeEnum},
+};
+
+use super::{any::AnyObject, tuple::TupleObject};
+
+/// Fields of [`NDArray`]
+pub struct NDArrayFields<'ctx, F: FieldTraversal<'ctx>> {
+    pub data: F::Output<Ptr<Int<Byte>>>,
+    pub itemsize: F::Output<Int<SizeT>>,
+    pub ndims: F::Output<Int<SizeT>>,
+    pub shape: F::Output<Ptr<Int<SizeT>>>,
+    pub strides: F::Output<Ptr<Int<SizeT>>>,
+}
+
+/// A strided ndarray in NAC3.
+///
+/// See IRRT implementation for details about its fields.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct NDArray;
+
+impl<'ctx> StructKind<'ctx> for NDArray {
+    type Fields<F: FieldTraversal<'ctx>> = NDArrayFields<'ctx, F>;
+
+    fn iter_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+        Self::Fields {
+            data: traversal.add_auto("data"),
+            itemsize: traversal.add_auto("itemsize"),
+            ndims: traversal.add_auto("ndims"),
+            shape: traversal.add_auto("shape"),
+            strides: traversal.add_auto("strides"),
+        }
+    }
+}
+
+/// A NAC3 Python ndarray object.
+#[derive(Debug, Clone, Copy)]
+pub struct NDArrayObject<'ctx> {
+    pub dtype: Type,
+    pub ndims: u64,
+    pub instance: Instance<'ctx, Ptr<Struct<NDArray>>>,
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Attempt to convert an [`AnyObject`] into an [`NDArrayObject`].
+    pub fn from_object<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        object: AnyObject<'ctx>,
+    ) -> NDArrayObject<'ctx> {
+        let (dtype, ndims) = unpack_ndarray_var_tys(&mut ctx.unifier, object.ty);
+        let ndims = extract_ndims(&ctx.unifier, ndims);
+        Self::from_value_and_unpacked_types(generator, ctx, object.value, dtype, ndims)
+    }
+
+    /// Like [`NDArrayObject::from_object`] but you directly supply the ndarray's
+    /// `dtype` and `ndims`.
+    pub fn from_value_and_unpacked_types<V: BasicValue<'ctx>, G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        value: V,
+        dtype: Type,
+        ndims: u64,
+    ) -> Self {
+        let value = Ptr(Struct(NDArray)).check_value(generator, ctx.ctx, value).unwrap();
+        NDArrayObject { dtype, ndims, instance: value }
+    }
+
+    /// Get this ndarray's `ndims` as an LLVM constant.
+    pub fn ndims_llvm<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &'ctx Context,
+    ) -> Instance<'ctx, Int<SizeT>> {
+        Int(SizeT).const_int(generator, ctx, self.ndims, false)
+    }
+
+    /// Allocate an ndarray on the stack given its `ndims` and `dtype`.
+    ///
+    /// `shape` and `strides` will be automatically allocated onto the stack.
+    ///
+    /// The returned ndarray's content will be:
+    /// - `data`: uninitialized.
+    /// - `itemsize`: set to the `sizeof()` of `dtype`.
+    /// - `ndims`: set to the value of  `ndims`.
+    /// - `shape`: allocated with an array of length `ndims` with uninitialized values.
+    /// - `strides`: allocated with an array of length `ndims` with uninitialized values.
+    pub fn alloca<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        ndims: u64,
+    ) -> Self {
+        let ndarray = Struct(NDArray).alloca(generator, ctx);
+
+        let itemsize = ctx.get_llvm_type(generator, dtype).size_of().unwrap();
+        let itemsize = Int(SizeT).z_extend_or_truncate(generator, ctx, itemsize);
+        ndarray.set(ctx, |f| f.itemsize, itemsize);
+
+        let ndims_val = Int(SizeT).const_int(generator, ctx.ctx, ndims, false);
+        ndarray.set(ctx, |f| f.ndims, ndims_val);
+
+        let shape = Int(SizeT).array_alloca(generator, ctx, ndims_val.value);
+        ndarray.set(ctx, |f| f.shape, shape);
+
+        let strides = Int(SizeT).array_alloca(generator, ctx, ndims_val.value);
+        ndarray.set(ctx, |f| f.strides, strides);
+
+        NDArrayObject { dtype, ndims, instance: ndarray }
+    }
+
+    /// Convenience function. Allocate an [`NDArrayObject`] with a statically known shape.
+    ///
+    /// The returned [`NDArrayObject`]'s `data` and `strides` are uninitialized.
+    pub fn alloca_constant_shape<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        shape: &[u64],
+    ) -> Self {
+        let ndarray = NDArrayObject::alloca(generator, ctx, dtype, shape.len() as u64);
+
+        // Write shape
+        let dst_shape = ndarray.instance.get(generator, ctx, |f| f.shape);
+        for (i, dim) in shape.iter().enumerate() {
+            let dim = Int(SizeT).const_int(generator, ctx.ctx, *dim, false);
+            dst_shape.offset_const(ctx, i64::try_from(i).unwrap()).store(ctx, dim);
+        }
+
+        ndarray
+    }
+
+    /// Convenience function. Allocate an [`NDArrayObject`] with a dynamically known shape.
+    ///
+    /// The returned [`NDArrayObject`]'s `data` and `strides` are uninitialized.
+    pub fn alloca_dynamic_shape<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        dtype: Type,
+        shape: &[Instance<'ctx, Int<SizeT>>],
+    ) -> Self {
+        let ndarray = NDArrayObject::alloca(generator, ctx, dtype, shape.len() as u64);
+
+        // Write shape
+        let dst_shape = ndarray.instance.get(generator, ctx, |f| f.shape);
+        for (i, dim) in shape.iter().enumerate() {
+            dst_shape.offset_const(ctx, i64::try_from(i).unwrap()).store(ctx, *dim);
+        }
+
+        ndarray
+    }
+
+    /// Initialize an ndarray's `data` by allocating a buffer on the stack.
+    /// The allocated data buffer is considered to be *owned* by the ndarray.
+    ///
+    /// `strides` of the ndarray will also be updated with `set_strides_by_shape`.
+    ///
+    /// `shape` and `itemsize` of the ndarray ***must*** be initialized first.
+    pub fn create_data<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) {
+        let nbytes = self.nbytes(generator, ctx);
+
+        let data = Int(Byte).array_alloca(generator, ctx, nbytes.value);
+        self.instance.set(ctx, |f| f.data, data);
+
+        self.set_strides_contiguous(generator, ctx);
+    }
+
+    /// Copy shape dimensions from an array.
+    pub fn copy_shape_from_array<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) {
+        let num_items = self.ndims_llvm(generator, ctx.ctx).value;
+        self.instance.get(generator, ctx, |f| f.shape).copy_from(generator, ctx, shape, num_items);
+    }
+
+    /// Copy shape dimensions from an ndarray.
+    /// Panics if `ndims` mismatches.
+    pub fn copy_shape_from_ndarray<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        src_ndarray: NDArrayObject<'ctx>,
+    ) {
+        assert_eq!(self.ndims, src_ndarray.ndims);
+        let src_shape = src_ndarray.instance.get(generator, ctx, |f| f.shape);
+        self.copy_shape_from_array(generator, ctx, src_shape);
+    }
+
+    /// Copy strides dimensions from an array.
+    pub fn copy_strides_from_array<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        strides: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) {
+        let num_items = self.ndims_llvm(generator, ctx.ctx).value;
+        self.instance
+            .get(generator, ctx, |f| f.strides)
+            .copy_from(generator, ctx, strides, num_items);
+    }
+
+    /// Copy strides dimensions from an ndarray.
+    /// Panics if `ndims` mismatches.
+    pub fn copy_strides_from_ndarray<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        src_ndarray: NDArrayObject<'ctx>,
+    ) {
+        assert_eq!(self.ndims, src_ndarray.ndims);
+        let src_strides = src_ndarray.instance.get(generator, ctx, |f| f.strides);
+        self.copy_strides_from_array(generator, ctx, src_strides);
+    }
+
+    /// Get the `np.size()` of this ndarray.
+    pub fn size<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<SizeT>> {
+        call_nac3_ndarray_size(generator, ctx, self.instance)
+    }
+
+    /// Get the `ndarray.nbytes` of this ndarray.
+    pub fn nbytes<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<SizeT>> {
+        call_nac3_ndarray_nbytes(generator, ctx, self.instance)
+    }
+
+    /// Get the `len()` of this ndarray.
+    pub fn len<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<SizeT>> {
+        call_nac3_ndarray_len(generator, ctx, self.instance)
+    }
+
+    /// Check if this ndarray is C-contiguous.
+    ///
+    /// See NumPy's `flags["C_CONTIGUOUS"]`: <https://numpy.org/doc/stable/reference/generated/numpy.ndarray.flags.html#numpy.ndarray.flags>
+    pub fn is_c_contiguous<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<Bool>> {
+        call_nac3_ndarray_is_c_contiguous(generator, ctx, self.instance)
+    }
+
+    /// Get the pointer to the n-th (0-based) element.
+    ///
+    /// The returned pointer has the element type of the LLVM type of this ndarray's `dtype`.
+    pub fn get_nth_pelement<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        nth: Instance<'ctx, Int<SizeT>>,
+    ) -> PointerValue<'ctx> {
+        let elem_ty = ctx.get_llvm_type(generator, self.dtype);
+
+        let p = call_nac3_ndarray_get_nth_pelement(generator, ctx, self.instance, nth);
+        ctx.builder
+            .build_pointer_cast(p.value, elem_ty.ptr_type(AddressSpace::default()), "")
+            .unwrap()
+    }
+
+    /// Get the n-th (0-based) scalar.
+    pub fn get_nth_scalar<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        nth: Instance<'ctx, Int<SizeT>>,
+    ) -> AnyObject<'ctx> {
+        let ptr = self.get_nth_pelement(generator, ctx, nth);
+        let value = ctx.builder.build_load(ptr, "").unwrap();
+        AnyObject { ty: self.dtype, value }
+    }
+
+    /// Get the pointer to the element indexed by `indices`.
+    ///
+    /// The returned pointer has the element type of the LLVM type of this ndarray's `dtype`.
+    pub fn get_pelement_by_indices<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        indices: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) -> PointerValue<'ctx> {
+        let elem_ty = ctx.get_llvm_type(generator, self.dtype);
+
+        let p = call_nac3_ndarray_get_pelement_by_indices(generator, ctx, self.instance, indices);
+        ctx.builder
+            .build_pointer_cast(p.value, elem_ty.ptr_type(AddressSpace::default()), "")
+            .unwrap()
+    }
+
+    /// Get the scalar indexed by `indices`.
+    pub fn get_scalar_by_indices<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        indices: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) -> AnyObject<'ctx> {
+        let ptr = self.get_pelement_by_indices(generator, ctx, indices);
+        let value = ctx.builder.build_load(ptr, "").unwrap();
+        AnyObject { ty: self.dtype, value }
+    }
+
+    /// Call [`call_nac3_ndarray_set_strides_by_shape`] on this ndarray to update `strides`.
+    ///
+    /// Update the ndarray's strides to make the ndarray contiguous.
+    pub fn set_strides_contiguous<G: CodeGenerator + ?Sized>(
+        self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) {
+        call_nac3_ndarray_set_strides_by_shape(generator, ctx, self.instance);
+    }
+
+    /// Clone/Copy this ndarray - Allocate a new ndarray with the same shape as this ndarray and copy the contents over.
+    ///
+    /// The new ndarray will own its data and will be C-contiguous.
+    #[must_use]
+    pub fn make_copy<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Self {
+        let clone = NDArrayObject::alloca(generator, ctx, self.dtype, self.ndims);
+
+        let shape = self.instance.gep(ctx, |f| f.shape).load(generator, ctx);
+        clone.copy_shape_from_array(generator, ctx, shape);
+        clone.create_data(generator, ctx);
+        clone.copy_data_from(generator, ctx, *self);
+        clone
+    }
+
+    /// Copy data from another ndarray.
+    ///
+    /// This ndarray and `src` is that their `np.size()` should be the same. Their shapes
+    /// do not matter. The copying order is determined by how their flattened views look.
+    ///
+    /// Panics if the `dtype`s of ndarrays are different.
+    pub fn copy_data_from<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        src: NDArrayObject<'ctx>,
+    ) {
+        assert!(ctx.unifier.unioned(self.dtype, src.dtype), "self and src dtype should match");
+        call_nac3_ndarray_copy_data(generator, ctx, src.instance, self.instance);
+    }
+
+    /// Returns true if this ndarray is unsized - `ndims == 0` and only contains a scalar.
+    #[must_use]
+    pub fn is_unsized(&self) -> bool {
+        self.ndims == 0
+    }
+
+    /// If this ndarray is unsized, return its sole value as an [`AnyObject`].
+    /// Otherwise, do nothing and return the ndarray itself.
+    pub fn split_unsized<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> ScalarOrNDArray<'ctx> {
+        if self.is_unsized() {
+            // NOTE: `np.size(self) == 0` here is never possible.
+            let zero = Int(SizeT).const_0(generator, ctx.ctx);
+            let value = self.get_nth_scalar(generator, ctx, zero).value;
+
+            ScalarOrNDArray::Scalar(AnyObject { ty: self.dtype, value })
+        } else {
+            ScalarOrNDArray::NDArray(*self)
+        }
+    }
+
+    /// Fill the ndarray with a scalar.
+    ///
+    /// `fill_value` must have the same LLVM type as the `dtype` of this ndarray.
+    pub fn fill<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        value: BasicValueEnum<'ctx>,
+    ) {
+        // TODO: It is possible to optimize this by exploiting contiguous strides with memset.
+        // Probably best to implement in IRRT.
+        self.foreach(generator, ctx, |generator, ctx, _hooks, nditer| {
+            let p = nditer.get_pointer(generator, ctx);
+            ctx.builder.build_store(p, value).unwrap();
+            Ok(())
+        })
+        .unwrap();
+    }
+
+    /// Create the shape tuple of this ndarray like `np.shape(<ndarray>)`.
+    ///
+    /// The returned integers in the tuple are in int32.
+    pub fn make_shape_tuple<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> TupleObject<'ctx> {
+        // TODO: Return a tuple of SizeT
+
+        let mut objects = Vec::with_capacity(self.ndims as usize);
+
+        for i in 0..self.ndims {
+            let dim = self
+                .instance
+                .get(generator, ctx, |f| f.shape)
+                .get_index_const(generator, ctx, i64::try_from(i).unwrap())
+                .truncate_or_bit_cast(generator, ctx, Int32);
+
+            objects.push(AnyObject {
+                ty: ctx.primitives.int32,
+                value: dim.value.as_basic_value_enum(),
+            });
+        }
+
+        TupleObject::from_objects(generator, ctx, objects)
+    }
+
+    /// Create the strides tuple of this ndarray like `<ndarray>.strides`.
+    ///
+    /// The returned integers in the tuple are in int32.
+    pub fn make_strides_tuple<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> TupleObject<'ctx> {
+        // TODO: Return a tuple of SizeT.
+
+        let mut objects = Vec::with_capacity(self.ndims as usize);
+
+        for i in 0..self.ndims {
+            let dim = self
+                .instance
+                .get(generator, ctx, |f| f.strides)
+                .get_index_const(generator, ctx, i64::try_from(i).unwrap())
+                .truncate_or_bit_cast(generator, ctx, Int32);
+
+            objects.push(AnyObject {
+                ty: ctx.primitives.int32,
+                value: dim.value.as_basic_value_enum(),
+            });
+        }
+
+        TupleObject::from_objects(generator, ctx, objects)
+    }
+
+    /// Create an unsized ndarray to contain `object`.
+    pub fn make_unsized<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        object: AnyObject<'ctx>,
+    ) -> NDArrayObject<'ctx> {
+        // We have to put the value on the stack to get a data pointer.
+        let data = ctx.builder.build_alloca(object.value.get_type(), "make_unsized").unwrap();
+        ctx.builder.build_store(data, object.value).unwrap();
+        let data = Ptr(Int(Byte)).pointer_cast(generator, ctx, data);
+
+        let ndarray = NDArrayObject::alloca(generator, ctx, object.ty, 0);
+        ndarray.instance.set(ctx, |f| f.data, data);
+        ndarray
+    }
+    /// Check if this `NDArray` can be used as an `out` ndarray for an operation.
+    ///
+    /// Raise an exception if the shapes do not match.
+    pub fn assert_can_be_written_by_out<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        out_ndims: u64,
+        out_shape: Instance<'ctx, Ptr<Int<SizeT>>>,
+    ) {
+        let ndarray_ndims = self.ndims_llvm(generator, ctx.ctx);
+        let ndarray_shape = self.instance.get(generator, ctx, |f| f.shape);
+
+        let output_ndims = Int(SizeT).const_int(generator, ctx.ctx, out_ndims, false);
+        let output_shape = out_shape;
+
+        call_nac3_ndarray_util_assert_output_shape_same(
+            generator,
+            ctx,
+            ndarray_ndims,
+            ndarray_shape,
+            output_ndims,
+            output_shape,
+        );
+    }
+}
+
+/// A convenience enum for implementing functions that acts on scalars or ndarrays or both.
+#[derive(Debug, Clone, Copy)]
+pub enum ScalarOrNDArray<'ctx> {
+    Scalar(AnyObject<'ctx>),
+    NDArray(NDArrayObject<'ctx>),
+}
+
+impl<'ctx> TryFrom<&ScalarOrNDArray<'ctx>> for AnyObject<'ctx> {
+    type Error = ();
+
+    fn try_from(value: &ScalarOrNDArray<'ctx>) -> Result<Self, Self::Error> {
+        match value {
+            ScalarOrNDArray::Scalar(scalar) => Ok(*scalar),
+            ScalarOrNDArray::NDArray(_ndarray) => Err(()),
+        }
+    }
+}
+
+impl<'ctx> TryFrom<&ScalarOrNDArray<'ctx>> for NDArrayObject<'ctx> {
+    type Error = ();
+
+    fn try_from(value: &ScalarOrNDArray<'ctx>) -> Result<Self, Self::Error> {
+        match value {
+            ScalarOrNDArray::Scalar(_scalar) => Err(()),
+            ScalarOrNDArray::NDArray(ndarray) => Ok(*ndarray),
+        }
+    }
+}
+
+impl<'ctx> ScalarOrNDArray<'ctx> {
+    /// Split on `object` either into a scalar or an ndarray.
+    ///
+    /// If `object` is an ndarray, [`ScalarOrNDArray::NDArray`].
+    ///
+    /// For everything else, it is wrapped with [`ScalarOrNDArray::Scalar`].
+    pub fn split_object<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        object: AnyObject<'ctx>,
+    ) -> ScalarOrNDArray<'ctx> {
+        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)
+            }
+            _ => ScalarOrNDArray::Scalar(object),
+        }
+    }
+
+    /// Get the underlying [`BasicValueEnum<'ctx>`] of this [`ScalarOrNDArray`].
+    #[must_use]
+    pub fn to_basic_value_enum(self) -> BasicValueEnum<'ctx> {
+        match self {
+            ScalarOrNDArray::Scalar(scalar) => scalar.value,
+            ScalarOrNDArray::NDArray(ndarray) => ndarray.instance.value.as_basic_value_enum(),
+        }
+    }
+
+    /// Convert this [`ScalarOrNDArray`] to an ndarray - behaves like `np.asarray`.
+    /// - If this is an ndarray, the ndarray is returned.
+    /// - If this is a scalar, this function returns new ndarray created with [`NDArrayObject::make_unsized`].
+    pub fn to_ndarray<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> NDArrayObject<'ctx> {
+        match self {
+            ScalarOrNDArray::NDArray(ndarray) => *ndarray,
+            ScalarOrNDArray::Scalar(scalar) => NDArrayObject::make_unsized(generator, ctx, *scalar),
+        }
+    }
+
+    /// Get the dtype of the ndarray created if this were called with [`ScalarOrNDArray::to_ndarray`].
+    #[must_use]
+    pub fn get_dtype(&self) -> Type {
+        match self {
+            ScalarOrNDArray::NDArray(ndarray) => ndarray.dtype,
+            ScalarOrNDArray::Scalar(scalar) => scalar.ty,
+        }
+    }
+}
+
+/// An helper enum specifying how a function should produce its output.
+///
+/// Many functions in NumPy has an optional `out` parameter (e.g., `matmul`). If `out` is specified
+/// with an ndarray, the result of a function will be written to `out`. If `out` is not specified, a function will
+/// create a new ndarray and store the result in it.
+#[derive(Debug, Clone, Copy)]
+pub enum NDArrayOut<'ctx> {
+    /// Tell a function should create a new ndarray with the expected element type `dtype`.
+    NewNDArray { dtype: Type },
+    /// Tell a function to write the result to `ndarray`.
+    WriteToNDArray { ndarray: NDArrayObject<'ctx> },
+}
+
+impl<'ctx> NDArrayOut<'ctx> {
+    /// Get the dtype of this output.
+    #[must_use]
+    pub fn get_dtype(&self) -> Type {
+        match self {
+            NDArrayOut::NewNDArray { dtype } => *dtype,
+            NDArrayOut::WriteToNDArray { ndarray } => ndarray.dtype,
+        }
+    }
+}

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

@@ -0,0 +1,179 @@
+use inkwell::{types::BasicType, values::PointerValue, AddressSpace};
+
+use crate::codegen::{
+    irrt::{call_nac3_nditer_has_element, call_nac3_nditer_initialize, call_nac3_nditer_next},
+    model::*,
+    object::any::AnyObject,
+    stmt::{gen_for_callback, BreakContinueHooks},
+    CodeGenContext, CodeGenerator,
+};
+
+use super::NDArrayObject;
+
+/// Fields of [`NDIter`]
+pub struct NDIterFields<'ctx, F: FieldTraversal<'ctx>> {
+    pub ndims: F::Output<Int<SizeT>>,
+    pub shape: F::Output<Ptr<Int<SizeT>>>,
+    pub strides: F::Output<Ptr<Int<SizeT>>>,
+
+    pub indices: F::Output<Ptr<Int<SizeT>>>,
+    pub nth: F::Output<Int<SizeT>>,
+    pub element: F::Output<Ptr<Int<Byte>>>,
+
+    pub size: F::Output<Int<SizeT>>,
+}
+
+/// An IRRT helper structure used to iterate through an ndarray.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct NDIter;
+
+impl<'ctx> StructKind<'ctx> for NDIter {
+    type Fields<F: FieldTraversal<'ctx>> = NDIterFields<'ctx, F>;
+
+    fn iter_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+        Self::Fields {
+            ndims: traversal.add_auto("ndims"),
+            shape: traversal.add_auto("shape"),
+            strides: traversal.add_auto("strides"),
+
+            indices: traversal.add_auto("indices"),
+            nth: traversal.add_auto("nth"),
+            element: traversal.add_auto("element"),
+
+            size: traversal.add_auto("size"),
+        }
+    }
+}
+
+/// A helper structure with a convenient interface to interact with [`NDIter`].
+#[derive(Debug, Clone)]
+pub struct NDIterHandle<'ctx> {
+    instance: Instance<'ctx, Ptr<Struct<NDIter>>>,
+    /// The ndarray this [`NDIter`] to iterating over.
+    ndarray: NDArrayObject<'ctx>,
+    /// The current indices of [`NDIter`].
+    indices: Instance<'ctx, Ptr<Int<SizeT>>>,
+}
+
+impl<'ctx> NDIterHandle<'ctx> {
+    /// Allocate an [`NDIter`] that iterates through an ndarray.
+    pub fn new<G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        ndarray: NDArrayObject<'ctx>,
+    ) -> Self {
+        let nditer = Struct(NDIter).alloca(generator, ctx);
+        let ndims = ndarray.ndims_llvm(generator, ctx.ctx);
+
+        // The caller has the responsibility to allocate 'indices' for `NDIter`.
+        let indices = Int(SizeT).array_alloca(generator, ctx, ndims.value);
+        call_nac3_nditer_initialize(generator, ctx, nditer, ndarray.instance, indices);
+
+        NDIterHandle { ndarray, instance: nditer, indices }
+    }
+
+    /// Is the current iteration valid?
+    ///
+    /// If true, then `element`, `indices` and `nth` contain details about the current element.
+    ///
+    /// If `ndarray` is unsized, this returns true only for the first iteration.
+    /// If `ndarray` is 0-sized, this always returns false.
+    #[must_use]
+    pub fn has_element<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<Bool>> {
+        call_nac3_nditer_has_element(generator, ctx, self.instance)
+    }
+
+    /// Go to the next element. If `has_element()` is false, then this has undefined behavior.
+    ///
+    /// If `ndarray` is unsized, this can only be called once.
+    /// If `ndarray` is 0-sized, this can never be called.
+    pub fn next<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) {
+        call_nac3_nditer_next(generator, ctx, self.instance);
+    }
+
+    /// Get pointer to the current element.
+    #[must_use]
+    pub fn get_pointer<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> PointerValue<'ctx> {
+        let elem_ty = ctx.get_llvm_type(generator, self.ndarray.dtype);
+
+        let p = self.instance.get(generator, ctx, |f| f.element);
+        ctx.builder
+            .build_pointer_cast(p.value, elem_ty.ptr_type(AddressSpace::default()), "element")
+            .unwrap()
+    }
+
+    /// Get the value of the current element.
+    #[must_use]
+    pub fn get_scalar<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> AnyObject<'ctx> {
+        let p = self.get_pointer(generator, ctx);
+        let value = ctx.builder.build_load(p, "value").unwrap();
+        AnyObject { ty: self.ndarray.dtype, value }
+    }
+
+    /// Get the index of the current element if this ndarray were a flat ndarray.
+    #[must_use]
+    pub fn get_index<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<SizeT>> {
+        self.instance.get(generator, ctx, |f| f.nth)
+    }
+
+    /// Get the indices of the current element.
+    #[must_use]
+    pub fn get_indices(&self) -> Instance<'ctx, Ptr<Int<SizeT>>> {
+        self.indices
+    }
+}
+
+impl<'ctx> NDArrayObject<'ctx> {
+    /// Iterate through every element in the ndarray.
+    ///
+    /// `body` has access to [`BreakContinueHooks`] to short-circuit and [`NDIterHandle`] to
+    /// get properties of the current iteration (e.g., the current element, indices, etc.)
+    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_element(generator, ctx).value),
+            |generator, ctx, hooks, nditer| body(generator, ctx, hooks, nditer),
+            |generator, ctx, nditer| {
+                nditer.next(generator, ctx);
+                Ok(())
+            },
+        )
+    }
+}

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

@@ -0,0 +1,105 @@
+use util::gen_for_model;
+
+use crate::{
+    codegen::{
+        model::*,
+        object::{any::AnyObject, list::ListObject, tuple::TupleObject},
+        CodeGenContext, CodeGenerator,
+    },
+    typecheck::typedef::TypeEnum,
+};
+
+/// Parse a NumPy-like "int sequence" input and return the int sequence as an array and its length.
+///
+/// * `sequence` - The `sequence` parameter.
+/// * `sequence_ty` - The typechecker type of `sequence`
+///
+/// The `sequence` argument type may only be one of the following:
+///   1. A list of `int32`;   e.g., `np.empty([600, 800, 3])`
+///   2. A tuple of `int32`;  e.g., `np.empty((600, 800, 3))`
+///   3. A scalar `int32`;    e.g., `np.empty(3)`, this is functionally equivalent to `np.empty([3])`
+///
+/// All `int32` values will be sign-extended to `SizeT`.
+pub fn parse_numpy_int_sequence<'ctx, G: CodeGenerator + ?Sized>(
+    generator: &mut G,
+    ctx: &mut CodeGenContext<'ctx, '_>,
+    input_sequence: AnyObject<'ctx>,
+) -> (Instance<'ctx, Int<SizeT>>, Instance<'ctx, Ptr<Int<SizeT>>>) {
+    let zero = Int(SizeT).const_0(generator, ctx.ctx);
+    let one = Int(SizeT).const_1(generator, ctx.ctx);
+
+    // The result `list` to return.
+    match &*ctx.unifier.get_ty(input_sequence.ty) {
+        TypeEnum::TObj { obj_id, .. }
+            if *obj_id == ctx.primitives.list.obj_id(&ctx.unifier).unwrap() =>
+        {
+            // 1. A list of `int32`; e.g., `np.empty([600, 800, 3])`
+
+            // Check `input_sequence`
+            let input_sequence = ListObject::from_object(generator, ctx, input_sequence);
+
+            let len = input_sequence.instance.get(generator, ctx, |f| f.len);
+            let result = Int(SizeT).array_alloca(generator, ctx, len.value);
+
+            // Load all the `int32`s from the input_sequence, cast them to `SizeT`, and store them into `result`
+            gen_for_model(generator, ctx, zero, len, one, |generator, ctx, _hooks, i| {
+                // Load the i-th int32 in the input sequence
+                let int = input_sequence
+                    .instance
+                    .get(generator, ctx, |f| f.items)
+                    .get_index(generator, ctx, i.value)
+                    .value
+                    .into_int_value();
+
+                // Cast to SizeT
+                let int = Int(SizeT).s_extend_or_bit_cast(generator, ctx, int);
+
+                // Store
+                result.set_index(ctx, i.value, int);
+
+                Ok(())
+            })
+            .unwrap();
+
+            (len, result)
+        }
+        TypeEnum::TTuple { .. } => {
+            // 2. A tuple of ints; e.g., `np.empty((600, 800, 3))`
+
+            let input_sequence = TupleObject::from_object(ctx, input_sequence);
+
+            let len = input_sequence.len(generator, ctx);
+
+            let result = Int(SizeT).array_alloca(generator, ctx, len.value);
+
+            for i in 0..input_sequence.num_elements() {
+                // Get the i-th element off of the tuple and load it into `result`.
+                let int = input_sequence.index(ctx, i).value.into_int_value();
+                let int = Int(SizeT).s_extend_or_bit_cast(generator, ctx, int);
+
+                result.set_index_const(ctx, i64::try_from(i).unwrap(), int);
+            }
+
+            (len, result)
+        }
+        TypeEnum::TObj { obj_id, .. }
+            if *obj_id == ctx.primitives.int32.obj_id(&ctx.unifier).unwrap() =>
+        {
+            // 3. A scalar int; e.g., `np.empty(3)`, this is functionally equivalent to `np.empty([3])`
+            let input_int = input_sequence.value.into_int_value();
+
+            let len = Int(SizeT).const_1(generator, ctx.ctx);
+            let result = Int(SizeT).array_alloca(generator, ctx, len.value);
+            let int = Int(SizeT).s_extend_or_bit_cast(generator, ctx, input_int);
+
+            // Storing into result[0]
+            result.store(ctx, int);
+
+            (len, result)
+        }
+        _ => panic!(
+            "encountered unknown sequence type: {}",
+            ctx.unifier.stringify(input_sequence.ty)
+        ),
+    }
+}

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

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

nac3core/src/codegen/object/tuple.rs

@@ -0,0 +1,99 @@
+use inkwell::values::StructValue;
+use itertools::Itertools;
+
+use crate::{
+    codegen::{model::*, CodeGenContext, CodeGenerator},
+    typecheck::typedef::{Type, TypeEnum},
+};
+
+use super::any::AnyObject;
+
+/// A NAC3 tuple object.
+///
+/// NOTE: This struct has no copy trait.
+#[derive(Debug, Clone)]
+pub struct TupleObject<'ctx> {
+    /// The type of the tuple.
+    pub tys: Vec<Type>,
+    /// The underlying LLVM struct value of this tuple.
+    pub value: StructValue<'ctx>,
+}
+
+impl<'ctx> TupleObject<'ctx> {
+    pub fn from_object(ctx: &mut CodeGenContext<'ctx, '_>, object: AnyObject<'ctx>) -> Self {
+        // TODO: Keep `is_vararg_ctx` from TTuple?
+
+        // Sanity check on object type.
+        let TypeEnum::TTuple { ty: tys, .. } = &*ctx.unifier.get_ty(object.ty) else {
+            panic!(
+                "Expected type to be a TypeEnum::TTuple, got {}",
+                ctx.unifier.stringify(object.ty)
+            );
+        };
+
+        // Check number of fields
+        let value = object.value.into_struct_value();
+        let value_num_fields = value.get_type().count_fields() as usize;
+        assert!(
+            value_num_fields == tys.len(),
+            "Tuple type has {} item(s), but the LLVM struct value has {} field(s)",
+            tys.len(),
+            value_num_fields
+        );
+
+        TupleObject { tys: tys.clone(), value }
+    }
+
+    /// Convenience function. Create a [`TupleObject`] from an iterator of objects.
+    pub fn from_objects<I, G: CodeGenerator + ?Sized>(
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+        objects: I,
+    ) -> Self
+    where
+        I: IntoIterator<Item = AnyObject<'ctx>>,
+    {
+        let (values, tys): (Vec<_>, Vec<_>) =
+            objects.into_iter().map(|object| (object.value, object.ty)).unzip();
+
+        let llvm_tys = tys.iter().map(|ty| ctx.get_llvm_type(generator, *ty)).collect_vec();
+        let llvm_tuple_ty = ctx.ctx.struct_type(&llvm_tys, false);
+
+        let pllvm_tuple = ctx.builder.build_alloca(llvm_tuple_ty, "tuple").unwrap();
+        for (i, val) in values.into_iter().enumerate() {
+            let pval = ctx.builder.build_struct_gep(pllvm_tuple, i as u32, "value").unwrap();
+            ctx.builder.build_store(pval, val).unwrap();
+        }
+
+        let value = ctx.builder.build_load(pllvm_tuple, "").unwrap().into_struct_value();
+        TupleObject { tys, value }
+    }
+
+    #[must_use]
+    pub fn num_elements(&self) -> usize {
+        self.tys.len()
+    }
+
+    /// Get the `len()` of this tuple.
+    #[must_use]
+    pub fn len<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &mut CodeGenContext<'ctx, '_>,
+    ) -> Instance<'ctx, Int<SizeT>> {
+        Int(SizeT).const_int(generator, ctx.ctx, self.num_elements() as u64, false)
+    }
+
+    /// Get the `i`-th (0-based) object in this tuple.
+    pub fn index(&self, ctx: &mut CodeGenContext<'ctx, '_>, i: usize) -> AnyObject<'ctx> {
+        assert!(
+            i < self.num_elements(),
+            "Tuple object with length {} have index {i}",
+            self.num_elements()
+        );
+
+        let value = ctx.builder.build_extract_value(self.value, i as u32, "tuple[{i}]").unwrap();
+        let ty = self.tys[i];
+        AnyObject { ty, value }
+    }
+}

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

@@ -0,0 +1 @@
+pub mod slice;

nac3core/src/codegen/object/utils/slice.rs

@@ -0,0 +1,125 @@
+use crate::codegen::{model::*, CodeGenContext, CodeGenerator};
+
+/// Fields of [`Slice`]
+#[derive(Debug, Clone)]
+pub struct SliceFields<'ctx, F: FieldTraversal<'ctx>, N: IntKind<'ctx>> {
+    pub start_defined: F::Output<Int<Bool>>,
+    pub start: F::Output<Int<N>>,
+    pub stop_defined: F::Output<Int<Bool>>,
+    pub stop: F::Output<Int<N>>,
+    pub step_defined: F::Output<Int<Bool>>,
+    pub step: F::Output<Int<N>>,
+}
+
+/// An IRRT representation of an (unresolved) slice.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct Slice<N>(pub N);
+
+impl<'ctx, N: IntKind<'ctx>> StructKind<'ctx> for Slice<N> {
+    type Fields<F: FieldTraversal<'ctx>> = SliceFields<'ctx, F, N>;
+
+    fn iter_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
+        Self::Fields {
+            start_defined: traversal.add_auto("start_defined"),
+            start: traversal.add("start", Int(self.0)),
+            stop_defined: traversal.add_auto("stop_defined"),
+            stop: traversal.add("stop", Int(self.0)),
+            step_defined: traversal.add_auto("step_defined"),
+            step: traversal.add("step", Int(self.0)),
+        }
+    }
+}
+
+/// A Rust structure that has [`Slice`] utilities and looks like a [`Slice`] but
+/// `start`, `stop` and `step` are held by LLVM registers only and possibly
+/// [`Option::None`] if unspecified.
+#[derive(Debug, Clone)]
+pub struct RustSlice<'ctx, N: IntKind<'ctx>> {
+    // It is possible that `start`, `stop`, and `step` are all `None`.
+    // We need to know the `int_kind` even when that is the case.
+    pub int_kind: N,
+    pub start: Option<Instance<'ctx, Int<N>>>,
+    pub stop: Option<Instance<'ctx, Int<N>>>,
+    pub step: Option<Instance<'ctx, Int<N>>>,
+}
+
+impl<'ctx, N: IntKind<'ctx>> RustSlice<'ctx, N> {
+    /// Write the contents to an LLVM [`Slice`].
+    pub fn write_to_slice<G: CodeGenerator + ?Sized>(
+        &self,
+        generator: &mut G,
+        ctx: &CodeGenContext<'ctx, '_>,
+        dst_slice_ptr: Instance<'ctx, Ptr<Struct<Slice<N>>>>,
+    ) {
+        let false_ = Int(Bool).const_false(generator, ctx.ctx);
+        let true_ = Int(Bool).const_true(generator, ctx.ctx);
+
+        match self.start {
+            Some(start) => {
+                dst_slice_ptr.gep(ctx, |f| f.start_defined).store(ctx, true_);
+                dst_slice_ptr.gep(ctx, |f| f.start).store(ctx, start);
+            }
+            None => dst_slice_ptr.gep(ctx, |f| f.start_defined).store(ctx, false_),
+        }
+
+        match self.stop {
+            Some(stop) => {
+                dst_slice_ptr.gep(ctx, |f| f.stop_defined).store(ctx, true_);
+                dst_slice_ptr.gep(ctx, |f| f.stop).store(ctx, stop);
+            }
+            None => dst_slice_ptr.gep(ctx, |f| f.stop_defined).store(ctx, false_),
+        }
+
+        match self.step {
+            Some(step) => {
+                dst_slice_ptr.gep(ctx, |f| f.step_defined).store(ctx, true_);
+                dst_slice_ptr.gep(ctx, |f| f.step).store(ctx, step);
+            }
+            None => dst_slice_ptr.gep(ctx, |f| f.step_defined).store(ctx, false_),
+        }
+    }
+}
+
+pub mod util {
+    use nac3parser::ast::Expr;
+
+    use crate::{
+        codegen::{model::*, CodeGenContext, CodeGenerator},
+        typecheck::typedef::Type,
+    };
+
+    use super::RustSlice;
+
+    /// Generate LLVM IR for an [`ExprKind::Slice`] and convert it into a [`RustSlice`].
+    #[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<RustSlice<'ctx, Int32>, String> {
+        let mut help = |value_expr: &Option<Box<Expr<Option<Type>>>>| -> Result<_, String> {
+            Ok(match value_expr {
+                None => None,
+                Some(value_expr) => {
+                    let value_expr = generator
+                        .gen_expr(ctx, value_expr)?
+                        .unwrap()
+                        .to_basic_value_enum(ctx, generator, ctx.primitives.int32)?;
+
+                    let value_expr =
+                        Int(Int32).check_value(generator, ctx.ctx, value_expr).unwrap();
+
+                    Some(value_expr)
+                }
+            })
+        };
+
+        let start = help(lower)?;
+        let stop = help(upper)?;
+        let step = help(step)?;
+
+        Ok(RustSlice { int_kind: Int32, start, stop, step })
+    }
+}

nac3core/src/codegen/stmt.rs

@@ -1,15 +1,19 @@
 use super::{
-    super::symbol_resolver::ValueEnum,
-    expr::destructure_range,
+    classes::{ArrayLikeIndexer, ArraySliceValue, ListValue, RangeValue},
+    expr::{destructure_range, gen_binop_expr},
+    gen_in_range_check,
     irrt::{handle_slice_indices, list_slice_assignment},
+    macros::codegen_unreachable,
+    object::{
+        any::AnyObject,
+        ndarray::{
+            indexing::util::gen_ndarray_subscript_ndindices, NDArrayObject, ScalarOrNDArray,
+        },
+    },
     CodeGenContext, CodeGenerator,
 };
 use crate::{
-    codegen::{
-        classes::{ArrayLikeIndexer, ArraySliceValue, ListValue, RangeValue},
-        expr::gen_binop_expr,
-        gen_in_range_check,
-    },
+    symbol_resolver::ValueEnum,
     toplevel::{DefinitionId, TopLevelDef},
     typecheck::{
         magic_methods::Binop,
@@ -121,7 +125,7 @@ pub fn gen_store_target<'ctx, G: CodeGenerator>(
                 return Ok(None);
             };
             let BasicValueEnum::PointerValue(ptr) = val else {
-                unreachable!();
+                codegen_unreachable!(ctx);
             };
             unsafe {
                 ctx.builder.build_in_bounds_gep(
@@ -135,7 +139,7 @@ pub fn gen_store_target<'ctx, G: CodeGenerator>(
             }
             .unwrap()
         }
-        _ => unreachable!(),
+        _ => codegen_unreachable!(ctx),
     }))
 }
 
@@ -176,6 +180,14 @@ pub fn gen_assign<'ctx, G: CodeGenerator>(
                 }
             }
             let val = value.to_basic_value_enum(ctx, generator, target.custom.unwrap())?;
+
+            // Perform i1 <-> i8 conversion as needed
+            let val = if ctx.unifier.unioned(target.custom.unwrap(), ctx.primitives.bool) {
+                generator.bool_to_i8(ctx, val.into_int_value()).into()
+            } else {
+                val
+            };
+
             ctx.builder.build_store(ptr, val).unwrap();
         }
     };
@@ -193,12 +205,12 @@ pub fn gen_assign_target_list<'ctx, G: CodeGenerator>(
     // Deconstruct the tuple `value`
     let BasicValueEnum::StructValue(tuple) = value.to_basic_value_enum(ctx, generator, value_ty)?
     else {
-        unreachable!()
+        codegen_unreachable!(ctx)
     };
 
     // NOTE: Currently, RHS's type is forced to be a Tuple by the type inferencer.
     let TypeEnum::TTuple { ty: tuple_tys, .. } = &*ctx.unifier.get_ty(value_ty) else {
-        unreachable!();
+        codegen_unreachable!(ctx);
     };
 
     assert_eq!(tuple.get_type().count_fields() as usize, tuple_tys.len());
@@ -258,7 +270,7 @@ pub fn gen_assign_target_list<'ctx, G: CodeGenerator>(
             // Now assign with that sub-tuple to the starred target.
             generator.gen_assign(ctx, target, ValueEnum::Dynamic(sub_tuple_val), sub_tuple_ty)?;
         } else {
-            unreachable!() // The typechecker ensures this
+            codegen_unreachable!(ctx) // The typechecker ensures this
         }
 
         // Handle assignment after the starred target
@@ -306,7 +318,9 @@ pub fn gen_setitem<'ctx, G: CodeGenerator>(
 
             if let ExprKind::Slice { .. } = &key.node {
                 // Handle assigning to a slice
-                let ExprKind::Slice { lower, upper, step } = &key.node else { unreachable!() };
+                let ExprKind::Slice { lower, upper, step } = &key.node else {
+                    codegen_unreachable!(ctx)
+                };
                 let Some((start, end, step)) = handle_slice_indices(
                     lower,
                     upper,
@@ -401,7 +415,47 @@ pub fn gen_setitem<'ctx, G: CodeGenerator>(
             if *obj_id == ctx.primitives.ndarray.obj_id(&ctx.unifier).unwrap() =>
         {
             // Handle NDArray item assignment
-            todo!("ndarray subscript assignment is not yet implemented");
+            // Process target
+            let target = generator
+                .gen_expr(ctx, target)?
+                .unwrap()
+                .to_basic_value_enum(ctx, generator, target_ty)?;
+            let target = AnyObject { value: target, ty: target_ty };
+
+            // Process key
+            let key = gen_ndarray_subscript_ndindices(generator, ctx, key)?;
+
+            // Process value
+            let value = value.to_basic_value_enum(ctx, generator, value_ty)?;
+            let value = AnyObject { value, ty: value_ty };
+
+            /*
+                Reference code:
+                ```python
+                target = target[key]
+                value = np.asarray(value)
+
+                shape = np.broadcast_shape((target, value))
+
+                target = np.broadcast_to(target, shape)
+                value = np.broadcast_to(value, shape)
+
+                ...and finally copy 1-1 from value to target.
+                ```
+            */
+
+            let target = NDArrayObject::from_object(generator, ctx, target);
+            let target = target.index(generator, ctx, &key);
+
+            let value =
+                ScalarOrNDArray::split_object(generator, ctx, value).to_ndarray(generator, ctx);
+
+            let broadcast_result = NDArrayObject::broadcast(generator, ctx, &[target, value]);
+
+            let target = broadcast_result.ndarrays[0];
+            let value = broadcast_result.ndarrays[1];
+
+            target.copy_data_from(generator, ctx, value);
         }
         _ => {
             panic!("encountered unknown target type: {}", ctx.unifier.stringify(target_ty));
@@ -416,7 +470,9 @@ pub fn gen_for<G: CodeGenerator>(
     ctx: &mut CodeGenContext<'_, '_>,
     stmt: &Stmt<Option<Type>>,
 ) -> Result<(), String> {
-    let StmtKind::For { iter, target, body, orelse, .. } = &stmt.node else { unreachable!() };
+    let StmtKind::For { iter, target, body, orelse, .. } = &stmt.node else {
+        codegen_unreachable!(ctx)
+    };
 
     // var_assignment static values may be changed in another branch
     // if so, remove the static value as it may not be correct in this branch
@@ -458,7 +514,7 @@ pub fn gen_for<G: CodeGenerator>(
             let Some(target_i) =
                 generator.gen_store_target(ctx, target, Some("for.target.addr"))?
             else {
-                unreachable!()
+                codegen_unreachable!(ctx)
             };
             let (start, stop, step) = destructure_range(ctx, iter_val);
 
@@ -901,7 +957,7 @@ pub fn gen_while<G: CodeGenerator>(
     ctx: &mut CodeGenContext<'_, '_>,
     stmt: &Stmt<Option<Type>>,
 ) -> Result<(), String> {
-    let StmtKind::While { test, body, orelse, .. } = &stmt.node else { unreachable!() };
+    let StmtKind::While { test, body, orelse, .. } = &stmt.node else { codegen_unreachable!(ctx) };
 
     // var_assignment static values may be changed in another branch
     // if so, remove the static value as it may not be correct in this branch
@@ -931,7 +987,7 @@ pub fn gen_while<G: CodeGenerator>(
 
         return Ok(());
     };
-    let BasicValueEnum::IntValue(test) = test else { unreachable!() };
+    let BasicValueEnum::IntValue(test) = test else { codegen_unreachable!(ctx) };
 
     ctx.builder
         .build_conditional_branch(generator.bool_to_i1(ctx, test), body_bb, orelse_bb)
@@ -1079,7 +1135,7 @@ pub fn gen_if<G: CodeGenerator>(
     ctx: &mut CodeGenContext<'_, '_>,
     stmt: &Stmt<Option<Type>>,
 ) -> Result<(), String> {
-    let StmtKind::If { test, body, orelse, .. } = &stmt.node else { unreachable!() };
+    let StmtKind::If { test, body, orelse, .. } = &stmt.node else { codegen_unreachable!(ctx) };
 
     // var_assignment static values may be changed in another branch
     // if so, remove the static value as it may not be correct in this branch
@@ -1202,11 +1258,11 @@ pub fn exn_constructor<'ctx>(
     let zelf_id = if let TypeEnum::TObj { obj_id, .. } = &*ctx.unifier.get_ty(zelf_ty) {
         obj_id.0
     } else {
-        unreachable!()
+        codegen_unreachable!(ctx)
     };
     let defs = ctx.top_level.definitions.read();
     let def = defs[zelf_id].read();
-    let TopLevelDef::Class { name: zelf_name, .. } = &*def else { unreachable!() };
+    let TopLevelDef::Class { name: zelf_name, .. } = &*def else { codegen_unreachable!(ctx) };
     let exception_name = format!("{}:{}", ctx.resolver.get_exception_id(zelf_id), zelf_name);
     unsafe {
         let id_ptr = ctx.builder.build_in_bounds_gep(zelf, &[zero, zero], "exn.id").unwrap();
@@ -1314,7 +1370,7 @@ pub fn gen_try<'ctx, 'a, G: CodeGenerator>(
     target: &Stmt<Option<Type>>,
 ) -> Result<(), String> {
     let StmtKind::Try { body, handlers, orelse, finalbody, .. } = &target.node else {
-        unreachable!()
+        codegen_unreachable!(ctx)
     };
 
     // if we need to generate anything related to exception, we must have personality defined
@@ -1391,7 +1447,7 @@ pub fn gen_try<'ctx, 'a, G: CodeGenerator>(
             if let TypeEnum::TObj { obj_id, .. } = &*ctx.unifier.get_ty(type_.custom.unwrap()) {
                 *obj_id
             } else {
-                unreachable!()
+                codegen_unreachable!(ctx)
             };
         let exception_name = format!("{}:{}", ctx.resolver.get_exception_id(obj_id.0), exn_name);
         let exn_id = ctx.resolver.get_string_id(&exception_name);
@@ -1663,6 +1719,23 @@ pub fn gen_return<G: CodeGenerator>(
     } else {
         None
     };
+
+    // Remap boolean return type into i1
+    let value = value.map(|ret_val| {
+        // The "return type" of a sret function is in the first parameter
+        let expected_ty = if ctx.need_sret {
+            func.get_type().get_param_types()[0]
+        } else {
+            func.get_type().get_return_type().unwrap()
+        };
+
+        if matches!(expected_ty, BasicTypeEnum::IntType(ty) if ty.get_bit_width() == 1) {
+            generator.bool_to_i1(ctx, ret_val.into_int_value()).into()
+        } else {
+            ret_val
+        }
+    });
+
     if let Some(return_target) = ctx.return_target {
         if let Some(value) = value {
             ctx.builder.build_store(ctx.return_buffer.unwrap(), value).unwrap();
@@ -1673,25 +1746,6 @@ pub fn gen_return<G: CodeGenerator>(
         ctx.builder.build_store(ctx.return_buffer.unwrap(), value.unwrap()).unwrap();
         ctx.builder.build_return(None).unwrap();
     } else {
-        // Remap boolean return type into i1
-        let value = value.map(|v| {
-            let expected_ty = func.get_type().get_return_type().unwrap();
-            let ret_val = v.as_basic_value_enum();
-
-            if expected_ty.is_int_type() && ret_val.is_int_value() {
-                let ret_type = expected_ty.into_int_type();
-                let ret_val = ret_val.into_int_value();
-
-                if ret_type.get_bit_width() == 1 && ret_val.get_type().get_bit_width() != 1 {
-                    generator.bool_to_i1(ctx, ret_val)
-                } else {
-                    ret_val
-                }
-                .into()
-            } else {
-                ret_val
-            }
-        });
         let value = value.as_ref().map(|v| v as &dyn BasicValue);
         ctx.builder.build_return(value).unwrap();
     }
@@ -1760,7 +1814,30 @@ pub fn gen_stmt<G: CodeGenerator>(
         StmtKind::Try { .. } => gen_try(generator, ctx, stmt)?,
         StmtKind::Raise { exc, .. } => {
             if let Some(exc) = exc {
-                let exc = if let Some(v) = generator.gen_expr(ctx, exc)? {
+                let exn = if let ExprKind::Name { id, .. } = &exc.node {
+                    // Handle "raise Exception" short form
+                    let def_id = ctx.resolver.get_identifier_def(*id).map_err(|e| {
+                        format!("{} (at {})", e.iter().next().unwrap(), exc.location)
+                    })?;
+                    let def = ctx.top_level.definitions.read();
+                    let TopLevelDef::Class { constructor, .. } = *def[def_id.0].read() else {
+                        return Err(format!("Failed to resolve symbol {id} (at {})", exc.location));
+                    };
+
+                    let TypeEnum::TFunc(signature) =
+                        ctx.unifier.get_ty(constructor.unwrap()).as_ref().clone()
+                    else {
+                        return Err(format!("Failed to resolve symbol {id} (at {})", exc.location));
+                    };
+
+                    generator
+                        .gen_call(ctx, None, (&signature, def_id), Vec::default())?
+                        .map(Into::into)
+                } else {
+                    generator.gen_expr(ctx, exc)?
+                };
+
+                let exc = if let Some(v) = exn {
                     v.to_basic_value_enum(ctx, generator, exc.custom.unwrap())?
                 } else {
                     return Ok(());

nac3core/src/toplevel/builtins.rs

@@ -1,6 +1,6 @@
 use std::iter::once;
 
-use helper::{debug_assert_prim_is_allowed, make_exception_fields, PrimDefDetails};
+use helper::{debug_assert_prim_is_allowed, extract_ndims, make_exception_fields, PrimDefDetails};
 use indexmap::IndexMap;
 use inkwell::{
     attributes::{Attribute, AttributeLoc},
@@ -9,13 +9,19 @@ use inkwell::{
     IntPredicate,
 };
 use itertools::Either;
+use numpy::unpack_ndarray_var_tys;
 use strum::IntoEnumIterator;
 
 use crate::{
     codegen::{
         builtin_fns,
         classes::{ProxyValue, RangeValue},
+        model::*,
         numpy::*,
+        object::{
+            any::AnyObject,
+            ndarray::{shape_util::parse_numpy_int_sequence, NDArrayObject},
+        },
         stmt::exn_constructor,
     },
     symbol_resolver::SymbolValue,
@@ -511,6 +517,14 @@ impl<'a> BuiltinBuilder<'a> {
             | PrimDef::FunNpEye
             | PrimDef::FunNpIdentity => self.build_ndarray_other_factory_function(prim),
 
+            PrimDef::FunNpSize | PrimDef::FunNpShape | PrimDef::FunNpStrides => {
+                self.build_ndarray_property_getter_function(prim)
+            }
+
+            PrimDef::FunNpBroadcastTo | PrimDef::FunNpTranspose | PrimDef::FunNpReshape => {
+                self.build_ndarray_view_function(prim)
+            }
+
             PrimDef::FunStr => self.build_str_function(),
 
             PrimDef::FunFloor | PrimDef::FunFloor64 | PrimDef::FunCeil | PrimDef::FunCeil64 => {
@@ -576,10 +590,6 @@ impl<'a> BuiltinBuilder<'a> {
             | PrimDef::FunNpHypot
             | PrimDef::FunNpNextAfter => self.build_np_2ary_function(prim),
 
-            PrimDef::FunNpTranspose | PrimDef::FunNpReshape => {
-                self.build_np_sp_ndarray_function(prim)
-            }
-
             PrimDef::FunNpDot
             | PrimDef::FunNpLinalgCholesky
             | PrimDef::FunNpLinalgQr
@@ -1385,6 +1395,171 @@ impl<'a> BuiltinBuilder<'a> {
         }
     }
 
+    fn build_ndarray_property_getter_function(&mut self, prim: PrimDef) -> TopLevelDef {
+        debug_assert_prim_is_allowed(
+            prim,
+            &[PrimDef::FunNpSize, PrimDef::FunNpShape, PrimDef::FunNpStrides],
+        );
+
+        let in_ndarray_ty = self.unifier.get_fresh_var_with_range(
+            &[self.primitives.ndarray],
+            Some("T".into()),
+            None,
+        );
+
+        match prim {
+            PrimDef::FunNpSize => create_fn_by_codegen(
+                self.unifier,
+                &into_var_map([in_ndarray_ty]),
+                prim.name(),
+                self.primitives.int32,
+                &[(in_ndarray_ty.ty, "a")],
+                Box::new(|ctx, obj, fun, args, generator| {
+                    assert!(obj.is_none());
+                    assert_eq!(args.len(), 1);
+
+                    let ndarray_ty = fun.0.args[0].ty;
+                    let ndarray =
+                        args[0].1.clone().to_basic_value_enum(ctx, generator, ndarray_ty)?;
+                    let ndarray = AnyObject { ty: ndarray_ty, value: ndarray };
+                    let ndarray = NDArrayObject::from_object(generator, ctx, ndarray);
+
+                    let size =
+                        ndarray.size(generator, ctx).truncate_or_bit_cast(generator, ctx, Int32);
+                    Ok(Some(size.value.as_basic_value_enum()))
+                }),
+            ),
+            PrimDef::FunNpShape | PrimDef::FunNpStrides => {
+                // The function signatures of `np_shape` an `np_size` are the same.
+                // Mixed together for convenience.
+
+                // The return type is a tuple of variable length depending on the ndims of the input ndarray.
+                let ret_ty = self.unifier.get_dummy_var().ty; // Handled by special folding
+
+                create_fn_by_codegen(
+                    self.unifier,
+                    &into_var_map([in_ndarray_ty]),
+                    prim.name(),
+                    ret_ty,
+                    &[(in_ndarray_ty.ty, "a")],
+                    Box::new(move |ctx, obj, fun, args, generator| {
+                        assert!(obj.is_none());
+                        assert_eq!(args.len(), 1);
+
+                        let ndarray_ty = fun.0.args[0].ty;
+                        let ndarray =
+                            args[0].1.clone().to_basic_value_enum(ctx, generator, ndarray_ty)?;
+
+                        let ndarray = AnyObject { ty: ndarray_ty, value: ndarray };
+                        let ndarray = NDArrayObject::from_object(generator, ctx, ndarray);
+
+                        let result_tuple = match prim {
+                            PrimDef::FunNpShape => ndarray.make_shape_tuple(generator, ctx),
+                            PrimDef::FunNpStrides => ndarray.make_strides_tuple(generator, ctx),
+                            _ => unreachable!(),
+                        };
+
+                        Ok(Some(result_tuple.value.as_basic_value_enum()))
+                    }),
+                )
+            }
+            _ => unreachable!(),
+        }
+    }
+
+    /// Build np/sp functions that take as input `NDArray` only
+    fn build_ndarray_view_function(&mut self, prim: PrimDef) -> TopLevelDef {
+        debug_assert_prim_is_allowed(
+            prim,
+            &[PrimDef::FunNpBroadcastTo, PrimDef::FunNpTranspose, PrimDef::FunNpReshape],
+        );
+
+        let in_ndarray_ty = self.unifier.get_fresh_var_with_range(
+            &[self.primitives.ndarray],
+            Some("T".into()),
+            None,
+        );
+
+        match prim {
+            PrimDef::FunNpTranspose => {
+                create_fn_by_codegen(
+                    self.unifier,
+                    &into_var_map([in_ndarray_ty]),
+                    prim.name(),
+                    in_ndarray_ty.ty,
+                    &[(in_ndarray_ty.ty, "x")],
+                    Box::new(move |ctx, _, fun, args, generator| {
+                        let arg_ty = fun.0.args[0].ty;
+                        let arg_val =
+                            args[0].1.clone().to_basic_value_enum(ctx, generator, arg_ty)?;
+
+                        let arg = AnyObject { ty: arg_ty, value: arg_val };
+                        let ndarray = NDArrayObject::from_object(generator, ctx, arg);
+
+                        let ndarray = ndarray.transpose(generator, ctx, None); // TODO: Add axes argument
+                        Ok(Some(ndarray.instance.value.as_basic_value_enum()))
+                    }),
+                )
+            }
+
+            // NOTE: on `ndarray_factory_fn_shape_arg_tvar` and
+            // the `param_ty` for `create_fn_by_codegen`.
+            //
+            // Similar to `build_ndarray_from_shape_factory_function` we delegate the responsibility of typechecking
+            // to [`typecheck::type_inferencer::Inferencer::fold_numpy_function_call_shape_argument`],
+            // and use a dummy [`TypeVar`] `ndarray_factory_fn_shape_arg_tvar` as a placeholder for `param_ty`.
+            PrimDef::FunNpBroadcastTo | PrimDef::FunNpReshape => {
+                // These two functions have the same function signature.
+                // Mixed together for convenience.
+
+                let ret_ty = self.unifier.get_dummy_var().ty; // Handled by special holding
+
+                create_fn_by_codegen(
+                    self.unifier,
+                    &VarMap::new(),
+                    prim.name(),
+                    ret_ty,
+                    &[
+                        (in_ndarray_ty.ty, "x"),
+                        (self.ndarray_factory_fn_shape_arg_tvar.ty, "shape"), // Handled by special folding
+                    ],
+                    Box::new(move |ctx, _, fun, args, generator| {
+                        let ndarray_ty = fun.0.args[0].ty;
+                        let ndarray_val =
+                            args[0].1.clone().to_basic_value_enum(ctx, generator, ndarray_ty)?;
+
+                        let shape_ty = fun.0.args[1].ty;
+                        let shape_val =
+                            args[1].1.clone().to_basic_value_enum(ctx, generator, shape_ty)?;
+
+                        let ndarray = AnyObject { value: ndarray_val, ty: ndarray_ty };
+                        let ndarray = NDArrayObject::from_object(generator, ctx, ndarray);
+
+                        let shape = AnyObject { value: shape_val, ty: shape_ty };
+                        let (_, shape) = parse_numpy_int_sequence(generator, ctx, shape);
+
+                        // The ndims after reshaping is gotten from the return type of the call.
+                        let (_, ndims) = unpack_ndarray_var_tys(&mut ctx.unifier, fun.0.ret);
+                        let ndims = extract_ndims(&ctx.unifier, ndims);
+
+                        let new_ndarray = match prim {
+                            PrimDef::FunNpBroadcastTo => {
+                                ndarray.broadcast_to(generator, ctx, ndims, shape)
+                            }
+                            PrimDef::FunNpReshape => {
+                                ndarray.reshape_or_copy(generator, ctx, ndims, shape)
+                            }
+                            _ => unreachable!(),
+                        };
+                        Ok(Some(new_ndarray.instance.value.as_basic_value_enum()))
+                    }),
+                )
+            }
+
+            _ => unreachable!(),
+        }
+    }
+
     /// Build the `str()` function.
     fn build_str_function(&mut self) -> TopLevelDef {
         let prim = PrimDef::FunStr;
@@ -1872,57 +2047,6 @@ impl<'a> BuiltinBuilder<'a> {
         }
     }
 
-    /// Build np/sp functions that take as input `NDArray` only
-    fn build_np_sp_ndarray_function(&mut self, prim: PrimDef) -> TopLevelDef {
-        debug_assert_prim_is_allowed(prim, &[PrimDef::FunNpTranspose, PrimDef::FunNpReshape]);
-
-        match prim {
-            PrimDef::FunNpTranspose => {
-                let ndarray_ty = self.unifier.get_fresh_var_with_range(
-                    &[self.ndarray_num_ty],
-                    Some("T".into()),
-                    None,
-                );
-                create_fn_by_codegen(
-                    self.unifier,
-                    &into_var_map([ndarray_ty]),
-                    prim.name(),
-                    ndarray_ty.ty,
-                    &[(ndarray_ty.ty, "x")],
-                    Box::new(move |ctx, _, fun, args, generator| {
-                        let arg_ty = fun.0.args[0].ty;
-                        let arg_val =
-                            args[0].1.clone().to_basic_value_enum(ctx, generator, arg_ty)?;
-                        Ok(Some(ndarray_transpose(generator, ctx, (arg_ty, arg_val))?))
-                    }),
-                )
-            }
-
-            // NOTE: on `ndarray_factory_fn_shape_arg_tvar` and
-            // the `param_ty` for `create_fn_by_codegen`.
-            //
-            // Similar to `build_ndarray_from_shape_factory_function` we delegate the responsibility of typechecking
-            // to [`typecheck::type_inferencer::Inferencer::fold_numpy_function_call_shape_argument`],
-            // and use a dummy [`TypeVar`] `ndarray_factory_fn_shape_arg_tvar` as a placeholder for `param_ty`.
-            PrimDef::FunNpReshape => create_fn_by_codegen(
-                self.unifier,
-                &VarMap::new(),
-                prim.name(),
-                self.ndarray_num_ty,
-                &[(self.ndarray_num_ty, "x"), (self.ndarray_factory_fn_shape_arg_tvar.ty, "shape")],
-                Box::new(move |ctx, _, fun, args, generator| {
-                    let x1_ty = fun.0.args[0].ty;
-                    let x1_val = args[0].1.clone().to_basic_value_enum(ctx, generator, x1_ty)?;
-                    let x2_ty = fun.0.args[1].ty;
-                    let x2_val = args[1].1.clone().to_basic_value_enum(ctx, generator, x2_ty)?;
-                    Ok(Some(ndarray_reshape(generator, ctx, (x1_ty, x1_val), (x2_ty, x2_val))?))
-                }),
-            ),
-
-            _ => unreachable!(),
-        }
-    }
-
     /// Build `np_linalg` and `sp_linalg` functions
     ///
     /// The input to these functions must be floating point `NDArray`

nac3core/src/toplevel/composer.rs

@@ -23,7 +23,7 @@ impl Default for ComposerConfig {
     }
 }
 
-type DefAst = (Arc<RwLock<TopLevelDef>>, Option<Stmt<()>>);
+pub type DefAst = (Arc<RwLock<TopLevelDef>>, Option<Stmt<()>>);
 pub struct TopLevelComposer {
     // list of top level definitions, same as top level context
     pub definition_ast_list: Vec<DefAst>,
@@ -1822,7 +1822,12 @@ impl TopLevelComposer {
                         if *name != init_str_id {
                             unreachable!("must be init function here")
                         }
-                        let all_inited = Self::get_all_assigned_field(body.as_slice())?;
+
+                        let all_inited = Self::get_all_assigned_field(
+                            object_id.0,
+                            definition_ast_list,
+                            body.as_slice(),
+                        )?;
                         for (f, _, _) in fields {
                             if !all_inited.contains(f) {
                                 return Err(HashSet::from([

nac3core/src/toplevel/helper.rs

@@ -3,6 +3,7 @@ use std::convert::TryInto;
 use crate::symbol_resolver::SymbolValue;
 use crate::toplevel::numpy::unpack_ndarray_var_tys;
 use crate::typecheck::typedef::{into_var_map, iter_type_vars, Mapping, TypeVarId, VarMap};
+use ast::ExprKind;
 use nac3parser::ast::{Constant, Location};
 use strum::IntoEnumIterator;
 use strum_macros::EnumIter;
@@ -52,6 +53,16 @@ pub enum PrimDef {
     FunNpEye,
     FunNpIdentity,
 
+    // NumPy ndarray property getters
+    FunNpSize,
+    FunNpShape,
+    FunNpStrides,
+
+    // NumPy ndarray view functions
+    FunNpBroadcastTo,
+    FunNpTranspose,
+    FunNpReshape,
+
     // Miscellaneous NumPy & SciPy functions
     FunNpRound,
     FunNpFloor,
@@ -99,8 +110,6 @@ pub enum PrimDef {
     FunNpLdExp,
     FunNpHypot,
     FunNpNextAfter,
-    FunNpTranspose,
-    FunNpReshape,
 
     // Linalg functions
     FunNpDot,
@@ -238,6 +247,16 @@ impl PrimDef {
             PrimDef::FunNpEye => fun("np_eye", None),
             PrimDef::FunNpIdentity => fun("np_identity", None),
 
+            // NumPy NDArray property getters,
+            PrimDef::FunNpSize => fun("np_size", None),
+            PrimDef::FunNpShape => fun("np_shape", None),
+            PrimDef::FunNpStrides => fun("np_strides", None),
+
+            // NumPy NDArray view functions
+            PrimDef::FunNpBroadcastTo => fun("np_broadcast_to", None),
+            PrimDef::FunNpTranspose => fun("np_transpose", None),
+            PrimDef::FunNpReshape => fun("np_reshape", None),
+
             // Miscellaneous NumPy & SciPy functions
             PrimDef::FunNpRound => fun("np_round", None),
             PrimDef::FunNpFloor => fun("np_floor", None),
@@ -285,8 +304,6 @@ 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),
@@ -733,7 +750,16 @@ impl TopLevelComposer {
         )
     }
 
-    pub fn get_all_assigned_field(stmts: &[Stmt<()>]) -> Result<HashSet<StrRef>, HashSet<String>> {
+    /// This function returns the fields that have been initialized in the `__init__` function of a class
+    /// The function takes as input:
+    ///  * `class_id`: The `object_id` of the class whose function is being evaluated (check `TopLevelDef::Class`)
+    ///  * `definition_ast_list`: A list of ast definitions and statements defined in `TopLevelComposer`
+    ///  * `stmts`: The body of function being parsed. Each statment is analyzed to check varaible initialization statements
+    pub fn get_all_assigned_field(
+        class_id: usize,
+        definition_ast_list: &Vec<DefAst>,
+        stmts: &[Stmt<()>],
+    ) -> Result<HashSet<StrRef>, HashSet<String>> {
         let mut result = HashSet::new();
         for s in stmts {
             match &s.node {
@@ -769,30 +795,138 @@ impl TopLevelComposer {
                 // TODO: do not check for For and While?
                 ast::StmtKind::For { body, orelse, .. }
                 | ast::StmtKind::While { body, orelse, .. } => {
-                    result.extend(Self::get_all_assigned_field(body.as_slice())?);
-                    result.extend(Self::get_all_assigned_field(orelse.as_slice())?);
+                    result.extend(Self::get_all_assigned_field(
+                        class_id,
+                        definition_ast_list,
+                        body.as_slice(),
+                    )?);
+                    result.extend(Self::get_all_assigned_field(
+                        class_id,
+                        definition_ast_list,
+                        orelse.as_slice(),
+                    )?);
                 }
                 ast::StmtKind::If { body, orelse, .. } => {
-                    let inited_for_sure = Self::get_all_assigned_field(body.as_slice())?
-                        .intersection(&Self::get_all_assigned_field(orelse.as_slice())?)
-                        .copied()
-                        .collect::<HashSet<_>>();
+                    let inited_for_sure = Self::get_all_assigned_field(
+                        class_id,
+                        definition_ast_list,
+                        body.as_slice(),
+                    )?
+                    .intersection(&Self::get_all_assigned_field(
+                        class_id,
+                        definition_ast_list,
+                        orelse.as_slice(),
+                    )?)
+                    .copied()
+                    .collect::<HashSet<_>>();
                     result.extend(inited_for_sure);
                 }
                 ast::StmtKind::Try { body, orelse, finalbody, .. } => {
-                    let inited_for_sure = Self::get_all_assigned_field(body.as_slice())?
-                        .intersection(&Self::get_all_assigned_field(orelse.as_slice())?)
-                        .copied()
-                        .collect::<HashSet<_>>();
+                    let inited_for_sure = Self::get_all_assigned_field(
+                        class_id,
+                        definition_ast_list,
+                        body.as_slice(),
+                    )?
+                    .intersection(&Self::get_all_assigned_field(
+                        class_id,
+                        definition_ast_list,
+                        orelse.as_slice(),
+                    )?)
+                    .copied()
+                    .collect::<HashSet<_>>();
                     result.extend(inited_for_sure);
-                    result.extend(Self::get_all_assigned_field(finalbody.as_slice())?);
+                    result.extend(Self::get_all_assigned_field(
+                        class_id,
+                        definition_ast_list,
+                        finalbody.as_slice(),
+                    )?);
                 }
                 ast::StmtKind::With { body, .. } => {
-                    result.extend(Self::get_all_assigned_field(body.as_slice())?);
+                    result.extend(Self::get_all_assigned_field(
+                        class_id,
+                        definition_ast_list,
+                        body.as_slice(),
+                    )?);
+                }
+                // Variables Initialized in function calls
+                ast::StmtKind::Expr { value, .. } => {
+                    let ExprKind::Call { func, .. } = &value.node else {
+                        continue;
+                    };
+                    let ExprKind::Attribute { value, attr, .. } = &func.node else {
+                        continue;
+                    };
+                    let ExprKind::Name { id, .. } = &value.node else {
+                        continue;
+                    };
+                    // Need to consider the two cases:
+                    //  Case 1) Call to class function i.e. id = `self`
+                    //  Case 2) Call to class ancestor function i.e. id = ancestor_name
+                    // We leave checking whether function in case 2 belonged to class ancestor or not to type checker
+                    //
+                    // According to current handling of `self`, function definition are fixed and do not change regardless
+                    // of which object is passed as `self` i.e. virtual polymorphism is not supported
+                    // Therefore, we change class id for case 2 to reflect behavior of our compiler
+
+                    let class_name = if *id == "self".into() {
+                        let ast::StmtKind::ClassDef { name, .. } =
+                            &definition_ast_list[class_id].1.as_ref().unwrap().node
+                        else {
+                            unreachable!()
+                        };
+                        name
+                    } else {
+                        id
+                    };
+
+                    let parent_method = definition_ast_list.iter().find_map(|def| {
+                        let (
+                            class_def,
+                            Some(ast::Located {
+                                node: ast::StmtKind::ClassDef { name, body, .. },
+                                ..
+                            }),
+                        ) = &def
+                        else {
+                            return None;
+                        };
+                        let TopLevelDef::Class { object_id: class_id, .. } = &*class_def.read()
+                        else {
+                            unreachable!()
+                        };
+
+                        if name == class_name {
+                            body.iter().find_map(|m| {
+                                let ast::StmtKind::FunctionDef { name, body, .. } = &m.node else {
+                                    return None;
+                                };
+                                if *name == *attr {
+                                    return Some((body.clone(), class_id.0));
+                                }
+                                None
+                            })
+                        } else {
+                            None
+                        }
+                    });
+
+                    // If method body is none then method does not exist
+                    if let Some((method_body, class_id)) = parent_method {
+                        result.extend(Self::get_all_assigned_field(
+                            class_id,
+                            definition_ast_list,
+                            method_body.as_slice(),
+                        )?);
+                    } else {
+                        return Err(HashSet::from([format!(
+                            "{}.{} not found in class {class_name} at {}",
+                            *id, *attr, value.location
+                        )]));
+                    }
                 }
                 ast::StmtKind::Pass { .. }
                 | ast::StmtKind::Assert { .. }
-                | ast::StmtKind::Expr { .. } => {}
+                | ast::StmtKind::AnnAssign { .. } => {}
 
                 _ => {
                     unimplemented!()
@@ -1000,3 +1134,23 @@ 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/typecheck/magic_methods.rs

@@ -520,6 +520,23 @@ pub fn typeof_binop(
         }
 
         Operator::MatMult => {
+            // NOTE: NumPy matmul's LHS and RHS must both be ndarrays. Scalars are not allowed.
+            match (&*unifier.get_ty(lhs), &*unifier.get_ty(rhs)) {
+                (
+                    TypeEnum::TObj { obj_id: lhs_obj_id, .. },
+                    TypeEnum::TObj { obj_id: rhs_obj_id, .. },
+                ) if *lhs_obj_id == primitives.ndarray.obj_id(unifier).unwrap()
+                    && *rhs_obj_id == primitives.ndarray.obj_id(unifier).unwrap() =>
+                {
+                    // LHS and RHS have valid types
+                }
+                _ => {
+                    let lhs_str = unifier.stringify(lhs);
+                    let rhs_str = unifier.stringify(rhs);
+                    return Err(format!("ndarray.__matmul__ only accepts ndarray operands, but left operand has type {lhs_str}, and right operand has type {rhs_str}"));
+                }
+            }
+
             let (_, lhs_ndims) = unpack_ndarray_var_tys(unifier, lhs);
             let lhs_ndims = match &*unifier.get_ty_immutable(lhs_ndims) {
                 TypeEnum::TLiteral { values, .. } => {

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::type_annotation::TypeAnnotation;
 use crate::{
     symbol_resolver::{SymbolResolver, SymbolValue},
     toplevel::{
@@ -102,6 +103,7 @@ pub struct Inferencer<'a> {
 }
 
 type InferenceError = HashSet<String>;
+type OverrideResult = Result<Option<ast::Expr<Option<Type>>>, InferenceError>;
 
 struct NaiveFolder();
 impl Fold<()> for NaiveFolder {
@@ -1181,6 +1183,45 @@ impl<'a> Inferencer<'a> {
             }));
         }
 
+        if ["np_shape".into(), "np_strides".into()].contains(id) && args.len() == 1 {
+            let ndarray = self.fold_expr(args.remove(0))?;
+
+            let ndims = arraylike_get_ndims(self.unifier, ndarray.custom.unwrap());
+
+            // Make a tuple of size `ndims` full of int32 (TODO: Make it usize)
+            let ret_ty = TypeEnum::TTuple {
+                ty: iter::repeat(self.primitives.int32).take(ndims as usize).collect_vec(),
+                is_vararg_ctx: false,
+            };
+            let ret_ty = self.unifier.add_ty(ret_ty);
+
+            let func_ty = TypeEnum::TFunc(FunSignature {
+                args: vec![FuncArg {
+                    name: "a".into(),
+                    default_value: None,
+                    ty: ndarray.custom.unwrap(),
+                    is_vararg: false,
+                }],
+                ret: ret_ty,
+                vars: VarMap::new(),
+            });
+            let func_ty = self.unifier.add_ty(func_ty);
+
+            return Ok(Some(Located {
+                location,
+                custom: Some(ret_ty),
+                node: ExprKind::Call {
+                    func: Box::new(Located {
+                        custom: Some(func_ty),
+                        location: func.location,
+                        node: ExprKind::Name { id: *id, ctx: *ctx },
+                    }),
+                    args: vec![ndarray],
+                    keywords: vec![],
+                },
+            }));
+        }
+
         if id == &"np_dot".into() {
             let arg0 = self.fold_expr(args.remove(0))?;
             let arg1 = self.fold_expr(args.remove(0))?;
@@ -1502,7 +1543,7 @@ impl<'a> Inferencer<'a> {
             }));
         }
         // 2-argument ndarray n-dimensional factory functions
-        if id == &"np_reshape".into() && args.len() == 2 {
+        if ["np_reshape".into(), "np_broadcast_to".into()].contains(id) && args.len() == 2 {
             let arg0 = self.fold_expr(args.remove(0))?;
 
             let shape_expr = args.remove(0);
@@ -1672,6 +1713,86 @@ impl<'a> Inferencer<'a> {
         Ok(None)
     }
 
+    /// Checks whether a class method is calling parent function
+    /// Returns [`None`] if its not a call to parent method, otherwise
+    /// returns a new `func` with class name replaced by `self` and method resolved to its `DefinitionID`
+    ///
+    /// e.g.  A.f1(self, ...) returns Some(self.{DefintionID(f1)})
+    fn check_overriding(&mut self, func: &ast::Expr<()>, args: &[ast::Expr<()>]) -> OverrideResult {
+        // `self` must be first argument for call to parent method
+        if let Some(Located { node: ExprKind::Name { id, .. }, .. }) = &args.first() {
+            if *id != "self".into() {
+                return Ok(None);
+            }
+        } else {
+            return Ok(None);
+        }
+
+        let Located {
+            node: ExprKind::Attribute { value, attr: method_name, ctx }, location, ..
+        } = func
+        else {
+            return Ok(None);
+        };
+        let ExprKind::Name { id: class_name, ctx: class_ctx } = &value.node else {
+            return Ok(None);
+        };
+        let zelf = &self.fold_expr(args[0].clone())?;
+
+        // Check whether the method belongs to class ancestors
+        let def_id = self.unifier.get_ty(zelf.custom.unwrap());
+        let TypeEnum::TObj { obj_id, .. } = def_id.as_ref() else { unreachable!() };
+        let defs = self.top_level.definitions.read();
+        let res = {
+            if let TopLevelDef::Class { ancestors, .. } = &*defs[obj_id.0].read() {
+                let res = ancestors.iter().find_map(|f| {
+                    let TypeAnnotation::CustomClass { id, .. } = f else { unreachable!() };
+                    let TopLevelDef::Class { name, methods, .. } = &*defs[id.0].read() else {
+                        unreachable!()
+                    };
+                    // Class names are stored as `__module__.class`
+                    let name = name.to_string();
+                    let (_, name) = name.rsplit_once('.').unwrap();
+                    if name == class_name.to_string() {
+                        return methods.iter().find_map(|f| {
+                            if f.0 == *method_name {
+                                return Some(*f);
+                            }
+                            None
+                        });
+                    }
+                    None
+                });
+                res
+            } else {
+                None
+            }
+        };
+
+        match res {
+            Some(r) => {
+                let mut new_func = func.clone();
+                let mut new_value = value.clone();
+                new_value.node = ExprKind::Name { id: "self".into(), ctx: *class_ctx };
+                new_func.node =
+                    ExprKind::Attribute { value: new_value.clone(), attr: *method_name, ctx: *ctx };
+
+                let mut new_func = self.fold_expr(new_func)?;
+
+                let ExprKind::Attribute { value, .. } = new_func.node else { unreachable!() };
+                new_func.node =
+                    ExprKind::Attribute { value, attr: r.2 .0.to_string().into(), ctx: *ctx };
+                new_func.custom = Some(r.1);
+
+                Ok(Some(new_func))
+            }
+            None => report_error(
+                format!("Ancestor method [{class_name}.{method_name}] should be defined with same decorator as its overridden version").as_str(),
+                *location,
+            ),
+        }
+    }
+
     fn fold_call(
         &mut self,
         location: Location,
@@ -1685,8 +1806,20 @@ impl<'a> Inferencer<'a> {
             return Ok(spec_call_func);
         }
 
-        let func = Box::new(self.fold_expr(func)?);
-        let args = args.into_iter().map(|v| self.fold_expr(v)).collect::<Result<Vec<_>, _>>()?;
+        // Check for call to parent method
+        let override_res = self.check_overriding(&func, &args)?;
+        let is_override = override_res.is_some();
+        let func = if is_override { override_res.unwrap() } else { self.fold_expr(func)? };
+        let func = Box::new(func);
+
+        let mut args =
+            args.into_iter().map(|v| self.fold_expr(v)).collect::<Result<Vec<_>, _>>()?;
+
+        // TODO: Handle passing of self to functions to allow runtime lookup of functions to be called
+        // Currently removing `self` and using compile time function definitions
+        if is_override {
+            args.remove(0);
+        }
         let keywords = keywords
             .into_iter()
             .map(|v| fold::fold_keyword(self, v))

nac3standalone/demo/demo.c

@@ -7,11 +7,11 @@
 #include <string.h>
 
 double dbl_nan(void) {
-  return NAN;
+    return NAN;
 }
 
 double dbl_inf(void) {
-  return INFINITY;
+    return INFINITY;
 }
 
 void output_bool(bool x) {
@@ -19,19 +19,19 @@ void output_bool(bool x) {
 }
 
 void output_int32(int32_t x) {
-    printf("%"PRId32"\n", x);
+    printf("%" PRId32 "\n", x);
 }
 
 void output_int64(int64_t x) {
-    printf("%"PRId64"\n", x);
+    printf("%" PRId64 "\n", x);
 }
 
 void output_uint32(uint32_t x) {
-    printf("%"PRIu32"\n", x);
+    printf("%" PRIu32 "\n", x);
 }
 
 void output_uint64(uint64_t x) {
-    printf("%"PRIu64"\n", x);
+    printf("%" PRIu64 "\n", x);
 }
 
 void output_float64(double x) {
@@ -52,7 +52,7 @@ void output_range(int32_t range[3]) {
 }
 
 void output_asciiart(int32_t x) {
-    static const char *chars = " .,-:;i+hHM$*#@    ";
+    static const char* chars = " .,-:;i+hHM$*#@    ";
     if (x < 0) {
         putchar('\n');
     } else {
@@ -61,12 +61,12 @@ void output_asciiart(int32_t x) {
 }
 
 struct cslice {
-    void *data;
+    void* data;
     size_t len;
 };
 
-void output_int32_list(struct cslice *slice) {
-    const int32_t *data = (int32_t *) slice->data;
+void output_int32_list(struct cslice* slice) {
+    const int32_t* data = (int32_t*)slice->data;
 
     putchar('[');
     for (size_t i = 0; i < slice->len; ++i) {
@@ -80,23 +80,23 @@ void output_int32_list(struct cslice *slice) {
     putchar('\n');
 }
 
-void output_str(struct cslice *slice) {
-    const char *data = (const char *) slice->data;
+void output_str(struct cslice* slice) {
+    const char* data = (const char*)slice->data;
 
     for (size_t i = 0; i < slice->len; ++i) {
         putchar(data[i]);
     }
 }
 
-void output_strln(struct cslice *slice) {
+void output_strln(struct cslice* slice) {
     output_str(slice);
     putchar('\n');
 }
 
-uint64_t dbg_stack_address(__attribute__((unused)) struct cslice *slice) {
+uint64_t dbg_stack_address(__attribute__((unused)) struct cslice* slice) {
     int i;
-    void *ptr = (void *) &i;
-    return (uintptr_t) ptr;
+    void* ptr = (void*)&i;
+    return (uintptr_t)ptr;
 }
 
 uint32_t __nac3_personality(uint32_t state, uint32_t exception_object, uint32_t context) {
@@ -119,11 +119,12 @@ struct Exception {
 
 uint32_t __nac3_raise(struct Exception* e) {
     printf("__nac3_raise called. Exception details:\n");
-    printf("          ID: %"PRIu32"\n", e->id);
-    printf("    Location: %*s:%"PRIu32":%"PRIu32"\n" , (int) e->file.len, (const char*) e->file.data, e->line, e->column);
-    printf("    Function: %*s\n" , (int) e->function.len, (const char*) e->function.data);
-    printf("     Message: \"%*s\"\n" , (int) e->message.len, (const char*) e->message.data);
-    printf("      Params: {0}=%"PRId64", {1}=%"PRId64", {2}=%"PRId64"\n", e->param[0], e->param[1], e->param[2]);
+    printf("          ID: %" PRIu32 "\n", e->id);
+    printf("    Location: %*s:%" PRIu32 ":%" PRIu32 "\n", (int)e->file.len, (const char*)e->file.data, e->line,
+           e->column);
+    printf("    Function: %*s\n", (int)e->function.len, (const char*)e->function.data);
+    printf("     Message: \"%*s\"\n", (int)e->message.len, (const char*)e->message.data);
+    printf("      Params: {0}=%" PRId64 ", {1}=%" PRId64 ", {2}=%" PRId64 "\n", e->param[0], e->param[1], e->param[2]);
     exit(101);
     __builtin_unreachable();
 }

nac3standalone/demo/interpret_demo.py

@@ -179,6 +179,16 @@ def patch(module):
     module.np_identity = np.identity
     module.np_array = np.array
 
+    # NumPy NDArray view functions
+    module.np_broadcast_to = np.broadcast_to
+    module.np_transpose = np.transpose
+    module.np_reshape = np.reshape
+
+    # NumPy NDArray property getters
+    module.np_size = np.size
+    module.np_shape = np.shape
+    module.np_strides = lambda ndarray: ndarray.strides
+
     # NumPy Math functions
     module.np_isnan = np.isnan
     module.np_isinf = np.isinf
@@ -218,8 +228,6 @@ def patch(module):
     module.np_ldexp = np.ldexp
     module.np_hypot = np.hypot
     module.np_nextafter = np.nextafter
-    module.np_transpose = np.transpose
-    module.np_reshape = np.reshape
 
     # SciPy Math functions
     module.sp_spec_erf = special.erf

nac3standalone/demo/src/assignment.py

@@ -9,6 +9,7 @@ def output_bool(x: bool):
 def example1():
     x, *ys, z = (1, 2, 3, 4, 5)
     output_int32(x)
+    output_int32(len(ys))
     output_int32(ys[0])
     output_int32(ys[1])
     output_int32(ys[2])
@@ -18,12 +19,14 @@ def example2():
     x, y, *zs = (1, 2, 3, 4, 5)
     output_int32(x)
     output_int32(y)
+    output_int32(len(zs))
     output_int32(zs[0])
     output_int32(zs[1])
     output_int32(zs[2])
 
 def example3():
     *xs, y, z = (1, 2, 3, 4, 5)
+    output_int32(len(xs))
     output_int32(xs[0])
     output_int32(xs[1])
     output_int32(xs[2])
@@ -31,6 +34,12 @@ def example3():
     output_int32(z)
 
 def example4():
+    *xs, y, z = (4, 5)
+    output_int32(len(xs))
+    output_int32(y)
+    output_int32(z)
+
+def example5():
     # Example from: https://docs.python.org/3/reference/simple_stmts.html#assignment-statements
     x = [0, 1]
     i = 0
@@ -44,7 +53,7 @@ class A:
     def __init__(self):
         self.value = 1000
 
-def example5():
+def example6():
     ws = [88, 7, 8]
     a = A()
     x, [y, *ys, a.value], ws[0], (ws[0],) = 1, (2, False, 4, 5), 99, (6,)
@@ -63,4 +72,5 @@ def run() -> int32:
     example3()
     example4()
     example5()
+    example6()
     return 0

nac3standalone/demo/src/inheritance.py

@@ -10,23 +10,58 @@ class A:
     def __init__(self, a: int32):
         self.a = a
     
-    def f1(self):
-        self.f2()
-    
-    def f2(self):
+    def output_all_fields(self):
         output_int32(self.a)
+    
+    def set_a(self, a: int32):
+        self.a = a
 
 class B(A):
     b: int32
 
     def __init__(self, b: int32):
-        self.a = b + 1
+        A.__init__(self, b + 1)
+        self.set_b(b)
+    
+    def output_parent_fields(self):
+        A.output_all_fields(self)
+
+    def output_all_fields(self):
+        A.output_all_fields(self)
+        output_int32(self.b)
+    
+    def set_b(self, b: int32):
         self.b = b
 
+class C(B):
+    c: int32
+
+    def __init__(self, c: int32):
+        B.__init__(self, c + 1)
+        self.c = c
+    
+    def output_parent_fields(self):
+        B.output_all_fields(self)
+
+    def output_all_fields(self):
+        B.output_all_fields(self)
+        output_int32(self.c)
+    
+    def set_c(self, c: int32):
+        self.c = c
 
 def run() -> int32:
-    aaa = A(5)
-    bbb = B(2)
-    aaa.f1()
-    bbb.f1()
+    ccc = C(10)
+    ccc.output_all_fields()
+    ccc.set_a(1)
+    ccc.set_b(2)
+    ccc.set_c(3)
+    ccc.output_all_fields()
+
+    bbb = B(10)
+    bbb.set_a(9)
+    bbb.set_b(8)
+    bbb.output_all_fields()
+    ccc.output_all_fields()
+
     return 0

nac3standalone/src/basic_symbol_resolver.rs

@@ -15,7 +15,6 @@ use std::{collections::HashMap, sync::Arc};
 pub struct ResolverInternal {
     pub id_to_type: Mutex<HashMap<StrRef, Type>>,
     pub id_to_def: Mutex<HashMap<StrRef, DefinitionId>>,
-    pub class_names: Mutex<HashMap<StrRef, Type>>,
     pub module_globals: Mutex<HashMap<StrRef, SymbolValue>>,
     pub str_store: Mutex<HashMap<String, i32>>,
 }

nac3standalone/src/main.rs

@@ -306,7 +306,6 @@ fn main() {
     let internal_resolver: Arc<ResolverInternal> = ResolverInternal {
         id_to_type: builtins_ty.into(),
         id_to_def: builtins_def.into(),
-        class_names: Mutex::default(),
         module_globals: Mutex::default(),
         str_store: Mutex::default(),
     }
@@ -314,6 +313,15 @@ 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, 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 +426,8 @@ 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 +447,9 @@ 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 +458,7 @@ 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 +472,7 @@ 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");