lyken
/
nac3
forked from M-Labs/nac3
1
0
Fork 0
Code Pull Requests Activity

WIP: core/ndstrides: checkpoint 9

This commit is contained in:
lyken 2024-08-13 17:00:48 +08:00
parent 82edcd9390
commit 1d7184708f
No known key found for this signature in database
GPG Key ID: 3BD5FC6AC8325DD8
65 changed files with 7832 additions and 2719 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

25
nac3artiq/src/codegen.rs
View File

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

11
nac3artiq/src/lib.rs
View File

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

44
nac3core/build.rs
View File

@ -29,28 +29,28 @@ fn compile_irrt_cpp() {
* Compiling for WASM32 and filtering the output with regex is the closest we can get.
*/
let irrt_cpp_path = irrt_dir.join("irrt.cpp");
let flags: &[&str] = &[
"--target=wasm32",
"-x",
"c++",
"-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(),
];
let mut flags = vec![];
flags.push("--target=wasm32");
flags.extend(&["-x", "c++"]);
flags.extend(&["-fno-discard-value-names", "-fno-exceptions", "-fno-rtti"]);
flags.push("-emit-llvm");
flags.push("-S");
flags.extend(&["-Wall", "-Wextra"]);
flags.extend(&["-o", "-"]);
flags.extend(&["-I", irrt_dir.to_str().unwrap()]);
flags.push(irrt_cpp_path.to_str().unwrap());
match env::var("PROFILE").as_deref() {
Ok("debug") => {
flags.push("-O0");
flags.push("-DIRRT_DEBUG");
}
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());

26
nac3core/irrt/irrt/core.hpp
View File

@ -4,7 +4,7 @@
#include <irrt/util.hpp>
// NDArray indices are always `uint32_t`.
using NDIndex = uint32_t;
using NDIndexInt = uint32_t;
// The type of an index or a value describing the length of a
// range/slice is always `int32_t`.
using SliceIndex = int32_t;
@ -43,7 +43,7 @@ SizeT __nac3_ndarray_calc_size_impl(const SizeT* list_data, SizeT list_len,
template <typename SizeT>
void __nac3_ndarray_calc_nd_indices_impl(SizeT index, const SizeT* dims,
SizeT num_dims, NDIndex* idxs) {
SizeT num_dims, NDIndexInt* idxs) {
SizeT stride = 1;
for (SizeT dim = 0; dim < num_dims; dim++) {
SizeT i = num_dims - dim - 1;
@ -55,7 +55,7 @@ void __nac3_ndarray_calc_nd_indices_impl(SizeT index, const SizeT* dims,
template <typename SizeT>
SizeT __nac3_ndarray_flatten_index_impl(const SizeT* dims, SizeT num_dims,
const NDIndex* indices,
const NDIndexInt* indices,
SizeT num_indices) {
SizeT idx = 0;
SizeT stride = 1;
@ -104,8 +104,8 @@ void __nac3_ndarray_calc_broadcast_impl(const SizeT* lhs_dims, SizeT lhs_ndims,
template <typename SizeT>
void __nac3_ndarray_calc_broadcast_idx_impl(const SizeT* src_dims,
SizeT src_ndims,
const NDIndex* in_idx,
NDIndex* out_idx) {
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];
@ -293,24 +293,24 @@ uint64_t __nac3_ndarray_calc_size64(const uint64_t* list_data,
}
void __nac3_ndarray_calc_nd_indices(uint32_t index, const uint32_t* dims,
uint32_t num_dims, NDIndex* idxs) {
uint32_t num_dims, NDIndexInt* idxs) {
__nac3_ndarray_calc_nd_indices_impl(index, dims, num_dims, idxs);
}
void __nac3_ndarray_calc_nd_indices64(uint64_t index, const uint64_t* dims,
uint64_t num_dims, NDIndex* idxs) {
uint64_t num_dims, NDIndexInt* idxs) {
__nac3_ndarray_calc_nd_indices_impl(index, dims, num_dims, idxs);
}
uint32_t __nac3_ndarray_flatten_index(const uint32_t* dims, uint32_t num_dims,
const NDIndex* indices,
const NDIndexInt* indices,
uint32_t num_indices) {
return __nac3_ndarray_flatten_index_impl(dims, num_dims, indices,
num_indices);
}
uint64_t __nac3_ndarray_flatten_index64(const uint64_t* dims, uint64_t num_dims,
const NDIndex* indices,
const NDIndexInt* indices,
uint64_t num_indices) {
return __nac3_ndarray_flatten_index_impl(dims, num_dims, indices,
num_indices);
@ -333,16 +333,16 @@ void __nac3_ndarray_calc_broadcast64(const uint64_t* lhs_dims,
void __nac3_ndarray_calc_broadcast_idx(const uint32_t* src_dims,
uint32_t src_ndims,
const NDIndex* in_idx,
NDIndex* out_idx) {
const NDIndexInt* in_idx,
NDIndexInt* out_idx) {
__nac3_ndarray_calc_broadcast_idx_impl(src_dims, src_ndims, in_idx,
out_idx);
}
void __nac3_ndarray_calc_broadcast_idx64(const uint64_t* src_dims,
uint64_t src_ndims,
const NDIndex* in_idx,
NDIndex* out_idx) {
const NDIndexInt* in_idx,
NDIndexInt* out_idx) {
__nac3_ndarray_calc_broadcast_idx_impl(src_dims, src_ndims, in_idx,
out_idx);
}

9
nac3core/irrt/irrt/cslice.hpp Normal file
View File

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

15
nac3core/irrt/irrt/debug.hpp Normal file
View File

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

123
nac3core/irrt/irrt/exception.hpp Normal file
View File

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

17
nac3core/irrt/irrt/list.hpp Normal file
View File

@ -0,0 +1,17 @@
#pragma once
#include <irrt/int_defs.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

119
nac3core/irrt/irrt/ndarray/array.hpp Normal file
View File

@ -0,0 +1,119 @@
#pragma once
#include <irrt/int_defs.hpp>
#include <irrt/list.hpp>
#include <irrt/ndarray/basic.hpp>
#include <irrt/ndarray/def.hpp>
namespace {
namespace ndarray {
namespace array {
// TODO: Document me
template <typename SizeT>
void set_and_validate_list_shape_helper(SizeT axis, List<SizeT>* list,
SizeT ndims, SizeT* shape) {
if (shape[axis] == -1) {
// Dimension is unspecified. Set it.
shape[axis] = list->len;
} else {
// Dimension is specified. Check.
if (shape[axis] != list->len) {
// Mismatch, throw an error.
// NOTE: NumPy's error message is more complex and needs more PARAMS to display.
raise_exception(SizeT, EXN_VALUE_ERROR,
"The requested array has an inhomogenous shape "
"after {0} dimension(s).",
axis, shape[axis], list->len);
}
}
if (axis + 1 == ndims) {
// `list` has type `list[ItemType]`
// Do nothing
} else {
// `list` has type `list[list[...]]`
List<SizeT>** lists = (List<SizeT>**)(list->items);
for (SizeT i = 0; i < list->len; i++) {
set_and_validate_list_shape_helper<SizeT>(axis + 1, lists[i], ndims,
shape);
}
}
}
// TODO: Document me
template <typename SizeT>
void set_and_validate_list_shape(List<SizeT>* list, SizeT ndims, SizeT* shape) {
for (SizeT axis = 0; axis < ndims; axis++) {
shape[axis] = -1; // Sentinel to say this dimension is unspecified.
}
set_and_validate_list_shape_helper<SizeT>(0, list, ndims, shape);
}
// TODO: Document me
template <typename SizeT>
void write_list_to_array_helper(SizeT axis, SizeT* index, List<SizeT>* list,
NDArray<SizeT>* ndarray) {
debug_assert_eq(SizeT, list->len, ndarray->shape[axis]);
if (IRRT_DEBUG_ASSERT_BOOL) {
if (!ndarray::basic::is_c_contiguous(ndarray)) {
raise_debug_assert(SizeT, "ndarray is not C-contiguous", ndarray->strides[0],
ndarray->strides[1], NO_PARAM);
}
}
if (axis + 1 == ndarray->ndims) {
// `list` has type `list[ItemType]`
// `ndarray` is contiguous, so we can do this, and this is fast.
uint8_t* dst = ndarray->data + (ndarray->itemsize * (*index));
__builtin_memcpy(dst, list->items, ndarray->itemsize * list->len);
*index += list->len;
} else {
// `list` has type `list[list[...]]`
List<SizeT>** lists = (List<SizeT>**)(list->items);
for (SizeT i = 0; i < list->len; i++) {
write_list_to_array_helper<SizeT>(axis + 1, index, lists[i],
ndarray);
}
}
}
// TODO: Document me
template <typename SizeT>
void write_list_to_array(List<SizeT>* list, NDArray<SizeT>* ndarray) {
// done after set_and_validate(list, ndims, shape), list is well-formed
// ndarray->data is allocated and owned
// ndarray->itemsize is set
// ndarray->ndims is set
// ndarray->shape is set
// ndarray->strides is ???
SizeT index = 0;
write_list_to_array_helper<SizeT>((SizeT)0, &index, list, ndarray);
}
} // namespace array
} // namespace ndarray
} // namespace
extern "C" {
using namespace ndarray::array;
void __nac3_array_set_and_validate_list_shape(List<int32_t>* list,
int32_t ndims, int32_t* shape) {
set_and_validate_list_shape(list, ndims, shape);
}
void __nac3_array_set_and_validate_list_shape64(List<int64_t>* list,
int64_t ndims, int64_t* shape) {
set_and_validate_list_shape(list, ndims, shape);
}
void __nac3_array_write_list_to_array(List<int32_t>* list,
NDArray<int32_t>* ndarray) {
write_list_to_array(list, ndarray);
}
void __nac3_array_write_list_to_array64(List<int64_t>* list,
NDArray<int64_t>* ndarray) {
write_list_to_array(list, ndarray);
}
}

345
nac3core/irrt/irrt/ndarray/basic.hpp Normal file
View File

@ -0,0 +1,345 @@
#pragma once
#include <irrt/exception.hpp>
#include <irrt/int_defs.hpp>
#include <irrt/ndarray/def.hpp>
namespace {
namespace ndarray {
namespace basic {
/**
* @brief Asserts that `shape` does not contain negative dimensions.
*
* @param ndims Number of dimensions in `shape`
* @param shape The shape to check on
*/
template <typename SizeT>
void assert_shape_no_negative(SizeT ndims, const SizeT* shape) {
for (SizeT axis = 0; axis < ndims; axis++) {
if (shape[axis] < 0) {
raise_exception(SizeT, EXN_VALUE_ERROR,
"negative dimensions are not allowed; axis {0} "
"has dimension {1}",
axis, shape[axis], NO_PARAM);
}
}
}
/**
* @brief Check two shapes are the same in the context of writing outputting to an ndarray.
*
* This function throws error messages for output shape mismatches.
*/
template <typename SizeT>
void assert_output_shape_same(SizeT ndarray_ndims, const SizeT* ndarray_shape,
SizeT output_ndims, const SizeT* output_shape) {
if (ndarray_ndims != output_ndims) {
// There is no corresponding NumPy error message like this.
raise_exception(
SizeT, EXN_VALUE_ERROR,
"Cannot write output of ndims {0} to an ndarray with ndims {1}",
output_ndims, ndarray_ndims, NO_PARAM);
}
for (SizeT axis = 0; axis < ndarray_ndims; axis++) {
if (ndarray_shape[axis] != output_shape[axis]) {
// There is no corresponding NumPy error message like this.
raise_exception(
SizeT, EXN_VALUE_ERROR,
"Mismatched dimensions on axis {0}, output has "
"dimension {1}, but destination ndarray has dimension {2}.",
axis, output_shape[axis], ndarray_shape[axis]);
}
}
}
/**
* @brief Returns the number of elements of an ndarray given its shape.
*
* @param ndims Number of dimensions in `shape`
* @param shape The shape of the ndarray
*/
template <typename SizeT>
SizeT calc_size_from_shape(SizeT ndims, const SizeT* shape) {
SizeT size = 1;
for (SizeT axis = 0; axis < ndims; axis++) size *= shape[axis];
return size;
}
/**
* @brief Compute the array indices of the `nth` (0-based) element of an ndarray given only its shape.
*
* @param ndims Number of elements in `shape` and `indices`
* @param shape The shape of the ndarray
* @param indices The returned indices indexing the ndarray with shape `shape`.
* @param nth The index of the element of interest.
*/
template <typename SizeT>
void set_indices_by_nth(SizeT ndims, const SizeT* shape, SizeT* indices,
SizeT nth) {
for (SizeT i = 0; i < ndims; i++) {
SizeT axis = ndims - i - 1;
SizeT dim = shape[axis];
indices[axis] = nth % dim;
nth /= dim;
}
}
/**
* @brief Return the number of elements of an `ndarray`
*
* This function corresponds to `<an_ndarray>.size`
*/
template <typename SizeT>
SizeT size(const NDArray<SizeT>* ndarray) {
return calc_size_from_shape(ndarray->ndims, ndarray->shape);
}
/**
* @brief Return of the number of its content of an `ndarray`.
*
* This function corresponds to `<an_ndarray>.nbytes`.
*/
template <typename SizeT>
SizeT nbytes(const NDArray<SizeT>* ndarray) {
return size(ndarray) * ndarray->itemsize;
}
/**
* @brief Get the `len()` of an ndarray, and asserts that `ndarray` is a sized object.
*
* This function corresponds to `<an_ndarray>.__len__`.
*
* @param dst_length The returned result
*/
template <typename SizeT>
SizeT len(const NDArray<SizeT>* ndarray) {
// numpy prohibits `__len__` on unsized objects
if (ndarray->ndims == 0) {
raise_exception(SizeT, EXN_TYPE_ERROR, "len() of unsized object",
NO_PARAM, NO_PARAM, NO_PARAM);
} else {
return ndarray->shape[0];
}
}
/**
* @brief Return a boolean indicating if `ndarray` is (C-)contiguous.
*
* You may want to see: ndarray's rules for C-contiguity: https://github.com/numpy/numpy/blob/df256d0d2f3bc6833699529824781c58f9c6e697/numpy/core/src/multiarray/flagsobject.c#L95C1-L99C45
*/
template <typename SizeT>
bool is_c_contiguous(const NDArray<SizeT>* ndarray) {
// Other references:
// - tinynumpy's implementation: https://github.com/wadetb/tinynumpy/blob/0d23d22e07062ffab2afa287374c7b366eebdda1/tinynumpy/tinynumpy.py#L102
// - ndarray's flags["C_CONTIGUOUS"]: https://numpy.org/doc/stable/reference/generated/numpy.ndarray.flags.html#numpy.ndarray.flags
// - ndarray's rules for C-contiguity: https://github.com/numpy/numpy/blob/df256d0d2f3bc6833699529824781c58f9c6e697/numpy/core/src/multiarray/flagsobject.c#L95C1-L99C45
// From https://github.com/numpy/numpy/blob/df256d0d2f3bc6833699529824781c58f9c6e697/numpy/core/src/multiarray/flagsobject.c#L95C1-L99C45:
//
// The traditional rule is that for an array to be flagged as C contiguous,
// the following must hold:
//
// strides[-1] == itemsize
// strides[i] == shape[i+1] * strides[i + 1]
// [...]
// According to these rules, a 0- or 1-dimensional array is either both
// C- and F-contiguous, or neither; and an array with 2+ dimensions
// can be C- or F- contiguous, or neither, but not both. Though there
// there are exceptions for arrays with zero or one item, in the first
// case the check is relaxed up to and including the first dimension
// with shape[i] == 0. In the second case `strides == itemsize` will
// can be true for all dimensions and both flags are set.
if (ndarray->ndims == 0) {
return true;
}
if (ndarray->strides[ndarray->ndims - 1] != ndarray->itemsize) {
return false;
}
for (SizeT i = 1; i < ndarray->ndims; i++) {
SizeT axis_i = ndarray->ndims - i - 1;
if (ndarray->strides[axis_i] !=
ndarray->shape[axis_i + 1] * ndarray->strides[axis_i + 1]) {
return false;
}
}
return true;
}
/**
* @brief Return the pointer to the element indexed by `indices`.
*/
template <typename SizeT>
uint8_t* get_pelement_by_indices(const NDArray<SizeT>* ndarray,
const SizeT* indices) {
uint8_t* element = ndarray->data;
for (SizeT dim_i = 0; dim_i < ndarray->ndims; dim_i++)
element += indices[dim_i] * ndarray->strides[dim_i];
return element;
}
/**
* @brief Convenience function. Like `get_pelement_by_indices` but
* reinterprets the element pointer.
*/
template <typename SizeT, typename T>
T* get_ptr(const NDArray<SizeT>* ndarray, const SizeT* indices) {
return (T*)get_pelement_by_indices(ndarray, indices);
}
/**
* @brief Return the pointer to the nth (0-based) element in a flattened view of `ndarray`.
*
* This function does no bound check.
*/
template <typename SizeT>
uint8_t* get_nth_pelement(const NDArray<SizeT>* ndarray, SizeT nth) {
uint8_t* element = ndarray->data;
for (SizeT i = 0; i < ndarray->ndims; i++) {
SizeT axis = ndarray->ndims - i - 1;
SizeT dim = ndarray->shape[axis];
element += ndarray->strides[axis] * (nth % dim);
nth /= dim;
}
return element;
}
/**
* @brief Update the strides of an ndarray given an ndarray `shape`
* and assuming that the ndarray is fully c-contagious.
*
* You might want to read https://ajcr.net/stride-guide-part-1/.
*/
template <typename SizeT>
void set_strides_by_shape(NDArray<SizeT>* ndarray) {
SizeT stride_product = 1;
for (SizeT i = 0; i < ndarray->ndims; i++) {
SizeT axis = ndarray->ndims - i - 1;
ndarray->strides[axis] = stride_product * ndarray->itemsize;
stride_product *= ndarray->shape[axis];
}
}
/**
* @brief Set an element in `ndarray`.
*
* @param pelement Pointer to the element in `ndarray` to be set.
* @param pvalue Pointer to the value `pelement` will be set to.
*/
template <typename SizeT>
void set_pelement_value(NDArray<SizeT>* ndarray, uint8_t* pelement,
const uint8_t* pvalue) {
__builtin_memcpy(pelement, pvalue, ndarray->itemsize);
}
/**
* @brief Copy data from one ndarray to another of the exact same size and itemsize.
*
* Both ndarrays will be viewed in their flatten views when copying the elements.
*/
template <typename SizeT>
void copy_data(const NDArray<SizeT>* src_ndarray, NDArray<SizeT>* dst_ndarray) {
// TODO: Make this faster with memcpy
debug_assert_eq(SizeT, src_ndarray->itemsize, dst_ndarray->itemsize);
for (SizeT i = 0; i < size(src_ndarray); i++) {
auto src_element = ndarray::basic::get_nth_pelement(src_ndarray, i);
auto dst_element = ndarray::basic::get_nth_pelement(dst_ndarray, i);
ndarray::basic::set_pelement_value(dst_ndarray, dst_element,
src_element);
}
}
} // namespace basic
} // namespace ndarray
} // namespace
extern "C" {
using namespace ndarray::basic;
void __nac3_ndarray_util_assert_shape_no_negative(int32_t ndims,
int32_t* shape) {
assert_shape_no_negative(ndims, shape);
}
void __nac3_ndarray_util_assert_shape_no_negative64(int64_t ndims,
int64_t* shape) {
assert_shape_no_negative(ndims, shape);
}
void __nac3_ndarray_util_assert_output_shape_same(int32_t ndarray_ndims,
const int32_t* ndarray_shape,
int32_t output_ndims,
const int32_t* output_shape) {
assert_output_shape_same(ndarray_ndims, ndarray_shape, output_ndims,
output_shape);
}
void __nac3_ndarray_util_assert_output_shape_same64(
int64_t ndarray_ndims, const int64_t* ndarray_shape, int64_t output_ndims,
const int64_t* output_shape) {
assert_output_shape_same(ndarray_ndims, ndarray_shape, output_ndims,
output_shape);
}
uint32_t __nac3_ndarray_size(NDArray<int32_t>* ndarray) {
return size(ndarray);
}
uint64_t __nac3_ndarray_size64(NDArray<int64_t>* ndarray) {
return size(ndarray);
}
uint32_t __nac3_ndarray_nbytes(NDArray<int32_t>* ndarray) {
return nbytes(ndarray);
}
uint64_t __nac3_ndarray_nbytes64(NDArray<int64_t>* ndarray) {
return nbytes(ndarray);
}
int32_t __nac3_ndarray_len(NDArray<int32_t>* ndarray) { return len(ndarray); }
int64_t __nac3_ndarray_len64(NDArray<int64_t>* ndarray) { return len(ndarray); }
bool __nac3_ndarray_is_c_contiguous(NDArray<int32_t>* ndarray) {
return is_c_contiguous(ndarray);
}
bool __nac3_ndarray_is_c_contiguous64(NDArray<int64_t>* ndarray) {
return is_c_contiguous(ndarray);
}
uint8_t* __nac3_ndarray_get_nth_pelement(const NDArray<int32_t>* ndarray,
int32_t nth) {
return get_nth_pelement(ndarray, nth);
}
uint8_t* __nac3_ndarray_get_nth_pelement64(const NDArray<int64_t>* ndarray,
int64_t nth) {
return get_nth_pelement(ndarray, nth);
}
void __nac3_ndarray_set_strides_by_shape(NDArray<int32_t>* ndarray) {
set_strides_by_shape(ndarray);
}
void __nac3_ndarray_set_strides_by_shape64(NDArray<int64_t>* ndarray) {
set_strides_by_shape(ndarray);
}
void __nac3_ndarray_copy_data(NDArray<int32_t>* src_ndarray,
NDArray<int32_t>* dst_ndarray) {
copy_data(src_ndarray, dst_ndarray);
}
void __nac3_ndarray_copy_data64(NDArray<int64_t>* src_ndarray,
NDArray<int64_t>* dst_ndarray) {
copy_data(src_ndarray, dst_ndarray);
}
}

171
nac3core/irrt/irrt/ndarray/broadcast.hpp Normal file
View File

@ -0,0 +1,171 @@
#pragma once
#include <irrt/int_defs.hpp>
#include <irrt/ndarray/def.hpp>
#include <irrt/slice.hpp>
namespace {
template <typename SizeT>
struct ShapeEntry {
SizeT ndims;
SizeT* shape;
};
} // namespace
namespace {
namespace ndarray {
namespace broadcast {
/**
* @brief Return true if `src_shape` can broadcast to `dst_shape`.
*
* See https://numpy.org/doc/stable/user/basics.broadcasting.html
*/
template <typename SizeT>
bool can_broadcast_shape_to(SizeT target_ndims, const SizeT* target_shape,
SizeT src_ndims, const SizeT* src_shape) {
if (src_ndims > target_ndims) {
return false;
}
for (SizeT i = 0; i < src_ndims; i++) {
SizeT target_dim = target_shape[target_ndims - i - 1];
SizeT src_dim = src_shape[src_ndims - i - 1];
if (!(src_dim == 1 || target_dim == src_dim)) {
return false;
}
}
return true;
}
/**
* @brief Performs `np.broadcast_shapes(<shapes>)`
*
* @param num_shapes Number of entries in `shapes`
* @param shapes The list of shape to do `np.broadcast_shapes` on.
* @param dst_ndims The length of `dst_shape`.
* `dst_ndims` must be `max([shape.ndims for shape in shapes])`, but the caller has to calculate it/provide it.
* for this function since they should already know in order to allocate `dst_shape` in the first place.
* @param dst_shape The resulting shape. Must be pre-allocated by the caller. This function calculate the result
* of `np.broadcast_shapes` and write it here.
*/
template <typename SizeT>
void broadcast_shapes(SizeT num_shapes, const ShapeEntry<SizeT>* shapes,
SizeT dst_ndims, SizeT* dst_shape) {
for (SizeT dst_axis = 0; dst_axis < dst_ndims; dst_axis++) {
dst_shape[dst_axis] = 1;
}
#ifdef IRRT_DEBUG_ASSERT
SizeT max_ndims_found = 0;
#endif
for (SizeT i = 0; i < num_shapes; i++) {
ShapeEntry<SizeT> entry = shapes[i];
// Check pre-condition: `dst_ndims` must be `max([shape.ndims for shape in shapes])`
debug_assert(SizeT, entry.ndims <= dst_ndims);
#ifdef IRRT_DEBUG_ASSERT
max_ndims_found = max(max_ndims_found, entry.ndims);
#endif
for (SizeT j = 0; j < entry.ndims; j++) {
SizeT entry_axis = entry.ndims - j - 1;
SizeT dst_axis = dst_ndims - j - 1;
SizeT entry_dim = entry.shape[entry_axis];
SizeT dst_dim = dst_shape[dst_axis];
if (dst_dim == 1) {
dst_shape[dst_axis] = entry_dim;
} else if (entry_dim == 1 || entry_dim == dst_dim) {
// Do nothing
} else {
raise_exception(SizeT, EXN_VALUE_ERROR,
"shape mismatch: objects cannot be broadcast "
"to a single shape.",
NO_PARAM, NO_PARAM, NO_PARAM);
}
}
}
// Check pre-condition: `dst_ndims` must be `max([shape.ndims for shape in shapes])`
debug_assert_eq(SizeT, max_ndims_found, dst_ndims);
}
/**
* @brief Perform `np.broadcast_to(<ndarray>, <target_shape>)` and appropriate assertions.
*
* This function attempts to broadcast `src_ndarray` to a new shape defined by `dst_ndarray.shape`,
* and return the result by modifying `dst_ndarray`.
*
* # Notes on `dst_ndarray`
* The caller is responsible for allocating space for the resulting ndarray.
* Here is what this function expects from `dst_ndarray` when called:
* - `dst_ndarray->data` does not have to be initialized.
* - `dst_ndarray->itemsize` does not have to be initialized.
* - `dst_ndarray->ndims` must be initialized, determining the length of `dst_ndarray->shape`
* - `dst_ndarray->shape` must be allocated, and must contain the desired target broadcast shape.
* - `dst_ndarray->strides` must be allocated, through it can contain uninitialized values.
* When this function call ends:
* - `dst_ndarray->data` is set to `src_ndarray->data` (`dst_ndarray` is just a view to `src_ndarray`)
* - `dst_ndarray->itemsize` is set to `src_ndarray->itemsize`
* - `dst_ndarray->ndims` is unchanged.
* - `dst_ndarray->shape` is unchanged.
* - `dst_ndarray->strides` is updated accordingly by how ndarray broadcast_to works.
*/
template <typename SizeT>
void broadcast_to(const NDArray<SizeT>* src_ndarray,
NDArray<SizeT>* dst_ndarray) {
if (!ndarray::broadcast::can_broadcast_shape_to(
dst_ndarray->ndims, dst_ndarray->shape, src_ndarray->ndims,
src_ndarray->shape)) {
raise_exception(SizeT, EXN_VALUE_ERROR,
"operands could not be broadcast together", NO_PARAM,
NO_PARAM, NO_PARAM);
}
dst_ndarray->data = src_ndarray->data;
dst_ndarray->itemsize = src_ndarray->itemsize;
for (SizeT i = 0; i < dst_ndarray->ndims; i++) {
SizeT src_axis = src_ndarray->ndims - i - 1;
SizeT dst_axis = dst_ndarray->ndims - i - 1;
if (src_axis < 0 || (src_ndarray->shape[src_axis] == 1 &&
dst_ndarray->shape[dst_axis] != 1)) {
// Freeze the steps in-place
dst_ndarray->strides[dst_axis] = 0;
} else {
dst_ndarray->strides[dst_axis] = src_ndarray->strides[src_axis];
}
}
}
} // namespace broadcast
} // namespace ndarray
} // namespace
extern "C" {
using namespace ndarray::broadcast;
void __nac3_ndarray_broadcast_to(NDArray<int32_t>* src_ndarray,
NDArray<int32_t>* dst_ndarray) {
broadcast_to(src_ndarray, dst_ndarray);
}
void __nac3_ndarray_broadcast_to64(NDArray<int64_t>* src_ndarray,
NDArray<int64_t>* dst_ndarray) {
broadcast_to(src_ndarray, dst_ndarray);
}
void __nac3_ndarray_broadcast_shapes(int32_t num_shapes,
const ShapeEntry<int32_t>* shapes,
int32_t dst_ndims, int32_t* dst_shape) {
broadcast_shapes(num_shapes, shapes, dst_ndims, dst_shape);
}
void __nac3_ndarray_broadcast_shapes64(int64_t num_shapes,
const ShapeEntry<int64_t>* shapes,
int64_t dst_ndims, int64_t* dst_shape) {
broadcast_shapes(num_shapes, shapes, dst_ndims, dst_shape);
}
}

44
nac3core/irrt/irrt/ndarray/def.hpp Normal file
View File

@ -0,0 +1,44 @@
#pragma once
namespace {
/**
* @brief The NDArray object
*
* The official numpy implementations: https://github.com/numpy/numpy/blob/735a477f0bc2b5b84d0e72d92f224bde78d4e069/doc/source/reference/c-api/types-and-structures.rst
*/
template <typename SizeT>
struct NDArray {
/**
* @brief The underlying data this `ndarray` is pointing to.
*
* Must be set to `nullptr` to indicate that this NDArray's `data` is uninitialized.
*/
uint8_t* data;
/**
* @brief The number of bytes of a single element in `data`.
*/
SizeT itemsize;
/**
* @brief The number of dimensions of this shape.
*/
SizeT ndims;
/**
* @brief The NDArray shape, with length equal to `ndims`.
*
* Note that it may contain 0.
*/
SizeT* shape;
/**
* @brief Array strides, with length equal to `ndims`
*
* The stride values are in units of bytes, not number of elements.
*
* Note that `strides` can have negative values.
*/
SizeT* strides;
};
} // namespace

221
nac3core/irrt/irrt/ndarray/indexing.hpp Normal file
View File

@ -0,0 +1,221 @@
#pragma once
#include <irrt/exception.hpp>
#include <irrt/int_defs.hpp>
#include <irrt/ndarray/basic.hpp>
#include <irrt/ndarray/def.hpp>
#include <irrt/slice.hpp>
namespace {
typedef uint8_t NDIndexType;
/**
* @brief A single element index
*
* `data` points to a `SliceIndex`.
*/
const NDIndexType ND_INDEX_TYPE_SINGLE_ELEMENT = 0;
/**
* @brief A slice index
*
* `data` points to a `UserRange`.
*/
const NDIndexType ND_INDEX_TYPE_SLICE = 1;
/**
* @brief `np.newaxis` / `None`
*
* `data` is unused.
*/
const NDIndexType ND_INDEX_TYPE_NEWAXIS = 2;
/**
* @brief `Ellipsis` / `...`
*
* `data` is unused.
*/
const NDIndexType ND_INDEX_TYPE_ELLIPSIS = 3;
/**
* @brief An index used in ndarray indexing
*/
struct NDIndex {
/**
* @brief Enum tag to specify the type of index.
*
* Please see comments of each enum constant.
*/
NDIndexType type;
/**
* @brief The accompanying data associated with `type`.
*
* Please see comments of each enum constant.
*/
uint8_t* data;
};
} // namespace
namespace {
namespace ndarray {
namespace indexing {
/**
* @brief Perform ndarray "basic indexing" (https://numpy.org/doc/stable/user/basics.indexing.html#basic-indexing)
*
* This is function very similar to performing `dst_ndarray = src_ndarray[indexes]` in Python (where the variables
* can all be found in the parameter of this function).
*
* In other words, this function takes in an ndarray (`src_ndarray`), index it with `indexes`, and return the
* indexed array (by writing the result to `dst_ndarray`).
*
* This function also does proper assertions on `indexes`.
*
* # Notes on `dst_ndarray`
* The caller is responsible for allocating space for the resulting ndarray.
* Here is what this function expects from `dst_ndarray` when called:
* - `dst_ndarray->data` does not have to be initialized.
* - `dst_ndarray->itemsize` does not have to be initialized.
* - `dst_ndarray->ndims` must be initialized, and it must be equal to the expected `ndims` of the `dst_ndarray` after
* indexing `src_ndarray` with `indexes`.
* - `dst_ndarray->shape` must be allocated, through it can contain uninitialized values.
* - `dst_ndarray->strides` must be allocated, through it can contain uninitialized values.
* When this function call ends:
* - `dst_ndarray->data` is set to `src_ndarray->data` (`dst_ndarray` is just a view to `src_ndarray`)
* - `dst_ndarray->itemsize` is set to `src_ndarray->itemsize`
* - `dst_ndarray->ndims` is unchanged.
* - `dst_ndarray->shape` is updated according to how `src_ndarray` is indexed.
* - `dst_ndarray->strides` is updated accordingly by how ndarray indexing works.
*
* @param indexes Indexes to index `src_ndarray`, ordered in the same way you would write them in Python.
* @param src_ndarray The NDArray to be indexed.
* @param dst_ndarray The resulting NDArray after indexing. Further details in the comments above,
*/
template <typename SizeT>
void index(SizeT num_indexes, const NDIndex* indexes,
const NDArray<SizeT>* src_ndarray, NDArray<SizeT>* dst_ndarray) {
// First, validate `indexes`.
// Expected value of `dst_ndarray->ndims`.
SizeT expected_dst_ndims = src_ndarray->ndims;
// To check for "too many indices for array: array is ?-dimensional, but ? were indexed"
SizeT num_indexed = 0;
// There may be ellipsis `...` in `indexes`. There can only be 0 or 1 ellipsis.
SizeT num_ellipsis = 0;
for (SizeT i = 0; i < num_indexes; i++) {
if (indexes[i].type == ND_INDEX_TYPE_SINGLE_ELEMENT) {
expected_dst_ndims--;
num_indexed++;
} else if (indexes[i].type == ND_INDEX_TYPE_SLICE) {
num_indexed++;
} else if (indexes[i].type == ND_INDEX_TYPE_NEWAXIS) {
expected_dst_ndims++;
} else if (indexes[i].type == ND_INDEX_TYPE_ELLIPSIS) {
num_ellipsis++;
if (num_ellipsis > 1) {
raise_exception(
SizeT, EXN_INDEX_ERROR,
"an index can only have a single ellipsis ('...')",
NO_PARAM, NO_PARAM, NO_PARAM);
}
} else {
__builtin_unreachable();
}
}
debug_assert_eq(SizeT, expected_dst_ndims, dst_ndarray->ndims);
if (src_ndarray->ndims - num_indexed < 0) {
raise_exception(SizeT, EXN_INDEX_ERROR,
"too many indices for array: array is {0}-dimensional, "
"but {1} were indexed",
src_ndarray->ndims, num_indexes, NO_PARAM);
}
dst_ndarray->data = src_ndarray->data;
dst_ndarray->itemsize = src_ndarray->itemsize;
// Reference code: https://github.com/wadetb/tinynumpy/blob/0d23d22e07062ffab2afa287374c7b366eebdda1/tinynumpy/tinynumpy.py#L652
SizeT src_axis = 0;
SizeT dst_axis = 0;
for (SliceIndex i = 0; i < num_indexes; i++) {
const NDIndex* index = &indexes[i];
if (index->type == ND_INDEX_TYPE_SINGLE_ELEMENT) {
SliceIndex input = *((SliceIndex*)index->data);
SliceIndex k = slice::resolve_index_in_length(
src_ndarray->shape[src_axis], input);
if (k == slice::OUT_OF_BOUNDS) {
raise_exception(SizeT, EXN_INDEX_ERROR,
"index {0} is out of bounds for axis {1} "
"with size {2}",
input, src_axis, src_ndarray->shape[src_axis]);
}
dst_ndarray->data += k * src_ndarray->strides[src_axis];
src_axis++;
} else if (index->type == ND_INDEX_TYPE_SLICE) {
UserSlice* input = (UserSlice*)index->data;
Slice slice;
input->indices_checked<SizeT>(src_ndarray->shape[src_axis], &slice);
dst_ndarray->data +=
(SizeT)slice.start * src_ndarray->strides[src_axis];
dst_ndarray->strides[dst_axis] =
((SizeT)slice.step) * src_ndarray->strides[src_axis];
dst_ndarray->shape[dst_axis] = (SizeT)slice.len();
dst_axis++;
src_axis++;
} else if (index->type == ND_INDEX_TYPE_NEWAXIS) {
dst_ndarray->strides[dst_axis] = 0;
dst_ndarray->shape[dst_axis] = 1;
dst_axis++;
} else if (index->type == ND_INDEX_TYPE_ELLIPSIS) {
// The number of ':' entries this '...' implies.
SizeT ellipsis_size = src_ndarray->ndims - num_indexed;
for (SizeT j = 0; j < ellipsis_size; j++) {
dst_ndarray->strides[dst_axis] = src_ndarray->strides[src_axis];
dst_ndarray->shape[dst_axis] = src_ndarray->shape[src_axis];
dst_axis++;
src_axis++;
}
} else {
__builtin_unreachable();
}
}
for (; dst_axis < dst_ndarray->ndims; dst_axis++, src_axis++) {
dst_ndarray->shape[dst_axis] = src_ndarray->shape[src_axis];
dst_ndarray->strides[dst_axis] = src_ndarray->strides[src_axis];
}
debug_assert_eq(SizeT, src_ndarray->ndims, src_axis);
debug_assert_eq(SizeT, dst_ndarray->ndims, dst_axis);
}
} // namespace indexing
} // namespace ndarray
} // namespace
extern "C" {
using namespace ndarray::indexing;
void __nac3_ndarray_index(int32_t num_indexes, NDIndex* indexes,
NDArray<int32_t>* src_ndarray,
NDArray<int32_t>* dst_ndarray) {
index(num_indexes, indexes, src_ndarray, dst_ndarray);
}
void __nac3_ndarray_index64(int64_t num_indexes, NDIndex* indexes,
NDArray<int64_t>* src_ndarray,
NDArray<int64_t>* dst_ndarray) {
index(num_indexes, indexes, src_ndarray, dst_ndarray);
}
}

63
nac3core/irrt/irrt/ndarray/iter.hpp Normal file
View File

@ -0,0 +1,63 @@
#pragma once
#include <irrt/int_defs.hpp>
namespace {
/**
* @brief Helper struct to enumerate through all indices under a shape.
*
* i.e., If `shape` is `[3, 2]`, by repeating `next()`, then you get:
* - `[0, 0]`
* - `[0, 1]`
* - `[1, 0]`
* - `[1, 1]`
* - `[2, 0]`
* - `[2, 1]`
* - end.
*
* Interesting cases:
* - If ndims == 0, there is one enumeration.
* - If shape contains zeroes, there are no enumerations.
*/
template <typename SizeT>
struct IndicesIter {
SizeT ndims;
SizeT* shape;
SizeT* indices;
SizeT size; // Product of shape
SizeT nth; // The nth (0-based) index of the current indices.
IndicesIter(SizeT ndims, SizeT* shape, SizeT* indices) {
this->ndims = ndims;
this->shape = shape;
this->indices = indices;
reset();
this->size = 1;
for (SizeT i = 0; i < ndims; i++) {
this->size *= shape[i];
}
}
void reset() {
for (SizeT axis = 0; axis < ndims; axis++) indices[axis] = 0;
nth = 0;
}
bool ok() { return nth < size; }
void next() {
for (SizeT i = 0; i < ndims; i++) {
SizeT axis = ndims - i - 1;
indices[axis]++;
if (indices[axis] >= shape[axis]) {
indices[axis] = 0;
} else {
break;
}
}
nth++;
}
};
} // namespace

192
nac3core/irrt/irrt/ndarray/product.hpp Normal file
View File

@ -0,0 +1,192 @@
#pragma once
#include <irrt/int_defs.hpp>
#include <irrt/ndarray/basic.hpp>
#include <irrt/ndarray/broadcast.hpp>
#include <irrt/ndarray/iter.hpp>
// NOTE: Everything would be much easier and elegant if einsum is implemented.
namespace {
namespace ndarray {
namespace matmul {
/*
* In einsum notation, the output is the broadcasts performed by `np.einsum("...ij,...jk->...ik", a, b)`.
*
* Example:
* Suppose `a_shape == [99, 1, 97, 4, 2]`
* and `b_shape == [ 1, 98, 1, 2, 5]`,
*
* ...then `new_a_shape == [99, 98, 97, 4, 2]`,
* `new_b_shape == [99, 98, 97, 2, 5]`,
* and `dst_shape == [99, 98, 97, 4, 5]`.
* ^^^^^^^^^^ ^^^^
* (by broadcast) (4x2 @ 2x5 => 4x5)
*/
template <typename SizeT>
void calculate_shapes(SizeT a_ndims, SizeT* a_shape, SizeT b_ndims,
SizeT* b_shape, SizeT final_ndims, SizeT* new_a_shape,
SizeT* new_b_shape, SizeT* dst_shape) {
debug_assert(SizeT, a_ndims >= 2);
debug_assert(SizeT, b_ndims >= 2);
debug_assert_eq(SizeT, max(a_ndims, b_ndims), final_ndims);
const SizeT num_entries = 2;
ShapeEntry<SizeT> entries[num_entries] = {
{.ndims = a_ndims - 2, .shape = a_shape},
{.ndims = b_ndims - 2, .shape = b_shape}};
ndarray::broadcast::broadcast_shapes<SizeT>(num_entries, entries,
final_ndims - 2, new_a_shape);
ndarray::broadcast::broadcast_shapes<SizeT>(num_entries, entries,
final_ndims - 2, new_b_shape);
ndarray::broadcast::broadcast_shapes<SizeT>(num_entries, entries,
final_ndims - 2, dst_shape);
new_a_shape[final_ndims - 2] = a_shape[a_ndims - 2];
new_a_shape[final_ndims - 1] = a_shape[a_ndims - 1];
new_b_shape[final_ndims - 2] = b_shape[b_ndims - 2];
new_b_shape[final_ndims - 1] = b_shape[b_ndims - 1];
dst_shape[final_ndims - 2] = a_shape[a_ndims - 2];
dst_shape[final_ndims - 1] = b_shape[b_ndims - 1];
}
/**
* @brief Perform `np.matmul(a, b)` but the inputs are both rank >=2 matrices and `a.shape[:-2] == b.shape[:-2]`.
*
* The compatibility of `a` and `b` (for their `.shape[-2:]`) are asserted.
*
* Also see https://numpy.org/doc/stable/reference/generated/numpy.matmul.html#numpy-matmul.
*
* This function expects `dst_ndarray` to contain the following content when called:
* - `dst_ndarray->data` is allocated. Can be uninitialized.
* - `dst_ndarray->itemsize` is set to `sizeof(T)`.
* - `dst_ndarray->ndims` is set appropriately.
* - `dst_ndarray->shape` is set appropriately.
* - `dst_ndarray->strides` is ignored.
*
* Moreover, the shapes of `a_ndarray`, `b_ndarray`, and `dst_ndarray` **must be the same**. This implies
*/
template <typename SizeT, typename T>
void matmul_at_least_2d(NDArray<SizeT>* a_ndarray, NDArray<SizeT>* b_ndarray,
NDArray<SizeT>* dst_ndarray) {
// All inputs' ndims should be >= 2 and be the same.
debug_assert_eq(SizeT, a_ndarray->ndims, b_ndarray->ndims);
debug_assert_eq(SizeT, a_ndarray->ndims, dst_ndarray->ndims);
debug_assert(SizeT, a_ndarray->ndims >= 2);
debug_assert_eq(SizeT, a_ndarray->itemsize, sizeof(T));
debug_assert_eq(SizeT, b_ndarray->itemsize, sizeof(T));
debug_assert_eq(SizeT, dst_ndarray->itemsize, sizeof(T));
if (IRRT_DEBUG_ASSERT_BOOL) {
// Check that the shapes are the same.
for (SizeT i = 0; i < a_ndarray->ndims - 2; i++) {
if (dst_ndarray->shape[0] != a_ndarray->shape[0]) {
raise_debug_assert(
SizeT, "Bad shape. At axis {0}, a has {1}, dst has {2}", i,
a_ndarray->shape[i], dst_ndarray->shape[i]);
}
if (dst_ndarray->shape[0] != b_ndarray->shape[0]) {
raise_debug_assert(
SizeT, "Bad shape. At axis {0}, b has {1}, dst has {2}", i,
b_ndarray->shape[i], dst_ndarray->shape[i]);
}
}
}
// Number of dimensions dedicated to stacking
// e.g., [4, 6, 1, 2, 3]
// ^^^^^^^ count these
const SizeT u = a_ndarray->ndims - 2; // Alias
SizeT* a_mat_shape = a_ndarray->shape + u;
SizeT* b_mat_shape = b_ndarray->shape + u;
SizeT* dst_mat_shape = dst_ndarray->shape + u;
// Assert that dst_ndarray has the correct shape
debug_assert_eq(SizeT, dst_mat_shape[0], a_mat_shape[0]);
debug_assert_eq(SizeT, dst_mat_shape[1], b_mat_shape[1]);
// Check that a and b are compatible for matmul
if (a_mat_shape[1] != b_mat_shape[0]) {
// This is a custom error message. Different from NumPy.
raise_exception(
SizeT, EXN_VALUE_ERROR,
"Cannot multiply LHS (shape ?x{0}) with RHS (shape {1}x?})",
a_mat_shape[1], b_mat_shape[0], NO_PARAM);
}
// Iterate through shape[:-2]. i.e,
// Given a = [5, 4, 3, m, p] and b = [5, 4, 3, p, n]. We iterate through [5, 4, 3].
SizeT* indices =
(SizeT*)__builtin_alloca(sizeof(SizeT) * dst_ndarray->ndims);
SizeT* mat_indices = indices + u;
IndicesIter<SizeT> iter(u, dst_ndarray->shape, indices);
for (; iter.ok(); iter.next()) {
for (SizeT i = 0; i < dst_mat_shape[0]; i++) {
for (SizeT j = 0; j < dst_mat_shape[1]; j++) {
// `indices` is being reused to index into different ndarrays.
mat_indices[0] = i;
mat_indices[1] = j;
T* d = ndarray::basic::get_ptr<SizeT, T>(dst_ndarray, indices);
*d = 0;
for (SizeT k = 0; k < a_ndarray->shape[1]; k++) {
mat_indices[0] = i;
mat_indices[1] = k;
T* a =
ndarray::basic::get_ptr<SizeT, T>(a_ndarray, indices);
mat_indices[0] = k;
mat_indices[1] = j;
T* b =
ndarray::basic::get_ptr<SizeT, T>(b_ndarray, indices);
*d += (*a) * (*b);
}
}
}
}
}
} // namespace matmul
} // namespace ndarray
} // namespace
extern "C" {
using namespace ndarray::matmul;
void __nac3_ndarray_matmul_calculate_shapes(int32_t a_ndims, int32_t* a_shape,
int32_t b_ndims, int32_t* b_shape,
int32_t final_ndims,
int32_t* new_a_shape,
int32_t* new_b_shape,
int32_t* dst_shape) {
calculate_shapes(a_ndims, a_shape, b_ndims, b_shape, final_ndims,
new_a_shape, new_b_shape, dst_shape);
}
void __nac3_ndarray_matmul_calculate_shapes64(int64_t a_ndims, int64_t* a_shape,
int64_t b_ndims, int64_t* b_shape,
int64_t final_ndims,
int64_t* new_a_shape,
int64_t* new_b_shape,
int64_t* dst_shape) {
calculate_shapes(a_ndims, a_shape, b_ndims, b_shape, final_ndims,
new_a_shape, new_b_shape, dst_shape);
}
void __nac3_ndarray_float64_matmul_at_least_2d(NDArray<int32_t>* a_ndarray,
NDArray<int32_t>* b_ndarray,
NDArray<int32_t>* dst_ndarray) {
matmul_at_least_2d<int32_t, double>(a_ndarray, b_ndarray, dst_ndarray);
}
void __nac3_ndarray_float64_matmul_at_least_2d64(
NDArray<int64_t>* a_ndarray, NDArray<int64_t>* b_ndarray,
NDArray<int64_t>* dst_ndarray) {
matmul_at_least_2d<int64_t, double>(a_ndarray, b_ndarray, dst_ndarray);
}
}

106
nac3core/irrt/irrt/ndarray/reshape.hpp Normal file
View File

@ -0,0 +1,106 @@
#pragma once
#include <irrt/int_defs.hpp>
#include <irrt/ndarray/def.hpp>
namespace {
namespace ndarray {
namespace reshape {
/**
* @brief Perform assertions on and resolve unknown dimensions in `new_shape` in `np.reshape(<ndarray>, new_shape)`
*
* If `new_shape` indeed contains unknown dimensions (specified with `-1`, just like numpy), `new_shape` will be
* modified to contain the resolved dimension.
*
* To perform assertions on and resolve unknown dimensions in `new_shape`, we don't need the actual
* `<ndarray>` object itself, but only the `.size` of the `<ndarray>`.
*
* @param size The `.size` of `<ndarray>`
* @param new_ndims Number of elements in `new_shape`
* @param new_shape Target shape to reshape to
*/
template <typename SizeT>
void resolve_and_check_new_shape(SizeT size, SizeT new_ndims,
SizeT* new_shape) {
// Is there a -1 in `new_shape`?
bool neg1_exists = false;
// Location of -1, only initialized if `neg1_exists` is true
SizeT neg1_axis_i;
// The computed ndarray size of `new_shape`
SizeT new_size = 1;
for (SizeT axis_i = 0; axis_i < new_ndims; axis_i++) {
SizeT dim = new_shape[axis_i];
if (dim < 0) {
if (dim == -1) {
if (neg1_exists) {
// Multiple `-1` found. Throw an error.
raise_exception(SizeT, EXN_VALUE_ERROR,
"can only specify one unknown dimension",
NO_PARAM, NO_PARAM, NO_PARAM);
} else {
neg1_exists = true;
neg1_axis_i = axis_i;
}
} else {
// TODO: What? In `np.reshape` any negative dimensions is
// treated like its `-1`.
//
// Try running `np.zeros((3, 4)).reshape((-999, 2))`
//
// It is not documented by numpy.
// Throw an error for now...
raise_exception(
SizeT, EXN_VALUE_ERROR,
"Found non -1 negative dimension {0} on axis {1}", dim,
axis_i, NO_PARAM);
}
} else {
new_size *= dim;
}
}
bool can_reshape;
if (neg1_exists) {
// Let `x` be the unknown dimension
// Solve `x * <new_size> = <size>`
if (new_size == 0 && size == 0) {
// `x` has infinitely many solutions
can_reshape = false;
} else if (new_size == 0 && size != 0) {
// `x` has no solutions
can_reshape = false;
} else if (size % new_size != 0) {
// `x` has no integer solutions
can_reshape = false;
} else {
can_reshape = true;
new_shape[neg1_axis_i] = size / new_size; // Resolve dimension
}
} else {
can_reshape = (new_size == size);
}
if (!can_reshape) {
raise_exception(SizeT, EXN_VALUE_ERROR,
"cannot reshape array of size {0} into given shape",
size, NO_PARAM, NO_PARAM);
}
}
} // namespace reshape
} // namespace ndarray
} // namespace
extern "C" {
void __nac3_ndarray_resolve_and_check_new_shape(int32_t size, int32_t new_ndims,
int32_t* new_shape) {
ndarray::reshape::resolve_and_check_new_shape(size, new_ndims, new_shape);
}
void __nac3_ndarray_resolve_and_check_new_shape64(int64_t size,
int64_t new_ndims,
int64_t* new_shape) {
ndarray::reshape::resolve_and_check_new_shape(size, new_ndims, new_shape);
}
}

145
nac3core/irrt/irrt/ndarray/transpose.hpp Normal file
View File

@ -0,0 +1,145 @@
#pragma once
#include <irrt/int_defs.hpp>
#include <irrt/ndarray/def.hpp>
#include <irrt/slice.hpp>
/*
* Notes on `np.transpose(<array>, <axes>)`
*
* TODO: `axes`, if specified, can actually contain negative indices,
* but it is not documented in numpy.
*
* Supporting it for now.
*/
namespace {
namespace ndarray {
namespace transpose {
/**
* @brief Do assertions on `<axes>` in `np.transpose(<array>, <axes>)`.
*
* Note that `np.transpose`'s `<axe>` argument is optional. If the argument
* is specified but the user, use this function to do assertions on it.
*
* @param ndims The number of dimensions of `<array>`
* @param num_axes Number of elements in `<axes>` as specified by the user.
* This should be equal to `ndims`. If not, a "ValueError: axes don't match array" is thrown.
* @param axes The user specified `<axes>`.
*/
template <typename SizeT>
void assert_transpose_axes(SizeT ndims, SizeT num_axes, const SizeT* axes) {
if (ndims != num_axes) {
raise_exception(SizeT, EXN_VALUE_ERROR, "axes don't match array",
NO_PARAM, NO_PARAM, NO_PARAM);
}
// TODO: Optimize this
bool* axe_specified = (bool*)__builtin_alloca(sizeof(bool) * ndims);
for (SizeT i = 0; i < ndims; i++) axe_specified[i] = false;
for (SizeT i = 0; i < ndims; i++) {
SizeT axis = slice::resolve_index_in_length(ndims, axes[i]);
if (axis == slice::OUT_OF_BOUNDS) {
// TODO: numpy actually throws a `numpy.exceptions.AxisError`
raise_exception(
SizeT, EXN_VALUE_ERROR,
"axis {0} is out of bounds for array of dimension {1}", axis,
ndims, NO_PARAM);
}
if (axe_specified[axis]) {
raise_exception(SizeT, EXN_VALUE_ERROR,
"repeated axis in transpose", NO_PARAM, NO_PARAM,
NO_PARAM);
}
axe_specified[axis] = true;
}
}
/**
* @brief Create a transpose view of `src_ndarray` and perform proper assertions.
*
* This function is very similar to doing `dst_ndarray = np.transpose(src_ndarray, <axes>)`.
* If `<axes>` is supposed to be `None`, caller can pass in a `nullptr` to `<axes>`.
*
* The transpose view created is returned by modifying `dst_ndarray`.
*
* The caller is responsible for setting up `dst_ndarray` before calling this function.
* Here is what this function expects from `dst_ndarray` when called:
* - `dst_ndarray->data` does not have to be initialized.
* - `dst_ndarray->itemsize` does not have to be initialized.
* - `dst_ndarray->ndims` must be initialized, must be equal to `src_ndarray->ndims`.
* - `dst_ndarray->shape` must be allocated, through it can contain uninitialized values.
* - `dst_ndarray->strides` must be allocated, through it can contain uninitialized values.
* When this function call ends:
* - `dst_ndarray->data` is set to `src_ndarray->data` (`dst_ndarray` is just a view to `src_ndarray`)
* - `dst_ndarray->itemsize` is set to `src_ndarray->itemsize`
* - `dst_ndarray->ndims` is unchanged
* - `dst_ndarray->shape` is updated according to how `np.transpose` works
* - `dst_ndarray->strides` is updated according to how `np.transpose` works
*
* @param src_ndarray The NDArray to build a transpose view on
* @param dst_ndarray The resulting NDArray after transpose. Further details in the comments above,
* @param num_axes Number of elements in axes. Unused if `axes` is nullptr.
* @param axes Axes permutation. Set it to `nullptr` if `<axes>` is `None`.
*/
template <typename SizeT>
void transpose(const NDArray<SizeT>* src_ndarray, NDArray<SizeT>* dst_ndarray,
SizeT num_axes, const SizeT* axes) {
debug_assert_eq(SizeT, src_ndarray->ndims, dst_ndarray->ndims);
const auto ndims = src_ndarray->ndims;
if (axes != nullptr) assert_transpose_axes(ndims, num_axes, axes);
dst_ndarray->data = src_ndarray->data;
dst_ndarray->itemsize = src_ndarray->itemsize;
// Check out https://ajcr.net/stride-guide-part-2/ to see how `np.transpose` works behind the scenes.
if (axes == nullptr) {
// `np.transpose(<array>, axes=None)`
/*
* Minor note: `np.transpose(<array>, axes=None)` is equivalent to
* `np.transpose(<array>, axes=[N-1, N-2, ..., 0])` - basically it
* is reversing the order of strides and shape.
*
* This is a fast implementation to handle this special (but very common) case.
*/
for (SizeT axis = 0; axis < ndims; axis++) {
dst_ndarray->shape[axis] = src_ndarray->shape[ndims - axis - 1];
dst_ndarray->strides[axis] = src_ndarray->strides[ndims - axis - 1];
}
} else {
// `np.transpose(<array>, <axes>)`
// Permute strides and shape according to `axes`, while resolving negative indices in `axes`
for (SizeT axis = 0; axis < ndims; axis++) {
// `i` cannot be OUT_OF_BOUNDS because of assertions
SizeT i = slice::resolve_index_in_length(ndims, axes[axis]);
dst_ndarray->shape[axis] = src_ndarray->shape[i];
dst_ndarray->strides[axis] = src_ndarray->strides[i];
}
}
}
} // namespace transpose
} // namespace ndarray
} // namespace
extern "C" {
using namespace ndarray::transpose;
void __nac3_ndarray_transpose(const NDArray<int32_t>* src_ndarray,
NDArray<int32_t>* dst_ndarray, int32_t num_axes,
const int32_t* axes) {
transpose(src_ndarray, dst_ndarray, num_axes, axes);
}
void __nac3_ndarray_transpose64(const NDArray<int64_t>* src_ndarray,
NDArray<int64_t>* dst_ndarray, int64_t num_axes,
const int64_t* axes) {
transpose(src_ndarray, dst_ndarray, num_axes, axes);
}
}

165
nac3core/irrt/irrt/slice.hpp Normal file
View File

@ -0,0 +1,165 @@
#pragma once
#include <irrt/int_defs.hpp>
#include <irrt/slice.hpp>
#include <irrt/util.hpp>
#include "exception.hpp"
// The type of an index or a value describing the length of a
// range/slice is always `int32_t`.
using SliceIndex = int32_t;
namespace {
/**
* @brief A Python-like slice with resolved indices.
*
* "Resolved indices" means that `start` and `stop` must be positive and are
* bound to a known length.
*/
struct Slice {
SliceIndex start;
SliceIndex stop;
SliceIndex step;
/**
* @brief Calculate and return the length / the number of the slice.
*
* If this were a Python range, this function would be `len(range(start, stop, step))`.
*/
SliceIndex len() {
SliceIndex diff = stop - start;
if (diff > 0 && step > 0) {
return ((diff - 1) / step) + 1;
} else if (diff < 0 && step < 0) {
return ((diff + 1) / step) + 1;
} else {
return 0;
}
}
};
namespace slice {
/**
* @brief Resolve a slice index under a given length like Python indexing.
*
* In Python, if you have a `list` of length 100, `list[-1]` resolves to
* `list[99]`, so `resolve_index_in_length_clamped(100, -1)` returns `99`.
*
* If `length` is 0, 0 is returned for any value of `index`.
*
* If `index` is out of bounds, clamps the returned value between `0` and
* `length - 1` (inclusive).
*
*/
SliceIndex resolve_index_in_length_clamped(SliceIndex length,
SliceIndex index) {
if (index < 0) {
return max<SliceIndex>(length + index, 0);
} else {
return min<SliceIndex>(length, index);
}
}
const SliceIndex OUT_OF_BOUNDS = -1;
/**
* @brief Like `resolve_index_in_length_clamped`, but returns `OUT_OF_BOUNDS`
* if `index` is out of bounds.
*/
SliceIndex resolve_index_in_length(SliceIndex length, SliceIndex index) {
SliceIndex resolved = index < 0 ? length + index : index;
if (0 <= resolved && resolved < length) {
return resolved;
} else {
return OUT_OF_BOUNDS;
}
}
} // namespace slice
/**
* @brief A Python-like slice with **unresolved** indices.
*/
struct UserSlice {
bool start_defined;
SliceIndex start;
bool stop_defined;
SliceIndex stop;
bool step_defined;
SliceIndex step;
UserSlice() { this->reset(); }
void reset() {
this->start_defined = false;
this->stop_defined = false;
this->step_defined = false;
}
void set_start(SliceIndex start) {
this->start_defined = true;
this->start = start;
}
void set_stop(SliceIndex stop) {
this->stop_defined = true;
this->stop = stop;
}
void set_step(SliceIndex step) {
this->step_defined = true;
this->step = step;
}
/**
* @brief Resolve this slice.
*
* In Python, this would be `slice(start, stop, step).indices(length)`.
*
* @return A `Slice` with the resolved indices.
*/
Slice indices(SliceIndex length) {
Slice result;
result.step = step_defined ? step : 1;
bool step_is_negative = result.step < 0;
if (start_defined) {
result.start =
slice::resolve_index_in_length_clamped(length, start);
} else {
result.start = step_is_negative ? length - 1 : 0;
}
if (stop_defined) {
result.stop = slice::resolve_index_in_length_clamped(length, stop);
} else {
result.stop = step_is_negative ? -1 : length;
}
return result;
}
/**
* @brief Like `.indices()` but with assertions.
*/
template <typename SizeT>
void indices_checked(SliceIndex length, Slice* result) {
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);
}
*result = this->indices(length);
}
};
} // namespace

80
nac3core/irrt/irrt/util.hpp
View File

@ -18,4 +18,84 @@ bool arrays_match(int len, T* as, T* bs) {
}
return true;
}
namespace cstr_utils {
/**
* @brief Return true if `str` is empty.
*/
bool is_empty(const char* str) { return str[0] == '\0'; }
/**
* @brief Implementation of `strcmp()`
*/
int8_t compare(const char* a, const char* b) {
uint32_t i = 0;
while (true) {
if (a[i] < b[i]) {
return -1;
} else if (a[i] > b[i]) {
return 1;
} else {
if (a[i] == '\0') {
return 0;
} else {
i++;
}
}
}
}
/**
* @brief Return true two strings have the same content.
*/
int8_t equal(const char* a, const char* b) { return compare(a, b) == 0; }
/**
* @brief Implementation of `strlen()`.
*/
uint32_t length(const char* str) {
uint32_t length = 0;
while (*str != '\0') {
length++;
str++;
}
return length;
}
/**
* @brief Copy a null-terminated string to a buffer with limited size and guaranteed null-termination.
*
* `dst_max_size` must be greater than 0, otherwise this function has undefined behavior.
*
* This function attempts to copy everything from `src` from `dst`, and *always* null-terminates `dst`.
*
* If the size of `dst` is too small, the final byte (`dst[dst_max_size - 1]`) of `dst` will be set to
* the null terminator.
*
* @param src String to copy from.
* @param dst Buffer to copy string to.
* @param dst_max_size
* Number of bytes of this buffer, including the space needed for the null terminator.
* Must be greater than 0.
* @return If `dst` is too small to contain everything in `src`.
*/
bool copy(const char* src, char* dst, uint32_t dst_max_size) {
for (uint32_t i = 0; i < dst_max_size; i++) {
bool is_last = i + 1 == dst_max_size;
if (is_last && src[i] != '\0') {
dst[i] = '\0';
return false;
}
if (src[i] == '\0') {
dst[i] = '\0';
return true;
}
dst[i] = src[i];
}
__builtin_unreachable();
}
} // namespace cstr_utils
} // namespace

20
nac3core/irrt/irrt_everything.hpp
View File

@ -1,5 +1,23 @@
#pragma once
#ifdef IRRT_DEBUG
#define IRRT_DEBUG_ASSERT
#define IRRT_DEBUG_ASSERT_BOOL true
#else
#define IRRT_DEBUG_ASSERT_BOOL false
#endif
#include <irrt/core.hpp>
#include <irrt/debug.hpp>
#include <irrt/exception.hpp>
#include <irrt/int_defs.hpp>
#include <irrt/util.hpp>
#include <irrt/ndarray/array.hpp>
#include <irrt/ndarray/basic.hpp>
#include <irrt/ndarray/broadcast.hpp>
#include <irrt/ndarray/def.hpp>
#include <irrt/ndarray/indexing.hpp>
#include <irrt/ndarray/iter.hpp>
#include <irrt/ndarray/product.hpp>
#include <irrt/ndarray/reshape.hpp>
#include <irrt/ndarray/transpose.hpp>
#include <irrt/util.hpp>

13
nac3core/irrt/irrt_test.cpp
View File

@ -1,12 +1,25 @@
// This file will be compiled like a real C++ program,
// and we do have the luxury to use the standard libraries.
// That is if the nix flakes do not have issues... especially on msys2...
#include <cstdint>
#include <cstdio>
#include <cstdlib>
// Special macro to inform `#include <irrt/*>` that we are testing.
#define IRRT_TESTING
// Note that failure unit tests are not supported.
#include <test/test_core.hpp>
#include <test/test_ndarray_basic.hpp>
#include <test/test_ndarray_broadcast.hpp>
#include <test/test_ndarray_indexing.hpp>
int main() {
test::core::run();
test::ndarray_basic::run();
test::ndarray_indexing::run();
test::ndarray_broadcast::run();
return 0;
}

4
nac3core/irrt/test/test_core.hpp
View File

@ -7,8 +7,8 @@ namespace core {
void test_int_exp() {
BEGIN_TEST();
assert_values_match(125, __nac3_int_exp_impl<int32_t>(5, 3));
assert_values_match(3125, __nac3_int_exp_impl<int32_t>(5, 5));
assert_values_match(125L, __nac3_int_exp_impl<int64_t>(5, 3));
assert_values_match(3125L, __nac3_int_exp_impl<int64_t>(5, 5));
}
void run() { test_int_exp(); }

30
nac3core/irrt/test/test_ndarray_basic.hpp Normal file
View File

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

127
nac3core/irrt/test/test_ndarray_broadcast.hpp Normal file
View File

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

165
nac3core/irrt/test/test_ndarray_indexing.hpp Normal file
View File

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

15
nac3core/irrt/test/util.hpp
View File

@ -6,6 +6,11 @@
template <class T>
void print_value(const T& value);
template <>
void print_value(const bool& value) {
printf("%s", value ? "true" : "false");
}
template <>
void print_value(const int8_t& value) {
printf("%d", value);
@ -16,6 +21,11 @@ void print_value(const int32_t& value) {
printf("%d", value);
}
template <>
void print_value(const int64_t& value) {
printf("%d", value);
}
template <>
void print_value(const uint8_t& value) {
printf("%u", value);
@ -26,6 +36,11 @@ void print_value(const uint32_t& value) {
printf("%u", value);
}
template <>
void print_value(const uint64_t& value) {
printf("%d", value);
}
template <>
void print_value(const float& value) {
printf("%f", value);

1916
nac3core/src/codegen/builtin_fns.rs
View File

File diff suppressed because it is too large Load Diff

651
nac3core/src/codegen/expr.rs
View File

@ -2,7 +2,7 @@ use crate::{
codegen::{
classes::{
ArrayLikeIndexer, ArrayLikeValue, ListType, ListValue, NDArrayValue, ProxyType,
ProxyValue, RangeValue, TypedArrayLikeAccessor, UntypedArrayLikeAccessor,
ProxyValue, RangeValue, UntypedArrayLikeAccessor,
},
concrete_type::{ConcreteFuncArg, ConcreteTypeEnum, ConcreteTypeStore},
gen_in_range_check, get_llvm_abi_type, get_llvm_type, get_va_count_arg_name,
@ -12,6 +12,7 @@ use crate::{
call_memcpy_generic,
},
need_sret, numpy,
object::ndarray::{scalar::split_scalar_or_ndarray, NDArrayObject, NDArrayOut},
stmt::{
gen_for_callback_incrementing, gen_if_callback, gen_if_else_expr_callback, gen_raise,
gen_var,
@ -19,11 +20,7 @@ use crate::{
CodeGenContext, CodeGenTask, CodeGenerator,
},
symbol_resolver::{SymbolValue, ValueEnum},
toplevel::{
helper::PrimDef,
numpy::{make_ndarray_ty, unpack_ndarray_var_tys},
DefinitionId, TopLevelDef,
},
toplevel::{helper::PrimDef, numpy::unpack_ndarray_var_tys, DefinitionId, TopLevelDef},
typecheck::{
magic_methods::{Binop, BinopVariant, HasOpInfo},
typedef::{FunSignature, FuncArg, Type, TypeEnum, TypeVarId, Unifier, VarMap},
@ -32,7 +29,10 @@ use crate::{
use inkwell::{
attributes::{Attribute, AttributeLoc},
types::{AnyType, BasicType, BasicTypeEnum},
values::{BasicValueEnum, CallSiteValue, FunctionValue, IntValue, PointerValue, StructValue},
values::{
BasicValue, BasicValueEnum, CallSiteValue, FunctionValue, IntValue, PointerValue,
StructValue,
},
AddressSpace, IntPredicate, OptimizationLevel,
};
use itertools::{chain, izip, Either, Itertools};
@ -43,6 +43,12 @@ use nac3parser::ast::{
use std::iter::{repeat, repeat_with};
use std::{collections::HashMap, convert::TryInto, iter::once, iter::zip};
use super::{
model::*,
object::ndarray::indexing::util::gen_ndarray_subscript_ndindexes,
structure::{CSlice, Exception, ExceptionId},
};
pub fn get_subst_key(
unifier: &mut Unifier,
obj: Option<Type>,
@ -298,24 +304,7 @@ impl<'ctx, 'a> CodeGenContext<'ctx, 'a> {
None
}
}
Constant::Str(v) => {
assert!(self.unifier.unioned(ty, self.primitives.str));
if let Some(v) = self.const_strings.get(v) {
Some(*v)
} else {
let str_ptr = self
.builder
.build_global_string_ptr(v, "const")
.map(|v| v.as_pointer_value().into())
.unwrap();
let size = generator.get_size_type(self.ctx).const_int(v.len() as u64, false);
let ty = self.get_llvm_type(generator, self.primitives.str);
let val =
ty.into_struct_type().const_named_struct(&[str_ptr, size.into()]).into();
self.const_strings.insert(v.to_string(), val);
Some(val)
}
}
Constant::Str(s) => Some(self.gen_string(generator, s).value.into()),
Constant::Ellipsis => {
let msg = self.gen_string(generator, "NotImplementedError");
@ -582,71 +571,70 @@ impl<'ctx, 'a> CodeGenContext<'ctx, 'a> {
}
/// Helper function for generating a LLVM variable storing a [String].
pub fn gen_string<G, S>(&mut self, generator: &mut G, s: S) -> StructValue<'ctx>
pub fn gen_string<G>(&mut self, generator: &mut G, string: &str) -> Struct<'ctx, CSlice>
where
G: CodeGenerator + ?Sized,
S: Into<String>,
{
self.gen_const(generator, &Constant::Str(s.into()), self.primitives.str)
.map(BasicValueEnum::into_struct_value)
.unwrap()
self.const_strings.get(string).copied().unwrap_or_else(|| {
let sizet_model = IntModel(SizeT);
let pbyte_model = PtrModel(IntModel(Byte));
let cslice_model = StructModel(CSlice);
let base = self.builder.build_global_string_ptr(string, "constant_string").unwrap();
let base = pbyte_model.believe_value(base.as_pointer_value());
let len = sizet_model.constant(generator, self.ctx, string.len() as u64);
let cslice = cslice_model.create_const(generator, self.ctx, base, len);
self.const_strings.insert(string.to_owned(), cslice);
cslice
})
}
pub fn raise_exn<G: CodeGenerator + ?Sized>(
&mut self,
generator: &mut G,
name: &str,
msg: BasicValueEnum<'ctx>,
params: [Option<IntValue<'ctx>>; 3],
msg: Struct<'ctx, CSlice>,
params: [Option<Int<'ctx, Int64>>; 3],
loc: Location,
) {
let zelf = if let Some(exception_val) = self.exception_val {
exception_val
} else {
let ty = self.get_llvm_type(generator, self.primitives.exception).into_pointer_type();
let zelf_ty: BasicTypeEnum = ty.get_element_type().into_struct_type().into();
let zelf = generator.gen_var_alloc(self, zelf_ty, Some("exn")).unwrap();
*self.exception_val.insert(zelf)
};
let int32 = self.ctx.i32_type();
let zero = int32.const_zero();
unsafe {
let id_ptr = self.builder.build_in_bounds_gep(zelf, &[zero, zero], "exn.id").unwrap();
let id = self.resolver.get_string_id(name);
self.builder.build_store(id_ptr, int32.const_int(id as u64, false)).unwrap();
let ptr = self
.builder
.build_in_bounds_gep(zelf, &[zero, int32.const_int(5, false)], "exn.msg")
.unwrap();
self.builder.build_store(ptr, msg).unwrap();
let i64_zero = self.ctx.i64_type().const_zero();
for (i, attr_ind) in [6, 7, 8].iter().enumerate() {
let ptr = self
.builder
.build_in_bounds_gep(
zelf,
&[zero, int32.const_int(*attr_ind, false)],
"exn.param",
)
.unwrap();
let val = params[i].map_or(i64_zero, |v| {
self.builder.build_int_s_extend(v, self.ctx.i64_type(), "sext").unwrap()
});
self.builder.build_store(ptr, val).unwrap();
let exn_model = StructModel(Exception);
let exn_id_model = IntModel(ExceptionId::default());
let exn_id =
exn_id_model.constant(generator, self.ctx, self.resolver.get_string_id(name) as u64);
let exn = self.exception_val.unwrap_or_else(|| {
let exn = exn_model.var_alloca(generator, self, Some("exn")).unwrap();
*self.exception_val.insert(exn)
});
exn.set(self, |f| f.id, exn_id);
exn.set(self, |f| f.msg, msg);
for (i, param) in params.iter().enumerate() {
if let Some(param) = param {
exn.set(self, |f| f.params[i], *param);
}
}
gen_raise(generator, self, Some(&zelf.into()), loc);
gen_raise(generator, self, Some(exn), loc);
}
pub fn make_assert<G: CodeGenerator + ?Sized>(
&mut self,
generator: &mut G,
cond: IntValue<'ctx>,
cond: IntValue<'ctx>, // IntType can have arbitrary bit width
err_name: &str,
err_msg: &str,
params: [Option<IntValue<'ctx>>; 3],
loc: Location,
) {
let param_model = IntModel(Int64);
let params =
params.map(|p| p.map(|p| param_model.check_value(generator, self.ctx, p).unwrap()));
let err_msg = self.gen_string(generator, err_msg);
self.make_assert_impl(generator, cond, err_name, err_msg.into(), params, loc);
}
@ -654,26 +642,36 @@ impl<'ctx, 'a> CodeGenContext<'ctx, 'a> {
pub fn make_assert_impl<G: CodeGenerator + ?Sized>(
&mut self,
generator: &mut G,
cond: IntValue<'ctx>,
cond: IntValue<'ctx>, // IntType can have arbitrary bit width
err_name: &str,
err_msg: BasicValueEnum<'ctx>,
params: [Option<IntValue<'ctx>>; 3],
err_msg: Struct<'ctx, CSlice>,
params: [Option<Int<'ctx, Int64>>; 3],
loc: Location,
) {
let i1 = self.ctx.bool_type();
let i1_true = i1.const_all_ones();
// we assume that the condition is most probably true, so the normal path is the most
// probable path
// even if this assumption is violated, it does not matter as exception unwinding is
// slow anyway...
let cond = call_expect(self, cond, i1_true, Some("expect"));
let bool_model = IntModel(Bool);
// We assume that the condition is most probably true, so the normal path is the most
// probable path even if this assumption is violated, it does not matter as exception unwinding is.
let cond = call_expect(
self,
generator.bool_to_i1(self, cond),
bool_model.const_true(generator, self.ctx).value,
Some("expect"),
);
let current_bb = self.builder.get_insert_block().unwrap();
let current_fun = current_bb.get_parent().unwrap();
let then_block = self.ctx.insert_basic_block_after(current_bb, "succ");
let exn_block = self.ctx.append_basic_block(current_fun, "fail");
self.builder.build_conditional_branch(cond, then_block, exn_block).unwrap();
// Inserting into `exn_block`
self.builder.position_at_end(exn_block);
self.raise_exn(generator, err_name, err_msg, params, loc);
// Continuation
self.builder.position_at_end(then_block);
}
}
@ -1539,99 +1537,49 @@ pub fn gen_binop_expr_with_values<'ctx, G: CodeGenerator>(
} else if ty1.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id())
|| ty2.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id())
{
let llvm_usize = generator.get_size_type(ctx.ctx);
let left =
split_scalar_or_ndarray(generator, ctx, left_val, ty1).as_ndarray(generator, ctx);
let right =
split_scalar_or_ndarray(generator, ctx, right_val, ty2).as_ndarray(generator, ctx);
let is_ndarray1 = ty1.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id());
let is_ndarray2 = ty2.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id());
debug_assert!(ctx.unifier.unioned(left.dtype, right.dtype)); // Typechecker ensures this.
if is_ndarray1 && is_ndarray2 {
let (ndarray_dtype1, _) = unpack_ndarray_var_tys(&mut ctx.unifier, ty1);
let (ndarray_dtype2, _) = unpack_ndarray_var_tys(&mut ctx.unifier, ty2);
let common_dtype = left.dtype;
assert!(ctx.unifier.unioned(ndarray_dtype1, ndarray_dtype2));
let out = match op.variant {
BinopVariant::Normal => NDArrayOut::NewNDArray { dtype: common_dtype },
BinopVariant::AugAssign => NDArrayOut::WriteToNDArray { ndarray: left },
};
let left_val =
NDArrayValue::from_ptr_val(left_val.into_pointer_value(), llvm_usize, None);
let right_val =
NDArrayValue::from_ptr_val(right_val.into_pointer_value(), llvm_usize, None);
let res = if op.base == Operator::MatMult {
// MatMult is the only binop which is not an elementwise op
numpy::ndarray_matmul_2d(
generator,
ctx,
ndarray_dtype1,
match op.variant {
BinopVariant::Normal => None,
BinopVariant::AugAssign => Some(left_val),
},
left_val,
right_val,
)?
} else {
numpy::ndarray_elementwise_binop_impl(
generator,
ctx,
ndarray_dtype1,
match op.variant {
BinopVariant::Normal => None,
BinopVariant::AugAssign => Some(left_val),
},
(left_val.as_base_value().into(), false),
(right_val.as_base_value().into(), false),
|generator, ctx, (lhs, rhs)| {
gen_binop_expr_with_values(
generator,
ctx,
(&Some(ndarray_dtype1), lhs),
op,
(&Some(ndarray_dtype2), rhs),
ctx.current_loc,
)?
.unwrap()
.to_basic_value_enum(
ctx,
generator,
ndarray_dtype1,
)
},
)?
};
Ok(Some(res.as_base_value().into()))
if op.base == Operator::MatMult {
// Handle `left @ right`
let result = NDArrayObject::matmul(generator, ctx, left, right, out)
.split_unsized(generator, ctx)
.to_basic_value_enum();
Ok(Some(ValueEnum::Dynamic(result)))
} else {
let (ndarray_dtype, _) =
unpack_ndarray_var_tys(&mut ctx.unifier, if is_ndarray1 { ty1 } else { ty2 });
let ndarray_val = NDArrayValue::from_ptr_val(
if is_ndarray1 { left_val } else { right_val }.into_pointer_value(),
llvm_usize,
None,
);
let res = numpy::ndarray_elementwise_binop_impl(
// For other operators like +, -, etc...; do them element-wise-ly
let result = NDArrayObject::broadcasting_starmap(
generator,
ctx,
ndarray_dtype,
match op.variant {
BinopVariant::Normal => None,
BinopVariant::AugAssign => Some(ndarray_val),
},
(left_val, !is_ndarray1),
(right_val, !is_ndarray2),
|generator, ctx, (lhs, rhs)| {
&[left, right],
out,
|generator, ctx, _i, scalars| {
let left = scalars[0];
let right = scalars[1];
gen_binop_expr_with_values(
generator,
ctx,
(&Some(ndarray_dtype), lhs),
(&Some(left.dtype), left.value),
op,
(&Some(ndarray_dtype), rhs),
(&Some(right.dtype), right.value),
ctx.current_loc,
)?
.unwrap()
.to_basic_value_enum(ctx, generator, ndarray_dtype)
.to_basic_value_enum(ctx, generator, common_dtype)
},
)?;
Ok(Some(res.as_base_value().into()))
Ok(Some(ValueEnum::Dynamic(result.value.value.as_basic_value_enum())))
}
} else {
let left_ty_enum = ctx.unifier.get_ty_immutable(left_ty.unwrap());
@ -2266,338 +2214,6 @@ pub fn gen_cmpop_expr<'ctx, G: CodeGenerator>(
)
}
/// Generates code for a subscript expression on an `ndarray`.
///
/// * `ty` - The `Type` of the `NDArray` elements.
/// * `ndims` - The `Type` of the `NDArray` number-of-dimensions `Literal`.
/// * `v` - The `NDArray` value.
/// * `slice` - The slice expression used to subscript into the `ndarray`.
fn gen_ndarray_subscript_expr<'ctx, G: CodeGenerator>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
ty: Type,
ndims: Type,
v: NDArrayValue<'ctx>,
slice: &Expr<Option<Type>>,
) -> Result<Option<ValueEnum<'ctx>>, String> {
let llvm_i1 = ctx.ctx.bool_type();
let llvm_i32 = ctx.ctx.i32_type();
let llvm_usize = generator.get_size_type(ctx.ctx);
let TypeEnum::TLiteral { values, .. } = &*ctx.unifier.get_ty_immutable(ndims) else {
unreachable!()
};
let ndims = values
.iter()
.map(|ndim| u64::try_from(ndim.clone()).map_err(|()| ndim.clone()))
.collect::<Result<Vec<_>, _>>()
.map_err(|val| {
format!(
"Expected non-negative literal for ndarray.ndims, got {}",
i128::try_from(val).unwrap()
)
})?;
assert!(!ndims.is_empty());
// The number of dimensions subscripted by the index expression.
// Slicing a ndarray will yield the same number of dimensions, whereas indexing into a
// dimension will remove a dimension.
let subscripted_dims = match &slice.node {
ExprKind::Tuple { elts, .. } => elts.iter().fold(0, |acc, value_subexpr| {
if let ExprKind::Slice { .. } = &value_subexpr.node {
acc
} else {
acc + 1
}
}),
ExprKind::Slice { .. } => 0,
_ => 1,
};
let ndarray_ndims_ty = ctx.unifier.get_fresh_literal(
ndims.iter().map(|v| SymbolValue::U64(v - subscripted_dims)).collect(),
None,
);
let ndarray_ty =
make_ndarray_ty(&mut ctx.unifier, &ctx.primitives, Some(ty), Some(ndarray_ndims_ty));
let llvm_pndarray_t = ctx.get_llvm_type(generator, ndarray_ty).into_pointer_type();
let llvm_ndarray_t = llvm_pndarray_t.get_element_type().into_struct_type();
let llvm_ndarray_data_t = ctx.get_llvm_type(generator, ty).as_basic_type_enum();
let sizeof_elem = llvm_ndarray_data_t.size_of().unwrap();
// Check that len is non-zero
let len = v.load_ndims(ctx);
ctx.make_assert(
generator,
ctx.builder.build_int_compare(IntPredicate::SGT, len, llvm_usize.const_zero(), "").unwrap(),
"0:IndexError",
"too many indices for array: array is {0}-dimensional but 1 were indexed",
[Some(len), None, None],
slice.location,
);
// Normalizes a possibly-negative index to its corresponding positive index
let normalize_index = |generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
index: IntValue<'ctx>,
dim: u64| {
gen_if_else_expr_callback(
generator,
ctx,
|_, ctx| {
Ok(ctx
.builder
.build_int_compare(IntPredicate::SGE, index, index.get_type().const_zero(), "")
.unwrap())
},
|_, _| Ok(Some(index)),
|generator, ctx| {
let llvm_i32 = ctx.ctx.i32_type();
let len = unsafe {
v.dim_sizes().get_typed_unchecked(
ctx,
generator,
&llvm_usize.const_int(dim, true),
None,
)
};
let index = ctx
.builder
.build_int_add(
len,
ctx.builder.build_int_s_extend(index, llvm_usize, "").unwrap(),
"",
)
.unwrap();
Ok(Some(ctx.builder.build_int_truncate(index, llvm_i32, "").unwrap()))
},
)
.map(|v| v.map(BasicValueEnum::into_int_value))
};
// Converts a slice expression into a slice-range tuple
let expr_to_slice = |generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
node: &ExprKind<Option<Type>>,
dim: u64| {
match node {
ExprKind::Constant { value: Constant::Int(v), .. } => {
let Some(index) =
normalize_index(generator, ctx, llvm_i32.const_int(*v as u64, true), dim)?
else {
return Ok(None);
};
Ok(Some((index, index, llvm_i32.const_int(1, true))))
}
ExprKind::Slice { lower, upper, step } => {
let dim_sz = unsafe {
v.dim_sizes().get_typed_unchecked(
ctx,
generator,
&llvm_usize.const_int(dim, false),
None,
)
};
handle_slice_indices(lower, upper, step, ctx, generator, dim_sz)
}
_ => {
let Some(index) = generator.gen_expr(ctx, slice)? else { return Ok(None) };
let index = index
.to_basic_value_enum(ctx, generator, slice.custom.unwrap())?
.into_int_value();
let Some(index) = normalize_index(generator, ctx, index, dim)? else {
return Ok(None);
};
Ok(Some((index, index, llvm_i32.const_int(1, true))))
}
}
};
let make_indices_arr = |generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>|
-> Result<_, String> {
Ok(if let ExprKind::Tuple { elts, .. } = &slice.node {
let llvm_int_ty = ctx.get_llvm_type(generator, elts[0].custom.unwrap());
let index_addr = generator.gen_array_var_alloc(
ctx,
llvm_int_ty,
llvm_usize.const_int(elts.len() as u64, false),
None,
)?;
for (i, elt) in elts.iter().enumerate() {
let Some(index) = generator.gen_expr(ctx, elt)? else {
return Ok(None);
};
let index = index
.to_basic_value_enum(ctx, generator, elt.custom.unwrap())?
.into_int_value();
let Some(index) = normalize_index(generator, ctx, index, 0)? else {
return Ok(None);
};
let store_ptr = unsafe {
index_addr.ptr_offset_unchecked(
ctx,
generator,
&llvm_usize.const_int(i as u64, false),
None,
)
};
ctx.builder.build_store(store_ptr, index).unwrap();
}
Some(index_addr)
} else if let Some(index) = generator.gen_expr(ctx, slice)? {
let llvm_int_ty = ctx.get_llvm_type(generator, slice.custom.unwrap());
let index_addr = generator.gen_array_var_alloc(
ctx,
llvm_int_ty,
llvm_usize.const_int(1u64, false),
None,
)?;
let index =
index.to_basic_value_enum(ctx, generator, slice.custom.unwrap())?.into_int_value();
let Some(index) = normalize_index(generator, ctx, index, 0)? else { return Ok(None) };
let store_ptr = unsafe {
index_addr.ptr_offset_unchecked(ctx, generator, &llvm_usize.const_zero(), None)
};
ctx.builder.build_store(store_ptr, index).unwrap();
Some(index_addr)
} else {
None
})
};
Ok(Some(if ndims.len() == 1 && ndims[0] - subscripted_dims == 0 {
let Some(index_addr) = make_indices_arr(generator, ctx)? else { return Ok(None) };
v.data().get(ctx, generator, &index_addr, None).into()
} else {
match &slice.node {
ExprKind::Tuple { elts, .. } => {
let slices = elts
.iter()
.enumerate()
.map(|(dim, elt)| expr_to_slice(generator, ctx, &elt.node, dim as u64))
.take_while_inclusive(|slice| slice.as_ref().is_ok_and(Option::is_some))
.collect::<Result<Vec<_>, _>>()?;
if slices.len() < elts.len() {
return Ok(None);
}
let slices = slices.into_iter().map(Option::unwrap).collect_vec();
numpy::ndarray_sliced_copy(generator, ctx, ty, v, &slices)?.as_base_value().into()
}
ExprKind::Slice { .. } => {
let Some(slice) = expr_to_slice(generator, ctx, &slice.node, 0)? else {
return Ok(None);
};
numpy::ndarray_sliced_copy(generator, ctx, ty, v, &[slice])?.as_base_value().into()
}
_ => {
// Accessing an element from a multi-dimensional `ndarray`
let Some(index_addr) = make_indices_arr(generator, ctx)? else { return Ok(None) };
// Create a new array, remove the top dimension from the dimension-size-list, and copy the
// elements over
let subscripted_ndarray =
generator.gen_var_alloc(ctx, llvm_ndarray_t.into(), None)?;
let ndarray = NDArrayValue::from_ptr_val(subscripted_ndarray, llvm_usize, None);
let num_dims = v.load_ndims(ctx);
ndarray.store_ndims(
ctx,
generator,
ctx.builder
.build_int_sub(num_dims, llvm_usize.const_int(1, false), "")
.unwrap(),
);
let ndarray_num_dims = ndarray.load_ndims(ctx);
ndarray.create_dim_sizes(ctx, llvm_usize, ndarray_num_dims);
let ndarray_num_dims = ctx
.builder
.build_int_z_extend_or_bit_cast(
ndarray.load_ndims(ctx),
llvm_usize.size_of().get_type(),
"",
)
.unwrap();
let v_dims_src_ptr = unsafe {
v.dim_sizes().ptr_offset_unchecked(
ctx,
generator,
&llvm_usize.const_int(1, false),
None,
)
};
call_memcpy_generic(
ctx,
ndarray.dim_sizes().base_ptr(ctx, generator),
v_dims_src_ptr,
ctx.builder
.build_int_mul(ndarray_num_dims, llvm_usize.size_of(), "")
.map(Into::into)
.unwrap(),
llvm_i1.const_zero(),
);
let ndarray_num_elems = call_ndarray_calc_size(
generator,
ctx,
&ndarray.dim_sizes().as_slice_value(ctx, generator),
(None, None),
);
let ndarray_num_elems = ctx
.builder
.build_int_z_extend_or_bit_cast(ndarray_num_elems, sizeof_elem.get_type(), "")
.unwrap();
ndarray.create_data(ctx, llvm_ndarray_data_t, ndarray_num_elems);
let v_data_src_ptr = v.data().ptr_offset(ctx, generator, &index_addr, None);
call_memcpy_generic(
ctx,
ndarray.data().base_ptr(ctx, generator),
v_data_src_ptr,
ctx.builder
.build_int_mul(
ndarray_num_elems,
llvm_ndarray_data_t.size_of().unwrap(),
"",
)
.map(Into::into)
.unwrap(),
llvm_i1.const_zero(),
);
ndarray.as_base_value().into()
}
}
}))
}
/// See [`CodeGenerator::gen_expr`].
pub fn gen_expr<'ctx, G: CodeGenerator>(
generator: &mut G,
@ -3237,18 +2853,22 @@ pub fn gen_expr<'ctx, G: CodeGenerator>(
v.data().get(ctx, generator, &index, None).into()
}
}
TypeEnum::TObj { obj_id, params, .. } if *obj_id == PrimDef::NDArray.id() => {
let (ty, ndims) = params.iter().map(|(_, ty)| ty).collect_tuple().unwrap();
let v = if let Some(v) = generator.gen_expr(ctx, value)? {
v.to_basic_value_enum(ctx, generator, value.custom.unwrap())?
.into_pointer_value()
} else {
TypeEnum::TObj { obj_id, .. } if *obj_id == PrimDef::NDArray.id() => {
let Some(ndarray) = generator.gen_expr(ctx, value)? else {
return Ok(None);
};
let v = NDArrayValue::from_ptr_val(v, usize, None);
return gen_ndarray_subscript_expr(generator, ctx, *ty, *ndims, v, slice);
let ndarray_ty = value.custom.unwrap();
let ndarray = ndarray.to_basic_value_enum(ctx, generator, ndarray_ty)?;
let ndarray =
NDArrayObject::from_value_and_type(generator, ctx, ndarray, ndarray_ty);
let indexes = gen_ndarray_subscript_ndindexes(generator, ctx, slice)?;
let result = ndarray
.index(generator, ctx, &indexes, "index_result")
.split_unsized(generator, ctx)
.to_basic_value_enum();
return Ok(Some(ValueEnum::Dynamic(result)));
}
TypeEnum::TTuple { .. } => {
let index: u32 =
@ -3291,3 +2911,44 @@ pub fn gen_expr<'ctx, G: CodeGenerator>(
_ => unimplemented!(),
}))
}
/// Generate LLVM IR for an [`ExprKind::Slice`]
#[allow(clippy::type_complexity)]
pub fn gen_slice<'ctx, G: CodeGenerator>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
lower: &Option<Box<Expr<Option<Type>>>>,
upper: &Option<Box<Expr<Option<Type>>>>,
step: &Option<Box<Expr<Option<Type>>>>,
) -> Result<
(
Option<Instance<'ctx, IntModel<Int32>>>,
Option<Instance<'ctx, IntModel<Int32>>>,
Option<Instance<'ctx, IntModel<Int32>>>,
),
String,
> {
let i32_model = IntModel(Int32); // TODO: Switch to usize
let mut help = |value_expr: &Option<Box<Expr<Option<Type>>>>| -> Result<_, String> {
Ok(match value_expr {
None => None,
Some(value_expr) => {
let value_expr = generator
.gen_expr(ctx, value_expr)?
.unwrap()
.to_basic_value_enum(ctx, generator, ctx.primitives.int32)?;
let value_expr = i32_model.check_value(generator, ctx.ctx, value_expr).unwrap();
Some(value_expr)
}
})
};
let lower = help(lower)?;
let upper = help(upper)?;
let step = help(step)?;
Ok((lower, upper, step))
}

328
nac3core/src/codegen/irrt/mod.rs
View File

@ -1,7 +1,13 @@
use crate::symbol_resolver::SymbolResolver;
use crate::typecheck::typedef::Type;
mod test;
pub mod util;
use super::model::*;
use super::object::ndarray::broadcast::ShapeEntry;
use super::object::ndarray::indexing::NDIndex;
use super::structure::{List, NDArray};
use super::{
classes::{
ArrayLikeIndexer, ArrayLikeValue, ArraySliceValue, ListValue, NDArrayValue,
@ -11,6 +17,7 @@ use super::{
};
use crate::codegen::classes::TypedArrayLikeAccessor;
use crate::codegen::stmt::gen_for_callback_incrementing;
use inkwell::values::BasicValue;
use inkwell::{
attributes::{Attribute, AttributeLoc},
context::Context,
@ -22,6 +29,8 @@ use inkwell::{
};
use itertools::Either;
use nac3parser::ast::Expr;
use util::function::CallFunction;
use util::get_sizet_dependent_function_name;
#[must_use]
pub fn load_irrt(ctx: &Context) -> Module {
@ -416,14 +425,27 @@ pub fn list_slice_assignment<'ctx, G: CodeGenerator + ?Sized>(
.unwrap();
let cond_1 = ctx.builder.build_and(dest_step_eq_one, src_slt_dest, "slice_cond_1").unwrap();
let cond = ctx.builder.build_or(src_eq_dest, cond_1, "slice_cond").unwrap();
ctx.make_assert(
generator,
cond,
"0:ValueError",
"attempt to assign sequence of size {0} to slice of size {1} with step size {2}",
[Some(src_slice_len), Some(dest_slice_len), Some(dest_idx.2)],
ctx.current_loc,
);
// TODO: Temporary fix. Rewrite `list_slice_assignment` later
// Exception params should have been i64
{
let param_model = IntModel(Int64);
let src_slice_len =
param_model.s_extend_or_bit_cast(generator, ctx, src_slice_len, "src_slice_len");
let dest_slice_len =
param_model.s_extend_or_bit_cast(generator, ctx, dest_slice_len, "dest_slice_len");
let dest_idx_2 = param_model.s_extend_or_bit_cast(generator, ctx, dest_idx.2, "dest_idx_2");
ctx.make_assert(
generator,
cond,
"0:ValueError",
"attempt to assign sequence of size {0} to slice of size {1} with step size {2}",
[Some(src_slice_len.value), Some(dest_slice_len.value), Some(dest_idx_2.value)],
ctx.current_loc,
);
}
let new_len = {
let args = vec![
@ -875,7 +897,7 @@ pub fn call_ndarray_calc_broadcast<'ctx, G: CodeGenerator + ?Sized>(
}
/// Generates a call to `__nac3_ndarray_calc_broadcast_idx`. Returns an [`ArrayAllocaValue`]
/// containing the indices used for accessing `array` corresponding to the index of the broadcasted
/// containing the indices used for accessing `array` corresponding to the index of the broadcast
/// array `broadcast_idx`.
pub fn call_ndarray_calc_broadcast_index<
'ctx,
@ -930,3 +952,291 @@ pub fn call_ndarray_calc_broadcast_index<
Box::new(|_, v| v.into()),
)
}
pub fn call_nac3_throw_dummy_error<G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &CodeGenContext<'_, '_>,
) {
let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_throw_dummy_error");
CallFunction::begin(generator, ctx, &name).returning_void();
}
/// Initialize all global `EXN_*` exception IDs in IRRT with the [`SymbolResolver`].
pub fn setup_irrt_exceptions<'ctx>(
ctx: &'ctx Context,
module: &Module<'ctx>,
symbol_resolver: &dyn SymbolResolver,
) {
let exn_id_type = ctx.i32_type();
let errors = &[
("EXN_INDEX_ERROR", "0:IndexError"),
("EXN_VALUE_ERROR", "0:ValueError"),
("EXN_ASSERTION_ERROR", "0:AssertionError"),
("EXN_RUNTIME_ERROR", "0:RuntimeError"),
("EXN_TYPE_ERROR", "0:TypeError"),
];
for (irrt_name, symbol_name) in errors {
let exn_id = symbol_resolver.get_string_id(symbol_name);
let exn_id = exn_id_type.const_int(exn_id as u64, false).as_basic_value_enum();
let global = module.get_global(irrt_name).unwrap_or_else(|| {
panic!("Exception symbol name '{irrt_name}' should exist in the IRRT LLVM module")
});
global.set_initializer(&exn_id);
}
}
pub fn call_nac3_ndarray_util_assert_shape_no_negative<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
ndims: Int<'ctx, SizeT>,
shape: Ptr<'ctx, IntModel<SizeT>>,
) {
let name = get_sizet_dependent_function_name(
generator,
ctx,
"__nac3_ndarray_util_assert_shape_no_negative",
);
CallFunction::begin(generator, ctx, &name).arg(ndims).arg(shape).returning_void();
}
pub fn call_nac3_ndarray_util_assert_output_shape_same<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
ndarray_ndims: Int<'ctx, SizeT>,
ndarray_shape: Ptr<'ctx, IntModel<SizeT>>,
output_ndims: Int<'ctx, SizeT>,
output_shape: Ptr<'ctx, IntModel<SizeT>>,
) {
let name = get_sizet_dependent_function_name(
generator,
ctx,
"__nac3_ndarray_util_assert_output_shape_same",
);
CallFunction::begin(generator, ctx, &name)
.arg(ndarray_ndims)
.arg(ndarray_shape)
.arg(output_ndims)
.arg(output_shape)
.returning_void();
}
pub fn call_nac3_ndarray_size<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
pndarray: Ptr<'ctx, StructModel<NDArray>>,
) -> Int<'ctx, SizeT> {
let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_size");
CallFunction::begin(generator, ctx, &name).arg(pndarray).returning_auto("size")
}
pub fn call_nac3_ndarray_nbytes<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
pndarray: Ptr<'ctx, StructModel<NDArray>>,
) -> Int<'ctx, SizeT> {
let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_nbytes");
CallFunction::begin(generator, ctx, &name).arg(pndarray).returning_auto("nbytes")
}
pub fn call_nac3_ndarray_len<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
pndarray: Ptr<'ctx, StructModel<NDArray>>,
) -> Int<'ctx, SizeT> {
let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_len");
CallFunction::begin(generator, ctx, &name).arg(pndarray).returning_auto("len")
}
pub fn call_nac3_ndarray_is_c_contiguous<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
ndarray_ptr: Ptr<'ctx, StructModel<NDArray>>,
) -> Int<'ctx, Bool> {
let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_is_c_contiguous");
CallFunction::begin(generator, ctx, &name).arg(ndarray_ptr).returning_auto("is_c_contiguous")
}
pub fn call_nac3_ndarray_get_nth_pelement<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
pndarray: Ptr<'ctx, StructModel<NDArray>>,
index: Int<'ctx, SizeT>,
) -> Ptr<'ctx, IntModel<Byte>> {
let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_get_nth_pelement");
CallFunction::begin(generator, ctx, &name).arg(pndarray).arg(index).returning_auto("pelement")
}
pub fn call_nac3_ndarray_set_strides_by_shape<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
pdnarray: Ptr<'ctx, StructModel<NDArray>>,
) {
let name =
get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_set_strides_by_shape");
CallFunction::begin(generator, ctx, &name).arg(pdnarray).returning_void();
}
pub fn call_nac3_ndarray_copy_data<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
src_ndarray: Ptr<'ctx, StructModel<NDArray>>,
dst_ndarray: Ptr<'ctx, StructModel<NDArray>>,
) {
let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_copy_data");
CallFunction::begin(generator, ctx, &name).arg(src_ndarray).arg(dst_ndarray).returning_void();
}
pub fn call_nac3_ndarray_index<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
num_indexes: Int<'ctx, SizeT>,
indexes: Ptr<'ctx, StructModel<NDIndex>>,
src_ndarray: Ptr<'ctx, StructModel<NDArray>>,
dst_ndarray: Ptr<'ctx, StructModel<NDArray>>,
) {
let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_index");
CallFunction::begin(generator, ctx, &name)
.arg(num_indexes)
.arg(indexes)
.arg(src_ndarray)
.arg(dst_ndarray)
.returning_void();
}
pub fn call_nac3_ndarray_broadcast_to<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
src_ndarray: Ptr<'ctx, StructModel<NDArray>>,
dst_ndarray: Ptr<'ctx, StructModel<NDArray>>,
) {
let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_broadcast_to");
CallFunction::begin(generator, ctx, &name).arg(src_ndarray).arg(dst_ndarray).returning_void();
}
pub fn call_nac3_ndarray_broadcast_shapes<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
num_shape_entries: Int<'ctx, SizeT>,
shape_entries: Ptr<'ctx, StructModel<ShapeEntry>>,
dst_ndims: Int<'ctx, SizeT>,
dst_shape: Ptr<'ctx, IntModel<SizeT>>,
) {
let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_broadcast_shapes");
CallFunction::begin(generator, ctx, &name)
.arg(num_shape_entries)
.arg(shape_entries)
.arg(dst_ndims)
.arg(dst_shape)
.returning_void();
}
pub fn call_nac3_ndarray_resolve_and_check_new_shape<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
size: Int<'ctx, SizeT>,
new_ndims: Int<'ctx, SizeT>,
new_shape: Ptr<'ctx, IntModel<SizeT>>,
) {
let name = get_sizet_dependent_function_name(
generator,
ctx,
"__nac3_ndarray_resolve_and_check_new_shape",
);
CallFunction::begin(generator, ctx, &name)
.arg(size)
.arg(new_ndims)
.arg(new_shape)
.returning_void();
}
pub fn call_nac3_ndarray_transpose<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
src_ndarray: Ptr<'ctx, StructModel<NDArray>>,
dst_ndarray: Ptr<'ctx, StructModel<NDArray>>,
num_axes: Int<'ctx, SizeT>,
axes: Ptr<'ctx, IntModel<SizeT>>,
) {
let name = get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_transpose");
CallFunction::begin(generator, ctx, &name)
.arg(src_ndarray)
.arg(dst_ndarray)
.arg(num_axes)
.arg(axes)
.returning_void();
}
#[allow(clippy::too_many_arguments)]
pub fn call_nac3_ndarray_matmul_calculate_shapes<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
a_ndims: Int<'ctx, SizeT>,
a_shape: Ptr<'ctx, IntModel<SizeT>>,
b_ndims: Int<'ctx, SizeT>,
b_shape: Ptr<'ctx, IntModel<SizeT>>,
final_ndims: Int<'ctx, SizeT>,
new_a_shape: Ptr<'ctx, IntModel<SizeT>>,
new_b_shape: Ptr<'ctx, IntModel<SizeT>>,
dst_shape: Ptr<'ctx, IntModel<SizeT>>,
) {
let name =
get_sizet_dependent_function_name(generator, ctx, "__nac3_ndarray_matmul_calculate_shapes");
CallFunction::begin(generator, ctx, &name)
.arg(a_ndims)
.arg(a_shape)
.arg(b_ndims)
.arg(b_shape)
.arg(final_ndims)
.arg(new_a_shape)
.arg(new_b_shape)
.arg(dst_shape)
.returning_void();
}
pub fn call_nac3_ndarray_float64_matmul_at_least_2d<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
a_ndarray: Ptr<'ctx, StructModel<NDArray>>,
b_ndarray: Ptr<'ctx, StructModel<NDArray>>,
dst_ndarray: Ptr<'ctx, StructModel<NDArray>>,
) {
let name = get_sizet_dependent_function_name(
generator,
ctx,
"__nac3_ndarray_float64_matmul_at_least_2d",
);
CallFunction::begin(generator, ctx, &name)
.arg(a_ndarray)
.arg(b_ndarray)
.arg(dst_ndarray)
.returning_void();
}
pub fn call_nac3_array_set_and_validate_list_shape<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
list: Ptr<'ctx, StructModel<List<IntModel<Byte>>>>,
ndims: Int<'ctx, SizeT>,
shape: Ptr<'ctx, IntModel<SizeT>>,
) {
let name = get_sizet_dependent_function_name(
generator,
ctx,
"__nac3_array_set_and_validate_list_shape",
);
CallFunction::begin(generator, ctx, &name).arg(list).arg(ndims).arg(shape).returning_void();
}
pub fn call_nac3_array_write_list_to_array<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
list: Ptr<'ctx, StructModel<List<IntModel<Byte>>>>,
ndarray: Ptr<'ctx, StructModel<NDArray>>,
) {
let name =
get_sizet_dependent_function_name(generator, ctx, "__nac3_array_write_list_to_array");
CallFunction::begin(generator, ctx, &name).arg(list).arg(ndarray).returning_void();
}

107
nac3core/src/codegen/irrt/util.rs Normal file
View File

@ -0,0 +1,107 @@
use crate::codegen::{CodeGenContext, CodeGenerator};
// 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 mod function {
use crate::codegen::{model::*, CodeGenContext, CodeGenerator};
use inkwell::{
types::{BasicMetadataTypeEnum, BasicType, FunctionType},
values::{AnyValue, BasicMetadataValueEnum, BasicValue, BasicValueEnum, CallSiteValue},
};
use itertools::Itertools;
#[derive(Debug, Clone, Copy)]
struct Arg<'ctx> {
ty: BasicMetadataTypeEnum<'ctx>,
val: BasicMetadataValueEnum<'ctx>,
}
/// Helper structure to reduce IRRT Inkwell function call boilerplate
pub struct CallFunction<'ctx, 'a, 'b, 'c, 'd, G: CodeGenerator + ?Sized> {
generator: &'d mut G,
ctx: &'b CodeGenContext<'ctx, 'a>,
/// Function name
name: &'c str,
/// Call arguments
args: Vec<Arg<'ctx>>,
}
impl<'ctx, 'a, 'b, 'c, 'd, G: CodeGenerator + ?Sized> CallFunction<'ctx, 'a, 'b, 'c, 'd, G> {
pub fn begin(
generator: &'d mut G,
ctx: &'b CodeGenContext<'ctx, 'a>,
name: &'c str,
) -> Self {
CallFunction { generator, ctx, name, args: Vec::new() }
}
/// Push a call argument to the function call.
#[allow(clippy::needless_pass_by_value)]
#[must_use]
pub fn arg<M: Model<'ctx>>(mut self, arg: Instance<'ctx, M>) -> Self {
let arg = Arg {
ty: arg.model.get_type(self.generator, self.ctx.ctx).as_basic_type_enum().into(),
val: arg.value.as_basic_value_enum().into(),
};
self.args.push(arg);
self
}
/// Call the function and expect the function to return a value of type of `return_model`.
#[must_use]
pub fn returning<M: Model<'ctx>>(self, name: &str, return_model: M) -> Instance<'ctx, M> {
let ret_ty = return_model.get_type(self.generator, self.ctx.ctx);
let ret = self.get_function(|tys| ret_ty.fn_type(tys, false), name);
let ret = BasicValueEnum::try_from(ret.as_any_value_enum()).unwrap(); // Must work
let ret = return_model.check_value(self.generator, self.ctx.ctx, ret).unwrap(); // Must work
ret
}
/// Like [`CallFunction::returning_`] but `return_model` is automatically inferred.
#[must_use]
pub fn returning_auto<M: Model<'ctx> + Default>(self, name: &str) -> Instance<'ctx, M> {
self.returning(name, M::default())
}
/// Call the function and expect the function to return a void-type.
pub fn returning_void(self) {
let ret_ty = self.ctx.ctx.void_type();
let _ = self.get_function(|tys| ret_ty.fn_type(tys, false), "");
}
fn get_function<F>(&self, make_fn_type: F, return_value_name: &str) -> CallSiteValue<'ctx>
where
F: FnOnce(&[BasicMetadataTypeEnum<'ctx>]) -> FunctionType<'ctx>,
{
// Get the LLVM function, declare the function if it doesn't exist - it will be defined by other
// components of NAC3.
let func = self.ctx.module.get_function(self.name).unwrap_or_else(|| {
let tys = self.args.iter().map(|arg| arg.ty).collect_vec();
let fn_type = make_fn_type(&tys);
self.ctx.module.add_function(self.name, fn_type, None)
});
let vals = self.args.iter().map(|arg| arg.val).collect_vec();
self.ctx.builder.build_call(func, &vals, return_value_name).unwrap()
}
}
}

58
nac3core/src/codegen/mod.rs
View File

@ -1,7 +1,7 @@
use crate::{
codegen::classes::{ListType, NDArrayType, ProxyType, RangeType},
codegen::classes::{ListType, ProxyType, RangeType},
symbol_resolver::{StaticValue, SymbolResolver},
toplevel::{helper::PrimDef, numpy::unpack_ndarray_var_tys, TopLevelContext, TopLevelDef},
toplevel::{helper::PrimDef, TopLevelContext, TopLevelDef},
typecheck::{
type_inferencer::{CodeLocation, PrimitiveStore},
typedef::{CallId, FuncArg, Type, TypeEnum, Unifier},
@ -24,6 +24,7 @@ use inkwell::{
AddressSpace, IntPredicate, OptimizationLevel,
};
use itertools::Itertools;
use model::*;
use nac3parser::ast::{Location, Stmt, StrRef};
use parking_lot::{Condvar, Mutex};
use std::collections::{HashMap, HashSet};
@ -32,6 +33,7 @@ use std::sync::{
Arc,
};
use std::thread;
use structure::{CSlice, Exception, NDArray};
pub mod builtin_fns;
pub mod classes;
@ -41,8 +43,12 @@ pub mod extern_fns;
mod generator;
pub mod irrt;
pub mod llvm_intrinsics;
pub mod model;
pub mod numpy;
pub mod numpy_new;
pub mod object;
pub mod stmt;
pub mod structure;
#[cfg(test)]
mod test;
@ -168,11 +174,11 @@ pub struct CodeGenContext<'ctx, 'a> {
pub registry: &'a WorkerRegistry,
/// Cache for constant strings.
pub const_strings: HashMap<String, BasicValueEnum<'ctx>>,
pub const_strings: HashMap<String, Struct<'ctx, CSlice>>,
/// [`BasicBlock`] containing all `alloca` statements for the current function.
pub init_bb: BasicBlock<'ctx>,
pub exception_val: Option<PointerValue<'ctx>>,
pub exception_val: Option<Ptr<'ctx, StructModel<Exception>>>,
/// The header and exit basic blocks of a loop in this context. See
/// <https://llvm.org/docs/LoopTerminology.html> for explanation of these terminology.
@ -489,12 +495,8 @@ fn get_llvm_type<'ctx, G: CodeGenerator + ?Sized>(
}
TObj { obj_id, .. } if *obj_id == PrimDef::NDArray.id() => {
let (dtype, _) = unpack_ndarray_var_tys(unifier, ty);
let element_type = get_llvm_type(
ctx, module, generator, unifier, top_level, type_cache, dtype,
);
NDArrayType::new(generator, ctx, element_type).as_base_type().into()
let pndarray_model = PtrModel(StructModel(NDArray));
pndarray_model.get_type(generator, ctx).as_basic_type_enum()
}
_ => unreachable!(
@ -700,43 +702,19 @@ pub fn gen_func_impl<
..primitives
};
let mut type_cache: HashMap<_, _> = [
let cslice_model = StructModel(CSlice);
let pexn_model = PtrModel(StructModel(Exception));
let mut type_cache: HashMap<_, BasicTypeEnum<'ctx>> = [
(primitives.int32, context.i32_type().into()),
(primitives.int64, context.i64_type().into()),
(primitives.uint32, context.i32_type().into()),
(primitives.uint64, context.i64_type().into()),
(primitives.float, context.f64_type().into()),
(primitives.bool, context.i8_type().into()),
(primitives.str, {
let name = "str";
match module.get_struct_type(name) {
None => {
let str_type = context.opaque_struct_type("str");
let fields = [
context.i8_type().ptr_type(AddressSpace::default()).into(),
generator.get_size_type(context).into(),
];
str_type.set_body(&fields, false);
str_type.into()
}
Some(t) => t.as_basic_type_enum(),
}
}),
(primitives.str, cslice_model.get_type(generator, context).into()),
(primitives.range, RangeType::new(context).as_base_type().into()),
(primitives.exception, {
let name = "Exception";
if let Some(t) = module.get_struct_type(name) {
t.ptr_type(AddressSpace::default()).as_basic_type_enum()
} else {
let exception = context.opaque_struct_type("Exception");
let int32 = context.i32_type().into();
let int64 = context.i64_type().into();
let str_ty = module.get_struct_type("str").unwrap().as_basic_type_enum();
let fields = [int32, str_ty, int32, int32, str_ty, str_ty, int64, int64, int64];
exception.set_body(&fields, false);
exception.ptr_type(AddressSpace::default()).as_basic_type_enum()
}
}),
(primitives.exception, pexn_model.get_type(generator, context).into()),
]
.iter()
.copied()

40
nac3core/src/codegen/model/any.rs Normal file
View File

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

121
nac3core/src/codegen/model/core.rs Normal file
View File

@ -0,0 +1,121 @@
use std::fmt;
use inkwell::{context::Context, types::*, values::*};
use super::*;
use crate::codegen::{CodeGenContext, CodeGenerator};
#[derive(Debug, Clone)]
pub struct ModelError(pub String);
impl ModelError {
pub(super) fn under_context(mut self, context: &str) -> Self {
self.0.push_str(" ... in ");
self.0.push_str(context);
self
}
}
pub trait Model<'ctx>: fmt::Debug + Clone + Copy {
type Value: BasicValue<'ctx> + TryFrom<BasicValueEnum<'ctx>>;
type Type: BasicType<'ctx>;
/// Return the [`BasicType`] of this model.
fn get_type<G: CodeGenerator + ?Sized>(&self, generator: &G, ctx: &'ctx Context) -> Self::Type;
/// Check if a [`BasicType`] is the same type of this model.
fn check_type<T: BasicType<'ctx>, G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &'ctx Context,
ty: T,
) -> Result<(), ModelError>;
/// Create an instance from a value with [`Instance::model`] being this model.
///
/// Caller must make sure the type of `value` and the type of this `model` are equivalent.
fn believe_value(&self, value: Self::Value) -> Instance<'ctx, Self> {
Instance { model: *self, value }
}
/// Check if a [`BasicValue`]'s type is equivalent to the type of this model.
/// Wrap it into an [`Instance`] if it is.
fn check_value<V: BasicValue<'ctx>, G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &'ctx Context,
value: V,
) -> Result<Instance<'ctx, Self>, ModelError> {
let value = value.as_basic_value_enum();
self.check_type(generator, ctx, value.get_type())
.map_err(|err| err.under_context(format!("the value {value:?}").as_str()))?;
let Ok(value) = Self::Value::try_from(value) else {
unreachable!("check_type() has bad implementation")
};
Ok(self.believe_value(value))
}
// Allocate a value on the stack and return its pointer.
fn alloca<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
name: &str,
) -> Ptr<'ctx, Self> {
let pmodel = PtrModel(*self);
let p = ctx.builder.build_alloca(self.get_type(generator, ctx.ctx), name).unwrap();
pmodel.believe_value(p)
}
// Allocate an array on the stack and return its pointer.
fn array_alloca<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
len: IntValue<'ctx>,
name: &str,
) -> Ptr<'ctx, Self> {
let pmodel = PtrModel(*self);
let p =
ctx.builder.build_array_alloca(self.get_type(generator, ctx.ctx), len, name).unwrap();
pmodel.believe_value(p)
}
fn var_alloca<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
name: Option<&str>,
) -> Result<Ptr<'ctx, Self>, String> {
let pmodel = PtrModel(*self);
let ty = self.get_type(generator, ctx.ctx).as_basic_type_enum();
let p = generator.gen_var_alloc(ctx, ty, name)?;
Ok(pmodel.believe_value(p))
}
fn array_var_alloca<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
len: IntValue<'ctx>,
name: Option<&'ctx str>,
) -> Result<Ptr<'ctx, Self>, String> {
// TODO: Remove ArraySliceValue
let pmodel = PtrModel(*self);
let ty = self.get_type(generator, ctx.ctx).as_basic_type_enum();
let p = generator.gen_array_var_alloc(ctx, ty, len, name)?;
Ok(pmodel.believe_value(PointerValue::from(p)))
}
}
#[derive(Debug, Clone, Copy)]
pub struct Instance<'ctx, M: Model<'ctx>> {
/// The model of this instance.
pub model: M,
/// The value of this instance.
///
/// Caller must make sure the type of `value` and the type of this `model` are equivalent,
/// down to having the same [`IntType::get_bit_width`] in case of [`IntType`] for example.
pub value: M::Value,
}

276
nac3core/src/codegen/model/int.rs Normal file
View File

@ -0,0 +1,276 @@
use std::fmt;
use inkwell::{context::Context, types::IntType, values::IntValue, IntPredicate};
use crate::codegen::{CodeGenContext, CodeGenerator};
use super::*;
pub trait IntKind<'ctx>: fmt::Debug + Clone + Copy {
fn get_int_type<G: CodeGenerator + ?Sized>(
&self,
generator: &G,
ctx: &'ctx Context,
) -> IntType<'ctx>;
}
#[derive(Debug, Clone, Copy, Default)]
pub struct Bool;
#[derive(Debug, Clone, Copy, Default)]
pub struct Byte;
#[derive(Debug, Clone, Copy, Default)]
pub struct Int32;
#[derive(Debug, Clone, Copy, Default)]
pub struct Int64;
#[derive(Debug, Clone, Copy, Default)]
pub struct SizeT;
impl<'ctx> IntKind<'ctx> for Bool {
fn get_int_type<G: CodeGenerator + ?Sized>(
&self,
_generator: &G,
ctx: &'ctx Context,
) -> IntType<'ctx> {
ctx.bool_type()
}
}
impl<'ctx> IntKind<'ctx> for Byte {
fn get_int_type<G: CodeGenerator + ?Sized>(
&self,
_generator: &G,
ctx: &'ctx Context,
) -> IntType<'ctx> {
ctx.i8_type()
}
}
impl<'ctx> IntKind<'ctx> for Int32 {
fn get_int_type<G: CodeGenerator + ?Sized>(
&self,
_generator: &G,
ctx: &'ctx Context,
) -> IntType<'ctx> {
ctx.i32_type()
}
}
impl<'ctx> IntKind<'ctx> for Int64 {
fn get_int_type<G: CodeGenerator + ?Sized>(
&self,
_generator: &G,
ctx: &'ctx Context,
) -> IntType<'ctx> {
ctx.i64_type()
}
}
impl<'ctx> IntKind<'ctx> for SizeT {
fn get_int_type<G: CodeGenerator + ?Sized>(
&self,
generator: &G,
ctx: &'ctx Context,
) -> IntType<'ctx> {
generator.get_size_type(ctx)
}
}
#[derive(Debug, Clone, Copy)]
pub struct AnyInt<'ctx>(pub IntType<'ctx>);
impl<'ctx> IntKind<'ctx> for AnyInt<'ctx> {
fn get_int_type<G: CodeGenerator + ?Sized>(
&self,
_generator: &G,
_ctx: &'ctx Context,
) -> IntType<'ctx> {
self.0
}
}
#[derive(Debug, Clone, Copy, Default)]
pub struct IntModel<N>(pub N);
pub type Int<'ctx, N> = Instance<'ctx, IntModel<N>>;
impl<'ctx, N: IntKind<'ctx>> Model<'ctx> for IntModel<N> {
type Value = IntValue<'ctx>;
type Type = IntType<'ctx>;
#[must_use]
fn get_type<G: CodeGenerator + ?Sized>(&self, generator: &G, ctx: &'ctx Context) -> Self::Type {
self.0.get_int_type(generator, ctx)
}
fn check_type<T: inkwell::types::BasicType<'ctx>, G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &'ctx Context,
ty: T,
) -> Result<(), ModelError> {
let ty = ty.as_basic_type_enum();
let Ok(ty) = IntType::try_from(ty) else {
return Err(ModelError(format!("Expecting IntType, but got {ty:?}")));
};
let exp_ty = self.0.get_int_type(generator, ctx);
if ty.get_bit_width() != exp_ty.get_bit_width() {
return Err(ModelError(format!(
"Expecting IntType to have {} bit(s), but got {} bit(s)",
exp_ty.get_bit_width(),
ty.get_bit_width()
)));
}
Ok(())
}
}
impl<'ctx, N: IntKind<'ctx>> IntModel<N> {
pub fn constant<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &'ctx Context,
value: u64,
) -> Int<'ctx, N> {
let value = self.get_type(generator, ctx).const_int(value, false);
self.believe_value(value)
}
pub fn const_0<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &'ctx Context,
) -> Int<'ctx, N> {
self.constant(generator, ctx, 0)
}
pub fn const_1<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &'ctx Context,
) -> Int<'ctx, N> {
self.constant(generator, ctx, 1)
}
pub fn const_all_1s<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &'ctx Context,
) -> Int<'ctx, N> {
let value = self.get_type(generator, ctx).const_all_ones();
self.believe_value(value)
}
pub fn s_extend_or_bit_cast<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
value: IntValue<'ctx>,
name: &str,
) -> Int<'ctx, N> {
let value = ctx
.builder
.build_int_s_extend_or_bit_cast(value, self.get_type(generator, ctx.ctx), name)
.unwrap();
self.believe_value(value)
}
pub fn truncate<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
value: IntValue<'ctx>,
name: &str,
) -> Int<'ctx, N> {
let value =
ctx.builder.build_int_truncate(value, self.get_type(generator, ctx.ctx), name).unwrap();
self.believe_value(value)
}
}
impl IntModel<Bool> {
#[must_use]
pub fn const_false<'ctx, G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &'ctx Context,
) -> Int<'ctx, Bool> {
self.constant(generator, ctx, 0)
}
#[must_use]
pub fn const_true<'ctx, G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &'ctx Context,
) -> Int<'ctx, Bool> {
self.constant(generator, ctx, 1)
}
}
impl<'ctx, N: IntKind<'ctx>> Int<'ctx, N> {
pub fn s_extend_or_bit_cast<NewN: IntKind<'ctx>, G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
to_int_kind: NewN,
name: &str,
) -> Int<'ctx, NewN> {
IntModel(to_int_kind).s_extend_or_bit_cast(generator, ctx, self.value, name)
}
pub fn truncate<NewN: IntKind<'ctx>, G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
to_int_kind: NewN,
name: &str,
) -> Int<'ctx, NewN> {
IntModel(to_int_kind).truncate(generator, ctx, self.value, name)
}
#[must_use]
pub fn add(
&self,
ctx: &CodeGenContext<'ctx, '_>,
other: Int<'ctx, N>,
name: &str,
) -> Int<'ctx, N> {
let value = ctx.builder.build_int_add(self.value, other.value, name).unwrap();
self.model.believe_value(value)
}
#[must_use]
pub fn sub(
&self,
ctx: &CodeGenContext<'ctx, '_>,
other: Int<'ctx, N>,
name: &str,
) -> Int<'ctx, N> {
let value = ctx.builder.build_int_sub(self.value, other.value, name).unwrap();
self.model.believe_value(value)
}
#[must_use]
pub fn mul(
&self,
ctx: &CodeGenContext<'ctx, '_>,
other: Int<'ctx, N>,
name: &str,
) -> Int<'ctx, N> {
let value = ctx.builder.build_int_mul(self.value, other.value, name).unwrap();
self.model.believe_value(value)
}
pub fn compare(
&self,
ctx: &CodeGenContext<'ctx, '_>,
op: IntPredicate,
other: Int<'ctx, N>,
name: &str,
) -> Int<'ctx, Bool> {
let bool_model = IntModel(Bool);
let value = ctx.builder.build_int_compare(op, self.value, other.value, name).unwrap();
bool_model.believe_value(value)
}
}

12
nac3core/src/codegen/model/mod.rs Normal file
View File

@ -0,0 +1,12 @@
mod any;
mod core;
mod int;
mod ptr;
mod structure;
pub mod util;
pub use any::*;
pub use core::*;
pub use int::*;
pub use ptr::*;
pub use structure::*;

143
nac3core/src/codegen/model/ptr.rs Normal file
View File

@ -0,0 +1,143 @@
use inkwell::{
context::Context,
types::{BasicType, BasicTypeEnum, PointerType},
values::{IntValue, PointerValue},
AddressSpace,
};
use crate::codegen::{CodeGenContext, CodeGenerator};
use super::*;
#[derive(Debug, Clone, Copy, Default)]
pub struct PtrModel<Element>(pub Element);
pub type Ptr<'ctx, Element> = Instance<'ctx, PtrModel<Element>>;
impl<'ctx, Element: Model<'ctx>> Model<'ctx> for PtrModel<Element> {
type Value = PointerValue<'ctx>;
type Type = PointerType<'ctx>;
fn get_type<G: CodeGenerator + ?Sized>(&self, generator: &G, ctx: &'ctx Context) -> Self::Type {
self.0.get_type(generator, ctx).ptr_type(AddressSpace::default())
}
fn check_type<T: BasicType<'ctx>, G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &'ctx Context,
ty: T,
) -> Result<(), ModelError> {
let ty = ty.as_basic_type_enum();
let Ok(ty) = PointerType::try_from(ty) else {
return Err(ModelError(format!("Expecting PointerType, but got {ty:?}")));
};
let elem_ty = ty.get_element_type();
let Ok(elem_ty) = BasicTypeEnum::try_from(elem_ty) else {
return Err(ModelError(format!(
"Expecting pointer element type to be a BasicTypeEnum, but got {elem_ty:?}"
)));
};
// TODO: inkwell `get_element_type()` will be deprecated.
// Remove the check for `get_element_type()` when the time comes.
self.0
.check_type(generator, ctx, elem_ty)
.map_err(|err| err.under_context("a PointerType"))?;
Ok(())
}
}
impl<'ctx, Element: Model<'ctx>> PtrModel<Element> {
/// Return a ***constant*** nullptr.
pub fn nullptr<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &'ctx Context,
) -> Ptr<'ctx, Element> {
let ptr = self.get_type(generator, ctx).const_null();
self.believe_value(ptr)
}
/// Cast a pointer into this model with [`inkwell::builder::Builder::build_pointer_cast`]
pub fn pointer_cast<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
ptr: PointerValue<'ctx>,
name: &str,
) -> Ptr<'ctx, Element> {
let ptr =
ctx.builder.build_pointer_cast(ptr, self.get_type(generator, ctx.ctx), name).unwrap();
self.believe_value(ptr)
}
}
impl<'ctx, Element: Model<'ctx>> Ptr<'ctx, Element> {
/// Offset the pointer by [`inkwell::builder::Builder::build_in_bounds_gep`].
#[must_use]
pub fn offset<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
offset: IntValue<'ctx>,
name: &str,
) -> Ptr<'ctx, Element> {
let new_ptr =
unsafe { ctx.builder.build_in_bounds_gep(self.value, &[offset], name).unwrap() };
self.model.check_value(generator, ctx.ctx, new_ptr).unwrap()
}
// Load the `i`-th element (0-based) on the array with [`inkwell::builder::Builder::build_in_bounds_gep`].
pub fn ix<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
i: IntValue<'ctx>,
name: &str,
) -> Instance<'ctx, Element> {
self.offset(generator, ctx, i, name).load(generator, ctx, name)
}
/// Load the value with [`inkwell::builder::Builder::build_load`].
pub fn load<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
name: &str,
) -> Instance<'ctx, Element> {
let value = ctx.builder.build_load(self.value, name).unwrap();
self.model.0.check_value(generator, ctx.ctx, value).unwrap() // If unwrap() panics, there is a logic error.
}
/// Store a value with [`inkwell::builder::Builder::build_store`].
pub fn store(&self, ctx: &CodeGenContext<'ctx, '_>, value: Instance<'ctx, Element>) {
ctx.builder.build_store(self.value, value.value).unwrap();
}
/// Return a casted pointer of element type `NewElement` with [`inkwell::builder::Builder::build_pointer_cast`].
pub fn pointer_cast<NewElement: Model<'ctx>, G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
new_model: NewElement,
name: &str,
) -> Ptr<'ctx, NewElement> {
PtrModel(new_model).pointer_cast(generator, ctx, self.value, name)
}
/// Check if the pointer is null with [`inkwell::builder::Builder::build_is_null`].
pub fn is_null(&self, ctx: &CodeGenContext<'ctx, '_>, name: &str) -> Int<'ctx, Bool> {
let bool_model = IntModel(Bool);
let value = ctx.builder.build_is_null(self.value, name).unwrap();
bool_model.believe_value(value)
}
/// Check if the pointer is not null with [`inkwell::builder::Builder::build_is_not_null`].
pub fn is_not_null(&self, ctx: &CodeGenContext<'ctx, '_>, name: &str) -> Int<'ctx, Bool> {
let bool_model = IntModel(Bool);
let value = ctx.builder.build_is_not_null(self.value, name).unwrap();
bool_model.believe_value(value)
}
}

220
nac3core/src/codegen/model/structure.rs Normal file
View File

@ -0,0 +1,220 @@
use std::fmt;
use inkwell::{
context::Context,
types::{BasicType, BasicTypeEnum, StructType},
values::StructValue,
};
use crate::codegen::{CodeGenContext, CodeGenerator};
use super::*;
#[derive(Debug, Clone, Copy)]
pub struct GepField<M> {
pub gep_index: u64,
pub name: &'static str,
pub model: M,
}
pub trait FieldTraversal<'ctx> {
type Out<M>;
fn add<M: Model<'ctx>>(&mut self, name: &'static str, model: M) -> Self::Out<M>;
/// Like [`FieldTraversal::visit`] but [`Model`] is automatically inferred.
fn add_auto<M: Model<'ctx> + Default>(&mut self, name: &'static str) -> Self::Out<M> {
self.add(name, M::default())
}
}
pub struct GepFieldTraversal {
gep_index_counter: u64,
}
impl<'ctx> FieldTraversal<'ctx> for GepFieldTraversal {
type Out<M> = GepField<M>;
fn add<M: Model<'ctx>>(&mut self, name: &'static str, model: M) -> Self::Out<M> {
let gep_index = self.gep_index_counter;
self.gep_index_counter += 1;
Self::Out { gep_index, name, model }
}
}
struct TypeFieldTraversal<'ctx, 'a, G: CodeGenerator + ?Sized> {
generator: &'a G,
ctx: &'ctx Context,
field_types: Vec<BasicTypeEnum<'ctx>>,
}
impl<'ctx, 'a, G: CodeGenerator + ?Sized> FieldTraversal<'ctx> for TypeFieldTraversal<'ctx, 'a, G> {
type Out<M> = ();
fn add<M: Model<'ctx>>(&mut self, _name: &'static str, model: M) -> Self::Out<M> {
let t = model.get_type(self.generator, self.ctx).as_basic_type_enum();
self.field_types.push(t);
}
}
struct CheckTypeFieldTraversal<'ctx, 'a, G: CodeGenerator + ?Sized> {
generator: &'a mut G,
ctx: &'ctx Context,
index: u32,
scrutinee: StructType<'ctx>,
errors: Vec<ModelError>,
}
impl<'ctx, 'a, G: CodeGenerator + ?Sized> FieldTraversal<'ctx>
for CheckTypeFieldTraversal<'ctx, 'a, G>
{
type Out<M> = ();
fn add<M: Model<'ctx>>(&mut self, name: &'static str, model: M) -> Self::Out<M> {
let i = self.index;
self.index += 1;
if let Some(t) = self.scrutinee.get_field_type_at_index(i) {
if let Err(err) = model.check_type(self.generator, self.ctx, t) {
self.errors.push(err.under_context(format!("field #{i} '{name}'").as_str()));
}
} // Otherwise, it will be caught
}
}
pub trait StructKind<'ctx>: fmt::Debug + Clone + Copy {
type Fields<F: FieldTraversal<'ctx>>;
fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F>;
fn fields(&self) -> Self::Fields<GepFieldTraversal> {
self.traverse_fields(&mut GepFieldTraversal { gep_index_counter: 0 })
}
fn get_struct_type<G: CodeGenerator + ?Sized>(
&self,
generator: &G,
ctx: &'ctx Context,
) -> StructType<'ctx> {
let mut traversal = TypeFieldTraversal { generator, ctx, field_types: Vec::new() };
self.traverse_fields(&mut traversal);
ctx.struct_type(&traversal.field_types, false)
}
}
#[derive(Debug, Clone, Copy, Default)]
pub struct StructModel<S>(pub S);
pub type Struct<'ctx, S> = Instance<'ctx, StructModel<S>>;
impl<'ctx, S: StructKind<'ctx>> Model<'ctx> for StructModel<S> {
type Value = StructValue<'ctx>;
type Type = StructType<'ctx>;
fn get_type<G: CodeGenerator + ?Sized>(&self, generator: &G, ctx: &'ctx Context) -> Self::Type {
self.0.get_struct_type(generator, ctx)
}
fn check_type<T: BasicType<'ctx>, G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &'ctx Context,
ty: T,
) -> Result<(), ModelError> {
let ty = ty.as_basic_type_enum();
let Ok(ty) = StructType::try_from(ty) else {
return Err(ModelError(format!("Expecting StructType, but got {ty:?}")));
};
let mut traversal =
CheckTypeFieldTraversal { generator, ctx, index: 0, errors: Vec::new(), scrutinee: ty };
self.0.traverse_fields(&mut traversal);
let exp_num_fields = traversal.index;
let got_num_fields = u32::try_from(ty.get_field_types().len()).unwrap();
if exp_num_fields != got_num_fields {
return Err(ModelError(format!(
"Expecting StructType with {exp_num_fields} field(s), but got {got_num_fields}"
)));
}
if !traversal.errors.is_empty() {
return Err(traversal.errors[0].clone()); // TODO: Return other errors as well
}
Ok(())
}
}
impl<'ctx, S: StructKind<'ctx>> Struct<'ctx, S> {
pub fn get_field<G: CodeGenerator + ?Sized, M, GetField>(
&self,
generator: &mut G,
ctx: &'ctx Context,
get_field: GetField,
) -> Instance<'ctx, M>
where
M: Model<'ctx>,
GetField: FnOnce(S::Fields<GepFieldTraversal>) -> GepField<M>,
{
let field = get_field(self.model.0.fields());
let val = self.value.get_field_at_index(field.gep_index as u32).unwrap();
field.model.check_value(generator, ctx, val).unwrap()
}
}
impl<'ctx, S: StructKind<'ctx>> Ptr<'ctx, StructModel<S>> {
pub fn gep<M, GetField>(
&self,
ctx: &CodeGenContext<'ctx, '_>,
get_field: GetField,
) -> Ptr<'ctx, M>
where
M: Model<'ctx>,
GetField: FnOnce(S::Fields<GepFieldTraversal>) -> GepField<M>,
{
let field = get_field(self.model.0 .0.fields());
let llvm_i32 = ctx.ctx.i32_type(); // i64 would segfault
let ptr = unsafe {
ctx.builder
.build_in_bounds_gep(
self.value,
&[llvm_i32.const_zero(), llvm_i32.const_int(field.gep_index, false)],
field.name,
)
.unwrap()
};
let ptr_model = PtrModel(field.model);
ptr_model.believe_value(ptr)
}
/// Convenience function equivalent to `.gep(...).load(...)`.
pub fn get<M, GetField, G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
get_field: GetField,
name: &str,
) -> Instance<'ctx, M>
where
M: Model<'ctx>,
GetField: FnOnce(S::Fields<GepFieldTraversal>) -> GepField<M>,
{
self.gep(ctx, get_field).load(generator, ctx, name)
}
/// Convenience function equivalent to `.gep(...).store(...)`.
pub fn set<M, GetField>(
&self,
ctx: &CodeGenContext<'ctx, '_>,
get_field: GetField,
value: Instance<'ctx, M>,
) where
M: Model<'ctx>,
GetField: FnOnce(S::Fields<GepFieldTraversal>) -> GepField<M>,
{
self.gep(ctx, get_field).store(ctx, value);
}
}

91
nac3core/src/codegen/model/util.rs Normal file
View File

@ -0,0 +1,91 @@
use inkwell::{types::BasicType, values::IntValue};
/// `llvm.memcpy` but under the [`Model`] abstraction
use crate::codegen::{
llvm_intrinsics::call_memcpy_generic,
stmt::{gen_for_callback_incrementing, BreakContinueHooks},
CodeGenContext, CodeGenerator,
};
use super::*;
/// Convenience function.
///
/// Like [`call_memcpy_generic`] but with model abstractions and `is_volatile` set to `false`.
pub fn call_memcpy_model<'ctx, Item: Model<'ctx> + Default, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
dst_array: Ptr<'ctx, Item>,
src_array: Ptr<'ctx, Item>,
num_items: IntValue<'ctx>,
) {
let itemsize = Item::default().get_type(generator, ctx.ctx).size_of().unwrap();
let totalsize = ctx.builder.build_int_mul(itemsize, num_items, "totalsize").unwrap(); // TODO: Int types may not match.
let is_volatile = ctx.ctx.bool_type().const_zero();
call_memcpy_generic(ctx, dst_array.value, src_array.value, totalsize, is_volatile);
}
/// Like [`gen_for_callback_incrementing`] with [`Model`] abstractions.
/// The [`IntKind`] is automatically inferred.
pub fn gen_for_model_auto<'ctx, 'a, G, F, I>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, 'a>,
start: Int<'ctx, I>,
stop: Int<'ctx, I>,
step: Int<'ctx, I>,
body: F,
) -> Result<(), String>
where
G: CodeGenerator + ?Sized,
F: FnOnce(
&mut G,
&mut CodeGenContext<'ctx, 'a>,
BreakContinueHooks<'ctx>,
Int<'ctx, I>,
) -> Result<(), String>,
I: IntKind<'ctx> + Default,
{
let int_model = IntModel(I::default());
gen_for_callback_incrementing(
generator,
ctx,
None,
start.value,
(stop.value, false),
|g, ctx, hooks, i| {
let i = int_model.believe_value(i);
body(g, ctx, hooks, i)
},
step.value,
)
}
/// Like [`gen_if_callback`] with [`Model`] abstractions and without the `else` block.
pub fn gen_if_model<'ctx, 'a, G, ThenFn>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, 'a>,
cond: Int<'ctx, Bool>,
then: ThenFn,
) -> Result<(), String>
where
G: CodeGenerator + ?Sized,
ThenFn: FnOnce(&mut G, &mut CodeGenContext<'ctx, 'a>) -> Result<(), String>,
{
let current_bb = ctx.builder.get_insert_block().unwrap();
let then_bb = ctx.ctx.insert_basic_block_after(current_bb, "if.then");
let end_bb = ctx.ctx.insert_basic_block_after(then_bb, "if.end");
// Inserting into `current_bb`.
ctx.builder.build_conditional_branch(cond.value, then_bb, end_bb).unwrap();
// Inserting into `then_bb`
ctx.builder.position_at_end(then_bb);
then(generator, ctx)?;
ctx.builder.build_unconditional_branch(end_bb).unwrap();
// Reposition to `end_bb` for continuation.
ctx.builder.position_at_end(end_bb);
Ok(())
}

4
nac3core/src/codegen/numpy.rs
View File

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

522
nac3core/src/codegen/numpy_new.rs Normal file
View File

@ -0,0 +1,522 @@
// TODO: Replace numpy.rs
use inkwell::values::{BasicValue, BasicValueEnum};
use nac3parser::ast::StrRef;
use crate::{
codegen::object::{ndarray::scalar::split_scalar_or_ndarray, tuple::TupleObject},
symbol_resolver::ValueEnum,
toplevel::{
numpy::{extract_ndims, unpack_ndarray_var_tys},
DefinitionId,
},
typecheck::typedef::{FunSignature, Type},
};
use super::{
irrt::call_nac3_ndarray_util_assert_shape_no_negative,
model::*,
object::ndarray::{shape_util::parse_numpy_int_sequence, NDArrayObject},
CodeGenContext, CodeGenerator,
};
/// Get the zero value in `np.zeros()` of a `dtype`.
fn ndarray_zero_value<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
dtype: Type,
) -> BasicValueEnum<'ctx> {
if [ctx.primitives.int32, ctx.primitives.uint32]
.iter()
.any(|ty| ctx.unifier.unioned(dtype, *ty))
{
ctx.ctx.i32_type().const_zero().into()
} else if [ctx.primitives.int64, ctx.primitives.uint64]
.iter()
.any(|ty| ctx.unifier.unioned(dtype, *ty))
{
ctx.ctx.i64_type().const_zero().into()
} else if ctx.unifier.unioned(dtype, ctx.primitives.float) {
ctx.ctx.f64_type().const_zero().into()
} else if ctx.unifier.unioned(dtype, ctx.primitives.bool) {
ctx.ctx.bool_type().const_zero().into()
} else if ctx.unifier.unioned(dtype, ctx.primitives.str) {
ctx.gen_string(generator, "").value.into()
} else {
unreachable!()
}
}
/// 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, '_>,
elem_ty: Type,
) -> BasicValueEnum<'ctx> {
if [ctx.primitives.int32, ctx.primitives.uint32]
.iter()
.any(|ty| ctx.unifier.unioned(elem_ty, *ty))
{
let is_signed = ctx.unifier.unioned(elem_ty, 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(elem_ty, *ty))
{
let is_signed = ctx.unifier.unioned(elem_ty, ctx.primitives.int64);
ctx.ctx.i64_type().const_int(1, is_signed).into()
} else if ctx.unifier.unioned(elem_ty, ctx.primitives.float) {
ctx.ctx.f64_type().const_float(1.0).into()
} else if ctx.unifier.unioned(elem_ty, ctx.primitives.bool) {
ctx.ctx.bool_type().const_int(1, false).into()
} else if ctx.unifier.unioned(elem_ty, ctx.primitives.str) {
ctx.gen_string(generator, "1").value.into()
} else {
unreachable!()
}
}
/// Helper function to create an ndarray with uninitialized values.
///
/// * `ndarray_ty` - The [`Type`] of the ndarray
/// * `shape` - The user input shape argument
/// * `shape_ty` - The [`Type`] of the shape argument
///
/// This function does data validation the `shape` input.
fn create_empty_ndarray<'ctx, G>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
ndarray_ty: Type,
shape: BasicValueEnum<'ctx>,
shape_ty: Type,
) -> NDArrayObject<'ctx>
where
G: CodeGenerator + ?Sized,
{
let (_, shape) = parse_numpy_int_sequence(generator, ctx, shape, shape_ty);
let ndarray =
NDArrayObject::alloca_uninitialized_of_type(generator, ctx, ndarray_ty, "ndarray");
// Validate `shape`
let ndims = ndarray.get_ndims(generator, ctx.ctx);
call_nac3_ndarray_util_assert_shape_no_negative(generator, ctx, ndims, shape);
// Setup `ndarray` with `shape`
ndarray.copy_shape_from_array(generator, ctx, shape);
ndarray.create_data(generator, ctx); // `shape` has to be set
ndarray
}
/// Generates LLVM IR for `np.empty`.
pub fn gen_ndarray_empty<'ctx>(
ctx: &mut CodeGenContext<'ctx, '_>,
obj: &Option<(Type, ValueEnum<'ctx>)>,
fun: (&FunSignature, DefinitionId),
args: &[(Option<StrRef>, ValueEnum<'ctx>)],
generator: &mut dyn CodeGenerator,
) -> Result<BasicValueEnum<'ctx>, String> {
assert!(obj.is_none());
assert_eq!(args.len(), 1);
// Parse arguments
let shape_ty = fun.0.args[0].ty;
let shape = args[0].1.clone().to_basic_value_enum(ctx, generator, shape_ty)?;
// Implementation
let ndarray_ty = fun.0.ret;
let ndarray = create_empty_ndarray(generator, ctx, ndarray_ty, shape, shape_ty);
Ok(ndarray.value.value.as_basic_value_enum())
}
/// Generates LLVM IR for `np.zero`.
pub fn gen_ndarray_zeros<'ctx>(
ctx: &mut CodeGenContext<'ctx, '_>,
obj: &Option<(Type, ValueEnum<'ctx>)>,
fun: (&FunSignature, DefinitionId),
args: &[(Option<StrRef>, ValueEnum<'ctx>)],
generator: &mut dyn CodeGenerator,
) -> Result<BasicValueEnum<'ctx>, String> {
assert!(obj.is_none());
assert_eq!(args.len(), 1);
// Parse arguments
let shape_ty = fun.0.args[0].ty;
let shape = args[0].1.clone().to_basic_value_enum(ctx, generator, shape_ty)?;
// Implementation
let ndarray_ty = fun.0.ret;
let ndarray = create_empty_ndarray(generator, ctx, ndarray_ty, shape, shape_ty);
let fill_value = ndarray_zero_value(generator, ctx, ndarray.dtype);
ndarray.fill(generator, ctx, fill_value);
Ok(ndarray.value.value.as_basic_value_enum())
}
/// Generates LLVM IR for `np.ones`.
pub fn gen_ndarray_ones<'ctx>(
ctx: &mut CodeGenContext<'ctx, '_>,
obj: &Option<(Type, ValueEnum<'ctx>)>,
fun: (&FunSignature, DefinitionId),
args: &[(Option<StrRef>, ValueEnum<'ctx>)],
generator: &mut dyn CodeGenerator,
) -> Result<BasicValueEnum<'ctx>, String> {
assert!(obj.is_none());
assert_eq!(args.len(), 1);
// Parse arguments
let shape_ty = fun.0.args[0].ty;
let shape = args[0].1.clone().to_basic_value_enum(ctx, generator, shape_ty)?;
// Implementation
let ndarray_ty = fun.0.ret;
let ndarray = create_empty_ndarray(generator, ctx, ndarray_ty, shape, shape_ty);
let fill_value = ndarray_one_value(generator, ctx, ndarray.dtype);
ndarray.fill(generator, ctx, fill_value);
Ok(ndarray.value.value.as_basic_value_enum())
}
/// Generates LLVM IR for `np.full`.
pub fn gen_ndarray_full<'ctx>(
ctx: &mut CodeGenContext<'ctx, '_>,
obj: &Option<(Type, ValueEnum<'ctx>)>,
fun: (&FunSignature, DefinitionId),
args: &[(Option<StrRef>, ValueEnum<'ctx>)],
generator: &mut dyn CodeGenerator,
) -> Result<BasicValueEnum<'ctx>, String> {
assert!(obj.is_none());
assert_eq!(args.len(), 2);
// Parse argument #1 shape
let shape_ty = fun.0.args[0].ty;
let shape = args[0].1.clone().to_basic_value_enum(ctx, generator, shape_ty)?;
// Parse argument #2 fill_value
let fill_value_ty = fun.0.args[1].ty;
let fill_value = args[1].1.clone().to_basic_value_enum(ctx, generator, fill_value_ty)?;
// Implementation
let ndarray_ty = fun.0.ret;
let ndarray = create_empty_ndarray(generator, ctx, ndarray_ty, shape, shape_ty);
ndarray.fill(generator, ctx, fill_value);
Ok(ndarray.value.value.as_basic_value_enum())
}
/// Generates LLVM IR for `np.broadcast_to`.
pub fn gen_ndarray_broadcast_to<'ctx>(
ctx: &mut CodeGenContext<'ctx, '_>,
obj: &Option<(Type, ValueEnum<'ctx>)>,
fun: (&FunSignature, DefinitionId),
args: &[(Option<StrRef>, ValueEnum<'ctx>)],
generator: &mut dyn CodeGenerator,
) -> Result<BasicValueEnum<'ctx>, String> {
assert!(obj.is_none());
assert_eq!(args.len(), 2);
// Parse argument #1 input
let input_ty = fun.0.args[0].ty;
let input = args[0].1.clone().to_basic_value_enum(ctx, generator, input_ty)?;
// Parse argument #2 shape
let shape_ty = fun.0.args[1].ty;
let shape = args[1].1.clone().to_basic_value_enum(ctx, generator, shape_ty)?;
// Define models
let sizet_model = IntModel(SizeT);
// Extract broadcast_ndims, this is the only way to get the
// ndims of the ndarray result statically.
let (_, broadcast_ndims_ty) = unpack_ndarray_var_tys(&mut ctx.unifier, fun.0.ret);
let broadcast_ndims = extract_ndims(&ctx.unifier, broadcast_ndims_ty);
// Process `input`
let in_ndarray =
split_scalar_or_ndarray(generator, ctx, input, input_ty).as_ndarray(generator, ctx);
// Process `shape`
let (_, broadcast_shape) = parse_numpy_int_sequence(generator, ctx, shape, shape_ty);
// NOTE: shape.size should equal to `broadcasted_ndims`.
let broadcast_ndims_llvm = sizet_model.constant(generator, ctx.ctx, broadcast_ndims);
call_nac3_ndarray_util_assert_shape_no_negative(
generator,
ctx,
broadcast_ndims_llvm,
broadcast_shape,
);
// Create broadcast view
let broadcast_ndarray =
in_ndarray.broadcast_to(generator, ctx, broadcast_ndims, broadcast_shape);
Ok(broadcast_ndarray.value.value.as_basic_value_enum())
}
/// Generates LLVM IR for `np.reshape`.
pub fn gen_ndarray_reshape<'ctx>(
ctx: &mut CodeGenContext<'ctx, '_>,
obj: &Option<(Type, ValueEnum<'ctx>)>,
fun: (&FunSignature, DefinitionId),
args: &[(Option<StrRef>, ValueEnum<'ctx>)],
generator: &mut dyn CodeGenerator,
) -> Result<BasicValueEnum<'ctx>, String> {
assert!(obj.is_none());
assert_eq!(args.len(), 2);
// Parse argument #1 input
let input_ty = fun.0.args[0].ty;
let input = args[0].1.clone().to_basic_value_enum(ctx, generator, input_ty)?;
// Parse argument #2 shape
let shape_ty = fun.0.args[1].ty;
let shape = args[1].1.clone().to_basic_value_enum(ctx, generator, shape_ty)?;
// Extract reshaped_ndims
let (_, reshaped_ndims_ty) = unpack_ndarray_var_tys(&mut ctx.unifier, fun.0.ret);
let reshaped_ndims = extract_ndims(&ctx.unifier, reshaped_ndims_ty);
// Process `input`
let in_ndarray =
split_scalar_or_ndarray(generator, ctx, input, input_ty).as_ndarray(generator, ctx);
// Process the shape input from user and resolve negative indices.
// The resulting `new_shape`'s size should be equal to reshaped_ndims.
// This is ensured by the typechecker.
let (_, new_shape) = parse_numpy_int_sequence(generator, ctx, shape, shape_ty);
let reshaped_ndarray = in_ndarray.reshape_or_copy(generator, ctx, reshaped_ndims, new_shape);
Ok(reshaped_ndarray.value.value.as_basic_value_enum())
}
/// Generates LLVM IR for `np.arange`.
pub fn gen_ndarray_arange<'ctx>(
ctx: &mut CodeGenContext<'ctx, '_>,
obj: &Option<(Type, ValueEnum<'ctx>)>,
fun: (&FunSignature, DefinitionId),
args: &[(Option<StrRef>, ValueEnum<'ctx>)],
generator: &mut dyn CodeGenerator,
) -> Result<BasicValueEnum<'ctx>, String> {
assert!(obj.is_none());
assert_eq!(args.len(), 1);
// Parse argument #1 len
let input_ty = fun.0.args[0].ty;
let input = args[0].1.clone().to_basic_value_enum(ctx, generator, input_ty)?.into_int_value();
// Define models
let sizet_model = IntModel(SizeT);
// Process input
let input = sizet_model.s_extend_or_bit_cast(generator, ctx, input, "input_dim");
// Allocate the resulting ndarray
let ndarray = NDArrayObject::alloca_uninitialized(
generator,
ctx,
ctx.primitives.float,
1, // ndims = 1
"arange_ndarray",
);
// `ndarray.shape[0] = input`
let zero = sizet_model.const_0(generator, ctx.ctx);
ndarray
.value
.get(generator, ctx, |f| f.shape, "shape")
.offset(generator, ctx, zero.value, "dim")
.store(ctx, input);
// Create data and set elements
ndarray.create_data(generator, ctx);
ndarray.foreach_pointer(generator, ctx, |_generator, ctx, _hooks, i, pelement| {
let val =
ctx.builder.build_unsigned_int_to_float(i.value, ctx.ctx.f64_type(), "val").unwrap();
ctx.builder.build_store(pelement, val).unwrap();
Ok(())
})?;
Ok(ndarray.value.value.as_basic_value_enum())
}
/// Generates LLVM IR for `np.size`.
pub fn gen_ndarray_size<'ctx>(
ctx: &mut CodeGenContext<'ctx, '_>,
obj: &Option<(Type, ValueEnum<'ctx>)>,
fun: (&FunSignature, DefinitionId),
args: &[(Option<StrRef>, ValueEnum<'ctx>)],
generator: &mut dyn CodeGenerator,
) -> Result<BasicValueEnum<'ctx>, String> {
assert!(obj.is_none());
assert_eq!(args.len(), 1);
let ndarray_ty = fun.0.args[0].ty;
let ndarray = args[0].1.clone().to_basic_value_enum(ctx, generator, ndarray_ty)?;
let ndarray = NDArrayObject::from_value_and_type(generator, ctx, ndarray, ndarray_ty);
let size = ndarray.size(generator, ctx).truncate(generator, ctx, Int32, "size");
Ok(size.value.as_basic_value_enum())
}
/// Generates LLVM IR for `np.shape`.
pub fn gen_ndarray_shape<'ctx>(
ctx: &mut CodeGenContext<'ctx, '_>,
obj: &Option<(Type, ValueEnum<'ctx>)>,
fun: (&FunSignature, DefinitionId),
args: &[(Option<StrRef>, ValueEnum<'ctx>)],
generator: &mut dyn CodeGenerator,
) -> Result<BasicValueEnum<'ctx>, String> {
assert!(obj.is_none());
assert_eq!(args.len(), 1);
// Parse argument #1 ndarray
let ndarray_ty = fun.0.args[0].ty;
let ndarray = args[0].1.clone().to_basic_value_enum(ctx, generator, ndarray_ty)?;
// Define models
let sizet_model = IntModel(SizeT);
// Process ndarray
let ndarray = NDArrayObject::from_value_and_type(generator, ctx, ndarray, ndarray_ty);
let mut items = Vec::with_capacity(ndarray.ndims as usize);
for i in 0..ndarray.ndims {
let i = sizet_model.constant(generator, ctx.ctx, i);
let dim =
ndarray.value.get(generator, ctx, |f| f.shape, "").ix(generator, ctx, i.value, "dim");
let dim = dim.truncate(generator, ctx, Int32, "dim"); // TODO: keep using SizeT
items.push((dim.value.as_basic_value_enum(), ctx.primitives.int32));
}
let shape = TupleObject::create(generator, ctx, items, "shape");
Ok(shape.value.as_basic_value_enum())
}
/// Generates LLVM IR for `<ndarray>.strides`.
pub fn gen_ndarray_strides<'ctx>(
ctx: &mut CodeGenContext<'ctx, '_>,
obj: &Option<(Type, ValueEnum<'ctx>)>,
fun: (&FunSignature, DefinitionId),
args: &[(Option<StrRef>, ValueEnum<'ctx>)],
generator: &mut dyn CodeGenerator,
) -> Result<BasicValueEnum<'ctx>, String> {
// TODO: This function looks exactly like `gen_ndarray_shapes`, code duplication?
assert!(obj.is_none());
assert_eq!(args.len(), 1);
// Parse argument #1 ndarray
let ndarray_ty = fun.0.args[0].ty;
let ndarray = args[0].1.clone().to_basic_value_enum(ctx, generator, ndarray_ty)?;
// Define models
let sizet_model = IntModel(SizeT);
// Process ndarray
let ndarray = NDArrayObject::from_value_and_type(generator, ctx, ndarray, ndarray_ty);
let mut items = Vec::with_capacity(ndarray.ndims as usize);
for i in 0..ndarray.ndims {
let i = sizet_model.constant(generator, ctx.ctx, i);
let dim =
ndarray.value.get(generator, ctx, |f| f.strides, "").ix(generator, ctx, i.value, "dim");
let dim = dim.truncate(generator, ctx, Int32, "dim"); // TODO: keep using SizeT
items.push((dim.value.as_basic_value_enum(), ctx.primitives.int32));
}
let strides = TupleObject::create(generator, ctx, items, "strides");
Ok(strides.value.as_basic_value_enum())
}
/// Generates LLVM IR for `np.transpose`.
pub fn gen_ndarray_transpose<'ctx>(
ctx: &mut CodeGenContext<'ctx, '_>,
obj: &Option<(Type, ValueEnum<'ctx>)>,
fun: (&FunSignature, DefinitionId),
args: &[(Option<StrRef>, ValueEnum<'ctx>)],
generator: &mut dyn CodeGenerator,
) -> Result<BasicValueEnum<'ctx>, String> {
// TODO: The implementation will be changed once default values start working again.
// Read the comment on this function in BuiltinBuilder.
// TODO: Change axes values to `SizeT`
assert!(obj.is_none());
assert_eq!(args.len(), 1);
// Parse argument #1 ndarray
let ndarray_ty = fun.0.args[0].ty;
let ndarray = args[0].1.clone().to_basic_value_enum(ctx, generator, ndarray_ty)?;
// Implementation
let ndarray = NDArrayObject::from_value_and_type(generator, ctx, ndarray, ndarray_ty);
let has_axes = args.len() >= 2;
let transposed_ndarray = if has_axes {
// Parse argument #2 axes
let in_axes_ty = fun.0.args[1].ty;
let in_axes = args[1].1.clone().to_basic_value_enum(ctx, generator, in_axes_ty)?;
let (_, axes) = parse_numpy_int_sequence(generator, ctx, in_axes, in_axes_ty);
ndarray.transpose(generator, ctx, Some(axes))
} else {
ndarray.transpose(generator, ctx, None)
};
Ok(transposed_ndarray.value.value.as_basic_value_enum())
}
pub fn gen_ndarray_array<'ctx>(
ctx: &mut CodeGenContext<'ctx, '_>,
obj: &Option<(Type, ValueEnum<'ctx>)>,
fun: (&FunSignature, DefinitionId),
args: &[(Option<StrRef>, ValueEnum<'ctx>)],
generator: &mut dyn CodeGenerator,
) -> Result<BasicValueEnum<'ctx>, String> {
assert!(obj.is_none());
assert!(matches!(args.len(), 1..=3));
let obj_ty = fun.0.args[0].ty;
let obj_arg = args[0].1.clone().to_basic_value_enum(ctx, generator, obj_ty)?;
let copy_arg = if let Some(arg) =
args.iter().find(|arg| arg.0.is_some_and(|name| name == fun.0.args[1].name))
{
let copy_ty = fun.0.args[1].ty;
arg.1.clone().to_basic_value_enum(ctx, generator, copy_ty)?
} else {
ctx.gen_symbol_val(
generator,
fun.0.args[1].default_value.as_ref().unwrap(),
fun.0.args[1].ty,
)
};
// The argument `ndmin` is completely ignored. We don't need to know its LLVM value.
// We simply make the output ndarray's ndims correct with `atleast_nd`.
let (dtype, ndims) = unpack_ndarray_var_tys(&mut ctx.unifier, fun.0.ret);
let output_ndims = extract_ndims(&mut ctx.unifier, ndims);
let copy = IntModel(Byte).check_value(generator, ctx.ctx, copy_arg).unwrap(); // NAC3 booleans are i8
let copy = copy.truncate(generator, ctx, Bool, "copy_bool");
let ndarray = NDArrayObject::from_np_array(generator, ctx, obj_arg, obj_ty, copy);
debug_assert!(ndarray.ndims <= output_ndims); // Sanity check on `ndims`
let ndarray = ndarray.atleast_nd(generator, ctx, output_ndims);
debug_assert!(ctx.unifier.unioned(ndarray.dtype, dtype)); // Sanity check on `dtype`
Ok(ndarray.value.value.as_basic_value_enum())
}

79
nac3core/src/codegen/object/list.rs Normal file
View File

@ -0,0 +1,79 @@
use inkwell::values::BasicValue;
use crate::{
codegen::{model::*, structure::List, CodeGenContext, CodeGenerator},
typecheck::typedef::{iter_type_vars, Type, TypeEnum},
};
/// A NAC3 Python List object.
#[derive(Debug, Clone, Copy)]
pub struct ListObject<'ctx> {
/// Typechecker type of the list items
pub item_type: Type,
pub value: Ptr<'ctx, StructModel<List<AnyModel<'ctx>>>>,
}
impl<'ctx> ListObject<'ctx> {
/// Create a [`ListObject`] from an LLVM value and its typechecker [`Type`].
pub fn from_value_and_type<V: BasicValue<'ctx>, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
list_val: V,
list_type: Type,
) -> Self {
// Check typechecker type and extract `item_type`
let item_type = match &*ctx.unifier.get_ty(list_type) {
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(list_type))
}
};
let item_model = AnyModel(ctx.get_llvm_type(generator, item_type));
let plist_model = PtrModel(StructModel(List { item: item_model }));
// Create object
let value = plist_model.check_value(generator, ctx.ctx, list_val).unwrap();
ListObject { item_type, value }
}
/// Get the `items` field as an opaque pointer.
pub fn get_opaque_items_ptr<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> Ptr<'ctx, IntModel<Byte>> {
self.value.get(generator, ctx, |f| f.items, "items").pointer_cast(
generator,
ctx,
IntModel(Byte),
"items_opaque",
)
}
/// Get the value of this [`ListObject`] as a list with opaque items.
///
/// This function allocates on the stack to create the list.
pub fn get_opaque_list_ptr<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> Ptr<'ctx, StructModel<List<IntModel<Byte>>>> {
let opaque_list_model = StructModel(List { item: IntModel(Byte) });
let opaque_list_ptr = opaque_list_model.alloca(generator, ctx, "opaque_list_ptr");
// Copy items pointer
let items = self.get_opaque_items_ptr(generator, ctx);
opaque_list_ptr.set(ctx, |f| f.items, items);
// Copy len
let len = self.value.get(generator, ctx, |f| f.len, "len");
opaque_list_ptr.set(ctx, |f| f.len, len);
opaque_list_ptr
}
}

3
nac3core/src/codegen/object/mod.rs Normal file
View File

@ -0,0 +1,3 @@
pub mod list;
pub mod ndarray;
pub mod tuple;

189
nac3core/src/codegen/object/ndarray/array.rs Normal file
View File

@ -0,0 +1,189 @@
use inkwell::values::BasicValueEnum;
use super::NDArrayObject;
use crate::{
codegen::{
irrt::{call_nac3_array_set_and_validate_list_shape, call_nac3_array_write_list_to_array},
model::*,
object::list::ListObject,
stmt::gen_if_else_expr_callback,
CodeGenContext, CodeGenerator,
},
toplevel::helper::{arraylike_flatten_element_type, arraylike_get_ndims},
typecheck::typedef::{Type, TypeEnum},
};
fn get_list_object_dtype_and_ndims<'ctx>(
ctx: &mut CodeGenContext<'ctx, '_>,
list: ListObject<'ctx>,
) -> (Type, u64) {
let dtype = arraylike_flatten_element_type(&mut ctx.unifier, list.item_type);
let ndims = arraylike_get_ndims(&mut ctx.unifier, list.item_type);
let ndims = ndims + 1; // To count `list` itself.
(dtype, ndims)
}
impl<'ctx> NDArrayObject<'ctx> {
fn from_np_array_list_copy<G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
list: ListObject<'ctx>,
) -> Self {
let sizet_model = IntModel(SizeT);
let (dtype, ndims_int) = get_list_object_dtype_and_ndims(ctx, list);
let list_value = list.get_opaque_list_ptr(generator, ctx);
// Validate `list` has a consistent shape.
// Raise an exception if `list` is something abnormal like `[[1, 2], [3]]`.
// If `list` has a consistent shape, deduce the shape and write it to `shape`.
let ndims = sizet_model.constant(generator, ctx.ctx, ndims_int);
let shape = sizet_model.array_alloca(generator, ctx, ndims.value, "shape");
call_nac3_array_set_and_validate_list_shape(generator, ctx, list_value, ndims, shape);
let ndarray = NDArrayObject::alloca_uninitialized(
generator,
ctx,
dtype,
ndims_int,
"ndarray_from_list",
);
ndarray.copy_shape_from_array(generator, ctx, shape);
ndarray.create_data(generator, ctx);
// Copy all contents from the list.
call_nac3_array_write_list_to_array(generator, ctx, list_value, ndarray.value);
ndarray
}
fn from_np_array_list_try_no_copy<G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
list: ListObject<'ctx>,
) -> Self {
// np_array without copy is only possible `list` is not nested.
// If `list` is `list[T]`, we can create an ndarray with `data` set
// to the array pointer of `list`.
let sizet_model = IntModel(SizeT);
let zero = sizet_model.const_0(generator, ctx.ctx);
let (dtype, ndims) = get_list_object_dtype_and_ndims(ctx, list);
if ndims == 1 {
// `list` is not nested, does not need to copy.
let ndarray = NDArrayObject::alloca_uninitialized(
generator,
ctx,
dtype,
1,
"ndarray_from_list_no_copy",
);
// Set data
let data = list.get_opaque_items_ptr(generator, ctx);
ndarray.value.set(ctx, |f| f.data, data);
// Set shape
// dim = list->len;
// shape[0] = dim;
let shape = ndarray.value.get(generator, ctx, |f| f.shape, "shape");
let dim = list.value.get(generator, ctx, |f| f.len, "dim");
shape.offset(generator, ctx, zero.value, "pdim").store(ctx, dim);
// Set strides, the `data` is contiguous
ndarray.update_strides_by_shape(generator, ctx);
// Done
ndarray
} else {
// `list` is nested, it is impossible to not copy.
NDArrayObject::from_np_array_list_copy(generator, ctx, list)
}
}
fn from_np_array_list<G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
list: ListObject<'ctx>,
copy: Int<'ctx, Bool>,
) -> Self {
let (dtype, ndims) = get_list_object_dtype_and_ndims(ctx, list);
let ndarray = gen_if_else_expr_callback(
generator,
ctx,
|_generator, _ctx| Ok(copy.value),
|generator, ctx| {
let ndarray = NDArrayObject::from_np_array_list_copy(generator, ctx, list);
Ok(Some(ndarray.value.value))
},
|generator, ctx| {
let ndarray = NDArrayObject::from_np_array_list_try_no_copy(generator, ctx, list);
Ok(Some(ndarray.value.value))
},
)
.unwrap()
.unwrap();
NDArrayObject::from_value_and_unpacked_types(generator, ctx, ndarray, dtype, ndims)
}
fn from_np_array_ndarray<G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
ndarray: NDArrayObject<'ctx>,
copy: Int<'ctx, Bool>,
) -> Self {
let ndarray_val = gen_if_else_expr_callback(
generator,
ctx,
|_generator, _ctx| Ok(copy.value),
|generator, ctx| {
let ndarray = ndarray.make_clone(generator, ctx, "np_array_copied_ndarray"); // Force copy
Ok(Some(ndarray.value.value))
},
|_generator, _ctx| {
// No need to copy. Return `ndarray` itself.
Ok(Some(ndarray.value.value))
},
)
.unwrap()
.unwrap();
NDArrayObject::from_value_and_unpacked_types(
generator,
ctx,
ndarray_val,
ndarray.dtype,
ndarray.ndims,
)
}
pub fn from_np_array<G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
object: BasicValueEnum<'ctx>,
object_ty: Type,
copy: Int<'ctx, Bool>,
) -> Self {
match &*ctx.unifier.get_ty(object_ty) {
TypeEnum::TObj { obj_id, .. }
if *obj_id == ctx.primitives.list.obj_id(&ctx.unifier).unwrap() =>
{
let list = ListObject::from_value_and_type(generator, ctx, object, object_ty);
NDArrayObject::from_np_array_list(generator, ctx, list, copy)
}
TypeEnum::TObj { obj_id, .. }
if *obj_id == ctx.primitives.ndarray.obj_id(&ctx.unifier).unwrap() =>
{
let ndarray = NDArrayObject::from_value_and_type(generator, ctx, object, object_ty);
NDArrayObject::from_np_array_ndarray(generator, ctx, ndarray, copy)
}
_ => panic!("Unrecognized object type: {}", ctx.unifier.stringify(object_ty)), // Typechecker ensures this
}
}
}

148
nac3core/src/codegen/object/ndarray/broadcast.rs Normal file
View File

@ -0,0 +1,148 @@
use itertools::Itertools;
use crate::{
codegen::{
irrt::{call_nac3_ndarray_broadcast_shapes, call_nac3_ndarray_broadcast_to},
model::*,
CodeGenContext, CodeGenerator,
},
toplevel::numpy::get_broadcast_all_ndims,
};
use super::NDArrayObject;
/// Fields of [`ShapeEntry`]
pub struct ShapeEntryFields<'ctx, F: FieldTraversal<'ctx>> {
pub ndims: F::Out<IntModel<SizeT>>,
pub shape: F::Out<PtrModel<IntModel<SizeT>>>,
}
/// An IRRT structure used in broadcasting.
#[derive(Debug, Clone, Copy, Default)]
pub struct ShapeEntry;
impl<'ctx> StructKind<'ctx> for ShapeEntry {
type Fields<F: FieldTraversal<'ctx>> = ShapeEntryFields<'ctx, F>;
fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
Self::Fields { ndims: traversal.add_auto("ndims"), shape: traversal.add_auto("shape") }
}
}
impl<'ctx> NDArrayObject<'ctx> {
/// Create a broadcast view on this ndarray with a target shape.
///
/// * `target_ndims` - The ndims type after broadcasting to the given shape.
/// The caller has to figure this out for this function.
/// * `target_shape` - An array pointer pointing to the target shape.
#[must_use]
pub fn broadcast_to<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
target_ndims: u64,
target_shape: Ptr<'ctx, IntModel<SizeT>>,
) -> Self {
let broadcast_ndarray = NDArrayObject::alloca_uninitialized(
generator,
ctx,
self.dtype,
target_ndims,
"broadcast_ndarray_to_dst",
);
broadcast_ndarray.copy_shape_from_array(generator, ctx, target_shape);
call_nac3_ndarray_broadcast_to(generator, ctx, self.value, broadcast_ndarray.value);
broadcast_ndarray
}
}
/// A result produced by [`broadcast_all_ndarrays`]
#[derive(Debug, Clone)]
pub struct BroadcastAllResult<'ctx> {
/// The statically known `ndims` of the broadcast result.
pub ndims: u64,
/// The broadcasting shape.
pub shape: Ptr<'ctx, IntModel<SizeT>>,
/// Broadcasted views on the inputs.
///
/// All of them will have `shape` [`BroadcastAllResult::shape`] and
/// `ndims` [`BroadcastAllResult::ndims`]. The length of the vector
/// is the same as the input.
pub ndarrays: Vec<NDArrayObject<'ctx>>,
}
pub fn broadcast_shapes<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
in_entries: &[(Ptr<'ctx, IntModel<SizeT>>, u64)],
broadcast_ndims: u64,
broadcast_shape: Ptr<'ctx, IntModel<SizeT>>,
) {
let sizet_model = IntModel(SizeT);
let shape_model = StructModel(ShapeEntry);
// Prepare input shape entries
let num_shape_entries =
sizet_model.constant(generator, ctx.ctx, u64::try_from(in_entries.len()).unwrap());
let shape_entries =
shape_model.array_alloca(generator, ctx, num_shape_entries.value, "shape_entries");
for (i, (in_shape, in_ndims)) in in_entries.iter().enumerate() {
let i = sizet_model.constant(generator, ctx.ctx, i as u64).value;
let pshape_entry = shape_entries.offset(generator, ctx, i, "shape_entry");
let in_ndims = sizet_model.constant(generator, ctx.ctx, *in_ndims);
pshape_entry.set(ctx, |f| f.ndims, in_ndims);
pshape_entry.set(ctx, |f| f.shape, *in_shape);
}
let broadcast_ndims = sizet_model.constant(generator, ctx.ctx, broadcast_ndims);
call_nac3_ndarray_broadcast_shapes(
generator,
ctx,
num_shape_entries,
shape_entries,
broadcast_ndims,
broadcast_shape,
);
}
impl<'ctx> NDArrayObject<'ctx> {
// TODO: DOCUMENT: Behaves like `np.broadcast()`, except returns results differently.
pub fn broadcast_all<G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
ndarrays: &[Self],
) -> BroadcastAllResult<'ctx> {
assert!(!ndarrays.is_empty());
let sizet_model = IntModel(SizeT);
let broadcast_ndims_int =
get_broadcast_all_ndims(ndarrays.iter().map(|ndarray| ndarray.ndims));
let broadcast_ndims = sizet_model.constant(generator, ctx.ctx, broadcast_ndims_int);
let broadcast_shape =
sizet_model.array_alloca(generator, ctx, broadcast_ndims.value, "broadcast_shape");
let shape_entries = ndarrays
.iter()
.map(|ndarray| (ndarray.value.get(generator, ctx, |f| f.shape, "shape"), ndarray.ndims))
.collect_vec();
broadcast_shapes(generator, ctx, &shape_entries, broadcast_ndims_int, broadcast_shape);
// Broadcast all the inputs to shape `dst_shape`.
let broadcast_ndarrays: Vec<_> = ndarrays
.iter()
.map(|ndarray| {
ndarray.broadcast_to(generator, ctx, broadcast_ndims_int, broadcast_shape)
})
.collect_vec();
BroadcastAllResult {
ndims: broadcast_ndims_int,
shape: broadcast_shape,
ndarrays: broadcast_ndarrays,
}
}
}

563
nac3core/src/codegen/object/ndarray/functions.rs Normal file
View File

@ -0,0 +1,563 @@
use inkwell::{
values::{BasicValue, FloatValue, IntValue},
FloatPredicate, IntPredicate,
};
use itertools::Itertools;
use crate::{
codegen::{
llvm_intrinsics,
model::{
util::{gen_for_model_auto, gen_if_model},
*,
},
CodeGenContext, CodeGenerator,
},
typecheck::typedef::Type,
};
use super::{scalar::ScalarObject, NDArrayObject};
/// Convenience function to crash the program when types of arguments are not supported.
/// Used to be debugged with a stacktrace.
fn unsupported_type<I>(ctx: &CodeGenContext<'_, '_>, tys: I) -> !
where
I: IntoIterator<Item = Type>,
{
unreachable!(
"unsupported types found '{}'",
tys.into_iter().map(|ty| format!("'{}'", ctx.unifier.stringify(ty))).join(", "),
)
}
#[derive(Debug, Clone, Copy)]
pub enum FloorOrCeil {
Floor,
Ceil,
}
#[derive(Debug, Clone, Copy)]
pub enum MinOrMax {
Min,
Max,
}
fn signed_ints(ctx: &CodeGenContext<'_, '_>) -> Vec<Type> {
vec![ctx.primitives.int32, ctx.primitives.int64]
}
fn unsigned_ints(ctx: &CodeGenContext<'_, '_>) -> Vec<Type> {
vec![ctx.primitives.uint32, ctx.primitives.uint64]
}
fn ints(ctx: &CodeGenContext<'_, '_>) -> Vec<Type> {
vec![ctx.primitives.int32, ctx.primitives.int64, ctx.primitives.uint32, ctx.primitives.uint64]
}
fn int_like(ctx: &CodeGenContext<'_, '_>) -> Vec<Type> {
vec![
ctx.primitives.bool,
ctx.primitives.int32,
ctx.primitives.int64,
ctx.primitives.uint32,
ctx.primitives.uint64,
]
}
fn cast_to_int_conversion<'ctx, 'a, G, HandleFloatFn>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, 'a>,
scalar: ScalarObject<'ctx>,
ret_int_dtype: Type,
handle_float: HandleFloatFn,
) -> ScalarObject<'ctx>
where
G: CodeGenerator + ?Sized,
HandleFloatFn:
FnOnce(&mut G, &mut CodeGenContext<'ctx, 'a>, FloatValue<'ctx>) -> IntValue<'ctx>,
{
let ret_int_dtype_llvm = ctx.get_llvm_type(generator, ret_int_dtype).into_int_type();
let result = if ctx.unifier.unioned(scalar.dtype, ctx.primitives.float) {
// Special handling for floats
let n = scalar.value.into_float_value();
handle_float(generator, ctx, n)
} else if ctx.unifier.unioned_any(scalar.dtype, int_like(ctx)) {
let n = scalar.value.into_int_value();
if n.get_type().get_bit_width() <= ret_int_dtype_llvm.get_bit_width() {
ctx.builder.build_int_z_extend(n, ret_int_dtype_llvm, "zext").unwrap()
} else {
ctx.builder.build_int_truncate(n, ret_int_dtype_llvm, "trunc").unwrap()
}
} else {
unsupported_type(ctx, [scalar.dtype]);
};
assert_eq!(ret_int_dtype_llvm.get_bit_width(), result.get_type().get_bit_width()); // Sanity check
ScalarObject { value: result.into(), dtype: ret_int_dtype }
}
impl<'ctx> ScalarObject<'ctx> {
/// Convenience function. Assume this scalar has typechecker type float64, get its underlying LLVM value.
///
/// Panic if the type is wrong.
pub fn into_float64(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> FloatValue<'ctx> {
if ctx.unifier.unioned(self.dtype, ctx.primitives.float) {
self.value.into_float_value() // self.value must be a FloatValue
} else {
panic!("not a float type")
}
}
/// Convenience function. Assume this scalar has typechecker type int32, get its underlying LLVM value.
///
/// Panic if the type is wrong.
pub fn into_int32(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> IntValue<'ctx> {
if ctx.unifier.unioned(self.dtype, ctx.primitives.int32) {
let value = self.value.into_int_value();
debug_assert_eq!(value.get_type().get_bit_width(), 32); // Sanity check
value
} else {
panic!("not a float type")
}
}
/// Compare two scalars. Only int-to-int and float-to-float comparisons are allowed.
/// Panic otherwise.
pub fn compare<G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
lhs: ScalarObject<'ctx>,
rhs: ScalarObject<'ctx>,
int_predicate: IntPredicate,
float_predicate: FloatPredicate,
name: &str,
) -> Int<'ctx, Bool> {
if !ctx.unifier.unioned(lhs.dtype, rhs.dtype) {
unsupported_type(ctx, [lhs.dtype, rhs.dtype]);
}
let bool_model = IntModel(Bool);
let common_ty = lhs.dtype;
let result = if ctx.unifier.unioned(common_ty, ctx.primitives.float) {
let lhs = lhs.value.into_float_value();
let rhs = rhs.value.into_float_value();
ctx.builder.build_float_compare(float_predicate, lhs, rhs, name).unwrap()
} else if ctx.unifier.unioned_any(common_ty, int_like(ctx)) {
let lhs = lhs.value.into_int_value();
let rhs = rhs.value.into_int_value();
ctx.builder.build_int_compare(int_predicate, lhs, rhs, name).unwrap()
} else {
unsupported_type(ctx, [lhs.dtype, rhs.dtype]);
};
bool_model.check_value(generator, ctx.ctx, result).unwrap()
}
/// Invoke NAC3's builtin `int32()`.
#[must_use]
pub fn cast_to_int32<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> Self {
cast_to_int_conversion(
generator,
ctx,
*self,
ctx.primitives.int32,
|_generator, ctx, input| {
let n =
ctx.builder.build_float_to_signed_int(input, ctx.ctx.i64_type(), "").unwrap();
ctx.builder.build_int_truncate(n, ctx.ctx.i32_type(), "conv").unwrap()
},
)
}
/// Invoke NAC3's builtin `int64()`.
#[must_use]
pub fn cast_to_int64<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> Self {
cast_to_int_conversion(
generator,
ctx,
*self,
ctx.primitives.int64,
|_generator, ctx, input| {
ctx.builder.build_float_to_signed_int(input, ctx.ctx.i64_type(), "").unwrap()
},
)
}
/// Invoke NAC3's builtin `uint32()`.
#[must_use]
pub fn cast_to_uint32<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> Self {
cast_to_int_conversion(
generator,
ctx,
*self,
ctx.primitives.uint32,
|_generator, ctx, n| {
let n_gez = ctx
.builder
.build_float_compare(FloatPredicate::OGE, n, n.get_type().const_zero(), "")
.unwrap();
let to_int32 =
ctx.builder.build_float_to_signed_int(n, ctx.ctx.i32_type(), "").unwrap();
let to_uint64 =
ctx.builder.build_float_to_unsigned_int(n, ctx.ctx.i64_type(), "").unwrap();
ctx.builder
.build_select(
n_gez,
ctx.builder.build_int_truncate(to_uint64, ctx.ctx.i32_type(), "").unwrap(),
to_int32,
"conv",
)
.unwrap()
.into_int_value()
},
)
}
/// Invoke NAC3's builtin `uint64()`.
#[must_use]
pub fn cast_to_uint64<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> Self {
cast_to_int_conversion(
generator,
ctx,
*self,
ctx.primitives.uint64,
|_generator, ctx, n| {
let val_gez = ctx
.builder
.build_float_compare(FloatPredicate::OGE, n, n.get_type().const_zero(), "")
.unwrap();
let to_int64 =
ctx.builder.build_float_to_signed_int(n, ctx.ctx.i64_type(), "").unwrap();
let to_uint64 =
ctx.builder.build_float_to_unsigned_int(n, ctx.ctx.i64_type(), "").unwrap();
ctx.builder
.build_select(val_gez, to_uint64, to_int64, "conv")
.unwrap()
.into_int_value()
},
)
}
/// Invoke NAC3's builtin `bool()`.
#[must_use]
pub fn cast_to_bool(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> Self {
// TODO: Why is the original code being so lax about i1 and i8 for the returned int type?
let result = if ctx.unifier.unioned(self.dtype, ctx.primitives.bool) {
self.value.into_int_value()
} else if ctx.unifier.unioned_any(self.dtype, ints(ctx)) {
let n = self.value.into_int_value();
ctx.builder
.build_int_compare(inkwell::IntPredicate::NE, n, n.get_type().const_zero(), "bool")
.unwrap()
} else if ctx.unifier.unioned(self.dtype, ctx.primitives.float) {
let n = self.value.into_float_value();
ctx.builder
.build_float_compare(FloatPredicate::UNE, n, n.get_type().const_zero(), "bool")
.unwrap()
} else {
unsupported_type(ctx, [self.dtype])
};
ScalarObject { dtype: ctx.primitives.bool, value: result.as_basic_value_enum() }
}
/// Invoke NAC3's builtin `float()`.
#[must_use]
pub fn cast_to_float(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> Self {
let llvm_f64 = ctx.ctx.f64_type();
let result: FloatValue<'_> = if ctx.unifier.unioned(self.dtype, ctx.primitives.float) {
self.value.into_float_value()
} else if ctx
.unifier
.unioned_any(self.dtype, [signed_ints(ctx).as_slice(), &[ctx.primitives.bool]].concat())
{
let n = self.value.into_int_value();
ctx.builder.build_signed_int_to_float(n, llvm_f64, "sitofp").unwrap()
} else if ctx.unifier.unioned_any(self.dtype, unsigned_ints(ctx)) {
let n = self.value.into_int_value();
ctx.builder.build_unsigned_int_to_float(n, llvm_f64, "uitofp").unwrap()
} else {
unsupported_type(ctx, [self.dtype]);
};
ScalarObject { value: result.as_basic_value_enum(), dtype: ctx.primitives.float }
}
/// Invoke NAC3's builtin `round()`.
#[must_use]
pub fn round<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
ret_int_dtype: Type,
) -> Self {
let ret_int_dtype_llvm = ctx.get_llvm_type(generator, ret_int_dtype).into_int_type();
let result = if ctx.unifier.unioned(self.dtype, ctx.primitives.float) {
let n = self.value.into_float_value();
let n = llvm_intrinsics::call_float_round(ctx, n, None);
ctx.builder.build_float_to_signed_int(n, ret_int_dtype_llvm, "round").unwrap()
} else {
unsupported_type(ctx, [self.dtype, ret_int_dtype])
};
ScalarObject { dtype: ret_int_dtype, value: result.as_basic_value_enum() }
}
/// Invoke NAC3's builtin `np_round()`.
///
/// NOTE: `np.round()` has different behaviors than `round()` in terms of their result
/// on "tie" cases and return type.
#[must_use]
pub fn np_round(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> Self {
let result = if ctx.unifier.unioned(self.dtype, ctx.primitives.float) {
let n = self.value.into_float_value();
llvm_intrinsics::call_float_rint(ctx, n, None)
} else {
unsupported_type(ctx, [self.dtype])
};
ScalarObject { dtype: ctx.primitives.float, value: result.as_basic_value_enum() }
}
/// Invoke NAC3's builtin `min()` or `max()`.
pub fn min_or_max(
ctx: &mut CodeGenContext<'ctx, '_>,
kind: MinOrMax,
a: Self,
b: Self,
) -> Self {
if !ctx.unifier.unioned(a.dtype, b.dtype) {
unsupported_type(ctx, [a.dtype, b.dtype])
}
let common_dtype = a.dtype;
if ctx.unifier.unioned(common_dtype, ctx.primitives.float) {
let function = match kind {
MinOrMax::Min => llvm_intrinsics::call_float_minnum,
MinOrMax::Max => llvm_intrinsics::call_float_maxnum,
};
let result =
function(ctx, a.value.into_float_value(), b.value.into_float_value(), None);
ScalarObject { value: result.as_basic_value_enum(), dtype: ctx.primitives.float }
} else if ctx.unifier.unioned_any(
common_dtype,
[unsigned_ints(ctx).as_slice(), &[ctx.primitives.bool]].concat(),
) {
// Treating bool has an unsigned int since that is convenient
let function = match kind {
MinOrMax::Min => llvm_intrinsics::call_int_umin,
MinOrMax::Max => llvm_intrinsics::call_int_umax,
};
let result = function(ctx, a.value.into_int_value(), b.value.into_int_value(), None);
ScalarObject { value: result.as_basic_value_enum(), dtype: common_dtype }
} else {
unsupported_type(ctx, [common_dtype])
}
}
/// Invoke NAC3's builtin `floor()` or `ceil()`.
///
/// * `ret_int_dtype` - The type of int to return.
///
/// Takes in a float/int and returns an int of type `ret_int_dtype`
#[must_use]
pub fn floor_or_ceil<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
kind: FloorOrCeil,
ret_int_dtype: Type,
) -> Self {
let ret_int_dtype_llvm = ctx.get_llvm_type(generator, ret_int_dtype).into_int_type();
if ctx.unifier.unioned(self.dtype, ctx.primitives.float) {
let function = match kind {
FloorOrCeil::Floor => llvm_intrinsics::call_float_floor,
FloorOrCeil::Ceil => llvm_intrinsics::call_float_ceil,
};
let n = self.value.into_float_value();
let n = function(ctx, n, None);
let n = ctx.builder.build_float_to_signed_int(n, ret_int_dtype_llvm, "").unwrap();
ScalarObject { dtype: ret_int_dtype, value: n.as_basic_value_enum() }
} else {
unsupported_type(ctx, [self.dtype])
}
}
/// Invoke NAC3's builtin `np_floor()`/ `np_ceil()`.
///
/// Takes in a float/int and returns a float64 result.
#[must_use]
pub fn np_floor_or_ceil(&self, ctx: &mut CodeGenContext<'ctx, '_>, kind: FloorOrCeil) -> Self {
if ctx.unifier.unioned(self.dtype, ctx.primitives.float) {
let function = match kind {
FloorOrCeil::Floor => llvm_intrinsics::call_float_floor,
FloorOrCeil::Ceil => llvm_intrinsics::call_float_ceil,
};
let n = self.value.into_float_value();
let n = function(ctx, n, None);
ScalarObject { dtype: ctx.primitives.float, value: n.as_basic_value_enum() }
} else {
unsupported_type(ctx, [self.dtype])
}
}
/// Invoke NAC3's builtin `abs()`.
#[must_use]
pub fn abs(&self, ctx: &mut CodeGenContext<'ctx, '_>) -> Self {
if ctx.unifier.unioned(self.dtype, ctx.primitives.float) {
let n = self.value.into_float_value();
let n = llvm_intrinsics::call_float_fabs(ctx, n, Some("abs"));
ScalarObject { value: n.into(), dtype: ctx.primitives.float }
} else if ctx.unifier.unioned_any(self.dtype, ints(ctx)) {
let n = self.value.into_int_value();
let is_poisoned = ctx.ctx.bool_type().const_zero(); // is_poisoned = false
let n = llvm_intrinsics::call_int_abs(ctx, n, is_poisoned, Some("abs"));
ScalarObject { value: n.into(), dtype: self.dtype }
} else {
unsupported_type(ctx, [self.dtype])
}
}
}
impl<'ctx> NDArrayObject<'ctx> {
/// Helper function to implement NAC3's builtin `np_min()`, `np_max()`, `np_argmin()`, and `np_argmax()`.
///
/// Generate LLVM IR to find the extremum and index of the **first** extremum value.
///
/// Care has also been taken to make the error messages match that of NumPy.
fn min_max_argmin_argmax_helper<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
kind: MinOrMax,
on_empty_err_msg: &str,
) -> (ScalarObject<'ctx>, Int<'ctx, SizeT>) {
let sizet_model = IntModel(SizeT);
let dtype_llvm = ctx.get_llvm_type(generator, self.dtype);
// If the ndarray is empty, throw an error.
let is_empty = self.is_empty(generator, ctx);
ctx.make_assert(
generator,
is_empty.value,
"0:ValueError",
on_empty_err_msg,
[None, None, None],
ctx.current_loc,
);
// Setup and initialize the extremum to be the first element in the ndarray
let pextremum_index = sizet_model.alloca(generator, ctx, "extremum_index");
let pextremum = ctx.builder.build_alloca(dtype_llvm, "extremum").unwrap();
let zero = sizet_model.const_0(generator, ctx.ctx);
pextremum_index.store(ctx, zero);
let first_scalar = self.get_nth(generator, ctx, zero);
ctx.builder.build_store(pextremum, first_scalar.value).unwrap();
// Find extremum
let start = sizet_model.const_1(generator, ctx.ctx); // Start on 1
let stop = self.size(generator, ctx);
let step = sizet_model.const_1(generator, ctx.ctx);
gen_for_model_auto(generator, ctx, start, stop, step, |generator, ctx, _hooks, i| {
// Worth reading on "Notes" in <https://numpy.org/doc/stable/reference/generated/numpy.min.html#numpy.min>
// on how `NaN` values have to be handled.
let scalar = self.get_nth(generator, ctx, i);
let old_extremum = ctx.builder.build_load(pextremum, "current_extremum").unwrap();
let old_extremum = ScalarObject { dtype: self.dtype, value: old_extremum };
let new_extremum = ScalarObject::min_or_max(ctx, kind, old_extremum, scalar);
// Check if new_extremum is more extreme than old_extremum.
let update_index = ScalarObject::compare(
generator,
ctx,
new_extremum,
old_extremum,
IntPredicate::NE,
FloatPredicate::ONE,
"",
);
gen_if_model(generator, ctx, update_index, |_generator, ctx| {
pextremum_index.store(ctx, i);
Ok(())
})
.unwrap();
Ok(())
})
.unwrap();
// Finally return the extremum and extremum index.
let extremum_index = pextremum_index.load(generator, ctx, "extremum_index");
let extremum = ctx.builder.build_load(pextremum, "extremum_value").unwrap();
let extremum = ScalarObject { dtype: self.dtype, value: extremum };
(extremum, extremum_index)
}
/// Invoke NAC3's builtin `np_min()` or `np_max()`.
pub fn min_or_max<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
kind: MinOrMax,
) -> ScalarObject<'ctx> {
let on_empty_err_msg = format!(
"zero-size array to reduction operation {} which has no identity",
match kind {
MinOrMax::Min => "minimum",
MinOrMax::Max => "maximum",
}
);
self.min_max_argmin_argmax_helper(generator, ctx, kind, &on_empty_err_msg).0
}
/// Invoke NAC3's builtin `np_argmin()` or `np_argmax()`.
pub fn argmin_or_argmax<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
kind: MinOrMax,
) -> Int<'ctx, SizeT> {
let on_empty_err_msg = format!(
"attempt to get {} of an empty sequence",
match kind {
MinOrMax::Min => "argmin",
MinOrMax::Max => "argmax",
}
);
self.min_max_argmin_argmax_helper(generator, ctx, kind, &on_empty_err_msg).1
}
}

322
nac3core/src/codegen/object/ndarray/indexing.rs Normal file
View File

@ -0,0 +1,322 @@
use crate::codegen::{irrt::call_nac3_ndarray_index, model::*, CodeGenContext, CodeGenerator};
use super::{scalar::ScalarOrNDArray, NDArrayObject};
pub type NDIndexType = Byte;
/// Fields of [`NDIndex`]
#[derive(Debug, Clone, Copy)]
pub struct NDIndexFields<'ctx, F: FieldTraversal<'ctx>> {
pub type_: F::Out<IntModel<NDIndexType>>, // Defined to be uint8_t in IRRT
pub data: F::Out<PtrModel<IntModel<Byte>>>,
}
/// An IRRT representation fo an ndarray subscript index.
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
pub struct NDIndex;
impl<'ctx> StructKind<'ctx> for NDIndex {
type Fields<F: FieldTraversal<'ctx>> = NDIndexFields<'ctx, F>;
fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
Self::Fields { type_: traversal.add_auto("type"), data: traversal.add_auto("data") }
}
}
/// Fields of [`UserSlice`]
#[derive(Debug, Clone)]
pub struct UserSliceFields<'ctx, F: FieldTraversal<'ctx>> {
pub start_defined: F::Out<IntModel<Bool>>,
pub start: F::Out<IntModel<Int32>>,
pub stop_defined: F::Out<IntModel<Bool>>,
pub stop: F::Out<IntModel<Int32>>,
pub step_defined: F::Out<IntModel<Bool>>,
pub step: F::Out<IntModel<Int32>>,
}
/// An IRRT representation of a user slice.
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
pub struct UserSlice;
impl<'ctx> StructKind<'ctx> for UserSlice {
type Fields<F: FieldTraversal<'ctx>> = UserSliceFields<'ctx, F>;
fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
Self::Fields {
start_defined: traversal.add_auto("start_defined"),
start: traversal.add_auto("start"),
stop_defined: traversal.add_auto("stop_defined"),
stop: traversal.add_auto("stop"),
step_defined: traversal.add_auto("step_defined"),
step: traversal.add_auto("step"),
}
}
}
/// A convenience structure to prepare a [`UserSlice`].
#[derive(Debug, Clone)]
pub struct RustUserSlice<'ctx> {
pub start: Option<Int<'ctx, Int32>>,
pub stop: Option<Int<'ctx, Int32>>,
pub step: Option<Int<'ctx, Int32>>,
}
impl<'ctx> RustUserSlice<'ctx> {
/// Write the contents to an LLVM [`UserSlice`].
pub fn write_to_user_slice<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
dst_slice_ptr: Ptr<'ctx, StructModel<UserSlice>>,
) {
let bool_model = IntModel(Bool);
let false_ = bool_model.constant(generator, ctx.ctx, 0);
let true_ = bool_model.constant(generator, ctx.ctx, 1);
// TODO: Code duplication. Probably okay...?
match self.start {
Some(start) => {
dst_slice_ptr.gep(ctx, |f| f.start_defined).store(ctx, true_);
dst_slice_ptr.gep(ctx, |f| f.start).store(ctx, start);
}
None => dst_slice_ptr.gep(ctx, |f| f.start_defined).store(ctx, false_),
}
match self.stop {
Some(stop) => {
dst_slice_ptr.gep(ctx, |f| f.stop_defined).store(ctx, true_);
dst_slice_ptr.gep(ctx, |f| f.stop).store(ctx, stop);
}
None => dst_slice_ptr.gep(ctx, |f| f.stop_defined).store(ctx, false_),
}
match self.step {
Some(step) => {
dst_slice_ptr.gep(ctx, |f| f.step_defined).store(ctx, true_);
dst_slice_ptr.gep(ctx, |f| f.step).store(ctx, step);
}
None => dst_slice_ptr.gep(ctx, |f| f.step_defined).store(ctx, false_),
}
}
}
// A convenience enum variant to store the content and type of an NDIndex in high level.
#[derive(Debug, Clone)]
pub enum RustNDIndex<'ctx> {
SingleElement(Int<'ctx, Int32>), // TODO: To be SizeT
Slice(RustUserSlice<'ctx>),
NewAxis,
Ellipsis,
}
impl<'ctx> RustNDIndex<'ctx> {
/// Get the value to set `NDIndex::type` for this variant.
fn get_type_id(&self) -> u64 {
// Defined in IRRT, must be in sync
match self {
RustNDIndex::SingleElement(_) => 0,
RustNDIndex::Slice(_) => 1,
RustNDIndex::NewAxis => 2,
RustNDIndex::Ellipsis => 3,
}
}
/// Write the contents to an LLVM [`NDIndex`].
fn write_to_ndindex<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
dst_ndindex_ptr: Ptr<'ctx, StructModel<NDIndex>>,
) {
let ndindex_type_model = IntModel(NDIndexType::default());
let i32_model = IntModel(Int32);
let user_slice_model = StructModel(UserSlice);
// Set `dst_ndindex_ptr->type`
dst_ndindex_ptr
.gep(ctx, |f| f.type_)
.store(ctx, ndindex_type_model.constant(generator, ctx.ctx, self.get_type_id()));
// Set `dst_ndindex_ptr->data`
match self {
RustNDIndex::SingleElement(in_index) => {
let index_ptr = i32_model.alloca(generator, ctx, "index");
index_ptr.store(ctx, *in_index);
dst_ndindex_ptr
.gep(ctx, |f| f.data)
.store(ctx, index_ptr.pointer_cast(generator, ctx, IntModel(Byte), ""));
}
RustNDIndex::Slice(in_rust_slice) => {
let user_slice_ptr = user_slice_model.alloca(generator, ctx, "user_slice");
in_rust_slice.write_to_user_slice(generator, ctx, user_slice_ptr);
dst_ndindex_ptr
.gep(ctx, |f| f.data)
.store(ctx, user_slice_ptr.pointer_cast(generator, ctx, IntModel(Byte), ""));
}
RustNDIndex::NewAxis | RustNDIndex::Ellipsis => {}
}
}
/// Allocate an array of `NDIndex`es on the stack and return its stack pointer.
pub fn alloca_ndindexes<G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &CodeGenContext<'ctx, '_>,
in_ndindexes: &[RustNDIndex<'ctx>],
) -> (Int<'ctx, SizeT>, Ptr<'ctx, StructModel<NDIndex>>) {
let sizet_model = IntModel(SizeT);
let ndindex_model = StructModel(NDIndex);
let num_ndindexes = sizet_model.constant(generator, ctx.ctx, in_ndindexes.len() as u64);
let ndindexes =
ndindex_model.array_alloca(generator, ctx, num_ndindexes.value, "ndindexes");
for (i, in_ndindex) in in_ndindexes.iter().enumerate() {
let i = sizet_model.constant(generator, ctx.ctx, i as u64);
let pndindex = ndindexes.offset(generator, ctx, i.value, "");
in_ndindex.write_to_ndindex(generator, ctx, pndindex);
}
(num_ndindexes, ndindexes)
}
}
impl<'ctx> NDArrayObject<'ctx> {
/// Get the ndims [`Type`] after indexing with a given slice.
#[must_use]
pub fn deduce_ndims_after_indexing_with(&self, indexes: &[RustNDIndex<'ctx>]) -> u64 {
let mut ndims = self.ndims;
for index in indexes {
match index {
RustNDIndex::SingleElement(_) => {
ndims -= 1; // Single elements decrements ndims
}
RustNDIndex::NewAxis => {
ndims += 1; // `np.newaxis` / `none` adds a new axis
}
RustNDIndex::Ellipsis | RustNDIndex::Slice(_) => {}
}
}
ndims
}
/// Index into the ndarray, and return a newly-allocated view on this ndarray.
///
/// This function behaves like NumPy's ndarray indexing, but if the indexes index
/// into a single element, an unsized ndarray is returned.
#[must_use]
pub fn index<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
indexes: &[RustNDIndex<'ctx>],
name: &str,
) -> Self {
let dst_ndims = self.deduce_ndims_after_indexing_with(indexes);
let dst_ndarray =
NDArrayObject::alloca_uninitialized(generator, ctx, self.dtype, dst_ndims, name);
let (num_indexes, indexes) = RustNDIndex::alloca_ndindexes(generator, ctx, indexes);
call_nac3_ndarray_index(
generator,
ctx,
num_indexes,
indexes,
self.value,
dst_ndarray.value,
);
dst_ndarray
}
}
pub mod util {
use itertools::Itertools;
use nac3parser::ast::{Constant, Expr, ExprKind};
use crate::{
codegen::{expr::gen_slice, model::*, CodeGenContext, CodeGenerator},
typecheck::typedef::{Type, TypeEnum},
};
use super::{RustNDIndex, RustUserSlice};
/// Generate LLVM code to transform an ndarray subscript expression to
/// its list of [`RustNDIndex`]
///
/// i.e.,
/// ```python
/// my_ndarray[::3, 1, :2:]
/// ^^^^^^^^^^^ Then these into a three `RustNDIndex`es
/// ```
pub fn gen_ndarray_subscript_ndindexes<'ctx, G: CodeGenerator>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
subscript: &Expr<Option<Type>>,
) -> Result<Vec<RustNDIndex<'ctx>>, String> {
// TODO: Support https://numpy.org/doc/stable/user/basics.indexing.html#dimensional-indexing-tools
let i32_model = IntModel(Int32);
// Annoying notes about `slice`
// - `my_array[5]`
// - slice is a `Constant`
// - `my_array[:5]`
// - slice is a `Slice`
// - `my_array[:]`
// - slice is a `Slice`, but lower upper step would all be `Option::None`
// - `my_array[:, :]`
// - slice is now a `Tuple` of two `Slice`-s
//
// In summary:
// - when there is a comma "," within [], `slice` will be a `Tuple` of the entries.
// - when there is not comma "," within [] (i.e., just a single entry), `slice` will be that entry itself.
//
// So we first "flatten" out the slice expression
let index_exprs = match &subscript.node {
ExprKind::Tuple { elts, .. } => elts.iter().collect_vec(),
_ => vec![subscript],
};
// Process all index expressions
let mut rust_ndindexes: Vec<RustNDIndex> = Vec::with_capacity(index_exprs.len()); // Not using iterators here because `?` is used here.
for index_expr in index_exprs {
// NOTE: Currently nac3core's slices do not have an object representation,
// so the code/implementation looks awkward - we have to do pattern matching on the expression
let ndindex = if let ExprKind::Slice { lower, upper, step } = &index_expr.node {
// Handle slices
// Helper function here to deduce code duplication
let (lower, upper, step) = gen_slice(generator, ctx, lower, upper, step)?;
RustNDIndex::Slice(RustUserSlice { start: lower, stop: upper, step })
} else if let ExprKind::Constant { value: Constant::Ellipsis, .. } = &index_expr.node {
// Handle '...'
RustNDIndex::Ellipsis
} else {
match &*ctx.unifier.get_ty(index_expr.custom.unwrap()) {
TypeEnum::TObj { obj_id, .. }
if *obj_id == ctx.primitives.option.obj_id(&ctx.unifier).unwrap() =>
{
// Handle `np.newaxis` / `None`
RustNDIndex::NewAxis
}
_ => {
// Treat and handle everything else as a single element index.
let index =
generator.gen_expr(ctx, index_expr)?.unwrap().to_basic_value_enum(
ctx,
generator,
ctx.primitives.int32, // Must be int32, this checks for illegal values
)?;
let index = i32_model.check_value(generator, ctx.ctx, index).unwrap();
RustNDIndex::SingleElement(index)
}
}
};
rust_ndindexes.push(ndindex);
}
Ok(rust_ndindexes)
}
}

180
nac3core/src/codegen/object/ndarray/mapping.rs Normal file
View File

@ -0,0 +1,180 @@
use inkwell::values::BasicValueEnum;
use itertools::Itertools;
use util::gen_for_model_auto;
use crate::{
codegen::{
model::*,
object::ndarray::{NDArrayObject, ScalarObject},
CodeGenContext, CodeGenerator,
},
typecheck::typedef::Type,
};
use super::{scalar::ScalarOrNDArray, NDArrayOut};
impl<'ctx> NDArrayObject<'ctx> {
/// TODO: Document me. Has complex behavior.
/// and explain why `ret_dtype` has to be specified beforehand.
pub fn broadcasting_starmap<'a, G, MappingFn>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, 'a>,
ndarrays: &[Self],
out: NDArrayOut<'ctx>,
mapping: MappingFn,
) -> Result<Self, String>
where
G: CodeGenerator + ?Sized,
MappingFn: FnOnce(
&mut G,
&mut CodeGenContext<'ctx, 'a>,
Int<'ctx, SizeT>,
&[ScalarObject<'ctx>],
) -> Result<BasicValueEnum<'ctx>, String>,
{
let sizet_model = IntModel(SizeT);
// Broadcast inputs
let broadcast_result = NDArrayObject::broadcast_all(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_uninitialized(
generator,
ctx,
dtype,
broadcast_result.ndims,
"mapped_ndarray",
);
result_ndarray.copy_shape_from_array(generator, ctx, broadcast_result.shape);
result_ndarray.create_data(generator, ctx);
result_ndarray
}
NDArrayOut::WriteToNDArray { ndarray: result_ndarray } => {
// Use an existing ndarray.
// Check that its shape is compatible with the broadcast shape.
result_ndarray.check_can_be_written_by_out(
generator,
ctx,
broadcast_result.ndims,
broadcast_result.shape,
);
result_ndarray
}
};
// Map element-wise and store results into `mapped_ndarray`.
let start = sizet_model.const_0(generator, ctx.ctx);
let stop = broadcast_result.ndarrays[0].size(generator, ctx); // They all should have the same `np.size`.
let step = sizet_model.const_1(generator, ctx.ctx);
gen_for_model_auto(generator, ctx, start, stop, step, move |generator, ctx, _hooks, i| {
let elements =
ndarrays.iter().map(|ndarray| ndarray.get_nth(generator, ctx, i)).collect_vec();
let ret = mapping(generator, ctx, i, &elements)?;
let pret = out_ndarray.get_nth_pointer(generator, ctx, i, "pret");
ctx.builder.build_store(pret, ret).unwrap();
Ok(())
})?;
Ok(out_ndarray)
}
pub fn map<'a, G, Mapping>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, 'a>,
out: NDArrayOut<'ctx>,
mapping: Mapping,
) -> Result<Self, String>
where
G: CodeGenerator + ?Sized,
Mapping: FnOnce(
&mut G,
&mut CodeGenContext<'ctx, 'a>,
Int<'ctx, SizeT>,
ScalarObject<'ctx>,
) -> Result<BasicValueEnum<'ctx>, String>,
{
NDArrayObject::broadcasting_starmap(
generator,
ctx,
&[*self],
out,
|generator, ctx, i, scalars| mapping(generator, ctx, i, scalars[0]),
)
}
}
impl<'ctx> ScalarOrNDArray<'ctx> {
/// TODO: Document me. Has complex behavior.
/// and explain why `ret_dtype` has to be specified beforehand.
pub fn broadcasting_starmap<'a, G, MappingFn>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, 'a>,
inputs: &[Self],
ret_dtype: Type,
mapping: MappingFn,
) -> Result<Self, String>
where
G: CodeGenerator + ?Sized,
MappingFn: FnOnce(
&mut G,
&mut CodeGenContext<'ctx, 'a>,
Int<'ctx, SizeT>,
&[ScalarObject<'ctx>],
) -> Result<BasicValueEnum<'ctx>, String>,
{
let sizet_model = IntModel(SizeT);
// Check if all inputs are ScalarObjects
let all_scalars: Option<Vec<_>> =
inputs.iter().map(ScalarObject::try_from).try_collect().ok();
if let Some(scalars) = all_scalars {
let i = sizet_model.const_0(generator, ctx.ctx); // Pass 0 as the index
let scalar =
ScalarObject { value: mapping(generator, ctx, i, &scalars)?, dtype: ret_dtype };
Ok(ScalarOrNDArray::Scalar(scalar))
} else {
// Promote all input to ndarrays and map through them.
let inputs = inputs.iter().map(|input| input.as_ndarray(generator, ctx)).collect_vec();
let ndarray = NDArrayObject::broadcasting_starmap(
generator,
ctx,
&inputs,
NDArrayOut::NewNDArray { dtype: ret_dtype },
mapping,
)?;
Ok(ScalarOrNDArray::NDArray(ndarray))
}
}
pub fn map<'a, G, Mapping>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, 'a>,
ret_dtype: Type,
mapping: Mapping,
) -> Result<Self, String>
where
G: CodeGenerator + ?Sized,
Mapping: FnOnce(
&mut G,
&mut CodeGenContext<'ctx, 'a>,
Int<'ctx, SizeT>,
ScalarObject<'ctx>,
) -> Result<BasicValueEnum<'ctx>, String>,
{
ScalarOrNDArray::broadcasting_starmap(
generator,
ctx,
&[*self],
ret_dtype,
|generator, ctx, i, scalars| mapping(generator, ctx, i, scalars[0]),
)
}
}

617
nac3core/src/codegen/object/ndarray/mod.rs Normal file
View File

@ -0,0 +1,617 @@
pub mod array;
pub mod broadcast;
pub mod functions;
pub mod indexing;
pub mod mapping;
pub mod product;
pub mod scalar;
pub mod shape_util;
use crate::{
codegen::{
irrt::{
call_nac3_ndarray_copy_data, call_nac3_ndarray_get_nth_pelement,
call_nac3_ndarray_is_c_contiguous, call_nac3_ndarray_len, call_nac3_ndarray_nbytes,
call_nac3_ndarray_resolve_and_check_new_shape, call_nac3_ndarray_set_strides_by_shape,
call_nac3_ndarray_size, call_nac3_ndarray_transpose,
call_nac3_ndarray_util_assert_output_shape_same,
},
model::*,
stmt::BreakContinueHooks,
structure::NDArray,
CodeGenContext, CodeGenerator,
},
toplevel::numpy::{extract_ndims, unpack_ndarray_var_tys},
typecheck::typedef::Type,
};
use indexing::RustNDIndex;
use inkwell::{
context::Context,
types::BasicType,
values::{BasicValue, BasicValueEnum, PointerValue},
AddressSpace, IntPredicate,
};
use scalar::{ScalarObject, ScalarOrNDArray};
use util::{call_memcpy_model, gen_for_model_auto};
pub struct NpArrayFields<'ctx, F: FieldTraversal<'ctx>> {
pub data: F::Out<PtrModel<IntModel<Byte>>>,
pub itemsize: F::Out<IntModel<SizeT>>,
pub ndims: F::Out<IntModel<SizeT>>,
pub shape: F::Out<PtrModel<IntModel<SizeT>>>,
pub strides: F::Out<PtrModel<IntModel<SizeT>>>,
}
/// A NAC3 Python ndarray object.
#[derive(Debug, Clone, Copy)]
pub struct NDArrayObject<'ctx> {
pub dtype: Type,
pub ndims: u64,
pub value: Ptr<'ctx, StructModel<NDArray>>,
}
impl<'ctx> NDArrayObject<'ctx> {
/// Create an [`NDArrayObject`] from an LLVM value and its typechecker [`Type`].
pub fn from_value_and_type<V: BasicValue<'ctx>, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
value: V,
ty: Type,
) -> Self {
let (dtype, ndims) = unpack_ndarray_var_tys(&mut ctx.unifier, ty);
let ndims = extract_ndims(&ctx.unifier, ndims);
Self::from_value_and_unpacked_types(generator, ctx, value, dtype, ndims)
}
/// Like [`NDArrayObject::from_value_and_type`] but you directly supply the ndarray's
/// `dtype` and `ndims`.
pub fn from_value_and_unpacked_types<V: BasicValue<'ctx>, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
value: V,
dtype: Type,
ndims: u64,
) -> Self {
let pndarray_model = PtrModel(StructModel(NDArray));
let value = pndarray_model.check_value(generator, ctx.ctx, value).unwrap();
NDArrayObject { dtype, ndims, value }
}
/// Get the `np.size()` of this ndarray.
pub fn size<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> Int<'ctx, SizeT> {
call_nac3_ndarray_size(generator, ctx, self.value)
}
/// Get the `ndarray.nbytes` of this ndarray.
pub fn nbytes<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> Int<'ctx, SizeT> {
call_nac3_ndarray_nbytes(generator, ctx, self.value)
}
/// Get the `len()` of this ndarray.
pub fn len<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> Int<'ctx, SizeT> {
call_nac3_ndarray_len(generator, ctx, self.value)
}
/// Check if this ndarray is C-contiguous.
///
/// See NumPy's `flags["C_CONTIGUOUS"]`: <https://numpy.org/doc/stable/reference/generated/numpy.ndarray.flags.html#numpy.ndarray.flags>
pub fn is_c_contiguous<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> Int<'ctx, Bool> {
call_nac3_ndarray_is_c_contiguous(generator, ctx, self.value)
}
/// 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_pointer<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
nth: Int<'ctx, SizeT>,
name: &str,
) -> PointerValue<'ctx> {
let elem_ty = ctx.get_llvm_type(generator, self.dtype);
let p = call_nac3_ndarray_get_nth_pelement(generator, ctx, self.value, nth);
ctx.builder
.build_pointer_cast(p.value, elem_ty.ptr_type(AddressSpace::default()), name)
.unwrap()
}
/// Get the n-th (0-based) scalar.
pub fn get_nth<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
nth: Int<'ctx, SizeT>,
) -> ScalarObject<'ctx> {
let p = self.get_nth_pointer(generator, ctx, nth, "value");
let value = ctx.builder.build_load(p, "value").unwrap();
ScalarObject { dtype: self.dtype, value }
}
/// Call [`call_nac3_ndarray_set_strides_by_shape`] on this ndarray to update `strides`.
///
/// Please refer to the IRRT implementation to see its purpose.
pub fn update_strides_by_shape<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) {
call_nac3_ndarray_set_strides_by_shape(generator, ctx, self.value);
}
/// 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.value, self.value);
}
/// Allocate an ndarray on the stack given its `ndims` and `dtype`.
///
/// `shape` and `strides` will be automatically allocated on the stack.
///
/// The returned ndarray's content will be:
/// - `data`: set to `nullptr`.
/// - `itemsize`: set to the `sizeof()` of `dtype`.
/// - `ndims`: set to the value of `ndims`.
/// - `shape`: allocated with an array of length `ndims` with uninitialized values.
/// - `strides`: allocated with an array of length `ndims` with uninitialized values.
pub fn alloca_uninitialized<G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
dtype: Type,
ndims: u64,
name: &str,
) -> Self {
let sizet_model = IntModel(SizeT);
let ndarray_model = StructModel(NDArray);
let ndarray_data_model = PtrModel(IntModel(Byte));
let pndarray = ndarray_model.alloca(generator, ctx, name);
let data = ndarray_data_model.nullptr(generator, ctx.ctx);
pndarray.set(ctx, |f| f.data, data);
let itemsize = ctx.get_llvm_type(generator, dtype).size_of().unwrap();
let itemsize =
sizet_model.s_extend_or_bit_cast(generator, ctx, itemsize, "alloca_itemsize");
pndarray.set(ctx, |f| f.itemsize, itemsize);
let ndims_val = sizet_model.constant(generator, ctx.ctx, ndims);
pndarray.set(ctx, |f| f.ndims, ndims_val);
let shape = sizet_model.array_alloca(generator, ctx, ndims_val.value, "alloca_shape");
pndarray.set(ctx, |f| f.shape, shape);
let strides = sizet_model.array_alloca(generator, ctx, ndims_val.value, "alloca_strides");
pndarray.set(ctx, |f| f.strides, strides);
NDArrayObject { dtype, ndims, value: pndarray }
}
/// Convenience function.
/// Like [`NDArrayObject::alloca_uninitialized`] but directly takes the typechecker type of the ndarray.
pub fn alloca_uninitialized_of_type<G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
ndarray_ty: Type,
name: &str,
) -> Self {
let (dtype, ndims) = unpack_ndarray_var_tys(&mut ctx.unifier, ndarray_ty);
let ndims = extract_ndims(&ctx.unifier, ndims);
Self::alloca_uninitialized(generator, ctx, dtype, ndims, name)
}
/// Clone this ndaarray - 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_clone<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
name: &str,
) -> Self {
let clone =
NDArrayObject::alloca_uninitialized(generator, ctx, self.dtype, self.ndims, name);
let shape = self.value.gep(ctx, |f| f.shape).load(generator, ctx, "shape");
clone.copy_shape_from_array(generator, ctx, shape);
clone.create_data(generator, ctx);
clone.copy_data_from(generator, ctx, *self);
clone
}
/// Get this ndarray's `ndims` as an LLVM constant.
pub fn get_ndims<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &'ctx Context,
) -> Int<'ctx, SizeT> {
let sizet_model = IntModel(SizeT);
sizet_model.constant(generator, ctx, self.ndims)
}
/// Get if this ndarray's `np.size` is `0` - containing no content.
pub fn is_empty<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> Int<'ctx, Bool> {
let sizet_model = IntModel(SizeT);
let size = self.size(generator, ctx);
size.compare(ctx, IntPredicate::EQ, sizet_model.const_0(generator, ctx.ctx), "is_empty")
}
/// Return true if this ndarray is unsized - `ndims == 0` and only contains a scalar.
///
/// This is a staticially known property of ndarrays. This is why it is returning
/// a Rust boolean instead of a [`BasicValue`].
#[must_use]
pub fn is_unsized(&self) -> bool {
self.ndims == 0
}
/// If this ndarray is unsized, return its sole value as a [`ScalarObject`]. Otherwise, do nothing.
pub fn split_unsized<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> ScalarOrNDArray<'ctx> {
if self.is_unsized() {
// NOTE: `np.size(self) == 0` here is never possible.
let sizet_model = IntModel(SizeT);
let zero = sizet_model.const_0(generator, ctx.ctx);
ScalarOrNDArray::Scalar(self.get_nth(generator, ctx, zero))
} else {
ScalarOrNDArray::NDArray(*self)
}
}
/// Initialize an ndarray's `data` by allocating a buffer on the stack.
/// The allocated data buffer is considered to be *owned* by the ndarray.
///
/// `strides` of the ndarray will also be updated with `set_strides_by_shape`.
///
/// `shape` and `itemsize` of the ndarray ***must*** be initialized first.
pub fn create_data<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) {
let byte_model = IntModel(Byte);
let nbytes = self.nbytes(generator, ctx);
let data = byte_model.array_alloca(generator, ctx, nbytes.value, "data");
self.value.set(ctx, |f| f.data, data);
self.update_strides_by_shape(generator, ctx);
}
/// Copy shape dimensions from an array.
pub fn copy_shape_from_array<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
src_shape: Ptr<'ctx, IntModel<SizeT>>,
) {
let dst_shape = self.value.get(generator, ctx, |f| f.shape, "dst_shape");
let num_items = self.get_ndims(generator, ctx.ctx).value;
call_memcpy_model(generator, ctx, dst_shape, src_shape, num_items);
}
/// Copy shape dimensions from an ndarray.
/// Panics if `ndims` mismatches.
pub fn copy_shape_from_ndarray<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
src_ndarray: NDArrayObject<'ctx>,
) {
assert_eq!(self.ndims, src_ndarray.ndims);
let src_shape = src_ndarray.value.get(generator, ctx, |f| f.shape, "src_shape");
self.copy_shape_from_array(generator, ctx, src_shape);
}
/// Copy strides dimensions from an array.
pub fn copy_strides_from_array<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
src_strides: Ptr<'ctx, IntModel<SizeT>>,
) {
let dst_strides = self.value.get(generator, ctx, |f| f.strides, "dst_strides");
let num_items = self.get_ndims(generator, ctx.ctx).value;
call_memcpy_model(generator, ctx, dst_strides, src_strides, num_items);
}
/// Copy strides dimensions from an ndarray.
/// Panics if `ndims` mismatches.
pub fn copy_strides_from_ndarray<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
src_ndarray: NDArrayObject<'ctx>,
) {
assert_eq!(self.ndims, src_ndarray.ndims);
let src_strides = src_ndarray.value.get(generator, ctx, |f| f.strides, "src_strides");
self.copy_strides_from_array(generator, ctx, src_strides);
}
/// Iterate through every element pointer in the ndarray in its flatten view.
///
/// `body` also access to [`BreakContinueHooks`] to short-circuit and an element's
/// index. The given element pointer also has been casted to the LLVM type of this ndarray's `dtype`.
pub fn foreach_pointer<'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>,
Int<'ctx, SizeT>,
PointerValue<'ctx>,
) -> Result<(), String>,
{
let sizet_model = IntModel(SizeT);
let start = sizet_model.const_0(generator, ctx.ctx);
let stop = self.size(generator, ctx);
let step = sizet_model.const_1(generator, ctx.ctx);
gen_for_model_auto(generator, ctx, start, stop, step, |generator, ctx, hooks, i| {
let pelement = self.get_nth_pointer(generator, ctx, i, "element");
body(generator, ctx, hooks, i, pelement)
})
}
/// Iterate through every scalar in this ndarray.
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>,
Int<'ctx, SizeT>,
ScalarObject<'ctx>,
) -> Result<(), String>,
{
self.foreach_pointer(generator, ctx, |generator, ctx, hooks, i, p| {
let value = ctx.builder.build_load(p, "value").unwrap();
let scalar = ScalarObject { dtype: self.dtype, value };
body(generator, ctx, hooks, i, scalar)
})
}
/// Make sure the ndarray is at least `ndmin`-dimensional.
///
/// If this ndarray's `ndims` is less than `ndmin`, a view is created on this with 1s prepended to the shape.
/// If this ndarray's `ndims` is not less than `ndmin`, this function does nothing and return this ndarray.
#[must_use]
pub fn atleast_nd<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
ndmin: u64,
) -> Self {
if self.ndims < ndmin {
let mut indices = vec![];
for _ in self.ndims..ndmin {
indices.push(RustNDIndex::NewAxis);
}
indices.push(RustNDIndex::Ellipsis);
self.index(generator, ctx, &indices, "atleast_nd_ndarray")
} else {
*self
}
}
/// Fill the ndarray with a value.
///
/// `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, '_>,
fill_value: BasicValueEnum<'ctx>,
) {
self.foreach_pointer(generator, ctx, |_generator, ctx, _hooks, _i, pelement| {
ctx.builder.build_store(pelement, fill_value).unwrap();
Ok(())
})
.unwrap();
}
/// Create a reshaped view on this ndarray like `np.reshape()`.
///
/// If there is a `-1` in `new_shape`, it will be resolved; `new_shape` would **NOT** be modified as a result.
///
/// If reshape without copying is impossible, this function will allocate a new ndarray and copy contents.
///
/// * `new_ndims` - The number of dimensions of `new_shape` as a [`Type`].
/// * `new_shape` - The target shape to do `np.reshape()`.
#[must_use]
pub fn reshape_or_copy<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
new_ndims: u64,
new_shape: Ptr<'ctx, IntModel<SizeT>>,
) -> Self {
// TODO: The current criterion for whether to do a full copy or not is by checking `is_c_contiguous`,
// but this is not optimal. 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_uninitialized(
generator,
ctx,
self.dtype,
new_ndims,
"reshaped_ndarray",
);
dst_ndarray.copy_shape_from_array(generator, ctx, new_shape);
let size = self.size(generator, ctx);
let new_ndims = dst_ndarray.get_ndims(generator, ctx.ctx);
call_nac3_ndarray_resolve_and_check_new_shape(generator, ctx, size, new_ndims, new_shape);
let is_c_contiguous = self.is_c_contiguous(generator, ctx);
ctx.builder.build_conditional_branch(is_c_contiguous.value, then_bb, else_bb).unwrap();
// Inserting into then_bb: reshape is possible without copying
ctx.builder.position_at_end(then_bb);
dst_ndarray.update_strides_by_shape(generator, ctx);
dst_ndarray.value.set(ctx, |f| f.data, self.value.get(generator, ctx, |f| f.data, "data"));
ctx.builder.build_unconditional_branch(end_bb).unwrap();
// Inserting into else_bb: reshape is impossible without copying
ctx.builder.position_at_end(else_bb);
dst_ndarray.create_data(generator, ctx);
dst_ndarray.copy_data_from(generator, ctx, *self);
ctx.builder.build_unconditional_branch(end_bb).unwrap();
// Reposition for continuation
ctx.builder.position_at_end(end_bb);
dst_ndarray
}
/// Create a flattened view of this ndarray, like `np.ravel()`.
///
/// Uses [`NDArrayObject::reshape_or_copy`] under-the-hood so ndarray may or may not be copied.
#[must_use]
pub fn ravel_or_copy<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> Self {
// Define models
let sizet_model = IntModel(SizeT);
let num0 = sizet_model.const_0(generator, ctx.ctx);
let num1 = sizet_model.const_1(generator, ctx.ctx);
let num_neg1 = sizet_model.const_all_1s(generator, ctx.ctx);
// Create `[-1]` and pass to `reshape_or_copy`.
let new_shape = sizet_model.array_alloca(generator, ctx, num1.value, "new_shape");
new_shape.offset(generator, ctx, num0.value, "").store(ctx, num_neg1);
self.reshape_or_copy(generator, ctx, 1, new_shape)
}
/// Create a transposed view on this ndarray like `np.transpose(<ndarray>, <axes> = None)`.
/// * `axes` - If specified, should be an array of the permutation (negative indices are **allowed**).
#[must_use]
pub fn transpose<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
axes: Option<Ptr<'ctx, IntModel<SizeT>>>,
) -> Self {
// Define models
let sizet_model = IntModel(SizeT);
let transposed_ndarray = NDArrayObject::alloca_uninitialized(
generator,
ctx,
self.dtype,
self.ndims,
"transposed_ndarray",
);
let num_axes = self.get_ndims(generator, ctx.ctx);
// `axes = nullptr` if `axes` is unspecified.
let axes = axes.unwrap_or_else(|| PtrModel(sizet_model).nullptr(generator, ctx.ctx));
call_nac3_ndarray_transpose(
generator,
ctx,
self.value,
transposed_ndarray.value,
num_axes,
axes,
);
transposed_ndarray
}
/// Check if this NDArray can be used as an `out` ndarray for an operation.
///
/// Raise an exception if the shapes do not match.
pub fn check_can_be_written_by_out<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
out_ndims: u64,
out_shape: Ptr<'ctx, IntModel<SizeT>>,
) {
let sizet_model = IntModel(SizeT);
let ndarray_ndims = self.get_ndims(generator, ctx.ctx);
let ndarray_shape = self.value.get(generator, ctx, |f| f.shape, "shape");
let output_ndims = sizet_model.constant(generator, ctx.ctx, out_ndims);
let output_shape = out_shape;
call_nac3_ndarray_util_assert_output_shape_same(
generator,
ctx,
ndarray_ndims,
ndarray_shape,
output_ndims,
output_shape,
);
}
}
/// TODO: Document me
#[derive(Debug, Clone, Copy)]
pub enum NDArrayOut<'ctx> {
NewNDArray { dtype: Type },
WriteToNDArray { ndarray: NDArrayObject<'ctx> },
}

159
nac3core/src/codegen/object/ndarray/product.rs Normal file
View File

@ -0,0 +1,159 @@
use std::cmp::max;
use crate::codegen::{
irrt::{
call_nac3_ndarray_float64_matmul_at_least_2d, call_nac3_ndarray_matmul_calculate_shapes,
},
model::*,
object::ndarray::indexing::RustNDIndex,
CodeGenContext, CodeGenerator,
};
use super::{NDArrayObject, NDArrayOut};
impl<'ctx> NDArrayObject<'ctx> {
/// TODO: Document me
fn np_matmul_helper<G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
a: Self,
b: Self,
) -> Self {
assert!(a.ndims >= 2);
assert!(b.ndims >= 2);
assert!(ctx.unifier.unioned(ctx.primitives.float, a.dtype));
assert!(ctx.unifier.unioned(ctx.primitives.float, b.dtype));
let sizet_model = IntModel(SizeT);
let final_ndims_int = max(a.ndims, b.ndims);
let a_ndims = a.get_ndims(generator, ctx.ctx);
let a_shape = a.value.get(generator, ctx, |f| f.shape, "a_shape");
let b_ndims = b.get_ndims(generator, ctx.ctx);
let b_shape = b.value.get(generator, ctx, |f| f.shape, "b_shape");
let final_ndims = sizet_model.constant(generator, ctx.ctx, final_ndims_int);
let new_a_shape =
sizet_model.array_alloca(generator, ctx, final_ndims.value, "new_a_shape");
let new_b_shape =
sizet_model.array_alloca(generator, ctx, final_ndims.value, "new_b_shape");
let dst_shape = sizet_model.array_alloca(generator, ctx, final_ndims.value, "dst_shape");
call_nac3_ndarray_matmul_calculate_shapes(
generator,
ctx,
a_ndims,
a_shape,
b_ndims,
b_shape,
final_ndims,
new_a_shape,
new_b_shape,
dst_shape,
);
let new_a = a.broadcast_to(generator, ctx, final_ndims_int, new_a_shape);
let new_b = b.broadcast_to(generator, ctx, final_ndims_int, new_b_shape);
let dst = NDArrayObject::alloca_uninitialized(
generator,
ctx,
ctx.primitives.float,
final_ndims_int,
"matmul_result",
);
dst.copy_shape_from_array(generator, ctx, dst_shape);
dst.create_data(generator, ctx);
call_nac3_ndarray_float64_matmul_at_least_2d(
generator,
ctx,
new_a.value,
new_b.value,
dst.value,
);
dst
}
/// Perform `np.matmul` according to the rules in
/// <https://numpy.org/doc/stable/reference/generated/numpy.matmul.html>.
///
/// This function always return an [`NDArrayObject`]. You may want to use [`NDArrayObject::split_unsized`].
pub fn matmul<G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
a: Self,
b: Self,
out: NDArrayOut<'ctx>,
) -> Self {
// Sanity check, but type inference should prevent this.
assert!(a.ndims > 0 && b.ndims > 0, "np.matmul disallows scalar input");
/*
If both arguments are 2-D they are multiplied like conventional matrices.
If either argument is N-D, N > 2, it is treated as a stack of matrices residing in the last two indexes and broadcast accordingly.
If the first argument is 1-D, it is promoted to a matrix by prepending a 1 to its dimensions. After matrix multiplication the prepended 1 is removed.
If the second argument is 1-D, it is promoted to a matrix by appending a 1 to its dimensions. After matrix multiplication the appended 1 is removed.
*/
let new_a = if a.ndims == 1 {
// Prepend 1 to its dimensions
a.index(generator, ctx, &[RustNDIndex::NewAxis, RustNDIndex::Ellipsis], "new_a")
} else {
a
};
let new_b = if b.ndims == 1 {
// Append 1 to its dimensions
b.index(generator, ctx, &[RustNDIndex::Ellipsis, RustNDIndex::NewAxis], "new_a")
} else {
b
};
// NOTE: `result` will always be a newly allocated ndarray.
// Current implementation cannot do in-place matrix muliplication.
let mut result = NDArrayObject::np_matmul_helper(generator, ctx, new_a, new_b);
let i32_model = IntModel(Int32); // TODO: Upgrade to SizeT
let zero = i32_model.const_0(generator, ctx.ctx);
if a.ndims == 1 {
// Remove the prepended 1
result = result.index(
generator,
ctx,
&[RustNDIndex::SingleElement(zero), RustNDIndex::Ellipsis],
"result_no_prepend_1",
);
}
if b.ndims == 1 {
// Remove the appended 1
result = result.index(
generator,
ctx,
&[RustNDIndex::Ellipsis, RustNDIndex::SingleElement(zero)],
"result_no_append_1",
);
}
match out {
NDArrayOut::NewNDArray { dtype } => {
// We don't support auto-casting right now, nor anything other than float64.
// Force the output dtype to be float64.
assert!(ctx.unifier.unioned(ctx.primitives.float, dtype));
result
}
NDArrayOut::WriteToNDArray { ndarray: out_ndarray } => {
// TODO: It is possible to check the shapes before computing the matmul to save resources.
let result_shape = result.value.get(generator, ctx, |f| f.shape, "result_shape");
out_ndarray.check_can_be_written_by_out(generator, ctx, result.ndims, result_shape);
// TODO: We can just set `out_ndarray.data` to `result.data`. Should we?
out_ndarray.copy_data_from(generator, ctx, result);
out_ndarray
}
}
}
}

143
nac3core/src/codegen/object/ndarray/scalar.rs Normal file
View File

@ -0,0 +1,143 @@
use inkwell::values::{BasicValue, BasicValueEnum};
use crate::{
codegen::{model::*, CodeGenContext, CodeGenerator},
typecheck::typedef::{Type, TypeEnum},
};
use super::NDArrayObject;
/// An LLVM numpy scalar with its [`Type`].
///
/// Intended to be used with [`ScalarOrNDArray`].
///
/// A scalar does not have to be an actual number. It could be arbitrary objects.
#[derive(Debug, Clone, Copy)]
pub struct ScalarObject<'ctx> {
pub dtype: Type,
pub value: BasicValueEnum<'ctx>,
}
impl<'ctx> ScalarObject<'ctx> {
/// Promote this scalar to an unsized ndarray (like doing `np.asarray`).
///
/// The scalar value is allocated onto the stack, and the ndarray's `data` will point to that
/// allocated value.
pub fn as_ndarray<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> NDArrayObject<'ctx> {
let pbyte_model = PtrModel(IntModel(Byte));
// We have to put the value on the stack to get a data pointer.
let data = ctx.builder.build_alloca(self.value.get_type(), "as_ndarray_scalar").unwrap();
ctx.builder.build_store(data, self.value).unwrap();
let data = pbyte_model.pointer_cast(generator, ctx, data, "data");
let ndarray =
NDArrayObject::alloca_uninitialized(generator, ctx, self.dtype, 0, "scalar_ndarray");
ndarray.value.set(ctx, |f| f.data, data);
ndarray
}
}
/// A convenience enum for implementing scalar/ndarray agnostic utilities.
#[derive(Debug, Clone, Copy)]
pub enum ScalarOrNDArray<'ctx> {
Scalar(ScalarObject<'ctx>),
NDArray(NDArrayObject<'ctx>),
}
impl<'ctx> ScalarOrNDArray<'ctx> {
/// Get the underlying [`BasicValueEnum<'ctx>`] of this [`ScalarOrNDArray`].
#[must_use]
pub fn to_basic_value_enum(self) -> BasicValueEnum<'ctx> {
match self {
ScalarOrNDArray::Scalar(scalar) => scalar.value,
ScalarOrNDArray::NDArray(ndarray) => ndarray.value.value.as_basic_value_enum(),
}
}
#[must_use]
pub fn into_scalar(&self) -> ScalarObject<'ctx> {
match self {
ScalarOrNDArray::NDArray(_ndarray) => panic!("Got NDArray"),
ScalarOrNDArray::Scalar(scalar) => *scalar,
}
}
#[must_use]
pub fn into_ndarray(&self) -> NDArrayObject<'ctx> {
match self {
ScalarOrNDArray::NDArray(ndarray) => *ndarray,
ScalarOrNDArray::Scalar(_scalar) => panic!("Got Scalar"),
}
}
/// Convert this [`ScalarOrNDArray`] to an ndarray - behaves like `np.asarray`.
/// - If this is an ndarray, the ndarray is returned.
/// - If this is a scalar, an unsized ndarray view is created on it.
pub fn as_ndarray<G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
) -> NDArrayObject<'ctx> {
match self {
ScalarOrNDArray::NDArray(ndarray) => *ndarray,
ScalarOrNDArray::Scalar(scalar) => scalar.as_ndarray(generator, ctx),
}
}
#[must_use]
pub fn dtype(&self) -> Type {
match self {
ScalarOrNDArray::Scalar(scalar) => scalar.dtype,
ScalarOrNDArray::NDArray(ndarray) => ndarray.dtype,
}
}
}
impl<'ctx> TryFrom<&ScalarOrNDArray<'ctx>> for ScalarObject<'ctx> {
type Error = ();
fn try_from(value: &ScalarOrNDArray<'ctx>) -> Result<Self, Self::Error> {
match value {
ScalarOrNDArray::Scalar(scalar) => Ok(*scalar),
ScalarOrNDArray::NDArray(_ndarray) => Err(()),
}
}
}
impl<'ctx> TryFrom<&ScalarOrNDArray<'ctx>> for NDArrayObject<'ctx> {
type Error = ();
fn try_from(value: &ScalarOrNDArray<'ctx>) -> Result<Self, Self::Error> {
match value {
ScalarOrNDArray::Scalar(_scalar) => Err(()),
ScalarOrNDArray::NDArray(ndarray) => Ok(*ndarray),
}
}
}
/// Split an [`BasicValueEnum<'ctx>`] into a [`ScalarOrNDArray`] depending
/// on its [`Type`].
pub fn split_scalar_or_ndarray<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
input: BasicValueEnum<'ctx>,
input_ty: Type,
) -> ScalarOrNDArray<'ctx> {
match &*ctx.unifier.get_ty(input_ty) {
TypeEnum::TObj { obj_id, .. }
if *obj_id == ctx.primitives.ndarray.obj_id(&ctx.unifier).unwrap() =>
{
let ndarray = NDArrayObject::from_value_and_type(generator, ctx, input, input_ty);
ScalarOrNDArray::NDArray(ndarray)
}
_ => {
let scalar = ScalarObject { dtype: input_ty, value: input };
ScalarOrNDArray::Scalar(scalar)
}
}
}

112
nac3core/src/codegen/object/ndarray/shape_util.rs Normal file
View File

@ -0,0 +1,112 @@
use inkwell::values::BasicValueEnum;
use util::gen_for_model_auto;
use crate::{
codegen::{model::*, object::list::ListObject, CodeGenContext, CodeGenerator},
typecheck::typedef::{Type, 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: BasicValueEnum<'ctx>,
input_sequence_ty: Type,
) -> (Int<'ctx, SizeT>, Ptr<'ctx, IntModel<SizeT>>) {
let sizet_model = IntModel(SizeT);
let zero = sizet_model.const_0(generator, ctx.ctx);
let one = sizet_model.const_1(generator, ctx.ctx);
// The result `list` to return.
match &*ctx.unifier.get_ty(input_sequence_ty) {
TypeEnum::TObj { obj_id, .. }
if *obj_id == ctx.primitives.list.obj_id(&ctx.unifier).unwrap() =>
{
// 1. A list of `int32`; e.g., `np.empty([600, 800, 3])`
// Check `input_sequence`
let input_sequence =
ListObject::from_value_and_type(generator, ctx, input_sequence, input_sequence_ty);
let len = input_sequence.value.gep(ctx, |f| f.len).load(generator, ctx, "len");
let result = sizet_model.array_alloca(generator, ctx, len.value, "int_sequence");
// Load all the `int32`s from the input_sequence, cast them to `SizeT`, and store them into `result`
gen_for_model_auto(generator, ctx, zero, len, one, |generator, ctx, _hooks, i| {
// Load the i-th int32 in the input sequence
let int = input_sequence
.value
.get(generator, ctx, |f| f.items, "int")
.ix(generator, ctx, i.value, "int")
.value
.into_int_value();
// Cast to SizeT
let int = sizet_model.s_extend_or_bit_cast(generator, ctx, int, "int");
// Store
result.offset(generator, ctx, i.value, "int").store(ctx, int);
Ok(())
})
.unwrap();
(len, result)
}
TypeEnum::TTuple { ty: tuple_types, .. } => {
// 2. A tuple of ints; e.g., `np.empty((600, 800, 3))`
let input_sequence = input_sequence.into_struct_value(); // A tuple is a struct
let len_int = tuple_types.len();
let len = sizet_model.constant(generator, ctx.ctx, len_int as u64);
let result = sizet_model.array_alloca(generator, ctx, len.value, "int_sequence");
for i in 0..len_int {
// Get the i-th element off of the tuple and load it into `result`.
let int = ctx
.builder
.build_extract_value(input_sequence, i as u32, "int")
.unwrap()
.into_int_value();
let int = sizet_model.s_extend_or_bit_cast(generator, ctx, int, "int");
let offset = sizet_model.constant(generator, ctx.ctx, i as u64);
result.offset(generator, ctx, offset.value, "int").store(ctx, int);
}
(len, result)
}
TypeEnum::TObj { obj_id, .. }
if *obj_id == ctx.primitives.int32.obj_id(&ctx.unifier).unwrap() =>
{
// 3. A scalar int; e.g., `np.empty(3)`, this is functionally equivalent to `np.empty([3])`
let input_int = input_sequence.into_int_value();
let len = sizet_model.const_1(generator, ctx.ctx);
let result = sizet_model.array_alloca(generator, ctx, len.value, "int_sequence");
let int = sizet_model.s_extend_or_bit_cast(generator, ctx, input_int, "int");
// Storing into result[0]
result.store(ctx, int);
(len, result)
}
_ => panic!(
"encountered unknown sequence type: {}",
ctx.unifier.stringify(input_sequence_ty)
),
}
}

39
nac3core/src/codegen/object/tuple.rs Normal file
View File

@ -0,0 +1,39 @@
use inkwell::values::{BasicValueEnum, StructValue};
use itertools::Itertools;
use crate::{
codegen::{CodeGenContext, CodeGenerator},
typecheck::typedef::Type,
};
pub struct TupleObject<'ctx> {
pub tys: Vec<Type>,
pub value: StructValue<'ctx>,
}
impl<'ctx> TupleObject<'ctx> {
pub fn create<I, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
items: I,
name: &str,
) -> Self
where
I: IntoIterator<Item = (BasicValueEnum<'ctx>, Type)>,
{
let (vals, tys): (Vec<_>, Vec<_>) = items.into_iter().unzip();
// let tuple_ty = ctx.unifier.add_ty(TypeEnum::TTuple { ty: tys });
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 vals.into_iter().enumerate() {
// Store the dim value into the tuple
let pval = ctx.builder.build_struct_gep(pllvm_tuple, i as u32, "value").unwrap();
ctx.builder.build_store(pval, val).unwrap();
}
let value = ctx.builder.build_load(pllvm_tuple, name).unwrap().into_struct_value();
TupleObject { tys, value }
}
}

137
nac3core/src/codegen/stmt.rs
View File

@ -1,8 +1,13 @@
use super::model::*;
use super::object::ndarray::indexing::util::gen_ndarray_subscript_ndindexes;
use super::object::ndarray::scalar::split_scalar_or_ndarray;
use super::object::ndarray::NDArrayObject;
use super::{
super::symbol_resolver::ValueEnum,
expr::destructure_range,
irrt::{handle_slice_indices, list_slice_assignment},
CodeGenContext, CodeGenerator,
structure::{CSlice, Exception},
CodeGenContext, CodeGenerator, Int32, IntModel, Ptr, StructModel,
};
use crate::{
codegen::{
@ -401,7 +406,43 @@ 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 = NDArrayObject::from_value_and_type(generator, ctx, target, target_ty);
// Process key
let key = gen_ndarray_subscript_ndindexes(generator, ctx, key)?;
// Process value
let value = value.to_basic_value_enum(ctx, generator, value_ty)?;
/*
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 = target.index(generator, ctx, &key, "assign_target_ndarray");
let value =
split_scalar_or_ndarray(generator, ctx, value, value_ty).as_ndarray(generator, ctx);
let broadcast_result = NDArrayObject::broadcast_all(generator, ctx, &[target, value]);
let target = broadcast_result.ndarrays[0];
let value = broadcast_result.ndarrays[1];
target.copy_data_from(generator, ctx, value);
}
_ => {
panic!("encountered unknown target type: {}", ctx.unifier.stringify(target_ty));
@ -638,8 +679,12 @@ where
I: Clone,
InitFn: FnOnce(&mut G, &mut CodeGenContext<'ctx, 'a>) -> Result<I, String>,
CondFn: FnOnce(&mut G, &mut CodeGenContext<'ctx, 'a>, I) -> Result<IntValue<'ctx>, String>,
BodyFn:
FnOnce(&mut G, &mut CodeGenContext<'ctx, 'a>, BreakContinueHooks, I) -> Result<(), String>,
BodyFn: FnOnce(
&mut G,
&mut CodeGenContext<'ctx, 'a>,
BreakContinueHooks<'ctx>,
I,
) -> Result<(), String>,
UpdateFn: FnOnce(&mut G, &mut CodeGenContext<'ctx, 'a>, I) -> Result<(), String>,
{
let label = label.unwrap_or("for");
@ -719,7 +764,7 @@ where
BodyFn: FnOnce(
&mut G,
&mut CodeGenContext<'ctx, 'a>,
BreakContinueHooks,
BreakContinueHooks<'ctx>,
IntValue<'ctx>,
) -> Result<(), String>,
{
@ -1259,47 +1304,36 @@ pub fn exn_constructor<'ctx>(
pub fn gen_raise<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
exception: Option<&BasicValueEnum<'ctx>>,
exception: Option<Ptr<'ctx, StructModel<Exception>>>,
loc: Location,
) {
if let Some(exception) = exception {
unsafe {
let int32 = ctx.ctx.i32_type();
let zero = int32.const_zero();
let exception = exception.into_pointer_value();
let file_ptr = ctx
.builder
.build_in_bounds_gep(exception, &[zero, int32.const_int(1, false)], "file_ptr")
.unwrap();
let filename = ctx.gen_string(generator, loc.file.0);
ctx.builder.build_store(file_ptr, filename).unwrap();
let row_ptr = ctx
.builder
.build_in_bounds_gep(exception, &[zero, int32.const_int(2, false)], "row_ptr")
.unwrap();
ctx.builder.build_store(row_ptr, int32.const_int(loc.row as u64, false)).unwrap();
let col_ptr = ctx
.builder
.build_in_bounds_gep(exception, &[zero, int32.const_int(3, false)], "col_ptr")
.unwrap();
ctx.builder.build_store(col_ptr, int32.const_int(loc.column as u64, false)).unwrap();
if let Some(pexn) = exception {
let i32_model = IntModel(Int32);
let cslice_model = StructModel(CSlice);
let current_fun = ctx.builder.get_insert_block().unwrap().get_parent().unwrap();
let fun_name = ctx.gen_string(generator, current_fun.get_name().to_str().unwrap());
let name_ptr = ctx
.builder
.build_in_bounds_gep(exception, &[zero, int32.const_int(4, false)], "name_ptr")
.unwrap();
ctx.builder.build_store(name_ptr, fun_name).unwrap();
}
// Get and store filename
let filename = loc.file.0;
let filename = ctx.gen_string(generator, &String::from(filename)).value;
let filename = cslice_model.check_value(generator, ctx.ctx, filename).unwrap();
pexn.set(ctx, |f| f.filename, filename);
let row = i32_model.constant(generator, ctx.ctx, loc.row as u64);
pexn.set(ctx, |f| f.line, row);
let column = i32_model.constant(generator, ctx.ctx, loc.column as u64);
pexn.set(ctx, |f| f.column, column);
let current_fn = ctx.builder.get_insert_block().unwrap().get_parent().unwrap();
let fn_name = ctx.gen_string(generator, current_fn.get_name().to_str().unwrap());
pexn.set(ctx, |f| f.function, fn_name);
let raise = get_builtins(generator, ctx, "__nac3_raise");
let exception = *exception;
ctx.build_call_or_invoke(raise, &[exception], "raise");
ctx.build_call_or_invoke(raise, &[pexn.value.into()], "raise");
} else {
let resume = get_builtins(generator, ctx, "__nac3_resume");
ctx.build_call_or_invoke(resume, &[], "resume");
}
ctx.builder.build_unreachable().unwrap();
}
@ -1761,30 +1795,41 @@ pub fn gen_stmt<G: CodeGenerator>(
} else {
return Ok(());
};
gen_raise(generator, ctx, Some(&exc), stmt.location);
let pexn_model = PtrModel(StructModel(Exception));
let exn = pexn_model.check_value(generator, ctx.ctx, exc).unwrap();
gen_raise(generator, ctx, Some(exn), stmt.location);
} else {
gen_raise(generator, ctx, None, stmt.location);
}
}
StmtKind::Assert { test, msg, .. } => {
let test = if let Some(v) = generator.gen_expr(ctx, test)? {
v.to_basic_value_enum(ctx, generator, test.custom.unwrap())?
} else {
let byte_model = IntModel(Byte);
let cslice_model = StructModel(CSlice);
let Some(test) = generator.gen_expr(ctx, test)? else {
return Ok(());
};
let test = test.to_basic_value_enum(ctx, generator, ctx.primitives.bool)?;
let test = byte_model.check_value(generator, ctx.ctx, test).unwrap(); // Python `bool` is represented as `i8` in nac3core
// Check `msg`
let err_msg = match msg {
Some(msg) => {
if let Some(v) = generator.gen_expr(ctx, msg)? {
v.to_basic_value_enum(ctx, generator, msg.custom.unwrap())?
} else {
let Some(msg) = generator.gen_expr(ctx, msg)? else {
return Ok(());
}
};
let msg = msg.to_basic_value_enum(ctx, generator, ctx.primitives.str)?;
cslice_model.check_value(generator, ctx.ctx, msg).unwrap()
}
None => ctx.gen_string(generator, "").into(),
};
ctx.make_assert_impl(
generator,
generator.bool_to_i1(ctx, test.into_int_value()),
test.value,
"0:AssertionError",
err_msg,
[None, None, None],

148
nac3core/src/codegen/structure.rs Normal file
View File

@ -0,0 +1,148 @@
use inkwell::context::Context;
use crate::codegen::model::*;
use super::{object::ndarray::NpArrayFields, CodeGenerator};
/// Fields of [`CSlice`]
pub struct CSliceFields<'ctx, F: FieldTraversal<'ctx>> {
/// Pointer to data.
pub base: F::Out<PtrModel<IntModel<Byte>>>,
/// Number of bytes of data.
pub len: F::Out<IntModel<SizeT>>,
}
/// See <https://crates.io/crates/cslice>.
///
/// Additionally, see <https://github.com/m-labs/artiq/blob/b0d2705c385f64b6e6711c1726cd9178f40b598e/artiq/firmware/libeh/eh_artiq.rs>)
/// for ARTIQ-specific notes.
#[derive(Debug, Clone, Copy, Default)]
pub struct CSlice;
impl<'ctx> StructKind<'ctx> for CSlice {
type Fields<F: FieldTraversal<'ctx>> = CSliceFields<'ctx, F>;
fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
Self::Fields { base: traversal.add_auto("base"), len: traversal.add_auto("len") }
}
}
impl StructModel<CSlice> {
/// Create a [`CSlice`].
///
/// `base` and `len` must be LLVM global constants.
pub fn create_const<'ctx, G: CodeGenerator + ?Sized>(
&self,
generator: &mut G,
ctx: &'ctx Context,
base: Ptr<'ctx, IntModel<Byte>>,
len: Int<'ctx, SizeT>,
) -> Struct<'ctx, CSlice> {
let value = self
.0
.get_struct_type(generator, ctx)
.const_named_struct(&[base.value.into(), len.value.into()]);
self.believe_value(value)
}
}
/// The LLVM int type of an Exception ID.
pub type ExceptionId = Int32;
/// Fields of [`Exception<'ctx>`]
///
/// The definition came from `pub struct Exception<'a>` in
/// <https://github.com/m-labs/artiq/blob/master/artiq/firmware/libeh/eh_artiq.rs>.
pub struct ExceptionFields<'ctx, F: FieldTraversal<'ctx>> {
/// nac3core's ID of the exception
pub id: F::Out<IntModel<ExceptionId>>,
/// The name of the file this `Exception` was raised in.
pub filename: F::Out<StructModel<CSlice>>,
/// The line number in the file this `Exception` was raised in.
pub line: F::Out<IntModel<Int32>>,
/// The column number in the file this `Exception` was raised in.
pub column: F::Out<IntModel<Int32>>,
/// The name of the Python function this `Exception` was raised in.
pub function: F::Out<StructModel<CSlice>>,
/// The message of this Exception.
///
/// The message can optionally contain integer parameters `{0}`, `{1}`, and `{2}` in its string,
/// where they will be substituted by `params[0]`, `params[1]`, and `params[2]` respectively (as `int64_t`s).
/// Here is an example:
///
/// ```ignore
/// "Index {0} is out of bounds! List only has {1} element(s)."
/// ```
///
/// In this case, `params[0]` and `params[1]` must be specified, and `params[2]` is ***unused***.
/// Having only 3 parameters is a constraint in ARTIQ.
pub msg: F::Out<StructModel<CSlice>>,
pub params: [F::Out<IntModel<Int64>>; 3],
}
/// nac3core & ARTIQ's Exception
#[derive(Debug, Clone, Copy, Default)]
pub struct Exception;
impl<'ctx> StructKind<'ctx> for Exception {
type Fields<F: FieldTraversal<'ctx>> = ExceptionFields<'ctx, F>;
fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
Self::Fields {
id: traversal.add_auto("id"),
filename: traversal.add_auto("filename"),
line: traversal.add_auto("line"),
column: traversal.add_auto("column"),
function: traversal.add_auto("function"),
msg: traversal.add_auto("msg"),
params: [
traversal.add_auto("params[0]"),
traversal.add_auto("params[1]"),
traversal.add_auto("params[2]"),
],
}
}
}
/// Fields of [`List`]
pub struct ListFields<'ctx, F: FieldTraversal<'ctx>, Item: Model<'ctx>> {
/// Array pointer to content
pub items: F::Out<PtrModel<Item>>,
/// Number of items in the array
pub len: F::Out<IntModel<SizeT>>,
}
/// A list in NAC3.
#[derive(Debug, Clone, Copy, Default)]
pub struct List<Item> {
/// Model of the list items
pub item: Item,
}
impl<'ctx, Item: Model<'ctx>> StructKind<'ctx> for List<Item> {
type Fields<F: FieldTraversal<'ctx>> = ListFields<'ctx, F, Item>;
fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
Self::Fields {
items: traversal.add("data", PtrModel(self.item)),
len: traversal.add_auto("len"),
}
}
}
#[derive(Debug, Clone, Copy, Default)]
pub struct NDArray;
impl<'ctx> StructKind<'ctx> for NDArray {
type Fields<F: FieldTraversal<'ctx>> = NpArrayFields<'ctx, F>;
fn traverse_fields<F: FieldTraversal<'ctx>>(&self, traversal: &mut F) -> Self::Fields<F> {
Self::Fields {
data: traversal.add_auto("data"),
itemsize: traversal.add_auto("itemsize"),
ndims: traversal.add_auto("ndims"),
shape: traversal.add_auto("shape"),
strides: traversal.add_auto("strides"),
}
}
}

595
nac3core/src/toplevel/builtins.rs
View File

@ -15,11 +15,20 @@ use crate::{
codegen::{
builtin_fns,
classes::{ProxyValue, RangeValue},
extern_fns, irrt, llvm_intrinsics,
numpy::*,
numpy_new::{self, gen_ndarray_transpose},
object::ndarray::{
functions::{FloorOrCeil, MinOrMax},
scalar::{split_scalar_or_ndarray, ScalarObject, ScalarOrNDArray},
},
stmt::exn_constructor,
},
symbol_resolver::SymbolValue,
toplevel::{helper::PrimDef, numpy::make_ndarray_ty},
toplevel::{
helper::PrimDef,
numpy::{create_ndims, make_ndarray_ty},
},
typecheck::typedef::{into_var_map, iter_type_vars, TypeVar, VarMap},
};
@ -509,7 +518,16 @@ impl<'a> BuiltinBuilder<'a> {
PrimDef::FunNpArray
| PrimDef::FunNpFull
| PrimDef::FunNpEye
| PrimDef::FunNpIdentity => self.build_ndarray_other_factory_function(prim),
| PrimDef::FunNpIdentity
| PrimDef::FunNpArange => self.build_ndarray_other_factory_function(prim),
PrimDef::FunNpBroadcastTo | PrimDef::FunNpReshape | PrimDef::FunNpTranspose => {
self.build_ndarray_view_function(prim)
}
PrimDef::FunNpSize | PrimDef::FunNpShape | PrimDef::FunNpStrides => {
self.build_ndarray_property_getter_function(prim)
}
PrimDef::FunStr => self.build_str_function(),
@ -576,10 +594,6 @@ 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
@ -1070,16 +1084,34 @@ impl<'a> BuiltinBuilder<'a> {
let arg_ty = fun.0.args[0].ty;
let arg = args[0].1.clone().to_basic_value_enum(ctx, generator, arg_ty)?;
let func = match prim {
PrimDef::FunInt32 => builtin_fns::call_int32,
PrimDef::FunInt64 => builtin_fns::call_int64,
PrimDef::FunUInt32 => builtin_fns::call_uint32,
PrimDef::FunUInt64 => builtin_fns::call_uint64,
PrimDef::FunFloat => builtin_fns::call_float,
PrimDef::FunBool => builtin_fns::call_bool,
let ret_dtype = match prim {
PrimDef::FunInt32 => ctx.primitives.int32,
PrimDef::FunInt64 => ctx.primitives.int64,
PrimDef::FunUInt32 => ctx.primitives.uint32,
PrimDef::FunUInt64 => ctx.primitives.uint64,
PrimDef::FunFloat => ctx.primitives.float,
PrimDef::FunBool => ctx.primitives.bool,
_ => unreachable!(),
};
Ok(Some(func(generator, ctx, (arg_ty, arg))?))
let result = split_scalar_or_ndarray(generator, ctx, arg, arg_ty).map(
generator,
ctx,
ret_dtype,
|generator, ctx, _i, scalar| {
let result = match prim {
PrimDef::FunInt32 => scalar.cast_to_int32(generator, ctx),
PrimDef::FunInt64 => scalar.cast_to_int64(generator, ctx),
PrimDef::FunUInt32 => scalar.cast_to_uint32(generator, ctx),
PrimDef::FunUInt64 => scalar.cast_to_uint64(generator, ctx),
PrimDef::FunFloat => scalar.cast_to_float(ctx),
PrimDef::FunBool => scalar.cast_to_bool(ctx),
_ => unreachable!(),
};
Ok(result.value)
},
)?;
Ok(Some(result.to_basic_value_enum()))
},
)))),
loc: None,
@ -1130,20 +1162,23 @@ impl<'a> BuiltinBuilder<'a> {
let arg_ty = fun.0.args[0].ty;
let arg = args[0].1.clone().to_basic_value_enum(ctx, generator, arg_ty)?;
let ret_elem_ty = size_variant.of_int(&ctx.primitives);
Ok(Some(builtin_fns::call_round(generator, ctx, (arg_ty, arg), ret_elem_ty)?))
let ret_int_dtype = size_variant.of_int(&ctx.primitives);
let result = split_scalar_or_ndarray(generator, ctx, arg, arg_ty).map(
generator,
ctx,
ret_int_dtype,
|generator, ctx, _i, scalar| {
Ok(scalar.round(generator, ctx, ret_int_dtype).value)
},
)?;
Ok(Some(result.to_basic_value_enum()))
}),
)
}
/// Build the functions `ceil()` and `floor()` and their 64 bit variants.
fn build_ceil_floor_function(&mut self, prim: PrimDef) -> TopLevelDef {
#[derive(Clone, Copy)]
enum Kind {
Floor,
Ceil,
}
debug_assert_prim_is_allowed(
prim,
&[PrimDef::FunFloor, PrimDef::FunFloor64, PrimDef::FunCeil, PrimDef::FunCeil64],
@ -1151,10 +1186,10 @@ impl<'a> BuiltinBuilder<'a> {
let (size_variant, kind) = {
match prim {
PrimDef::FunFloor => (SizeVariant::Bits32, Kind::Floor),
PrimDef::FunFloor64 => (SizeVariant::Bits64, Kind::Floor),
PrimDef::FunCeil => (SizeVariant::Bits32, Kind::Ceil),
PrimDef::FunCeil64 => (SizeVariant::Bits64, Kind::Ceil),
PrimDef::FunFloor => (SizeVariant::Bits32, FloorOrCeil::Floor),
PrimDef::FunFloor64 => (SizeVariant::Bits64, FloorOrCeil::Floor),
PrimDef::FunCeil => (SizeVariant::Bits32, FloorOrCeil::Ceil),
PrimDef::FunCeil64 => (SizeVariant::Bits64, FloorOrCeil::Ceil),
_ => unreachable!(),
}
};
@ -1194,12 +1229,15 @@ impl<'a> BuiltinBuilder<'a> {
let arg_ty = fun.0.args[0].ty;
let arg = args[0].1.clone().to_basic_value_enum(ctx, generator, arg_ty)?;
let ret_elem_ty = size_variant.of_int(&ctx.primitives);
let func = match kind {
Kind::Ceil => builtin_fns::call_ceil,
Kind::Floor => builtin_fns::call_floor,
};
Ok(Some(func(generator, ctx, (arg_ty, arg), ret_elem_ty)?))
let result = split_scalar_or_ndarray(generator, ctx, arg, arg_ty).map(
generator,
ctx,
int_sized,
|generator, ctx, _i, scalar| {
Ok(scalar.floor_or_ceil(generator, ctx, kind, int_sized).value)
},
)?;
Ok(Some(result.to_basic_value_enum()))
}),
)
}
@ -1246,9 +1284,9 @@ impl<'a> BuiltinBuilder<'a> {
&[(self.ndarray_factory_fn_shape_arg_tvar.ty, "shape")],
Box::new(move |ctx, obj, fun, args, generator| {
let func = match prim {
PrimDef::FunNpNDArray | PrimDef::FunNpEmpty => gen_ndarray_empty,
PrimDef::FunNpZeros => gen_ndarray_zeros,
PrimDef::FunNpOnes => gen_ndarray_ones,
PrimDef::FunNpNDArray | PrimDef::FunNpEmpty => numpy_new::gen_ndarray_empty,
PrimDef::FunNpZeros => numpy_new::gen_ndarray_zeros,
PrimDef::FunNpOnes => numpy_new::gen_ndarray_ones,
_ => unreachable!(),
};
func(ctx, &obj, fun, &args, generator).map(|val| Some(val.as_basic_value_enum()))
@ -1262,7 +1300,13 @@ impl<'a> BuiltinBuilder<'a> {
fn build_ndarray_other_factory_function(&mut self, prim: PrimDef) -> TopLevelDef {
debug_assert_prim_is_allowed(
prim,
&[PrimDef::FunNpArray, PrimDef::FunNpFull, PrimDef::FunNpEye, PrimDef::FunNpIdentity],
&[
PrimDef::FunNpArray,
PrimDef::FunNpFull,
PrimDef::FunNpEye,
PrimDef::FunNpIdentity,
PrimDef::FunNpArange,
],
);
let PrimitiveStore { int32, bool, ndarray, .. } = *self.primitives;
@ -1304,7 +1348,7 @@ impl<'a> BuiltinBuilder<'a> {
resolver: None,
codegen_callback: Some(Arc::new(GenCall::new(Box::new(
|ctx, obj, fun, args, generator| {
gen_ndarray_array(ctx, &obj, fun, &args, generator)
numpy_new::gen_ndarray_array(ctx, &obj, fun, &args, generator)
.map(|val| Some(val.as_basic_value_enum()))
},
)))),
@ -1323,7 +1367,7 @@ impl<'a> BuiltinBuilder<'a> {
// type variable
&[(self.list_int32, "shape"), (tv.ty, "fill_value")],
Box::new(move |ctx, obj, fun, args, generator| {
gen_ndarray_full(ctx, &obj, fun, &args, generator)
numpy_new::gen_ndarray_full(ctx, &obj, fun, &args, generator)
.map(|val| Some(val.as_basic_value_enum()))
}),
)
@ -1381,6 +1425,152 @@ impl<'a> BuiltinBuilder<'a> {
.map(|val| Some(val.as_basic_value_enum()))
}),
),
PrimDef::FunNpArange => {
// TODO: Support `np.arange(start, stop, step)`
let ndims1 = create_ndims(self.unifier, 1);
let ndarray_float_1d = make_ndarray_ty(
self.unifier,
self.primitives,
Some(self.primitives.float),
Some(ndims1),
);
create_fn_by_codegen(
self.unifier,
&VarMap::new(),
prim.name(),
ndarray_float_1d,
&[(int32, "n")],
Box::new(|ctx, obj, fun, args, generator| {
numpy_new::gen_ndarray_arange(ctx, &obj, fun, &args, generator)
.map(|val| Some(val.as_basic_value_enum()))
}),
)
}
_ => unreachable!(),
}
}
fn build_ndarray_view_function(&mut self, prim: PrimDef) -> TopLevelDef {
debug_assert_prim_is_allowed(
prim,
&[PrimDef::FunNpBroadcastTo, PrimDef::FunNpReshape, PrimDef::FunNpTranspose],
);
match prim {
PrimDef::FunNpBroadcastTo | PrimDef::FunNpReshape => {
// `array_ty` can be ndarrays and arbitrary scalars and objects.
let array_tvar = self.unifier.get_dummy_var();
// The return type is handled by special folding in the type inferencer,
// since the returned `ndims` depends on input shape.
let return_tvar = self.unifier.get_dummy_var();
create_fn_by_codegen(
self.unifier,
&into_var_map([array_tvar, return_tvar]),
prim.name(),
return_tvar.ty,
&[
(array_tvar.ty, "array"),
(self.ndarray_factory_fn_shape_arg_tvar.ty, "shape"),
],
Box::new(move |ctx, obj, fun, args, generator| {
let f = match prim {
PrimDef::FunNpBroadcastTo => numpy_new::gen_ndarray_broadcast_to,
PrimDef::FunNpReshape => numpy_new::gen_ndarray_reshape,
_ => unreachable!(),
};
f(ctx, &obj, fun, &args, generator).map(Some)
}),
)
}
PrimDef::FunNpTranspose => {
// TODO: Allow tuple inputs.
// TODO: Support scalar inputs (difficult)
// TODO: Default values don't work for some reason.
// `axes` should have been `Option[List[int32]]` with default `None`.
// Workaround with some bogus types and values for now.
let axes_ty = self.list_int32;
TopLevelDef::Function {
name: prim.name().into(),
simple_name: prim.simple_name().into(),
signature: self.unifier.add_ty(TypeEnum::TFunc(FunSignature {
args: vec![
FuncArg {
name: "a".into(),
ty: self.primitives.ndarray,
default_value: None,
is_vararg: false,
},
FuncArg {
name: "axes".into(),
ty: axes_ty,
default_value: Some(SymbolValue::OptionNone), // Bogus
is_vararg: false,
},
],
ret: self.primitives.ndarray,
vars: VarMap::new(),
})),
var_id: Vec::default(),
instance_to_symbol: HashMap::default(),
instance_to_stmt: HashMap::default(),
resolver: None,
codegen_callback: Some(Arc::new(GenCall::new(Box::new(
|ctx, obj, fun, args, generator| {
gen_ndarray_transpose(ctx, &obj, fun, &args, generator).map(Some)
},
)))),
loc: None,
}
}
_ => unreachable!(),
}
}
fn build_ndarray_property_getter_function(&mut self, prim: PrimDef) -> TopLevelDef {
debug_assert_prim_is_allowed(
prim,
&[PrimDef::FunNpSize, PrimDef::FunNpShape, PrimDef::FunNpStrides],
);
match prim {
PrimDef::FunNpSize => {
// TODO: Make the return type usize
create_fn_by_codegen(
self.unifier,
&VarMap::new(),
prim.name(),
self.primitives.int32,
&[(self.primitives.ndarray, "a")],
Box::new(|ctx, obj, fun, args, generator| {
numpy_new::gen_ndarray_size(ctx, &obj, fun, &args, generator).map(Some)
}),
)
}
PrimDef::FunNpShape | PrimDef::FunNpStrides => {
// 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;
create_fn_by_codegen(
self.unifier,
&VarMap::new(),
prim.name(),
ret_ty,
&[(self.primitives.ndarray, "a")],
Box::new(move |ctx, obj, fun, args, generator| {
let f = match prim {
PrimDef::FunNpShape => numpy_new::gen_ndarray_shape,
PrimDef::FunNpStrides => numpy_new::gen_ndarray_strides,
_ => unreachable!(),
};
f(ctx, &obj, fun, &args, generator).map(Some)
}),
)
}
_ => unreachable!(),
}
}
@ -1432,12 +1622,22 @@ impl<'a> BuiltinBuilder<'a> {
let arg_ty = fun.0.args[0].ty;
let arg = args[0].1.clone().to_basic_value_enum(ctx, generator, arg_ty)?;
let func = match prim {
PrimDef::FunNpCeil => builtin_fns::call_ceil,
PrimDef::FunNpFloor => builtin_fns::call_floor,
let kind = match prim {
PrimDef::FunNpFloor => FloorOrCeil::Floor,
PrimDef::FunNpCeil => FloorOrCeil::Ceil,
_ => unreachable!(),
};
Ok(Some(func(generator, ctx, (arg_ty, arg), ctx.primitives.float)?))
let result = split_scalar_or_ndarray(generator, ctx, arg, arg_ty).map(
generator,
ctx,
ctx.primitives.float,
move |_generator, ctx, _i, scalar| {
let result = scalar.np_floor_or_ceil(ctx, kind);
Ok(result.value)
},
)?;
Ok(Some(result.to_basic_value_enum()))
}),
)
}
@ -1455,7 +1655,17 @@ impl<'a> BuiltinBuilder<'a> {
Box::new(|ctx, _, fun, args, generator| {
let arg_ty = fun.0.args[0].ty;
let arg = args[0].1.clone().to_basic_value_enum(ctx, generator, arg_ty)?;
Ok(Some(builtin_fns::call_numpy_round(generator, ctx, (arg_ty, arg))?))
let result = split_scalar_or_ndarray(generator, ctx, arg, arg_ty).map(
generator,
ctx,
ctx.primitives.float,
|_generator, ctx, _i, scalar| {
let result = scalar.np_round(ctx);
Ok(result.value)
},
)?;
Ok(Some(result.to_basic_value_enum()))
}),
)
}
@ -1528,16 +1738,21 @@ impl<'a> BuiltinBuilder<'a> {
codegen_callback: Some(Arc::new(GenCall::new(Box::new(
move |ctx, _, fun, args, generator| {
let m_ty = fun.0.args[0].ty;
let n_ty = fun.0.args[1].ty;
let m_val = args[0].1.clone().to_basic_value_enum(ctx, generator, m_ty)?;
let n_ty = fun.0.args[1].ty;
let n_val = args[1].1.clone().to_basic_value_enum(ctx, generator, n_ty)?;
let func = match prim {
PrimDef::FunMin => builtin_fns::call_min,
PrimDef::FunMax => builtin_fns::call_max,
let kind = match prim {
PrimDef::FunMin => MinOrMax::Min,
PrimDef::FunMax => MinOrMax::Max,
_ => unreachable!(),
};
Ok(Some(func(ctx, (m_ty, m_val), (n_ty, n_val))))
let m = ScalarObject { dtype: m_ty, value: m_val };
let n = ScalarObject { dtype: n_ty, value: n_val };
let result = ScalarObject::min_or_max(ctx, kind, m, n);
Ok(Some(result.value))
},
)))),
loc: None,
@ -1579,7 +1794,25 @@ impl<'a> BuiltinBuilder<'a> {
let a_ty = fun.0.args[0].ty;
let a = args[0].1.clone().to_basic_value_enum(ctx, generator, a_ty)?;
Ok(Some(builtin_fns::call_numpy_max_min(generator, ctx, (a_ty, a), prim.name())?))
let a = split_scalar_or_ndarray(generator, ctx, a, a_ty).as_ndarray(generator, ctx);
let result = match prim {
PrimDef::FunNpArgmin => a
.argmin_or_argmax(generator, ctx, MinOrMax::Min)
.value
.as_basic_value_enum(),
PrimDef::FunNpArgmax => a
.argmin_or_argmax(generator, ctx, MinOrMax::Max)
.value
.as_basic_value_enum(),
PrimDef::FunNpMin => {
a.min_or_max(generator, ctx, MinOrMax::Min).value.as_basic_value_enum()
}
PrimDef::FunNpMax => {
a.min_or_max(generator, ctx, MinOrMax::Max).value.as_basic_value_enum()
}
_ => unreachable!(),
};
Ok(Some(result))
}),
)
}
@ -1619,13 +1852,32 @@ impl<'a> BuiltinBuilder<'a> {
let x1_val = args[0].1.clone().to_basic_value_enum(ctx, generator, x1_ty)?;
let x2_val = args[1].1.clone().to_basic_value_enum(ctx, generator, x2_ty)?;
let func = match prim {
PrimDef::FunNpMinimum => builtin_fns::call_numpy_minimum,
PrimDef::FunNpMaximum => builtin_fns::call_numpy_maximum,
let kind = match prim {
PrimDef::FunNpMinimum => MinOrMax::Min,
PrimDef::FunNpMaximum => MinOrMax::Max,
_ => unreachable!(),
};
Ok(Some(func(generator, ctx, (x1_ty, x1_val), (x2_ty, x2_val))?))
let x1 = split_scalar_or_ndarray(generator, ctx, x1_val, x1_ty);
let x2 = split_scalar_or_ndarray(generator, ctx, x2_val, x2_ty);
// NOTE: x1.dtype() and x2.dtype() should be the same
let common_ty = x1.dtype();
let result = ScalarOrNDArray::broadcasting_starmap(
generator,
ctx,
&[x1, x2],
common_ty,
|_generator, ctx, _i, scalars| {
let x1 = scalars[0];
let x2 = scalars[1];
let result = ScalarObject::min_or_max(ctx, kind, x1, x2);
Ok(result.value)
},
)?;
Ok(Some(result.to_basic_value_enum()))
},
)))),
loc: None,
@ -1636,6 +1888,7 @@ impl<'a> BuiltinBuilder<'a> {
fn build_abs_function(&mut self) -> TopLevelDef {
let prim = PrimDef::FunAbs;
let num_ty = self.num_ty; // To move into codegen_callback
TopLevelDef::Function {
name: prim.name().into(),
simple_name: prim.simple_name().into(),
@ -1654,11 +1907,17 @@ impl<'a> BuiltinBuilder<'a> {
instance_to_stmt: HashMap::default(),
resolver: None,
codegen_callback: Some(Arc::new(GenCall::new(Box::new(
|ctx, _, fun, args, generator| {
move |ctx, _, fun, args, generator| {
let n_ty = fun.0.args[0].ty;
let n_val = args[0].1.clone().to_basic_value_enum(ctx, generator, n_ty)?;
Ok(Some(builtin_fns::call_abs(generator, ctx, (n_ty, n_val))?))
let result = split_scalar_or_ndarray(generator, ctx, n_val, n_ty).map(
generator,
ctx,
num_ty.ty,
|_generator, ctx, _i, scalar| Ok(scalar.abs(ctx).value),
)?;
Ok(Some(result.to_basic_value_enum()))
},
)))),
loc: None,
@ -1681,13 +1940,23 @@ impl<'a> BuiltinBuilder<'a> {
let x_ty = fun.0.args[0].ty;
let x_val = args[0].1.clone().to_basic_value_enum(ctx, generator, x_ty)?;
let func = match prim {
PrimDef::FunNpIsInf => builtin_fns::call_numpy_isinf,
PrimDef::FunNpIsNan => builtin_fns::call_numpy_isnan,
let function = match prim {
PrimDef::FunNpIsInf => irrt::call_isnan,
PrimDef::FunNpIsNan => irrt::call_isinf,
_ => unreachable!(),
};
Ok(Some(func(generator, ctx, (x_ty, x_val))?))
let result = split_scalar_or_ndarray(generator, ctx, x_val, x_ty).map(
generator,
ctx,
ctx.primitives.bool,
|generator, ctx, _i, scalar| {
let n = scalar.into_float64(ctx);
let n = function(generator, ctx, n);
Ok(n.as_basic_value_enum())
},
)?;
Ok(Some(result.to_basic_value_enum()))
}),
)
}
@ -1745,49 +2014,58 @@ impl<'a> BuiltinBuilder<'a> {
let arg_ty = fun.0.args[0].ty;
let arg_val = args[0].1.clone().to_basic_value_enum(ctx, generator, arg_ty)?;
let func = match prim {
PrimDef::FunNpSin => builtin_fns::call_numpy_sin,
PrimDef::FunNpCos => builtin_fns::call_numpy_cos,
PrimDef::FunNpTan => builtin_fns::call_numpy_tan,
let result = split_scalar_or_ndarray(generator, ctx, arg_val, arg_ty).map(
generator,
ctx,
ctx.primitives.float,
|_generator, ctx, _i, scalar| {
let n = scalar.into_float64(ctx);
let n = match prim {
PrimDef::FunNpSin => llvm_intrinsics::call_float_sin(ctx, n, None),
PrimDef::FunNpCos => llvm_intrinsics::call_float_cos(ctx, n, None),
PrimDef::FunNpTan => extern_fns::call_tan(ctx, n, None),
PrimDef::FunNpArcsin => builtin_fns::call_numpy_arcsin,
PrimDef::FunNpArccos => builtin_fns::call_numpy_arccos,
PrimDef::FunNpArctan => builtin_fns::call_numpy_arctan,
PrimDef::FunNpArcsin => extern_fns::call_asin(ctx, n, None),
PrimDef::FunNpArccos => extern_fns::call_acos(ctx, n, None),
PrimDef::FunNpArctan => extern_fns::call_atan(ctx, n, None),
PrimDef::FunNpSinh => builtin_fns::call_numpy_sinh,
PrimDef::FunNpCosh => builtin_fns::call_numpy_cosh,
PrimDef::FunNpTanh => builtin_fns::call_numpy_tanh,
PrimDef::FunNpSinh => extern_fns::call_sinh(ctx, n, None),
PrimDef::FunNpCosh => extern_fns::call_cosh(ctx, n, None),
PrimDef::FunNpTanh => extern_fns::call_tanh(ctx, n, None),
PrimDef::FunNpArcsinh => builtin_fns::call_numpy_arcsinh,
PrimDef::FunNpArccosh => builtin_fns::call_numpy_arccosh,
PrimDef::FunNpArctanh => builtin_fns::call_numpy_arctanh,
PrimDef::FunNpArcsinh => extern_fns::call_asinh(ctx, n, None),
PrimDef::FunNpArccosh => extern_fns::call_acosh(ctx, n, None),
PrimDef::FunNpArctanh => extern_fns::call_atanh(ctx, n, None),
PrimDef::FunNpExp => builtin_fns::call_numpy_exp,
PrimDef::FunNpExp2 => builtin_fns::call_numpy_exp2,
PrimDef::FunNpExpm1 => builtin_fns::call_numpy_expm1,
PrimDef::FunNpExp => llvm_intrinsics::call_float_exp(ctx, n, None),
PrimDef::FunNpExp2 => llvm_intrinsics::call_float_exp2(ctx, n, None),
PrimDef::FunNpExpm1 => extern_fns::call_expm1(ctx, n, None),
PrimDef::FunNpLog => builtin_fns::call_numpy_log,
PrimDef::FunNpLog2 => builtin_fns::call_numpy_log2,
PrimDef::FunNpLog10 => builtin_fns::call_numpy_log10,
PrimDef::FunNpLog => llvm_intrinsics::call_float_log(ctx, n, None),
PrimDef::FunNpLog2 => llvm_intrinsics::call_float_log2(ctx, n, None),
PrimDef::FunNpLog10 => llvm_intrinsics::call_float_log10(ctx, n, None),
PrimDef::FunNpSqrt => builtin_fns::call_numpy_sqrt,
PrimDef::FunNpCbrt => builtin_fns::call_numpy_cbrt,
PrimDef::FunNpSqrt => llvm_intrinsics::call_float_sqrt(ctx, n, None),
PrimDef::FunNpCbrt => extern_fns::call_cbrt(ctx, n, None),
PrimDef::FunNpFabs => builtin_fns::call_numpy_fabs,
PrimDef::FunNpRint => builtin_fns::call_numpy_rint,
PrimDef::FunNpFabs => llvm_intrinsics::call_float_fabs(ctx, n, None),
PrimDef::FunNpRint => llvm_intrinsics::call_float_rint(ctx, n, None),
PrimDef::FunSpSpecErf => builtin_fns::call_scipy_special_erf,
PrimDef::FunSpSpecErfc => builtin_fns::call_scipy_special_erfc,
PrimDef::FunSpSpecErf => extern_fns::call_erf(ctx, n, None),
PrimDef::FunSpSpecErfc => extern_fns::call_erfc(ctx, n, None),
PrimDef::FunSpSpecGamma => builtin_fns::call_scipy_special_gamma,
PrimDef::FunSpSpecGammaln => builtin_fns::call_scipy_special_gammaln,
PrimDef::FunSpSpecGamma => irrt::call_gamma(ctx, n),
PrimDef::FunSpSpecGammaln => irrt::call_gammaln(ctx, n),
PrimDef::FunSpSpecJ0 => builtin_fns::call_scipy_special_j0,
PrimDef::FunSpSpecJ1 => builtin_fns::call_scipy_special_j1,
PrimDef::FunSpSpecJ0 => irrt::call_j0(ctx, n),
PrimDef::FunSpSpecJ1 => extern_fns::call_j1(ctx, n, None),
_ => unreachable!(),
};
Ok(Some(func(generator, ctx, (arg_ty, arg_val))?))
_ => unreachable!(),
};
Ok(n.as_basic_value_enum())
},
)?;
Ok(Some(result.to_basic_value_enum()))
}),
)
}
@ -1809,20 +2087,20 @@ impl<'a> BuiltinBuilder<'a> {
let PrimitiveStore { float, int32, .. } = *self.primitives;
// The argument types of the two input arguments are controlled here.
let (x1_ty, x2_ty) = match prim {
// The argument types of the two input arguments + the return type
let (x1_dtype, x2_dtype, ret_dtype) = match prim {
PrimDef::FunNpArctan2
| PrimDef::FunNpCopysign
| PrimDef::FunNpFmax
| PrimDef::FunNpFmin
| PrimDef::FunNpHypot
| PrimDef::FunNpNextAfter => (float, float),
PrimDef::FunNpLdExp => (float, int32),
| PrimDef::FunNpNextAfter => (float, float, float),
PrimDef::FunNpLdExp => (float, int32, float),
_ => unreachable!(),
};
let x1_ty = self.new_type_or_ndarray_ty(x1_ty);
let x2_ty = self.new_type_or_ndarray_ty(x2_ty);
let x1_ty = self.new_type_or_ndarray_ty(x1_dtype);
let x2_ty = self.new_type_or_ndarray_ty(x2_dtype);
let param_ty = &[(x1_ty.ty, "x1"), (x2_ty.ty, "x2")];
let ret_ty = self.unifier.get_fresh_var(None, None);
@ -1851,78 +2129,77 @@ impl<'a> BuiltinBuilder<'a> {
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)?;
let func = match prim {
PrimDef::FunNpArctan2 => builtin_fns::call_numpy_arctan2,
PrimDef::FunNpCopysign => builtin_fns::call_numpy_copysign,
PrimDef::FunNpFmax => builtin_fns::call_numpy_fmax,
PrimDef::FunNpFmin => builtin_fns::call_numpy_fmin,
PrimDef::FunNpLdExp => builtin_fns::call_numpy_ldexp,
PrimDef::FunNpHypot => builtin_fns::call_numpy_hypot,
PrimDef::FunNpNextAfter => builtin_fns::call_numpy_nextafter,
_ => unreachable!(),
};
let x1 = split_scalar_or_ndarray(generator, ctx, x1_val, x1_ty);
let x2 = split_scalar_or_ndarray(generator, ctx, x2_val, x2_ty);
Ok(Some(func(generator, ctx, (x1_ty, x1_val), (x2_ty, x2_val))?))
let result = ScalarOrNDArray::broadcasting_starmap(
generator,
ctx,
&[x1, x2],
ret_dtype,
|_generator, ctx, _i, scalars| {
let x1 = scalars[0];
let x2 = scalars[1];
// TODO: This looks ugly
let result = match prim {
PrimDef::FunNpArctan2 => {
let x1 = x1.into_float64(ctx);
let x2 = x2.into_float64(ctx);
extern_fns::call_atan2(ctx, x1, x2, None).as_basic_value_enum()
}
PrimDef::FunNpCopysign => {
let x1 = x1.into_float64(ctx);
let x2 = x2.into_float64(ctx);
llvm_intrinsics::call_float_copysign(ctx, x1, x2, None)
.as_basic_value_enum()
}
PrimDef::FunNpFmax => {
let x1 = x1.into_float64(ctx);
let x2 = x2.into_float64(ctx);
llvm_intrinsics::call_float_maxnum(ctx, x1, x2, None)
.as_basic_value_enum()
}
PrimDef::FunNpFmin => {
let x1 = x1.into_float64(ctx);
let x2 = x2.into_float64(ctx);
llvm_intrinsics::call_float_minnum(ctx, x1, x2, None)
.as_basic_value_enum()
}
PrimDef::FunNpHypot => {
let x1 = x1.into_float64(ctx);
let x2 = x2.into_float64(ctx);
extern_fns::call_hypot(ctx, x1, x2, None).as_basic_value_enum()
}
PrimDef::FunNpNextAfter => {
let x1 = x1.into_float64(ctx);
let x2 = x2.into_float64(ctx);
extern_fns::call_nextafter(ctx, x1, x2, None)
.as_basic_value_enum()
}
PrimDef::FunNpLdExp => {
let x1 = x1.into_float64(ctx);
let x2 = x2.into_int32(ctx);
extern_fns::call_ldexp(ctx, x1, x2, None).as_basic_value_enum()
}
_ => unreachable!(),
};
Ok(result)
},
)?;
Ok(Some(result.to_basic_value_enum()))
},
)))),
loc: None,
}
}
/// 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`

26
nac3core/src/toplevel/helper.rs
View File

@ -51,6 +51,17 @@ pub enum PrimDef {
FunNpArray,
FunNpEye,
FunNpIdentity,
FunNpArange,
// NumPy view functions
FunNpBroadcastTo,
FunNpReshape,
FunNpTranspose,
// NumPy NDArray property getters
FunNpSize,
FunNpShape,
FunNpStrides,
// Miscellaneous NumPy & SciPy functions
FunNpRound,
@ -99,8 +110,6 @@ pub enum PrimDef {
FunNpLdExp,
FunNpHypot,
FunNpNextAfter,
FunNpTranspose,
FunNpReshape,
// Linalg functions
FunNpDot,
@ -237,6 +246,17 @@ impl PrimDef {
PrimDef::FunNpArray => fun("np_array", None),
PrimDef::FunNpEye => fun("np_eye", None),
PrimDef::FunNpIdentity => fun("np_identity", None),
PrimDef::FunNpArange => fun("np_arange", None),
// NumPy view functions
PrimDef::FunNpBroadcastTo => fun("np_broadcast_to", None),
PrimDef::FunNpReshape => fun("np_reshape", None),
PrimDef::FunNpTranspose => fun("np_transpose", None),
// NumPy NDArray property getters
PrimDef::FunNpSize => fun("np_size", None),
PrimDef::FunNpShape => fun("np_shape", None),
PrimDef::FunNpStrides => fun("np_strides", None),
// Miscellaneous NumPy & SciPy functions
PrimDef::FunNpRound => fun("np_round", None),
@ -285,8 +305,6 @@ impl PrimDef {
PrimDef::FunNpLdExp => fun("np_ldexp", None),
PrimDef::FunNpHypot => fun("np_hypot", None),
PrimDef::FunNpNextAfter => fun("np_nextafter", None),
PrimDef::FunNpTranspose => fun("np_transpose", None),
PrimDef::FunNpReshape => fun("np_reshape", None),
// Linalg functions
PrimDef::FunNpDot => fun("np_dot", None),

1
nac3core/src/toplevel/mod.rs
View File

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

60
nac3core/src/toplevel/numpy.rs
View File

@ -1,11 +1,12 @@
use crate::{
symbol_resolver::SymbolValue,
toplevel::helper::PrimDef,
typecheck::{
type_inferencer::PrimitiveStore,
typedef::{Type, TypeEnum, TypeVarId, Unifier, VarMap},
},
};
use itertools::Itertools;
use itertools::{Either, Itertools};
/// Creates a `ndarray` [`Type`] with the given type arguments.
///
@ -83,3 +84,60 @@ pub fn unpack_ndarray_var_ids(unifier: &mut Unifier, ndarray: Type) -> (TypeVarI
pub fn unpack_ndarray_var_tys(unifier: &mut Unifier, ndarray: Type) -> (Type, Type) {
unpack_ndarray_tvars(unifier, ndarray).into_iter().map(|v| v.1).collect_tuple().unwrap()
}
/// 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)
}
/// Return the ndims after broadcasting ndarrays of different ndims.
///
/// Panics if the input list is empty.
pub fn get_broadcast_all_ndims<I>(ndims: I) -> u64
where
I: IntoIterator<Item = u64>,
{
ndims.into_iter().max().unwrap()
}
pub fn split_scalar_or_ndarray_type(
unifier: &mut Unifier,
primitives: &PrimitiveStore,
ty: Type,
) -> Either<Type, (Type, Type)> {
match &*unifier.get_ty(ty) {
TypeEnum::TObj { obj_id, .. } if *obj_id == primitives.ndarray.obj_id(unifier).unwrap() => {
Either::Right(unpack_ndarray_var_tys(unifier, ty))
}
_ => Either::Left(ty),
}
}
pub fn split_as_ndarray_type(
unifier: &mut Unifier,
primitives: &PrimitiveStore,
ty: Type,
) -> (Type, Type) {
match split_scalar_or_ndarray_type(unifier, primitives, ty) {
Either::Left(dtype) => {
let ndims = unifier.get_fresh_literal(vec![SymbolValue::U64(0)], None);
(dtype, ndims)
}
Either::Right((dtype, ndims)) => (dtype, ndims),
}
}

46
nac3core/src/toplevel/option.rs Normal file
View File

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

2
nac3core/src/typecheck/function_check.rs
View File

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

66
nac3core/src/typecheck/magic_methods.rs
View File

@ -1,6 +1,9 @@
use crate::symbol_resolver::SymbolValue;
use crate::toplevel::helper::PrimDef;
use crate::toplevel::numpy::{make_ndarray_ty, unpack_ndarray_var_tys};
use crate::toplevel::numpy::{
extract_ndims, make_ndarray_ty, split_as_ndarray_type, split_scalar_or_ndarray_type,
unpack_ndarray_var_tys,
};
use crate::typecheck::{
type_inferencer::*,
typedef::{FunSignature, FuncArg, Type, TypeEnum, Unifier, VarMap},
@ -520,36 +523,41 @@ pub fn typeof_binop(
}
Operator::MatMult => {
let (_, lhs_ndims) = unpack_ndarray_var_tys(unifier, lhs);
let lhs_ndims = match &*unifier.get_ty_immutable(lhs_ndims) {
TypeEnum::TLiteral { values, .. } => {
assert_eq!(values.len(), 1);
u64::try_from(values[0].clone()).unwrap()
let (lhs_dtype, lhs_ndims) = split_as_ndarray_type(unifier, primitives, lhs);
let (rhs_dtype, rhs_ndims) = split_as_ndarray_type(unifier, primitives, rhs);
let lhs_ndims = extract_ndims(unifier, lhs_ndims);
let rhs_ndims = extract_ndims(unifier, rhs_ndims);
if !(unifier.unioned(lhs_dtype, primitives.float)
&& unifier.unioned(rhs_dtype, primitives.float))
{
return Err(format!(
"ndarray.__matmul__ only supports float64 operations, but LHS has type {} and RHS has type {}",
unifier.stringify(lhs),
unifier.stringify(rhs)
));
}
let result_ndims = match (lhs_ndims, rhs_ndims) {
(0, _) | (_, 0) => {
return Err(
"ndarray.__matmul__ does not allow unsized ndarray input".to_string()
)
}
_ => unreachable!(),
};
let (_, rhs_ndims) = unpack_ndarray_var_tys(unifier, rhs);
let rhs_ndims = match &*unifier.get_ty_immutable(rhs_ndims) {
TypeEnum::TLiteral { values, .. } => {
assert_eq!(values.len(), 1);
u64::try_from(values[0].clone()).unwrap()
}
_ => unreachable!(),
(1, 1) => 0,
(1, _) => rhs_ndims - 1,
(_, 1) => lhs_ndims - 1,
(m, n) => max(m, n),
};
match (lhs_ndims, rhs_ndims) {
(2, 2) => typeof_ndarray_broadcast(unifier, primitives, lhs, rhs)?,
(lhs, rhs) if lhs == 0 || rhs == 0 => {
return Err(format!(
"Input operand {} does not have enough dimensions (has {lhs}, requires {rhs})",
u8::from(rhs == 0)
))
}
(lhs, rhs) => {
return Err(format!(
"ndarray.__matmul__ on {lhs}D and {rhs}D operands not supported"
))
}
if result_ndims == 0 {
// If the result is unsized, NumPy returns a scalar.
primitives.float
} else {
let result_ndims_ty =
unifier.get_fresh_literal(vec![SymbolValue::U64(result_ndims)], None);
make_ndarray_ty(unifier, primitives, Some(primitives.float), Some(result_ndims_ty))
}
}
@ -748,7 +756,7 @@ pub fn set_primitives_magic_methods(store: &PrimitiveStore, unifier: &mut Unifie
impl_div(unifier, store, ndarray_t, &[ndarray_t, ndarray_dtype_t], None);
impl_floordiv(unifier, store, ndarray_t, &[ndarray_unsized_t, ndarray_unsized_dtype_t], None);
impl_mod(unifier, store, ndarray_t, &[ndarray_unsized_t, ndarray_unsized_dtype_t], None);
impl_matmul(unifier, store, ndarray_t, &[ndarray_t], Some(ndarray_t));
impl_matmul(unifier, store, ndarray_t, &[ndarray_unsized_t], Some(ndarray_t));
impl_sign(unifier, store, ndarray_t, Some(ndarray_t));
impl_invert(unifier, store, ndarray_t, Some(ndarray_t));
impl_eq(unifier, store, ndarray_t, &[ndarray_unsized_t, ndarray_unsized_dtype_t], None);

94
nac3core/src/typecheck/type_inferencer/mod.rs
View File

@ -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::{
@ -12,16 +12,18 @@ use super::{
RecordField, RecordKey, Type, TypeEnum, TypeVar, Unifier, VarMap,
},
};
use crate::toplevel::option::make_option_ty;
use crate::{
symbol_resolver::{SymbolResolver, SymbolValue},
toplevel::{
helper::{arraylike_flatten_element_type, arraylike_get_ndims, PrimDef},
numpy::{make_ndarray_ty, unpack_ndarray_var_tys},
numpy::{extract_ndims, make_ndarray_ty, unpack_ndarray_var_tys},
TopLevelContext, TopLevelDef,
},
typecheck::typedef::Mapping,
};
use itertools::{izip, Itertools};
use nac3parser::ast::Constant;
use nac3parser::ast::{
self,
fold::{self, Fold},
@ -1325,7 +1327,7 @@ impl<'a> Inferencer<'a> {
arg_ty.obj_id(self.unifier).is_some_and(|id| id == PrimDef::NDArray.id())
}) {
// typeof_ndarray_broadcast requires both dtypes to be the same, but ldexp accepts
// (float, int32), so convert it to align with the dtype of the first arg
// (float, int32), so convert it to align with t#he dtype of the first arg
let arg1_ty = if id == &"np_ldexp".into() {
if arg1_ty.obj_id(self.unifier).is_some_and(|id| id == PrimDef::NDArray.id()) {
let (_, ndims) = unpack_ndarray_var_tys(self.unifier, arg1_ty);
@ -1501,7 +1503,7 @@ impl<'a> Inferencer<'a> {
},
}));
}
// 2-argument ndarray n-dimensional factory functions
if id == &"np_reshape".into() && args.len() == 2 {
let arg0 = self.fold_expr(args.remove(0))?;
@ -1546,6 +1548,51 @@ impl<'a> Inferencer<'a> {
},
}));
}
if ["np_shape".into(), "np_strides".into()].contains(id) && args.len() == 1 {
// Output tuple size depends on input ndarray's ndims.
let ndarray = self.fold_expr(args.remove(0))?;
let (_, ndims) = unpack_ndarray_var_tys(self.unifier, ndarray.custom.unwrap());
let ndims = extract_ndims(self.unifier, ndims);
// Create a tuple of size `ndims` full of int32
// TODO: Make it usize
let ret_ty = TypeEnum::TTuple {
ty: iter::repeat(self.primitives.int32).take(ndims as usize).collect_vec(),
is_vararg_ctx: false,
};
let ret_ty = self.unifier.add_ty(ret_ty);
let func_ty = TypeEnum::TFunc(FunSignature {
args: vec![FuncArg {
name: "a".into(),
default_value: None,
ty: ndarray.custom.unwrap(),
is_vararg: false,
}],
ret: ret_ty,
vars: VarMap::new(),
});
let func_ty = self.unifier.add_ty(func_ty);
return Ok(Some(Located {
location,
custom: Some(ret_ty),
node: ExprKind::Call {
func: Box::new(Located {
custom: Some(func_ty),
location: func.location,
node: ExprKind::Name { id: *id, ctx: *ctx },
}),
args: vec![ndarray],
keywords: vec![],
},
}));
}
// 2-argument ndarray n-dimensional creation functions
if id == &"np_full".into() && args.len() == 2 {
let ExprKind::List { elts, .. } = &args[0].node else {
@ -2264,14 +2311,39 @@ impl<'a> Inferencer<'a> {
// We will also take the opportunity to deduce `dims_to_subtract` as well
let mut dims_to_subtract: i128 = 0;
for index in indices {
if let ExprKind::Slice { lower, upper, step } = &index.node {
for v in [lower.as_ref(), upper.as_ref(), step.as_ref()].iter().flatten() {
self.constrain(v.custom.unwrap(), self.primitives.int32, &v.location)?;
match &index.node {
ExprKind::Slice { lower, upper, step } => {
// Handle slices
for v in [lower.as_ref(), upper.as_ref(), step.as_ref()].iter().flatten() {
self.constrain(v.custom.unwrap(), self.primitives.int32, &v.location)?;
}
}
ExprKind::Constant { value: Constant::Ellipsis, .. } => {
// Handle `...`.
// See https://git.m-labs.hk/M-Labs/nac3/issues/486
// Force `...` to have `()` (completely bogus) to make it concrete.
let empty_tuple = TypeEnum::TTuple { ty: vec![], is_vararg_ctx: false };
let empty_tuple = self.unifier.add_ty(empty_tuple);
self.unify(index.custom.unwrap(), empty_tuple, &index.location)?;
}
ExprKind::Name { id, .. } if id == &"none".into() => {
// Handle `np.newaxis` / `None`.
dims_to_subtract -= 1;
// "none" itself has type `Option[T]`, and since we have a stray `T` (non-concrete type).
// We will force the type to be `Option[()]` to make it concrete. (TODO: is there a void type?)
let empty_tuple = TypeEnum::TTuple { ty: vec![], is_vararg_ctx: false };
let empty_tuple = self.unifier.add_ty(empty_tuple);
let expected_type =
make_option_ty(self.unifier, self.primitives, Some(empty_tuple));
self.unify(index.custom.unwrap(), expected_type, &index.location)?;
}
_ => {
// Treat anything else as an integer index, and force unify their type to int32.
dims_to_subtract += 1;
self.unify(index.custom.unwrap(), self.primitives.int32, &index.location)?;
}
} else {
// Treat anything else as an integer index, and force unify their type to int32.
self.unify(index.custom.unwrap(), self.primitives.int32, &index.location)?;
dims_to_subtract += 1;
}
}

8
nac3core/src/typecheck/typedef/mod.rs
View File

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

10
nac3standalone/demo/interpret_demo.py
View File

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

20
nac3standalone/src/main.rs
View File

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