nac3/nac3core/src/codegen/mod.rs

1016 lines
37 KiB
Rust
Raw Normal View History

2021-08-11 14:37:26 +08:00
use crate::{
2021-11-20 19:50:25 +08:00
symbol_resolver::{StaticValue, SymbolResolver},
2024-02-27 13:39:05 +08:00
toplevel::{
helper::PRIMITIVE_DEF_IDS,
numpy::unpack_ndarray_tvars,
TopLevelContext,
TopLevelDef,
},
2021-08-11 14:37:26 +08:00
typecheck::{
type_inferencer::{CodeLocation, PrimitiveStore},
typedef::{CallId, FuncArg, Type, TypeEnum, Unifier},
2021-08-11 14:37:26 +08:00
},
};
use crossbeam::channel::{unbounded, Receiver, Sender};
2021-08-11 14:37:26 +08:00
use inkwell::{
AddressSpace,
IntPredicate,
OptimizationLevel,
attributes::{Attribute, AttributeLoc},
2021-08-11 14:37:26 +08:00
basic_block::BasicBlock,
builder::Builder,
context::Context,
module::Module,
2023-09-11 13:14:35 +08:00
passes::PassBuilderOptions,
targets::{CodeModel, RelocMode, Target, TargetMachine, TargetTriple},
types::{AnyType, BasicType, BasicTypeEnum},
values::{BasicValueEnum, FunctionValue, IntValue, PhiValue, PointerValue},
2022-04-14 15:40:12 +08:00
debug_info::{
DebugInfoBuilder, DICompileUnit, DISubprogram, AsDIScope, DIFlagsConstants, DIScope
},
2021-08-11 14:37:26 +08:00
};
use itertools::Itertools;
use nac3parser::ast::{Stmt, StrRef, Location};
2021-11-20 19:50:25 +08:00
use parking_lot::{Condvar, Mutex};
use std::collections::{HashMap, HashSet};
2021-08-13 16:20:14 +08:00
use std::sync::{
atomic::{AtomicBool, Ordering},
Arc,
};
use std::thread;
2021-08-11 14:37:26 +08:00
pub mod classes;
pub mod concrete_type;
pub mod expr;
mod generator;
pub mod irrt;
pub mod llvm_intrinsics;
pub mod numpy;
2022-02-21 18:27:46 +08:00
pub mod stmt;
2021-08-11 14:37:26 +08:00
2021-08-12 13:55:15 +08:00
#[cfg(test)]
mod test;
use concrete_type::{ConcreteType, ConcreteTypeEnum, ConcreteTypeStore};
pub use generator::{CodeGenerator, DefaultCodeGenerator};
2021-11-20 19:50:25 +08:00
#[derive(Default)]
pub struct StaticValueStore {
pub lookup: HashMap<Vec<(usize, u64)>, usize>,
pub store: Vec<HashMap<usize, Arc<dyn StaticValue + Send + Sync>>>,
}
pub type VarValue<'ctx> = (PointerValue<'ctx>, Option<Arc<dyn StaticValue + Send + Sync>>, i64);
/// Additional options for LLVM during codegen.
#[derive(Clone, Debug, Eq, PartialEq)]
pub struct CodeGenLLVMOptions {
/// The optimization level to apply on the generated LLVM IR.
pub opt_level: OptimizationLevel,
/// Options related to the target machine.
pub target: CodeGenTargetMachineOptions,
}
/// Additional options for code generation for the target machine.
#[derive(Clone, Debug, Eq, PartialEq)]
pub struct CodeGenTargetMachineOptions {
/// The target machine triple.
pub triple: String,
/// The target machine CPU.
pub cpu: String,
/// Additional target machine features.
pub features: String,
/// Relocation mode for code generation.
pub reloc_mode: RelocMode,
/// Code model for code generation.
pub code_model: CodeModel,
}
impl CodeGenTargetMachineOptions {
2023-12-08 17:43:32 +08:00
/// Creates an instance of [`CodeGenTargetMachineOptions`] using the triple of the host machine.
/// Other options are set to defaults.
2023-12-08 17:43:32 +08:00
#[must_use]
pub fn from_host_triple() -> CodeGenTargetMachineOptions {
CodeGenTargetMachineOptions {
triple: TargetMachine::get_default_triple().as_str().to_string_lossy().into_owned(),
cpu: String::default(),
features: String::default(),
reloc_mode: RelocMode::Default,
code_model: CodeModel::Default,
}
}
2023-12-08 17:43:32 +08:00
/// Creates an instance of [`CodeGenTargetMachineOptions`] using the properties of the host
/// machine. Other options are set to defaults.
2023-12-08 17:43:32 +08:00
#[must_use]
pub fn from_host() -> CodeGenTargetMachineOptions {
CodeGenTargetMachineOptions {
cpu: TargetMachine::get_host_cpu_name().to_string(),
features: TargetMachine::get_host_cpu_features().to_string(),
..CodeGenTargetMachineOptions::from_host_triple()
}
}
2023-12-08 17:43:32 +08:00
/// Creates a [`TargetMachine`] using the target options specified by this struct.
///
2023-12-08 17:43:32 +08:00
/// See [`Target::create_target_machine`].
#[must_use]
pub fn create_target_machine(
&self,
level: OptimizationLevel,
) -> Option<TargetMachine> {
let triple = TargetTriple::create(self.triple.as_str());
let target = Target::from_triple(&triple)
2023-12-06 11:49:02 +08:00
.unwrap_or_else(|_| panic!("could not create target from target triple {}", self.triple));
target.create_target_machine(
&triple,
self.cpu.as_str(),
self.features.as_str(),
level,
self.reloc_mode,
self.code_model
)
}
}
2021-08-12 13:55:15 +08:00
pub struct CodeGenContext<'ctx, 'a> {
2023-10-18 13:40:37 +08:00
/// The LLVM context associated with [this context][CodeGenContext].
2021-08-11 14:37:26 +08:00
pub ctx: &'ctx Context,
2023-10-18 13:40:37 +08:00
/// The [Builder] instance for creating LLVM IR statements.
2021-08-11 14:37:26 +08:00
pub builder: Builder<'ctx>,
/// The [DebugInfoBuilder], [compilation unit information][DICompileUnit], and
/// [scope information][DIScope] of this context.
2022-04-14 15:40:12 +08:00
pub debug_info: (DebugInfoBuilder<'ctx>, DICompileUnit<'ctx>, DIScope<'ctx>),
2023-10-18 13:40:37 +08:00
/// The module for which [this context][CodeGenContext] is generating into.
2021-08-11 14:37:26 +08:00
pub module: Module<'ctx>,
2023-10-18 13:40:37 +08:00
/// The [TopLevelContext] associated with [this context][CodeGenContext].
2021-08-12 13:55:15 +08:00
pub top_level: &'a TopLevelContext,
2021-08-11 14:37:26 +08:00
pub unifier: Unifier,
2021-10-16 18:08:13 +08:00
pub resolver: Arc<dyn SymbolResolver + Send + Sync>,
2021-11-20 19:50:25 +08:00
pub static_value_store: Arc<Mutex<StaticValueStore>>,
2023-10-18 13:40:37 +08:00
/// A [HashMap] containing the mapping between the names of variables currently in-scope and
/// its value information.
2021-11-20 19:50:25 +08:00
pub var_assignment: HashMap<StrRef, VarValue<'ctx>>,
2023-10-18 13:40:37 +08:00
///
2021-08-11 14:37:26 +08:00
pub type_cache: HashMap<Type, BasicTypeEnum<'ctx>>,
pub primitives: PrimitiveStore,
2021-09-22 16:04:25 +08:00
pub calls: Arc<HashMap<CodeLocation, CallId>>,
pub registry: &'a WorkerRegistry,
2023-10-18 13:40:37 +08:00
/// Cache for constant strings.
pub const_strings: HashMap<String, BasicValueEnum<'ctx>>,
2023-10-18 13:40:37 +08:00
/// [BasicBlock] containing all `alloca` statements for the current function.
2021-08-11 14:37:26 +08:00
pub init_bb: BasicBlock<'ctx>,
pub exception_val: Option<PointerValue<'ctx>>,
2023-10-18 13:40:37 +08:00
/// The header and exit basic blocks of a loop in this context. See
/// https://llvm.org/docs/LoopTerminology.html for explanation of these terminology.
pub loop_target: Option<(BasicBlock<'ctx>, BasicBlock<'ctx>)>,
2023-10-18 13:40:37 +08:00
/// The target [BasicBlock] to jump to when performing stack unwind.
pub unwind_target: Option<BasicBlock<'ctx>>,
2023-10-18 13:40:37 +08:00
/// The target [BasicBlock] to jump to before returning from the function.
///
/// If this field is [None] when generating a return from a function, `ret` with no argument can
/// be emitted.
pub return_target: Option<BasicBlock<'ctx>>,
2023-10-18 13:40:37 +08:00
/// The [PointerValue] containing the return value of the function.
pub return_buffer: Option<PointerValue<'ctx>>,
2023-10-18 13:40:37 +08:00
// outer catch clauses
pub outer_catch_clauses:
Option<(Vec<Option<BasicValueEnum<'ctx>>>, BasicBlock<'ctx>, PhiValue<'ctx>)>,
2023-10-18 13:40:37 +08:00
/// Whether `sret` is needed for the first parameter of the function.
///
/// See [need_sret].
pub need_sret: bool,
2023-10-18 13:40:37 +08:00
/// The current source location.
pub current_loc: Location,
}
impl<'ctx, 'a> CodeGenContext<'ctx, 'a> {
2023-10-18 13:40:37 +08:00
/// Whether the [current basic block][Builder::get_insert_block] referenced by `builder`
/// contains a [terminator statement][BasicBlock::get_terminator].
pub fn is_terminated(&self) -> bool {
2023-12-08 17:43:32 +08:00
self.builder.get_insert_block().and_then(BasicBlock::get_terminator).is_some()
}
2021-08-11 14:37:26 +08:00
}
type Fp = Box<dyn Fn(&Module) + Send + Sync>;
pub struct WithCall {
fp: Fp,
}
impl WithCall {
2023-12-08 17:43:32 +08:00
#[must_use]
pub fn new(fp: Fp) -> WithCall {
WithCall { fp }
}
2023-12-06 11:49:02 +08:00
pub fn run(&self, m: &Module) {
2023-12-08 17:43:32 +08:00
(self.fp)(m);
}
}
pub struct WorkerRegistry {
sender: Arc<Sender<Option<CodeGenTask>>>,
receiver: Arc<Receiver<Option<CodeGenTask>>>,
2023-10-18 13:40:37 +08:00
/// Whether any thread in this registry has panicked.
2021-08-13 16:20:14 +08:00
panicked: AtomicBool,
2023-10-18 13:40:37 +08:00
/// The total number of tasks queued or completed in the registry.
task_count: Mutex<usize>,
2023-10-18 13:40:37 +08:00
/// The number of threads available for this registry.
thread_count: usize,
wait_condvar: Condvar,
2021-11-20 19:50:25 +08:00
top_level_ctx: Arc<TopLevelContext>,
static_value_store: Arc<Mutex<StaticValueStore>>,
2023-10-18 13:40:37 +08:00
/// LLVM-related options for code generation.
pub llvm_options: CodeGenLLVMOptions,
}
impl WorkerRegistry {
2023-10-18 13:40:37 +08:00
/// Creates workers for this registry.
2023-12-08 17:43:32 +08:00
#[must_use]
pub fn create_workers<G: CodeGenerator + Send + 'static>(
generators: Vec<Box<G>>,
top_level_ctx: Arc<TopLevelContext>,
llvm_options: &CodeGenLLVMOptions,
2023-12-08 17:43:32 +08:00
f: &Arc<WithCall>,
2021-08-13 16:20:14 +08:00
) -> (Arc<WorkerRegistry>, Vec<thread::JoinHandle<()>>) {
let (sender, receiver) = unbounded();
let task_count = Mutex::new(0);
let wait_condvar = Condvar::new();
2021-11-20 19:50:25 +08:00
// init: 0 to be empty
2023-12-08 17:43:32 +08:00
let mut static_value_store = StaticValueStore::default();
static_value_store.lookup.insert(Vec::default(), 0);
static_value_store.store.push(HashMap::default());
2021-11-20 19:50:25 +08:00
let registry = Arc::new(WorkerRegistry {
sender: Arc::new(sender),
receiver: Arc::new(receiver),
thread_count: generators.len(),
2021-08-13 16:20:14 +08:00
panicked: AtomicBool::new(false),
2021-11-20 19:50:25 +08:00
static_value_store: Arc::new(Mutex::new(static_value_store)),
task_count,
wait_condvar,
2021-11-20 19:50:25 +08:00
top_level_ctx,
llvm_options: llvm_options.clone(),
});
2021-08-13 16:20:14 +08:00
let mut handles = Vec::new();
2023-12-08 17:43:32 +08:00
for mut generator in generators {
let registry = registry.clone();
2021-08-13 16:20:14 +08:00
let registry2 = registry.clone();
let f = f.clone();
2021-08-13 16:20:14 +08:00
let handle = thread::spawn(move || {
2023-12-08 17:43:32 +08:00
registry.worker_thread(generator.as_mut(), &f);
});
2021-08-13 16:20:14 +08:00
let handle = thread::spawn(move || {
if let Err(e) = handle.join() {
if let Some(e) = e.downcast_ref::<&'static str>() {
2023-12-08 17:43:32 +08:00
eprintln!("Got an error: {e}");
2021-08-13 16:20:14 +08:00
} else {
2023-12-08 17:43:32 +08:00
eprintln!("Got an unknown error: {e:?}");
2021-08-13 16:20:14 +08:00
}
registry2.panicked.store(true, Ordering::SeqCst);
registry2.wait_condvar.notify_all();
}
});
handles.push(handle);
}
2021-08-13 16:20:14 +08:00
(registry, handles)
}
2021-08-13 16:20:14 +08:00
pub fn wait_tasks_complete(&self, handles: Vec<thread::JoinHandle<()>>) {
{
let mut count = self.task_count.lock();
while *count != 0 {
2021-08-13 16:20:14 +08:00
if self.panicked.load(Ordering::SeqCst) {
break;
}
self.wait_condvar.wait(&mut count);
}
}
for _ in 0..self.thread_count {
self.sender.send(None).unwrap();
}
{
let mut count = self.task_count.lock();
while *count != self.thread_count {
2021-08-13 16:20:14 +08:00
if self.panicked.load(Ordering::SeqCst) {
break;
}
self.wait_condvar.wait(&mut count);
}
}
2021-08-13 16:20:14 +08:00
for handle in handles {
handle.join().unwrap();
}
2023-12-08 17:43:32 +08:00
assert!(!self.panicked.load(Ordering::SeqCst), "tasks panicked");
}
2023-12-08 17:43:32 +08:00
/// Adds a task to this [`WorkerRegistry`].
pub fn add_task(&self, task: CodeGenTask) {
*self.task_count.lock() += 1;
self.sender.send(Some(task)).unwrap();
}
2023-10-18 13:40:37 +08:00
/// Function executed by worker thread for generating IR for each function.
2023-12-08 17:43:32 +08:00
fn worker_thread<G: CodeGenerator>(&self, generator: &mut G, f: &Arc<WithCall>) {
let context = Context::create();
let mut builder = context.create_builder();
let mut module = context.create_module(generator.get_name());
2022-04-14 15:40:12 +08:00
module.add_basic_value_flag(
"Debug Info Version",
inkwell::module::FlagBehavior::Warning,
context.i32_type().const_int(3, false),
);
module.add_basic_value_flag(
"Dwarf Version",
inkwell::module::FlagBehavior::Warning,
context.i32_type().const_int(4, false),
);
let mut errors = HashSet::new();
while let Some(task) = self.receiver.recv().unwrap() {
match gen_func(&context, generator, self, builder, module, task) {
Ok(result) => {
builder = result.0;
module = result.1;
}
Err((old_builder, e)) => {
builder = old_builder;
errors.insert(e);
// create a new empty module just to continue codegen and collect errors
module = context.create_module(&format!("{}_recover", generator.get_name()));
}
}
*self.task_count.lock() -= 1;
self.wait_condvar.notify_all();
}
2023-12-08 17:43:32 +08:00
assert!(errors.is_empty(), "Codegen error: {}", errors.into_iter().sorted().join("\n----------\n"));
let result = module.verify();
if let Err(err) = result {
println!("{}", module.print_to_string().to_str().unwrap());
panic!("{}", err.to_string())
}
2023-09-11 13:14:35 +08:00
let pass_options = PassBuilderOptions::create();
2023-12-06 11:49:02 +08:00
let target_machine = self
.llvm_options
.target
.create_target_machine(self.llvm_options.opt_level)
.unwrap_or_else(|| panic!("could not create target machine from properties {:?}", self.llvm_options.target));
2023-09-11 13:14:35 +08:00
let passes = format!("default<O{}>", self.llvm_options.opt_level as u32);
let result = module.run_passes(passes.as_str(), &target_machine, pass_options);
if let Err(err) = result {
panic!("Failed to run optimization for module `{}`: {}",
module.get_name().to_str().unwrap(),
err.to_string());
}
f.run(&module);
let mut lock = self.task_count.lock();
*lock += 1;
self.wait_condvar.notify_all();
}
}
2021-08-11 14:37:26 +08:00
pub struct CodeGenTask {
pub subst: Vec<(Type, ConcreteType)>,
pub store: ConcreteTypeStore,
2021-08-11 14:37:26 +08:00
pub symbol_name: String,
pub signature: ConcreteType,
2021-09-22 16:04:25 +08:00
pub body: Arc<Vec<Stmt<Option<Type>>>>,
pub calls: Arc<HashMap<CodeLocation, CallId>>,
pub unifier_index: usize,
2021-10-16 18:08:13 +08:00
pub resolver: Arc<dyn SymbolResolver + Send + Sync>,
2021-11-20 19:50:25 +08:00
pub id: usize,
2021-08-11 14:37:26 +08:00
}
/// Retrieves the [LLVM type][BasicTypeEnum] corresponding to the [Type].
///
/// This function is used to obtain the in-memory representation of `ty`, e.g. a `bool` variable
/// would be represented by an `i8`.
2024-02-20 18:07:55 +08:00
#[allow(clippy::too_many_arguments)]
2021-08-11 14:37:26 +08:00
fn get_llvm_type<'ctx>(
ctx: &'ctx Context,
module: &Module<'ctx>,
generator: &mut dyn CodeGenerator,
2021-08-11 14:37:26 +08:00
unifier: &mut Unifier,
top_level: &TopLevelContext,
type_cache: &mut HashMap<Type, BasicTypeEnum<'ctx>>,
ty: Type,
) -> BasicTypeEnum<'ctx> {
use TypeEnum::*;
// we assume the type cache should already contain primitive types,
// and they should be passed by value instead of passing as pointer.
2023-12-08 17:43:32 +08:00
type_cache.get(&unifier.get_representative(ty)).copied().unwrap_or_else(|| {
let ty_enum = unifier.get_ty(ty);
let result = match &*ty_enum {
2021-08-12 16:36:23 +08:00
TObj { obj_id, fields, .. } => {
2024-02-27 13:39:05 +08:00
// check to avoid treating non-class primitives as classes
if obj_id.0 <= PRIMITIVE_DEF_IDS.max_id().0 {
return match &*unifier.get_ty_immutable(ty) {
TObj { obj_id, params, .. } if *obj_id == PRIMITIVE_DEF_IDS.option => {
get_llvm_type(
ctx,
module,
generator,
unifier,
top_level,
type_cache,
*params.iter().next().unwrap().1,
)
2023-01-12 19:31:03 +08:00
.ptr_type(AddressSpace::default())
2024-02-27 13:39:05 +08:00
.into()
}
2024-02-27 13:39:05 +08:00
TObj { obj_id, .. } if *obj_id == PRIMITIVE_DEF_IDS.ndarray => {
let llvm_usize = generator.get_size_type(ctx);
let (dtype, _) = unpack_ndarray_tvars(unifier, ty);
let element_type = get_llvm_type(
ctx,
module,
generator,
unifier,
top_level,
type_cache,
dtype,
);
// struct NDArray { num_dims: size_t, dims: size_t*, data: T* }
//
// * num_dims: Number of dimensions in the array
// * dims: Pointer to an array containing the size of each dimension
// * data: Pointer to an array containing the array data
let fields = [
llvm_usize.into(),
llvm_usize.ptr_type(AddressSpace::default()).into(),
element_type.ptr_type(AddressSpace::default()).into(),
];
ctx.struct_type(&fields, false).ptr_type(AddressSpace::default()).into()
}
_ => unreachable!("LLVM type for primitive {} is missing", unifier.stringify(ty)),
}
}
2021-08-12 16:36:23 +08:00
// a struct with fields in the order of declaration
let top_level_defs = top_level.definitions.read();
let definition = top_level_defs.get(obj_id.0).unwrap();
let TopLevelDef::Class { fields: fields_list, .. } = &*definition.read() else {
2021-08-12 16:36:23 +08:00
unreachable!()
};
let name = unifier.stringify(ty);
let ty = if let Some(t) = module.get_struct_type(&name) {
t.ptr_type(AddressSpace::default()).into()
} else {
let struct_type = ctx.opaque_struct_type(&name);
type_cache.insert(
unifier.get_representative(ty),
struct_type.ptr_type(AddressSpace::default()).into()
);
let fields = fields_list
.iter()
.map(|f| {
get_llvm_type(
ctx,
module,
generator,
unifier,
top_level,
type_cache,
fields[&f.0].0,
)
})
.collect_vec();
struct_type.set_body(&fields, false);
struct_type.ptr_type(AddressSpace::default()).into()
};
return ty
2021-08-12 16:36:23 +08:00
}
TTuple { ty } => {
// a struct with fields in the order present in the tuple
let fields = ty
2021-08-11 14:37:26 +08:00
.iter()
.map(|ty| {
get_llvm_type(
2024-02-27 13:39:05 +08:00
ctx, module, generator, unifier, top_level, type_cache, *ty,
)
})
2021-08-11 14:37:26 +08:00
.collect_vec();
2022-01-20 04:14:22 +08:00
ctx.struct_type(&fields, false).into()
2021-08-12 16:36:23 +08:00
}
TList { ty } => {
// a struct with an integer and a pointer to an array
let element_type = get_llvm_type(
2024-02-27 13:39:05 +08:00
ctx, module, generator, unifier, top_level, type_cache, *ty,
);
2022-02-21 18:27:46 +08:00
let fields = [
2023-01-12 19:31:03 +08:00
element_type.ptr_type(AddressSpace::default()).into(),
2022-02-21 18:27:46 +08:00
generator.get_size_type(ctx).into(),
];
2023-01-12 19:31:03 +08:00
ctx.struct_type(&fields, false).ptr_type(AddressSpace::default()).into()
2021-08-12 16:36:23 +08:00
}
TVirtual { .. } => unimplemented!(),
_ => unreachable!("{}", ty_enum.get_type_name()),
};
type_cache.insert(unifier.get_representative(ty), result);
result
2021-08-11 14:37:26 +08:00
})
}
2023-12-08 17:43:32 +08:00
/// Retrieves the [LLVM type][`BasicTypeEnum`] corresponding to the [`Type`].
///
/// This function is used mainly to obtain the ABI representation of `ty`, e.g. a `bool` is
/// would be represented by an `i1`.
///
2023-12-08 17:43:32 +08:00
/// The difference between the in-memory representation (as returned by [`get_llvm_type`]) and the
/// ABI representation is that the in-memory representation must be at least byte-sized and must
/// be byte-aligned for the variable to be addressable in memory, whereas there is no such
/// restriction for ABI representations.
2024-02-20 18:07:55 +08:00
#[allow(clippy::too_many_arguments)]
fn get_llvm_abi_type<'ctx>(
ctx: &'ctx Context,
module: &Module<'ctx>,
generator: &mut dyn CodeGenerator,
unifier: &mut Unifier,
top_level: &TopLevelContext,
type_cache: &mut HashMap<Type, BasicTypeEnum<'ctx>>,
primitives: &PrimitiveStore,
ty: Type,
) -> BasicTypeEnum<'ctx> {
// If the type is used in the definition of a function, return `i1` instead of `i8` for ABI
// consistency.
return if unifier.unioned(ty, primitives.bool) {
ctx.bool_type().into()
} else {
2024-02-27 13:39:05 +08:00
get_llvm_type(ctx, module, generator, unifier, top_level, type_cache, ty)
}
}
2023-10-18 13:40:37 +08:00
/// Whether `sret` is needed for a return value with type `ty`.
///
/// When returning a large data structure (e.g. structures that do not fit in 1-2 native words of
/// the target processor) by value, a synthetic parameter with a pointer type will be passed in the
/// slot of the first parameter to act as the location of which the return value is passed into.
///
2023-12-08 17:43:32 +08:00
/// See <https://releases.llvm.org/14.0.0/docs/LangRef.html#parameter-attributes> for more
2023-10-18 13:40:37 +08:00
/// information.
2023-12-06 11:49:02 +08:00
fn need_sret(ty: BasicTypeEnum) -> bool {
fn need_sret_impl(ty: BasicTypeEnum, maybe_large: bool) -> bool {
match ty {
BasicTypeEnum::IntType(_) | BasicTypeEnum::PointerType(_) => false,
BasicTypeEnum::FloatType(_) if maybe_large => false,
BasicTypeEnum::StructType(ty) if maybe_large && ty.count_fields() <= 2 =>
2023-12-06 11:49:02 +08:00
ty.get_field_types().iter().any(|ty| need_sret_impl(*ty, false)),
_ => true,
}
}
2023-12-06 11:49:02 +08:00
need_sret_impl(ty, true)
}
/// Implementation for generating LLVM IR for a function.
pub fn gen_func_impl<'ctx, G: CodeGenerator, F: FnOnce(&mut G, &mut CodeGenContext) -> Result<(), String>> (
2021-08-12 16:36:23 +08:00
context: &'ctx Context,
generator: &mut G,
registry: &WorkerRegistry,
2021-08-12 16:36:23 +08:00
builder: Builder<'ctx>,
module: Module<'ctx>,
task: CodeGenTask,
codegen_function: F
) -> Result<(Builder<'ctx>, Module<'ctx>, FunctionValue<'ctx>), (Builder<'ctx>, String)> {
2021-11-20 19:50:25 +08:00
let top_level_ctx = registry.top_level_ctx.clone();
let static_value_store = registry.static_value_store.clone();
2021-08-11 14:37:26 +08:00
let (mut unifier, primitives) = {
let (unifier, primitives) = &top_level_ctx.unifiers.read()[task.unifier_index];
(Unifier::from_shared_unifier(unifier), *primitives)
2021-08-11 14:37:26 +08:00
};
unifier.put_primitive_store(&primitives);
unifier.top_level = Some(top_level_ctx.clone());
2021-08-11 14:37:26 +08:00
let mut cache = HashMap::new();
2023-12-08 17:43:32 +08:00
for (a, b) in &task.subst {
2021-08-11 14:37:26 +08:00
// this should be unification between variables and concrete types
// and should not cause any problem...
let b = task.store.to_unifier_type(&mut unifier, &primitives, *b, &mut cache);
2022-02-21 18:27:46 +08:00
unifier
.unify(*a, b)
.or_else(|err| {
if matches!(&*unifier.get_ty(*a), TypeEnum::TRigidVar { .. }) {
unifier.replace_rigid_var(*a, b);
Ok(())
} else {
Err(err)
}
})
2023-12-08 17:43:32 +08:00
.unwrap();
2021-08-11 14:37:26 +08:00
}
// rebuild primitive store with unique representatives
let primitives = PrimitiveStore {
int32: unifier.get_representative(primitives.int32),
int64: unifier.get_representative(primitives.int64),
2022-03-05 03:45:09 +08:00
uint32: unifier.get_representative(primitives.uint32),
uint64: unifier.get_representative(primitives.uint64),
2021-08-11 14:37:26 +08:00
float: unifier.get_representative(primitives.float),
bool: unifier.get_representative(primitives.bool),
none: unifier.get_representative(primitives.none),
2021-10-23 23:53:36 +08:00
range: unifier.get_representative(primitives.range),
2021-11-02 23:22:37 +08:00
str: unifier.get_representative(primitives.str),
exception: unifier.get_representative(primitives.exception),
option: unifier.get_representative(primitives.option),
..primitives
2021-08-11 14:37:26 +08:00
};
let mut type_cache: HashMap<_, _> = [
(primitives.int32, context.i32_type().into()),
(primitives.int64, context.i64_type().into()),
2022-03-05 03:45:09 +08:00
(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 = [
2023-01-12 19:31:03 +08:00
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()
}
}),
2023-01-12 19:31:03 +08:00
(primitives.range, context.i32_type().array_type(3).ptr_type(AddressSpace::default()).into()),
(primitives.exception, {
let name = "Exception";
2023-12-08 17:43:32 +08:00
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()
}
})
2021-08-11 14:37:26 +08:00
]
.iter()
2023-12-08 17:43:32 +08:00
.copied()
2021-08-11 14:37:26 +08:00
.collect();
// NOTE: special handling of option cannot use this type cache since it contains type var,
// handled inside get_llvm_type instead
2021-08-11 14:37:26 +08:00
let ConcreteTypeEnum::TFunc { args, ret, .. } =
task.store.get(task.signature) else {
unreachable!()
};
let (args, ret) = (
args.iter()
.map(|arg| FuncArg {
name: arg.name,
ty: task.store.to_unifier_type(&mut unifier, &primitives, arg.ty, &mut cache),
default_value: arg.default_value.clone(),
})
.collect_vec(),
task.store.to_unifier_type(&mut unifier, &primitives, *ret, &mut cache),
);
let ret_type = if unifier.unioned(ret, primitives.none) {
None
} else {
Some(get_llvm_abi_type(context, &module, generator, &mut unifier, top_level_ctx.as_ref(), &mut type_cache, &primitives, ret))
};
2023-12-06 11:49:02 +08:00
let has_sret = ret_type.map_or(false, |ty| need_sret(ty));
let mut params = args
2021-08-11 14:37:26 +08:00
.iter()
.map(|arg| {
get_llvm_abi_type(
2022-02-21 18:27:46 +08:00
context,
&module,
2022-02-21 18:27:46 +08:00
generator,
&mut unifier,
top_level_ctx.as_ref(),
&mut type_cache,
&primitives,
2022-02-21 18:27:46 +08:00
arg.ty,
)
.into()
2021-08-11 14:37:26 +08:00
})
.collect_vec();
if has_sret {
2023-01-12 19:31:03 +08:00
params.insert(0, ret_type.unwrap().ptr_type(AddressSpace::default()).into());
}
let fn_type = match ret_type {
Some(ret_type) if !has_sret => ret_type.fn_type(&params, false),
_ => context.void_type().fn_type(&params, false)
2021-08-11 14:37:26 +08:00
};
let symbol = &task.symbol_name;
2021-09-08 19:45:36 +08:00
let fn_val =
module.get_function(symbol).unwrap_or_else(|| module.add_function(symbol, fn_type, None));
if let Some(personality) = &top_level_ctx.personality_symbol {
2021-10-16 18:08:13 +08:00
let personality = module.get_function(personality).unwrap_or_else(|| {
let ty = context.i32_type().fn_type(&[], true);
2021-10-16 18:08:13 +08:00
module.add_function(personality, ty, None)
});
fn_val.set_personality_function(personality);
}
if has_sret {
fn_val.add_attribute(AttributeLoc::Param(0),
context.create_type_attribute(Attribute::get_named_enum_kind_id("sret"),
ret_type.unwrap().as_any_type_enum()));
}
2021-08-11 14:37:26 +08:00
let init_bb = context.append_basic_block(fn_val, "init");
builder.position_at_end(init_bb);
let body_bb = context.append_basic_block(fn_val, "body");
let mut var_assignment = HashMap::new();
2023-12-08 17:43:32 +08:00
let offset = u32::from(has_sret);
for (n, arg) in args.iter().enumerate() {
let param = fn_val.get_nth_param((n as u32) + offset).unwrap();
let local_type = get_llvm_type(
context,
&module,
generator,
&mut unifier,
top_level_ctx.as_ref(),
&mut type_cache,
arg.ty,
);
2024-02-19 19:30:25 +08:00
let alloca = builder
.build_alloca(local_type, &format!("{}.addr", &arg.name.to_string()))
.unwrap();
// Remap boolean parameters into i8
let param = if local_type.is_int_type() && param.is_int_value() {
let expected_ty = local_type.into_int_type();
let param_val = param.into_int_value();
if expected_ty.get_bit_width() == 8 && param_val.get_type().get_bit_width() == 1 {
2023-12-06 11:49:02 +08:00
bool_to_i8(&builder, context, param_val)
} else {
param_val
}.into()
} else {
param
};
2024-02-19 19:30:25 +08:00
builder.build_store(alloca, param).unwrap();
2021-11-20 19:50:25 +08:00
var_assignment.insert(arg.name, (alloca, None, 0));
}
let return_buffer = if has_sret {
Some(fn_val.get_nth_param(0).unwrap().into_pointer_value())
} else {
2024-02-19 19:30:25 +08:00
fn_type.get_return_type().map(|v| builder.build_alloca(v, "$ret").unwrap())
};
2021-11-20 19:50:25 +08:00
let static_values = {
let store = registry.static_value_store.lock();
store.store[task.id].clone()
};
2023-12-08 17:43:32 +08:00
for (k, v) in static_values {
2021-11-20 19:50:25 +08:00
let (_, static_val, _) = var_assignment.get_mut(&args[k].name).unwrap();
*static_val = Some(v);
2021-08-11 14:37:26 +08:00
}
2021-11-20 19:50:25 +08:00
2024-02-19 19:30:25 +08:00
builder.build_unconditional_branch(body_bb).unwrap();
2021-08-11 14:37:26 +08:00
builder.position_at_end(body_bb);
2022-04-14 15:40:12 +08:00
let (dibuilder, compile_unit) = module.create_debug_info_builder(
/* allow_unresolved */ true,
/* language */ inkwell::debug_info::DWARFSourceLanguage::Python,
/* filename */
&task
.body
2024-02-20 18:07:55 +08:00
.first()
2022-04-14 15:40:12 +08:00
.map_or_else(
|| "<nac3_internal>".to_string(),
|f| f.location.file.0.to_string(),
),
/* directory */ "",
/* producer */ "NAC3",
/* is_optimized */ registry.llvm_options.opt_level != OptimizationLevel::None,
2022-04-14 15:40:12 +08:00
/* compiler command line flags */ "",
/* runtime_ver */ 0,
/* split_name */ "",
/* kind */ inkwell::debug_info::DWARFEmissionKind::Full,
/* dwo_id */ 0,
/* split_debug_inling */ true,
/* debug_info_for_profiling */ false,
/* sysroot */ "",
/* sdk */ "",
);
let subroutine_type = dibuilder.create_subroutine_type(
compile_unit.get_file(),
Some(
dibuilder
.create_basic_type("_", 0_u64, 0x00, inkwell::debug_info::DIFlags::PUBLIC)
.unwrap()
.as_type(),
),
&[],
inkwell::debug_info::DIFlags::PUBLIC,
);
let (row, col) =
2024-02-20 18:07:55 +08:00
task.body.first().map_or_else(|| (0, 0), |b| (b.location.row, b.location.column));
2022-04-14 15:40:12 +08:00
let func_scope: DISubprogram<'_> = dibuilder.create_function(
/* scope */ compile_unit.as_debug_info_scope(),
/* func name */ symbol,
/* linkage_name */ None,
/* file */ compile_unit.get_file(),
/* line_no */ row as u32,
/* DIType */ subroutine_type,
/* is_local_to_unit */ false,
/* is_definition */ true,
/* scope_line */ row as u32,
/* flags */ inkwell::debug_info::DIFlags::PUBLIC,
/* is_optimized */ registry.llvm_options.opt_level != OptimizationLevel::None,
2022-04-14 15:40:12 +08:00
);
fn_val.set_subprogram(func_scope);
2021-08-11 14:37:26 +08:00
let mut code_gen_context = CodeGenContext {
2021-10-16 18:08:13 +08:00
ctx: context,
2021-08-11 14:37:26 +08:00
resolver: task.resolver,
top_level: top_level_ctx.as_ref(),
calls: task.calls,
loop_target: None,
return_target: None,
return_buffer,
unwind_target: None,
outer_catch_clauses: None,
2023-12-08 17:43:32 +08:00
const_strings: HashMap::default(),
registry,
2021-08-11 14:37:26 +08:00
var_assignment,
type_cache,
primitives,
init_bb,
2023-12-08 17:43:32 +08:00
exception_val: Option::default(),
2021-08-11 14:37:26 +08:00
builder,
module,
unifier,
2021-11-20 19:50:25 +08:00
static_value_store,
need_sret: has_sret,
2023-12-08 17:43:32 +08:00
current_loc: Location::default(),
2022-04-14 15:40:12 +08:00
debug_info: (dibuilder, compile_unit, func_scope.as_debug_info_scope()),
2021-08-11 14:37:26 +08:00
};
2022-04-14 15:40:12 +08:00
let loc = code_gen_context.debug_info.0.create_debug_location(
context,
row as u32,
col as u32,
func_scope.as_debug_info_scope(),
None
);
2023-03-08 15:19:09 +08:00
code_gen_context.builder.set_current_debug_location(loc);
2022-04-14 15:40:12 +08:00
let result = codegen_function(generator, &mut code_gen_context);
2021-09-19 22:54:06 +08:00
// after static analysis, only void functions can have no return at the end.
if !code_gen_context.is_terminated() {
2024-02-19 19:30:25 +08:00
code_gen_context.builder.build_return(None).unwrap();
2021-08-11 14:37:26 +08:00
}
2021-08-12 13:55:15 +08:00
2022-04-14 15:40:12 +08:00
code_gen_context.builder.unset_current_debug_location();
code_gen_context.debug_info.0.finalize();
2021-08-13 13:33:59 +08:00
let CodeGenContext { builder, module, .. } = code_gen_context;
if let Err(e) = result {
return Err((builder, e));
}
2021-08-13 13:33:59 +08:00
Ok((builder, module, fn_val))
2021-08-11 14:37:26 +08:00
}
/// Generates LLVM IR for a function.
///
2023-12-08 17:43:32 +08:00
/// * `context` - The [LLVM Context][`Context`] used in generating the function body.
/// * `generator` - The [`CodeGenerator`] for generating various program constructs.
/// * `registry` - The [`WorkerRegistry`] responsible for monitoring this function generation task.
/// * `builder` - The [`Builder`] used for generating LLVM IR.
/// * `module` - The [`Module`] of which the generated LLVM function will be inserted into.
/// * `task` - The [`CodeGenTask`] associated with this function generation task.
///
pub fn gen_func<'ctx, G: CodeGenerator>(
context: &'ctx Context,
generator: &mut G,
registry: &WorkerRegistry,
builder: Builder<'ctx>,
module: Module<'ctx>,
task: CodeGenTask,
) -> Result<(Builder<'ctx>, Module<'ctx>, FunctionValue<'ctx>), (Builder<'ctx>, String)> {
let body = task.body.clone();
gen_func_impl(context, generator, registry, builder, module, task, |generator, ctx| {
generator.gen_block(ctx, body.iter())
})
}
/// Converts the value of a boolean-like value `bool_value` into an `i1`.
fn bool_to_i1<'ctx>(builder: &Builder<'ctx>, bool_value: IntValue<'ctx>) -> IntValue<'ctx> {
2023-12-08 17:43:32 +08:00
if bool_value.get_type().get_bit_width() == 1 {
bool_value
} else {
2024-02-19 19:30:25 +08:00
builder
.build_int_compare(
IntPredicate::NE,
bool_value,
bool_value.get_type().const_zero(),
"tobool",
)
.unwrap()
}
}
/// Converts the value of a boolean-like value `bool_value` into an `i8`.
fn bool_to_i8<'ctx>(
builder: &Builder<'ctx>,
ctx: &'ctx Context,
bool_value: IntValue<'ctx>
) -> IntValue<'ctx> {
let value_bits = bool_value.get_type().get_bit_width();
match value_bits {
8 => bool_value,
2024-02-19 19:30:25 +08:00
1 => builder.build_int_z_extend(bool_value, ctx.i8_type(), "frombool").unwrap(),
_ => bool_to_i8(
builder,
ctx,
2024-02-19 19:30:25 +08:00
builder
.build_int_compare(
IntPredicate::NE,
bool_value,
bool_value.get_type().const_zero(),
"",
)
.unwrap()
),
}
}
/// Generates a sequence of IR which checks whether `value` does not exceed the upper bound of the
/// range as defined by `stop` and `step`.
///
/// Note that the generated IR will **not** check whether value is part of the range or whether
/// value exceeds the lower bound of the range (as evident by the missing `start` argument).
///
/// The generated IR is equivalent to the following Rust code:
///
/// ```rust,ignore
/// let sign = step > 0;
/// let (lo, hi) = if sign { (value, stop) } else { (stop, value) };
/// let cmp = lo < hi;
/// ```
///
2023-12-08 17:43:32 +08:00
/// Returns an `i1` [`IntValue`] representing the result of whether the `value` is in the range.
2023-12-06 11:49:02 +08:00
fn gen_in_range_check<'ctx>(
ctx: &CodeGenContext<'ctx, '_>,
value: IntValue<'ctx>,
stop: IntValue<'ctx>,
step: IntValue<'ctx>,
) -> IntValue<'ctx> {
2024-02-19 19:30:25 +08:00
let sign = ctx.builder.build_int_compare(IntPredicate::SGT, step, ctx.ctx.i32_type().const_zero(), "").unwrap();
let lo = ctx.builder.build_select(sign, value, stop, "")
.map(BasicValueEnum::into_int_value)
.unwrap();
let hi = ctx.builder.build_select(sign, stop, value, "")
.map(BasicValueEnum::into_int_value)
.unwrap();
ctx.builder.build_int_compare(IntPredicate::SLT, lo, hi, "cmp").unwrap()
}