Backend documentation work.

This commit is contained in:
2023-04-23 20:37:25 -07:00
parent 9b72fb7827
commit e79788bcd1

View File

@@ -1,3 +1,34 @@
//! # The compiler backend!
//!
//! This module is responsible for taking our intermediate representation from
//! [`crate::ir`] and turning it into Cranelift and then into object code that
//! can either be saved to disk or run in memory. Because the runtime functions
//! for NGR are very closely tied to the compiler implentation, we also include
//! information about these functions as part of the module.
//!
//! ## Using the `Backend`
//!
//! The backend of this compiler can be used in two modes: a static compilation
//! mode, where the goal is to write the compiled object to disk and then link
//! it later, and a JIT mode, where the goal is to write the compiled object to
//! memory and then run it. Both modes use the same `Backend` object, because
//! they share a lot of behaviors. However, you'll want to use different variants
//! based on your goals:
//!
//! * Use `Backend<ObjectModule>`, constructed via
//! [`Backend::object_file`](self::backend::Backend::object_file), if you
//! want to compile to an object file on disk, which you're then going to
//! link to later.
//! * Use `Backend<JITModule>`, constructed via
//! [`Backend::jit`](self::backend::Backend::jit), if you want to do
//! just-in-time compilation and are just going to run things immediately.
//!
//! ## Working with Runtime Functions
//!
//! For now, runtime functions are pretty easy to describe, because there's
//! only one. In the future, though, the
//! [`RuntimeFunctions`](self::backend::RuntimeFunctions) object is there to
//! help provide a clean interface to them all.
mod error; mod error;
mod eval; mod eval;
mod into_crane; mod into_crane;
@@ -16,6 +47,15 @@ use target_lexicon::Triple;
const EMPTY_DATUM: [u8; 8] = [0; 8]; const EMPTY_DATUM: [u8; 8] = [0; 8];
/// An object representing an active backend.
///
/// Internally, this object holds a bunch of state useful for compiling one
/// or more functions into an object file or memory. It can be passed around,
/// but cannot currently be duplicated because some of that state is not
/// easily duplicated. You should be able to share this across threads, assuming
/// normal Rust safety, but you should be thoughtful about transferring it across
/// processes in a JIT context due to some special cases in the runtime function
/// implementations.
pub struct Backend<M: Module> { pub struct Backend<M: Module> {
pub module: M, pub module: M,
data_ctx: DataContext, data_ctx: DataContext,
@@ -26,6 +66,12 @@ pub struct Backend<M: Module> {
} }
impl Backend<JITModule> { impl Backend<JITModule> {
/// Create a new JIT backend for compiling NGR into memory.
///
/// The provided output buffer is not for the compiled code, but for the output
/// of any `print` expressions that are evaluated. If set to `None`, the output
/// will be written to `stdout` as per normal, but if a String buffer is provided,
/// it will be extended by any `print` statements that happen during code execution.
pub fn jit(output_buffer: Option<String>) -> Result<Self, BackendError> { pub fn jit(output_buffer: Option<String>) -> Result<Self, BackendError> {
let platform = Triple::host(); let platform = Triple::host();
let isa_builder = isa::lookup(platform.clone())?; let isa_builder = isa::lookup(platform.clone())?;
@@ -50,12 +96,24 @@ impl Backend<JITModule> {
}) })
} }
/// Given a compiled function ID, get a pointer to where that function was written
/// in memory.
///
/// The data at this pointer should not be mutated unless you really, really,
/// really know what you're doing. It can be run by casting it into a Rust
/// `fn() -> ()`, and then calling it from normal Rust.
pub fn bytes(&self, function_id: FuncId) -> *const u8 { pub fn bytes(&self, function_id: FuncId) -> *const u8 {
self.module.get_finalized_function(function_id) self.module.get_finalized_function(function_id)
} }
} }
impl Backend<ObjectModule> { impl Backend<ObjectModule> {
/// Generate a backend for compiling into an object file for the given target.
///
/// This backend will generate a single output file per `Backend` object, although
/// that file may have multiple functions defined within it. Data between those
/// functions (in particular, strings) will be defined once and shared between
/// the different functions.
pub fn object_file(platform: Triple) -> Result<Self, BackendError> { pub fn object_file(platform: Triple) -> Result<Self, BackendError> {
let isa_builder = isa::lookup(platform.clone())?; let isa_builder = isa::lookup(platform.clone())?;
let mut settings_builder = settings::builder(); let mut settings_builder = settings::builder();
@@ -76,12 +134,22 @@ impl Backend<ObjectModule> {
}) })
} }
/// Given all the functions defined, return the bytes the object file should contain.
pub fn bytes(self) -> Result<Vec<u8>, BackendError> { pub fn bytes(self) -> Result<Vec<u8>, BackendError> {
self.module.finish().emit().map_err(Into::into) self.module.finish().emit().map_err(Into::into)
} }
} }
impl<M: Module> Backend<M> { impl<M: Module> Backend<M> {
/// Define a string within the current backend.
///
/// Note that this is a Cranelift [`DataId`], which then must be redeclared inside the
/// context of any functions or data items that want to use it. That being said, the
/// string value will be defined once in the file and then shared by all referencers.
///
/// This function will automatically add a null character (`'\0'`) to the end of the
/// string, to ensure that strings are non-terminated for interactions with other
/// languages.
pub fn define_string(&mut self, s: &str) -> Result<DataId, BackendError> { pub fn define_string(&mut self, s: &str) -> Result<DataId, BackendError> {
let name = format!("<string_constant>{}", s); let name = format!("<string_constant>{}", s);
let s0 = format!("{}\0", s); let s0 = format!("{}\0", s);
@@ -97,6 +165,11 @@ impl<M: Module> Backend<M> {
Ok(global_id) Ok(global_id)
} }
/// Define a global variable within the current backend.
///
/// These variables can be shared between functions, and will be exported from the
/// module itself as public data in the case of static compilation. There initial
/// value will be null.
pub fn define_variable(&mut self, name: String) -> Result<DataId, BackendError> { pub fn define_variable(&mut self, name: String) -> Result<DataId, BackendError> {
self.data_ctx.define(Box::new(EMPTY_DATUM)); self.data_ctx.define(Box::new(EMPTY_DATUM));
let id = self let id = self
@@ -108,6 +181,11 @@ impl<M: Module> Backend<M> {
Ok(id) Ok(id)
} }
/// Get a pointer to the output buffer for `print`ing, or `null`.
///
/// As suggested, returns `null` in the case where the user has not provided an
/// output buffer; it is your responsibility to check for this case and do
/// something sensible.
pub fn output_buffer_ptr(&mut self) -> *mut String { pub fn output_buffer_ptr(&mut self) -> *mut String {
if let Some(str) = self.output_buffer.as_mut() { if let Some(str) = self.output_buffer.as_mut() {
str as *mut String str as *mut String
@@ -116,6 +194,10 @@ impl<M: Module> Backend<M> {
} }
} }
/// Get any captured output `print`ed by the program during execution.
///
/// If an output buffer was not provided, or if the program has not done any
/// printing, then this function will return an empty string.
pub fn output(self) -> String { pub fn output(self) -> String {
if let Some(s) = self.output_buffer { if let Some(s) = self.output_buffer {
s s