Files
ngr/src/backend/into_crane.rs

714 lines
32 KiB
Rust

use crate::backend::error::BackendError;
use crate::backend::Backend;
use crate::eval::PrimitiveType;
use crate::ir::{Expression, Name, Primitive, Program, Type, Value, ValueOrRef};
use crate::syntax::{ConstantType, Location};
use cranelift_codegen::ir::{
self, entities, types, AbiParam, Function, GlobalValue, InstBuilder, MemFlags, Signature,
UserFuncName,
};
use cranelift_codegen::isa::CallConv;
use cranelift_codegen::Context;
use cranelift_frontend::{FunctionBuilder, FunctionBuilderContext};
use cranelift_module::{DataDescription, FuncId, Linkage, Module};
use std::collections::{hash_map, HashMap};
const VOID_REPR_TYPE: types::Type = types::I64;
/// When we're talking about variables, it's handy to just have a table that points
/// from a variable to "what to do if you want to reference this variable", which is
/// agnostic about whether the variable is local, global, an argument, etc. Since
/// the type of that function is a little bit annoying, we summarize it here.
pub enum ReferenceBuilder {
Global(types::Type, GlobalValue),
Local(types::Type, cranelift_frontend::Variable),
Argument(types::Type, entities::Value),
}
impl<M: Module> Backend<M> {
/// Translate the given IR type into an ABI parameter type for cranelift, as
/// best as possible.
fn translate_type(&self, t: &Type) -> AbiParam {
let (value_type, extension) = match t {
Type::Function(_, _) | Type::Structure(_) => (
types::Type::triple_pointer_type(&self.platform),
ir::ArgumentExtension::None,
),
Type::Primitive(PrimitiveType::Void) => (types::I8, ir::ArgumentExtension::None), // FIXME?
Type::Primitive(PrimitiveType::I8) => (types::I8, ir::ArgumentExtension::Sext),
Type::Primitive(PrimitiveType::I16) => (types::I16, ir::ArgumentExtension::Sext),
Type::Primitive(PrimitiveType::I32) => (types::I32, ir::ArgumentExtension::Sext),
Type::Primitive(PrimitiveType::I64) => (types::I64, ir::ArgumentExtension::Sext),
Type::Primitive(PrimitiveType::U8) => (types::I8, ir::ArgumentExtension::Uext),
Type::Primitive(PrimitiveType::U16) => (types::I16, ir::ArgumentExtension::Uext),
Type::Primitive(PrimitiveType::U32) => (types::I32, ir::ArgumentExtension::Uext),
Type::Primitive(PrimitiveType::U64) => (types::I64, ir::ArgumentExtension::Uext),
};
AbiParam {
value_type,
purpose: ir::ArgumentPurpose::Normal,
extension,
}
}
/// Compile the given program.
///
/// The returned value is a `FuncId` that represents a function that runs all the statements
/// found in the program, which will be compiled using the given function name. (If there
/// are no such statements, the function will do nothing.)
pub fn compile_program(
&mut self,
function_name: Name,
program: Program<Type>,
) -> Result<FuncId, BackendError> {
let mut variables = HashMap::new();
for (top_level_name, top_level_type) in program.get_top_level_variables() {
match top_level_type {
Type::Function(argument_types, return_type) => {
let func_id = self.declare_function(
top_level_name.current_name(),
Linkage::Export,
argument_types,
*return_type,
)?;
self.defined_functions.insert(top_level_name, func_id);
}
Type::Primitive(pt) => {
let data_id = self.module.declare_data(
top_level_name.current_name(),
Linkage::Export,
true,
false,
)?;
tracing::info!(name = %top_level_name, data_type = %pt, "defining top-level data");
self.module.define_data(data_id, &pt.blank_data())?;
let constant_type = ConstantType::from(pt);
self.defined_symbols
.insert(top_level_name, (data_id, constant_type.into()));
}
Type::Structure(mut fields) => {
let data_id = self.module.declare_data(
top_level_name.current_name(),
Linkage::Export,
true,
false,
)?;
tracing::info!(name = %top_level_name, "defining top-level data structure");
self.module.define_data(data_id, &fields.blank_data())?;
let pointer = self.module.target_config().pointer_type();
self.defined_symbols
.insert(top_level_name, (data_id, pointer));
}
}
}
let void = Type::Primitive(PrimitiveType::Void);
let main_func_id = self.declare_function(
function_name.current_name(),
Linkage::Export,
vec![],
void.clone(),
)?;
self.defined_functions
.insert(function_name.clone(), main_func_id);
for (_, function) in program.functions.into_iter() {
let func_id = self.compile_function(
&mut variables,
function.name.clone(),
&function.arguments,
function.return_type,
function.body,
)?;
self.defined_functions.insert(function.name, func_id);
}
self.compile_function(
&mut variables,
function_name,
&[],
void.clone(),
program.body,
)
}
fn declare_function(
&mut self,
name: &str,
linkage: Linkage,
argument_types: Vec<Type>,
return_type: Type,
) -> Result<FuncId, cranelift_module::ModuleError> {
tracing::info!(linkage = ?linkage, name, "Declaring function");
let basic_signature = Signature {
params: argument_types
.iter()
.map(|t| self.translate_type(t))
.collect(),
returns: if return_type == Type::Primitive(PrimitiveType::Void) {
vec![]
} else {
vec![self.translate_type(&return_type)]
},
call_conv: CallConv::triple_default(&self.platform),
};
// this generates the handle for the function that we'll eventually want to
// return to the user. For now, we declare all functions defined by this
// function as public/global/exported, although we may want to reconsider
// this decision later.
self.module
.declare_function(name, linkage, &basic_signature)
}
/// Compile the given function.
#[tracing::instrument(level = "debug", skip(self, variables, body))]
pub fn compile_function(
&mut self,
variables: &mut HashMap<Name, ReferenceBuilder>,
function_name: Name,
arguments: &[(Name, Type)],
return_type: Type,
body: Expression<Type>,
) -> Result<FuncId, BackendError> {
// reset the next variable counter. this value shouldn't matter; hopefully
// we won't be using close to 2^32 variables!
self.reset_local_variable_tracker();
let basic_signature = Signature {
params: arguments
.iter()
.map(|(_, t)| self.translate_type(t))
.collect(),
returns: if return_type == Type::Primitive(PrimitiveType::Void) {
vec![]
} else {
vec![self.translate_type(&return_type)]
},
call_conv: CallConv::triple_default(&self.platform),
};
// this generates the handle for the function that we'll eventually want to
// return to the user. For now, we declare all functions defined by this
// function as public/global/exported, although we may want to reconsider
// this decision later.
let func_id = match self.defined_functions.entry(function_name.clone()) {
hash_map::Entry::Occupied(entry) => *entry.get(),
hash_map::Entry::Vacant(vac) => {
tracing::warn!(name = ?function_name.current_name(), "compiling undeclared function");
let func_id = self.module.declare_function(
function_name.current_name(),
Linkage::Export,
&basic_signature,
)?;
vac.insert(func_id);
func_id
}
};
// Next we have to generate the compilation context for the rest of this
// function. Currently, we generate a fresh context for every function.
// Since we're only generating one function per `Program`, this makes
// complete sense. However, in the future, we may want to revisit this
// decision.
let mut ctx = Context::new();
let user_func_name = UserFuncName::user(0, func_id.as_u32());
ctx.func = Function::with_name_signature(user_func_name, basic_signature);
// At the outer-most scope of things, we'll put global variables we've defined
// elsewhere in the program.
for (name, (data_id, ty)) in self.defined_symbols.iter() {
let local_data = self.module.declare_data_in_func(*data_id, &mut ctx.func);
variables.insert(name.clone(), ReferenceBuilder::Global(*ty, local_data));
}
// Finally (!), we generate the function builder that we're going to use to
// make this function!
let mut fctx = FunctionBuilderContext::new();
let mut builder = FunctionBuilder::new(&mut ctx.func, &mut fctx);
// Make the initial block to put instructions in. Later, when we have control
// flow, we might add more blocks after this one. But, for now, we only have
// the one block.
let main_block = builder.create_block();
// add the block parameters, which should be the function parameters
for (name, ty) in arguments.iter() {
let constant_type = self.translate_type(ty).value_type;
let value = builder.append_block_param(main_block, constant_type);
variables.insert(
name.clone(),
ReferenceBuilder::Argument(constant_type, value),
);
}
builder.switch_to_block(main_block);
let (value, _) = self.compile_expression(body, variables, &mut builder)?;
// Now that we're done, inject a return function (one with no actual value; basically
// the equivalent of Rust's `return;`). We then seal the block (which lets Cranelift
// know that the block is done), and then finalize the function (which lets Cranelift
// know we're done with the function).
if return_type == Type::Primitive(PrimitiveType::Void) {
builder.ins().return_(&[]);
} else {
builder.ins().return_(&[value]);
};
builder.seal_block(main_block);
builder.finalize();
// This is a little odd. We want to tell the rest of Cranelift about this function,
// so we register it using the function ID and our builder context. However, the
// result of this function isn't actually super helpful. So we ignore it, unless
// it's an error.
self.module.define_function(func_id, &mut ctx)?;
// done!
Ok(func_id)
}
/// Compile an expression, returning the Cranelift Value for the expression and
/// its type.
#[tracing::instrument(level = "trace", skip(self, variables, builder))]
fn compile_expression(
&mut self,
expr: Expression<Type>,
variables: &mut HashMap<Name, ReferenceBuilder>,
builder: &mut FunctionBuilder,
) -> Result<(entities::Value, types::Type), BackendError> {
match expr {
Expression::Atomic(x) => self.compile_value_or_ref(x, variables, builder),
Expression::Cast(_, target_type, valref) => {
let val_is_signed = valref.type_of().is_signed();
let (val, val_type) = self.compile_value_or_ref(valref, variables, builder)?;
match (val_type, &target_type) {
(types::I8, Type::Primitive(PrimitiveType::I8)) => Ok((val, val_type)),
(types::I8, Type::Primitive(PrimitiveType::I16)) => {
if val_is_signed {
Ok((builder.ins().sextend(types::I16, val), types::I16))
} else {
Ok((builder.ins().uextend(types::I16, val), types::I16))
}
}
(types::I8, Type::Primitive(PrimitiveType::I32)) => {
if val_is_signed {
Ok((builder.ins().sextend(types::I32, val), types::I32))
} else {
Ok((builder.ins().uextend(types::I32, val), types::I32))
}
}
(types::I8, Type::Primitive(PrimitiveType::I64)) => {
if val_is_signed {
Ok((builder.ins().sextend(types::I64, val), types::I64))
} else {
Ok((builder.ins().uextend(types::I64, val), types::I64))
}
}
(types::I16, Type::Primitive(PrimitiveType::I16)) => Ok((val, val_type)),
(types::I16, Type::Primitive(PrimitiveType::I32)) => {
if val_is_signed {
Ok((builder.ins().sextend(types::I32, val), types::I32))
} else {
Ok((builder.ins().uextend(types::I32, val), types::I32))
}
}
(types::I16, Type::Primitive(PrimitiveType::I64)) => {
if val_is_signed {
Ok((builder.ins().sextend(types::I64, val), types::I64))
} else {
Ok((builder.ins().uextend(types::I64, val), types::I64))
}
}
(types::I32, Type::Primitive(PrimitiveType::I32)) => Ok((val, val_type)),
(types::I32, Type::Primitive(PrimitiveType::I64)) => {
if val_is_signed {
Ok((builder.ins().sextend(types::I64, val), types::I64))
} else {
Ok((builder.ins().uextend(types::I64, val), types::I64))
}
}
(types::I64, Type::Primitive(PrimitiveType::I64)) => Ok((val, val_type)),
(types::I8, Type::Primitive(PrimitiveType::U8)) => Ok((val, val_type)),
(types::I8, Type::Primitive(PrimitiveType::U16)) => {
Ok((builder.ins().uextend(types::I16, val), types::I16))
}
(types::I8, Type::Primitive(PrimitiveType::U32)) => {
Ok((builder.ins().uextend(types::I32, val), types::I32))
}
(types::I8, Type::Primitive(PrimitiveType::U64)) => {
Ok((builder.ins().uextend(types::I64, val), types::I64))
}
(types::I16, Type::Primitive(PrimitiveType::U16)) => Ok((val, val_type)),
(types::I16, Type::Primitive(PrimitiveType::U32)) => {
Ok((builder.ins().uextend(types::I32, val), types::I32))
}
(types::I16, Type::Primitive(PrimitiveType::U64)) => {
Ok((builder.ins().uextend(types::I64, val), types::I64))
}
(types::I32, Type::Primitive(PrimitiveType::U32)) => Ok((val, val_type)),
(types::I32, Type::Primitive(PrimitiveType::U64)) => {
Ok((builder.ins().uextend(types::I64, val), types::I64))
}
(types::I64, Type::Primitive(PrimitiveType::U64)) => Ok((val, val_type)),
(types::I64, Type::Primitive(PrimitiveType::Void)) => {
Ok((builder.ins().iconst(types::I64, 0), VOID_REPR_TYPE))
}
_ => Err(BackendError::InvalidTypeCast {
from: val_type,
to: target_type,
}),
}
}
Expression::Construct(_, ty, _, fields) => {
let Type::Structure(type_fields) = ty else {
panic!("Got to backend with non-structure type in structure construction?!");
};
let global_allocator =
Name::new("__global_allocation_pointer__", Location::manufactured());
let Some(ReferenceBuilder::Global(_, allocator_variable)) =
variables.get(&global_allocator)
else {
panic!("Couldn't find global allocation pointer");
};
let pointer_to = self.module.target_config().pointer_type();
let allocator_pointer = builder.ins().symbol_value(pointer_to, *allocator_variable);
let structure =
builder
.ins()
.load(pointer_to, MemFlags::new(), allocator_pointer, 0);
let structure_size = builder
.ins()
.iconst(pointer_to, type_fields.object_size() as i64);
let updated_allocator_value = builder.ins().iadd(structure, structure_size);
builder.ins().store(
MemFlags::new(),
updated_allocator_value,
allocator_pointer,
0,
);
for (field_name, field_value) in fields.into_iter() {
let (field_value, field_cranelift_type) =
self.compile_value_or_ref(field_value, variables, builder)?;
let Some((field_internal_type, offset)) =
type_fields.field_type_and_offset(&field_name)
else {
panic!("Got to backend with mismatched construction and type definition");
};
assert_eq!(
field_cranelift_type,
self.translate_type(field_internal_type).value_type
);
builder
.ins()
.store(MemFlags::new(), field_value, structure, offset);
}
Ok((structure, self.module.target_config().pointer_type()))
}
Expression::FieldRef(_, _, struct_type, val, field) => {
let (structure, _) = self.compile_value_or_ref(val, variables, builder)?;
let Type::Structure(fields) = struct_type else {
panic!("Got to backend with non-structure type in field reference?!");
};
let Some((field_type, offset)) = fields.field_type_and_offset(&field) else {
panic!("Got to backend with invalid field for structure type?!");
};
let field_cranelift_type = self.translate_type(field_type).value_type;
let value =
builder
.ins()
.load(field_cranelift_type, MemFlags::new(), structure, offset);
Ok((value, field_cranelift_type))
}
Expression::Block(_, _, mut exprs) => match exprs.pop() {
None => Ok((builder.ins().iconst(types::I64, 0), VOID_REPR_TYPE)),
Some(last) => {
for inner in exprs {
// we can ignore all of these return values and such, because we
// don't actually use them anywhere
self.compile_expression(inner, variables, builder)?;
}
// instead, we just return the last one
self.compile_expression(last, variables, builder)
}
},
Expression::Bind(_, name, _, expr) => {
let (value, value_type) = self.compile_expression(*expr, variables, builder)?;
let variable = self.generate_local();
match variables.get(&name) {
Some(ReferenceBuilder::Global(_, global_value)) => {
let pointer = self.module.target_config().pointer_type();
let pointer_to = builder.ins().symbol_value(pointer, *global_value);
builder.ins().store(MemFlags::new(), value, pointer_to, 0);
}
Some(ReferenceBuilder::Argument(_, _)) => {
panic!("Attempt to mutate an argument {}", name)
}
Some(ReferenceBuilder::Local(_, _)) => {
panic!("Attempt to mutate local {}", name);
}
None => {
builder.declare_var(variable, value_type);
builder.def_var(variable, value);
variables.insert(name, ReferenceBuilder::Local(value_type, variable));
}
}
Ok((value, value_type))
}
Expression::Call(_, final_type, function, args) => {
// Get a reference to the string we want to print.
let var_name = match args[0] {
ValueOrRef::Ref(_, _, ref name) => name.to_string(),
ValueOrRef::Value(_, _, _) => "<unknown>".to_string(),
ValueOrRef::Primitive(_, _, n) => format!("<primitive {}>", n),
};
let var_type = args[0].type_of();
let (arguments, argument_types): (Vec<_>, Vec<_>) = args
.into_iter()
.map(|x| self.compile_value_or_ref(x, variables, builder))
.collect::<Result<Vec<(_, _)>, BackendError>>()?
.into_iter()
.unzip();
match function {
ValueOrRef::Value(_, _, _) => {
panic!("Can't use a value for a function")
}
ValueOrRef::Ref(_, result_type, name) => {
match self.defined_functions.get(&name) {
None => panic!("Couldn't find function {} to call", name),
Some(function) => {
let func_ref =
self.module.declare_func_in_func(*function, builder.func);
let call = builder.ins().call(func_ref, &arguments);
let results = builder.inst_results(call);
match results {
[] => Ok((builder.ins().iconst(types::I64, 0), VOID_REPR_TYPE)),
[result] => {
Ok((*result, self.translate_type(&result_type).value_type))
}
_ => panic!("don't support multi-value returns yet"),
}
}
}
}
ValueOrRef::Primitive(_, _, prim) => match prim {
Primitive::Plus => {
assert_eq!(2, arguments.len());
Ok((
builder.ins().iadd(arguments[0], arguments[1]),
argument_types[0],
))
}
Primitive::Minus => {
assert_eq!(2, arguments.len());
Ok((
builder.ins().isub(arguments[0], arguments[1]),
argument_types[0],
))
}
Primitive::Times => {
assert_eq!(2, arguments.len());
Ok((
builder.ins().imul(arguments[0], arguments[1]),
argument_types[0],
))
}
Primitive::Divide if final_type.is_signed() => {
assert_eq!(2, arguments.len());
Ok((
builder.ins().sdiv(arguments[0], arguments[1]),
argument_types[0],
))
}
Primitive::Divide => {
assert_eq!(2, arguments.len());
Ok((
builder.ins().udiv(arguments[0], arguments[1]),
argument_types[0],
))
}
Primitive::Negate => {
assert_eq!(1, arguments.len());
Ok((builder.ins().ineg(arguments[0]), argument_types[0]))
}
Primitive::Print => {
// Get the output buffer (or null) from our general compilation context.
let buffer_ptr = self.output_buffer_ptr();
let buffer_ptr = builder.ins().iconst(types::I64, buffer_ptr as i64);
assert_eq!(1, arguments.len());
let string_data_id = self.string_reference(&var_name)?;
let local_name_ref = self
.module
.declare_data_in_func(string_data_id, builder.func);
let name_ptr = builder.ins().symbol_value(types::I64, local_name_ref);
let (repr_val, casted_val) = match var_type {
Type::Structure(_) => (ConstantType::I64 as i64, arguments[0]),
Type::Function(_, _) => (ConstantType::I64 as i64, arguments[0]),
Type::Primitive(pt) => {
let constant_type = pt.into();
let new_val = match constant_type {
ConstantType::U64
| ConstantType::I64
| ConstantType::Void => arguments[0],
ConstantType::I8
| ConstantType::I16
| ConstantType::I32 => {
builder.ins().sextend(types::I64, arguments[0])
}
ConstantType::U8
| ConstantType::U16
| ConstantType::U32 => {
builder.ins().uextend(types::I64, arguments[0])
}
};
(constant_type as i64, new_val)
}
};
let vtype_repr = builder.ins().iconst(types::I64, repr_val);
// Finally, we can generate the call to print.
let print_func_ref = self.runtime_functions.include_runtime_function(
"print",
&mut self.module,
builder.func,
)?;
builder.ins().call(
print_func_ref,
&[buffer_ptr, name_ptr, vtype_repr, casted_val],
);
Ok((builder.ins().iconst(types::I64, 0), VOID_REPR_TYPE))
}
},
}
}
}
}
/// Compile a value or reference into Cranelift, returning the Cranelift Value for
/// the expression and its type.
#[tracing::instrument(level = "trace", skip(self, variables, builder))]
fn compile_value_or_ref(
&self,
value_or_ref: ValueOrRef<Type>,
variables: &HashMap<Name, ReferenceBuilder>,
builder: &mut FunctionBuilder,
) -> Result<(entities::Value, types::Type), BackendError> {
match value_or_ref {
ValueOrRef::Value(_, _, val) => match val {
Value::I8(_, v) => {
// Cranelift does a funny thing where it checks you aren't using bits in the I64
// we provide above the size of the type we provide. So, in this case, we can only
// set the low 8 bits of the i64. This restriction creates a bit of a problem when
// casting direction from i8 to i64, because Rust will (helpfully) sign-extend the
// negative number for us. Which sets the high bits, which makes Cranelift unhappy.
// So first we cast the i8 as u8, to get rid of the whole concept of sign extension,
// and *then* we cast to i64.
Ok((builder.ins().iconst(types::I8, v as u8 as i64), types::I8))
}
Value::I16(_, v) => Ok((
// see above note for the "... as ... as"
builder.ins().iconst(types::I16, v as u16 as i64),
types::I16,
)),
Value::I32(_, v) => Ok((
// see above note for the "... as ... as"
builder.ins().iconst(types::I32, v as u32 as i64),
types::I32,
)),
Value::I64(_, v) => Ok((builder.ins().iconst(types::I64, v), types::I64)),
Value::U8(_, v) => Ok((builder.ins().iconst(types::I8, v as i64), types::I8)),
Value::U16(_, v) => Ok((builder.ins().iconst(types::I16, v as i64), types::I16)),
Value::U32(_, v) => Ok((builder.ins().iconst(types::I32, v as i64), types::I32)),
Value::U64(_, v) => Ok((builder.ins().iconst(types::I64, v as i64), types::I64)),
Value::Void => Ok((builder.ins().iconst(types::I64, 0i64), VOID_REPR_TYPE)),
},
ValueOrRef::Ref(_, _, name) => match variables.get(&name) {
None => Err(BackendError::VariableLookupFailure(name)),
Some(ReferenceBuilder::Global(ty, gv)) => {
let pointer_to = self.module.target_config().pointer_type();
let pointer_value = builder.ins().symbol_value(pointer_to, *gv);
let value = builder.ins().load(*ty, MemFlags::new(), pointer_value, 0);
Ok((value, *ty))
}
Some(ReferenceBuilder::Argument(ctype, val)) => Ok((*val, *ctype)),
Some(ReferenceBuilder::Local(ctype, var)) => {
let value = builder.use_var(*var);
Ok((value, *ctype))
}
},
ValueOrRef::Primitive(_, _, _) => {
unimplemented!()
}
}
}
}
impl PrimitiveType {
fn blank_data(&self) -> DataDescription {
let (size, alignment) = match self {
PrimitiveType::Void => (8, 8),
PrimitiveType::U8 => (1, 1),
PrimitiveType::U16 => (2, 2),
PrimitiveType::U32 => (4, 4),
PrimitiveType::U64 => (8, 8),
PrimitiveType::I8 => (1, 1),
PrimitiveType::I16 => (2, 2),
PrimitiveType::I32 => (4, 4),
PrimitiveType::I64 => (8, 8),
};
let mut result = DataDescription::new();
result.define_zeroinit(size);
result.set_align(alignment);
result
}
}