roc/compiler/gen/src/llvm/build.rs

1943 lines
67 KiB
Rust

use crate::layout_id::LayoutIds;
use crate::llvm::build_list::{
allocate_list, empty_polymorphic_list, list_append, list_concat, list_get_unsafe, list_join,
list_len, list_prepend, list_repeat, list_reverse, list_set, list_single,
};
use crate::llvm::compare::{build_eq, build_neq};
use crate::llvm::convert::{basic_type_from_layout, collection, get_fn_type, ptr_int};
use bumpalo::collections::Vec;
use bumpalo::Bump;
use inkwell::basic_block::BasicBlock;
use inkwell::builder::Builder;
use inkwell::context::Context;
use inkwell::memory_buffer::MemoryBuffer;
use inkwell::module::{Linkage, Module};
use inkwell::passes::{PassManager, PassManagerBuilder};
use inkwell::types::{BasicTypeEnum, FunctionType, IntType, StructType};
use inkwell::values::BasicValueEnum::{self, *};
use inkwell::values::{FloatValue, FunctionValue, IntValue, PointerValue, StructValue};
use inkwell::AddressSpace;
use inkwell::{IntPredicate, OptimizationLevel};
use roc_collections::all::{ImMap, MutSet};
use roc_module::low_level::LowLevel;
use roc_module::symbol::{Interns, Symbol};
use roc_mono::ir::JoinPointId;
use roc_mono::layout::{Builtin, Layout, MemoryMode};
use target_lexicon::CallingConvention;
/// This is for Inkwell's FunctionValue::verify - we want to know the verification
/// output in debug builds, but we don't want it to print to stdout in release builds!
#[cfg(debug_assertions)]
const PRINT_FN_VERIFICATION_OUTPUT: bool = true;
#[cfg(not(debug_assertions))]
const PRINT_FN_VERIFICATION_OUTPUT: bool = false;
#[derive(Debug, Clone, Copy)]
pub enum OptLevel {
Normal,
Optimize,
}
// pub type Scope<'a, 'ctx> = ImMap<Symbol, (Layout<'a>, PointerValue<'ctx>)>;
#[derive(Default, Debug, Clone, PartialEq)]
pub struct Scope<'a, 'ctx> {
symbols: ImMap<Symbol, (Layout<'a>, PointerValue<'ctx>)>,
join_points: ImMap<JoinPointId, (BasicBlock<'ctx>, &'a [PointerValue<'ctx>])>,
}
impl<'a, 'ctx> Scope<'a, 'ctx> {
fn get(&self, symbol: &Symbol) -> Option<&(Layout<'a>, PointerValue<'ctx>)> {
self.symbols.get(symbol)
}
fn insert(&mut self, symbol: Symbol, value: (Layout<'a>, PointerValue<'ctx>)) {
self.symbols.insert(symbol, value);
}
fn remove(&mut self, symbol: &Symbol) {
self.symbols.remove(symbol);
}
/*
fn get_join_point(&self, symbol: &JoinPointId) -> Option<&PhiValue<'ctx>> {
self.join_points.get(symbol)
}
fn remove_join_point(&mut self, symbol: &JoinPointId) {
self.join_points.remove(symbol);
}
fn get_mut_join_point(&mut self, symbol: &JoinPointId) -> Option<&mut PhiValue<'ctx>> {
self.join_points.get_mut(symbol)
}
fn insert_join_point(&mut self, symbol: JoinPointId, value: PhiValue<'ctx>) {
self.join_points.insert(symbol, value);
}
*/
}
pub struct Env<'a, 'ctx, 'env> {
pub arena: &'a Bump,
pub context: &'ctx Context,
pub builder: &'env Builder<'ctx>,
pub module: &'ctx Module<'ctx>,
pub interns: Interns,
pub ptr_bytes: u32,
pub leak: bool,
pub exposed_to_host: MutSet<Symbol>,
}
impl<'a, 'ctx, 'env> Env<'a, 'ctx, 'env> {
pub fn ptr_int(&self) -> IntType<'ctx> {
ptr_int(self.context, self.ptr_bytes)
}
}
pub fn module_from_builtins<'ctx>(ctx: &'ctx Context, module_name: &str) -> Module<'ctx> {
let memory_buffer =
MemoryBuffer::create_from_memory_range(include_bytes!("builtins.bc"), module_name);
let module = Module::parse_bitcode_from_buffer(&memory_buffer, ctx)
.unwrap_or_else(|err| panic!("Unable to import builtins bitcode. LLVM error: {:?}", err));
// Add LLVM intrinsics.
add_intrinsics(ctx, &module);
module
}
fn add_intrinsics<'ctx>(ctx: &'ctx Context, module: &Module<'ctx>) {
// List of all supported LLVM intrinsics:
//
// https://releases.llvm.org/10.0.0/docs/LangRef.html#standard-c-library-intrinsics
let i64_type = ctx.i64_type();
let f64_type = ctx.f64_type();
add_intrinsic(
module,
LLVM_SQRT_F64,
f64_type.fn_type(&[f64_type.into()], false),
);
add_intrinsic(
module,
LLVM_LROUND_I64_F64,
i64_type.fn_type(&[f64_type.into()], false),
);
add_intrinsic(
module,
LLVM_FABS_F64,
f64_type.fn_type(&[f64_type.into()], false),
);
add_intrinsic(
module,
LLVM_SIN_F64,
f64_type.fn_type(&[f64_type.into()], false),
);
add_intrinsic(
module,
LLVM_COS_F64,
f64_type.fn_type(&[f64_type.into()], false),
);
}
static LLVM_SQRT_F64: &str = "llvm.sqrt.f64";
static LLVM_LROUND_I64_F64: &str = "llvm.lround.i64.f64";
static LLVM_FABS_F64: &str = "llvm.fabs.f64";
static LLVM_SIN_F64: &str = "llvm.sin.f64";
static LLVM_COS_F64: &str = "llvm.cos.f64";
fn add_intrinsic<'ctx>(
module: &Module<'ctx>,
intrinsic_name: &'static str,
fn_type: FunctionType<'ctx>,
) -> FunctionValue<'ctx> {
let fn_val = module.add_function(intrinsic_name, fn_type, None);
// LLVM intrinsics always use the C calling convention, because
// they are implemented in C libraries
fn_val.set_call_conventions(C_CALL_CONV);
fn_val
}
pub fn construct_optimization_passes<'a>(
module: &'a Module,
opt_level: OptLevel,
) -> (PassManager<Module<'a>>, PassManager<FunctionValue<'a>>) {
let mpm = PassManager::create(());
let fpm = PassManager::create(module);
// tail-call elimination is always on
fpm.add_instruction_combining_pass();
fpm.add_tail_call_elimination_pass();
let pmb = PassManagerBuilder::create();
match opt_level {
OptLevel::Normal => {
pmb.set_optimization_level(OptimizationLevel::None);
}
OptLevel::Optimize => {
// this threshold seems to do what we want
pmb.set_inliner_with_threshold(2);
// TODO figure out which of these actually help
// function passes
fpm.add_cfg_simplification_pass();
mpm.add_cfg_simplification_pass();
fpm.add_jump_threading_pass();
mpm.add_jump_threading_pass();
fpm.add_memcpy_optimize_pass(); // this one is very important
fpm.add_licm_pass();
}
}
pmb.populate_module_pass_manager(&mpm);
pmb.populate_function_pass_manager(&fpm);
fpm.initialize();
// For now, we have just one of each
(mpm, fpm)
}
pub fn build_exp_literal<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
literal: &roc_mono::ir::Literal<'a>,
) -> BasicValueEnum<'ctx> {
use roc_mono::ir::Literal::*;
match literal {
Int(num) => env.context.i64_type().const_int(*num as u64, true).into(),
Float(num) => env.context.f64_type().const_float(*num).into(),
Bool(b) => env.context.bool_type().const_int(*b as u64, false).into(),
Byte(b) => env.context.i8_type().const_int(*b as u64, false).into(),
Str(str_literal) => {
if str_literal.is_empty() {
panic!("TODO build an empty string in LLVM");
} else {
let ctx = env.context;
let builder = env.builder;
let str_len = str_literal.len() + 1/* TODO drop the +1 when we have structs and this is no longer a NUL-terminated CString.*/;
let byte_type = ctx.i8_type();
let nul_terminator = byte_type.const_zero();
let len_val = ctx.i64_type().const_int(str_len as u64, false);
let ptr = env
.builder
.build_array_malloc(ctx.i8_type(), len_val, "str_ptr")
.unwrap();
// TODO check if malloc returned null; if so, runtime error for OOM!
// Copy the bytes from the string literal into the array
for (index, byte) in str_literal.bytes().enumerate() {
let index_val = ctx.i64_type().const_int(index as u64, false);
let elem_ptr =
unsafe { builder.build_in_bounds_gep(ptr, &[index_val], "byte") };
builder.build_store(elem_ptr, byte_type.const_int(byte as u64, false));
}
// Add a NUL terminator at the end.
// TODO: Instead of NUL-terminating, return a struct
// with the pointer and also the length and capacity.
let index_val = ctx.i64_type().const_int(str_len as u64 - 1, false);
let elem_ptr =
unsafe { builder.build_in_bounds_gep(ptr, &[index_val], "nul_terminator") };
builder.build_store(elem_ptr, nul_terminator);
BasicValueEnum::PointerValue(ptr)
}
}
}
}
pub fn build_exp_expr<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
layout_ids: &mut LayoutIds<'a>,
scope: &Scope<'a, 'ctx>,
parent: FunctionValue<'ctx>,
expr: &roc_mono::ir::Expr<'a>,
) -> BasicValueEnum<'ctx> {
use roc_mono::ir::CallType::*;
use roc_mono::ir::Expr::*;
match expr {
Literal(literal) => build_exp_literal(env, literal),
RunLowLevel(op, symbols) => run_low_level(env, scope, parent, *op, symbols),
FunctionCall {
call_type: ByName(name),
layout,
args,
..
} => {
let mut arg_tuples: Vec<BasicValueEnum> = Vec::with_capacity_in(args.len(), env.arena);
for symbol in args.iter() {
arg_tuples.push(load_symbol(env, scope, symbol));
}
call_with_args(
env,
layout_ids,
layout,
*name,
parent,
arg_tuples.into_bump_slice(),
)
}
FunctionCall {
call_type: ByPointer(name),
layout: _,
args,
..
} => {
let sub_expr = load_symbol(env, scope, name);
let mut arg_vals: Vec<BasicValueEnum> = Vec::with_capacity_in(args.len(), env.arena);
for arg in args.iter() {
arg_vals.push(load_symbol(env, scope, arg));
}
let call = match sub_expr {
BasicValueEnum::PointerValue(ptr) => {
env.builder.build_call(ptr, arg_vals.as_slice(), "tmp")
}
non_ptr => {
panic!(
"Tried to call by pointer, but encountered a non-pointer: {:?}",
non_ptr
);
}
};
if env.exposed_to_host.contains(name) {
// If this is an external-facing function, use the C calling convention.
call.set_call_convention(C_CALL_CONV);
} else {
// If it's an internal-only function, use the fast calling conention.
call.set_call_convention(FAST_CALL_CONV);
}
call.try_as_basic_value()
.left()
.unwrap_or_else(|| panic!("LLVM error: Invalid call by pointer."))
}
Struct(sorted_fields) => {
let ctx = env.context;
let builder = env.builder;
let ptr_bytes = env.ptr_bytes;
// Determine types
let num_fields = sorted_fields.len();
let mut field_types = Vec::with_capacity_in(num_fields, env.arena);
let mut field_vals = Vec::with_capacity_in(num_fields, env.arena);
for symbol in sorted_fields.iter() {
// Zero-sized fields have no runtime representation.
// The layout of the struct expects them to be dropped!
let (field_expr, field_layout) = load_symbol_and_layout(env, scope, symbol);
if field_layout.stack_size(ptr_bytes) != 0 {
field_types.push(basic_type_from_layout(
env.arena,
env.context,
&field_layout,
env.ptr_bytes,
));
field_vals.push(field_expr);
}
}
// If the record has only one field that isn't zero-sized,
// unwrap it. This is what the layout expects us to do.
if field_vals.len() == 1 {
field_vals.pop().unwrap()
} else {
// Create the struct_type
let struct_type = ctx.struct_type(field_types.into_bump_slice(), false);
let mut struct_val = struct_type.const_zero().into();
// Insert field exprs into struct_val
for (index, field_val) in field_vals.into_iter().enumerate() {
struct_val = builder
.build_insert_value(struct_val, field_val, index as u32, "insert_field")
.unwrap();
}
BasicValueEnum::StructValue(struct_val.into_struct_value())
}
}
Tag {
union_size,
arguments,
..
} if *union_size == 1 => {
let it = arguments.iter();
let ctx = env.context;
let ptr_bytes = env.ptr_bytes;
let builder = env.builder;
// Determine types
let num_fields = arguments.len() + 1;
let mut field_types = Vec::with_capacity_in(num_fields, env.arena);
let mut field_vals = Vec::with_capacity_in(num_fields, env.arena);
for field_symbol in it {
let (val, field_layout) = load_symbol_and_layout(env, scope, field_symbol);
// Zero-sized fields have no runtime representation.
// The layout of the struct expects them to be dropped!
if field_layout.stack_size(ptr_bytes) != 0 {
let field_type = basic_type_from_layout(
env.arena,
env.context,
&field_layout,
env.ptr_bytes,
);
field_types.push(field_type);
field_vals.push(val);
}
}
// If the struct has only one field that isn't zero-sized,
// unwrap it. This is what the layout expects us to do.
if field_vals.len() == 1 {
field_vals.pop().unwrap()
} else {
// Create the struct_type
let struct_type = ctx.struct_type(field_types.into_bump_slice(), false);
let mut struct_val = struct_type.const_zero().into();
// Insert field exprs into struct_val
for (index, field_val) in field_vals.into_iter().enumerate() {
struct_val = builder
.build_insert_value(struct_val, field_val, index as u32, "insert_field")
.unwrap();
}
BasicValueEnum::StructValue(struct_val.into_struct_value())
}
}
Tag {
arguments,
tag_layout,
union_size,
..
} => {
debug_assert!(*union_size > 1);
let ptr_size = env.ptr_bytes;
let mut filler = tag_layout.stack_size(ptr_size);
let ctx = env.context;
let builder = env.builder;
// Determine types
let num_fields = arguments.len() + 1;
let mut field_types = Vec::with_capacity_in(num_fields, env.arena);
let mut field_vals = Vec::with_capacity_in(num_fields, env.arena);
for field_symbol in arguments.iter() {
let (val, field_layout) = load_symbol_and_layout(env, scope, field_symbol);
let field_size = field_layout.stack_size(ptr_size);
// Zero-sized fields have no runtime representation.
// The layout of the struct expects them to be dropped!
if field_size != 0 {
let field_type =
basic_type_from_layout(env.arena, env.context, field_layout, ptr_size);
field_types.push(field_type);
field_vals.push(val);
filler -= field_size;
}
}
// TODO verify that this is required (better safe than sorry)
if filler > 0 {
field_types.push(env.context.i8_type().array_type(filler).into());
}
// Create the struct_type
let struct_type = ctx.struct_type(field_types.into_bump_slice(), false);
let mut struct_val = struct_type.const_zero().into();
// Insert field exprs into struct_val
for (index, field_val) in field_vals.into_iter().enumerate() {
struct_val = builder
.build_insert_value(struct_val, field_val, index as u32, "insert_field")
.unwrap();
}
// How we create tag values
//
// The memory layout of tags can be different. e.g. in
//
// [ Ok Int, Err Str ]
//
// the `Ok` tag stores a 64-bit integer, the `Err` tag stores a struct.
// All tags of a union must have the same length, for easy addressing (e.g. array lookups).
// So we need to ask for the maximum of all tag's sizes, even if most tags won't use
// all that memory, and certainly won't use it in the same way (the tags have fields of
// different types/sizes)
//
// In llvm, we must be explicit about the type of value we're creating: we can't just
// make a unspecified block of memory. So what we do is create a byte array of the
// desired size. Then when we know which tag we have (which is here, in this function),
// we need to cast that down to the array of bytes that llvm expects
//
// There is the bitcast instruction, but it doesn't work for arrays. So we need to jump
// through some hoops using store and load to get this to work: the array is put into a
// one-element struct, which can be cast to the desired type.
//
// This tricks comes from
// https://github.com/raviqqe/ssf/blob/bc32aae68940d5bddf5984128e85af75ca4f4686/ssf-llvm/src/expression_compiler.rs#L116
let internal_type =
basic_type_from_layout(env.arena, env.context, tag_layout, env.ptr_bytes);
cast_basic_basic(
builder,
struct_val.into_struct_value().into(),
internal_type,
)
}
AccessAtIndex {
index,
structure,
is_unwrapped,
..
} if *is_unwrapped => {
use inkwell::values::BasicValueEnum::*;
let builder = env.builder;
// Get Struct val
// Since this is a one-element tag union, we get the underlying value
// right away. However, that struct might have only one field which
// is not zero-sized, which would make it unwrapped. If that happens,
// we must be
match load_symbol(env, scope, structure) {
StructValue(argument) => builder
.build_extract_value(
argument,
*index as u32,
env.arena.alloc(format!("tag_field_access_{}_", index)),
)
.unwrap(),
other => {
// If it's not a Struct, that means it was unwrapped,
// so we should return it directly.
other
}
}
}
AccessAtIndex {
index,
structure,
field_layouts,
..
} => {
let builder = env.builder;
// Determine types, assumes the descriminant is in the field layouts
let num_fields = field_layouts.len();
let mut field_types = Vec::with_capacity_in(num_fields, env.arena);
let ptr_bytes = env.ptr_bytes;
for field_layout in field_layouts.iter() {
let field_type =
basic_type_from_layout(env.arena, env.context, &field_layout, ptr_bytes);
field_types.push(field_type);
}
// Create the struct_type
let struct_type = env
.context
.struct_type(field_types.into_bump_slice(), false);
// cast the argument bytes into the desired shape for this tag
let argument = load_symbol(env, scope, structure).into_struct_value();
let struct_value = cast_struct_struct(builder, argument, struct_type);
builder
.build_extract_value(struct_value, *index as u32, "")
.expect("desired field did not decode")
}
EmptyArray => empty_polymorphic_list(env),
Array { elem_layout, elems } => list_literal(env, scope, elem_layout, elems),
FunctionPointer(symbol, layout) => {
let fn_name = layout_ids
.get(*symbol, layout)
.to_symbol_string(*symbol, &env.interns);
let ptr = env
.module
.get_function(fn_name.as_str())
.unwrap_or_else(|| panic!("Could not get pointer to unknown function {:?}", symbol))
.as_global_value()
.as_pointer_value();
BasicValueEnum::PointerValue(ptr)
}
RuntimeErrorFunction(_) => todo!(),
}
}
fn list_literal<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
scope: &Scope<'a, 'ctx>,
elem_layout: &Layout<'a>,
elems: &&[Symbol],
) -> BasicValueEnum<'ctx> {
let ctx = env.context;
let builder = env.builder;
let len_u64 = elems.len() as u64;
let elem_bytes = elem_layout.stack_size(env.ptr_bytes) as u64;
let ptr = {
let bytes_len = elem_bytes * len_u64;
let len_type = env.ptr_int();
let len = len_type.const_int(bytes_len, false);
allocate_list(env, elem_layout, len)
// TODO check if malloc returned null; if so, runtime error for OOM!
};
// Copy the elements from the list literal into the array
for (index, symbol) in elems.iter().enumerate() {
let val = load_symbol(env, scope, symbol);
let index_val = ctx.i64_type().const_int(index as u64, false);
let elem_ptr = unsafe { builder.build_in_bounds_gep(ptr, &[index_val], "index") };
builder.build_store(elem_ptr, val);
}
let ptr_bytes = env.ptr_bytes;
let int_type = ptr_int(ctx, ptr_bytes);
let ptr_as_int = builder.build_ptr_to_int(ptr, int_type, "list_cast_ptr");
let struct_type = collection(ctx, ptr_bytes);
let len = BasicValueEnum::IntValue(env.ptr_int().const_int(len_u64, false));
let mut struct_val;
// Store the pointer
struct_val = builder
.build_insert_value(
struct_type.get_undef(),
ptr_as_int,
Builtin::WRAPPER_PTR,
"insert_ptr",
)
.unwrap();
// Store the length
struct_val = builder
.build_insert_value(struct_val, len, Builtin::WRAPPER_LEN, "insert_len")
.unwrap();
// Bitcast to an array of raw bytes
builder.build_bitcast(
struct_val.into_struct_value(),
collection(ctx, ptr_bytes),
"cast_collection",
)
}
pub fn build_exp_stmt<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
layout_ids: &mut LayoutIds<'a>,
scope: &mut Scope<'a, 'ctx>,
parent: FunctionValue<'ctx>,
stmt: &roc_mono::ir::Stmt<'a>,
) -> BasicValueEnum<'ctx> {
use roc_mono::ir::Stmt::*;
match stmt {
Let(symbol, expr, layout, cont) => {
let context = &env.context;
let val = build_exp_expr(env, layout_ids, &scope, parent, &expr);
let expr_bt = basic_type_from_layout(env.arena, context, &layout, env.ptr_bytes);
let alloca =
create_entry_block_alloca(env, parent, expr_bt, symbol.ident_string(&env.interns));
env.builder.build_store(alloca, val);
// Make a new scope which includes the binding we just encountered.
// This should be done *after* compiling the bound expr, since any
// recursive (in the LetRec sense) bindings should already have
// been extracted as procedures. Nothing in here should need to
// access itself!
// scope = scope.clone();
scope.insert(*symbol, (layout.clone(), alloca));
let result = build_exp_stmt(env, layout_ids, scope, parent, cont);
scope.remove(symbol);
result
}
Ret(symbol) => {
let value = load_symbol(env, scope, symbol);
if let Some(block) = env.builder.get_insert_block() {
if block.get_terminator().is_none() {
env.builder.build_return(Some(&value));
}
}
value
}
Cond {
branching_symbol,
pass: pass_stmt,
fail: fail_stmt,
ret_layout,
..
} => {
let ret_type =
basic_type_from_layout(env.arena, env.context, &ret_layout, env.ptr_bytes);
let cond_expr = load_symbol(env, scope, branching_symbol);
match cond_expr {
IntValue(value) => {
// This is a call tobuild_basic_phi2, except inlined to prevent
// problems with lifetimes and closures involving layout_ids.
let builder = env.builder;
let context = env.context;
// build blocks
let then_block = context.append_basic_block(parent, "then");
let else_block = context.append_basic_block(parent, "else");
let mut blocks: std::vec::Vec<(
&dyn inkwell::values::BasicValue<'_>,
inkwell::basic_block::BasicBlock<'_>,
)> = std::vec::Vec::with_capacity(2);
let cont_block = context.append_basic_block(parent, "condbranchcont");
builder.build_conditional_branch(value, then_block, else_block);
// build then block
builder.position_at_end(then_block);
let then_val = build_exp_stmt(env, layout_ids, scope, parent, pass_stmt);
if then_block.get_terminator().is_none() {
builder.build_unconditional_branch(cont_block);
let then_block = builder.get_insert_block().unwrap();
blocks.push((&then_val, then_block));
}
// build else block
builder.position_at_end(else_block);
let else_val = build_exp_stmt(env, layout_ids, scope, parent, fail_stmt);
if else_block.get_terminator().is_none() {
let else_block = builder.get_insert_block().unwrap();
builder.build_unconditional_branch(cont_block);
blocks.push((&else_val, else_block));
}
// emit merge block
if blocks.is_empty() {
// SAFETY there are no other references to this block in this case
unsafe {
cont_block.delete().unwrap();
}
// return garbage value
context.i64_type().const_int(0, false).into()
} else {
builder.position_at_end(cont_block);
let phi = builder.build_phi(ret_type, "branch");
// phi.add_incoming(&[(&then_val, then_block), (&else_val, else_block)]);
phi.add_incoming(&blocks);
phi.as_basic_value()
}
}
_ => panic!(
"Tried to make a branch out of an invalid condition: cond_expr = {:?}",
cond_expr,
),
}
}
Switch {
branches,
default_branch,
ret_layout,
cond_layout,
cond_symbol,
} => {
let ret_type =
basic_type_from_layout(env.arena, env.context, &ret_layout, env.ptr_bytes);
let switch_args = SwitchArgsIr {
cond_layout: cond_layout.clone(),
cond_symbol: *cond_symbol,
branches,
default_branch,
ret_type,
};
build_switch_ir(env, layout_ids, scope, parent, switch_args)
}
Join {
id,
parameters,
remainder,
continuation,
} => {
let builder = env.builder;
let context = env.context;
let mut joinpoint_args = Vec::with_capacity_in(parameters.len(), env.arena);
for param in parameters.iter() {
let btype =
basic_type_from_layout(env.arena, env.context, &param.layout, env.ptr_bytes);
joinpoint_args.push(create_entry_block_alloca(
env,
parent,
btype,
"joinpointarg",
));
}
// create new block
let cont_block = context.append_basic_block(parent, "joinpointcont");
// store this join point
let joinpoint_args = joinpoint_args.into_bump_slice();
scope.join_points.insert(*id, (cont_block, joinpoint_args));
// construct the blocks that may jump to this join point
build_exp_stmt(env, layout_ids, scope, parent, remainder);
for (ptr, param) in joinpoint_args.iter().zip(parameters.iter()) {
scope.insert(param.symbol, (param.layout.clone(), *ptr));
}
let phi_block = builder.get_insert_block().unwrap();
// put the cont block at the back
builder.position_at_end(cont_block);
// put the continuation in
let result = build_exp_stmt(env, layout_ids, scope, parent, continuation);
// remove this join point again
scope.join_points.remove(&id);
cont_block.move_after(phi_block).unwrap();
result
}
Jump(join_point, arguments) => {
let builder = env.builder;
let context = env.context;
let (cont_block, argument_pointers) = scope.join_points.get(join_point).unwrap();
for (pointer, argument) in argument_pointers.iter().zip(arguments.iter()) {
let value = load_symbol(env, scope, argument);
builder.build_store(*pointer, value);
}
builder.build_unconditional_branch(*cont_block);
// This doesn't currently do anything
context.i64_type().const_zero().into()
}
Inc(symbol, cont) => {
let (value, layout) = load_symbol_and_layout(env, scope, symbol);
let layout = layout.clone();
match layout {
Layout::Builtin(Builtin::List(MemoryMode::Refcounted, _)) => {
increment_refcount_list(env, value.into_struct_value());
build_exp_stmt(env, layout_ids, scope, parent, cont)
}
_ => build_exp_stmt(env, layout_ids, scope, parent, cont),
}
}
Dec(symbol, cont) => {
let (value, layout) = load_symbol_and_layout(env, scope, symbol);
let layout = layout.clone();
if layout.contains_refcounted() {
decrement_refcount_layout(env, parent, value, &layout);
}
build_exp_stmt(env, layout_ids, scope, parent, cont)
}
_ => todo!("unsupported expr {:?}", stmt),
}
}
fn refcount_is_one_comparison<'ctx>(
builder: &Builder<'ctx>,
context: &'ctx Context,
refcount: IntValue<'ctx>,
) -> IntValue<'ctx> {
let refcount_one: IntValue<'ctx> = context.i64_type().const_int((std::usize::MAX) as _, false);
// Note: Check for refcount < refcount_1 as the "true" condition,
// to avoid misprediction. (In practice this should usually pass,
// and CPUs generally default to predicting that a forward jump
// shouldn't be taken; that is, they predict "else" won't be taken.)
builder.build_int_compare(
IntPredicate::EQ,
refcount,
refcount_one,
"refcount_one_check",
)
}
#[allow(dead_code)]
fn list_get_refcount_ptr<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
list_wrapper: StructValue<'ctx>,
) -> PointerValue<'ctx> {
let builder = env.builder;
let ctx = env.context;
// pointer to usize
let ptr_bytes = env.ptr_bytes;
let int_type = ptr_int(ctx, ptr_bytes);
// fetch the pointer to the array data, as an integer
let ptr_as_int = builder
.build_extract_value(list_wrapper, Builtin::WRAPPER_PTR, "read_list_ptr")
.unwrap()
.into_int_value();
// subtract ptr_size, to access the refcount
let refcount_ptr = builder.build_int_sub(
ptr_as_int,
ctx.i64_type().const_int(env.ptr_bytes as u64, false),
"make_refcount_ptr",
);
builder.build_int_to_ptr(
refcount_ptr,
int_type.ptr_type(AddressSpace::Generic),
"get_refcount_ptr",
)
}
fn decrement_refcount_layout<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
parent: FunctionValue<'ctx>,
value: BasicValueEnum<'ctx>,
layout: &Layout<'a>,
) {
use Layout::*;
match layout {
Builtin(builtin) => decrement_refcount_builtin(env, parent, value, builtin),
Struct(layouts) => {
let wrapper_struct = value.into_struct_value();
for (i, field_layout) in layouts.iter().enumerate() {
if field_layout.contains_refcounted() {
let field_ptr = env
.builder
.build_extract_value(wrapper_struct, i as u32, "decrement_struct_field")
.unwrap();
decrement_refcount_layout(env, parent, field_ptr, field_layout)
}
}
}
Union(tags) => {
debug_assert!(!tags.is_empty());
let wrapper_struct = value.into_struct_value();
// read the tag_id
let tag_id = env
.builder
.build_extract_value(wrapper_struct, 0, "read_tag_id")
.unwrap()
.into_int_value();
// next, make a jump table for all possible values of the tag_id
let mut cases = Vec::with_capacity_in(tags.len(), env.arena);
let merge_block = env.context.append_basic_block(parent, "decrement_merge");
for (tag_id, field_layouts) in tags.iter().enumerate() {
let block = env.context.append_basic_block(parent, "tag_id_decrement");
env.builder.position_at_end(block);
for (i, field_layout) in field_layouts.iter().enumerate() {
if field_layout.contains_refcounted() {
let field_ptr = env
.builder
.build_extract_value(wrapper_struct, i as u32, "decrement_struct_field")
.unwrap();
decrement_refcount_layout(env, parent, field_ptr, field_layout)
}
}
env.builder.build_unconditional_branch(merge_block);
cases.push((env.context.i8_type().const_int(tag_id as u64, false), block));
}
let (_, default_block) = cases.pop().unwrap();
env.builder.build_switch(tag_id, default_block, &cases);
env.builder.position_at_end(merge_block);
}
FunctionPointer(_, _) | Pointer(_) => {}
}
}
#[inline(always)]
fn decrement_refcount_builtin<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
parent: FunctionValue<'ctx>,
value: BasicValueEnum<'ctx>,
builtin: &Builtin<'a>,
) {
use Builtin::*;
match builtin {
List(MemoryMode::Refcounted, element_layout) => {
if element_layout.contains_refcounted() {
// TODO decrement all values
}
let wrapper_struct = value.into_struct_value();
decrement_refcount_list(env, parent, wrapper_struct);
}
List(MemoryMode::Unique, _element_layout) => {
// do nothing
}
Set(element_layout) => {
if element_layout.contains_refcounted() {
// TODO decrement all values
}
let wrapper_struct = value.into_struct_value();
decrement_refcount_list(env, parent, wrapper_struct);
}
Map(key_layout, value_layout) => {
if key_layout.contains_refcounted() || value_layout.contains_refcounted() {
// TODO decrement all values
}
let wrapper_struct = value.into_struct_value();
decrement_refcount_list(env, parent, wrapper_struct);
}
_ => {}
}
}
fn increment_refcount_list<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
original_wrapper: StructValue<'ctx>,
) {
let builder = env.builder;
let ctx = env.context;
let refcount_ptr = list_get_refcount_ptr(env, original_wrapper);
let refcount = env
.builder
.build_load(refcount_ptr, "get_refcount")
.into_int_value();
// our refcount 0 is actually usize::MAX, so incrementing the refcount means decrementing this value.
let decremented = env.builder.build_int_sub(
refcount,
ctx.i64_type().const_int(1 as u64, false),
"incremented_refcount",
);
// Mutate the new array in-place to change the element.
builder.build_store(refcount_ptr, decremented);
}
fn decrement_refcount_list<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
parent: FunctionValue<'ctx>,
original_wrapper: StructValue<'ctx>,
) {
let builder = env.builder;
let ctx = env.context;
let refcount_ptr = list_get_refcount_ptr(env, original_wrapper);
let refcount = env
.builder
.build_load(refcount_ptr, "get_refcount")
.into_int_value();
let comparison = refcount_is_one_comparison(builder, env.context, refcount);
// build blocks
let then_block = ctx.append_basic_block(parent, "then");
let else_block = ctx.append_basic_block(parent, "else");
let cont_block = ctx.append_basic_block(parent, "dec_ref_branchcont");
builder.build_conditional_branch(comparison, then_block, else_block);
// build then block
{
builder.position_at_end(then_block);
// our refcount 0 is actually usize::MAX, so decrementing the refcount means incrementing this value.
let decremented = env.builder.build_int_add(
ctx.i64_type().const_int(1 as u64, false),
refcount,
"decremented_refcount",
);
// Mutate the new array in-place to change the element.
builder.build_store(refcount_ptr, decremented);
builder.build_unconditional_branch(cont_block);
}
// build else block
{
builder.position_at_end(else_block);
if !env.leak {
let free = builder.build_free(refcount_ptr);
builder.insert_instruction(&free, None);
}
builder.build_unconditional_branch(cont_block);
}
// emit merge block
builder.position_at_end(cont_block);
}
pub fn load_symbol<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
scope: &Scope<'a, 'ctx>,
symbol: &Symbol,
) -> BasicValueEnum<'ctx> {
match scope.get(symbol) {
Some((_, ptr)) => env
.builder
.build_load(*ptr, symbol.ident_string(&env.interns)),
None => panic!("There was no entry for {:?} in scope {:?}", symbol, scope),
}
}
pub fn load_symbol_and_layout<'a, 'ctx, 'env, 'b>(
env: &Env<'a, 'ctx, 'env>,
scope: &'b Scope<'a, 'ctx>,
symbol: &Symbol,
) -> (BasicValueEnum<'ctx>, &'b Layout<'a>) {
match scope.get(symbol) {
Some((layout, ptr)) => (
env.builder
.build_load(*ptr, symbol.ident_string(&env.interns)),
layout,
),
None => panic!("There was no entry for {:?} in scope {:?}", symbol, scope),
}
}
/// Cast a struct to another struct of the same (or smaller?) size
fn cast_struct_struct<'ctx>(
builder: &Builder<'ctx>,
from_value: StructValue<'ctx>,
to_type: StructType<'ctx>,
) -> StructValue<'ctx> {
cast_basic_basic(builder, from_value.into(), to_type.into()).into_struct_value()
}
/// Cast a value to another value of the same (or smaller?) size
fn cast_basic_basic<'ctx>(
builder: &Builder<'ctx>,
from_value: BasicValueEnum<'ctx>,
to_type: BasicTypeEnum<'ctx>,
) -> BasicValueEnum<'ctx> {
use inkwell::types::BasicType;
// store the value in memory
let argument_pointer = builder.build_alloca(from_value.get_type(), "");
builder.build_store(argument_pointer, from_value);
// then read it back as a different type
let to_type_pointer = builder
.build_bitcast(
argument_pointer,
to_type.ptr_type(inkwell::AddressSpace::Generic),
"cast_basic_basic",
)
.into_pointer_value();
builder.build_load(to_type_pointer, "")
}
fn extract_tag_discriminant<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
from_value: StructValue<'ctx>,
) -> IntValue<'ctx> {
let struct_type = env
.context
.struct_type(&[env.context.i64_type().into()], false);
let struct_value = cast_struct_struct(env.builder, from_value, struct_type);
env.builder
.build_extract_value(struct_value, 0, "")
.expect("desired field did not decode")
.into_int_value()
}
struct SwitchArgsIr<'a, 'ctx> {
pub cond_symbol: Symbol,
pub cond_layout: Layout<'a>,
pub branches: &'a [(u64, roc_mono::ir::Stmt<'a>)],
pub default_branch: &'a roc_mono::ir::Stmt<'a>,
pub ret_type: BasicTypeEnum<'ctx>,
}
fn build_switch_ir<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
layout_ids: &mut LayoutIds<'a>,
scope: &Scope<'a, 'ctx>,
parent: FunctionValue<'ctx>,
switch_args: SwitchArgsIr<'a, 'ctx>,
) -> BasicValueEnum<'ctx> {
let arena = env.arena;
let builder = env.builder;
let context = env.context;
let SwitchArgsIr {
branches,
cond_symbol,
mut cond_layout,
default_branch,
ret_type,
..
} = switch_args;
let mut copy = scope.clone();
let scope = &mut copy;
let cond_symbol = &cond_symbol;
let cont_block = context.append_basic_block(parent, "cont");
// Build the condition
let cond = match cond_layout {
Layout::Builtin(Builtin::Float64) => {
// float matches are done on the bit pattern
cond_layout = Layout::Builtin(Builtin::Int64);
let full_cond = load_symbol(env, scope, cond_symbol);
builder
.build_bitcast(full_cond, env.context.i64_type(), "")
.into_int_value()
}
Layout::Union(_) => {
// we match on the discriminant, not the whole Tag
cond_layout = Layout::Builtin(Builtin::Int64);
let full_cond = load_symbol(env, scope, cond_symbol).into_struct_value();
extract_tag_discriminant(env, full_cond)
}
Layout::Builtin(_) => load_symbol(env, scope, cond_symbol).into_int_value(),
other => todo!("Build switch value from layout: {:?}", other),
};
// Build the cases
let mut incoming = Vec::with_capacity_in(branches.len(), arena);
let mut cases = Vec::with_capacity_in(branches.len(), arena);
for (int, _) in branches.iter() {
// Switch constants must all be same type as switch value!
// e.g. this is incorrect, and will trigger a LLVM warning:
//
// switch i8 %apple1, label %default [
// i64 2, label %branch2
// i64 0, label %branch0
// i64 1, label %branch1
// ]
//
// they either need to all be i8, or i64
let int_val = match cond_layout {
Layout::Builtin(Builtin::Int128) => context.i128_type().const_int(*int as u64, false), /* TODO file an issue: you can't currently have an int literal bigger than 64 bits long, and also (as we see here), you can't currently have (at least in Inkwell) a when-branch with an i128 literal in its pattren */
Layout::Builtin(Builtin::Int64) => context.i64_type().const_int(*int as u64, false),
Layout::Builtin(Builtin::Int32) => context.i32_type().const_int(*int as u64, false),
Layout::Builtin(Builtin::Int16) => context.i16_type().const_int(*int as u64, false),
Layout::Builtin(Builtin::Int8) => context.i8_type().const_int(*int as u64, false),
Layout::Builtin(Builtin::Int1) => context.bool_type().const_int(*int as u64, false),
_ => panic!("Can't cast to cond_layout = {:?}", cond_layout),
};
let block = context.append_basic_block(parent, format!("branch{}", int).as_str());
cases.push((int_val, block));
}
let default_block = context.append_basic_block(parent, "default");
builder.build_switch(cond, default_block, &cases);
for ((_, branch_expr), (_, block)) in branches.iter().zip(cases) {
builder.position_at_end(block);
let branch_val = build_exp_stmt(env, layout_ids, scope, parent, branch_expr);
if block.get_terminator().is_none() {
builder.build_unconditional_branch(cont_block);
incoming.push((branch_val, block));
}
}
// The block for the conditional's default branch.
builder.position_at_end(default_block);
let default_val = build_exp_stmt(env, layout_ids, scope, parent, default_branch);
if default_block.get_terminator().is_none() {
builder.build_unconditional_branch(cont_block);
incoming.push((default_val, default_block));
}
// emit merge block
if incoming.is_empty() {
unsafe {
cont_block.delete().unwrap();
}
// produce unused garbage value
context.i64_type().const_zero().into()
} else {
builder.position_at_end(cont_block);
let phi = builder.build_phi(ret_type, "branch");
for (branch_val, block) in incoming {
phi.add_incoming(&[(&Into::<BasicValueEnum>::into(branch_val), block)]);
}
phi.as_basic_value()
}
}
/// TODO could this be added to Inkwell itself as a method on BasicValueEnum?
fn set_name(bv_enum: BasicValueEnum<'_>, name: &str) {
match bv_enum {
ArrayValue(val) => val.set_name(name),
IntValue(val) => val.set_name(name),
FloatValue(val) => val.set_name(name),
PointerValue(val) => val.set_name(name),
StructValue(val) => val.set_name(name),
VectorValue(val) => val.set_name(name),
}
}
/// Creates a new stack allocation instruction in the entry block of the function.
pub fn create_entry_block_alloca<'a, 'ctx>(
env: &Env<'a, 'ctx, '_>,
parent: FunctionValue<'_>,
basic_type: BasicTypeEnum<'ctx>,
name: &str,
) -> PointerValue<'ctx> {
let builder = env.context.create_builder();
let entry = parent.get_first_basic_block().unwrap();
match entry.get_first_instruction() {
Some(first_instr) => builder.position_before(&first_instr),
None => builder.position_at_end(entry),
}
builder.build_alloca(basic_type, name)
}
pub fn build_proc_header<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
layout_ids: &mut LayoutIds<'a>,
symbol: Symbol,
layout: &Layout<'a>,
proc: &roc_mono::ir::Proc<'a>,
) -> (FunctionValue<'ctx>, Vec<'a, BasicTypeEnum<'ctx>>) {
let args = proc.args;
let arena = env.arena;
let context = &env.context;
let ret_type = basic_type_from_layout(arena, context, &proc.ret_layout, env.ptr_bytes);
let mut arg_basic_types = Vec::with_capacity_in(args.len(), arena);
let mut arg_symbols = Vec::new_in(arena);
for (layout, arg_symbol) in args.iter() {
let arg_type = basic_type_from_layout(arena, env.context, &layout, env.ptr_bytes);
arg_basic_types.push(arg_type);
arg_symbols.push(arg_symbol);
}
let fn_type = get_fn_type(&ret_type, &arg_basic_types);
let fn_name = layout_ids
.get(symbol, layout)
.to_symbol_string(symbol, &env.interns);
let fn_val = env
.module
.add_function(fn_name.as_str(), fn_type, Some(Linkage::Private));
if env.exposed_to_host.contains(&symbol) {
// If this is an external-facing function, it'll use the C calling convention
// and external linkage.
fn_val.set_linkage(Linkage::External);
fn_val.set_call_conventions(C_CALL_CONV);
} else {
// If it's an internal-only function, it should use the fast calling conention.
fn_val.set_call_conventions(FAST_CALL_CONV);
}
(fn_val, arg_basic_types)
}
pub fn build_proc<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
layout_ids: &mut LayoutIds<'a>,
proc: roc_mono::ir::Proc<'a>,
fn_val: FunctionValue<'ctx>,
arg_basic_types: Vec<'a, BasicTypeEnum<'ctx>>,
) {
let args = proc.args;
let context = &env.context;
// Add a basic block for the entry point
let entry = context.append_basic_block(fn_val, "entry");
let builder = env.builder;
builder.position_at_end(entry);
let mut scope = Scope::default();
// Add args to scope
for ((arg_val, arg_type), (layout, arg_symbol)) in
fn_val.get_param_iter().zip(arg_basic_types).zip(args)
{
set_name(arg_val, arg_symbol.ident_string(&env.interns));
let alloca =
create_entry_block_alloca(env, fn_val, arg_type, arg_symbol.ident_string(&env.interns));
builder.build_store(alloca, arg_val);
scope.insert(*arg_symbol, (layout.clone(), alloca));
}
let body = build_exp_stmt(env, layout_ids, &mut scope, fn_val, &proc.body);
// only add a return if codegen did not already add one
if let Some(block) = builder.get_insert_block() {
if block.get_terminator().is_none() {
builder.build_return(Some(&body));
}
}
}
pub fn verify_fn(fn_val: FunctionValue<'_>) {
if !fn_val.verify(PRINT_FN_VERIFICATION_OUTPUT) {
unsafe {
fn_val.delete();
}
panic!("Invalid generated fn_val.")
}
}
// #[allow(clippy::cognitive_complexity)]
#[inline(always)]
fn call_with_args<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
layout_ids: &mut LayoutIds<'a>,
layout: &Layout<'a>,
symbol: Symbol,
_parent: FunctionValue<'ctx>,
args: &[BasicValueEnum<'ctx>],
) -> BasicValueEnum<'ctx> {
let fn_name = layout_ids
.get(symbol, layout)
.to_symbol_string(symbol, &env.interns);
let fn_val = env
.module
.get_function(fn_name.as_str())
.unwrap_or_else(|| {
if symbol.is_builtin() {
panic!("Unrecognized builtin function: {:?}", symbol)
} else {
panic!("Unrecognized non-builtin function: {:?}", symbol)
}
});
let call = env.builder.build_call(fn_val, args, "call");
call.set_call_convention(fn_val.get_call_conventions());
call.try_as_basic_value()
.left()
.unwrap_or_else(|| panic!("LLVM error: Invalid call by name for name {:?}", symbol))
}
fn call_intrinsic<'a, 'ctx, 'env>(
intrinsic_name: &'static str,
env: &Env<'a, 'ctx, 'env>,
args: &[(BasicValueEnum<'ctx>, &'a Layout<'a>)],
) -> BasicValueEnum<'ctx> {
let fn_val = env
.module
.get_function(intrinsic_name)
.unwrap_or_else(|| panic!("Unrecognized intrinsic function: {}", intrinsic_name));
let mut arg_vals: Vec<BasicValueEnum> = Vec::with_capacity_in(args.len(), env.arena);
for (arg, _layout) in args.iter() {
arg_vals.push(*arg);
}
let call = env
.builder
.build_call(fn_val, arg_vals.into_bump_slice(), "call");
call.set_call_convention(fn_val.get_call_conventions());
call.try_as_basic_value().left().unwrap_or_else(|| {
panic!(
"LLVM error: Invalid call by name for intrinsic {}",
intrinsic_name
)
})
}
pub enum InPlace {
InPlace,
Clone,
}
/// Translates a target_lexicon::Triple to a LLVM calling convention u32
/// as described in https://llvm.org/doxygen/namespacellvm_1_1CallingConv.html
pub fn get_call_conventions(cc: CallingConvention) -> u32 {
use CallingConvention::*;
// For now, we're returning 0 for the C calling convention on all of these.
// Not sure if we should be picking something more specific!
match cc {
SystemV => C_CALL_CONV,
WasmBasicCAbi => C_CALL_CONV,
WindowsFastcall => C_CALL_CONV,
}
}
/// Source: https://llvm.org/doxygen/namespacellvm_1_1CallingConv.html
pub static C_CALL_CONV: u32 = 0;
pub static FAST_CALL_CONV: u32 = 8;
pub static COLD_CALL_CONV: u32 = 9;
fn run_low_level<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
scope: &Scope<'a, 'ctx>,
parent: FunctionValue<'ctx>,
op: LowLevel,
args: &[Symbol],
) -> BasicValueEnum<'ctx> {
use LowLevel::*;
match op {
ListLen => {
// List.len : List * -> Int
debug_assert_eq!(args.len(), 1);
let arg = load_symbol(env, scope, &args[0]);
list_len(env.builder, arg.into_struct_value()).into()
}
ListSingle => {
// List.single : a -> List a
debug_assert_eq!(args.len(), 1);
let (arg, arg_layout) = load_symbol_and_layout(env, scope, &args[0]);
list_single(env, arg, arg_layout)
}
ListRepeat => {
// List.repeat : Int, elem -> List elem
debug_assert_eq!(args.len(), 2);
let list_len = load_symbol(env, scope, &args[0]).into_int_value();
let (elem, elem_layout) = load_symbol_and_layout(env, scope, &args[1]);
list_repeat(env, parent, list_len, elem, elem_layout)
}
ListReverse => {
// List.reverse : List elem -> List elem
debug_assert_eq!(args.len(), 1);
let list = &args[0];
list_reverse(env, parent, scope, list)
}
ListConcat => list_concat(env, scope, parent, args),
ListAppend => {
// List.append : List elem, elem -> List elem
debug_assert_eq!(args.len(), 2);
let original_wrapper = load_symbol(env, scope, &args[0]).into_struct_value();
let (elem, elem_layout) = load_symbol_and_layout(env, scope, &args[1]);
list_append(env, original_wrapper, elem, elem_layout)
}
ListPrepend => {
// List.prepend : List elem, elem -> List elem
debug_assert_eq!(args.len(), 2);
let original_wrapper = load_symbol(env, scope, &args[0]).into_struct_value();
let (elem, elem_layout) = load_symbol_and_layout(env, scope, &args[1]);
list_prepend(env, original_wrapper, elem, elem_layout)
}
ListJoin => {
// List.join : List (List elem) -> List elem
debug_assert_eq!(args.len(), 1);
let (list, outer_list_layout) = load_symbol_and_layout(env, scope, &args[0]);
let outer_wrapper_struct = list.into_struct_value();
list_join(env, parent, outer_wrapper_struct, outer_list_layout)
}
NumAbs | NumNeg | NumRound | NumSqrtUnchecked | NumSin | NumCos | NumToFloat => {
debug_assert_eq!(args.len(), 1);
let (arg, arg_layout) = load_symbol_and_layout(env, scope, &args[0]);
match arg_layout {
Layout::Builtin(arg_builtin) => {
use roc_mono::layout::Builtin::*;
match arg_builtin {
Int128 | Int64 | Int32 | Int16 | Int8 => {
build_int_unary_op(env, arg.into_int_value(), arg_layout, op)
}
Float128 | Float64 | Float32 | Float16 => {
build_float_unary_op(env, arg.into_float_value(), arg_layout, op)
}
_ => {
unreachable!("Compiler bug: tried to run numeric operation {:?} on invalid builtin layout: ({:?})", op, arg_layout);
}
}
}
_ => {
unreachable!(
"Compiler bug: tried to run numeric operation {:?} on invalid layout: {:?}",
op, arg_layout
);
}
}
}
NumAdd | NumSub | NumMul | NumLt | NumLte | NumGt | NumGte | NumRemUnchecked
| NumDivUnchecked => {
debug_assert_eq!(args.len(), 2);
let (lhs_arg, lhs_layout) = load_symbol_and_layout(env, scope, &args[0]);
let (rhs_arg, rhs_layout) = load_symbol_and_layout(env, scope, &args[1]);
match (lhs_layout, rhs_layout) {
(Layout::Builtin(lhs_builtin), Layout::Builtin(rhs_builtin))
if lhs_builtin == rhs_builtin =>
{
use roc_mono::layout::Builtin::*;
match lhs_builtin {
Int128 | Int64 | Int32 | Int16 | Int8 => build_int_binop(
env,
lhs_arg.into_int_value(),
lhs_layout,
rhs_arg.into_int_value(),
rhs_layout,
op,
),
Float128 | Float64 | Float32 | Float16 => build_float_binop(
env,
lhs_arg.into_float_value(),
lhs_layout,
rhs_arg.into_float_value(),
rhs_layout,
op,
),
_ => {
unreachable!("Compiler bug: tried to run numeric operation {:?} on invalid builtin layout: ({:?})", op, lhs_layout);
}
}
}
_ => {
unreachable!("Compiler bug: tried to run numeric operation {:?} on invalid layouts. The 2 layouts were: ({:?}) and ({:?})", op, lhs_layout, rhs_layout);
}
}
}
Eq => {
debug_assert_eq!(args.len(), 2);
let (lhs_arg, lhs_layout) = load_symbol_and_layout(env, scope, &args[0]);
let (rhs_arg, rhs_layout) = load_symbol_and_layout(env, scope, &args[1]);
build_eq(env, lhs_arg, rhs_arg, lhs_layout, rhs_layout)
}
NotEq => {
debug_assert_eq!(args.len(), 2);
let (lhs_arg, lhs_layout) = load_symbol_and_layout(env, scope, &args[0]);
let (rhs_arg, rhs_layout) = load_symbol_and_layout(env, scope, &args[1]);
build_neq(env, lhs_arg, rhs_arg, lhs_layout, rhs_layout)
}
And => {
// The (&&) operator
debug_assert_eq!(args.len(), 2);
let lhs_arg = load_symbol(env, scope, &args[0]);
let rhs_arg = load_symbol(env, scope, &args[1]);
let bool_val = env.builder.build_and(
lhs_arg.into_int_value(),
rhs_arg.into_int_value(),
"bool_and",
);
BasicValueEnum::IntValue(bool_val)
}
Or => {
// The (||) operator
debug_assert_eq!(args.len(), 2);
let lhs_arg = load_symbol(env, scope, &args[0]);
let rhs_arg = load_symbol(env, scope, &args[1]);
let bool_val = env.builder.build_or(
lhs_arg.into_int_value(),
rhs_arg.into_int_value(),
"bool_or",
);
BasicValueEnum::IntValue(bool_val)
}
Not => {
// The (!) operator
debug_assert_eq!(args.len(), 1);
let arg = load_symbol(env, scope, &args[0]);
let bool_val = env.builder.build_not(arg.into_int_value(), "bool_not");
BasicValueEnum::IntValue(bool_val)
}
ListGetUnsafe => {
// List.get : List elem, Int -> [ Ok elem, OutOfBounds ]*
debug_assert_eq!(args.len(), 2);
let (wrapper_struct, list_layout) = load_symbol_and_layout(env, scope, &args[0]);
let wrapper_struct = wrapper_struct.into_struct_value();
let elem_index = load_symbol(env, scope, &args[1]).into_int_value();
list_get_unsafe(env, list_layout, elem_index, wrapper_struct)
}
ListSet => {
let (list_symbol, list_layout) = load_symbol_and_layout(env, scope, &args[0]);
let in_place = match &list_layout {
Layout::Builtin(Builtin::List(MemoryMode::Unique, _)) => InPlace::InPlace,
_ => InPlace::Clone,
};
list_set(
parent,
&[
(list_symbol, list_layout),
(load_symbol_and_layout(env, scope, &args[1])),
(load_symbol_and_layout(env, scope, &args[2])),
],
env,
in_place,
)
}
ListSetInPlace => list_set(
parent,
&[
(load_symbol_and_layout(env, scope, &args[0])),
(load_symbol_and_layout(env, scope, &args[1])),
(load_symbol_and_layout(env, scope, &args[2])),
],
env,
InPlace::InPlace,
),
}
}
fn build_int_binop<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
lhs: IntValue<'ctx>,
_lhs_layout: &Layout<'a>,
rhs: IntValue<'ctx>,
_rhs_layout: &Layout<'a>,
op: LowLevel,
) -> BasicValueEnum<'ctx> {
use inkwell::IntPredicate::*;
use roc_module::low_level::LowLevel::*;
let bd = env.builder;
match op {
NumAdd => bd.build_int_add(lhs, rhs, "add_int").into(),
NumSub => bd.build_int_sub(lhs, rhs, "sub_int").into(),
NumMul => bd.build_int_mul(lhs, rhs, "mul_int").into(),
NumGt => bd.build_int_compare(SGT, lhs, rhs, "int_gt").into(),
NumGte => bd.build_int_compare(SGE, lhs, rhs, "int_gte").into(),
NumLt => bd.build_int_compare(SLT, lhs, rhs, "int_lt").into(),
NumLte => bd.build_int_compare(SLE, lhs, rhs, "int_lte").into(),
NumRemUnchecked => bd.build_int_signed_rem(lhs, rhs, "rem_int").into(),
NumDivUnchecked => bd.build_int_signed_div(lhs, rhs, "div_int").into(),
_ => {
unreachable!("Unrecognized int binary operation: {:?}", op);
}
}
}
fn build_float_binop<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
lhs: FloatValue<'ctx>,
_lhs_layout: &Layout<'a>,
rhs: FloatValue<'ctx>,
_rhs_layout: &Layout<'a>,
op: LowLevel,
) -> BasicValueEnum<'ctx> {
use inkwell::FloatPredicate::*;
use roc_module::low_level::LowLevel::*;
let bd = env.builder;
match op {
NumAdd => bd.build_float_add(lhs, rhs, "add_float").into(),
NumSub => bd.build_float_sub(lhs, rhs, "sub_float").into(),
NumMul => bd.build_float_mul(lhs, rhs, "mul_float").into(),
NumGt => bd.build_float_compare(OGT, lhs, rhs, "float_gt").into(),
NumGte => bd.build_float_compare(OGE, lhs, rhs, "float_gte").into(),
NumLt => bd.build_float_compare(OLT, lhs, rhs, "float_lt").into(),
NumLte => bd.build_float_compare(OLE, lhs, rhs, "float_lte").into(),
NumRemUnchecked => bd.build_float_rem(lhs, rhs, "rem_float").into(),
NumDivUnchecked => bd.build_float_div(lhs, rhs, "div_float").into(),
_ => {
unreachable!("Unrecognized int binary operation: {:?}", op);
}
}
}
fn build_int_unary_op<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
arg: IntValue<'ctx>,
arg_layout: &Layout<'a>,
op: LowLevel,
) -> BasicValueEnum<'ctx> {
use roc_module::low_level::LowLevel::*;
let bd = env.builder;
match op {
NumNeg => bd.build_int_neg(arg, "negate_int").into(),
NumAbs => {
// This is how libc's abs() is implemented - it uses no branching!
//
// abs = \arg ->
// shifted = arg >>> 63
//
// (xor arg shifted) - shifted
let ctx = env.context;
let shifted_name = "abs_shift_right";
let shifted_alloca = {
let bits_to_shift = ((arg_layout.stack_size(env.ptr_bytes) as u64) * 8) - 1;
let shift_val = ctx.i64_type().const_int(bits_to_shift, false);
let shifted = bd.build_right_shift(arg, shift_val, true, shifted_name);
let alloca = bd.build_alloca(
basic_type_from_layout(env.arena, ctx, arg_layout, env.ptr_bytes),
"#int_abs_help",
);
// shifted = arg >>> 63
bd.build_store(alloca, shifted);
alloca
};
let xored_arg = bd.build_xor(
arg,
bd.build_load(shifted_alloca, shifted_name).into_int_value(),
"xor_arg_shifted",
);
BasicValueEnum::IntValue(bd.build_int_sub(
xored_arg,
bd.build_load(shifted_alloca, shifted_name).into_int_value(),
"sub_xored_shifted",
))
}
NumToFloat => {
// TODO specialize this to be not just for i64!
let builtin_fn_name = "i64_to_f64_";
let fn_val = env
.module
.get_function(builtin_fn_name)
.unwrap_or_else(|| panic!("Unrecognized builtin function: {:?} - if you're working on the Roc compiler, do you need to rebuild the bitcode? See compiler/builtins/bitcode/README.md", builtin_fn_name));
let call = env
.builder
.build_call(fn_val, &[arg.into()], "call_builtin");
call.set_call_convention(fn_val.get_call_conventions());
call.try_as_basic_value()
.left()
.unwrap_or_else(|| panic!("LLVM error: Invalid call for low-level op {:?}", op))
}
_ => {
unreachable!("Unrecognized int unary operation: {:?}", op);
}
}
}
fn build_float_unary_op<'a, 'ctx, 'env>(
env: &Env<'a, 'ctx, 'env>,
arg: FloatValue<'ctx>,
arg_layout: &Layout<'a>,
op: LowLevel,
) -> BasicValueEnum<'ctx> {
use roc_module::low_level::LowLevel::*;
let bd = env.builder;
match op {
NumNeg => bd.build_float_neg(arg, "negate_float").into(),
NumAbs => call_intrinsic(LLVM_FABS_F64, env, &[(arg.into(), arg_layout)]),
NumSqrtUnchecked => call_intrinsic(LLVM_SQRT_F64, env, &[(arg.into(), arg_layout)]),
NumRound => call_intrinsic(LLVM_LROUND_I64_F64, env, &[(arg.into(), arg_layout)]),
NumSin => call_intrinsic(LLVM_SIN_F64, env, &[(arg.into(), arg_layout)]),
NumCos => call_intrinsic(LLVM_COS_F64, env, &[(arg.into(), arg_layout)]),
NumToFloat => arg.into(), /* Converting from Float to Float is a no-op */
_ => {
unreachable!("Unrecognized int unary operation: {:?}", op);
}
}
}