holey-bytes/cranelift-backend/src/x86_64.rs
Jakub Doka fa2c4bcd8f
save
Signed-off-by: Jakub Doka <jakub.doka2@gmail.com>
2025-01-10 10:30:42 +01:00

318 lines
10 KiB
Rust

// The classification code for the x86_64 ABI is taken from the clay language
// https://github.com/jckarter/clay/blob/db0bd2702ab0b6e48965cd85f8859bbd5f60e48e/compiler/externals.cpp
use {crate::AbiMeta, hblang::ty};
pub fn build_systemv_signature(
sig: hblang::ty::Sig,
types: &hblang::ty::Types,
signature: &mut cranelift_codegen::ir::Signature,
arg_lens: &mut Vec<AbiMeta>,
) -> bool {
let mut alloca = Alloca::new();
alloca.next(false, sig.ret, types, &mut signature.returns);
let stack_ret = signature.returns.len() == 1
&& signature.returns[0].purpose == cranelift_codegen::ir::ArgumentPurpose::StructReturn;
if stack_ret {
signature.params.append(&mut signature.returns);
arg_lens.push(AbiMeta { arg_count: signature.params.len(), trough_mem: true });
} else {
arg_lens.push(AbiMeta { arg_count: signature.returns.len(), trough_mem: false });
}
let mut args = sig.args.args();
while let Some(arg) = args.next_value(types) {
let prev = signature.params.len();
let trough_mem = alloca.next(true, arg, types, &mut signature.params);
arg_lens.push(AbiMeta { arg_count: signature.params.len() - prev, trough_mem });
}
stack_ret
}
/// Classification of "eightbyte" components.
// N.B., the order of the variants is from general to specific,
// such that `unify(a, b)` is the "smaller" of `a` and `b`.
#[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Debug)]
enum Class {
Int,
Sse,
SseUp,
}
#[derive(Clone, Copy, Debug)]
struct Memory;
// Currently supported vector size (AVX-512).
const LARGEST_VECTOR_SIZE: usize = 512;
const MAX_EIGHTBYTES: usize = LARGEST_VECTOR_SIZE / 64;
fn classify_arg(
cx: &hblang::ty::Types,
arg: hblang::ty::Id,
) -> Result<([Option<Class>; MAX_EIGHTBYTES], [u32; MAX_EIGHTBYTES]), Memory> {
fn classify(
cx: &hblang::ty::Types,
layout: hblang::ty::Id,
cls: &mut [Option<Class>],
sizes: &mut [u32],
off: hblang::ty::Offset,
) -> Result<(), Memory> {
let size = cx.size_of(layout);
if off & (cx.align_of(layout) - 1) != 0 {
if size != 0 {
return Err(Memory);
}
return Ok(());
}
let mut c = match layout.expand() {
_ if size == 0 => return Ok(()),
_ if layout.is_integer() || layout.is_pointer() || layout == ty::Id::BOOL => Class::Int,
_ if layout.is_float() => Class::Sse,
hblang::ty::Kind::Struct(s) => {
for (f, foff) in hblang::ty::OffsetIter::new(s, cx).into_iter(cx) {
classify(cx, f.ty, cls, sizes, off + foff)?;
}
return Ok(());
}
hblang::ty::Kind::Tuple(tuple) => {
for (&ty, foff) in hblang::ty::OffsetIter::new(tuple, cx).into_iter(cx) {
classify(cx, ty, cls, sizes, off + foff)?;
}
return Ok(());
}
hblang::ty::Kind::Enum(_) => Class::Int,
hblang::ty::Kind::Union(union) => {
for f in cx.union_fields(union) {
classify(cx, f.ty, cls, sizes, off)?;
}
return Ok(());
}
hblang::ty::Kind::Slice(slice) if let Some(len) = cx.ins.slices[slice].len() => {
for i in 0..len as u32 {
classify(
cx,
cx.ins.slices[slice].elem,
cls,
sizes,
off + i * cx.size_of(cx.ins.slices[slice].elem),
)?;
}
return Ok(());
}
hblang::ty::Kind::Slice(_) => {
classify(cx, hblang::ty::Id::UINT, cls, sizes, off)?;
classify(cx, hblang::ty::Id::UINT, cls, sizes, off + 8)?;
return Ok(());
}
hblang::ty::Kind::Opt(opt) => {
let base = cx.ins.opts[opt].base;
if cx.nieche_of(base).is_some() {
classify(cx, base, cls, sizes, off)?;
} else {
classify(cx, hblang::ty::Id::BOOL, cls, sizes, off)?;
classify(cx, base, cls, sizes, off + cx.align_of(base))?;
}
return Ok(());
}
ty => unimplemented!("{ty:?}"),
};
// Fill in `cls` for scalars (Int/Sse) and vectors (Sse).
let first = (off / 8) as usize;
let last = ((off + size - 1) / 8) as usize;
for (cls, sz) in cls[first..=last].iter_mut().zip(&mut sizes[first..=last]) {
*cls = Some(cls.map_or(c, |old| old.min(c)));
*sz = size;
// Everything after the first Sse "eightbyte"
// component is the upper half of a register.
if c == Class::Sse {
c = Class::SseUp;
}
}
Ok(())
}
let size = cx.size_of(arg);
let n = ((size + 7) / 8) as usize;
if n > MAX_EIGHTBYTES {
return Err(Memory);
}
let mut cls = [None; MAX_EIGHTBYTES];
let mut sizes = [0; MAX_EIGHTBYTES];
classify(cx, arg, &mut cls, &mut sizes, 0)?;
if n > 2 {
if cls[0] != Some(Class::Sse) {
return Err(Memory);
}
if cls[1..n].iter().any(|&c| c != Some(Class::SseUp)) {
return Err(Memory);
}
} else {
let mut i = 0;
while i < n {
if cls[i] == Some(Class::SseUp) {
cls[i] = Some(Class::Sse);
} else if cls[i] == Some(Class::Sse) {
i += 1;
while i != n && cls[i] == Some(Class::SseUp) {
i += 1;
}
} else {
i += 1;
}
}
}
Ok((cls, sizes))
}
fn reg_component(
cls: &[Option<Class>],
sizes: &[u32],
i: &mut usize,
_size: hblang::ty::Size,
) -> Option<cranelift_codegen::ir::Type> {
if *i >= cls.len() {
return None;
}
let size = sizes[*i];
match cls[*i] {
None => None,
Some(Class::Int) => {
*i += 1;
Some(if size < 8 {
cranelift_codegen::ir::Type::int(size as u16 * 8).unwrap()
} else {
cranelift_codegen::ir::types::I64
})
}
Some(Class::Sse) => {
let vec_len =
1 + cls[*i + 1..].iter().take_while(|&&c| c == Some(Class::SseUp)).count();
*i += vec_len;
Some(if vec_len == 1 {
match size {
4 => cranelift_codegen::ir::types::F32,
_ => cranelift_codegen::ir::types::F64,
}
} else {
cranelift_codegen::ir::types::I64.by(vec_len as _).unwrap()
})
}
Some(c) => unreachable!("reg_component: unhandled class {:?}", c),
}
}
fn cast_target(
cls: &[Option<Class>],
sizes: &[u32],
size: hblang::ty::Size,
dest: &mut Vec<cranelift_codegen::ir::AbiParam>,
) {
let mut i = 0;
let lo = reg_component(cls, sizes, &mut i, size).unwrap();
let offset = 8 * (i as u32);
dest.push(cranelift_codegen::ir::AbiParam::new(lo));
if size > offset {
if let Some(hi) = reg_component(cls, sizes, &mut i, size - offset) {
dest.push(cranelift_codegen::ir::AbiParam::new(hi));
}
}
assert_eq!(reg_component(cls, sizes, &mut i, 0), None);
}
const MAX_INT_REGS: usize = 6; // RDI, RSI, RDX, RCX, R8, R9
const MAX_SSE_REGS: usize = 8; // XMM0-7
pub struct Alloca {
int_regs: usize,
sse_regs: usize,
}
impl Alloca {
pub fn new() -> Self {
Self { int_regs: MAX_INT_REGS, sse_regs: MAX_SSE_REGS }
}
pub fn next(
&mut self,
is_arg: bool,
arg: hblang::ty::Id,
cx: &hblang::ty::Types,
dest: &mut Vec<cranelift_codegen::ir::AbiParam>,
) -> bool {
if cx.size_of(arg) == 0 {
return false;
}
let mut cls_or_mem = classify_arg(cx, arg);
if is_arg {
if let Ok((cls, _)) = cls_or_mem {
let mut needed_int = 0;
let mut needed_sse = 0;
for c in cls {
match c {
Some(Class::Int) => needed_int += 1,
Some(Class::Sse) => needed_sse += 1,
_ => {}
}
}
match (self.int_regs.checked_sub(needed_int), self.sse_regs.checked_sub(needed_sse))
{
(Some(left_int), Some(left_sse)) => {
self.int_regs = left_int;
self.sse_regs = left_sse;
}
_ => {
// Not enough registers for this argument, so it will be
// passed on the stack, but we only mark aggregates
// explicitly as indirect `byval` arguments, as LLVM will
// automatically put immediates on the stack itself.
if arg.is_aggregate(cx) {
cls_or_mem = Err(Memory);
}
}
}
}
}
match cls_or_mem {
Err(Memory) => {
if is_arg {
dest.push(cranelift_codegen::ir::AbiParam::new(
cranelift_codegen::ir::types::I64,
));
} else {
dest.push(cranelift_codegen::ir::AbiParam::special(
cranelift_codegen::ir::types::I64,
cranelift_codegen::ir::ArgumentPurpose::StructReturn,
));
}
true
}
Ok((ref cls, ref sizes)) => {
// split into sized chunks passed individually
if arg.is_aggregate(cx) {
cast_target(cls, sizes, cx.size_of(arg), dest);
} else {
dest.push(cranelift_codegen::ir::AbiParam::new(
reg_component(cls, sizes, &mut 0, cx.size_of(arg)).unwrap(),
));
}
false
}
}
}
}