mirror of
https://github.com/rust-lang/rust.git
synced 2024-11-22 14:55:26 +00:00
Simplify FixedSizeEncoding using const generics.
This commit is contained in:
parent
b9287a83c5
commit
b4cf2cdf87
@ -7,6 +7,7 @@
|
||||
#![feature(proc_macro_internals)]
|
||||
#![feature(macro_metavar_expr)]
|
||||
#![feature(min_specialization)]
|
||||
#![feature(slice_as_chunks)]
|
||||
#![feature(try_blocks)]
|
||||
#![feature(never_type)]
|
||||
#![recursion_limit = "256"]
|
||||
|
@ -16,76 +16,34 @@ use tracing::debug;
|
||||
/// Unchecked invariant: `Self::default()` should encode as `[0; BYTE_LEN]`,
|
||||
/// but this has no impact on safety.
|
||||
pub(super) trait FixedSizeEncoding: Default {
|
||||
const BYTE_LEN: usize;
|
||||
/// This should be `[u8; BYTE_LEN]`;
|
||||
type ByteArray;
|
||||
|
||||
// FIXME(eddyb) convert to and from `[u8; Self::BYTE_LEN]` instead,
|
||||
// once that starts being allowed by the compiler (i.e. lazy normalization).
|
||||
fn from_bytes(b: &[u8]) -> Self;
|
||||
fn write_to_bytes(self, b: &mut [u8]);
|
||||
|
||||
// FIXME(eddyb) make these generic functions, or at least defaults here.
|
||||
// (same problem as above, needs `[u8; Self::BYTE_LEN]`)
|
||||
// For now, a macro (`fixed_size_encoding_byte_len_and_defaults`) is used.
|
||||
|
||||
/// Read a `Self` value (encoded as `Self::BYTE_LEN` bytes),
|
||||
/// from `&b[i * Self::BYTE_LEN..]`, returning `None` if `i`
|
||||
/// is not in bounds, or `Some(Self::from_bytes(...))` otherwise.
|
||||
fn maybe_read_from_bytes_at(b: &[u8], i: usize) -> Option<Self>;
|
||||
/// Write a `Self` value (encoded as `Self::BYTE_LEN` bytes),
|
||||
/// at `&mut b[i * Self::BYTE_LEN..]`, using `Self::write_to_bytes`.
|
||||
fn write_to_bytes_at(self, b: &mut [u8], i: usize);
|
||||
}
|
||||
|
||||
// HACK(eddyb) this shouldn't be needed (see comments on the methods above).
|
||||
macro_rules! fixed_size_encoding_byte_len_and_defaults {
|
||||
($byte_len:expr) => {
|
||||
const BYTE_LEN: usize = $byte_len;
|
||||
fn maybe_read_from_bytes_at(b: &[u8], i: usize) -> Option<Self> {
|
||||
const BYTE_LEN: usize = $byte_len;
|
||||
// HACK(eddyb) ideally this would be done with fully safe code,
|
||||
// but slicing `[u8]` with `i * N..` is optimized worse, due to the
|
||||
// possibility of `i * N` overflowing, than indexing `[[u8; N]]`.
|
||||
let b = unsafe {
|
||||
std::slice::from_raw_parts(b.as_ptr() as *const [u8; BYTE_LEN], b.len() / BYTE_LEN)
|
||||
};
|
||||
b.get(i).map(|b| FixedSizeEncoding::from_bytes(b))
|
||||
}
|
||||
fn write_to_bytes_at(self, b: &mut [u8], i: usize) {
|
||||
const BYTE_LEN: usize = $byte_len;
|
||||
// HACK(eddyb) ideally this would be done with fully safe code,
|
||||
// see similar comment in `read_from_bytes_at` for why it can't yet.
|
||||
let b = unsafe {
|
||||
std::slice::from_raw_parts_mut(
|
||||
b.as_mut_ptr() as *mut [u8; BYTE_LEN],
|
||||
b.len() / BYTE_LEN,
|
||||
)
|
||||
};
|
||||
self.write_to_bytes(&mut b[i]);
|
||||
}
|
||||
};
|
||||
fn from_bytes(b: &Self::ByteArray) -> Self;
|
||||
fn write_to_bytes(self, b: &mut Self::ByteArray);
|
||||
}
|
||||
|
||||
impl FixedSizeEncoding for u32 {
|
||||
fixed_size_encoding_byte_len_and_defaults!(4);
|
||||
type ByteArray = [u8; 4];
|
||||
|
||||
fn from_bytes(b: &[u8]) -> Self {
|
||||
let mut bytes = [0; Self::BYTE_LEN];
|
||||
bytes.copy_from_slice(&b[..Self::BYTE_LEN]);
|
||||
Self::from_le_bytes(bytes)
|
||||
#[inline]
|
||||
fn from_bytes(b: &[u8; 4]) -> Self {
|
||||
Self::from_le_bytes(*b)
|
||||
}
|
||||
|
||||
fn write_to_bytes(self, b: &mut [u8]) {
|
||||
b[..Self::BYTE_LEN].copy_from_slice(&self.to_le_bytes());
|
||||
#[inline]
|
||||
fn write_to_bytes(self, b: &mut [u8; 4]) {
|
||||
*b = self.to_le_bytes();
|
||||
}
|
||||
}
|
||||
|
||||
macro_rules! fixed_size_enum {
|
||||
($ty:ty { $(($($pat:tt)*))* }) => {
|
||||
impl FixedSizeEncoding for Option<$ty> {
|
||||
fixed_size_encoding_byte_len_and_defaults!(1);
|
||||
type ByteArray = [u8;1];
|
||||
|
||||
#[inline]
|
||||
fn from_bytes(b: &[u8]) -> Self {
|
||||
fn from_bytes(b: &[u8;1]) -> Self {
|
||||
use $ty::*;
|
||||
if b[0] == 0 {
|
||||
return None;
|
||||
@ -97,7 +55,7 @@ macro_rules! fixed_size_enum {
|
||||
}
|
||||
|
||||
#[inline]
|
||||
fn write_to_bytes(self, b: &mut [u8]) {
|
||||
fn write_to_bytes(self, b: &mut [u8;1]) {
|
||||
use $ty::*;
|
||||
b[0] = match self {
|
||||
None => 0,
|
||||
@ -184,30 +142,30 @@ fixed_size_enum! {
|
||||
|
||||
// We directly encode `DefPathHash` because a `Lazy` would encur a 25% cost.
|
||||
impl FixedSizeEncoding for Option<DefPathHash> {
|
||||
fixed_size_encoding_byte_len_and_defaults!(16);
|
||||
type ByteArray = [u8; 16];
|
||||
|
||||
#[inline]
|
||||
fn from_bytes(b: &[u8]) -> Self {
|
||||
Some(DefPathHash(Fingerprint::from_le_bytes(b.try_into().unwrap())))
|
||||
fn from_bytes(b: &[u8; 16]) -> Self {
|
||||
Some(DefPathHash(Fingerprint::from_le_bytes(*b)))
|
||||
}
|
||||
|
||||
#[inline]
|
||||
fn write_to_bytes(self, b: &mut [u8]) {
|
||||
fn write_to_bytes(self, b: &mut [u8; 16]) {
|
||||
let Some(DefPathHash(fingerprint)) = self else {
|
||||
panic!("Trying to encode absent DefPathHash.")
|
||||
};
|
||||
b[..Self::BYTE_LEN].copy_from_slice(&fingerprint.to_le_bytes());
|
||||
*b = fingerprint.to_le_bytes();
|
||||
}
|
||||
}
|
||||
|
||||
// We directly encode RawDefId because using a `Lazy` would incur a 50% overhead in the worst case.
|
||||
impl FixedSizeEncoding for Option<RawDefId> {
|
||||
fixed_size_encoding_byte_len_and_defaults!(2 * u32::BYTE_LEN);
|
||||
type ByteArray = [u8; 8];
|
||||
|
||||
#[inline]
|
||||
fn from_bytes(b: &[u8]) -> Self {
|
||||
let krate = u32::from_bytes(&b[0..4]);
|
||||
let index = u32::from_bytes(&b[4..8]);
|
||||
fn from_bytes(b: &[u8; 8]) -> Self {
|
||||
let krate = u32::from_le_bytes(b[0..4].try_into().unwrap());
|
||||
let index = u32::from_le_bytes(b[4..8].try_into().unwrap());
|
||||
if krate == 0 {
|
||||
return None;
|
||||
}
|
||||
@ -215,14 +173,14 @@ impl FixedSizeEncoding for Option<RawDefId> {
|
||||
}
|
||||
|
||||
#[inline]
|
||||
fn write_to_bytes(self, b: &mut [u8]) {
|
||||
fn write_to_bytes(self, b: &mut [u8; 8]) {
|
||||
match self {
|
||||
None => 0u32.write_to_bytes(b),
|
||||
None => *b = [0; 8],
|
||||
Some(RawDefId { krate, index }) => {
|
||||
// CrateNum is less than `CrateNum::MAX_AS_U32`.
|
||||
debug_assert!(krate < u32::MAX);
|
||||
(1 + krate).write_to_bytes(&mut b[0..4]);
|
||||
index.write_to_bytes(&mut b[4..8]);
|
||||
b[0..4].copy_from_slice(&(1 + krate).to_le_bytes());
|
||||
b[4..8].copy_from_slice(&index.to_le_bytes());
|
||||
}
|
||||
}
|
||||
}
|
||||
@ -232,44 +190,51 @@ impl FixedSizeEncoding for Option<RawDefId> {
|
||||
// generic `Lazy<T>` impl, but in the general case we might not need / want to
|
||||
// fit every `usize` in `u32`.
|
||||
impl<T> FixedSizeEncoding for Option<Lazy<T>> {
|
||||
fixed_size_encoding_byte_len_and_defaults!(u32::BYTE_LEN);
|
||||
type ByteArray = [u8; 4];
|
||||
|
||||
fn from_bytes(b: &[u8]) -> Self {
|
||||
Some(Lazy::from_position(NonZeroUsize::new(u32::from_bytes(b) as usize)?))
|
||||
#[inline]
|
||||
fn from_bytes(b: &[u8; 4]) -> Self {
|
||||
let position = NonZeroUsize::new(u32::from_bytes(b) as usize)?;
|
||||
Some(Lazy::from_position(position))
|
||||
}
|
||||
|
||||
fn write_to_bytes(self, b: &mut [u8]) {
|
||||
#[inline]
|
||||
fn write_to_bytes(self, b: &mut [u8; 4]) {
|
||||
let position = self.map_or(0, |lazy| lazy.position.get());
|
||||
let position: u32 = position.try_into().unwrap();
|
||||
|
||||
position.write_to_bytes(b)
|
||||
}
|
||||
}
|
||||
|
||||
impl<T> FixedSizeEncoding for Option<Lazy<[T]>> {
|
||||
fixed_size_encoding_byte_len_and_defaults!(u32::BYTE_LEN * 2);
|
||||
type ByteArray = [u8; 8];
|
||||
|
||||
fn from_bytes(b: &[u8]) -> Self {
|
||||
Some(Lazy::from_position_and_meta(
|
||||
<Option<Lazy<T>>>::from_bytes(b)?.position,
|
||||
u32::from_bytes(&b[u32::BYTE_LEN..]) as usize,
|
||||
))
|
||||
#[inline]
|
||||
fn from_bytes(b: &[u8; 8]) -> Self {
|
||||
let ([ref position_bytes, ref meta_bytes],[])= b.as_chunks::<4>() else { panic!() };
|
||||
let position = NonZeroUsize::new(u32::from_bytes(position_bytes) as usize)?;
|
||||
let len = u32::from_bytes(meta_bytes) as usize;
|
||||
Some(Lazy::from_position_and_meta(position, len))
|
||||
}
|
||||
|
||||
fn write_to_bytes(self, b: &mut [u8]) {
|
||||
self.map(|lazy| Lazy::<T>::from_position(lazy.position)).write_to_bytes(b);
|
||||
#[inline]
|
||||
fn write_to_bytes(self, b: &mut [u8; 8]) {
|
||||
let ([ref mut position_bytes, ref mut meta_bytes],[])= b.as_chunks_mut::<4>() else { panic!() };
|
||||
|
||||
let position = self.map_or(0, |lazy| lazy.position.get());
|
||||
let position: u32 = position.try_into().unwrap();
|
||||
position.write_to_bytes(position_bytes);
|
||||
|
||||
let len = self.map_or(0, |lazy| lazy.meta);
|
||||
let len: u32 = len.try_into().unwrap();
|
||||
|
||||
len.write_to_bytes(&mut b[u32::BYTE_LEN..]);
|
||||
len.write_to_bytes(meta_bytes);
|
||||
}
|
||||
}
|
||||
|
||||
/// Random-access table (i.e. offering constant-time `get`/`set`), similar to
|
||||
/// `Vec<Option<T>>`, but without requiring encoding or decoding all the values
|
||||
/// eagerly and in-order.
|
||||
/// A total of `(max_idx + 1) * <Option<T> as FixedSizeEncoding>::BYTE_LEN` bytes
|
||||
/// A total of `(max_idx + 1)` times `Option<T> as FixedSizeEncoding>::ByteArray`
|
||||
/// are used for a table, where `max_idx` is the largest index passed to
|
||||
/// `TableBuilder::set`.
|
||||
pub(super) struct Table<I: Idx, T>
|
||||
@ -287,12 +252,8 @@ pub(super) struct TableBuilder<I: Idx, T>
|
||||
where
|
||||
Option<T>: FixedSizeEncoding,
|
||||
{
|
||||
// FIXME(eddyb) use `IndexVec<I, [u8; <Option<T>>::BYTE_LEN]>` instead of
|
||||
// `Vec<u8>`, once that starts working (i.e. lazy normalization).
|
||||
// Then again, that has the downside of not allowing `TableBuilder::encode` to
|
||||
// obtain a `&[u8]` entirely in safe code, for writing the bytes out.
|
||||
bytes: Vec<u8>,
|
||||
_marker: PhantomData<(fn(&I), T)>,
|
||||
blocks: IndexVec<I, <Option<T> as FixedSizeEncoding>::ByteArray>,
|
||||
_marker: PhantomData<T>,
|
||||
}
|
||||
|
||||
impl<I: Idx, T> Default for TableBuilder<I, T>
|
||||
@ -300,7 +261,7 @@ where
|
||||
Option<T>: FixedSizeEncoding,
|
||||
{
|
||||
fn default() -> Self {
|
||||
TableBuilder { bytes: vec![], _marker: PhantomData }
|
||||
TableBuilder { blocks: Default::default(), _marker: PhantomData }
|
||||
}
|
||||
}
|
||||
|
||||
@ -308,25 +269,29 @@ impl<I: Idx, T> TableBuilder<I, T>
|
||||
where
|
||||
Option<T>: FixedSizeEncoding,
|
||||
{
|
||||
pub(crate) fn set(&mut self, i: I, value: T) {
|
||||
pub(crate) fn set<const N: usize>(&mut self, i: I, value: T)
|
||||
where
|
||||
Option<T>: FixedSizeEncoding<ByteArray = [u8; N]>,
|
||||
{
|
||||
// FIXME(eddyb) investigate more compact encodings for sparse tables.
|
||||
// On the PR @michaelwoerister mentioned:
|
||||
// > Space requirements could perhaps be optimized by using the HAMT `popcnt`
|
||||
// > trick (i.e. divide things into buckets of 32 or 64 items and then
|
||||
// > store bit-masks of which item in each bucket is actually serialized).
|
||||
let i = i.index();
|
||||
let needed = (i + 1) * <Option<T>>::BYTE_LEN;
|
||||
if self.bytes.len() < needed {
|
||||
self.bytes.resize(needed, 0);
|
||||
}
|
||||
|
||||
Some(value).write_to_bytes_at(&mut self.bytes, i);
|
||||
self.blocks.ensure_contains_elem(i, || [0; N]);
|
||||
Some(value).write_to_bytes(&mut self.blocks[i]);
|
||||
}
|
||||
|
||||
pub(crate) fn encode(&self, buf: &mut Encoder) -> Lazy<Table<I, T>> {
|
||||
pub(crate) fn encode<const N: usize>(&self, buf: &mut Encoder) -> Lazy<Table<I, T>>
|
||||
where
|
||||
Option<T>: FixedSizeEncoding<ByteArray = [u8; N]>,
|
||||
{
|
||||
let pos = buf.position();
|
||||
buf.emit_raw_bytes(&self.bytes).unwrap();
|
||||
Lazy::from_position_and_meta(NonZeroUsize::new(pos as usize).unwrap(), self.bytes.len())
|
||||
for block in &self.blocks {
|
||||
buf.emit_raw_bytes(block).unwrap();
|
||||
}
|
||||
let num_bytes = self.blocks.len() * N;
|
||||
Lazy::from_position_and_meta(NonZeroUsize::new(pos as usize).unwrap(), num_bytes)
|
||||
}
|
||||
}
|
||||
|
||||
@ -334,6 +299,7 @@ impl<I: Idx, T> LazyMeta for Table<I, T>
|
||||
where
|
||||
Option<T>: FixedSizeEncoding,
|
||||
{
|
||||
/// Number of bytes in the data stream.
|
||||
type Meta = usize;
|
||||
}
|
||||
|
||||
@ -343,16 +309,28 @@ where
|
||||
{
|
||||
/// Given the metadata, extract out the value at a particular index (if any).
|
||||
#[inline(never)]
|
||||
pub(super) fn get<'a, 'tcx, M: Metadata<'a, 'tcx>>(&self, metadata: M, i: I) -> Option<T> {
|
||||
pub(super) fn get<'a, 'tcx, M: Metadata<'a, 'tcx>, const N: usize>(
|
||||
&self,
|
||||
metadata: M,
|
||||
i: I,
|
||||
) -> Option<T>
|
||||
where
|
||||
Option<T>: FixedSizeEncoding<ByteArray = [u8; N]>,
|
||||
{
|
||||
debug!("Table::lookup: index={:?} len={:?}", i, self.meta);
|
||||
|
||||
let start = self.position.get();
|
||||
let bytes = &metadata.blob()[start..start + self.meta];
|
||||
<Option<T>>::maybe_read_from_bytes_at(bytes, i.index())?
|
||||
let (bytes, []) = bytes.as_chunks::<N>() else { panic!() };
|
||||
let bytes = bytes.get(i.index())?;
|
||||
FixedSizeEncoding::from_bytes(bytes)
|
||||
}
|
||||
|
||||
/// Size of the table in entries, including possible gaps.
|
||||
pub(super) fn size(&self) -> usize {
|
||||
self.meta / <Option<T>>::BYTE_LEN
|
||||
pub(super) fn size<const N: usize>(&self) -> usize
|
||||
where
|
||||
Option<T>: FixedSizeEncoding<ByteArray = [u8; N]>,
|
||||
{
|
||||
self.meta / N
|
||||
}
|
||||
}
|
||||
|
Loading…
Reference in New Issue
Block a user