Auto merge of #93984 - nnethercote:ChunkedBitSet, r=Mark-Simulacrum

Introduce `ChunkedBitSet` and use it for some dataflow analyses.

This reduces peak memory usage significantly for some programs with very
large functions.

r? `@ghost`
This commit is contained in:
bors 2022-02-23 01:26:07 +00:00
commit bafe8d06e0
14 changed files with 806 additions and 75 deletions

View File

@ -22,7 +22,7 @@ use rustc_errors::{Applicability, Diagnostic, DiagnosticBuilder, ErrorReported};
use rustc_hir as hir;
use rustc_hir::def_id::LocalDefId;
use rustc_hir::Node;
use rustc_index::bit_set::BitSet;
use rustc_index::bit_set::ChunkedBitSet;
use rustc_index::vec::IndexVec;
use rustc_infer::infer::{InferCtxt, TyCtxtInferExt};
use rustc_middle::mir::{
@ -1667,7 +1667,7 @@ impl<'cx, 'tcx> MirBorrowckCtxt<'cx, 'tcx> {
location: Location,
desired_action: InitializationRequiringAction,
place_span: (PlaceRef<'tcx>, Span),
maybe_uninits: &BitSet<MovePathIndex>,
maybe_uninits: &ChunkedBitSet<MovePathIndex>,
from: u64,
to: u64,
) {

View File

@ -37,6 +37,8 @@ extern crate cfg_if;
#[macro_use]
extern crate rustc_macros;
pub use rustc_index::static_assert_size;
#[inline(never)]
#[cold]
pub fn cold_path<F: FnOnce() -> R, R>(f: F) -> R {

View File

@ -1,11 +1,3 @@
/// Type size assertion. The first argument is a type and the second argument is its expected size.
#[macro_export]
macro_rules! static_assert_size {
($ty:ty, $size:expr) => {
const _: [(); $size] = [(); ::std::mem::size_of::<$ty>()];
};
}
#[macro_export]
macro_rules! enum_from_u32 {
($(#[$attr:meta])* pub enum $name:ident {

View File

@ -5,16 +5,37 @@ use std::iter;
use std::marker::PhantomData;
use std::mem;
use std::ops::{BitAnd, BitAndAssign, BitOrAssign, Bound, Not, Range, RangeBounds, Shl};
use std::rc::Rc;
use std::slice;
use rustc_macros::{Decodable, Encodable};
use Chunk::*;
#[cfg(test)]
mod tests;
pub type Word = u64;
pub const WORD_BYTES: usize = mem::size_of::<Word>();
pub const WORD_BITS: usize = WORD_BYTES * 8;
type Word = u64;
const WORD_BYTES: usize = mem::size_of::<Word>();
const WORD_BITS: usize = WORD_BYTES * 8;
// The choice of chunk size has some trade-offs.
//
// A big chunk size tends to favour cases where many large `ChunkedBitSet`s are
// present, because they require fewer `Chunk`s, reducing the number of
// allocations and reducing peak memory usage. Also, fewer chunk operations are
// required, though more of them might be `Mixed`.
//
// A small chunk size tends to favour cases where many small `ChunkedBitSet`s
// are present, because less space is wasted at the end of the final chunk (if
// it's not full).
const CHUNK_WORDS: usize = 32;
const CHUNK_BITS: usize = CHUNK_WORDS * WORD_BITS; // 2048 bits
/// ChunkSize is small to keep `Chunk` small. The static assertion ensures it's
/// not too small.
type ChunkSize = u16;
const _: () = assert!(CHUNK_BITS <= ChunkSize::MAX as usize);
pub trait BitRelations<Rhs> {
fn union(&mut self, other: &Rhs) -> bool;
@ -121,19 +142,12 @@ impl<T: Idx> BitSet<T> {
/// Clear all elements.
#[inline]
pub fn clear(&mut self) {
for word in &mut self.words {
*word = 0;
}
self.words.fill(0);
}
/// Clear excess bits in the final word.
fn clear_excess_bits(&mut self) {
let num_bits_in_final_word = self.domain_size % WORD_BITS;
if num_bits_in_final_word > 0 {
let mask = (1 << num_bits_in_final_word) - 1;
let final_word_idx = self.words.len() - 1;
self.words[final_word_idx] &= mask;
}
clear_excess_bits_in_final_word(self.domain_size, &mut self.words);
}
/// Count the number of set bits in the set.
@ -203,9 +217,7 @@ impl<T: Idx> BitSet<T> {
/// Sets all bits to true.
pub fn insert_all(&mut self) {
for word in &mut self.words {
*word = !0;
}
self.words.fill(!0);
self.clear_excess_bits();
}
@ -328,6 +340,407 @@ impl<T: Idx> BitRelations<BitSet<T>> for BitSet<T> {
}
}
/// A fixed-size bitset type with a partially dense, partially sparse
/// representation. The bitset is broken into chunks, and chunks that are all
/// zeros or all ones are represented and handled very efficiently.
///
/// This type is especially efficient for sets that typically have a large
/// `domain_size` with significant stretches of all zeros or all ones, and also
/// some stretches with lots of 0s and 1s mixed in a way that causes trouble
/// for `IntervalSet`.
///
/// `T` is an index type, typically a newtyped `usize` wrapper, but it can also
/// just be `usize`.
///
/// All operations that involve an element will panic if the element is equal
/// to or greater than the domain size. All operations that involve two bitsets
/// will panic if the bitsets have differing domain sizes.
#[derive(Debug, PartialEq, Eq)]
pub struct ChunkedBitSet<T> {
domain_size: usize,
/// The chunks. Each one contains exactly CHUNK_BITS values, except the
/// last one which contains 1..=CHUNK_BITS values.
chunks: Box<[Chunk]>,
marker: PhantomData<T>,
}
// Note: the chunk domain size is duplicated in each variant. This is a bit
// inconvenient, but it allows the type size to be smaller than if we had an
// outer struct containing a chunk domain size plus the `Chunk`, because the
// compiler can place the chunk domain size after the tag.
#[derive(Clone, Debug, PartialEq, Eq)]
enum Chunk {
/// A chunk that is all zeros; we don't represent the zeros explicitly.
Zeros(ChunkSize),
/// A chunk that is all ones; we don't represent the ones explicitly.
Ones(ChunkSize),
/// A chunk that has a mix of zeros and ones, which are represented
/// explicitly and densely. It never has all zeros or all ones.
///
/// If this is the final chunk there may be excess, unused words. This
/// turns out to be both simpler and have better performance than
/// allocating the minimum number of words, largely because we avoid having
/// to store the length, which would make this type larger. These excess
/// words are always be zero, as are any excess bits in the final in-use
/// word.
///
/// The second field is the count of 1s set in the chunk, and must satisfy
/// `0 < count < chunk_domain_size`.
///
/// The words are within an `Rc` because it's surprisingly common to
/// duplicate an entire chunk, e.g. in `ChunkedBitSet::clone_from()`, or
/// when a `Mixed` chunk is union'd into a `Zeros` chunk. When we do need
/// to modify a chunk we use `Rc::make_mut`.
Mixed(ChunkSize, ChunkSize, Rc<[Word; CHUNK_WORDS]>),
}
// This type is used a lot. Make sure it doesn't unintentionally get bigger.
#[cfg(all(target_arch = "x86_64", target_pointer_width = "64"))]
crate::static_assert_size!(Chunk, 16);
impl<T> ChunkedBitSet<T> {
pub fn domain_size(&self) -> usize {
self.domain_size
}
#[cfg(test)]
fn assert_valid(&self) {
if self.domain_size == 0 {
assert!(self.chunks.is_empty());
return;
}
assert!((self.chunks.len() - 1) * CHUNK_BITS <= self.domain_size);
assert!(self.chunks.len() * CHUNK_BITS >= self.domain_size);
for chunk in self.chunks.iter() {
chunk.assert_valid();
}
}
}
impl<T: Idx> ChunkedBitSet<T> {
/// Creates a new bitset with a given `domain_size` and chunk kind.
fn new(domain_size: usize, is_empty: bool) -> Self {
let chunks = if domain_size == 0 {
Box::new([])
} else {
// All the chunks have a chunk_domain_size of `CHUNK_BITS` except
// the final one.
let final_chunk_domain_size = {
let n = domain_size % CHUNK_BITS;
if n == 0 { CHUNK_BITS } else { n }
};
let mut chunks =
vec![Chunk::new(CHUNK_BITS, is_empty); num_chunks(domain_size)].into_boxed_slice();
*chunks.last_mut().unwrap() = Chunk::new(final_chunk_domain_size, is_empty);
chunks
};
ChunkedBitSet { domain_size, chunks, marker: PhantomData }
}
/// Creates a new, empty bitset with a given `domain_size`.
#[inline]
pub fn new_empty(domain_size: usize) -> Self {
ChunkedBitSet::new(domain_size, /* is_empty */ true)
}
/// Creates a new, filled bitset with a given `domain_size`.
#[inline]
pub fn new_filled(domain_size: usize) -> Self {
ChunkedBitSet::new(domain_size, /* is_empty */ false)
}
#[cfg(test)]
fn chunks(&self) -> &[Chunk] {
&self.chunks
}
/// Count the number of bits in the set.
pub fn count(&self) -> usize {
self.chunks.iter().map(|chunk| chunk.count()).sum()
}
/// Returns `true` if `self` contains `elem`.
#[inline]
pub fn contains(&self, elem: T) -> bool {
assert!(elem.index() < self.domain_size);
let chunk = &self.chunks[chunk_index(elem)];
match &chunk {
Zeros(_) => false,
Ones(_) => true,
Mixed(_, _, words) => {
let (word_index, mask) = chunk_word_index_and_mask(elem);
(words[word_index] & mask) != 0
}
}
}
/// Insert `elem`. Returns whether the set has changed.
pub fn insert(&mut self, elem: T) -> bool {
assert!(elem.index() < self.domain_size);
let chunk_index = chunk_index(elem);
let chunk = &mut self.chunks[chunk_index];
match *chunk {
Zeros(chunk_domain_size) => {
if chunk_domain_size > 1 {
// We take some effort to avoid copying the words.
let words = Rc::<[Word; CHUNK_WORDS]>::new_zeroed();
// SAFETY: `words` can safely be all zeroes.
let mut words = unsafe { words.assume_init() };
let words_ref = Rc::get_mut(&mut words).unwrap();
let (word_index, mask) = chunk_word_index_and_mask(elem);
words_ref[word_index] |= mask;
*chunk = Mixed(chunk_domain_size, 1, words);
} else {
*chunk = Ones(chunk_domain_size);
}
true
}
Ones(_) => false,
Mixed(chunk_domain_size, ref mut count, ref mut words) => {
// We skip all the work if the bit is already set.
let (word_index, mask) = chunk_word_index_and_mask(elem);
if (words[word_index] & mask) == 0 {
*count += 1;
if *count < chunk_domain_size {
let words = Rc::make_mut(words);
words[word_index] |= mask;
} else {
*chunk = Ones(chunk_domain_size);
}
true
} else {
false
}
}
}
}
/// Sets all bits to true.
pub fn insert_all(&mut self) {
for chunk in self.chunks.iter_mut() {
*chunk = match *chunk {
Zeros(chunk_domain_size)
| Ones(chunk_domain_size)
| Mixed(chunk_domain_size, ..) => Ones(chunk_domain_size),
}
}
}
/// Returns `true` if the set has changed.
pub fn remove(&mut self, elem: T) -> bool {
assert!(elem.index() < self.domain_size);
let chunk_index = chunk_index(elem);
let chunk = &mut self.chunks[chunk_index];
match *chunk {
Zeros(_) => false,
Ones(chunk_domain_size) => {
if chunk_domain_size > 1 {
// We take some effort to avoid copying the words.
let words = Rc::<[Word; CHUNK_WORDS]>::new_zeroed();
// SAFETY: `words` can safely be all zeroes.
let mut words = unsafe { words.assume_init() };
let words_ref = Rc::get_mut(&mut words).unwrap();
// Set only the bits in use.
let num_words = num_words(chunk_domain_size as usize);
words_ref[..num_words].fill(!0);
clear_excess_bits_in_final_word(
chunk_domain_size as usize,
&mut words_ref[..num_words],
);
let (word_index, mask) = chunk_word_index_and_mask(elem);
words_ref[word_index] &= !mask;
*chunk = Mixed(chunk_domain_size, chunk_domain_size - 1, words);
} else {
*chunk = Zeros(chunk_domain_size);
}
true
}
Mixed(chunk_domain_size, ref mut count, ref mut words) => {
// We skip all the work if the bit is already clear.
let (word_index, mask) = chunk_word_index_and_mask(elem);
if (words[word_index] & mask) != 0 {
*count -= 1;
if *count > 0 {
let words = Rc::make_mut(words);
words[word_index] &= !mask;
} else {
*chunk = Zeros(chunk_domain_size);
}
true
} else {
false
}
}
}
}
bit_relations_inherent_impls! {}
}
impl<T: Idx> BitRelations<ChunkedBitSet<T>> for ChunkedBitSet<T> {
fn union(&mut self, other: &ChunkedBitSet<T>) -> bool {
assert_eq!(self.domain_size, other.domain_size);
debug_assert_eq!(self.chunks.len(), other.chunks.len());
let mut changed = false;
for (mut self_chunk, other_chunk) in self.chunks.iter_mut().zip(other.chunks.iter()) {
match (&mut self_chunk, &other_chunk) {
(_, Zeros(_)) | (Ones(_), _) => {}
(Zeros(self_chunk_domain_size), Ones(other_chunk_domain_size))
| (Mixed(self_chunk_domain_size, ..), Ones(other_chunk_domain_size))
| (Zeros(self_chunk_domain_size), Mixed(other_chunk_domain_size, ..)) => {
// `other_chunk` fully overwrites `self_chunk`
debug_assert_eq!(self_chunk_domain_size, other_chunk_domain_size);
*self_chunk = other_chunk.clone();
changed = true;
}
(
Mixed(
self_chunk_domain_size,
ref mut self_chunk_count,
ref mut self_chunk_words,
),
Mixed(_other_chunk_domain_size, _other_chunk_count, other_chunk_words),
) => {
// First check if the operation would change
// `self_chunk.words`. If not, we can avoid allocating some
// words, and this happens often enough that it's a
// performance win. Also, we only need to operate on the
// in-use words, hence the slicing.
let op = |a, b| a | b;
let num_words = num_words(*self_chunk_domain_size as usize);
if bitwise_changes(
&self_chunk_words[0..num_words],
&other_chunk_words[0..num_words],
op,
) {
let self_chunk_words = Rc::make_mut(self_chunk_words);
let has_changed = bitwise(
&mut self_chunk_words[0..num_words],
&other_chunk_words[0..num_words],
op,
);
debug_assert!(has_changed);
*self_chunk_count = self_chunk_words[0..num_words]
.iter()
.map(|w| w.count_ones() as ChunkSize)
.sum();
if *self_chunk_count == *self_chunk_domain_size {
*self_chunk = Ones(*self_chunk_domain_size);
}
changed = true;
}
}
}
}
changed
}
fn subtract(&mut self, _other: &ChunkedBitSet<T>) -> bool {
unimplemented!("implement if/when necessary");
}
fn intersect(&mut self, _other: &ChunkedBitSet<T>) -> bool {
unimplemented!("implement if/when necessary");
}
}
impl<T: Idx> BitRelations<HybridBitSet<T>> for ChunkedBitSet<T> {
fn union(&mut self, other: &HybridBitSet<T>) -> bool {
// FIXME: this is slow if `other` is dense, and could easily be
// improved, but it hasn't been a problem in practice so far.
assert_eq!(self.domain_size, other.domain_size());
sequential_update(|elem| self.insert(elem), other.iter())
}
fn subtract(&mut self, other: &HybridBitSet<T>) -> bool {
// FIXME: this is slow if `other` is dense, and could easily be
// improved, but it hasn't been a problem in practice so far.
assert_eq!(self.domain_size, other.domain_size());
sequential_update(|elem| self.remove(elem), other.iter())
}
fn intersect(&mut self, _other: &HybridBitSet<T>) -> bool {
unimplemented!("implement if/when necessary");
}
}
impl<T> Clone for ChunkedBitSet<T> {
fn clone(&self) -> Self {
ChunkedBitSet {
domain_size: self.domain_size,
chunks: self.chunks.clone(),
marker: PhantomData,
}
}
/// WARNING: this implementation of clone_from will panic if the two
/// bitsets have different domain sizes. This constraint is not inherent to
/// `clone_from`, but it works with the existing call sites and allows a
/// faster implementation, which is important because this function is hot.
fn clone_from(&mut self, from: &Self) {
assert_eq!(self.domain_size, from.domain_size);
debug_assert_eq!(self.chunks.len(), from.chunks.len());
self.chunks.clone_from(&from.chunks)
}
}
impl Chunk {
#[cfg(test)]
fn assert_valid(&self) {
match *self {
Zeros(chunk_domain_size) | Ones(chunk_domain_size) => {
assert!(chunk_domain_size as usize <= CHUNK_BITS);
}
Mixed(chunk_domain_size, count, ref words) => {
assert!(chunk_domain_size as usize <= CHUNK_BITS);
assert!(0 < count && count < chunk_domain_size);
// Check the number of set bits matches `count`.
assert_eq!(
words.iter().map(|w| w.count_ones() as ChunkSize).sum::<ChunkSize>(),
count
);
// Check the not-in-use words are all zeroed.
let num_words = num_words(chunk_domain_size as usize);
if num_words < CHUNK_WORDS {
assert_eq!(
words[num_words..]
.iter()
.map(|w| w.count_ones() as ChunkSize)
.sum::<ChunkSize>(),
0
);
}
}
}
}
fn new(chunk_domain_size: usize, is_empty: bool) -> Self {
debug_assert!(chunk_domain_size <= CHUNK_BITS);
let chunk_domain_size = chunk_domain_size as ChunkSize;
if is_empty { Zeros(chunk_domain_size) } else { Ones(chunk_domain_size) }
}
/// Count the number of 1s in the chunk.
fn count(&self) -> usize {
match *self {
Zeros(_) => 0,
Ones(chunk_domain_size) => chunk_domain_size as usize,
Mixed(_, count, _) => count as usize,
}
}
}
// Applies a function to mutate a bitset, and returns true if any
// of the applications return true
fn sequential_update<T: Idx>(
@ -642,6 +1055,23 @@ where
changed != 0
}
/// Does this bitwise operation change `out_vec`?
#[inline]
fn bitwise_changes<Op>(out_vec: &[Word], in_vec: &[Word], op: Op) -> bool
where
Op: Fn(Word, Word) -> Word,
{
assert_eq!(out_vec.len(), in_vec.len());
for (out_elem, in_elem) in iter::zip(out_vec, in_vec) {
let old_val = *out_elem;
let new_val = op(old_val, *in_elem);
if old_val != new_val {
return true;
}
}
false
}
const SPARSE_MAX: usize = 8;
/// A fixed-size bitset type with a sparse representation and a maximum of
@ -1136,18 +1566,7 @@ impl<R: Idx, C: Idx> BitMatrix<R, C> {
for index in start..end {
words[index] = !0;
}
self.clear_excess_bits(row);
}
/// Clear excess bits in the final word of the row.
fn clear_excess_bits(&mut self, row: R) {
let num_bits_in_final_word = self.num_columns % WORD_BITS;
if num_bits_in_final_word > 0 {
let mask = (1 << num_bits_in_final_word) - 1;
let (_, end) = self.range(row);
let final_word_idx = end - 1;
self.words[final_word_idx] &= mask;
}
clear_excess_bits_in_final_word(self.num_columns, &mut self.words[..end]);
}
/// Gets a slice of the underlying words.
@ -1339,6 +1758,12 @@ fn num_words<T: Idx>(domain_size: T) -> usize {
(domain_size.index() + WORD_BITS - 1) / WORD_BITS
}
#[inline]
fn num_chunks<T: Idx>(domain_size: T) -> usize {
assert!(domain_size.index() > 0);
(domain_size.index() + CHUNK_BITS - 1) / CHUNK_BITS
}
#[inline]
fn word_index_and_mask<T: Idx>(elem: T) -> (usize, Word) {
let elem = elem.index();
@ -1347,6 +1772,25 @@ fn word_index_and_mask<T: Idx>(elem: T) -> (usize, Word) {
(word_index, mask)
}
#[inline]
fn chunk_index<T: Idx>(elem: T) -> usize {
elem.index() / CHUNK_BITS
}
#[inline]
fn chunk_word_index_and_mask<T: Idx>(elem: T) -> (usize, Word) {
let chunk_elem = elem.index() % CHUNK_BITS;
word_index_and_mask(chunk_elem)
}
fn clear_excess_bits_in_final_word(domain_size: usize, words: &mut [Word]) {
let num_bits_in_final_word = domain_size % WORD_BITS;
if num_bits_in_final_word > 0 {
let mask = (1 << num_bits_in_final_word) - 1;
words[words.len() - 1] &= mask;
}
}
#[inline]
fn max_bit(word: Word) -> usize {
WORD_BITS - 1 - word.leading_zeros() as usize

View File

@ -147,6 +147,201 @@ fn hybrid_bitset() {
assert!(dense0.is_empty());
}
#[test]
fn chunked_bitset() {
let mut b0 = ChunkedBitSet::<usize>::new_empty(0);
let b0b = b0.clone();
assert_eq!(b0, ChunkedBitSet { domain_size: 0, chunks: Box::new([]), marker: PhantomData });
// There are no valid insert/remove/contains operations on a 0-domain
// bitset, but we can test `union`.
b0.assert_valid();
assert!(!b0.union(&b0b));
assert_eq!(b0.chunks(), vec![]);
assert_eq!(b0.count(), 0);
b0.assert_valid();
//-----------------------------------------------------------------------
let mut b1 = ChunkedBitSet::<usize>::new_empty(1);
assert_eq!(
b1,
ChunkedBitSet { domain_size: 1, chunks: Box::new([Zeros(1)]), marker: PhantomData }
);
b1.assert_valid();
assert!(!b1.contains(0));
assert_eq!(b1.count(), 0);
assert!(b1.insert(0));
assert!(b1.contains(0));
assert_eq!(b1.count(), 1);
assert_eq!(b1.chunks(), [Ones(1)]);
assert!(!b1.insert(0));
assert!(b1.remove(0));
assert!(!b1.contains(0));
assert_eq!(b1.count(), 0);
assert_eq!(b1.chunks(), [Zeros(1)]);
b1.assert_valid();
//-----------------------------------------------------------------------
let mut b100 = ChunkedBitSet::<usize>::new_filled(100);
assert_eq!(
b100,
ChunkedBitSet { domain_size: 100, chunks: Box::new([Ones(100)]), marker: PhantomData }
);
b100.assert_valid();
for i in 0..100 {
assert!(b100.contains(i));
}
assert_eq!(b100.count(), 100);
assert!(b100.remove(3));
assert!(b100.insert(3));
assert_eq!(b100.chunks(), vec![Ones(100)]);
assert!(
b100.remove(20) && b100.remove(30) && b100.remove(40) && b100.remove(99) && b100.insert(30)
);
assert_eq!(b100.count(), 97);
assert!(!b100.contains(20) && b100.contains(30) && !b100.contains(99) && b100.contains(50));
assert_eq!(
b100.chunks(),
vec![Mixed(
100,
97,
#[rustfmt::skip]
Rc::new([
0b11111111_11111111_11111110_11111111_11111111_11101111_11111111_11111111,
0b00000000_00000000_00000000_00000111_11111111_11111111_11111111_11111111,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0,
])
)],
);
b100.assert_valid();
let mut num_removed = 0;
for i in 0..100 {
if b100.remove(i) {
num_removed += 1;
}
}
assert_eq!(num_removed, 97);
assert_eq!(b100.chunks(), vec![Zeros(100)]);
b100.assert_valid();
//-----------------------------------------------------------------------
let mut b2548 = ChunkedBitSet::<usize>::new_empty(2548);
assert_eq!(
b2548,
ChunkedBitSet {
domain_size: 2548,
chunks: Box::new([Zeros(2048), Zeros(500)]),
marker: PhantomData,
}
);
b2548.assert_valid();
b2548.insert(14);
b2548.remove(14);
assert_eq!(b2548.chunks(), vec![Zeros(2048), Zeros(500)]);
b2548.insert_all();
for i in 0..2548 {
assert!(b2548.contains(i));
}
assert_eq!(b2548.count(), 2548);
assert_eq!(b2548.chunks(), vec![Ones(2048), Ones(500)]);
b2548.assert_valid();
//-----------------------------------------------------------------------
let mut b4096 = ChunkedBitSet::<usize>::new_empty(4096);
assert_eq!(
b4096,
ChunkedBitSet {
domain_size: 4096,
chunks: Box::new([Zeros(2048), Zeros(2048)]),
marker: PhantomData,
}
);
b4096.assert_valid();
for i in 0..4096 {
assert!(!b4096.contains(i));
}
assert!(b4096.insert(0) && b4096.insert(4095) && !b4096.insert(4095));
assert!(
b4096.contains(0) && !b4096.contains(2047) && !b4096.contains(2048) && b4096.contains(4095)
);
assert_eq!(
b4096.chunks(),
#[rustfmt::skip]
vec![
Mixed(2048, 1, Rc::new([
1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
])),
Mixed(2048, 1, Rc::new([
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0x8000_0000_0000_0000
])),
],
);
assert_eq!(b4096.count(), 2);
b4096.assert_valid();
//-----------------------------------------------------------------------
let mut b10000 = ChunkedBitSet::<usize>::new_empty(10000);
assert_eq!(
b10000,
ChunkedBitSet {
domain_size: 10000,
chunks: Box::new([Zeros(2048), Zeros(2048), Zeros(2048), Zeros(2048), Zeros(1808),]),
marker: PhantomData,
}
);
b10000.assert_valid();
assert!(b10000.insert(3000) && b10000.insert(5000));
assert_eq!(
b10000.chunks(),
#[rustfmt::skip]
vec![
Zeros(2048),
Mixed(2048, 1, Rc::new([
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0x0100_0000_0000_0000, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
])),
Mixed(2048, 1, Rc::new([
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0x0100, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
])),
Zeros(2048),
Zeros(1808),
],
);
let mut b10000b = ChunkedBitSet::<usize>::new_empty(10000);
b10000b.clone_from(&b10000);
assert_eq!(b10000, b10000b);
for i in 6000..7000 {
b10000b.insert(i);
}
assert_eq!(b10000b.count(), 1002);
b10000b.assert_valid();
b10000b.clone_from(&b10000);
assert_eq!(b10000b.count(), 2);
for i in 2000..8000 {
b10000b.insert(i);
}
b10000.union(&b10000b);
assert_eq!(b10000.count(), 6000);
b10000.union(&b10000b);
assert_eq!(b10000.count(), 6000);
b10000.assert_valid();
b10000b.assert_valid();
}
#[test]
fn grow() {
let mut set: GrowableBitSet<usize> = GrowableBitSet::with_capacity(65);

View File

@ -1,11 +1,21 @@
#![feature(allow_internal_unstable)]
#![feature(bench_black_box)]
#![feature(extend_one)]
#![feature(min_specialization)]
#![feature(step_trait)]
#![feature(test)]
#![feature(let_else)]
#![feature(min_specialization)]
#![feature(new_uninit)]
#![feature(step_trait)]
#![feature(stmt_expr_attributes)]
#![feature(test)]
pub mod bit_set;
pub mod interval;
pub mod vec;
/// Type size assertion. The first argument is a type and the second argument is its expected size.
#[macro_export]
macro_rules! static_assert_size {
($ty:ty, $size:expr) => {
const _: [(); $size] = [(); ::std::mem::size_of::<$ty>()];
};
}

View File

@ -1,10 +1,12 @@
//! Random access inspection of the results of a dataflow analysis.
use crate::framework::BitSetExt;
use std::borrow::Borrow;
use std::cmp::Ordering;
#[cfg(debug_assertions)]
use rustc_index::bit_set::BitSet;
use rustc_index::vec::Idx;
use rustc_middle::mir::{self, BasicBlock, Location};
use super::{Analysis, Direction, Effect, EffectIndex, Results};
@ -209,13 +211,13 @@ where
}
}
impl<'mir, 'tcx, A, R, T> ResultsCursor<'mir, 'tcx, A, R>
impl<'mir, 'tcx, A, R> ResultsCursor<'mir, 'tcx, A, R>
where
A: Analysis<'tcx, Domain = BitSet<T>>,
T: Idx,
A: crate::GenKillAnalysis<'tcx>,
A::Domain: BitSetExt<A::Idx>,
R: Borrow<Results<'tcx, A>>,
{
pub fn contains(&self, elem: T) -> bool {
pub fn contains(&self, elem: A::Idx) -> bool {
self.get().contains(elem)
}
}

View File

@ -1,6 +1,7 @@
//! A solver for dataflow problems.
use std::borrow::BorrowMut;
use crate::framework::BitSetExt;
use std::ffi::OsString;
use std::path::PathBuf;
@ -91,7 +92,7 @@ where
impl<'a, 'tcx, A, D, T> Engine<'a, 'tcx, A>
where
A: GenKillAnalysis<'tcx, Idx = T, Domain = D>,
D: Clone + JoinSemiLattice + GenKill<T> + BorrowMut<BitSet<T>>,
D: Clone + JoinSemiLattice + GenKill<T> + BitSetExt<T>,
T: Idx,
{
/// Creates a new `Engine` to solve a gen-kill dataflow problem.
@ -106,7 +107,7 @@ where
// Otherwise, compute and store the cumulative transfer function for each block.
let identity = GenKillSet::identity(analysis.bottom_value(body).borrow().domain_size());
let identity = GenKillSet::identity(analysis.bottom_value(body).domain_size());
let mut trans_for_block = IndexVec::from_elem(identity, body.basic_blocks());
for (block, block_data) in body.basic_blocks().iter_enumerated() {
@ -115,7 +116,7 @@ where
}
let apply_trans = Box::new(move |bb: BasicBlock, state: &mut A::Domain| {
trans_for_block[bb].apply(state.borrow_mut());
trans_for_block[bb].apply(state);
});
Self::new(tcx, body, analysis, Some(apply_trans as Box<_>))

View File

@ -1,7 +1,7 @@
//! Custom formatting traits used when outputting Graphviz diagrams with the results of a dataflow
//! analysis.
use rustc_index::bit_set::{BitSet, HybridBitSet};
use rustc_index::bit_set::{BitSet, ChunkedBitSet, HybridBitSet};
use rustc_index::vec::Idx;
use std::fmt;
@ -133,6 +133,19 @@ where
}
}
impl<T, C> DebugWithContext<C> for ChunkedBitSet<T>
where
T: Idx + DebugWithContext<C>,
{
fn fmt_with(&self, _ctxt: &C, _f: &mut fmt::Formatter<'_>) -> fmt::Result {
unimplemented!("implement when/if needed");
}
fn fmt_diff_with(&self, _old: &Self, _ctxt: &C, _f: &mut fmt::Formatter<'_>) -> fmt::Result {
unimplemented!("implement when/if needed");
}
}
impl<T, C> DebugWithContext<C> for &'_ T
where
T: DebugWithContext<C>,

View File

@ -38,7 +38,8 @@
//! [Hasse diagram]: https://en.wikipedia.org/wiki/Hasse_diagram
//! [poset]: https://en.wikipedia.org/wiki/Partially_ordered_set
use rustc_index::bit_set::BitSet;
use crate::framework::BitSetExt;
use rustc_index::bit_set::{BitSet, ChunkedBitSet, HybridBitSet};
use rustc_index::vec::{Idx, IndexVec};
use std::iter;
@ -145,6 +146,18 @@ impl<T: Idx> MeetSemiLattice for BitSet<T> {
}
}
impl<T: Idx> JoinSemiLattice for ChunkedBitSet<T> {
fn join(&mut self, other: &Self) -> bool {
self.union(other)
}
}
impl<T: Idx> MeetSemiLattice for ChunkedBitSet<T> {
fn meet(&mut self, other: &Self) -> bool {
self.intersect(other)
}
}
/// The counterpart of a given semilattice `T` using the [inverse order].
///
/// The dual of a join-semilattice is a meet-semilattice and vice versa. For example, the dual of a
@ -155,15 +168,21 @@ impl<T: Idx> MeetSemiLattice for BitSet<T> {
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct Dual<T>(pub T);
impl<T> std::borrow::Borrow<T> for Dual<T> {
fn borrow(&self) -> &T {
&self.0
impl<T: Idx> BitSetExt<T> for Dual<BitSet<T>> {
fn domain_size(&self) -> usize {
self.0.domain_size()
}
}
impl<T> std::borrow::BorrowMut<T> for Dual<T> {
fn borrow_mut(&mut self) -> &mut T {
&mut self.0
fn contains(&self, elem: T) -> bool {
self.0.contains(elem)
}
fn union(&mut self, other: &HybridBitSet<T>) {
self.0.union(other);
}
fn subtract(&mut self, other: &HybridBitSet<T>) {
self.0.subtract(other);
}
}

View File

@ -30,10 +30,9 @@
//!
//! [gen-kill]: https://en.wikipedia.org/wiki/Data-flow_analysis#Bit_vector_problems
use std::borrow::BorrowMut;
use std::cmp::Ordering;
use rustc_index::bit_set::{BitSet, HybridBitSet};
use rustc_index::bit_set::{BitSet, ChunkedBitSet, HybridBitSet};
use rustc_index::vec::Idx;
use rustc_middle::mir::{self, BasicBlock, Location};
use rustc_middle::ty::TyCtxt;
@ -52,6 +51,51 @@ pub use self::engine::{Engine, Results};
pub use self::lattice::{JoinSemiLattice, MeetSemiLattice};
pub use self::visitor::{visit_results, ResultsVisitable, ResultsVisitor};
/// Analysis domains are all bitsets of various kinds. This trait holds
/// operations needed by all of them.
pub trait BitSetExt<T> {
fn domain_size(&self) -> usize;
fn contains(&self, elem: T) -> bool;
fn union(&mut self, other: &HybridBitSet<T>);
fn subtract(&mut self, other: &HybridBitSet<T>);
}
impl<T: Idx> BitSetExt<T> for BitSet<T> {
fn domain_size(&self) -> usize {
self.domain_size()
}
fn contains(&self, elem: T) -> bool {
self.contains(elem)
}
fn union(&mut self, other: &HybridBitSet<T>) {
self.union(other);
}
fn subtract(&mut self, other: &HybridBitSet<T>) {
self.subtract(other);
}
}
impl<T: Idx> BitSetExt<T> for ChunkedBitSet<T> {
fn domain_size(&self) -> usize {
self.domain_size()
}
fn contains(&self, elem: T) -> bool {
self.contains(elem)
}
fn union(&mut self, other: &HybridBitSet<T>) {
self.union(other);
}
fn subtract(&mut self, other: &HybridBitSet<T>) {
self.subtract(other);
}
}
/// Define the domain of a dataflow problem.
///
/// This trait specifies the lattice on which this analysis operates (the domain) as well as its
@ -303,7 +347,7 @@ pub trait GenKillAnalysis<'tcx>: Analysis<'tcx> {
impl<'tcx, A> Analysis<'tcx> for A
where
A: GenKillAnalysis<'tcx>,
A::Domain: GenKill<A::Idx> + BorrowMut<BitSet<A::Idx>>,
A::Domain: GenKill<A::Idx> + BitSetExt<A::Idx>,
{
fn apply_statement_effect(
&self,
@ -435,7 +479,7 @@ impl<T: Idx> GenKillSet<T> {
}
}
pub fn apply(&self, state: &mut BitSet<T>) {
pub fn apply(&self, state: &mut impl BitSetExt<T>) {
state.union(&self.gen);
state.subtract(&self.kill);
}
@ -463,6 +507,16 @@ impl<T: Idx> GenKill<T> for BitSet<T> {
}
}
impl<T: Idx> GenKill<T> for ChunkedBitSet<T> {
fn gen(&mut self, elem: T) {
self.insert(elem);
}
fn kill(&mut self, elem: T) {
self.remove(elem);
}
}
impl<T: Idx> GenKill<T> for lattice::Dual<BitSet<T>> {
fn gen(&mut self, elem: T) {
self.0.insert(elem);

View File

@ -2,7 +2,7 @@
//! bitvectors attached to each basic block, represented via a
//! zero-sized structure.
use rustc_index::bit_set::BitSet;
use rustc_index::bit_set::{BitSet, ChunkedBitSet};
use rustc_index::vec::Idx;
use rustc_middle::mir::visit::{MirVisitable, Visitor};
use rustc_middle::mir::{self, Body, Location};
@ -286,12 +286,12 @@ impl<'a, 'tcx> DefinitelyInitializedPlaces<'a, 'tcx> {
}
impl<'tcx> AnalysisDomain<'tcx> for MaybeInitializedPlaces<'_, 'tcx> {
type Domain = BitSet<MovePathIndex>;
type Domain = ChunkedBitSet<MovePathIndex>;
const NAME: &'static str = "maybe_init";
fn bottom_value(&self, _: &mir::Body<'tcx>) -> Self::Domain {
// bottom = uninitialized
BitSet::new_empty(self.move_data().move_paths.len())
ChunkedBitSet::new_empty(self.move_data().move_paths.len())
}
fn initialize_start_block(&self, _: &mir::Body<'tcx>, state: &mut Self::Domain) {
@ -417,13 +417,13 @@ impl<'tcx> GenKillAnalysis<'tcx> for MaybeInitializedPlaces<'_, 'tcx> {
}
impl<'tcx> AnalysisDomain<'tcx> for MaybeUninitializedPlaces<'_, 'tcx> {
type Domain = BitSet<MovePathIndex>;
type Domain = ChunkedBitSet<MovePathIndex>;
const NAME: &'static str = "maybe_uninit";
fn bottom_value(&self, _: &mir::Body<'tcx>) -> Self::Domain {
// bottom = initialized (start_block_effect counters this at outset)
BitSet::new_empty(self.move_data().move_paths.len())
ChunkedBitSet::new_empty(self.move_data().move_paths.len())
}
// sets on_entry bits for Arg places
@ -606,13 +606,13 @@ impl<'tcx> GenKillAnalysis<'tcx> for DefinitelyInitializedPlaces<'_, 'tcx> {
}
impl<'tcx> AnalysisDomain<'tcx> for EverInitializedPlaces<'_, 'tcx> {
type Domain = BitSet<InitIndex>;
type Domain = ChunkedBitSet<InitIndex>;
const NAME: &'static str = "ever_init";
fn bottom_value(&self, _: &mir::Body<'tcx>) -> Self::Domain {
// bottom = no initialized variables by default
BitSet::new_empty(self.move_data().inits.len())
ChunkedBitSet::new_empty(self.move_data().inits.len())
}
fn initialize_start_block(&self, body: &mir::Body<'tcx>, state: &mut Self::Domain) {

View File

@ -1,5 +1,3 @@
use std::borrow::Borrow;
use rustc_ast::ast;
use rustc_span::symbol::sym;
use rustc_span::Span;
@ -10,6 +8,7 @@ use rustc_middle::mir::MirPass;
use rustc_middle::mir::{self, Body, Local, Location};
use rustc_middle::ty::{self, Ty, TyCtxt};
use crate::framework::BitSetExt;
use crate::impls::{
DefinitelyInitializedPlaces, MaybeInitializedPlaces, MaybeLiveLocals, MaybeUninitializedPlaces,
};
@ -248,7 +247,7 @@ pub trait RustcPeekAt<'tcx>: Analysis<'tcx> {
impl<'tcx, A, D> RustcPeekAt<'tcx> for A
where
A: Analysis<'tcx, Domain = D> + HasMoveData<'tcx>,
D: JoinSemiLattice + Clone + Borrow<BitSet<MovePathIndex>>,
D: JoinSemiLattice + Clone + BitSetExt<MovePathIndex>,
{
fn peek_at(
&self,
@ -259,7 +258,7 @@ where
) {
match self.move_data().rev_lookup.find(place.as_ref()) {
LookupResult::Exact(peek_mpi) => {
let bit_state = flow_state.borrow().contains(peek_mpi);
let bit_state = flow_state.contains(peek_mpi);
debug!("rustc_peek({:?} = &{:?}) bit_state: {}", call.arg, place, bit_state);
if !bit_state {
tcx.sess.span_err(call.span, "rustc_peek: bit not set");

View File

@ -1,4 +1,4 @@
use rustc_index::bit_set::BitSet;
use rustc_index::bit_set::ChunkedBitSet;
use rustc_middle::mir::{Body, Field, Rvalue, Statement, StatementKind, TerminatorKind};
use rustc_middle::ty::subst::SubstsRef;
use rustc_middle::ty::{self, ParamEnv, Ty, TyCtxt, VariantDef};
@ -89,7 +89,7 @@ impl<'tcx> MirPass<'tcx> for RemoveUninitDrops {
fn is_needs_drop_and_init<'tcx>(
tcx: TyCtxt<'tcx>,
param_env: ParamEnv<'tcx>,
maybe_inits: &BitSet<MovePathIndex>,
maybe_inits: &ChunkedBitSet<MovePathIndex>,
move_data: &MoveData<'tcx>,
ty: Ty<'tcx>,
mpi: MovePathIndex,