Introduce ChunkedBitSet and use it for some dataflow analyses.

This reduces peak memory usage significantly for some programs with very
large functions, such as:
- `keccak`, `unicode_normalization`, and `match-stress-enum`, from
  the `rustc-perf` benchmark suite;
- `http-0.2.6` from crates.io.

The new type is used in the analyses where the bitsets can get huge
(e.g. 10s of thousands of bits): `MaybeInitializedPlaces`,
`MaybeUninitializedPlaces`, and `EverInitializedPlaces`.

Some refactoring was required in `rustc_mir_dataflow`. All existing
analysis domains are either `BitSet` or a trivial wrapper around
`BitSet`, and access in a few places is done via `Borrow<BitSet>` or
`BorrowMut<BitSet>`. Now that some of these domains are `ClusterBitSet`,
that no longer works. So this commit replaces the `Borrow`/`BorrowMut`
usage with a new trait `BitSetExt` containing the needed bitset
operations. The impls just forward these to the underlying bitset type.
This required fiddling with trait bounds in a few places.

The commit also:
- Moves `static_assert_size` from `rustc_data_structures` to
  `rustc_index` so it can be used in the latter; the former now
  re-exports it so existing users are unaffected.
- Factors out some common "clear excess bits in the final word"
  functionality in `bit_set.rs`.
- Uses `fill` in a few places instead of loops.

compiler/rustc_index/src/bit_set.rs

@@ -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