Previously, Item was a struct of a NonZeroU64, an Option which was usually unset or irrelevant, and a 4-variant enum. So collectively, the size of an Item was 24 bytes, but only 8 bytes were used for the most part. So this takes advantage of the fact that it is probably impossible to exhaust the total space of SbTags, and steals 3 bits from it to pack the whole struct into a single u64. This bit-packing means that we reduce peak memory usage when Miri goes memory-bound by ~3x. We also get CPU performance improvements of varying size, because not only are we simply accessing less memory, we can now compare a Vec<Item> using a memcmp because it does not have any padding.
386 lines
16 KiB
Rust
386 lines
16 KiB
Rust
use crate::stacked_borrows::{AccessKind, Item, Permission, SbTag, SbTagExtra};
|
|
use rustc_data_structures::fx::FxHashSet;
|
|
#[cfg(feature = "stack-cache")]
|
|
use std::ops::Range;
|
|
|
|
/// Exactly what cache size we should use is a difficult tradeoff. There will always be some
|
|
/// workload which has a `SbTag` working set which exceeds the size of the cache, and ends up
|
|
/// falling back to linear searches of the borrow stack very often.
|
|
/// The cost of making this value too large is that the loop in `Stack::insert` which ensures the
|
|
/// entries in the cache stay correct after an insert becomes expensive.
|
|
#[cfg(feature = "stack-cache")]
|
|
const CACHE_LEN: usize = 32;
|
|
|
|
/// Extra per-location state.
|
|
#[derive(Clone, Debug)]
|
|
pub struct Stack {
|
|
/// Used *mostly* as a stack; never empty.
|
|
/// Invariants:
|
|
/// * Above a `SharedReadOnly` there can only be more `SharedReadOnly`.
|
|
/// * Except for `Untagged`, no tag occurs in the stack more than once.
|
|
borrows: Vec<Item>,
|
|
/// If this is `Some(id)`, then the actual current stack is unknown. This can happen when
|
|
/// wildcard pointers are used to access this location. What we do know is that `borrows` are at
|
|
/// the top of the stack, and below it are arbitrarily many items whose `tag` is strictly less
|
|
/// than `id`.
|
|
/// When the bottom is unknown, `borrows` always has a `SharedReadOnly` or `Unique` at the bottom;
|
|
/// we never have the unknown-to-known boundary in an SRW group.
|
|
unknown_bottom: Option<SbTag>,
|
|
|
|
/// A small LRU cache of searches of the borrow stack.
|
|
#[cfg(feature = "stack-cache")]
|
|
cache: StackCache,
|
|
/// On a read, we need to disable all `Unique` above the granting item. We can avoid most of
|
|
/// this scan by keeping track of the region of the borrow stack that may contain `Unique`s.
|
|
#[cfg(feature = "stack-cache")]
|
|
unique_range: Range<usize>,
|
|
}
|
|
|
|
/// A very small cache of searches of the borrow stack
|
|
/// This maps items to locations in the borrow stack. Any use of this still needs to do a
|
|
/// probably-cold random access into the borrow stack to figure out what `Permission` an
|
|
/// `SbTag` grants. We could avoid this by also storing the `Permission` in the cache, but
|
|
/// most lookups into the cache are immediately followed by access of the full borrow stack anyway.
|
|
///
|
|
/// It may seem like maintaining this cache is a waste for small stacks, but
|
|
/// (a) iterating over small fixed-size arrays is super fast, and (b) empirically this helps *a lot*,
|
|
/// probably because runtime is dominated by large stacks.
|
|
#[cfg(feature = "stack-cache")]
|
|
#[derive(Clone, Debug)]
|
|
struct StackCache {
|
|
items: [Item; CACHE_LEN], // Hot in find_granting
|
|
idx: [usize; CACHE_LEN], // Hot in grant
|
|
}
|
|
|
|
#[cfg(feature = "stack-cache")]
|
|
impl StackCache {
|
|
/// When a tag is used, we call this function to add or refresh it in the cache.
|
|
///
|
|
/// We use the position in the cache to represent how recently a tag was used; the first position
|
|
/// is the most recently used tag. So an add shifts every element towards the end, and inserts
|
|
/// the new element at the start. We lose the last element.
|
|
/// This strategy is effective at keeping the most-accessed items in the cache, but it costs a
|
|
/// linear shift across the entire cache when we add a new tag.
|
|
fn add(&mut self, idx: usize, item: Item) {
|
|
self.items.copy_within(0..CACHE_LEN - 1, 1);
|
|
self.items[0] = item;
|
|
self.idx.copy_within(0..CACHE_LEN - 1, 1);
|
|
self.idx[0] = idx;
|
|
}
|
|
}
|
|
|
|
impl PartialEq for Stack {
|
|
fn eq(&self, other: &Self) -> bool {
|
|
// All the semantics of Stack are in self.borrows, everything else is caching
|
|
self.borrows == other.borrows
|
|
}
|
|
}
|
|
|
|
impl Eq for Stack {}
|
|
|
|
impl<'tcx> Stack {
|
|
/// Panics if any of the caching mechanisms have broken,
|
|
/// - The StackCache indices don't refer to the parallel items,
|
|
/// - There are no Unique items outside of first_unique..last_unique
|
|
#[cfg(feature = "expensive-debug-assertions")]
|
|
fn verify_cache_consistency(&self) {
|
|
// Only a full cache needs to be valid. Also see the comments in find_granting_cache
|
|
// and set_unknown_bottom.
|
|
if self.borrows.len() >= CACHE_LEN {
|
|
for (tag, stack_idx) in self.cache.items.iter().zip(self.cache.idx.iter()) {
|
|
assert_eq!(self.borrows[*stack_idx], *tag);
|
|
}
|
|
}
|
|
|
|
for (idx, item) in self.borrows.iter().enumerate() {
|
|
if item.perm() == Permission::Unique {
|
|
assert!(
|
|
self.unique_range.contains(&idx),
|
|
"{:?} {:?}",
|
|
self.unique_range,
|
|
self.borrows
|
|
);
|
|
}
|
|
}
|
|
}
|
|
|
|
/// Find the item granting the given kind of access to the given tag, and return where
|
|
/// it is on the stack. For wildcard tags, the given index is approximate, but if *no*
|
|
/// index is given it means the match was *not* in the known part of the stack.
|
|
/// `Ok(None)` indicates it matched the "unknown" part of the stack.
|
|
/// `Err` indicates it was not found.
|
|
pub(super) fn find_granting(
|
|
&mut self,
|
|
access: AccessKind,
|
|
tag: SbTagExtra,
|
|
exposed_tags: &FxHashSet<SbTag>,
|
|
) -> Result<Option<usize>, ()> {
|
|
#[cfg(feature = "expensive-debug-assertions")]
|
|
self.verify_cache_consistency();
|
|
|
|
let SbTagExtra::Concrete(tag) = tag else {
|
|
// Handle the wildcard case.
|
|
// Go search the stack for an exposed tag.
|
|
if let Some(idx) =
|
|
self.borrows
|
|
.iter()
|
|
.enumerate() // we also need to know *where* in the stack
|
|
.rev() // search top-to-bottom
|
|
.find_map(|(idx, item)| {
|
|
// If the item fits and *might* be this wildcard, use it.
|
|
if item.perm().grants(access) && exposed_tags.contains(&item.tag()) {
|
|
Some(idx)
|
|
} else {
|
|
None
|
|
}
|
|
})
|
|
{
|
|
return Ok(Some(idx));
|
|
}
|
|
// If we couldn't find it in the stack, check the unknown bottom.
|
|
return if self.unknown_bottom.is_some() { Ok(None) } else { Err(()) };
|
|
};
|
|
|
|
if let Some(idx) = self.find_granting_tagged(access, tag) {
|
|
return Ok(Some(idx));
|
|
}
|
|
|
|
// Couldn't find it in the stack; but if there is an unknown bottom it might be there.
|
|
let found = self.unknown_bottom.is_some_and(|&unknown_limit| {
|
|
tag.0 < unknown_limit.0 // unknown_limit is an upper bound for what can be in the unknown bottom.
|
|
});
|
|
if found { Ok(None) } else { Err(()) }
|
|
}
|
|
|
|
fn find_granting_tagged(&mut self, access: AccessKind, tag: SbTag) -> Option<usize> {
|
|
#[cfg(feature = "stack-cache")]
|
|
if let Some(idx) = self.find_granting_cache(access, tag) {
|
|
return Some(idx);
|
|
}
|
|
|
|
// If we didn't find the tag in the cache, fall back to a linear search of the
|
|
// whole stack, and add the tag to the cache.
|
|
for (stack_idx, item) in self.borrows.iter().enumerate().rev() {
|
|
if tag == item.tag() && item.perm().grants(access) {
|
|
#[cfg(feature = "stack-cache")]
|
|
self.cache.add(stack_idx, *item);
|
|
return Some(stack_idx);
|
|
}
|
|
}
|
|
None
|
|
}
|
|
|
|
#[cfg(feature = "stack-cache")]
|
|
fn find_granting_cache(&mut self, access: AccessKind, tag: SbTag) -> Option<usize> {
|
|
// This looks like a common-sense optimization; we're going to do a linear search of the
|
|
// cache or the borrow stack to scan the shorter of the two. This optimization is miniscule
|
|
// and this check actually ensures we do not access an invalid cache.
|
|
// When a stack is created and when items are removed from the top of the borrow stack, we
|
|
// need some valid value to populate the cache. In both cases, we try to use the bottom
|
|
// item. But when the stack is cleared in `set_unknown_bottom` there is nothing we could
|
|
// place in the cache that is correct. But due to the way we populate the cache in
|
|
// `StackCache::add`, we know that when the borrow stack has grown larger than the cache,
|
|
// every slot in the cache is valid.
|
|
if self.borrows.len() <= CACHE_LEN {
|
|
return None;
|
|
}
|
|
// Search the cache for the tag we're looking up
|
|
let cache_idx = self.cache.items.iter().position(|t| t.tag() == tag)?;
|
|
let stack_idx = self.cache.idx[cache_idx];
|
|
// If we found the tag, look up its position in the stack to see if it grants
|
|
// the required permission
|
|
if self.cache.items[cache_idx].perm().grants(access) {
|
|
// If it does, and it's not already in the most-recently-used position, re-insert it at
|
|
// the most-recently-used position. This technically reduces the efficiency of the
|
|
// cache by duplicating elements, but current benchmarks do not seem to benefit from
|
|
// avoiding this duplication.
|
|
// But if the tag is in position 1, avoiding the duplicating add is trivial.
|
|
// If it does, and it's not already in the most-recently-used position, move it there.
|
|
// Except if the tag is in position 1, this is equivalent to just a swap, so do that.
|
|
if cache_idx == 1 {
|
|
self.cache.items.swap(0, 1);
|
|
self.cache.idx.swap(0, 1);
|
|
} else if cache_idx > 1 {
|
|
self.cache.add(stack_idx, self.cache.items[cache_idx]);
|
|
}
|
|
Some(stack_idx)
|
|
} else {
|
|
// Tag is in the cache, but it doesn't grant the required permission
|
|
None
|
|
}
|
|
}
|
|
|
|
pub fn insert(&mut self, new_idx: usize, new: Item) {
|
|
self.borrows.insert(new_idx, new);
|
|
|
|
#[cfg(feature = "stack-cache")]
|
|
self.insert_cache(new_idx, new);
|
|
}
|
|
|
|
#[cfg(feature = "stack-cache")]
|
|
fn insert_cache(&mut self, new_idx: usize, new: Item) {
|
|
// Adjust the possibly-unique range if an insert occurs before or within it
|
|
if self.unique_range.start >= new_idx {
|
|
self.unique_range.start += 1;
|
|
}
|
|
if self.unique_range.end >= new_idx {
|
|
self.unique_range.end += 1;
|
|
}
|
|
if new.perm() == Permission::Unique {
|
|
// Make sure the possibly-unique range contains the new borrow
|
|
self.unique_range.start = self.unique_range.start.min(new_idx);
|
|
self.unique_range.end = self.unique_range.end.max(new_idx + 1);
|
|
}
|
|
|
|
// The above insert changes the meaning of every index in the cache >= new_idx, so now
|
|
// we need to find every one of those indexes and increment it.
|
|
// But if the insert is at the end (equivalent to a push), we can skip this step because
|
|
// it didn't change the position of any other items.
|
|
if new_idx != self.borrows.len() - 1 {
|
|
for idx in &mut self.cache.idx {
|
|
if *idx >= new_idx {
|
|
*idx += 1;
|
|
}
|
|
}
|
|
}
|
|
|
|
// This primes the cache for the next access, which is almost always the just-added tag.
|
|
self.cache.add(new_idx, new);
|
|
|
|
#[cfg(feature = "expensive-debug-assertions")]
|
|
self.verify_cache_consistency();
|
|
}
|
|
|
|
/// Construct a new `Stack` using the passed `Item` as the base tag.
|
|
pub fn new(item: Item) -> Self {
|
|
Stack {
|
|
borrows: vec![item],
|
|
unknown_bottom: None,
|
|
#[cfg(feature = "stack-cache")]
|
|
cache: StackCache { idx: [0; CACHE_LEN], items: [item; CACHE_LEN] },
|
|
#[cfg(feature = "stack-cache")]
|
|
unique_range: if item.perm() == Permission::Unique { 0..1 } else { 0..0 },
|
|
}
|
|
}
|
|
|
|
pub fn get(&self, idx: usize) -> Option<Item> {
|
|
self.borrows.get(idx).cloned()
|
|
}
|
|
|
|
#[allow(clippy::len_without_is_empty)] // Stacks are never empty
|
|
pub fn len(&self) -> usize {
|
|
self.borrows.len()
|
|
}
|
|
|
|
pub fn unknown_bottom(&self) -> Option<SbTag> {
|
|
self.unknown_bottom
|
|
}
|
|
|
|
pub fn set_unknown_bottom(&mut self, tag: SbTag) {
|
|
// We clear the borrow stack but the lookup cache doesn't support clearing per se. Instead,
|
|
// there is a check explained in `find_granting_cache` which protects against accessing the
|
|
// cache when it has been cleared and not yet refilled.
|
|
self.borrows.clear();
|
|
self.unknown_bottom = Some(tag);
|
|
}
|
|
|
|
/// Find all `Unique` elements in this borrow stack above `granting_idx`, pass a copy of them
|
|
/// to the `visitor`, then set their `Permission` to `Disabled`.
|
|
pub fn disable_uniques_starting_at<V: FnMut(Item) -> crate::InterpResult<'tcx>>(
|
|
&mut self,
|
|
disable_start: usize,
|
|
mut visitor: V,
|
|
) -> crate::InterpResult<'tcx> {
|
|
#[cfg(feature = "stack-cache")]
|
|
let unique_range = self.unique_range.clone();
|
|
#[cfg(not(feature = "stack-cache"))]
|
|
let unique_range = 0..self.len();
|
|
|
|
if disable_start <= unique_range.end {
|
|
let lower = unique_range.start.max(disable_start);
|
|
let upper = (unique_range.end + 1).min(self.borrows.len());
|
|
for item in &mut self.borrows[lower..upper] {
|
|
if item.perm() == Permission::Unique {
|
|
log::trace!("access: disabling item {:?}", item);
|
|
visitor(*item)?;
|
|
item.set_disabled();
|
|
for t in &mut self.cache.items {
|
|
if t.tag() == item.tag() {
|
|
t.set_disabled();
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
#[cfg(feature = "stack-cache")]
|
|
if disable_start < self.unique_range.start {
|
|
// We disabled all Unique items
|
|
self.unique_range.start = 0;
|
|
self.unique_range.end = 0;
|
|
} else {
|
|
// Truncate the range to disable_start. This is + 2 because we are only removing
|
|
// elements after disable_start, and this range does not include the end.
|
|
self.unique_range.end = self.unique_range.end.min(disable_start + 1);
|
|
}
|
|
|
|
#[cfg(feature = "expensive-debug-assertions")]
|
|
self.verify_cache_consistency();
|
|
|
|
Ok(())
|
|
}
|
|
|
|
/// Produces an iterator which iterates over `range` in reverse, and when dropped removes that
|
|
/// range of `Item`s from this `Stack`.
|
|
pub fn pop_items_after<V: FnMut(Item) -> crate::InterpResult<'tcx>>(
|
|
&mut self,
|
|
start: usize,
|
|
mut visitor: V,
|
|
) -> crate::InterpResult<'tcx> {
|
|
while self.borrows.len() > start {
|
|
let item = self.borrows.pop().unwrap();
|
|
visitor(item)?;
|
|
}
|
|
|
|
#[cfg(feature = "stack-cache")]
|
|
if !self.borrows.is_empty() {
|
|
// After we remove from the borrow stack, every aspect of our caching may be invalid, but it is
|
|
// also possible that the whole cache is still valid. So we call this method to repair what
|
|
// aspects of the cache are now invalid, instead of resetting the whole thing to a trivially
|
|
// valid default state.
|
|
let base_tag = self.borrows[0];
|
|
let mut removed = 0;
|
|
let mut cursor = 0;
|
|
// Remove invalid entries from the cache by rotating them to the end of the cache, then
|
|
// keep track of how many invalid elements there are and overwrite them with the base tag.
|
|
// The base tag here serves as a harmless default value.
|
|
for _ in 0..CACHE_LEN - 1 {
|
|
if self.cache.idx[cursor] >= start {
|
|
self.cache.idx[cursor..CACHE_LEN - removed].rotate_left(1);
|
|
self.cache.items[cursor..CACHE_LEN - removed].rotate_left(1);
|
|
removed += 1;
|
|
} else {
|
|
cursor += 1;
|
|
}
|
|
}
|
|
for i in CACHE_LEN - removed - 1..CACHE_LEN {
|
|
self.cache.idx[i] = 0;
|
|
self.cache.items[i] = base_tag;
|
|
}
|
|
|
|
if start < self.unique_range.start.saturating_sub(1) {
|
|
// We removed all the Unique items
|
|
self.unique_range = 0..0;
|
|
} else {
|
|
// Ensure the range doesn't extend past the new top of the stack
|
|
self.unique_range.end = self.unique_range.end.min(start + 1);
|
|
}
|
|
} else {
|
|
self.unique_range = 0..0;
|
|
}
|
|
|
|
#[cfg(feature = "expensive-debug-assertions")]
|
|
self.verify_cache_consistency();
|
|
Ok(())
|
|
}
|
|
}
|