Auto merge of #17179 - bholley:one_selector_allocation, r=emilio

shrink Rule and store all heap-allocated selector data inline

https://bugzilla.mozilla.org/show_bug.cgi?id=1370107

<!-- Reviewable:start -->
---
This change is [<img src="https://reviewable.io/review_button.svg" height="34" align="absmiddle" alt="Reviewable"/>](https://reviewable.io/reviews/servo/servo/17179)
<!-- Reviewable:end -->

Cargo.lock

@@ -2523,6 +2523,7 @@ dependencies = [
  "phf 0.7.21 (registry+https://github.com/rust-lang/crates.io-index)",
  "phf_codegen 0.7.21 (registry+https://github.com/rust-lang/crates.io-index)",
  "precomputed-hash 0.1.0 (registry+https://github.com/rust-lang/crates.io-index)",
+ "servo_arc 0.0.1",
  "size_of_test 0.0.1",
  "smallvec 0.4.0 (registry+https://github.com/rust-lang/crates.io-index)",
 ]
@@ -2688,6 +2689,15 @@ dependencies = [
  "url 1.4.1 (registry+https://github.com/rust-lang/crates.io-index)",
 ]
 
+[[package]]
+name = "servo_arc"
+version = "0.0.1"
+dependencies = [
+ "heapsize 0.4.0 (registry+https://github.com/rust-lang/crates.io-index)",
+ "nodrop 0.1.8 (registry+https://github.com/rust-lang/crates.io-index)",
+ "serde 0.9.15 (registry+https://github.com/rust-lang/crates.io-index)",
+]
+
 [[package]]
 name = "servo_atoms"
 version = "0.0.1"
@@ -2913,6 +2923,7 @@ dependencies = [
  "selectors 0.19.0",
  "serde 0.9.15 (registry+https://github.com/rust-lang/crates.io-index)",
  "serde_derive 0.9.15 (registry+https://github.com/rust-lang/crates.io-index)",
+ "servo_arc 0.0.1",
  "servo_atoms 0.0.1",
  "servo_config 0.0.1",
  "servo_url 0.0.1",

components/script/dom/element.rs

@@ -2061,7 +2061,7 @@ impl ElementMethods for Element {
             Err(()) => Err(Error::Syntax),
             Ok(selectors) => {
                 let mut ctx = MatchingContext::new(MatchingMode::Normal, None);
-                Ok(matches_selector_list(&selectors.0, &Root::from_ref(self), &mut ctx))
+                Ok(matches_selector_list(&selectors, &Root::from_ref(self), &mut ctx))
             }
         }
     }
@@ -2080,7 +2080,7 @@ impl ElementMethods for Element {
                 for element in root.inclusive_ancestors() {
                     if let Some(element) = Root::downcast::<Element>(element) {
                         let mut ctx = MatchingContext::new(MatchingMode::Normal, None);
-                        if matches_selector_list(&selectors.0, &element, &mut ctx) {
+                        if matches_selector_list(&selectors, &element, &mut ctx) {
                             return Ok(Some(element));
                         }
                     }

components/script/dom/node.rs

@@ -345,7 +345,7 @@ impl<'a> Iterator for QuerySelectorIterator {
     type Item = Root<Node>;
 
     fn next(&mut self) -> Option<Root<Node>> {
-        let selectors = &self.selectors.0;
+        let selectors = &self.selectors;
 
         // TODO(cgaebel): Is it worth it to build a bloom filter here
         // (instead of passing `None`)? Probably.
@@ -722,7 +722,7 @@ impl Node {
             Ok(selectors) => {
                 let mut ctx = MatchingContext::new(MatchingMode::Normal, None);
                 Ok(self.traverse_preorder().filter_map(Root::downcast).find(|element| {
-                    matches_selector_list(&selectors.0, element, &mut ctx)
+                    matches_selector_list(&selectors, element, &mut ctx)
                 }))
             }
         }

components/selectors/Cargo.toml

@@ -28,6 +28,7 @@ cssparser = "0.13.7"
 fnv = "1.0"
 phf = "0.7.18"
 precomputed-hash = "0.1"
+servo_arc = { path = "../servo_arc" }
 smallvec = "0.4"
 
 [dev-dependencies]

components/selectors/arcslice.rs

@@ -1,326 +0,0 @@
-/* Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
- * http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
- * <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
- * option. This file may not be copied, modified, or distributed
- * except according to those terms.
- *
- * See the COPYRIGHT file at the top-level directory of this distribution */
-//! A thread-safe reference-counted slice type.
-//!
-//! Forked from https://github.com/huonw/shared_slice , which doesn't work on
-//! rust stable.
-
-use std::{cmp, fmt, ops};
-use std::hash::{Hash, Hasher};
-use std::sync::{Arc, Weak};
-
-
-/// A reference-counted slice type.
-pub struct ArcSlice<T> {
-    data: *const [T],
-    counts: Arc<Box<[T]>>,
-}
-
-unsafe impl<T: Send + Sync> Send for ArcSlice<T> {}
-unsafe impl<T: Send + Sync> Sync for ArcSlice<T> {}
-
-/// A non-owning reference-counted slice type.
-///
-/// This is to `ArcSlice` as `std::sync::Weak` is to `std::sync::Arc`, and
-/// allows one to have cyclic references without stopping memory from
-/// being deallocated.
-pub struct WeakSlice<T> {
-    data: *const [T],
-    counts: Weak<Box<[T]>>,
-}
-unsafe impl<T: Send + Sync> Send for WeakSlice<T> {}
-unsafe impl<T: Send + Sync> Sync for WeakSlice<T> {}
-
-impl<T> ArcSlice<T> {
-    /// Construct a new `ArcSlice` containing the elements of `slice`.
-    ///
-    /// This reuses the allocation of `slice`.
-    pub fn new(slice: Box<[T]>) -> ArcSlice<T> {
-        ArcSlice {
-            data: &*slice,
-            counts: Arc::new(slice),
-        }
-    }
-
-    /// Downgrade self into a weak slice.
-    pub fn downgrade(&self) -> WeakSlice<T> {
-        WeakSlice {
-            data: self.data,
-            counts: Arc::downgrade(&self.counts)
-        }
-    }
-
-    /// Construct a new `ArcSlice` that only points to elements at
-    /// indices `lo` (inclusive) through `hi` (exclusive).
-    ///
-    /// This consumes `self` to avoid unnecessary reference-count
-    /// modifications. Use `.clone()` if it is necessary to refer to
-    /// `self` after calling this.
-    ///
-    /// # Panics
-    ///
-    /// Panics if `lo > hi` or if either are strictly greater than
-    /// `self.len()`.
-    pub fn slice(mut self, lo: usize, hi: usize) -> ArcSlice<T> {
-        self.data = &self[lo..hi];
-        self
-    }
-    /// Construct a new `ArcSlice` that only points to elements at
-    /// indices up to `hi` (exclusive).
-    ///
-    /// This consumes `self` to avoid unnecessary reference-count
-    /// modifications. Use `.clone()` if it is necessary to refer to
-    /// `self` after calling this.
-    ///
-    /// # Panics
-    ///
-    /// Panics if `hi > self.len()`.
-    pub fn slice_to(self, hi: usize) -> ArcSlice<T> {
-        self.slice(0, hi)
-    }
-    /// Construct a new `ArcSlice` that only points to elements at
-    /// indices starting at  `lo` (inclusive).
-    ///
-    /// This consumes `self` to avoid unnecessary reference-count
-    /// modifications. Use `.clone()` if it is necessary to refer to
-    /// `self` after calling this.
-    ///
-    /// # Panics
-    ///
-    /// Panics if `lo > self.len()`.
-    pub fn slice_from(self, lo: usize) -> ArcSlice<T> {
-        let hi = self.len();
-        self.slice(lo, hi)
-    }
-}
-
-impl<T> Clone for ArcSlice<T> {
-    fn clone(&self) -> ArcSlice<T> {
-        ArcSlice {
-            data: self.data,
-            counts: self.counts.clone()
-        }
-    }
-}
-
-impl<T> ops::Deref for ArcSlice<T> {
-    type Target = [T];
-    fn deref<'a>(&'a self) -> &'a [T] {
-        unsafe { &*self.data }
-    }
-}
-
-impl<T> AsRef<[T]> for ArcSlice<T> {
-    fn as_ref(&self) -> &[T] { &**self }
-}
-
-impl<T: PartialEq> PartialEq for ArcSlice<T> {
-    fn eq(&self, other: &ArcSlice<T>) -> bool { **self == **other }
-    fn ne(&self, other: &ArcSlice<T>) -> bool { **self != **other }
-}
-impl<T: Eq> Eq for ArcSlice<T> {}
-
-impl<T: PartialOrd> PartialOrd for ArcSlice<T> {
-    fn partial_cmp(&self, other: &ArcSlice<T>) -> Option<cmp::Ordering> {
-        (**self).partial_cmp(&**other)
-    }
-    fn lt(&self, other: &ArcSlice<T>) -> bool { **self < **other }
-    fn le(&self, other: &ArcSlice<T>) -> bool { **self <= **other }
-    fn gt(&self, other: &ArcSlice<T>) -> bool { **self > **other }
-    fn ge(&self, other: &ArcSlice<T>) -> bool { **self >= **other }
-}
-impl<T: Ord> Ord for ArcSlice<T> {
-    fn cmp(&self, other: &ArcSlice<T>) -> cmp::Ordering { (**self).cmp(&**other) }
-}
-
-impl<T: Hash> Hash for ArcSlice<T> {
-    fn hash<H: Hasher>(&self, state: &mut H) {
-        Hash::hash(&**self, state)
-    }
-}
-
-impl<T: fmt::Debug> fmt::Debug for ArcSlice<T> {
-    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        fmt::Debug::fmt(&**self, f)
-    }
-}
-
-impl<T> WeakSlice<T> {
-    /// Attempt to upgrade `self` to a strongly-counted `ArcSlice`.
-    ///
-    /// Returns `None` if this is not possible (the data has already
-    /// been freed).
-    pub fn upgrade(&self) -> Option<ArcSlice<T>> {
-        self.counts.upgrade().map(|counts| {
-            ArcSlice {
-                data: self.data,
-                counts: counts
-            }
-        })
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::cell::Cell;
-    use std::cmp::Ordering;
-    use std::sync::{Arc, Mutex};
-    use super::{ArcSlice, WeakSlice};
-    #[test]
-    fn clone() {
-        let x = ArcSlice::new(Box::new([Cell::new(false)]));
-        let y = x.clone();
-
-        assert_eq!(x[0].get(), false);
-        assert_eq!(y[0].get(), false);
-
-        x[0].set(true);
-        assert_eq!(x[0].get(), true);
-        assert_eq!(y[0].get(), true);
-    }
-
-    #[test]
-    fn test_upgrade_downgrade() {
-        let x = ArcSlice::new(Box::new([1]));
-        let y: WeakSlice<_> = x.downgrade();
-
-        assert_eq!(y.upgrade(), Some(x.clone()));
-
-        drop(x);
-
-        assert!(y.upgrade().is_none())
-    }
-
-    #[test]
-    fn test_total_cmp() {
-        let x = ArcSlice::new(Box::new([1, 2, 3]));
-        let y = ArcSlice::new(Box::new([1, 2, 3]));
-        let z = ArcSlice::new(Box::new([1, 2, 4]));
-
-        assert_eq!(x, x);
-        assert_eq!(x, y);
-        assert!(x != z);
-        assert!(y != z);
-
-        assert!(x < z);
-        assert!(x <= z);
-        assert!(!(x > z));
-        assert!(!(x >= z));
-
-        assert!(!(z < x));
-        assert!(!(z <= x));
-        assert!(z > x);
-        assert!(z >= x);
-
-        assert_eq!(x.partial_cmp(&x), Some(Ordering::Equal));
-        assert_eq!(x.partial_cmp(&y), Some(Ordering::Equal));
-        assert_eq!(x.partial_cmp(&z), Some(Ordering::Less));
-        assert_eq!(z.partial_cmp(&y), Some(Ordering::Greater));
-
-        assert_eq!(x.cmp(&x), Ordering::Equal);
-        assert_eq!(x.cmp(&y), Ordering::Equal);
-        assert_eq!(x.cmp(&z), Ordering::Less);
-        assert_eq!(z.cmp(&y), Ordering::Greater);
-    }
-
-    #[test]
-    fn test_partial_cmp() {
-        use std::f64;
-        let x = ArcSlice::new(Box::new([1.0, f64::NAN]));
-        let y = ArcSlice::new(Box::new([1.0, f64::NAN]));
-        let z = ArcSlice::new(Box::new([2.0, f64::NAN]));
-        let w = ArcSlice::new(Box::new([f64::NAN, 1.0]));
-        assert!(!(x == y));
-        assert!(x != y);
-
-        assert!(!(x < y));
-        assert!(!(x <= y));
-        assert!(!(x > y));
-        assert!(!(x >= y));
-
-        assert!(x < z);
-        assert!(x <= z);
-        assert!(!(x > z));
-        assert!(!(x >= z));
-
-        assert!(!(z < w));
-        assert!(!(z <= w));
-        assert!(!(z > w));
-        assert!(!(z >= w));
-
-        assert_eq!(x.partial_cmp(&x), None);
-        assert_eq!(x.partial_cmp(&y), None);
-        assert_eq!(x.partial_cmp(&z), Some(Ordering::Less));
-        assert_eq!(z.partial_cmp(&x), Some(Ordering::Greater));
-
-        assert_eq!(x.partial_cmp(&w), None);
-        assert_eq!(y.partial_cmp(&w), None);
-        assert_eq!(z.partial_cmp(&w), None);
-        assert_eq!(w.partial_cmp(&w), None);
-    }
-
-    #[test]
-    fn test_show() {
-        let x = ArcSlice::new(Box::new([1, 2]));
-        assert_eq!(format!("{:?}", x), "[1, 2]");
-
-        let y: ArcSlice<i32> = ArcSlice::new(Box::new([]));
-        assert_eq!(format!("{:?}", y), "[]");
-    }
-
-    #[test]
-    fn test_slice() {
-        let x = ArcSlice::new(Box::new([1, 2, 3]));
-        let real = [1, 2, 3];
-        for i in 0..3 + 1 {
-            for j in i..3 + 1 {
-                let slice: ArcSlice<_> = x.clone().slice(i, j);
-                assert_eq!(&*slice, &real[i..j]);
-            }
-            assert_eq!(&*x.clone().slice_to(i), &real[..i]);
-            assert_eq!(&*x.clone().slice_from(i), &real[i..]);
-        }
-    }
-
-
-    #[test]
-    fn test_send_sync() {
-        fn assert_send<T: Send>() {}
-        fn assert_sync<T: Send>() {}
-
-        assert_send::<ArcSlice<u8>>();
-        assert_sync::<ArcSlice<u8>>();
-        assert_send::<WeakSlice<u8>>();
-        assert_sync::<WeakSlice<u8>>();
-    }
-
-    #[test]
-    fn test_drop() {
-        let drop_flag = Arc::new(Mutex::new(0));
-        struct Foo(Arc<Mutex<i32>>);
-
-        impl Drop for Foo {
-            fn drop(&mut self) {
-                let mut n = self.0.lock().unwrap();
-                *n += 1;
-            }
-        }
-
-        let whole = ArcSlice::new(Box::new([Foo(drop_flag.clone()), Foo(drop_flag.clone())]));
-
-        drop(whole);
-        assert_eq!(*drop_flag.lock().unwrap(), 2);
-
-        *drop_flag.lock().unwrap() = 0;
-
-        let whole = ArcSlice::new(Box::new([Foo(drop_flag.clone()), Foo(drop_flag.clone())]));
-        let part = whole.slice(1, 2);
-        drop(part);
-        assert_eq!(*drop_flag.lock().unwrap(), 2);
-    }
-}

components/selectors/lib.rs

@@ -9,9 +9,9 @@ extern crate fnv;
 extern crate phf;
 extern crate precomputed_hash;
 #[cfg(test)] #[macro_use] extern crate size_of_test;
+extern crate servo_arc;
 extern crate smallvec;
 
-pub mod arcslice;
 pub mod attr;
 pub mod bloom;
 pub mod matching;

components/selectors/matching.rs

@@ -4,8 +4,8 @@
 
 use attr::{ParsedAttrSelectorOperation, AttrSelectorOperation, NamespaceConstraint};
 use bloom::BloomFilter;
-use parser::{Combinator, ComplexSelector, Component, LocalName};
+use parser::{AncestorHashes, Combinator, Component, LocalName};
-use parser::{Selector, SelectorInner, SelectorIter};
+use parser::{Selector, SelectorIter, SelectorList};
 use std::borrow::Borrow;
 use tree::Element;
 
@@ -152,27 +152,30 @@ impl<'a> MatchingContext<'a> {
     }
 }
 
-pub fn matches_selector_list<E>(selector_list: &[Selector<E::Impl>],
+pub fn matches_selector_list<E>(selector_list: &SelectorList<E::Impl>,
                                 element: &E,
                                 context: &mut MatchingContext)
                                 -> bool
     where E: Element
 {
-    selector_list.iter().any(|selector| {
+    selector_list.0.iter().any(|selector_and_hashes| {
-        matches_selector(&selector.inner,
+        matches_selector(&selector_and_hashes.selector,
+                         0,
+                         &selector_and_hashes.hashes,
                          element,
                          context,
                          &mut |_, _| {})
     })
 }
 
-fn may_match<E>(sel: &SelectorInner<E::Impl>,
+#[inline(always)]
+fn may_match<E>(hashes: &AncestorHashes,
                 bf: &BloomFilter)
                 -> bool
     where E: Element,
 {
     // Check against the list of precomputed hashes.
-    for hash in sel.ancestor_hashes.iter() {
+    for hash in hashes.0.iter() {
         // If we hit the 0 sentinel hash, that means the rest are zero as well.
         if *hash == 0 {
             break;
@@ -330,8 +333,18 @@ enum SelectorMatchingResult {
     NotMatchedGlobally,
 }
 
-/// Matches an inner selector.
+/// Matches a selector, fast-rejecting against a bloom filter.
-pub fn matches_selector<E, F>(selector: &SelectorInner<E::Impl>,
+///
+/// We accept an offset to allow consumers to represent and match against partial
+/// selectors (indexed from the right). We use this API design, rather than
+/// having the callers pass a SelectorIter, because creating a SelectorIter
+/// requires dereferencing the selector to get the length, which adds an
+/// unncessary cache miss for cases when we can fast-reject with AncestorHashes
+/// (which the caller can store inline with the selector pointer).
+#[inline(always)]
+pub fn matches_selector<E, F>(selector: &Selector<E::Impl>,
+                              offset: usize,
+                              hashes: &AncestorHashes,
                               element: &E,
                               context: &mut MatchingContext,
                               flags_setter: &mut F)
@@ -341,18 +354,17 @@ pub fn matches_selector<E, F>(selector: &SelectorInner<E::Impl>,
 {
     // Use the bloom filter to fast-reject.
     if let Some(filter) = context.bloom_filter {
-        if !may_match::<E>(&selector, filter) {
+        if !may_match::<E>(hashes, filter) {
             return false;
         }
     }
 
-    matches_complex_selector(&selector.complex, element, context, flags_setter)
+    matches_complex_selector(selector, offset, element, context, flags_setter)
 }
 
 /// Matches a complex selector.
-///
+pub fn matches_complex_selector<E, F>(complex_selector: &Selector<E::Impl>,
-/// Use `matches_selector` if you need to skip pseudos.
+                                      offset: usize,
-pub fn matches_complex_selector<E, F>(complex_selector: &ComplexSelector<E::Impl>,
                                       element: &E,
                                       context: &mut MatchingContext,
                                       flags_setter: &mut F)
@@ -360,11 +372,15 @@ pub fn matches_complex_selector<E, F>(complex_selector: &ComplexSelector<E::Impl
     where E: Element,
           F: FnMut(&E, ElementSelectorFlags),
 {
-    let mut iter = complex_selector.iter();
+    let mut iter = if offset == 0 {
+        complex_selector.iter()
+    } else {
+        complex_selector.iter_from(offset)
+    };
 
     if cfg!(debug_assertions) {
         if context.matching_mode == MatchingMode::ForStatelessPseudoElement {
-            assert!(complex_selector.iter().any(|c| {
+            assert!(iter.clone().any(|c| {
                 matches!(*c, Component::PseudoElement(..))
             }));
         }

components/selectors/size_of_tests.rs

@@ -11,10 +11,8 @@ use precomputed_hash::PrecomputedHash;
 use std::fmt;
 use visitor::SelectorVisitor;
 
-size_of_test!(size_of_selector, Selector<Impl>, 48);
+size_of_test!(size_of_selector, Selector<Impl>, 8);
 size_of_test!(size_of_pseudo_element, gecko_like_types::PseudoElement, 1);
-size_of_test!(size_of_selector_inner, SelectorInner<Impl>, 40);
-size_of_test!(size_of_complex_selector, ComplexSelector<Impl>, 24);
 
 size_of_test!(size_of_component, Component<Impl>, 32);
 size_of_test!(size_of_pseudo_class, PseudoClass, 24);

components/servo_arc/Cargo.toml

@@ -0,0 +1,18 @@
+[package]
+name = "servo_arc"
+version = "0.0.1"
+authors = ["The Servo Project Developers"]
+license = "MPL-2.0"
+publish = false
+
+[lib]
+name = "servo_arc"
+path = "lib.rs"
+
+[features]
+servo = ["serde", "heapsize"]
+
+[dependencies]
+heapsize = {version = "0.4.0", optional = true}
+serde = {version = "0.9", optional = true}
+nodrop = {version = "0.1.8"}

components/servo_arc/lib.rs

@@ -8,11 +8,13 @@
 // option. This file may not be copied, modified, or distributed
 // except according to those terms.
 
-//! Fork of Arc for the style system. This has the following advantages over std::Arc:
+//! Fork of Arc for Servo. This has the following advantages over std::Arc:
 //! * We don't waste storage on the weak reference count.
 //! * We don't do extra RMU operations to handle the possibility of weak references.
 //! * We can experiment with arena allocation (todo).
 //! * We can add methods to support our custom use cases [1].
+//! * We have support for dynamically-sized types (see from_header_and_iter).
+//! * We have support for thin arcs to unsized types (see ThinArc).
 //!
 //! [1] https://bugzilla.mozilla.org/show_bug.cgi?id=1360883
 
@@ -20,8 +22,12 @@
 // duplicate those here.
 #![allow(missing_docs)]
 
+#[cfg(feature = "servo")] extern crate serde;
+extern crate nodrop;
+
 #[cfg(feature = "servo")]
 use heapsize::HeapSizeOf;
+use nodrop::NoDrop;
 #[cfg(feature = "servo")]
 use serde::{Deserialize, Serialize};
 use std::{isize, usize};
@@ -30,8 +36,11 @@ use std::cmp::Ordering;
 use std::convert::From;
 use std::fmt;
 use std::hash::{Hash, Hasher};
+use std::iter::{ExactSizeIterator, Iterator};
 use std::mem;
 use std::ops::{Deref, DerefMut};
+use std::ptr;
+use std::slice;
 use std::sync::atomic;
 use std::sync::atomic::Ordering::{Acquire, Relaxed, Release};
 
@@ -412,3 +421,257 @@ impl<T: Serialize> Serialize for Arc<T>
         (**self).serialize(serializer)
     }
 }
+
+/// Structure to allow Arc-managing some fixed-sized data and a variably-sized
+/// slice in a single allocation.
+#[derive(Debug, Eq, PartialEq, PartialOrd)]
+pub struct HeaderSlice<H, T: ?Sized> {
+    /// The fixed-sized data.
+    pub header: H,
+
+    /// The dynamically-sized data.
+    pub slice: T,
+}
+
+#[inline(always)]
+fn divide_rounding_up(dividend: usize, divisor: usize) -> usize {
+    (dividend + divisor - 1) / divisor
+}
+
+impl<H, T> Arc<HeaderSlice<H, [T]>> {
+    /// Creates an Arc for a HeaderSlice using the given header struct and
+    /// iterator to generate the slice. The resulting Arc will be fat.
+    #[inline]
+    pub fn from_header_and_iter<I>(header: H, mut items: I) -> Self
+        where I: Iterator<Item=T> + ExactSizeIterator
+    {
+        use ::std::mem::size_of;
+        assert!(size_of::<T>() != 0, "Need to think about ZST");
+
+        // Compute the required size for the allocation.
+        let num_items = items.len();
+        let size = {
+            // First, determine the alignment of a hypothetical pointer to a
+            // HeaderSlice.
+            let fake_slice_ptr_align: usize = mem::align_of::<ArcInner<HeaderSlice<H, [T; 1]>>>();
+
+            // Next, synthesize a totally garbage (but properly aligned) pointer
+            // to a sequence of T.
+            let fake_slice_ptr = fake_slice_ptr_align as *const T;
+
+            // Convert that sequence to a fat pointer. The address component of
+            // the fat pointer will be garbage, but the length will be correct.
+            let fake_slice = unsafe { slice::from_raw_parts(fake_slice_ptr, num_items) };
+
+            // Pretend the garbage address points to our allocation target (with
+            // a trailing sequence of T), rather than just a sequence of T.
+            let fake_ptr = fake_slice as *const [T] as *const ArcInner<HeaderSlice<H, [T]>>;
+            let fake_ref: &ArcInner<HeaderSlice<H, [T]>> = unsafe { &*fake_ptr };
+
+            // Use size_of_val, which will combine static information about the
+            // type with the length from the fat pointer. The garbage address
+            // will not be used.
+            mem::size_of_val(fake_ref)
+        };
+
+        let ptr: *mut ArcInner<HeaderSlice<H, [T]>>;
+        unsafe {
+            // Allocate the buffer. We use Vec because the underlying allocation
+            // machinery isn't available in stable Rust.
+            //
+            // To avoid alignment issues, we allocate words rather than bytes,
+            // rounding up to the nearest word size.
+            let words_to_allocate = divide_rounding_up(size, size_of::<usize>());
+            let mut vec = Vec::<usize>::with_capacity(words_to_allocate);
+            vec.set_len(words_to_allocate);
+            let buffer = Box::into_raw(vec.into_boxed_slice()) as *mut usize as *mut u8;
+
+            // Synthesize the fat pointer. We do this by claiming we have a direct
+            // pointer to a [T], and then changing the type of the borrow. The key
+            // point here is that the length portion of the fat pointer applies
+            // only to the number of elements in the dynamically-sized portion of
+            // the type, so the value will be the same whether it points to a [T]
+            // or something else with a [T] as its last member.
+            let fake_slice: &mut [T] = slice::from_raw_parts_mut(buffer as *mut T, num_items);
+            ptr = fake_slice as *mut [T] as *mut ArcInner<HeaderSlice<H, [T]>>;
+
+            // Write the data.
+            ptr::write(&mut ((*ptr).count), atomic::AtomicUsize::new(1));
+            ptr::write(&mut ((*ptr).data.header), header);
+            let mut current: *mut T = &mut (*ptr).data.slice[0];
+            for _ in 0..num_items {
+                ptr::write(current, items.next().expect("ExactSizeIterator over-reported length"));
+                current = current.offset(1);
+            }
+            assert!(items.next().is_none(), "ExactSizeIterator under-reported length");
+
+            // We should have consumed the buffer exactly.
+            debug_assert!(current as *mut u8 == buffer.offset(size as isize));
+        }
+
+        // Return the fat Arc.
+        assert_eq!(size_of::<Self>(), size_of::<usize>() * 2, "The Arc will be fat");
+        Arc { ptr: ptr }
+    }
+}
+
+/// Header data with an inline length. Consumers that use HeaderWithLength as the
+/// Header type in HeaderSlice can take advantage of ThinArc.
+#[derive(Debug, Eq, PartialEq, PartialOrd)]
+pub struct HeaderWithLength<H> {
+    /// The fixed-sized data.
+    pub header: H,
+
+    /// The slice length.
+    length: usize,
+}
+
+impl<H> HeaderWithLength<H> {
+    /// Creates a new HeaderWithLength.
+    pub fn new(header: H, length: usize) -> Self {
+        HeaderWithLength {
+            header: header,
+            length: length,
+        }
+    }
+}
+
+pub struct ThinArc<H, T> {
+    ptr: *mut ArcInner<HeaderSlice<HeaderWithLength<H>, [T; 1]>>,
+}
+
+unsafe impl<H: Sync + Send, T: Sync + Send> Send for ThinArc<H, T> {}
+unsafe impl<H: Sync + Send, T: Sync + Send> Sync for ThinArc<H, T> {}
+
+// Synthesize a fat pointer from a thin pointer.
+//
+// See the comment around the analogous operation in from_header_and_iter.
+fn thin_to_thick<H, T>(thin: *mut ArcInner<HeaderSlice<HeaderWithLength<H>, [T; 1]>>)
+    -> *mut ArcInner<HeaderSlice<HeaderWithLength<H>, [T]>>
+{
+    let len = unsafe { (*thin).data.header.length };
+    let fake_slice: *mut [T] = unsafe {
+        slice::from_raw_parts_mut(thin as *mut T, len)
+    };
+
+    fake_slice as *mut ArcInner<HeaderSlice<HeaderWithLength<H>, [T]>>
+}
+
+impl<H, T> ThinArc<H, T> {
+    /// Temporarily converts |self| into a bonafide Arc and exposes it to the
+    /// provided callback. The refcount is not modified.
+    #[inline(always)]
+    pub fn with_arc<F, U>(&self, f: F) -> U
+        where F: FnOnce(&Arc<HeaderSlice<HeaderWithLength<H>, [T]>>) -> U
+    {
+        // Synthesize transient Arc, which never touches the refcount of the ArcInner.
+        let transient = NoDrop::new(Arc {
+            ptr: thin_to_thick(self.ptr)
+        });
+
+        // Expose the transient Arc to the callback, which may clone it if it wants.
+        let result = f(&transient);
+
+        // Forget the transient Arc to leave the refcount untouched.
+        mem::forget(transient);
+
+        // Forward the result.
+        result
+    }
+}
+
+impl<H, T> Deref for ThinArc<H, T> {
+    type Target = HeaderSlice<HeaderWithLength<H>, [T]>;
+    fn deref(&self) -> &Self::Target {
+        unsafe { &(*thin_to_thick(self.ptr)).data }
+    }
+}
+
+impl<H, T> Clone for ThinArc<H, T> {
+    fn clone(&self) -> Self {
+        ThinArc::with_arc(self, |a| Arc::into_thin(a.clone()))
+    }
+}
+
+impl<H, T> Drop for ThinArc<H, T> {
+    fn drop(&mut self) {
+        let _ = Arc::from_thin(ThinArc { ptr: self.ptr });
+    }
+}
+
+impl<H, T> Arc<HeaderSlice<HeaderWithLength<H>, [T]>> {
+    /// Converts an Arc into a ThinArc. This consumes the Arc, so the refcount
+    /// is not modified.
+    pub fn into_thin(a: Self) -> ThinArc<H, T> {
+        assert!(a.header.length == a.slice.len(),
+                "Length needs to be correct for ThinArc to work");
+        let fat_ptr: *mut ArcInner<HeaderSlice<HeaderWithLength<H>, [T]>> = a.ptr;
+        mem::forget(a);
+        let thin_ptr = fat_ptr as *mut [usize] as *mut usize;
+        ThinArc {
+            ptr: thin_ptr as *mut ArcInner<HeaderSlice<HeaderWithLength<H>, [T; 1]>>
+        }
+    }
+
+    /// Converts a ThinArc into an Arc. This consumes the ThinArc, so the refcount
+    /// is not modified.
+    pub fn from_thin(a: ThinArc<H, T>) -> Self {
+        let ptr = thin_to_thick(a.ptr);
+        mem::forget(a);
+        Arc {
+            ptr: ptr
+        }
+    }
+}
+
+impl<H: PartialEq, T: PartialEq> PartialEq for ThinArc<H, T> {
+    fn eq(&self, other: &ThinArc<H, T>) -> bool {
+        ThinArc::with_arc(self, |a| {
+            ThinArc::with_arc(other, |b| {
+                *a == *b
+            })
+        })
+    }
+}
+
+impl<H: Eq, T: Eq> Eq for ThinArc<H, T> {}
+
+#[cfg(test)]
+mod tests {
+    use std::clone::Clone;
+    use std::ops::Drop;
+    use std::sync::atomic;
+    use std::sync::atomic::Ordering::{Acquire, SeqCst};
+    use super::{Arc, HeaderWithLength, ThinArc};
+
+    #[derive(PartialEq)]
+    struct Canary(*mut atomic::AtomicUsize);
+
+    impl Drop for Canary {
+        fn drop(&mut self) {
+            unsafe {
+                match *self {
+                    Canary(c) => {
+                        (*c).fetch_add(1, SeqCst);
+                    }
+                }
+            }
+        }
+    }
+
+    #[test]
+    fn slices_and_thin() {
+        let mut canary = atomic::AtomicUsize::new(0);
+        let c = Canary(&mut canary as *mut atomic::AtomicUsize);
+        let v = vec![5, 6];
+        let header = HeaderWithLength::new(c, v.len());
+        {
+            let x = Arc::into_thin(Arc::from_header_and_iter(header, v.into_iter()));
+            let y = ThinArc::with_arc(&x, |q| q.clone());
+            let _ = y.clone();
+            let _ = x == x;
+            Arc::from_thin(x.clone());
+        }
+        assert!(canary.load(Acquire) == 1);
+    }
+}

components/style/Cargo.toml

@@ -61,6 +61,7 @@ rayon = "0.7.1"
 selectors = { path = "../selectors" }
 serde = {version = "0.9", optional = true}
 serde_derive = {version = "0.9", optional = true}
+servo_arc = { path = "../servo_arc" }
 servo_atoms = {path = "../atoms", optional = true}
 servo_config = {path = "../config", optional = true}
 smallvec = "0.4"

components/style/gecko/selector_parser.rs

@@ -8,7 +8,7 @@ use cssparser::{Parser, ToCss};
 use element_state::ElementState;
 use gecko_bindings::structs::CSSPseudoClassType;
 use selector_parser::{SelectorParser, PseudoElementCascadeType};
-use selectors::parser::{ComplexSelector, SelectorMethods};
+use selectors::parser::{Selector, SelectorMethods};
 use selectors::visitor::SelectorVisitor;
 use std::borrow::Cow;
 use std::fmt;
@@ -47,7 +47,7 @@ macro_rules! pseudo_class_name {
             ///
             /// TODO(emilio): We disallow combinators and pseudos here, so we
             /// should use SimpleSelector instead
-            MozAny(Box<[ComplexSelector<SelectorImpl>]>),
+            MozAny(Box<[Selector<SelectorImpl>]>),
         }
     }
 }
@@ -273,7 +273,7 @@ impl<'a> ::selectors::Parser for SelectorParser<'a> {
                     }, )*
                     "-moz-any" => {
                         let selectors = parser.parse_comma_separated(|input| {
-                            ComplexSelector::parse(self, input)
+                            Selector::parse(self, input)
                         })?;
                         // Selectors inside `:-moz-any` may not include combinators.
                         if selectors.iter().flat_map(|x| x.iter_raw()).any(|s| s.is_combinator()) {

components/style/gecko/wrapper.rs

@@ -1405,7 +1405,7 @@ impl<'le> ::selectors::Element for GeckoElement<'le> {
             NonTSPseudoClass::MozPlaceholder => false,
             NonTSPseudoClass::MozAny(ref sels) => {
                 sels.iter().any(|s| {
-                    matches_complex_selector(s, self, context, flags_setter)
+                    matches_complex_selector(s, 0, self, context, flags_setter)
                 })
             }
             NonTSPseudoClass::MozSystemMetric(ref s) |

components/style/invalidation/stylesheets.rs

@@ -218,7 +218,7 @@ impl StylesheetInvalidationSet {
         let mut scope: Option<InvalidationScope> = None;
 
         let mut scan = true;
-        let mut iter = selector.inner.complex.iter();
+        let mut iter = selector.iter();
 
         loop {
             for component in &mut iter {
@@ -262,8 +262,8 @@ impl StylesheetInvalidationSet {
         match *rule {
             Style(ref lock) => {
                 let style_rule = lock.read_with(guard);
-                for selector in &style_rule.selectors.0 {
+                for selector_and_hashes in &style_rule.selectors.0 {
-                    self.collect_scopes(selector);
+                    self.collect_scopes(&selector_and_hashes.selector);
                     if self.fully_invalid {
                         return;
                     }

components/style/lib.rs

@@ -72,8 +72,8 @@ extern crate pdqsort;
 #[cfg(feature = "gecko")] extern crate precomputed_hash;
 extern crate rayon;
 extern crate selectors;
-#[cfg(feature = "servo")] extern crate serde;
 #[cfg(feature = "servo")] #[macro_use] extern crate serde_derive;
+pub extern crate servo_arc;
 #[cfg(feature = "servo")] #[macro_use] extern crate servo_atoms;
 #[cfg(feature = "servo")] extern crate servo_config;
 #[cfg(feature = "servo")] extern crate servo_url;
@@ -129,7 +129,6 @@ pub mod sequential;
 pub mod sink;
 pub mod str;
 pub mod style_adjuster;
-pub mod stylearc;
 pub mod stylesheet_set;
 pub mod stylesheets;
 pub mod thread_state;
@@ -139,6 +138,10 @@ pub mod traversal;
 #[allow(non_camel_case_types)]
 pub mod values;
 
+// Compat shim for the old name when it lived in the style crate.
+// FIXME(bholley) Remove this.
+pub use servo_arc as stylearc;
+
 use std::fmt;
 use style_traits::ToCss;
 

components/style/restyle_hints.rs

@@ -22,7 +22,8 @@ use selectors::Element;
 use selectors::attr::{AttrSelectorOperation, NamespaceConstraint};
 use selectors::matching::{ElementSelectorFlags, MatchingContext, MatchingMode};
 use selectors::matching::{RelevantLinkStatus, VisitedHandlingMode, matches_selector};
-use selectors::parser::{Combinator, Component, Selector, SelectorInner, SelectorMethods};
+use selectors::parser::{AncestorHashes, Combinator, Component};
+use selectors::parser::{Selector, SelectorAndHashes, SelectorIter, SelectorMethods};
 use selectors::visitor::SelectorVisitor;
 use smallvec::SmallVec;
 use std::cell::Cell;
@@ -642,7 +643,7 @@ impl<'a, E> Element for ElementWrapper<'a, E>
             use selectors::matching::matches_complex_selector;
             if let NonTSPseudoClass::MozAny(ref selectors) = *pseudo_class {
                 return selectors.iter().any(|s| {
-                    matches_complex_selector(s, self, context, _setter)
+                    matches_complex_selector(s, 0, self, context, _setter)
                 })
             }
         }
@@ -866,8 +867,15 @@ impl Sensitivities {
 #[derive(Clone, Debug)]
 #[cfg_attr(feature = "servo", derive(HeapSizeOf))]
 pub struct Dependency {
+    /// The dependency selector.
     #[cfg_attr(feature = "servo", ignore_heap_size_of = "Arc")]
-    selector: SelectorInner<SelectorImpl>,
+    selector: Selector<SelectorImpl>,
+    /// The offset into the selector that we should match on.
+    selector_offset: usize,
+    /// The ancestor hashes associated with the above selector at the given
+    /// offset.
+    #[cfg_attr(feature = "servo", ignore_heap_size_of = "No heap data")]
+    hashes: AncestorHashes,
     /// The hint associated with this dependency.
     pub hint: RestyleHint,
     /// The sensitivities associated with this dependency.
@@ -875,8 +883,12 @@ pub struct Dependency {
 }
 
 impl SelectorMapEntry for Dependency {
-    fn selector(&self) -> &SelectorInner<SelectorImpl> {
+    fn selector(&self) -> SelectorIter<SelectorImpl> {
-        &self.selector
+        self.selector.iter_from(self.selector_offset)
+    }
+
+    fn hashes(&self) -> &AncestorHashes {
+        &self.hashes
     }
 }
 
@@ -940,9 +952,9 @@ pub enum HintComputationContext<'a, E: 'a>
 
 impl DependencySet {
     /// Adds a selector to this `DependencySet`.
-    pub fn note_selector(&mut self, selector: &Selector<SelectorImpl>) {
+    pub fn note_selector(&mut self, selector_and_hashes: &SelectorAndHashes<SelectorImpl>) {
         let mut combinator = None;
-        let mut iter = selector.inner.complex.iter();
+        let mut iter = selector_and_hashes.selector.iter();
         let mut index = 0;
         let mut child_combinators_seen = 0;
         let mut saw_descendant_combinator = false;
@@ -992,17 +1004,21 @@ impl DependencySet {
                     None => RestyleHint::for_self(),
                 };
 
-                let dep_selector = if sequence_start == 0 {
+                // Reuse the bloom hashes if this is the base selector. Otherwise,
-                    // Reuse the bloom hashes if this is the base selector.
+                // rebuild them.
-                    selector.inner.clone()
+                let hashes = if sequence_start == 0 {
+                    selector_and_hashes.hashes.clone()
                 } else {
-                    SelectorInner::new(selector.inner.complex.slice_from(sequence_start))
+                    let seq_iter = selector_and_hashes.selector.iter_from(sequence_start);
+                    AncestorHashes::from_iter(seq_iter)
                 };
 
                 self.dependencies.insert(Dependency {
                     sensitivities: visitor.sensitivities,
                     hint: hint,
-                    selector: dep_selector,
+                    selector: selector_and_hashes.selector.clone(),
+                    selector_offset: sequence_start,
+                    hashes: hashes,
                 });
             }
 
@@ -1130,14 +1146,20 @@ impl DependencySet {
                 MatchingContext::new_for_visited(MatchingMode::Normal, None,
                                                  VisitedHandlingMode::AllLinksUnvisited);
             let matched_then =
-                matches_selector(&dep.selector, &snapshot_el,
+                matches_selector(&dep.selector,
+                                 dep.selector_offset,
+                                 &dep.hashes,
+                                 &snapshot_el,
                                  &mut then_context,
                                  &mut |_, _| {});
             let mut now_context =
                 MatchingContext::new_for_visited(MatchingMode::Normal, bloom_filter,
                                                  VisitedHandlingMode::AllLinksUnvisited);
             let matches_now =
-                matches_selector(&dep.selector, el,
+                matches_selector(&dep.selector,
+                                 dep.selector_offset,
+                                 &dep.hashes,
+                                 el,
                                  &mut now_context,
                                  &mut |_, _| {});
 
@@ -1162,12 +1184,18 @@ impl DependencySet {
                dep.sensitivities.states.intersects(IN_VISITED_OR_UNVISITED_STATE) {
                 then_context.visited_handling = VisitedHandlingMode::RelevantLinkVisited;
                 let matched_then =
-                    matches_selector(&dep.selector, &snapshot_el,
+                    matches_selector(&dep.selector,
+                                     dep.selector_offset,
+                                     &dep.hashes,
+                                     &snapshot_el,
                                      &mut then_context,
                                      &mut |_, _| {});
                 now_context.visited_handling = VisitedHandlingMode::RelevantLinkVisited;
                 let matches_now =
-                    matches_selector(&dep.selector, el,
+                    matches_selector(&dep.selector,
+                                     dep.selector_offset,
+                                     &dep.hashes,
+                                     el,
                                      &mut now_context,
                                      &mut |_, _| {});
                 if matched_then != matches_now {

components/style/selector_map.rs

@@ -12,7 +12,7 @@ use pdqsort::sort_by;
 use rule_tree::CascadeLevel;
 use selector_parser::SelectorImpl;
 use selectors::matching::{matches_selector, MatchingContext, ElementSelectorFlags};
-use selectors::parser::{Component, Combinator, SelectorInner};
+use selectors::parser::{AncestorHashes, Component, Combinator, SelectorAndHashes, SelectorIter};
 use selectors::parser::LocalName as LocalNameSelector;
 use smallvec::VecLike;
 use std::borrow::Borrow;
@@ -22,13 +22,20 @@ use stylist::{ApplicableDeclarationBlock, Rule};
 
 /// A trait to abstract over a given selector map entry.
 pub trait SelectorMapEntry : Sized + Clone {
-    /// Get the selector we should use to index in the selector map.
+    /// Gets the selector we should use to index in the selector map.
-    fn selector(&self) -> &SelectorInner<SelectorImpl>;
+    fn selector(&self) -> SelectorIter<SelectorImpl>;
+
+    /// Gets the ancestor hashes associated with the selector.
+    fn hashes(&self) -> &AncestorHashes;
 }
 
-impl SelectorMapEntry for SelectorInner<SelectorImpl> {
+impl SelectorMapEntry for SelectorAndHashes<SelectorImpl> {
-    fn selector(&self) -> &SelectorInner<SelectorImpl> {
+    fn selector(&self) -> SelectorIter<SelectorImpl> {
-        self
+        self.selector.iter()
+    }
+
+    fn hashes(&self) -> &AncestorHashes {
+        &self.hashes
     }
 }
 
@@ -224,7 +231,9 @@ impl SelectorMap<Rule> {
               F: FnMut(&E, ElementSelectorFlags),
     {
         for rule in rules {
-            if matches_selector(&rule.selector.inner,
+            if matches_selector(&rule.selector,
+                                0,
+                                &rule.hashes,
                                 element,
                                 context,
                                 flags_setter) {
@@ -390,12 +399,12 @@ impl<T: SelectorMapEntry> SelectorMap<T> {
 ///
 /// Effectively, pseudo-elements are ignored, given only state pseudo-classes
 /// may appear before them.
-fn find_from_right<F, R>(selector: &SelectorInner<SelectorImpl>,
+#[inline(always)]
+fn find_from_right<F, R>(mut iter: SelectorIter<SelectorImpl>,
                          mut f: F)
                          -> Option<R>
     where F: FnMut(&Component<SelectorImpl>) -> Option<R>,
 {
-    let mut iter = selector.complex.iter();
     for ss in &mut iter {
         if let Some(r) = f(ss) {
             return Some(r)
@@ -414,9 +423,10 @@ fn find_from_right<F, R>(selector: &SelectorInner<SelectorImpl>,
 }
 
 /// Retrieve the first ID name in the selector, or None otherwise.
-pub fn get_id_name(selector: &SelectorInner<SelectorImpl>)
+#[inline(always)]
-               -> Option<Atom> {
+pub fn get_id_name(iter: SelectorIter<SelectorImpl>)
-    find_from_right(selector, |ss| {
+                   -> Option<Atom> {
+    find_from_right(iter, |ss| {
         // TODO(pradeep): Implement case-sensitivity based on the
         // document type and quirks mode.
         if let Component::ID(ref id) = *ss {
@@ -427,9 +437,10 @@ pub fn get_id_name(selector: &SelectorInner<SelectorImpl>)
 }
 
 /// Retrieve the FIRST class name in the selector, or None otherwise.
-pub fn get_class_name(selector: &SelectorInner<SelectorImpl>)
+#[inline(always)]
-                  -> Option<Atom> {
+pub fn get_class_name(iter: SelectorIter<SelectorImpl>)
-    find_from_right(selector, |ss| {
+                      -> Option<Atom> {
+    find_from_right(iter, |ss| {
         // TODO(pradeep): Implement case-sensitivity based on the
         // document type and quirks mode.
         if let Component::Class(ref class) = *ss {
@@ -440,9 +451,10 @@ pub fn get_class_name(selector: &SelectorInner<SelectorImpl>)
 }
 
 /// Retrieve the name if it is a type selector, or None otherwise.
-pub fn get_local_name(selector: &SelectorInner<SelectorImpl>)
+#[inline(always)]
-                  -> Option<LocalNameSelector<SelectorImpl>> {
+pub fn get_local_name(iter: SelectorIter<SelectorImpl>)
-    find_from_right(selector, |ss| {
+                      -> Option<LocalNameSelector<SelectorImpl>> {
+    find_from_right(iter, |ss| {
         if let Component::LocalName(ref n) = *ss {
             return Some(LocalNameSelector {
                 name: n.name.clone(),

components/style/stylist.rs

@@ -28,7 +28,8 @@ use selectors::attr::NamespaceConstraint;
 use selectors::bloom::BloomFilter;
 use selectors::matching::{ElementSelectorFlags, matches_selector, MatchingContext, MatchingMode};
 use selectors::matching::AFFECTED_BY_PRESENTATIONAL_HINTS;
-use selectors::parser::{Combinator, Component, Selector, SelectorInner, SelectorIter, SelectorMethods};
+use selectors::parser::{AncestorHashes, Combinator, Component, Selector, SelectorAndHashes};
+use selectors::parser::{SelectorIter, SelectorMethods};
 use selectors::visitor::SelectorVisitor;
 use shared_lock::{Locked, SharedRwLockReadGuard, StylesheetGuards};
 use sink::Push;
@@ -159,7 +160,7 @@ pub struct Stylist {
     /// on state that is not otherwise visible to the cache, like attributes or
     /// tree-structural state like child index and pseudos).
     #[cfg_attr(feature = "servo", ignore_heap_size_of = "Arc")]
-    selectors_for_cache_revalidation: SelectorMap<SelectorInner<SelectorImpl>>,
+    selectors_for_cache_revalidation: SelectorMap<SelectorAndHashes<SelectorImpl>>,
 
     /// The total number of selectors.
     num_selectors: usize,
@@ -454,10 +455,10 @@ impl Stylist {
                 CssRule::Style(ref locked) => {
                     let style_rule = locked.read_with(&guard);
                     self.num_declarations += style_rule.block.read_with(&guard).len();
-                    for selector in &style_rule.selectors.0 {
+                    for selector_and_hashes in &style_rule.selectors.0 {
                         self.num_selectors += 1;
 
-                        let map = if let Some(pseudo) = selector.pseudo_element() {
+                        let map = if let Some(pseudo) = selector_and_hashes.selector.pseudo_element() {
                             self.pseudos_map
                                 .entry(pseudo.canonical())
                                 .or_insert_with(PerPseudoElementSelectorMap::new)
@@ -466,20 +467,21 @@ impl Stylist {
                             self.element_map.borrow_for_origin(&stylesheet.origin)
                         };
 
-                        map.insert(Rule::new(selector.clone(),
+                        map.insert(Rule::new(selector_and_hashes.selector.clone(),
+                                             selector_and_hashes.hashes.clone(),
                                              locked.clone(),
                                              self.rules_source_order));
 
-                        self.dependencies.note_selector(selector);
+                        self.dependencies.note_selector(selector_and_hashes);
-                        if needs_revalidation(selector) {
+                        if needs_revalidation(&selector_and_hashes.selector) {
-                            self.selectors_for_cache_revalidation.insert(selector.inner.clone());
+                            self.selectors_for_cache_revalidation.insert(selector_and_hashes.clone());
                         }
-                        selector.visit(&mut AttributeAndStateDependencyVisitor {
+                        selector_and_hashes.selector.visit(&mut AttributeAndStateDependencyVisitor {
                             attribute_dependencies: &mut self.attribute_dependencies,
                             style_attribute_dependency: &mut self.style_attribute_dependency,
                             state_dependencies: &mut self.state_dependencies,
                         });
-                        selector.visit(&mut MappedIdVisitor {
+                        selector_and_hashes.selector.visit(&mut MappedIdVisitor {
                             mapped_ids: &mut self.mapped_ids,
                         });
                     }
@@ -1130,8 +1132,10 @@ impl Stylist {
         // the lookups, which means that the bitvecs are comparable. We verify
         // this in the caller by asserting that the bitvecs are same-length.
         let mut results = BitVec::new();
-        self.selectors_for_cache_revalidation.lookup(*element, &mut |selector| {
+        self.selectors_for_cache_revalidation.lookup(*element, &mut |selector_and_hashes| {
-            results.push(matches_selector(selector,
+            results.push(matches_selector(&selector_and_hashes.selector,
+                                          0,
+                                          &selector_and_hashes.hashes,
                                           element,
                                           &mut matching_context,
                                           flags_setter));
@@ -1410,6 +1414,9 @@ pub struct Rule {
     /// can ruin performance when there are a lot of rules.
     #[cfg_attr(feature = "servo", ignore_heap_size_of = "Arc")]
     pub selector: Selector<SelectorImpl>,
+    /// The ancestor hashes associated with the selector.
+    #[cfg_attr(feature = "servo", ignore_heap_size_of = "No heap data")]
+    pub hashes: AncestorHashes,
     /// The actual style rule.
     #[cfg_attr(feature = "servo", ignore_heap_size_of = "Arc")]
     pub style_rule: Arc<Locked<StyleRule>>,
@@ -1418,8 +1425,12 @@ pub struct Rule {
 }
 
 impl SelectorMapEntry for Rule {
-    fn selector(&self) -> &SelectorInner<SelectorImpl> {
+    fn selector(&self) -> SelectorIter<SelectorImpl> {
-        &self.selector.inner
+        self.selector.iter()
+    }
+
+    fn hashes(&self) -> &AncestorHashes {
+        &self.hashes
     }
 }
 
@@ -1444,12 +1455,14 @@ impl Rule {
 
     /// Creates a new Rule.
     pub fn new(selector: Selector<SelectorImpl>,
+               hashes: AncestorHashes,
                style_rule: Arc<Locked<StyleRule>>,
                source_order: usize)
                -> Self
     {
         Rule {
             selector: selector,
+            hashes: hashes,
             style_rule: style_rule,
             source_order: source_order,
         }

tests/unit/style/stylesheets.rs

@@ -90,26 +90,23 @@ fn test_parse_stylesheet() {
                 },
             }))),
             CssRule::Style(Arc::new(stylesheet.shared_lock.wrap(StyleRule {
-                selectors: SelectorList(vec![
+                selectors: SelectorList::from_vec(vec!(
-                    Selector::new_for_unit_testing(
+                    Selector::from_vec(vec!(
-                        SelectorInner::from_vec(vec![
+                        Component::DefaultNamespace(NsAtom::from("http://www.w3.org/1999/xhtml")),
-                            Component::DefaultNamespace(NsAtom::from("http://www.w3.org/1999/xhtml")),
+                        Component::LocalName(LocalName {
-                            Component::LocalName(LocalName {
+                            name: local_name!("input"),
-                                name: local_name!("input"),
+                            lower_name: local_name!("input"),
-                                lower_name: local_name!("input"),
+                        }),
-                            }),
+                        Component::AttributeInNoNamespace {
-                            Component::AttributeInNoNamespace {
+                            local_name: local_name!("type"),
-                                local_name: local_name!("type"),
+                            local_name_lower: local_name!("type"),
-                                local_name_lower: local_name!("type"),
+                            operator: AttrSelectorOperator::Equal,
-                                operator: AttrSelectorOperator::Equal,
+                            value: "hidden".to_owned(),
-                                value: "hidden".to_owned(),
+                            case_sensitivity: ParsedCaseSensitivity::AsciiCaseInsensitive,
-                                case_sensitivity: ParsedCaseSensitivity::AsciiCaseInsensitive,
+                            never_matches: false,
-                                never_matches: false,
+                        }
-                            }
+                    ), (0 << 20) + (1 << 10) + (1 << 0))
-                        ]),
+                )),
-                        (0 << 20) + (1 << 10) + (1 << 0)
-                    ),
-                ]),
                 block: Arc::new(stylesheet.shared_lock.wrap(block_from(vec![
                     (PropertyDeclaration::Display(longhands::display::SpecifiedValue::none),
                      Importance::Important),
@@ -123,28 +120,23 @@ fn test_parse_stylesheet() {
                 },
             }))),
             CssRule::Style(Arc::new(stylesheet.shared_lock.wrap(StyleRule {
-                selectors: SelectorList(vec![
+                selectors: SelectorList::from_vec(vec!(
-                    Selector::new_for_unit_testing(
+                    Selector::from_vec(vec!(
-                        SelectorInner::from_vec(vec![
                             Component::DefaultNamespace(NsAtom::from("http://www.w3.org/1999/xhtml")),
                             Component::LocalName(LocalName {
                                 name: local_name!("html"),
                                 lower_name: local_name!("html"),
                             }),
-                        ]),
+                        ), (0 << 20) + (0 << 10) + (1 << 0)),
-                        (0 << 20) + (0 << 10) + (1 << 0)
+                    Selector::from_vec(vec!(
+                        Component::DefaultNamespace(NsAtom::from("http://www.w3.org/1999/xhtml")),
+                        Component::LocalName(LocalName {
+                            name: local_name!("body"),
+                            lower_name: local_name!("body"),
+                        })
+                        ), (0 << 20) + (0 << 10) + (1 << 0)
                     ),
-                    Selector::new_for_unit_testing(
+                )),
-                        SelectorInner::from_vec(vec![
-                            Component::DefaultNamespace(NsAtom::from("http://www.w3.org/1999/xhtml")),
-                            Component::LocalName(LocalName {
-                                name: local_name!("body"),
-                                lower_name: local_name!("body"),
-                            }),
-                        ]),
-                        (0 << 20) + (0 << 10) + (1 << 0)
-                    ),
-                ]),
                 block: Arc::new(stylesheet.shared_lock.wrap(block_from(vec![
                     (PropertyDeclaration::Display(longhands::display::SpecifiedValue::block),
                      Importance::Normal),
@@ -155,18 +147,15 @@ fn test_parse_stylesheet() {
                 },
             }))),
             CssRule::Style(Arc::new(stylesheet.shared_lock.wrap(StyleRule {
-                selectors: SelectorList(vec![
+                selectors: SelectorList::from_vec(vec!(
-                    Selector::new_for_unit_testing(
+                    Selector::from_vec(vec!(
-                        SelectorInner::from_vec(vec![
+                        Component::DefaultNamespace(NsAtom::from("http://www.w3.org/1999/xhtml")),
-                            Component::DefaultNamespace(NsAtom::from("http://www.w3.org/1999/xhtml")),
+                        Component::ID(Atom::from("d1")),
-                            Component::ID(Atom::from("d1")),
+                        Component::Combinator(Combinator::Child),
-                            Component::Combinator(Combinator::Child),
+                        Component::DefaultNamespace(NsAtom::from("http://www.w3.org/1999/xhtml")),
-                            Component::DefaultNamespace(NsAtom::from("http://www.w3.org/1999/xhtml")),
+                        Component::Class(Atom::from("ok"))
-                            Component::Class(Atom::from("ok")),
+                    ), (1 << 20) + (1 << 10) + (0 << 0))
-                        ]),
+                )),
-                        (1 << 20) + (1 << 10) + (0 << 0)
-                    ),
-                ]),
                 block: Arc::new(stylesheet.shared_lock.wrap(block_from(vec![
                     (PropertyDeclaration::BackgroundColor(
                         longhands::background_color::SpecifiedValue {

tests/unit/style/stylist.rs

@@ -45,7 +45,7 @@ fn get_mock_rules(css_selectors: &[&str]) -> (Vec<Vec<Rule>>, SharedRwLock) {
         let guard = shared_lock.read();
         let rule = locked.read_with(&guard);
         rule.selectors.0.iter().map(|s| {
-            Rule::new(s.clone(), locked.clone(), i)
+            Rule::new(s.selector.clone(), s.hashes.clone(), locked.clone(), i)
         }).collect()
     }).collect(), shared_lock)
 }
@@ -68,7 +68,7 @@ fn parse_selectors(selectors: &[&str]) -> Vec<Selector<SelectorImpl>> {
              .map(|x| SelectorParser::parse_author_origin_no_namespace(x).unwrap().0
                                                                          .into_iter()
                                                                          .nth(0)
-                                                                         .unwrap())
+                                                                         .unwrap().selector)
              .collect()
 }
 
@@ -126,7 +126,6 @@ fn test_revalidation_selectors() {
         "p:first-child span",
     ]).into_iter()
       .filter(|s| needs_revalidation(&s))
-      .map(|s| s.inner.complex)
       .collect::<Vec<_>>();
 
     let reference = parse_selectors(&[
@@ -167,7 +166,6 @@ fn test_revalidation_selectors() {
         // Selectors in the ancestor chain (needed for cousin sharing).
         "p:first-child span",
     ]).into_iter()
-      .map(|s| s.inner.complex)
       .collect::<Vec<_>>();
 
     assert_eq!(test.len(), reference.len());
@@ -189,22 +187,22 @@ fn test_rule_ordering_same_specificity() {
 #[test]
 fn test_get_id_name() {
     let (rules_list, _) = get_mock_rules(&[".intro", "#top"]);
-    assert_eq!(selector_map::get_id_name(&rules_list[0][0].selector.inner), None);
+    assert_eq!(selector_map::get_id_name(rules_list[0][0].selector.iter()), None);
-    assert_eq!(selector_map::get_id_name(&rules_list[1][0].selector.inner), Some(Atom::from("top")));
+    assert_eq!(selector_map::get_id_name(rules_list[1][0].selector.iter()), Some(Atom::from("top")));
 }
 
 #[test]
 fn test_get_class_name() {
     let (rules_list, _) = get_mock_rules(&[".intro.foo", "#top"]);
-    assert_eq!(selector_map::get_class_name(&rules_list[0][0].selector.inner), Some(Atom::from("foo")));
+    assert_eq!(selector_map::get_class_name(rules_list[0][0].selector.iter()), Some(Atom::from("foo")));
-    assert_eq!(selector_map::get_class_name(&rules_list[1][0].selector.inner), None);
+    assert_eq!(selector_map::get_class_name(rules_list[1][0].selector.iter()), None);
 }
 
 #[test]
 fn test_get_local_name() {
     let (rules_list, _) = get_mock_rules(&["img.foo", "#top", "IMG", "ImG"]);
     let check = |i: usize, names: Option<(&str, &str)>| {
-        assert!(selector_map::get_local_name(&rules_list[i][0].selector.inner)
+        assert!(selector_map::get_local_name(rules_list[i][0].selector.iter())
                 == names.map(|(name, lower_name)| LocalNameSelector {
                         name: LocalName::from(name),
                         lower_name: LocalName::from(lower_name) }))

tests/unit/stylo/size_of.rs

@@ -19,6 +19,11 @@ fn size_of_selectors_dummy_types() {
     assert_eq!(align_of::<dummies::Atom>(), align_of::<style::Atom>());
 }
 
+// The size of this is critical to performance on the bloom-basic microbenchmark.
+// When iterating over a large Rule array, we want to be able to fast-reject
+// selectors (with the inline hashes) with as few cache misses as possible.
+size_of_test!(test_size_of_rule, style::stylist::Rule, 40);
+
 size_of_test!(test_size_of_property_declaration, style::properties::PropertyDeclaration, 32);
 
 // This is huge, but we allocate it on the stack and then never move it,