auto merge of #17801 : Gankro/rust/collections-stuff, r=sfackler

I previously avoided `#[inline]`ing anything assuming someone would come in and explain to me where this would be appropriate. Apparently no one *really* knows, so I'll just go the opposite way an inline everything assuming someone will come in and yell at me that such-and-such shouldn't be `#[inline]`.

==================

For posterity, iteration comparisons:

```
test btree::map::bench::iter_20                            ... bench:       971 ns/iter (+/- 30)
test btree::map::bench::iter_1000                          ... bench:     29445 ns/iter (+/- 480)
test btree::map::bench::iter_100000                        ... bench:   2929035 ns/iter (+/- 21551)

test treemap::bench::iter_20                               ... bench:       530 ns/iter (+/- 66)
test treemap::bench::iter_1000                             ... bench:     26287 ns/iter (+/- 825)
test treemap::bench::iter_100000                           ... bench:   7650084 ns/iter (+/- 356711)

test trie::bench_map::iter_20                              ... bench:       646 ns/iter (+/- 265)
test trie::bench_map::iter_1000                            ... bench:     43556 ns/iter (+/- 5014)
test trie::bench_map::iter_100000                          ... bench:  12988002 ns/iter (+/- 139676)
```

As you can see `btree` "scales" much better than `treemap`. `triemap` scales quite poorly.

Note that *completely* different results are given if the elements are inserted in order from the range [0, size]. In particular, TrieMap *completely* dominates in the sorted case. This suggests adding benches for both might be worthwhile. However unsorted is *probably* the more "normal" case, so I consider this "good enough" for now.

Author: bors

src/libcollections/btree/map.rs

@@ -29,6 +29,47 @@ use ringbuf::RingBuf;
 
 
 /// A map based on a B-Tree.
+///
+/// B-Trees represent a fundamental compromise between cache-efficiency and actually minimizing
+/// the amount of work performed in a search. In theory, a binary search tree (BST) is the optimal
+/// choice for a sorted map, as a perfectly balanced BST performs the theoretical minimum amount of
+/// comparisons necessary to find an element (log<sub>2</sub>n). However, in practice the way this
+/// is done is *very* inefficient for modern computer architectures. In particular, every element
+/// is stored in its own individually heap-allocated node. This means that every single insertion
+/// triggers a heap-allocation, and every single comparison should be a cache-miss. Since these
+/// are both notably expensive things to do in practice, we are forced to at very least reconsider
+/// the BST strategy.
+///
+/// A B-Tree instead makes each node contain B-1 to 2B-1 elements in a contiguous array. By doing
+/// this, we reduce the number of allocations by a factor of B, and improve cache effeciency in
+/// searches. However, this does mean that searches will have to do *more* comparisons on average.
+/// The precise number of comparisons depends on the node search strategy used. For optimal cache
+/// effeciency, one could search the nodes linearly. For optimal comparisons, one could search
+/// the node using binary search. As a compromise, one could also perform a linear search
+/// that initially only checks every i<sup>th</sup> element for some choice of i.
+///
+/// Currently, our implementation simply performs naive linear search. This provides excellent
+/// performance on *small* nodes of elements which are cheap to compare. However in the future we
+/// would like to further explore choosing the optimal search strategy based on the choice of B,
+/// and possibly other factors. Using linear search, searching for a random element is expected
+/// to take O(B log<sub>B</sub>n) comparisons, which is generally worse than a BST. In practice,
+/// however, performance is excellent. `BTreeMap` is able to readily outperform `TreeMap` under
+/// many workloads, and is competetive where it doesn't. BTreeMap also generally *scales* better
+/// than TreeMap, making it more appropriate for large datasets.
+///
+/// However, `TreeMap` may still be more appropriate to use in many contexts. If elements are very
+/// large or expensive to compare, `TreeMap` may be more appropriate. It won't allocate any
+/// more space than is needed, and will perform the minimal number of comparisons necessary.
+/// `TreeMap` also provides much better performance stability guarantees. Generally, very few
+/// changes need to be made to update a BST, and two updates are expected to take about the same
+/// amount of time on roughly equal sized BSTs. However a B-Tree's performance is much more
+/// amortized. If a node is overfull, it must be split into two nodes. If a node is underfull, it
+/// may be merged with another. Both of these operations are relatively expensive to perform, and
+/// it's possible to force one to occur at every single level of the tree in a single insertion or
+/// deletion. In fact, a malicious or otherwise unlucky sequence of insertions and deletions can
+/// force this degenerate behaviour to occur on every operation. While the total amount of work
+/// done on each operation isn't *catastrophic*, and *is* still bounded by O(B log<sub>B</sub>n),
+/// it is certainly much slower when it does.
 #[deriving(Clone)]
 pub struct BTreeMap<K, V> {
     root: Node<K, V>,
@@ -93,6 +134,8 @@ impl<K: Ord, V> BTreeMap<K, V> {
     }
 
     /// Makes a new empty BTreeMap with the given B.
+    ///
+    /// B cannot be less than 2.
     pub fn with_b(b: uint) -> BTreeMap<K, V> {
         assert!(b > 1, "B must be greater than 1");
         BTreeMap {
@@ -1145,9 +1188,12 @@ mod test {
 
 #[cfg(test)]
 mod bench {
-    use test::Bencher;
+    use std::prelude::*;
+    use std::rand::{weak_rng, Rng};
+    use test::{Bencher, black_box};
 
     use super::BTreeMap;
+    use MutableMap;
     use deque::bench::{insert_rand_n, insert_seq_n, find_rand_n, find_seq_n};
 
     #[bench]
@@ -1200,4 +1246,34 @@ mod bench {
         let mut m : BTreeMap<uint,uint> = BTreeMap::new();
         find_seq_n(10_000, &mut m, b);
     }
+
+    fn bench_iter(b: &mut Bencher, size: uint) {
+        let mut map = BTreeMap::<uint, uint>::new();
+        let mut rng = weak_rng();
+
+        for _ in range(0, size) {
+            map.swap(rng.gen(), rng.gen());
+        }
+
+        b.iter(|| {
+            for entry in map.iter() {
+                black_box(entry);
+            }
+        });
+    }
+
+    #[bench]
+    pub fn iter_20(b: &mut Bencher) {
+        bench_iter(b, 20);
+    }
+
+    #[bench]
+    pub fn iter_1000(b: &mut Bencher) {
+        bench_iter(b, 1000);
+    }
+
+    #[bench]
+    pub fn iter_100000(b: &mut Bencher) {
+        bench_iter(b, 100000);
+    }
 }

src/libcollections/btree/set.rs

@@ -23,6 +23,9 @@ use core::fmt::Show;
 use {Mutable, Set, MutableSet, MutableMap, Map};
 
 /// A set based on a B-Tree.
+///
+/// See BTreeMap's documentation for a detailed discussion of this collection's performance
+/// benefits and drawbacks.
 #[deriving(Clone, Hash, PartialEq, Eq, Ord, PartialOrd)]
 pub struct BTreeSet<T>{
     map: BTreeMap<T, ()>,
@@ -65,6 +68,8 @@ impl<T: Ord> BTreeSet<T> {
     }
 
     /// Makes a new BTreeSet with the given B.
+    ///
+    /// B cannot be less than 2.
     pub fn with_b(b: uint) -> BTreeSet<T> {
         BTreeSet { map: BTreeMap::with_b(b) }
     }

src/libcollections/treemap.rs

@@ -2232,9 +2232,12 @@ mod test_treemap {
 
 #[cfg(test)]
 mod bench {
-    use test::Bencher;
+    use std::prelude::*;
+    use std::rand::{weak_rng, Rng};
+    use test::{Bencher, black_box};
 
     use super::TreeMap;
+    use MutableMap;
     use deque::bench::{insert_rand_n, insert_seq_n, find_rand_n, find_seq_n};
 
     // Find seq
@@ -2288,6 +2291,36 @@ mod bench {
         let mut m : TreeMap<uint,uint> = TreeMap::new();
         find_seq_n(10_000, &mut m, b);
     }
+
+    fn bench_iter(b: &mut Bencher, size: uint) {
+        let mut map = TreeMap::<uint, uint>::new();
+        let mut rng = weak_rng();
+
+        for _ in range(0, size) {
+            map.swap(rng.gen(), rng.gen());
+        }
+
+        b.iter(|| {
+            for entry in map.iter() {
+                black_box(entry);
+            }
+        });
+    }
+
+    #[bench]
+    pub fn iter_20(b: &mut Bencher) {
+        bench_iter(b, 20);
+    }
+
+    #[bench]
+    pub fn iter_1000(b: &mut Bencher) {
+        bench_iter(b, 1000);
+    }
+
+    #[bench]
+    pub fn iter_100000(b: &mut Bencher) {
+        bench_iter(b, 100000);
+    }
 }
 
 #[cfg(test)]

src/libcollections/trie.rs

@@ -949,8 +949,8 @@ macro_rules! iterator_impl {
                 // rules, and are just manipulating raw pointers like there's no
                 // such thing as invalid pointers and memory unsafety. The
                 // reason is performance, without doing this we can get the
-                // bench_iter_large microbenchmark down to about 30000 ns/iter
-                // (using .unsafe_get to index self.stack directly, 38000
+                // (now replaced) bench_iter_large microbenchmark down to about
+                // 30000 ns/iter (using .unsafe_get to index self.stack directly, 38000
                 // ns/iter with [] checked indexing), but this smashes that down
                 // to 13500 ns/iter.
                 //
@@ -1459,31 +1459,39 @@ mod test_map {
 mod bench_map {
     use std::prelude::*;
     use std::rand::{weak_rng, Rng};
-    use test::Bencher;
+    use test::{Bencher, black_box};
 
     use MutableMap;
     use super::TrieMap;
 
-    #[bench]
-    fn bench_iter_small(b: &mut Bencher) {
-        let mut m = TrieMap::<uint>::new();
+    fn bench_iter(b: &mut Bencher, size: uint) {
+        let mut map = TrieMap::<uint>::new();
         let mut rng = weak_rng();
-        for _ in range(0u, 20) {
-            m.insert(rng.gen(), rng.gen());
+
+        for _ in range(0, size) {
+            map.swap(rng.gen(), rng.gen());
         }
 
-        b.iter(|| for _ in m.iter() {})
+        b.iter(|| {
+            for entry in map.iter() {
+                black_box(entry);
+            }
+        });
     }
 
     #[bench]
-    fn bench_iter_large(b: &mut Bencher) {
-        let mut m = TrieMap::<uint>::new();
-        let mut rng = weak_rng();
-        for _ in range(0u, 1000) {
-            m.insert(rng.gen(), rng.gen());
-        }
+    pub fn iter_20(b: &mut Bencher) {
+        bench_iter(b, 20);
+    }
 
-        b.iter(|| for _ in m.iter() {})
+    #[bench]
+    pub fn iter_1000(b: &mut Bencher) {
+        bench_iter(b, 1000);
+    }
+
+    #[bench]
+    pub fn iter_100000(b: &mut Bencher) {
+        bench_iter(b, 100000);
     }
 
     #[bench]