Make [u8]::reverse() 5x faster

Since LLVM doesn't vectorize the loop for us, do unaligned reads
of a larger type and use LLVM's bswap intrinsic to do the
reversing of the actual bytes.  cfg!-restricted to x86 and
x86_64, as I assume it wouldn't help on things like ARMv5.

Also makes [u16]::reverse() a more modest 1.5x faster by
loading/storing u32 and swapping the u16s with ROT16.

Thank you ptr::*_unaligned for making this easy :)

src/libcollections/benches/slice.rs

@@ -290,3 +290,24 @@ sort!(sort_unstable, sort_unstable_large_random, gen_random, 10000);
 sort!(sort_unstable, sort_unstable_large_big_random, gen_big_random, 10000);
 sort!(sort_unstable, sort_unstable_large_strings, gen_strings, 10000);
 sort_expensive!(sort_unstable_by, sort_unstable_large_random_expensive, gen_random, 10000);
+
+macro_rules! reverse {
+    ($name:ident, $ty:ident) => {
+        #[bench]
+        fn $name(b: &mut Bencher) {
+            // odd length and offset by 1 to be as unaligned as possible
+            let n = 0xFFFFF;
+            let mut v: Vec<_> =
+                (0..1+(n / mem::size_of::<$ty>() as u64))
+                .map(|x| x as $ty)
+                .collect();
+            b.iter(|| black_box(&mut v[1..]).reverse());
+            b.bytes = n;
+        }
+    }
+}
+
+reverse!(reverse_u8, u8);
+reverse!(reverse_u16, u16);
+reverse!(reverse_u32, u32);
+reverse!(reverse_u64, u64);

src/libcollections/tests/slice.rs

@@ -379,6 +379,16 @@ fn test_reverse() {
     let mut v3 = Vec::<i32>::new();
     v3.reverse();
     assert!(v3.is_empty());
+
+    // check the 1-byte-types path
+    let mut v = (-50..51i8).collect::<Vec<_>>();
+    v.reverse();
+    assert_eq!(v, (-50..51i8).rev().collect::<Vec<_>>());
+
+    // check the 2-byte-types path
+    let mut v = (-50..51i16).collect::<Vec<_>>();
+    v.reverse();
+    assert_eq!(v, (-50..51i16).rev().collect::<Vec<_>>());
 }
 
 #[test]

src/libcore/slice/mod.rs

@@ -539,6 +539,44 @@ impl<T> SliceExt for [T] {
     fn reverse(&mut self) {
         let mut i: usize = 0;
         let ln = self.len();
+
+        let fast_unaligned =
+            cfg!(any(target_arch = "x86", target_arch = "x86_64"));
+
+        if fast_unaligned && mem::size_of::<T>() == 1 {
+            // Single-byte read & write are comparatively slow. Instead,
+            // work in usize chunks and get bswap to do the hard work.
+            let chunk = mem::size_of::<usize>();
+            while i + chunk - 1 < ln / 2 {
+                unsafe {
+                    let pa: *mut T = self.get_unchecked_mut(i);
+                    let pb: *mut T = self.get_unchecked_mut(ln - i - chunk);
+                    let va = ptr::read_unaligned(pa as *mut usize);
+                    let vb = ptr::read_unaligned(pb as *mut usize);
+                    ptr::write_unaligned(pa as *mut usize, vb.swap_bytes());
+                    ptr::write_unaligned(pb as *mut usize, va.swap_bytes());
+                }
+                i += chunk;
+            }
+        }
+
+        if fast_unaligned && mem::size_of::<T>() == 2 {
+            // Not quite as good as the above, but still helpful.
+            // Same general idea, read bigger and do the swap in a register.
+            let chunk = mem::size_of::<u32>() / 2;
+            while i + chunk - 1 < ln / 2 {
+                unsafe {
+                    let pa: *mut T = self.get_unchecked_mut(i);
+                    let pb: *mut T = self.get_unchecked_mut(ln - i - chunk);
+                    let va = ptr::read_unaligned(pa as *mut u32);
+                    let vb = ptr::read_unaligned(pb as *mut u32);
+                    ptr::write_unaligned(pa as *mut u32, vb.rotate_left(16));
+                    ptr::write_unaligned(pb as *mut u32, va.rotate_left(16));
+                }
+                i += chunk;
+            }
+        }
+
         while i < ln / 2 {
             // Unsafe swap to avoid the bounds check in safe swap.
             unsafe {