1 // Copyright 2012-2017 The Rust Project Developers. See the COPYRIGHT
2 // file at the top-level directory of this distribution and at
3 // http://rust-lang.org/COPYRIGHT.
5 // Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
6 // http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
7 // <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
8 // option. This file may not be copied, modified, or distributed
9 // except according to those terms.
15 /// Rotation is much faster if it has access to a little bit of memory. This
16 /// union provides a RawVec-like interface, but to a fixed-size stack buffer.
17 #[allow(unions_with_drop_fields)]
19 /// Ensure this is appropriately aligned for T, and is big
20 /// enough for two elements even if T is enormous.
22 /// For normally-sized types, especially things like u8, having more
23 /// than 2 in the buffer is necessary for usefulness, so pad it out
24 /// enough to be helpful, but not so big as to risk overflow.
30 unsafe { mem::uninitialized() }
32 fn ptr(&self) -> *mut T {
33 unsafe { &self.typed as *const T as *mut T }
36 if mem::size_of::<T>() == 0 {
39 mem::size_of::<Self>() / mem::size_of::<T>()
44 /// Rotates the range `[mid-left, mid+right)` such that the element at `mid`
45 /// becomes the first element. Equivalently, rotates the range `left`
46 /// elements to the left or `right` elements to the right.
50 /// The specified range must be valid for reading and writing.
51 /// The type `T` must have non-zero size.
55 /// For longer rotations, swap the left-most `delta = min(left, right)`
56 /// elements with the right-most `delta` elements. LLVM vectorizes this,
57 /// which is profitable as we only reach this step for a "large enough"
58 /// rotation. Doing this puts `delta` elements on the larger side into the
59 /// correct position, leaving a smaller rotate problem. Demonstration:
62 /// [ 6 7 8 9 10 11 12 13 . 1 2 3 4 5 ]
63 /// 1 2 3 4 5 [ 11 12 13 . 6 7 8 9 10 ]
64 /// 1 2 3 4 5 [ 8 9 10 . 6 7 ] 11 12 13
65 /// 1 2 3 4 5 6 7 [ 10 . 8 9 ] 11 12 13
66 /// 1 2 3 4 5 6 7 [ 9 . 8 ] 10 11 12 13
67 /// 1 2 3 4 5 6 7 8 [ . ] 9 10 11 12 13
70 /// Once the rotation is small enough, copy some elements into a stack
71 /// buffer, `memmove` the others, and move the ones back from the buffer.
72 pub unsafe fn ptr_rotate<T>(mut left: usize, mid: *mut T, mut right: usize) {
74 let delta = cmp::min(left, right);
75 if delta <= RawArray::<T>::cap() {
80 mid.offset(-(left as isize)),
81 mid.offset((right-delta) as isize),
91 let rawarray = RawArray::new();
92 let buf = rawarray.ptr();
94 let dim = mid.offset(-(left as isize)).offset(right as isize);
96 ptr::copy_nonoverlapping(mid.offset(-(left as isize)), buf, left);
97 ptr::copy(mid, mid.offset(-(left as isize)), right);
98 ptr::copy_nonoverlapping(buf, dim, left);
101 ptr::copy_nonoverlapping(mid, buf, right);
102 ptr::copy(mid.offset(-(left as isize)), dim, left);
103 ptr::copy_nonoverlapping(buf, mid.offset(-(left as isize)), right);
107 unsafe fn ptr_swap_u8(a: *mut u8, b: *mut u8, n: usize) {
109 ptr::swap(a.offset(i as isize), b.offset(i as isize));
112 unsafe fn ptr_swap_u16(a: *mut u16, b: *mut u16, n: usize) {
114 ptr::swap(a.offset(i as isize), b.offset(i as isize));
117 unsafe fn ptr_swap_u32(a: *mut u32, b: *mut u32, n: usize) {
119 ptr::swap(a.offset(i as isize), b.offset(i as isize));
122 unsafe fn ptr_swap_u64(a: *mut u64, b: *mut u64, n: usize) {
124 ptr::swap(a.offset(i as isize), b.offset(i as isize));
128 unsafe fn ptr_swap_n<T>(a: *mut T, b: *mut T, n: usize) {
129 // Doing this as a generic is 16% & 40% slower in two of the `String`
130 // benchmarks, as (based on the block names) LLVM doesn't vectorize it.
131 // Since this is just operating on raw memory, dispatch to a version
132 // with appropriate alignment. Helps with code size as well, by
133 // avoiding monomorphizing different unrolled loops for `i32`,
134 // `u32`, `f32`, `[u32; 1]`, etc.
135 let size_of_t = mem::size_of::<T>();
136 let align_of_t = mem::align_of::<T>();
138 let a64 = mem::align_of::<u64>();
139 if a64 == 8 && align_of_t % a64 == 0 {
140 return ptr_swap_u64(a as *mut u64, b as *mut u64, n * (size_of_t / 8));
143 let a32 = mem::align_of::<u32>();
144 if a32 == 4 && align_of_t % a32 == 0 {
145 return ptr_swap_u32(a as *mut u32, b as *mut u32, n * (size_of_t / 4));
148 let a16 = mem::align_of::<u16>();
149 if a16 == 2 && align_of_t % a16 == 0 {
150 return ptr_swap_u16(a as *mut u16, b as *mut u16, n * (size_of_t / 2));
153 ptr_swap_u8(a as *mut u8, b as *mut u8, n * size_of_t);