1 //! # FastCpy
2 //!
3 //! The Rust Compiler calls `memcpy` for slices of unknown length.
4 //! This crate provides a faster implementation of `memcpy` for slices up to 32bytes (64bytes with `avx`).
5 //! If you know most of you copy operations are not too big you can use `fastcpy` to speed up your program.
6 //!
7 //! `fastcpy` is designed to contain not too much assembly, so the overhead is low.
8 //!
9 //! As fall back the standard `memcpy` is called
10 //!
11 //! ## Double Copy Trick
12 //! `fastcpy` employs a double copy trick to copy slices of length 4-32bytes (64bytes with `avx`).
13 //! E.g. Slice of length 6 can be copied with two uncoditional copy operations.
14 //!
15 //! /// [1, 2, 3, 4, 5, 6]
16 //! /// [1, 2, 3, 4]
17 //! ///       [3, 4, 5, 6]
18 //!
19 
20 #[inline]
slice_copy(src: *const u8, dst: *mut u8, num_bytes: usize)21 pub fn slice_copy(src: *const u8, dst: *mut u8, num_bytes: usize) {
22     if num_bytes < 4 {
23         short_copy(src, dst, num_bytes);
24         return;
25     }
26 
27     if num_bytes < 8 {
28         double_copy_trick::<4>(src, dst, num_bytes);
29         return;
30     }
31 
32     if num_bytes <= 16 {
33         double_copy_trick::<8>(src, dst, num_bytes);
34         return;
35     }
36 
37     //if num_bytes <= 32 {
38     //double_copy_trick::<16>(src, dst, num_bytes);
39     //return;
40     //}
41 
42     // /// The code will use the vmovdqu instruction to copy 32 bytes at a time.
43     //#[cfg(target_feature = "avx")]
44     //{
45     //if num_bytes <= 64 {
46     //double_copy_trick::<32>(src, dst, num_bytes);
47     //return;
48     //}
49     //}
50 
51     // For larger sizes we use the default, which calls memcpy
52     // memcpy does some virtual memory tricks to copy large chunks of memory.
53     //
54     // The theory should be that the checks above don't cost much relative to the copy call for
55     // larger copies.
56     // The bounds checks in `copy_from_slice` are elided.
57 
58     //unsafe { core::ptr::copy_nonoverlapping(src, dst, num_bytes) }
59     wild_copy_from_src::<16>(src, dst, num_bytes)
60 }
61 
62 // Inline never because otherwise we get a call to memcpy -.-
63 #[inline]
wild_copy_from_src<const SIZE: usize>( mut source: *const u8, mut dst: *mut u8, num_bytes: usize, )64 fn wild_copy_from_src<const SIZE: usize>(
65     mut source: *const u8,
66     mut dst: *mut u8,
67     num_bytes: usize,
68 ) {
69     // Note: if the compiler auto-vectorizes this it'll hurt performance!
70     // It's not the case for 16 bytes stepsize, but for 8 bytes.
71     let l_last = unsafe { source.add(num_bytes - SIZE) };
72     let r_last = unsafe { dst.add(num_bytes - SIZE) };
73     let num_bytes = (num_bytes / SIZE) * SIZE;
74 
75     unsafe {
76         let dst_ptr_end = dst.add(num_bytes);
77         loop {
78             core::ptr::copy_nonoverlapping(source, dst, SIZE);
79             source = source.add(SIZE);
80             dst = dst.add(SIZE);
81             if dst >= dst_ptr_end {
82                 break;
83             }
84         }
85     }
86 
87     unsafe {
88         core::ptr::copy_nonoverlapping(l_last, r_last, SIZE);
89     }
90 }
91 
92 #[inline]
short_copy(src: *const u8, dst: *mut u8, len: usize)93 fn short_copy(src: *const u8, dst: *mut u8, len: usize) {
94     unsafe {
95         *dst = *src;
96     }
97     if len >= 2 {
98         double_copy_trick::<2>(src, dst, len);
99     }
100 }
101 
102 #[inline(always)]
103 /// [1, 2, 3, 4, 5, 6]
104 /// [1, 2, 3, 4]
105 ///       [3, 4, 5, 6]
double_copy_trick<const SIZE: usize>(src: *const u8, dst: *mut u8, len: usize)106 fn double_copy_trick<const SIZE: usize>(src: *const u8, dst: *mut u8, len: usize) {
107     let l_end = unsafe { src.add(len - SIZE) };
108     let r_end = unsafe { dst.add(len - SIZE) };
109 
110     unsafe {
111         core::ptr::copy_nonoverlapping(src, dst, SIZE);
112         core::ptr::copy_nonoverlapping(l_end, r_end, SIZE);
113     }
114 }
115 
116 #[cfg(test)]
117 mod tests {
118     use super::slice_copy;
119     use alloc::vec::Vec;
120     use proptest::prelude::*;
121     proptest! {
122         #[test]
123         fn test_fast_short_slice_copy(left: Vec<u8>) {
124             if left.is_empty() {
125                 return Ok(());
126             }
127             let mut right = vec![0u8; left.len()];
128             slice_copy(left.as_ptr(), right.as_mut_ptr(), left.len());
129             prop_assert_eq!(&left, &right);
130         }
131     }
132 
133     #[test]
test_fast_short_slice_copy_edge_cases()134     fn test_fast_short_slice_copy_edge_cases() {
135         for len in 1..(512 * 2) {
136             let left = (0..len).map(|i| i as u8).collect::<Vec<_>>();
137             let mut right = vec![0u8; len];
138             slice_copy(left.as_ptr(), right.as_mut_ptr(), left.len());
139             assert_eq!(left, right);
140         }
141     }
142 
143     #[test]
test_fail2()144     fn test_fail2() {
145         let left = vec![
146             0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23,
147             24, 25, 26, 27, 28, 29, 30, 31, 32,
148         ];
149         let mut right = vec![0u8; left.len()];
150         slice_copy(left.as_ptr(), right.as_mut_ptr(), left.len());
151         assert_eq!(left, right);
152     }
153 
154     #[test]
test_fail()155     fn test_fail() {
156         let left = vec![
157             0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
158             0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
159             0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
160         ];
161         let mut right = vec![0u8; left.len()];
162         slice_copy(left.as_ptr(), right.as_mut_ptr(), left.len());
163         assert_eq!(left, right);
164     }
165 }
166