1*e4df2d5eSMatteo Croce // SPDX-License-Identifier: GPL-2.0-only
2*e4df2d5eSMatteo Croce /*
3*e4df2d5eSMatteo Croce * String functions optimized for hardware which doesn't
4*e4df2d5eSMatteo Croce * handle unaligned memory accesses efficiently.
5*e4df2d5eSMatteo Croce *
6*e4df2d5eSMatteo Croce * Copyright (C) 2021 Matteo Croce
7*e4df2d5eSMatteo Croce */
8*e4df2d5eSMatteo Croce
9*e4df2d5eSMatteo Croce #include <linux/types.h>
10*e4df2d5eSMatteo Croce #include <linux/module.h>
11*e4df2d5eSMatteo Croce
12*e4df2d5eSMatteo Croce /* Minimum size for a word copy to be convenient */
13*e4df2d5eSMatteo Croce #define BYTES_LONG sizeof(long)
14*e4df2d5eSMatteo Croce #define WORD_MASK (BYTES_LONG - 1)
15*e4df2d5eSMatteo Croce #define MIN_THRESHOLD (BYTES_LONG * 2)
16*e4df2d5eSMatteo Croce
17*e4df2d5eSMatteo Croce /* convenience union to avoid cast between different pointer types */
18*e4df2d5eSMatteo Croce union types {
19*e4df2d5eSMatteo Croce u8 *as_u8;
20*e4df2d5eSMatteo Croce unsigned long *as_ulong;
21*e4df2d5eSMatteo Croce uintptr_t as_uptr;
22*e4df2d5eSMatteo Croce };
23*e4df2d5eSMatteo Croce
24*e4df2d5eSMatteo Croce union const_types {
25*e4df2d5eSMatteo Croce const u8 *as_u8;
26*e4df2d5eSMatteo Croce unsigned long *as_ulong;
27*e4df2d5eSMatteo Croce uintptr_t as_uptr;
28*e4df2d5eSMatteo Croce };
29*e4df2d5eSMatteo Croce
memcpy(void * dest,const void * src,size_t count)30*e4df2d5eSMatteo Croce void *memcpy(void *dest, const void *src, size_t count)
31*e4df2d5eSMatteo Croce {
32*e4df2d5eSMatteo Croce union const_types s = { .as_u8 = src };
33*e4df2d5eSMatteo Croce union types d = { .as_u8 = dest };
34*e4df2d5eSMatteo Croce int distance = 0;
35*e4df2d5eSMatteo Croce
36*e4df2d5eSMatteo Croce if (count < MIN_THRESHOLD)
37*e4df2d5eSMatteo Croce goto copy_remainder;
38*e4df2d5eSMatteo Croce
39*e4df2d5eSMatteo Croce /* Copy a byte at time until destination is aligned. */
40*e4df2d5eSMatteo Croce for (; d.as_uptr & WORD_MASK; count--)
41*e4df2d5eSMatteo Croce *d.as_u8++ = *s.as_u8++;
42*e4df2d5eSMatteo Croce
43*e4df2d5eSMatteo Croce distance = s.as_uptr & WORD_MASK;
44*e4df2d5eSMatteo Croce
45*e4df2d5eSMatteo Croce if (distance) {
46*e4df2d5eSMatteo Croce unsigned long last, next;
47*e4df2d5eSMatteo Croce
48*e4df2d5eSMatteo Croce /*
49*e4df2d5eSMatteo Croce * s is distance bytes ahead of d, and d just reached
50*e4df2d5eSMatteo Croce * the alignment boundary. Move s backward to word align it
51*e4df2d5eSMatteo Croce * and shift data to compensate for distance, in order to do
52*e4df2d5eSMatteo Croce * word-by-word copy.
53*e4df2d5eSMatteo Croce */
54*e4df2d5eSMatteo Croce s.as_u8 -= distance;
55*e4df2d5eSMatteo Croce
56*e4df2d5eSMatteo Croce next = s.as_ulong[0];
57*e4df2d5eSMatteo Croce for (; count >= BYTES_LONG; count -= BYTES_LONG) {
58*e4df2d5eSMatteo Croce last = next;
59*e4df2d5eSMatteo Croce next = s.as_ulong[1];
60*e4df2d5eSMatteo Croce
61*e4df2d5eSMatteo Croce d.as_ulong[0] = last >> (distance * 8) |
62*e4df2d5eSMatteo Croce next << ((BYTES_LONG - distance) * 8);
63*e4df2d5eSMatteo Croce
64*e4df2d5eSMatteo Croce d.as_ulong++;
65*e4df2d5eSMatteo Croce s.as_ulong++;
66*e4df2d5eSMatteo Croce }
67*e4df2d5eSMatteo Croce
68*e4df2d5eSMatteo Croce /* Restore s with the original offset. */
69*e4df2d5eSMatteo Croce s.as_u8 += distance;
70*e4df2d5eSMatteo Croce } else {
71*e4df2d5eSMatteo Croce /*
72*e4df2d5eSMatteo Croce * If the source and dest lower bits are the same, do a simple
73*e4df2d5eSMatteo Croce * 32/64 bit wide copy.
74*e4df2d5eSMatteo Croce */
75*e4df2d5eSMatteo Croce for (; count >= BYTES_LONG; count -= BYTES_LONG)
76*e4df2d5eSMatteo Croce *d.as_ulong++ = *s.as_ulong++;
77*e4df2d5eSMatteo Croce }
78*e4df2d5eSMatteo Croce
79*e4df2d5eSMatteo Croce copy_remainder:
80*e4df2d5eSMatteo Croce while (count--)
81*e4df2d5eSMatteo Croce *d.as_u8++ = *s.as_u8++;
82*e4df2d5eSMatteo Croce
83*e4df2d5eSMatteo Croce return dest;
84*e4df2d5eSMatteo Croce }
85*e4df2d5eSMatteo Croce EXPORT_SYMBOL(memcpy);
86*e4df2d5eSMatteo Croce
87*e4df2d5eSMatteo Croce /*
88*e4df2d5eSMatteo Croce * Simply check if the buffer overlaps an call memcpy() in case,
89*e4df2d5eSMatteo Croce * otherwise do a simple one byte at time backward copy.
90*e4df2d5eSMatteo Croce */
memmove(void * dest,const void * src,size_t count)91*e4df2d5eSMatteo Croce void *memmove(void *dest, const void *src, size_t count)
92*e4df2d5eSMatteo Croce {
93*e4df2d5eSMatteo Croce if (dest < src || src + count <= dest)
94*e4df2d5eSMatteo Croce return memcpy(dest, src, count);
95*e4df2d5eSMatteo Croce
96*e4df2d5eSMatteo Croce if (dest > src) {
97*e4df2d5eSMatteo Croce const char *s = src + count;
98*e4df2d5eSMatteo Croce char *tmp = dest + count;
99*e4df2d5eSMatteo Croce
100*e4df2d5eSMatteo Croce while (count--)
101*e4df2d5eSMatteo Croce *--tmp = *--s;
102*e4df2d5eSMatteo Croce }
103*e4df2d5eSMatteo Croce return dest;
104*e4df2d5eSMatteo Croce }
105*e4df2d5eSMatteo Croce EXPORT_SYMBOL(memmove);
106*e4df2d5eSMatteo Croce
memset(void * s,int c,size_t count)107*e4df2d5eSMatteo Croce void *memset(void *s, int c, size_t count)
108*e4df2d5eSMatteo Croce {
109*e4df2d5eSMatteo Croce union types dest = { .as_u8 = s };
110*e4df2d5eSMatteo Croce
111*e4df2d5eSMatteo Croce if (count >= MIN_THRESHOLD) {
112*e4df2d5eSMatteo Croce unsigned long cu = (unsigned long)c;
113*e4df2d5eSMatteo Croce
114*e4df2d5eSMatteo Croce /* Compose an ulong with 'c' repeated 4/8 times */
115*e4df2d5eSMatteo Croce cu |= cu << 8;
116*e4df2d5eSMatteo Croce cu |= cu << 16;
117*e4df2d5eSMatteo Croce /* Suppress warning on 32 bit machines */
118*e4df2d5eSMatteo Croce cu |= (cu << 16) << 16;
119*e4df2d5eSMatteo Croce
120*e4df2d5eSMatteo Croce for (; count && dest.as_uptr & WORD_MASK; count--)
121*e4df2d5eSMatteo Croce *dest.as_u8++ = c;
122*e4df2d5eSMatteo Croce
123*e4df2d5eSMatteo Croce /* Copy using the largest size allowed */
124*e4df2d5eSMatteo Croce for (; count >= BYTES_LONG; count -= BYTES_LONG)
125*e4df2d5eSMatteo Croce *dest.as_ulong++ = cu;
126*e4df2d5eSMatteo Croce }
127*e4df2d5eSMatteo Croce
128*e4df2d5eSMatteo Croce /* copy the remainder */
129*e4df2d5eSMatteo Croce while (count--)
130*e4df2d5eSMatteo Croce *dest.as_u8++ = c;
131*e4df2d5eSMatteo Croce
132*e4df2d5eSMatteo Croce return s;
133*e4df2d5eSMatteo Croce }
134*e4df2d5eSMatteo Croce EXPORT_SYMBOL(memset);
135