summaryrefslogtreecommitdiffstats
path: root/arch/alpha/lib/memcpy.c
blob: dc708c73e28b0375f0f7872b12c0fd0334431e1a (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
/*
 *  linux/arch/alpha/lib/memcpy.c
 *
 *  Copyright (C) 1995  Linus Torvalds
 */

/*
 * This is a reasonably optimized memcpy() routine.
 */

/*
 * Note that the C code is written to be optimized into good assembly. However,
 * at this point gcc is unable to sanely compile "if (n >= 0)", resulting in a
 * explicit compare against 0 (instead of just using the proper "blt reg, xx" or
 * "bge reg, xx"). I hope alpha-gcc will be fixed to notice this eventually..
 */

#include <linux/types.h>

/*
 * This should be done in one go with ldq_u*2/mask/stq_u. Do it
 * with a macro so that we can fix it up later..
 */
#define ALIGN_DEST_TO8(d,s,n) \
	while (d & 7) { \
		if (n <= 0) return; \
		n--; \
		*(char *) d = *(char *) s; \
		d++; s++; \
	}

/*
 * This should similarly be done with ldq_u*2/mask/stq. The destination
 * is aligned, but we don't fill in a full quad-word
 */
#define DO_REST(d,s,n) \
	while (n > 0) { \
		n--; \
		*(char *) d = *(char *) s; \
		d++; s++; \
	}

/*
 * This should be done with ldq/mask/stq. The source and destination are
 * aligned, but we don't fill in a full quad-word
 */
#define DO_REST_ALIGNED(d,s,n) DO_REST(d,s,n)

/*
 * This does unaligned memory copies. We want to avoid storing to
 * an unaligned address, as that would do a read-modify-write cycle.
 * We also want to avoid double-reading the unaligned reads.
 *
 * Note the ordering to try to avoid load (and address generation) latencies.
 */
static inline void __memcpy_unaligned(unsigned long d, unsigned long s, long n)
{
	ALIGN_DEST_TO8(d,s,n);
	n -= 8;			/* to avoid compare against 8 in the loop */
	if (n >= 0) {
		unsigned long low_word, high_word;
		__asm__("ldq_u %0,%1":"=r" (low_word):"m" (*(unsigned long *) s));
		do {
			unsigned long tmp;
			__asm__("ldq_u %0,%1":"=r" (high_word):"m" (*(unsigned long *)(s+8)));
			n -= 8;
			__asm__("extql %1,%2,%0"
				:"=r" (low_word)
				:"r" (low_word), "r" (s));
			__asm__("extqh %1,%2,%0"
				:"=r" (tmp)
				:"r" (high_word), "r" (s));
			s += 8;
			*(unsigned long *) d = low_word | tmp;
			d += 8;
			low_word = high_word;
		} while (n >= 0);
	}
	n += 8;
	DO_REST(d,s,n);
}

/*
 * Hmm.. Strange. The __asm__ here is there to make gcc use an integer register
 * for the load-store. I don't know why, but it would seem that using a floating
 * point register for the move seems to slow things down (very small difference,
 * though).
 *
 * Note the ordering to try to avoid load (and address generation) latencies.
 */
static inline void __memcpy_aligned(unsigned long d, unsigned long s, long n)
{
	ALIGN_DEST_TO8(d,s,n);
	n -= 8;
	while (n >= 0) {
		unsigned long tmp;
		__asm__("ldq %0,%1":"=r" (tmp):"m" (*(unsigned long *) s));
		n -= 8;
		s += 8;
		*(unsigned long *) d = tmp;
		d += 8;
	}
	n += 8;
	DO_REST_ALIGNED(d,s,n);
}

void * memcpy(void * dest, const void *src, size_t n)
{
	if (!(((unsigned long) dest ^ (unsigned long) src) & 7)) {
		__memcpy_aligned((unsigned long) dest, (unsigned long) src, n);
		return dest;
	}
	__memcpy_unaligned((unsigned long) dest, (unsigned long) src, n);
	return dest;
}

/* For backward modules compatibility, define __memcpy.  */
asm("__memcpy = memcpy; .globl __memcpy");