1 /*
2  *  linux/arch/alpha/lib/memcpy.c
3  *
4  *  Copyright (C) 1995  Linus Torvalds
5  */
6 
7 /*
8  * This is a reasonably optimized memcpy() routine.
9  */
10 
11 /*
12  * Note that the C code is written to be optimized into good assembly. However,
13  * at this point gcc is unable to sanely compile "if (n >= 0)", resulting in a
14  * explicit compare against 0 (instead of just using the proper "blt reg, xx" or
15  * "bge reg, xx"). I hope alpha-gcc will be fixed to notice this eventually..
16  */
17 
18 #include <linux/types.h>
19 
20 /*
21  * This should be done in one go with ldq_u*2/mask/stq_u. Do it
22  * with a macro so that we can fix it up later..
23  */
24 #define ALIGN_DEST_TO8_UP(d,s,n) \
25 	while (d & 7) { \
26 		if (n <= 0) return; \
27 		n--; \
28 		*(char *) d = *(char *) s; \
29 		d++; s++; \
30 	}
31 #define ALIGN_DEST_TO8_DN(d,s,n) \
32 	while (d & 7) { \
33 		if (n <= 0) return; \
34 		n--; \
35 		d--; s--; \
36 		*(char *) d = *(char *) s; \
37 	}
38 
39 /*
40  * This should similarly be done with ldq_u*2/mask/stq. The destination
41  * is aligned, but we don't fill in a full quad-word
42  */
43 #define DO_REST_UP(d,s,n) \
44 	while (n > 0) { \
45 		n--; \
46 		*(char *) d = *(char *) s; \
47 		d++; s++; \
48 	}
49 #define DO_REST_DN(d,s,n) \
50 	while (n > 0) { \
51 		n--; \
52 		d--; s--; \
53 		*(char *) d = *(char *) s; \
54 	}
55 
56 /*
57  * This should be done with ldq/mask/stq. The source and destination are
58  * aligned, but we don't fill in a full quad-word
59  */
60 #define DO_REST_ALIGNED_UP(d,s,n) DO_REST_UP(d,s,n)
61 #define DO_REST_ALIGNED_DN(d,s,n) DO_REST_DN(d,s,n)
62 
63 /*
64  * This does unaligned memory copies. We want to avoid storing to
65  * an unaligned address, as that would do a read-modify-write cycle.
66  * We also want to avoid double-reading the unaligned reads.
67  *
68  * Note the ordering to try to avoid load (and address generation) latencies.
69  */
__memcpy_unaligned_up(unsigned long d,unsigned long s,long n)70 static inline void __memcpy_unaligned_up (unsigned long d, unsigned long s,
71 					  long n)
72 {
73 	ALIGN_DEST_TO8_UP(d,s,n);
74 	n -= 8;			/* to avoid compare against 8 in the loop */
75 	if (n >= 0) {
76 		unsigned long low_word, high_word;
77 		__asm__("ldq_u %0,%1":"=r" (low_word):"m" (*(unsigned long *) s));
78 		do {
79 			unsigned long tmp;
80 			__asm__("ldq_u %0,%1":"=r" (high_word):"m" (*(unsigned long *)(s+8)));
81 			n -= 8;
82 			__asm__("extql %1,%2,%0"
83 				:"=r" (low_word)
84 				:"r" (low_word), "r" (s));
85 			__asm__("extqh %1,%2,%0"
86 				:"=r" (tmp)
87 				:"r" (high_word), "r" (s));
88 			s += 8;
89 			*(unsigned long *) d = low_word | tmp;
90 			d += 8;
91 			low_word = high_word;
92 		} while (n >= 0);
93 	}
94 	n += 8;
95 	DO_REST_UP(d,s,n);
96 }
97 
__memcpy_unaligned_dn(unsigned long d,unsigned long s,long n)98 static inline void __memcpy_unaligned_dn (unsigned long d, unsigned long s,
99 					  long n)
100 {
101 	/* I don't understand AXP assembler well enough for this. -Tim */
102 	s += n;
103 	d += n;
104 	while (n--)
105 		* (char *) --d = * (char *) --s;
106 }
107 
108 /*
109  * Hmm.. Strange. The __asm__ here is there to make gcc use an integer register
110  * for the load-store. I don't know why, but it would seem that using a floating
111  * point register for the move seems to slow things down (very small difference,
112  * though).
113  *
114  * Note the ordering to try to avoid load (and address generation) latencies.
115  */
__memcpy_aligned_up(unsigned long d,unsigned long s,long n)116 static inline void __memcpy_aligned_up (unsigned long d, unsigned long s,
117 					long n)
118 {
119 	ALIGN_DEST_TO8_UP(d,s,n);
120 	n -= 8;
121 	while (n >= 0) {
122 		unsigned long tmp;
123 		__asm__("ldq %0,%1":"=r" (tmp):"m" (*(unsigned long *) s));
124 		n -= 8;
125 		s += 8;
126 		*(unsigned long *) d = tmp;
127 		d += 8;
128 	}
129 	n += 8;
130 	DO_REST_ALIGNED_UP(d,s,n);
131 }
__memcpy_aligned_dn(unsigned long d,unsigned long s,long n)132 static inline void __memcpy_aligned_dn (unsigned long d, unsigned long s,
133 					long n)
134 {
135 	s += n;
136 	d += n;
137 	ALIGN_DEST_TO8_DN(d,s,n);
138 	n -= 8;
139 	while (n >= 0) {
140 		unsigned long tmp;
141 		s -= 8;
142 		__asm__("ldq %0,%1":"=r" (tmp):"m" (*(unsigned long *) s));
143 		n -= 8;
144 		d -= 8;
145 		*(unsigned long *) d = tmp;
146 	}
147 	n += 8;
148 	DO_REST_ALIGNED_DN(d,s,n);
149 }
150 
memcpy(void * dest,const void * src,size_t n)151 void * memcpy(void * dest, const void *src, size_t n)
152 {
153 	if (!(((unsigned long) dest ^ (unsigned long) src) & 7)) {
154 		__memcpy_aligned_up ((unsigned long) dest, (unsigned long) src,
155 				     n);
156 		return dest;
157 	}
158 	__memcpy_unaligned_up ((unsigned long) dest, (unsigned long) src, n);
159 	return dest;
160 }
161 
162 /* For backward modules compatibility, define __memcpy.  */
163 asm("__memcpy = memcpy; .globl __memcpy");
164