xref: /linux/arch/x86/lib/memmove_64.S (revision f3539c12d8196ce0a1993364d30b3a18908470d1)
1/*
2 * Normally compiler builtins are used, but sometimes the compiler calls out
3 * of line code. Based on asm-i386/string.h.
4 *
5 * This assembly file is re-written from memmove_64.c file.
6 *	- Copyright 2011 Fenghua Yu <fenghua.yu@intel.com>
7 */
8#include <linux/linkage.h>
9#include <asm/cpufeatures.h>
10#include <asm/alternative-asm.h>
11
12#undef memmove
13
14/*
15 * Implement memmove(). This can handle overlap between src and dst.
16 *
17 * Input:
18 * rdi: dest
19 * rsi: src
20 * rdx: count
21 *
22 * Output:
23 * rax: dest
24 */
25.weak memmove
26
27ENTRY(memmove)
28ENTRY(__memmove)
29
30	/* Handle more 32 bytes in loop */
31	mov %rdi, %rax
32	cmp $0x20, %rdx
33	jb	1f
34
35	/* Decide forward/backward copy mode */
36	cmp %rdi, %rsi
37	jge .Lmemmove_begin_forward
38	mov %rsi, %r8
39	add %rdx, %r8
40	cmp %rdi, %r8
41	jg 2f
42
43.Lmemmove_begin_forward:
44	ALTERNATIVE "", "movq %rdx, %rcx; rep movsb; retq", X86_FEATURE_ERMS
45
46	/*
47	 * movsq instruction have many startup latency
48	 * so we handle small size by general register.
49	 */
50	cmp  $680, %rdx
51	jb	3f
52	/*
53	 * movsq instruction is only good for aligned case.
54	 */
55
56	cmpb %dil, %sil
57	je 4f
583:
59	sub $0x20, %rdx
60	/*
61	 * We gobble 32 bytes forward in each loop.
62	 */
635:
64	sub $0x20, %rdx
65	movq 0*8(%rsi), %r11
66	movq 1*8(%rsi), %r10
67	movq 2*8(%rsi), %r9
68	movq 3*8(%rsi), %r8
69	leaq 4*8(%rsi), %rsi
70
71	movq %r11, 0*8(%rdi)
72	movq %r10, 1*8(%rdi)
73	movq %r9, 2*8(%rdi)
74	movq %r8, 3*8(%rdi)
75	leaq 4*8(%rdi), %rdi
76	jae 5b
77	addq $0x20, %rdx
78	jmp 1f
79	/*
80	 * Handle data forward by movsq.
81	 */
82	.p2align 4
834:
84	movq %rdx, %rcx
85	movq -8(%rsi, %rdx), %r11
86	lea -8(%rdi, %rdx), %r10
87	shrq $3, %rcx
88	rep movsq
89	movq %r11, (%r10)
90	jmp 13f
91.Lmemmove_end_forward:
92
93	/*
94	 * Handle data backward by movsq.
95	 */
96	.p2align 4
977:
98	movq %rdx, %rcx
99	movq (%rsi), %r11
100	movq %rdi, %r10
101	leaq -8(%rsi, %rdx), %rsi
102	leaq -8(%rdi, %rdx), %rdi
103	shrq $3, %rcx
104	std
105	rep movsq
106	cld
107	movq %r11, (%r10)
108	jmp 13f
109
110	/*
111	 * Start to prepare for backward copy.
112	 */
113	.p2align 4
1142:
115	cmp $680, %rdx
116	jb 6f
117	cmp %dil, %sil
118	je 7b
1196:
120	/*
121	 * Calculate copy position to tail.
122	 */
123	addq %rdx, %rsi
124	addq %rdx, %rdi
125	subq $0x20, %rdx
126	/*
127	 * We gobble 32 bytes backward in each loop.
128	 */
1298:
130	subq $0x20, %rdx
131	movq -1*8(%rsi), %r11
132	movq -2*8(%rsi), %r10
133	movq -3*8(%rsi), %r9
134	movq -4*8(%rsi), %r8
135	leaq -4*8(%rsi), %rsi
136
137	movq %r11, -1*8(%rdi)
138	movq %r10, -2*8(%rdi)
139	movq %r9, -3*8(%rdi)
140	movq %r8, -4*8(%rdi)
141	leaq -4*8(%rdi), %rdi
142	jae 8b
143	/*
144	 * Calculate copy position to head.
145	 */
146	addq $0x20, %rdx
147	subq %rdx, %rsi
148	subq %rdx, %rdi
1491:
150	cmpq $16, %rdx
151	jb 9f
152	/*
153	 * Move data from 16 bytes to 31 bytes.
154	 */
155	movq 0*8(%rsi), %r11
156	movq 1*8(%rsi), %r10
157	movq -2*8(%rsi, %rdx), %r9
158	movq -1*8(%rsi, %rdx), %r8
159	movq %r11, 0*8(%rdi)
160	movq %r10, 1*8(%rdi)
161	movq %r9, -2*8(%rdi, %rdx)
162	movq %r8, -1*8(%rdi, %rdx)
163	jmp 13f
164	.p2align 4
1659:
166	cmpq $8, %rdx
167	jb 10f
168	/*
169	 * Move data from 8 bytes to 15 bytes.
170	 */
171	movq 0*8(%rsi), %r11
172	movq -1*8(%rsi, %rdx), %r10
173	movq %r11, 0*8(%rdi)
174	movq %r10, -1*8(%rdi, %rdx)
175	jmp 13f
17610:
177	cmpq $4, %rdx
178	jb 11f
179	/*
180	 * Move data from 4 bytes to 7 bytes.
181	 */
182	movl (%rsi), %r11d
183	movl -4(%rsi, %rdx), %r10d
184	movl %r11d, (%rdi)
185	movl %r10d, -4(%rdi, %rdx)
186	jmp 13f
18711:
188	cmp $2, %rdx
189	jb 12f
190	/*
191	 * Move data from 2 bytes to 3 bytes.
192	 */
193	movw (%rsi), %r11w
194	movw -2(%rsi, %rdx), %r10w
195	movw %r11w, (%rdi)
196	movw %r10w, -2(%rdi, %rdx)
197	jmp 13f
19812:
199	cmp $1, %rdx
200	jb 13f
201	/*
202	 * Move data for 1 byte.
203	 */
204	movb (%rsi), %r11b
205	movb %r11b, (%rdi)
20613:
207	retq
208ENDPROC(__memmove)
209ENDPROC(memmove)
210