xref: /linux/arch/x86/lib/memcpy_64.S (revision e0bf6c5ca2d3281f231c5f0c9bf145e9513644de)
1/* Copyright 2002 Andi Kleen */
2
3#include <linux/linkage.h>
4
5#include <asm/cpufeature.h>
6#include <asm/dwarf2.h>
7#include <asm/alternative-asm.h>
8
9/*
10 * memcpy - Copy a memory block.
11 *
12 * Input:
13 *  rdi destination
14 *  rsi source
15 *  rdx count
16 *
17 * Output:
18 * rax original destination
19 */
20
21/*
22 * memcpy_c() - fast string ops (REP MOVSQ) based variant.
23 *
24 * This gets patched over the unrolled variant (below) via the
25 * alternative instructions framework:
26 */
27	.section .altinstr_replacement, "ax", @progbits
28.Lmemcpy_c:
29	movq %rdi, %rax
30	movq %rdx, %rcx
31	shrq $3, %rcx
32	andl $7, %edx
33	rep movsq
34	movl %edx, %ecx
35	rep movsb
36	ret
37.Lmemcpy_e:
38	.previous
39
40/*
41 * memcpy_c_e() - enhanced fast string memcpy. This is faster and simpler than
42 * memcpy_c. Use memcpy_c_e when possible.
43 *
44 * This gets patched over the unrolled variant (below) via the
45 * alternative instructions framework:
46 */
47	.section .altinstr_replacement, "ax", @progbits
48.Lmemcpy_c_e:
49	movq %rdi, %rax
50	movq %rdx, %rcx
51	rep movsb
52	ret
53.Lmemcpy_e_e:
54	.previous
55
56.weak memcpy
57
58ENTRY(__memcpy)
59ENTRY(memcpy)
60	CFI_STARTPROC
61	movq %rdi, %rax
62
63	cmpq $0x20, %rdx
64	jb .Lhandle_tail
65
66	/*
67	 * We check whether memory false dependence could occur,
68	 * then jump to corresponding copy mode.
69	 */
70	cmp  %dil, %sil
71	jl .Lcopy_backward
72	subq $0x20, %rdx
73.Lcopy_forward_loop:
74	subq $0x20,	%rdx
75
76	/*
77	 * Move in blocks of 4x8 bytes:
78	 */
79	movq 0*8(%rsi),	%r8
80	movq 1*8(%rsi),	%r9
81	movq 2*8(%rsi),	%r10
82	movq 3*8(%rsi),	%r11
83	leaq 4*8(%rsi),	%rsi
84
85	movq %r8,	0*8(%rdi)
86	movq %r9,	1*8(%rdi)
87	movq %r10,	2*8(%rdi)
88	movq %r11,	3*8(%rdi)
89	leaq 4*8(%rdi),	%rdi
90	jae  .Lcopy_forward_loop
91	addl $0x20,	%edx
92	jmp  .Lhandle_tail
93
94.Lcopy_backward:
95	/*
96	 * Calculate copy position to tail.
97	 */
98	addq %rdx,	%rsi
99	addq %rdx,	%rdi
100	subq $0x20,	%rdx
101	/*
102	 * At most 3 ALU operations in one cycle,
103	 * so append NOPS in the same 16 bytes trunk.
104	 */
105	.p2align 4
106.Lcopy_backward_loop:
107	subq $0x20,	%rdx
108	movq -1*8(%rsi),	%r8
109	movq -2*8(%rsi),	%r9
110	movq -3*8(%rsi),	%r10
111	movq -4*8(%rsi),	%r11
112	leaq -4*8(%rsi),	%rsi
113	movq %r8,		-1*8(%rdi)
114	movq %r9,		-2*8(%rdi)
115	movq %r10,		-3*8(%rdi)
116	movq %r11,		-4*8(%rdi)
117	leaq -4*8(%rdi),	%rdi
118	jae  .Lcopy_backward_loop
119
120	/*
121	 * Calculate copy position to head.
122	 */
123	addl $0x20,	%edx
124	subq %rdx,	%rsi
125	subq %rdx,	%rdi
126.Lhandle_tail:
127	cmpl $16,	%edx
128	jb   .Lless_16bytes
129
130	/*
131	 * Move data from 16 bytes to 31 bytes.
132	 */
133	movq 0*8(%rsi), %r8
134	movq 1*8(%rsi),	%r9
135	movq -2*8(%rsi, %rdx),	%r10
136	movq -1*8(%rsi, %rdx),	%r11
137	movq %r8,	0*8(%rdi)
138	movq %r9,	1*8(%rdi)
139	movq %r10,	-2*8(%rdi, %rdx)
140	movq %r11,	-1*8(%rdi, %rdx)
141	retq
142	.p2align 4
143.Lless_16bytes:
144	cmpl $8,	%edx
145	jb   .Lless_8bytes
146	/*
147	 * Move data from 8 bytes to 15 bytes.
148	 */
149	movq 0*8(%rsi),	%r8
150	movq -1*8(%rsi, %rdx),	%r9
151	movq %r8,	0*8(%rdi)
152	movq %r9,	-1*8(%rdi, %rdx)
153	retq
154	.p2align 4
155.Lless_8bytes:
156	cmpl $4,	%edx
157	jb   .Lless_3bytes
158
159	/*
160	 * Move data from 4 bytes to 7 bytes.
161	 */
162	movl (%rsi), %ecx
163	movl -4(%rsi, %rdx), %r8d
164	movl %ecx, (%rdi)
165	movl %r8d, -4(%rdi, %rdx)
166	retq
167	.p2align 4
168.Lless_3bytes:
169	subl $1, %edx
170	jb .Lend
171	/*
172	 * Move data from 1 bytes to 3 bytes.
173	 */
174	movzbl (%rsi), %ecx
175	jz .Lstore_1byte
176	movzbq 1(%rsi), %r8
177	movzbq (%rsi, %rdx), %r9
178	movb %r8b, 1(%rdi)
179	movb %r9b, (%rdi, %rdx)
180.Lstore_1byte:
181	movb %cl, (%rdi)
182
183.Lend:
184	retq
185	CFI_ENDPROC
186ENDPROC(memcpy)
187ENDPROC(__memcpy)
188
189	/*
190	 * Some CPUs are adding enhanced REP MOVSB/STOSB feature
191	 * If the feature is supported, memcpy_c_e() is the first choice.
192	 * If enhanced rep movsb copy is not available, use fast string copy
193	 * memcpy_c() when possible. This is faster and code is simpler than
194	 * original memcpy().
195	 * Otherwise, original memcpy() is used.
196	 * In .altinstructions section, ERMS feature is placed after REG_GOOD
197         * feature to implement the right patch order.
198	 *
199	 * Replace only beginning, memcpy is used to apply alternatives,
200	 * so it is silly to overwrite itself with nops - reboot is the
201	 * only outcome...
202	 */
203	.section .altinstructions, "a"
204	altinstruction_entry __memcpy,.Lmemcpy_c,X86_FEATURE_REP_GOOD,\
205			     .Lmemcpy_e-.Lmemcpy_c,.Lmemcpy_e-.Lmemcpy_c
206	altinstruction_entry __memcpy,.Lmemcpy_c_e,X86_FEATURE_ERMS, \
207			     .Lmemcpy_e_e-.Lmemcpy_c_e,.Lmemcpy_e_e-.Lmemcpy_c_e
208	.previous
209