xref: /openbmc/linux/tools/arch/x86/lib/memcpy_64.S (revision 82e6fdd6)
1/* Copyright 2002 Andi Kleen */
2
3#include <linux/linkage.h>
4#include <asm/errno.h>
5#include <asm/cpufeatures.h>
6#include <asm/alternative-asm.h>
7#include <asm/export.h>
8
9/*
10 * We build a jump to memcpy_orig by default which gets NOPped out on
11 * the majority of x86 CPUs which set REP_GOOD. In addition, CPUs which
12 * have the enhanced REP MOVSB/STOSB feature (ERMS), change those NOPs
13 * to a jmp to memcpy_erms which does the REP; MOVSB mem copy.
14 */
15
16.weak memcpy
17
18/*
19 * memcpy - Copy a memory block.
20 *
21 * Input:
22 *  rdi destination
23 *  rsi source
24 *  rdx count
25 *
26 * Output:
27 * rax original destination
28 */
29ENTRY(__memcpy)
30ENTRY(memcpy)
31	ALTERNATIVE_2 "jmp memcpy_orig", "", X86_FEATURE_REP_GOOD, \
32		      "jmp memcpy_erms", X86_FEATURE_ERMS
33
34	movq %rdi, %rax
35	movq %rdx, %rcx
36	shrq $3, %rcx
37	andl $7, %edx
38	rep movsq
39	movl %edx, %ecx
40	rep movsb
41	ret
42ENDPROC(memcpy)
43ENDPROC(__memcpy)
44EXPORT_SYMBOL(memcpy)
45EXPORT_SYMBOL(__memcpy)
46
47/*
48 * memcpy_erms() - enhanced fast string memcpy. This is faster and
49 * simpler than memcpy. Use memcpy_erms when possible.
50 */
51ENTRY(memcpy_erms)
52	movq %rdi, %rax
53	movq %rdx, %rcx
54	rep movsb
55	ret
56ENDPROC(memcpy_erms)
57
58ENTRY(memcpy_orig)
59	movq %rdi, %rax
60
61	cmpq $0x20, %rdx
62	jb .Lhandle_tail
63
64	/*
65	 * We check whether memory false dependence could occur,
66	 * then jump to corresponding copy mode.
67	 */
68	cmp  %dil, %sil
69	jl .Lcopy_backward
70	subq $0x20, %rdx
71.Lcopy_forward_loop:
72	subq $0x20,	%rdx
73
74	/*
75	 * Move in blocks of 4x8 bytes:
76	 */
77	movq 0*8(%rsi),	%r8
78	movq 1*8(%rsi),	%r9
79	movq 2*8(%rsi),	%r10
80	movq 3*8(%rsi),	%r11
81	leaq 4*8(%rsi),	%rsi
82
83	movq %r8,	0*8(%rdi)
84	movq %r9,	1*8(%rdi)
85	movq %r10,	2*8(%rdi)
86	movq %r11,	3*8(%rdi)
87	leaq 4*8(%rdi),	%rdi
88	jae  .Lcopy_forward_loop
89	addl $0x20,	%edx
90	jmp  .Lhandle_tail
91
92.Lcopy_backward:
93	/*
94	 * Calculate copy position to tail.
95	 */
96	addq %rdx,	%rsi
97	addq %rdx,	%rdi
98	subq $0x20,	%rdx
99	/*
100	 * At most 3 ALU operations in one cycle,
101	 * so append NOPS in the same 16 bytes trunk.
102	 */
103	.p2align 4
104.Lcopy_backward_loop:
105	subq $0x20,	%rdx
106	movq -1*8(%rsi),	%r8
107	movq -2*8(%rsi),	%r9
108	movq -3*8(%rsi),	%r10
109	movq -4*8(%rsi),	%r11
110	leaq -4*8(%rsi),	%rsi
111	movq %r8,		-1*8(%rdi)
112	movq %r9,		-2*8(%rdi)
113	movq %r10,		-3*8(%rdi)
114	movq %r11,		-4*8(%rdi)
115	leaq -4*8(%rdi),	%rdi
116	jae  .Lcopy_backward_loop
117
118	/*
119	 * Calculate copy position to head.
120	 */
121	addl $0x20,	%edx
122	subq %rdx,	%rsi
123	subq %rdx,	%rdi
124.Lhandle_tail:
125	cmpl $16,	%edx
126	jb   .Lless_16bytes
127
128	/*
129	 * Move data from 16 bytes to 31 bytes.
130	 */
131	movq 0*8(%rsi), %r8
132	movq 1*8(%rsi),	%r9
133	movq -2*8(%rsi, %rdx),	%r10
134	movq -1*8(%rsi, %rdx),	%r11
135	movq %r8,	0*8(%rdi)
136	movq %r9,	1*8(%rdi)
137	movq %r10,	-2*8(%rdi, %rdx)
138	movq %r11,	-1*8(%rdi, %rdx)
139	retq
140	.p2align 4
141.Lless_16bytes:
142	cmpl $8,	%edx
143	jb   .Lless_8bytes
144	/*
145	 * Move data from 8 bytes to 15 bytes.
146	 */
147	movq 0*8(%rsi),	%r8
148	movq -1*8(%rsi, %rdx),	%r9
149	movq %r8,	0*8(%rdi)
150	movq %r9,	-1*8(%rdi, %rdx)
151	retq
152	.p2align 4
153.Lless_8bytes:
154	cmpl $4,	%edx
155	jb   .Lless_3bytes
156
157	/*
158	 * Move data from 4 bytes to 7 bytes.
159	 */
160	movl (%rsi), %ecx
161	movl -4(%rsi, %rdx), %r8d
162	movl %ecx, (%rdi)
163	movl %r8d, -4(%rdi, %rdx)
164	retq
165	.p2align 4
166.Lless_3bytes:
167	subl $1, %edx
168	jb .Lend
169	/*
170	 * Move data from 1 bytes to 3 bytes.
171	 */
172	movzbl (%rsi), %ecx
173	jz .Lstore_1byte
174	movzbq 1(%rsi), %r8
175	movzbq (%rsi, %rdx), %r9
176	movb %r8b, 1(%rdi)
177	movb %r9b, (%rdi, %rdx)
178.Lstore_1byte:
179	movb %cl, (%rdi)
180
181.Lend:
182	retq
183ENDPROC(memcpy_orig)
184
185#ifndef CONFIG_UML
186/*
187 * memcpy_mcsafe_unrolled - memory copy with machine check exception handling
188 * Note that we only catch machine checks when reading the source addresses.
189 * Writes to target are posted and don't generate machine checks.
190 */
191ENTRY(memcpy_mcsafe_unrolled)
192	cmpl $8, %edx
193	/* Less than 8 bytes? Go to byte copy loop */
194	jb .L_no_whole_words
195
196	/* Check for bad alignment of source */
197	testl $7, %esi
198	/* Already aligned */
199	jz .L_8byte_aligned
200
201	/* Copy one byte at a time until source is 8-byte aligned */
202	movl %esi, %ecx
203	andl $7, %ecx
204	subl $8, %ecx
205	negl %ecx
206	subl %ecx, %edx
207.L_copy_leading_bytes:
208	movb (%rsi), %al
209	movb %al, (%rdi)
210	incq %rsi
211	incq %rdi
212	decl %ecx
213	jnz .L_copy_leading_bytes
214
215.L_8byte_aligned:
216	/* Figure out how many whole cache lines (64-bytes) to copy */
217	movl %edx, %ecx
218	andl $63, %edx
219	shrl $6, %ecx
220	jz .L_no_whole_cache_lines
221
222	/* Loop copying whole cache lines */
223.L_cache_w0: movq (%rsi), %r8
224.L_cache_w1: movq 1*8(%rsi), %r9
225.L_cache_w2: movq 2*8(%rsi), %r10
226.L_cache_w3: movq 3*8(%rsi), %r11
227	movq %r8, (%rdi)
228	movq %r9, 1*8(%rdi)
229	movq %r10, 2*8(%rdi)
230	movq %r11, 3*8(%rdi)
231.L_cache_w4: movq 4*8(%rsi), %r8
232.L_cache_w5: movq 5*8(%rsi), %r9
233.L_cache_w6: movq 6*8(%rsi), %r10
234.L_cache_w7: movq 7*8(%rsi), %r11
235	movq %r8, 4*8(%rdi)
236	movq %r9, 5*8(%rdi)
237	movq %r10, 6*8(%rdi)
238	movq %r11, 7*8(%rdi)
239	leaq 64(%rsi), %rsi
240	leaq 64(%rdi), %rdi
241	decl %ecx
242	jnz .L_cache_w0
243
244	/* Are there any trailing 8-byte words? */
245.L_no_whole_cache_lines:
246	movl %edx, %ecx
247	andl $7, %edx
248	shrl $3, %ecx
249	jz .L_no_whole_words
250
251	/* Copy trailing words */
252.L_copy_trailing_words:
253	movq (%rsi), %r8
254	mov %r8, (%rdi)
255	leaq 8(%rsi), %rsi
256	leaq 8(%rdi), %rdi
257	decl %ecx
258	jnz .L_copy_trailing_words
259
260	/* Any trailing bytes? */
261.L_no_whole_words:
262	andl %edx, %edx
263	jz .L_done_memcpy_trap
264
265	/* Copy trailing bytes */
266	movl %edx, %ecx
267.L_copy_trailing_bytes:
268	movb (%rsi), %al
269	movb %al, (%rdi)
270	incq %rsi
271	incq %rdi
272	decl %ecx
273	jnz .L_copy_trailing_bytes
274
275	/* Copy successful. Return zero */
276.L_done_memcpy_trap:
277	xorq %rax, %rax
278	ret
279ENDPROC(memcpy_mcsafe_unrolled)
280EXPORT_SYMBOL_GPL(memcpy_mcsafe_unrolled)
281
282	.section .fixup, "ax"
283	/* Return -EFAULT for any failure */
284.L_memcpy_mcsafe_fail:
285	mov	$-EFAULT, %rax
286	ret
287
288	.previous
289
290	_ASM_EXTABLE_FAULT(.L_copy_leading_bytes, .L_memcpy_mcsafe_fail)
291	_ASM_EXTABLE_FAULT(.L_cache_w0, .L_memcpy_mcsafe_fail)
292	_ASM_EXTABLE_FAULT(.L_cache_w1, .L_memcpy_mcsafe_fail)
293	_ASM_EXTABLE_FAULT(.L_cache_w2, .L_memcpy_mcsafe_fail)
294	_ASM_EXTABLE_FAULT(.L_cache_w3, .L_memcpy_mcsafe_fail)
295	_ASM_EXTABLE_FAULT(.L_cache_w4, .L_memcpy_mcsafe_fail)
296	_ASM_EXTABLE_FAULT(.L_cache_w5, .L_memcpy_mcsafe_fail)
297	_ASM_EXTABLE_FAULT(.L_cache_w6, .L_memcpy_mcsafe_fail)
298	_ASM_EXTABLE_FAULT(.L_cache_w7, .L_memcpy_mcsafe_fail)
299	_ASM_EXTABLE_FAULT(.L_copy_trailing_words, .L_memcpy_mcsafe_fail)
300	_ASM_EXTABLE_FAULT(.L_copy_trailing_bytes, .L_memcpy_mcsafe_fail)
301#endif
302