xref: /linux/tools/arch/x86/lib/memcpy_64.S (revision 95298d63c67673c654c08952672d016212b26054)
1/* SPDX-License-Identifier: GPL-2.0-only */
2/* Copyright 2002 Andi Kleen */
3
4#include <linux/linkage.h>
5#include <asm/errno.h>
6#include <asm/cpufeatures.h>
7#include <asm/mcsafe_test.h>
8#include <asm/alternative-asm.h>
9#include <asm/export.h>
10
11/*
12 * We build a jump to memcpy_orig by default which gets NOPped out on
13 * the majority of x86 CPUs which set REP_GOOD. In addition, CPUs which
14 * have the enhanced REP MOVSB/STOSB feature (ERMS), change those NOPs
15 * to a jmp to memcpy_erms which does the REP; MOVSB mem copy.
16 */
17
18.weak memcpy
19
20/*
21 * memcpy - Copy a memory block.
22 *
23 * Input:
24 *  rdi destination
25 *  rsi source
26 *  rdx count
27 *
28 * Output:
29 * rax original destination
30 */
31SYM_FUNC_START_ALIAS(__memcpy)
32SYM_FUNC_START_LOCAL(memcpy)
33	ALTERNATIVE_2 "jmp memcpy_orig", "", X86_FEATURE_REP_GOOD, \
34		      "jmp memcpy_erms", X86_FEATURE_ERMS
35
36	movq %rdi, %rax
37	movq %rdx, %rcx
38	shrq $3, %rcx
39	andl $7, %edx
40	rep movsq
41	movl %edx, %ecx
42	rep movsb
43	ret
44SYM_FUNC_END(memcpy)
45SYM_FUNC_END_ALIAS(__memcpy)
46EXPORT_SYMBOL(memcpy)
47EXPORT_SYMBOL(__memcpy)
48
49/*
50 * memcpy_erms() - enhanced fast string memcpy. This is faster and
51 * simpler than memcpy. Use memcpy_erms when possible.
52 */
53SYM_FUNC_START(memcpy_erms)
54	movq %rdi, %rax
55	movq %rdx, %rcx
56	rep movsb
57	ret
58SYM_FUNC_END(memcpy_erms)
59
60SYM_FUNC_START(memcpy_orig)
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
185SYM_FUNC_END(memcpy_orig)
186
187#ifndef CONFIG_UML
188
189MCSAFE_TEST_CTL
190
191/*
192 * __memcpy_mcsafe - memory copy with machine check exception handling
193 * Note that we only catch machine checks when reading the source addresses.
194 * Writes to target are posted and don't generate machine checks.
195 */
196SYM_FUNC_START(__memcpy_mcsafe)
197	cmpl $8, %edx
198	/* Less than 8 bytes? Go to byte copy loop */
199	jb .L_no_whole_words
200
201	/* Check for bad alignment of source */
202	testl $7, %esi
203	/* Already aligned */
204	jz .L_8byte_aligned
205
206	/* Copy one byte at a time until source is 8-byte aligned */
207	movl %esi, %ecx
208	andl $7, %ecx
209	subl $8, %ecx
210	negl %ecx
211	subl %ecx, %edx
212.L_read_leading_bytes:
213	movb (%rsi), %al
214	MCSAFE_TEST_SRC %rsi 1 .E_leading_bytes
215	MCSAFE_TEST_DST %rdi 1 .E_leading_bytes
216.L_write_leading_bytes:
217	movb %al, (%rdi)
218	incq %rsi
219	incq %rdi
220	decl %ecx
221	jnz .L_read_leading_bytes
222
223.L_8byte_aligned:
224	movl %edx, %ecx
225	andl $7, %edx
226	shrl $3, %ecx
227	jz .L_no_whole_words
228
229.L_read_words:
230	movq (%rsi), %r8
231	MCSAFE_TEST_SRC %rsi 8 .E_read_words
232	MCSAFE_TEST_DST %rdi 8 .E_write_words
233.L_write_words:
234	movq %r8, (%rdi)
235	addq $8, %rsi
236	addq $8, %rdi
237	decl %ecx
238	jnz .L_read_words
239
240	/* Any trailing bytes? */
241.L_no_whole_words:
242	andl %edx, %edx
243	jz .L_done_memcpy_trap
244
245	/* Copy trailing bytes */
246	movl %edx, %ecx
247.L_read_trailing_bytes:
248	movb (%rsi), %al
249	MCSAFE_TEST_SRC %rsi 1 .E_trailing_bytes
250	MCSAFE_TEST_DST %rdi 1 .E_trailing_bytes
251.L_write_trailing_bytes:
252	movb %al, (%rdi)
253	incq %rsi
254	incq %rdi
255	decl %ecx
256	jnz .L_read_trailing_bytes
257
258	/* Copy successful. Return zero */
259.L_done_memcpy_trap:
260	xorl %eax, %eax
261.L_done:
262	ret
263SYM_FUNC_END(__memcpy_mcsafe)
264EXPORT_SYMBOL_GPL(__memcpy_mcsafe)
265
266	.section .fixup, "ax"
267	/*
268	 * Return number of bytes not copied for any failure. Note that
269	 * there is no "tail" handling since the source buffer is 8-byte
270	 * aligned and poison is cacheline aligned.
271	 */
272.E_read_words:
273	shll	$3, %ecx
274.E_leading_bytes:
275	addl	%edx, %ecx
276.E_trailing_bytes:
277	mov	%ecx, %eax
278	jmp	.L_done
279
280	/*
281	 * For write fault handling, given the destination is unaligned,
282	 * we handle faults on multi-byte writes with a byte-by-byte
283	 * copy up to the write-protected page.
284	 */
285.E_write_words:
286	shll	$3, %ecx
287	addl	%edx, %ecx
288	movl	%ecx, %edx
289	jmp mcsafe_handle_tail
290
291	.previous
292
293	_ASM_EXTABLE_FAULT(.L_read_leading_bytes, .E_leading_bytes)
294	_ASM_EXTABLE_FAULT(.L_read_words, .E_read_words)
295	_ASM_EXTABLE_FAULT(.L_read_trailing_bytes, .E_trailing_bytes)
296	_ASM_EXTABLE(.L_write_leading_bytes, .E_leading_bytes)
297	_ASM_EXTABLE(.L_write_words, .E_write_words)
298	_ASM_EXTABLE(.L_write_trailing_bytes, .E_trailing_bytes)
299#endif
300