xref: /linux/arch/x86/lib/memcpy_64.S (revision f4738f56d1dc62aaba69b33702a5ab098f1b8c63)
1/* SPDX-License-Identifier: GPL-2.0-only */
2/* Copyright 2002 Andi Kleen */
3
4#include <linux/linkage.h>
5#include <linux/cfi_types.h>
6#include <asm/errno.h>
7#include <asm/cpufeatures.h>
8#include <asm/alternative.h>
9#include <asm/export.h>
10
11.section .noinstr.text, "ax"
12
13/*
14 * memcpy - Copy a memory block.
15 *
16 * Input:
17 *  rdi destination
18 *  rsi source
19 *  rdx count
20 *
21 * Output:
22 * rax original destination
23 *
24 * The FSRM alternative should be done inline (avoiding the call and
25 * the disgusting return handling), but that would require some help
26 * from the compiler for better calling conventions.
27 *
28 * The 'rep movsb' itself is small enough to replace the call, but the
29 * two register moves blow up the code. And one of them is "needed"
30 * only for the return value that is the same as the source input,
31 * which the compiler could/should do much better anyway.
32 */
33SYM_TYPED_FUNC_START(__memcpy)
34	ALTERNATIVE "jmp memcpy_orig", "", X86_FEATURE_FSRM
35
36	movq %rdi, %rax
37	movq %rdx, %rcx
38	rep movsb
39	RET
40SYM_FUNC_END(__memcpy)
41EXPORT_SYMBOL(__memcpy)
42
43SYM_FUNC_ALIAS_MEMFUNC(memcpy, __memcpy)
44EXPORT_SYMBOL(memcpy)
45
46SYM_FUNC_START_LOCAL(memcpy_orig)
47	movq %rdi, %rax
48
49	cmpq $0x20, %rdx
50	jb .Lhandle_tail
51
52	/*
53	 * We check whether memory false dependence could occur,
54	 * then jump to corresponding copy mode.
55	 */
56	cmp  %dil, %sil
57	jl .Lcopy_backward
58	subq $0x20, %rdx
59.Lcopy_forward_loop:
60	subq $0x20,	%rdx
61
62	/*
63	 * Move in blocks of 4x8 bytes:
64	 */
65	movq 0*8(%rsi),	%r8
66	movq 1*8(%rsi),	%r9
67	movq 2*8(%rsi),	%r10
68	movq 3*8(%rsi),	%r11
69	leaq 4*8(%rsi),	%rsi
70
71	movq %r8,	0*8(%rdi)
72	movq %r9,	1*8(%rdi)
73	movq %r10,	2*8(%rdi)
74	movq %r11,	3*8(%rdi)
75	leaq 4*8(%rdi),	%rdi
76	jae  .Lcopy_forward_loop
77	addl $0x20,	%edx
78	jmp  .Lhandle_tail
79
80.Lcopy_backward:
81	/*
82	 * Calculate copy position to tail.
83	 */
84	addq %rdx,	%rsi
85	addq %rdx,	%rdi
86	subq $0x20,	%rdx
87	/*
88	 * At most 3 ALU operations in one cycle,
89	 * so append NOPS in the same 16 bytes trunk.
90	 */
91	.p2align 4
92.Lcopy_backward_loop:
93	subq $0x20,	%rdx
94	movq -1*8(%rsi),	%r8
95	movq -2*8(%rsi),	%r9
96	movq -3*8(%rsi),	%r10
97	movq -4*8(%rsi),	%r11
98	leaq -4*8(%rsi),	%rsi
99	movq %r8,		-1*8(%rdi)
100	movq %r9,		-2*8(%rdi)
101	movq %r10,		-3*8(%rdi)
102	movq %r11,		-4*8(%rdi)
103	leaq -4*8(%rdi),	%rdi
104	jae  .Lcopy_backward_loop
105
106	/*
107	 * Calculate copy position to head.
108	 */
109	addl $0x20,	%edx
110	subq %rdx,	%rsi
111	subq %rdx,	%rdi
112.Lhandle_tail:
113	cmpl $16,	%edx
114	jb   .Lless_16bytes
115
116	/*
117	 * Move data from 16 bytes to 31 bytes.
118	 */
119	movq 0*8(%rsi), %r8
120	movq 1*8(%rsi),	%r9
121	movq -2*8(%rsi, %rdx),	%r10
122	movq -1*8(%rsi, %rdx),	%r11
123	movq %r8,	0*8(%rdi)
124	movq %r9,	1*8(%rdi)
125	movq %r10,	-2*8(%rdi, %rdx)
126	movq %r11,	-1*8(%rdi, %rdx)
127	RET
128	.p2align 4
129.Lless_16bytes:
130	cmpl $8,	%edx
131	jb   .Lless_8bytes
132	/*
133	 * Move data from 8 bytes to 15 bytes.
134	 */
135	movq 0*8(%rsi),	%r8
136	movq -1*8(%rsi, %rdx),	%r9
137	movq %r8,	0*8(%rdi)
138	movq %r9,	-1*8(%rdi, %rdx)
139	RET
140	.p2align 4
141.Lless_8bytes:
142	cmpl $4,	%edx
143	jb   .Lless_3bytes
144
145	/*
146	 * Move data from 4 bytes to 7 bytes.
147	 */
148	movl (%rsi), %ecx
149	movl -4(%rsi, %rdx), %r8d
150	movl %ecx, (%rdi)
151	movl %r8d, -4(%rdi, %rdx)
152	RET
153	.p2align 4
154.Lless_3bytes:
155	subl $1, %edx
156	jb .Lend
157	/*
158	 * Move data from 1 bytes to 3 bytes.
159	 */
160	movzbl (%rsi), %ecx
161	jz .Lstore_1byte
162	movzbq 1(%rsi), %r8
163	movzbq (%rsi, %rdx), %r9
164	movb %r8b, 1(%rdi)
165	movb %r9b, (%rdi, %rdx)
166.Lstore_1byte:
167	movb %cl, (%rdi)
168
169.Lend:
170	RET
171SYM_FUNC_END(memcpy_orig)
172
173