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