1/* SPDX-License-Identifier: GPL-2.0 */
2/* Copyright 2002 Andi Kleen, SuSE Labs */
3
4#include <linux/linkage.h>
5#include <asm/cpufeatures.h>
6#include <asm/alternative.h>
7#include <asm/export.h>
8
9/*
10 * ISO C memset - set a memory block to a byte value. This function uses fast
11 * string to get better performance than the original function. The code is
12 * simpler and shorter than the original function as well.
13 *
14 * rdi   destination
15 * rsi   value (char)
16 * rdx   count (bytes)
17 *
18 * rax   original destination
19 */
20SYM_FUNC_START_WEAK(memset)
21SYM_FUNC_START(__memset)
22	/*
23	 * Some CPUs support enhanced REP MOVSB/STOSB feature. It is recommended
24	 * to use it when possible. If not available, use fast string instructions.
25	 *
26	 * Otherwise, use original memset function.
27	 */
28	ALTERNATIVE_2 "jmp memset_orig", "", X86_FEATURE_REP_GOOD, \
29		      "jmp memset_erms", X86_FEATURE_ERMS
30
31	movq %rdi,%r9
32	movq %rdx,%rcx
33	andl $7,%edx
34	shrq $3,%rcx
35	/* expand byte value  */
36	movzbl %sil,%esi
37	movabs $0x0101010101010101,%rax
38	imulq %rsi,%rax
39	rep stosq
40	movl %edx,%ecx
41	rep stosb
42	movq %r9,%rax
43	ret
44SYM_FUNC_END(__memset)
45SYM_FUNC_END_ALIAS(memset)
46EXPORT_SYMBOL(memset)
47EXPORT_SYMBOL(__memset)
48
49/*
50 * ISO C memset - set a memory block to a byte value. This function uses
51 * enhanced rep stosb to override the fast string function.
52 * The code is simpler and shorter than the fast string function as well.
53 *
54 * rdi   destination
55 * rsi   value (char)
56 * rdx   count (bytes)
57 *
58 * rax   original destination
59 */
60SYM_FUNC_START_LOCAL(memset_erms)
61	movq %rdi,%r9
62	movb %sil,%al
63	movq %rdx,%rcx
64	rep stosb
65	movq %r9,%rax
66	ret
67SYM_FUNC_END(memset_erms)
68
69SYM_FUNC_START_LOCAL(memset_orig)
70	movq %rdi,%r10
71
72	/* expand byte value  */
73	movzbl %sil,%ecx
74	movabs $0x0101010101010101,%rax
75	imulq  %rcx,%rax
76
77	/* align dst */
78	movl  %edi,%r9d
79	andl  $7,%r9d
80	jnz  .Lbad_alignment
81.Lafter_bad_alignment:
82
83	movq  %rdx,%rcx
84	shrq  $6,%rcx
85	jz	 .Lhandle_tail
86
87	.p2align 4
88.Lloop_64:
89	decq  %rcx
90	movq  %rax,(%rdi)
91	movq  %rax,8(%rdi)
92	movq  %rax,16(%rdi)
93	movq  %rax,24(%rdi)
94	movq  %rax,32(%rdi)
95	movq  %rax,40(%rdi)
96	movq  %rax,48(%rdi)
97	movq  %rax,56(%rdi)
98	leaq  64(%rdi),%rdi
99	jnz    .Lloop_64
100
101	/* Handle tail in loops. The loops should be faster than hard
102	   to predict jump tables. */
103	.p2align 4
104.Lhandle_tail:
105	movl	%edx,%ecx
106	andl    $63&(~7),%ecx
107	jz 		.Lhandle_7
108	shrl	$3,%ecx
109	.p2align 4
110.Lloop_8:
111	decl   %ecx
112	movq  %rax,(%rdi)
113	leaq  8(%rdi),%rdi
114	jnz    .Lloop_8
115
116.Lhandle_7:
117	andl	$7,%edx
118	jz      .Lende
119	.p2align 4
120.Lloop_1:
121	decl    %edx
122	movb 	%al,(%rdi)
123	leaq	1(%rdi),%rdi
124	jnz     .Lloop_1
125
126.Lende:
127	movq	%r10,%rax
128	ret
129
130.Lbad_alignment:
131	cmpq $7,%rdx
132	jbe	.Lhandle_7
133	movq %rax,(%rdi)	/* unaligned store */
134	movq $8,%r8
135	subq %r9,%r8
136	addq %r8,%rdi
137	subq %r8,%rdx
138	jmp .Lafter_bad_alignment
139.Lfinal:
140SYM_FUNC_END(memset_orig)
141