Loading...
1/* SPDX-License-Identifier: GPL-2.0 */
2/* Copyright 2002 Andi Kleen, SuSE Labs */
3
4#include <linux/export.h>
5#include <linux/linkage.h>
6#include <asm/cpufeatures.h>
7#include <asm/alternative.h>
8
9.section .noinstr.text, "ax"
10
11/*
12 * ISO C memset - set a memory block to a byte value. This function uses fast
13 * string to get better performance than the original function. The code is
14 * simpler and shorter than the original function as well.
15 *
16 * rdi destination
17 * rsi value (char)
18 * rdx count (bytes)
19 *
20 * rax original destination
21 *
22 * The FSRS alternative should be done inline (avoiding the call and
23 * the disgusting return handling), but that would require some help
24 * from the compiler for better calling conventions.
25 *
26 * The 'rep stosb' itself is small enough to replace the call, but all
27 * the register moves blow up the code. And two of them are "needed"
28 * only for the return value that is the same as the source input,
29 * which the compiler could/should do much better anyway.
30 */
31SYM_FUNC_START(__memset)
32 ALTERNATIVE "jmp memset_orig", "", X86_FEATURE_FSRS
33
34 movq %rdi,%r9
35 movb %sil,%al
36 movq %rdx,%rcx
37 rep stosb
38 movq %r9,%rax
39 RET
40SYM_FUNC_END(__memset)
41EXPORT_SYMBOL(__memset)
42
43SYM_FUNC_ALIAS_MEMFUNC(memset, __memset)
44EXPORT_SYMBOL(memset)
45
46SYM_FUNC_START_LOCAL(memset_orig)
47 movq %rdi,%r10
48
49 /* expand byte value */
50 movzbl %sil,%ecx
51 movabs $0x0101010101010101,%rax
52 imulq %rcx,%rax
53
54 /* align dst */
55 movl %edi,%r9d
56 andl $7,%r9d
57 jnz .Lbad_alignment
58.Lafter_bad_alignment:
59
60 movq %rdx,%rcx
61 shrq $6,%rcx
62 jz .Lhandle_tail
63
64 .p2align 4
65.Lloop_64:
66 decq %rcx
67 movq %rax,(%rdi)
68 movq %rax,8(%rdi)
69 movq %rax,16(%rdi)
70 movq %rax,24(%rdi)
71 movq %rax,32(%rdi)
72 movq %rax,40(%rdi)
73 movq %rax,48(%rdi)
74 movq %rax,56(%rdi)
75 leaq 64(%rdi),%rdi
76 jnz .Lloop_64
77
78 /* Handle tail in loops. The loops should be faster than hard
79 to predict jump tables. */
80 .p2align 4
81.Lhandle_tail:
82 movl %edx,%ecx
83 andl $63&(~7),%ecx
84 jz .Lhandle_7
85 shrl $3,%ecx
86 .p2align 4
87.Lloop_8:
88 decl %ecx
89 movq %rax,(%rdi)
90 leaq 8(%rdi),%rdi
91 jnz .Lloop_8
92
93.Lhandle_7:
94 andl $7,%edx
95 jz .Lende
96 .p2align 4
97.Lloop_1:
98 decl %edx
99 movb %al,(%rdi)
100 leaq 1(%rdi),%rdi
101 jnz .Lloop_1
102
103.Lende:
104 movq %r10,%rax
105 RET
106
107.Lbad_alignment:
108 cmpq $7,%rdx
109 jbe .Lhandle_7
110 movq %rax,(%rdi) /* unaligned store */
111 movq $8,%r8
112 subq %r9,%r8
113 addq %r8,%rdi
114 subq %r8,%rdx
115 jmp .Lafter_bad_alignment
116.Lfinal:
117SYM_FUNC_END(memset_orig)
1/* Copyright 2002 Andi Kleen, SuSE Labs */
2
3#include <linux/linkage.h>
4#include <asm/dwarf2.h>
5#include <asm/cpufeature.h>
6#include <asm/alternative-asm.h>
7
8/*
9 * ISO C memset - set a memory block to a byte value. This function uses fast
10 * string to get better performance than the original function. The code is
11 * simpler and shorter than the orignal function as well.
12 *
13 * rdi destination
14 * rsi value (char)
15 * rdx count (bytes)
16 *
17 * rax original destination
18 */
19 .section .altinstr_replacement, "ax", @progbits
20.Lmemset_c:
21 movq %rdi,%r9
22 movq %rdx,%rcx
23 andl $7,%edx
24 shrq $3,%rcx
25 /* expand byte value */
26 movzbl %sil,%esi
27 movabs $0x0101010101010101,%rax
28 imulq %rsi,%rax
29 rep stosq
30 movl %edx,%ecx
31 rep stosb
32 movq %r9,%rax
33 ret
34.Lmemset_e:
35 .previous
36
37/*
38 * ISO C memset - set a memory block to a byte value. This function uses
39 * enhanced rep stosb to override the fast string function.
40 * The code is simpler and shorter than the fast string function as well.
41 *
42 * rdi destination
43 * rsi value (char)
44 * rdx count (bytes)
45 *
46 * rax original destination
47 */
48 .section .altinstr_replacement, "ax", @progbits
49.Lmemset_c_e:
50 movq %rdi,%r9
51 movb %sil,%al
52 movq %rdx,%rcx
53 rep stosb
54 movq %r9,%rax
55 ret
56.Lmemset_e_e:
57 .previous
58
59ENTRY(memset)
60ENTRY(__memset)
61 CFI_STARTPROC
62 movq %rdi,%r10
63
64 /* expand byte value */
65 movzbl %sil,%ecx
66 movabs $0x0101010101010101,%rax
67 imulq %rcx,%rax
68
69 /* align dst */
70 movl %edi,%r9d
71 andl $7,%r9d
72 jnz .Lbad_alignment
73 CFI_REMEMBER_STATE
74.Lafter_bad_alignment:
75
76 movq %rdx,%rcx
77 shrq $6,%rcx
78 jz .Lhandle_tail
79
80 .p2align 4
81.Lloop_64:
82 decq %rcx
83 movq %rax,(%rdi)
84 movq %rax,8(%rdi)
85 movq %rax,16(%rdi)
86 movq %rax,24(%rdi)
87 movq %rax,32(%rdi)
88 movq %rax,40(%rdi)
89 movq %rax,48(%rdi)
90 movq %rax,56(%rdi)
91 leaq 64(%rdi),%rdi
92 jnz .Lloop_64
93
94 /* Handle tail in loops. The loops should be faster than hard
95 to predict jump tables. */
96 .p2align 4
97.Lhandle_tail:
98 movl %edx,%ecx
99 andl $63&(~7),%ecx
100 jz .Lhandle_7
101 shrl $3,%ecx
102 .p2align 4
103.Lloop_8:
104 decl %ecx
105 movq %rax,(%rdi)
106 leaq 8(%rdi),%rdi
107 jnz .Lloop_8
108
109.Lhandle_7:
110 andl $7,%edx
111 jz .Lende
112 .p2align 4
113.Lloop_1:
114 decl %edx
115 movb %al,(%rdi)
116 leaq 1(%rdi),%rdi
117 jnz .Lloop_1
118
119.Lende:
120 movq %r10,%rax
121 ret
122
123 CFI_RESTORE_STATE
124.Lbad_alignment:
125 cmpq $7,%rdx
126 jbe .Lhandle_7
127 movq %rax,(%rdi) /* unaligned store */
128 movq $8,%r8
129 subq %r9,%r8
130 addq %r8,%rdi
131 subq %r8,%rdx
132 jmp .Lafter_bad_alignment
133.Lfinal:
134 CFI_ENDPROC
135ENDPROC(memset)
136ENDPROC(__memset)
137
138 /* Some CPUs support enhanced REP MOVSB/STOSB feature.
139 * It is recommended to use this when possible.
140 *
141 * If enhanced REP MOVSB/STOSB feature is not available, use fast string
142 * instructions.
143 *
144 * Otherwise, use original memset function.
145 *
146 * In .altinstructions section, ERMS feature is placed after REG_GOOD
147 * feature to implement the right patch order.
148 */
149 .section .altinstructions,"a"
150 altinstruction_entry memset,.Lmemset_c,X86_FEATURE_REP_GOOD,\
151 .Lfinal-memset,.Lmemset_e-.Lmemset_c
152 altinstruction_entry memset,.Lmemset_c_e,X86_FEATURE_ERMS, \
153 .Lfinal-memset,.Lmemset_e_e-.Lmemset_c_e
154 .previous