blob: 3bb1914c4fa487aecd598e11b4bd74687287c441 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/* bzero.S: Simple prefetching memset, bzero, and clear_user
2 * implementations.
3 *
4 * Copyright (C) 2005 David S. Miller <davem@davemloft.net>
5 */
6
David S. Miller8695c372012-05-11 20:33:22 -07007#include <linux/linkage.h>
Al Virod3867f042016-01-16 21:39:30 -05008#include <asm/export.h>
David S. Miller8695c372012-05-11 20:33:22 -07009
Linus Torvalds1da177e2005-04-16 15:20:36 -070010 .text
11
David S. Miller8695c372012-05-11 20:33:22 -070012ENTRY(memset) /* %o0=buf, %o1=pat, %o2=len */
Linus Torvalds1da177e2005-04-16 15:20:36 -070013 and %o1, 0xff, %o3
14 mov %o2, %o1
15 sllx %o3, 8, %g1
16 or %g1, %o3, %o2
17 sllx %o2, 16, %g1
18 or %g1, %o2, %o2
19 sllx %o2, 32, %g1
20 ba,pt %xcc, 1f
21 or %g1, %o2, %o2
22
David S. Miller8695c372012-05-11 20:33:22 -070023ENTRY(__bzero) /* %o0=buf, %o1=len */
Linus Torvalds1da177e2005-04-16 15:20:36 -070024 clr %o2
251: mov %o0, %o3
26 brz,pn %o1, __bzero_done
27 cmp %o1, 16
28 bl,pn %icc, __bzero_tiny
29 prefetch [%o0 + 0x000], #n_writes
30 andcc %o0, 0x3, %g0
31 be,pt %icc, 2f
321: stb %o2, [%o0 + 0x00]
33 add %o0, 1, %o0
34 andcc %o0, 0x3, %g0
35 bne,pn %icc, 1b
36 sub %o1, 1, %o1
372: andcc %o0, 0x7, %g0
38 be,pt %icc, 3f
39 stw %o2, [%o0 + 0x00]
40 sub %o1, 4, %o1
41 add %o0, 4, %o0
423: and %o1, 0x38, %g1
43 cmp %o1, 0x40
44 andn %o1, 0x3f, %o4
45 bl,pn %icc, 5f
46 and %o1, 0x7, %o1
47 prefetch [%o0 + 0x040], #n_writes
48 prefetch [%o0 + 0x080], #n_writes
49 prefetch [%o0 + 0x0c0], #n_writes
50 prefetch [%o0 + 0x100], #n_writes
51 prefetch [%o0 + 0x140], #n_writes
524: prefetch [%o0 + 0x180], #n_writes
53 stx %o2, [%o0 + 0x00]
54 stx %o2, [%o0 + 0x08]
55 stx %o2, [%o0 + 0x10]
56 stx %o2, [%o0 + 0x18]
57 stx %o2, [%o0 + 0x20]
58 stx %o2, [%o0 + 0x28]
59 stx %o2, [%o0 + 0x30]
60 stx %o2, [%o0 + 0x38]
61 subcc %o4, 0x40, %o4
62 bne,pt %icc, 4b
63 add %o0, 0x40, %o0
64 brz,pn %g1, 6f
65 nop
665: stx %o2, [%o0 + 0x00]
67 subcc %g1, 8, %g1
68 bne,pt %icc, 5b
69 add %o0, 0x8, %o0
706: brz,pt %o1, __bzero_done
71 nop
72__bzero_tiny:
731: stb %o2, [%o0 + 0x00]
74 subcc %o1, 1, %o1
75 bne,pt %icc, 1b
76 add %o0, 1, %o0
77__bzero_done:
78 retl
79 mov %o3, %o0
David S. Miller8695c372012-05-11 20:33:22 -070080ENDPROC(__bzero)
81ENDPROC(memset)
Al Virod3867f042016-01-16 21:39:30 -050082EXPORT_SYMBOL(__bzero)
83EXPORT_SYMBOL(memset)
Linus Torvalds1da177e2005-04-16 15:20:36 -070084
85#define EX_ST(x,y) \
8698: x,y; \
David S. Miller4d000d52006-03-04 23:23:56 -080087 .section __ex_table,"a";\
Linus Torvalds1da177e2005-04-16 15:20:36 -070088 .align 4; \
David S. Miller40bdac72009-02-08 22:00:55 -080089 .word 98b, __retl_o1; \
Linus Torvalds1da177e2005-04-16 15:20:36 -070090 .text; \
91 .align 4;
92
David S. Miller8695c372012-05-11 20:33:22 -070093ENTRY(__clear_user) /* %o0=buf, %o1=len */
David S. Millerc857e3f2006-02-17 10:35:23 -080094 brz,pn %o1, __clear_user_done
Linus Torvalds1da177e2005-04-16 15:20:36 -070095 cmp %o1, 16
David S. Millerc857e3f2006-02-17 10:35:23 -080096 bl,pn %icc, __clear_user_tiny
Linus Torvalds1da177e2005-04-16 15:20:36 -070097 EX_ST(prefetcha [%o0 + 0x00] %asi, #n_writes)
98 andcc %o0, 0x3, %g0
99 be,pt %icc, 2f
1001: EX_ST(stba %g0, [%o0 + 0x00] %asi)
101 add %o0, 1, %o0
102 andcc %o0, 0x3, %g0
103 bne,pn %icc, 1b
104 sub %o1, 1, %o1
1052: andcc %o0, 0x7, %g0
106 be,pt %icc, 3f
107 EX_ST(stwa %g0, [%o0 + 0x00] %asi)
108 sub %o1, 4, %o1
109 add %o0, 4, %o0
1103: and %o1, 0x38, %g1
111 cmp %o1, 0x40
112 andn %o1, 0x3f, %o4
113 bl,pn %icc, 5f
114 and %o1, 0x7, %o1
115 EX_ST(prefetcha [%o0 + 0x040] %asi, #n_writes)
116 EX_ST(prefetcha [%o0 + 0x080] %asi, #n_writes)
117 EX_ST(prefetcha [%o0 + 0x0c0] %asi, #n_writes)
118 EX_ST(prefetcha [%o0 + 0x100] %asi, #n_writes)
119 EX_ST(prefetcha [%o0 + 0x140] %asi, #n_writes)
1204: EX_ST(prefetcha [%o0 + 0x180] %asi, #n_writes)
121 EX_ST(stxa %g0, [%o0 + 0x00] %asi)
122 EX_ST(stxa %g0, [%o0 + 0x08] %asi)
123 EX_ST(stxa %g0, [%o0 + 0x10] %asi)
124 EX_ST(stxa %g0, [%o0 + 0x18] %asi)
125 EX_ST(stxa %g0, [%o0 + 0x20] %asi)
126 EX_ST(stxa %g0, [%o0 + 0x28] %asi)
127 EX_ST(stxa %g0, [%o0 + 0x30] %asi)
128 EX_ST(stxa %g0, [%o0 + 0x38] %asi)
129 subcc %o4, 0x40, %o4
130 bne,pt %icc, 4b
131 add %o0, 0x40, %o0
132 brz,pn %g1, 6f
133 nop
1345: EX_ST(stxa %g0, [%o0 + 0x00] %asi)
135 subcc %g1, 8, %g1
136 bne,pt %icc, 5b
137 add %o0, 0x8, %o0
David S. Millerc857e3f2006-02-17 10:35:23 -08001386: brz,pt %o1, __clear_user_done
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139 nop
David S. Millerc857e3f2006-02-17 10:35:23 -0800140__clear_user_tiny:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411: EX_ST(stba %g0, [%o0 + 0x00] %asi)
142 subcc %o1, 1, %o1
143 bne,pt %icc, 1b
144 add %o0, 1, %o0
David S. Millerc857e3f2006-02-17 10:35:23 -0800145__clear_user_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146 retl
147 clr %o0
David S. Miller8695c372012-05-11 20:33:22 -0700148ENDPROC(__clear_user)
Al Virod3867f042016-01-16 21:39:30 -0500149EXPORT_SYMBOL(__clear_user)