xref: /openbmc/linux/arch/sparc/lib/memset.S (revision 981ab3f1)
1/* linux/arch/sparc/lib/memset.S: Sparc optimized memset, bzero and clear_user code
2 * Copyright (C) 1991,1996 Free Software Foundation
3 * Copyright (C) 1996,1997 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
4 * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
5 *
6 * Calls to memset returns initial %o0. Calls to bzero returns 0, if ok, and
7 * number of bytes not yet set if exception occurs and we were called as
8 * clear_user.
9 */
10
11#include <asm/ptrace.h>
12#include <asm/export.h>
13
14/* Work around cpp -rob */
15#define ALLOC #alloc
16#define EXECINSTR #execinstr
17#define EX(x,y,a,b) 				\
1898: 	x,y;					\
19	.section .fixup,ALLOC,EXECINSTR;	\
20	.align	4;				\
2199:	ba 30f;					\
22	 a, b, %o0;				\
23	.section __ex_table,ALLOC;		\
24	.align	4;				\
25	.word	98b, 99b;			\
26	.text;					\
27	.align	4
28
29#define EXT(start,end,handler) 			\
30	.section __ex_table,ALLOC;		\
31	.align	4;				\
32	.word	start, 0, end, handler;		\
33	.text;					\
34	.align	4
35
36/* Please don't change these macros, unless you change the logic
37 * in the .fixup section below as well.
38 * Store 64 bytes at (BASE + OFFSET) using value SOURCE. */
39#define ZERO_BIG_BLOCK(base, offset, source)    \
40	std	source, [base + offset + 0x00]; \
41	std	source, [base + offset + 0x08]; \
42	std	source, [base + offset + 0x10]; \
43	std	source, [base + offset + 0x18]; \
44	std	source, [base + offset + 0x20]; \
45	std	source, [base + offset + 0x28]; \
46	std	source, [base + offset + 0x30]; \
47	std	source, [base + offset + 0x38];
48
49#define ZERO_LAST_BLOCKS(base, offset, source)	\
50	std	source, [base - offset - 0x38]; \
51	std	source, [base - offset - 0x30]; \
52	std	source, [base - offset - 0x28]; \
53	std	source, [base - offset - 0x20]; \
54	std	source, [base - offset - 0x18]; \
55	std	source, [base - offset - 0x10]; \
56	std	source, [base - offset - 0x08]; \
57	std	source, [base - offset - 0x00];
58
59	.text
60	.align 4
61
62        .globl  __bzero_begin
63__bzero_begin:
64
65	.globl	__bzero
66	.type	__bzero,#function
67	.globl	memset
68	EXPORT_SYMBOL(__bzero)
69	EXPORT_SYMBOL(memset)
70	.globl	__memset_start, __memset_end
71__memset_start:
72memset:
73	mov	%o0, %g1
74	mov	1, %g4
75	and	%o1, 0xff, %g3
76	sll	%g3, 8, %g2
77	or	%g3, %g2, %g3
78	sll	%g3, 16, %g2
79	or	%g3, %g2, %g3
80	b	1f
81	 mov	%o2, %o1
823:
83	cmp	%o2, 3
84	be	2f
85	 EX(stb	%g3, [%o0], sub %o1, 0)
86
87	cmp	%o2, 2
88	be	2f
89	 EX(stb	%g3, [%o0 + 0x01], sub %o1, 1)
90
91	EX(stb	%g3, [%o0 + 0x02], sub %o1, 2)
922:
93	sub	%o2, 4, %o2
94	add	%o1, %o2, %o1
95	b	4f
96	 sub	%o0, %o2, %o0
97
98__bzero:
99	clr	%g4
100	mov	%g0, %g3
1011:
102	cmp	%o1, 7
103	bleu	7f
104	 andcc	%o0, 3, %o2
105
106	bne	3b
1074:
108	 andcc	%o0, 4, %g0
109
110	be	2f
111	 mov	%g3, %g2
112
113	EX(st	%g3, [%o0], sub %o1, 0)
114	sub	%o1, 4, %o1
115	add	%o0, 4, %o0
1162:
117	andcc	%o1, 0xffffff80, %o3	! Now everything is 8 aligned and o1 is len to run
118	be	9f
119	 andcc	%o1, 0x78, %o2
12010:
121	ZERO_BIG_BLOCK(%o0, 0x00, %g2)
122	subcc	%o3, 128, %o3
123	ZERO_BIG_BLOCK(%o0, 0x40, %g2)
12411:
125	EXT(10b, 11b, 20f)
126	bne	10b
127	 add	%o0, 128, %o0
128
129	orcc	%o2, %g0, %g0
1309:
131	be	13f
132	 andcc	%o1, 7, %o1
133
134	srl	%o2, 1, %o3
135	set	13f, %o4
136	sub	%o4, %o3, %o4
137	jmp	%o4
138	 add	%o0, %o2, %o0
139
14012:
141	ZERO_LAST_BLOCKS(%o0, 0x48, %g2)
142	ZERO_LAST_BLOCKS(%o0, 0x08, %g2)
14313:
144	be	8f
145	 andcc	%o1, 4, %g0
146
147	be	1f
148	 andcc	%o1, 2, %g0
149
150	EX(st	%g3, [%o0], and %o1, 7)
151	add	%o0, 4, %o0
1521:
153	be	1f
154	 andcc	%o1, 1, %g0
155
156	EX(sth	%g3, [%o0], and %o1, 3)
157	add	%o0, 2, %o0
1581:
159	bne,a	8f
160	 EX(stb	%g3, [%o0], and %o1, 1)
1618:
162	b	0f
163	 nop
1647:
165	be	13b
166	 orcc	%o1, 0, %g0
167
168	be	0f
1698:
170	 add	%o0, 1, %o0
171	subcc	%o1, 1, %o1
172	bne	8b
173	 EX(stb	%g3, [%o0 - 1], add %o1, 1)
1740:
175	andcc	%g4, 1, %g0
176	be	5f
177	 nop
178	retl
179	 mov	%g1, %o0
1805:
181	retl
182	 clr	%o0
183__memset_end:
184
185	.section .fixup,#alloc,#execinstr
186	.align	4
18720:
188	cmp	%g2, 8
189	bleu	1f
190	 and	%o1, 0x7f, %o1
191	sub	%g2, 9, %g2
192	add	%o3, 64, %o3
1931:
194	sll	%g2, 3, %g2
195	add	%o3, %o1, %o0
196	b 30f
197	 sub	%o0, %g2, %o0
19821:
199	mov	8, %o0
200	and	%o1, 7, %o1
201	sub	%o0, %g2, %o0
202	sll	%o0, 3, %o0
203	b 30f
204	 add	%o0, %o1, %o0
20530:
206/* %o4 is faulting address, %o5 is %pc where fault occurred */
207	save	%sp, -104, %sp
208	mov	%i5, %o0
209	mov	%i7, %o1
210	call	lookup_fault
211	 mov	%i4, %o2
212	ret
213	 restore
214
215	.globl __bzero_end
216__bzero_end:
217