xref: /openbmc/linux/arch/sparc/lib/memset.S (revision 1c2dd16a) 
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	.globl	memset
67	EXPORT_SYMBOL(__bzero)
68	EXPORT_SYMBOL(memset)
69	.globl	__memset_start, __memset_end
70__memset_start:
71memset:
72	mov	%o0, %g1
73	mov	1, %g4
74	and	%o1, 0xff, %g3
75	sll	%g3, 8, %g2
76	or	%g3, %g2, %g3
77	sll	%g3, 16, %g2
78	or	%g3, %g2, %g3
79	b	1f
80	 mov	%o2, %o1
813:
82	cmp	%o2, 3
83	be	2f
84	 EX(stb	%g3, [%o0], sub %o1, 0)
85
86	cmp	%o2, 2
87	be	2f
88	 EX(stb	%g3, [%o0 + 0x01], sub %o1, 1)
89
90	EX(stb	%g3, [%o0 + 0x02], sub %o1, 2)
912:
92	sub	%o2, 4, %o2
93	add	%o1, %o2, %o1
94	b	4f
95	 sub	%o0, %o2, %o0
96
97__bzero:
98	clr	%g4
99	mov	%g0, %g3
1001:
101	cmp	%o1, 7
102	bleu	7f
103	 andcc	%o0, 3, %o2
104
105	bne	3b
1064:
107	 andcc	%o0, 4, %g0
108
109	be	2f
110	 mov	%g3, %g2
111
112	EX(st	%g3, [%o0], sub %o1, 0)
113	sub	%o1, 4, %o1
114	add	%o0, 4, %o0
1152:
116	andcc	%o1, 0xffffff80, %o3	! Now everything is 8 aligned and o1 is len to run
117	be	9f
118	 andcc	%o1, 0x78, %o2
11910:
120	ZERO_BIG_BLOCK(%o0, 0x00, %g2)
121	subcc	%o3, 128, %o3
122	ZERO_BIG_BLOCK(%o0, 0x40, %g2)
12311:
124	EXT(10b, 11b, 20f)
125	bne	10b
126	 add	%o0, 128, %o0
127
128	orcc	%o2, %g0, %g0
1299:
130	be	13f
131	 andcc	%o1, 7, %o1
132
133	srl	%o2, 1, %o3
134	set	13f, %o4
135	sub	%o4, %o3, %o4
136	jmp	%o4
137	 add	%o0, %o2, %o0
138
13912:
140	ZERO_LAST_BLOCKS(%o0, 0x48, %g2)
141	ZERO_LAST_BLOCKS(%o0, 0x08, %g2)
14213:
143	be	8f
144	 andcc	%o1, 4, %g0
145
146	be	1f
147	 andcc	%o1, 2, %g0
148
149	EX(st	%g3, [%o0], and %o1, 7)
150	add	%o0, 4, %o0
1511:
152	be	1f
153	 andcc	%o1, 1, %g0
154
155	EX(sth	%g3, [%o0], and %o1, 3)
156	add	%o0, 2, %o0
1571:
158	bne,a	8f
159	 EX(stb	%g3, [%o0], and %o1, 1)
1608:
161	b	0f
162	 nop
1637:
164	be	13b
165	 orcc	%o1, 0, %g0
166
167	be	0f
1688:
169	 add	%o0, 1, %o0
170	subcc	%o1, 1, %o1
171	bne	8b
172	 EX(stb	%g3, [%o0 - 1], add %o1, 1)
1730:
174	andcc	%g4, 1, %g0
175	be	5f
176	 nop
177	retl
178	 mov	%g1, %o0
1795:
180	retl
181	 clr	%o0
182__memset_end:
183
184	.section .fixup,#alloc,#execinstr
185	.align	4
18620:
187	cmp	%g2, 8
188	bleu	1f
189	 and	%o1, 0x7f, %o1
190	sub	%g2, 9, %g2
191	add	%o3, 64, %o3
1921:
193	sll	%g2, 3, %g2
194	add	%o3, %o1, %o0
195	b 30f
196	 sub	%o0, %g2, %o0
19721:
198	mov	8, %o0
199	and	%o1, 7, %o1
200	sub	%o0, %g2, %o0
201	sll	%o0, 3, %o0
202	b 30f
203	 add	%o0, %o1, %o0
20430:
205/* %o4 is faulting address, %o5 is %pc where fault occurred */
206	save	%sp, -104, %sp
207	mov	%i5, %o0
208	mov	%i7, %o1
209	call	lookup_fault
210	 mov	%i4, %o2
211	ret
212	 restore
213
214	.globl __bzero_end
215__bzero_end:
216