xref: /openbmc/linux/arch/sparc/lib/memset.S (revision 8684014d) 
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
13/* Work around cpp -rob */
14#define ALLOC #alloc
15#define EXECINSTR #execinstr
16#define EX(x,y,a,b) 				\
1798: 	x,y;					\
18	.section .fixup,ALLOC,EXECINSTR;	\
19	.align	4;				\
2099:	ba 30f;					\
21	 a, b, %o0;				\
22	.section __ex_table,ALLOC;		\
23	.align	4;				\
24	.word	98b, 99b;			\
25	.text;					\
26	.align	4
27
28#define EXT(start,end,handler) 			\
29	.section __ex_table,ALLOC;		\
30	.align	4;				\
31	.word	start, 0, end, handler;		\
32	.text;					\
33	.align	4
34
35/* Please don't change these macros, unless you change the logic
36 * in the .fixup section below as well.
37 * Store 64 bytes at (BASE + OFFSET) using value SOURCE. */
38#define ZERO_BIG_BLOCK(base, offset, source)    \
39	std	source, [base + offset + 0x00]; \
40	std	source, [base + offset + 0x08]; \
41	std	source, [base + offset + 0x10]; \
42	std	source, [base + offset + 0x18]; \
43	std	source, [base + offset + 0x20]; \
44	std	source, [base + offset + 0x28]; \
45	std	source, [base + offset + 0x30]; \
46	std	source, [base + offset + 0x38];
47
48#define ZERO_LAST_BLOCKS(base, offset, source)	\
49	std	source, [base - offset - 0x38]; \
50	std	source, [base - offset - 0x30]; \
51	std	source, [base - offset - 0x28]; \
52	std	source, [base - offset - 0x20]; \
53	std	source, [base - offset - 0x18]; \
54	std	source, [base - offset - 0x10]; \
55	std	source, [base - offset - 0x08]; \
56	std	source, [base - offset - 0x00];
57
58	.text
59	.align 4
60
61        .globl  __bzero_begin
62__bzero_begin:
63
64	.globl	__bzero
65	.globl	memset
66	.globl	__memset_start, __memset_end
67__memset_start:
68memset:
69	mov	%o0, %g1
70	mov	1, %g4
71	and	%o1, 0xff, %g3
72	sll	%g3, 8, %g2
73	or	%g3, %g2, %g3
74	sll	%g3, 16, %g2
75	or	%g3, %g2, %g3
76	b	1f
77	 mov	%o2, %o1
783:
79	cmp	%o2, 3
80	be	2f
81	 EX(stb	%g3, [%o0], sub %o1, 0)
82
83	cmp	%o2, 2
84	be	2f
85	 EX(stb	%g3, [%o0 + 0x01], sub %o1, 1)
86
87	EX(stb	%g3, [%o0 + 0x02], sub %o1, 2)
882:
89	sub	%o2, 4, %o2
90	add	%o1, %o2, %o1
91	b	4f
92	 sub	%o0, %o2, %o0
93
94__bzero:
95	clr	%g4
96	mov	%g0, %g3
971:
98	cmp	%o1, 7
99	bleu	7f
100	 andcc	%o0, 3, %o2
101
102	bne	3b
1034:
104	 andcc	%o0, 4, %g0
105
106	be	2f
107	 mov	%g3, %g2
108
109	EX(st	%g3, [%o0], sub %o1, 0)
110	sub	%o1, 4, %o1
111	add	%o0, 4, %o0
1122:
113	andcc	%o1, 0xffffff80, %o3	! Now everything is 8 aligned and o1 is len to run
114	be	9f
115	 andcc	%o1, 0x78, %o2
11610:
117	ZERO_BIG_BLOCK(%o0, 0x00, %g2)
118	subcc	%o3, 128, %o3
119	ZERO_BIG_BLOCK(%o0, 0x40, %g2)
12011:
121	EXT(10b, 11b, 20f)
122	bne	10b
123	 add	%o0, 128, %o0
124
125	orcc	%o2, %g0, %g0
1269:
127	be	13f
128	 andcc	%o1, 7, %o1
129
130	srl	%o2, 1, %o3
131	set	13f, %o4
132	sub	%o4, %o3, %o4
133	jmp	%o4
134	 add	%o0, %o2, %o0
135
13612:
137	ZERO_LAST_BLOCKS(%o0, 0x48, %g2)
138	ZERO_LAST_BLOCKS(%o0, 0x08, %g2)
13913:
140	be	8f
141	 andcc	%o1, 4, %g0
142
143	be	1f
144	 andcc	%o1, 2, %g0
145
146	EX(st	%g3, [%o0], and %o1, 7)
147	add	%o0, 4, %o0
1481:
149	be	1f
150	 andcc	%o1, 1, %g0
151
152	EX(sth	%g3, [%o0], and %o1, 3)
153	add	%o0, 2, %o0
1541:
155	bne,a	8f
156	 EX(stb	%g3, [%o0], and %o1, 1)
1578:
158	b	0f
159	 nop
1607:
161	be	13b
162	 orcc	%o1, 0, %g0
163
164	be	0f
1658:
166	 add	%o0, 1, %o0
167	subcc	%o1, 1, %o1
168	bne	8b
169	 EX(stb	%g3, [%o0 - 1], add %o1, 1)
1700:
171	andcc	%g4, 1, %g0
172	be	5f
173	 nop
174	retl
175	 mov	%g1, %o0
1765:
177	retl
178	 clr	%o0
179__memset_end:
180
181	.section .fixup,#alloc,#execinstr
182	.align	4
18320:
184	cmp	%g2, 8
185	bleu	1f
186	 and	%o1, 0x7f, %o1
187	sub	%g2, 9, %g2
188	add	%o3, 64, %o3
1891:
190	sll	%g2, 3, %g2
191	add	%o3, %o1, %o0
192	b 30f
193	 sub	%o0, %g2, %o0
19421:
195	mov	8, %o0
196	and	%o1, 7, %o1
197	sub	%o0, %g2, %o0
198	sll	%o0, 3, %o0
199	b 30f
200	 add	%o0, %o1, %o0
20130:
202/* %o4 is faulting address, %o5 is %pc where fault occurred */
203	save	%sp, -104, %sp
204	mov	%i5, %o0
205	mov	%i7, %o1
206	call	lookup_fault
207	 mov	%i4, %o2
208	ret
209	 restore
210
211	.globl __bzero_end
212__bzero_end:
213