1/*
2 * blockops.S: Common block zero optimized routines.
3 *
4 * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
5 */
6
7#include <asm/page.h>
8
9	/* Zero out 64 bytes of memory at (buf + offset).
10	 * Assumes %g1 contains zero.
11	 */
12#define BLAST_BLOCK(buf, offset) \
13	std	%g0, [buf + offset + 0x38]; \
14	std	%g0, [buf + offset + 0x30]; \
15	std	%g0, [buf + offset + 0x28]; \
16	std	%g0, [buf + offset + 0x20]; \
17	std	%g0, [buf + offset + 0x18]; \
18	std	%g0, [buf + offset + 0x10]; \
19	std	%g0, [buf + offset + 0x08]; \
20	std	%g0, [buf + offset + 0x00];
21
22	/* Copy 32 bytes of memory at (src + offset) to
23	 * (dst + offset).
24	 */
25#define MIRROR_BLOCK(dst, src, offset, t0, t1, t2, t3, t4, t5, t6, t7) \
26	ldd	[src + offset + 0x18], t0; \
27	ldd	[src + offset + 0x10], t2; \
28	ldd	[src + offset + 0x08], t4; \
29	ldd	[src + offset + 0x00], t6; \
30	std	t0, [dst + offset + 0x18]; \
31	std	t2, [dst + offset + 0x10]; \
32	std	t4, [dst + offset + 0x08]; \
33	std	t6, [dst + offset + 0x00];
34
35	/* Profiling evidence indicates that memset() is
36	 * commonly called for blocks of size PAGE_SIZE,
37	 * and (2 * PAGE_SIZE) (for kernel stacks)
38	 * and with a second arg of zero.  We assume in
39	 * all of these cases that the buffer is aligned
40	 * on at least an 8 byte boundary.
41	 *
42	 * Therefore we special case them to make them
43	 * as fast as possible.
44	 */
45
46	.text
47	.align	4
48	.globl	bzero_1page, __copy_1page
49
50bzero_1page:
51/* NOTE: If you change the number of insns of this routine, please check
52 * arch/sparc/mm/hypersparc.S */
53	/* %o0 = buf */
54	or	%g0, %g0, %g1
55	or	%o0, %g0, %o1
56	or	%g0, (PAGE_SIZE >> 8), %g2
571:
58	BLAST_BLOCK(%o0, 0x00)
59	BLAST_BLOCK(%o0, 0x40)
60	BLAST_BLOCK(%o0, 0x80)
61	BLAST_BLOCK(%o0, 0xc0)
62	subcc	%g2, 1, %g2
63	bne	1b
64	 add	%o0, 0x100, %o0
65
66	retl
67	 nop
68
69__copy_1page:
70/* NOTE: If you change the number of insns of this routine, please check
71 * arch/sparc/mm/hypersparc.S */
72	/* %o0 = dst, %o1 = src */
73	or	%g0, (PAGE_SIZE >> 8), %g1
741:
75	MIRROR_BLOCK(%o0, %o1, 0x00, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
76	MIRROR_BLOCK(%o0, %o1, 0x20, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
77	MIRROR_BLOCK(%o0, %o1, 0x40, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
78	MIRROR_BLOCK(%o0, %o1, 0x60, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
79	MIRROR_BLOCK(%o0, %o1, 0x80, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
80	MIRROR_BLOCK(%o0, %o1, 0xa0, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
81	MIRROR_BLOCK(%o0, %o1, 0xc0, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
82	MIRROR_BLOCK(%o0, %o1, 0xe0, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5)
83	subcc	%g1, 1, %g1
84	add	%o0, 0x100, %o0
85	bne	1b
86	 add	%o1, 0x100, %o1
87
88	retl
89	 nop
90