1 | /* SPDX-License-Identifier: GPL-2.0 */ |
2 | /* |
3 | * blockops.S: Common block zero optimized routines. |
4 | * |
5 | * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu) |
6 | */ |
7 | |
8 | #include <linux/export.h> |
9 | #include <linux/linkage.h> |
10 | #include <asm/page.h> |
11 | |
12 | /* Zero out 64 bytes of memory at (buf + offset). |
13 | * Assumes %g1 contains zero. |
14 | */ |
15 | #define BLAST_BLOCK(buf, offset) \ |
16 | std %g0, [buf + offset + 0x38]; \ |
17 | std %g0, [buf + offset + 0x30]; \ |
18 | std %g0, [buf + offset + 0x28]; \ |
19 | std %g0, [buf + offset + 0x20]; \ |
20 | std %g0, [buf + offset + 0x18]; \ |
21 | std %g0, [buf + offset + 0x10]; \ |
22 | std %g0, [buf + offset + 0x08]; \ |
23 | std %g0, [buf + offset + 0x00]; |
24 | |
25 | /* Copy 32 bytes of memory at (src + offset) to |
26 | * (dst + offset). |
27 | */ |
28 | #define MIRROR_BLOCK(dst, src, offset, t0, t1, t2, t3, t4, t5, t6, t7) \ |
29 | ldd [src + offset + 0x18], t0; \ |
30 | ldd [src + offset + 0x10], t2; \ |
31 | ldd [src + offset + 0x08], t4; \ |
32 | ldd [src + offset + 0x00], t6; \ |
33 | std t0, [dst + offset + 0x18]; \ |
34 | std t2, [dst + offset + 0x10]; \ |
35 | std t4, [dst + offset + 0x08]; \ |
36 | std t6, [dst + offset + 0x00]; |
37 | |
38 | /* Profiling evidence indicates that memset() is |
39 | * commonly called for blocks of size PAGE_SIZE, |
40 | * and (2 * PAGE_SIZE) (for kernel stacks) |
41 | * and with a second arg of zero. We assume in |
42 | * all of these cases that the buffer is aligned |
43 | * on at least an 8 byte boundary. |
44 | * |
45 | * Therefore we special case them to make them |
46 | * as fast as possible. |
47 | */ |
48 | |
49 | .text |
50 | ENTRY(bzero_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 |
57 | 1: |
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 | ENDPROC(bzero_1page) |
69 | EXPORT_SYMBOL(bzero_1page) |
70 | |
71 | ENTRY(__copy_1page) |
72 | /* NOTE: If you change the number of insns of this routine, please check |
73 | * arch/sparc/mm/hypersparc.S */ |
74 | /* %o0 = dst, %o1 = src */ |
75 | or %g0, (PAGE_SIZE >> 8), %g1 |
76 | 1: |
77 | MIRROR_BLOCK(%o0, %o1, 0x00, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5) |
78 | MIRROR_BLOCK(%o0, %o1, 0x20, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5) |
79 | MIRROR_BLOCK(%o0, %o1, 0x40, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5) |
80 | MIRROR_BLOCK(%o0, %o1, 0x60, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5) |
81 | MIRROR_BLOCK(%o0, %o1, 0x80, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5) |
82 | MIRROR_BLOCK(%o0, %o1, 0xa0, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5) |
83 | MIRROR_BLOCK(%o0, %o1, 0xc0, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5) |
84 | MIRROR_BLOCK(%o0, %o1, 0xe0, %o2, %o3, %o4, %o5, %g2, %g3, %g4, %g5) |
85 | subcc %g1, 1, %g1 |
86 | add %o0, 0x100, %o0 |
87 | bne 1b |
88 | add %o1, 0x100, %o1 |
89 | |
90 | retl |
91 | nop |
92 | ENDPROC(__copy_1page) |
93 | EXPORT_SYMBOL(__copy_1page) |
94 | |