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 * Returns 0, if ok, and number of bytes not yet set if exception 7 * occurs and we were called as clear_user. 8 */ 9 10#include <asm/cprefix.h> 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 C_LABEL(__bzero), C_LABEL(__memset), 65 .globl C_LABEL(memset) 66 .globl C_LABEL(__memset_start), C_LABEL(__memset_end) 67C_LABEL(__memset_start): 68C_LABEL(__memset): 69C_LABEL(memset): 70 and %o1, 0xff, %g3 71 sll %g3, 8, %g2 72 or %g3, %g2, %g3 73 sll %g3, 16, %g2 74 or %g3, %g2, %g3 75 b 1f 76 mov %o2, %o1 773: 78 cmp %o2, 3 79 be 2f 80 EX(stb %g3, [%o0], sub %o1, 0) 81 82 cmp %o2, 2 83 be 2f 84 EX(stb %g3, [%o0 + 0x01], sub %o1, 1) 85 86 EX(stb %g3, [%o0 + 0x02], sub %o1, 2) 872: 88 sub %o2, 4, %o2 89 add %o1, %o2, %o1 90 b 4f 91 sub %o0, %o2, %o0 92 93C_LABEL(__bzero): 94 mov %g0, %g3 951: 96 cmp %o1, 7 97 bleu 7f 98 andcc %o0, 3, %o2 99 100 bne 3b 1014: 102 andcc %o0, 4, %g0 103 104 be 2f 105 mov %g3, %g2 106 107 EX(st %g3, [%o0], sub %o1, 0) 108 sub %o1, 4, %o1 109 add %o0, 4, %o0 1102: 111 andcc %o1, 0xffffff80, %o3 ! Now everything is 8 aligned and o1 is len to run 112 be 9f 113 andcc %o1, 0x78, %o2 11410: 115 ZERO_BIG_BLOCK(%o0, 0x00, %g2) 116 subcc %o3, 128, %o3 117 ZERO_BIG_BLOCK(%o0, 0x40, %g2) 11811: 119 EXT(10b, 11b, 20f) 120 bne 10b 121 add %o0, 128, %o0 122 123 orcc %o2, %g0, %g0 1249: 125 be 13f 126 andcc %o1, 7, %o1 127 128 srl %o2, 1, %o3 129 set 13f, %o4 130 sub %o4, %o3, %o4 131 jmp %o4 132 add %o0, %o2, %o0 133 13412: 135 ZERO_LAST_BLOCKS(%o0, 0x48, %g2) 136 ZERO_LAST_BLOCKS(%o0, 0x08, %g2) 13713: 138 be 8f 139 andcc %o1, 4, %g0 140 141 be 1f 142 andcc %o1, 2, %g0 143 144 EX(st %g3, [%o0], and %o1, 7) 145 add %o0, 4, %o0 1461: 147 be 1f 148 andcc %o1, 1, %g0 149 150 EX(sth %g3, [%o0], and %o1, 3) 151 add %o0, 2, %o0 1521: 153 bne,a 8f 154 EX(stb %g3, [%o0], and %o1, 1) 1558: 156 retl 157 clr %o0 1587: 159 be 13b 160 orcc %o1, 0, %g0 161 162 be 0f 1638: 164 add %o0, 1, %o0 165 subcc %o1, 1, %o1 166 bne,a 8b 167 EX(stb %g3, [%o0 - 1], add %o1, 1) 1680: 169 retl 170 clr %o0 171C_LABEL(__memset_end): 172 173 .section .fixup,#alloc,#execinstr 174 .align 4 17520: 176 cmp %g2, 8 177 bleu 1f 178 and %o1, 0x7f, %o1 179 sub %g2, 9, %g2 180 add %o3, 64, %o3 1811: 182 sll %g2, 3, %g2 183 add %o3, %o1, %o0 184 b 30f 185 sub %o0, %g2, %o0 18621: 187 mov 8, %o0 188 and %o1, 7, %o1 189 sub %o0, %g2, %o0 190 sll %o0, 3, %o0 191 b 30f 192 add %o0, %o1, %o0 19330: 194/* %o4 is faulting address, %o5 is %pc where fault occurred */ 195 save %sp, -104, %sp 196 mov %i5, %o0 197 mov %i7, %o1 198 call C_LABEL(lookup_fault) 199 mov %i4, %o2 200 ret 201 restore 202 203 .globl __bzero_end 204__bzero_end: 205