lh | 9ed821d | 2023-04-07 01:36:19 -0700 | [diff] [blame^] | 1 | /* strchr (str, ch) -- Return pointer to first occurrence of CH in STR. |
| 2 | For Intel 80x86, x>=3. |
| 3 | Copyright (C) 1994-2015 Free Software Foundation, Inc. |
| 4 | This file is part of the GNU C Library. |
| 5 | Contributed by Ulrich Drepper <drepper@gnu.ai.mit.edu> |
| 6 | Some optimisations by Alan Modra <Alan@SPRI.Levels.UniSA.Edu.Au> |
| 7 | |
| 8 | The GNU C Library is free software; you can redistribute it and/or |
| 9 | modify it under the terms of the GNU Lesser General Public |
| 10 | License as published by the Free Software Foundation; either |
| 11 | version 2.1 of the License, or (at your option) any later version. |
| 12 | |
| 13 | The GNU C Library is distributed in the hope that it will be useful, |
| 14 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 15 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 16 | Lesser General Public License for more details. |
| 17 | |
| 18 | You should have received a copy of the GNU Lesser General Public |
| 19 | License along with the GNU C Library; if not, see |
| 20 | <http://www.gnu.org/licenses/>. */ |
| 21 | |
| 22 | #include <sysdep.h> |
| 23 | #include "asm-syntax.h" |
| 24 | |
| 25 | #define PARMS 4+4 /* space for 1 saved reg */ |
| 26 | #define RTN PARMS |
| 27 | #define STR RTN |
| 28 | #define CHR STR+4 |
| 29 | |
| 30 | .text |
| 31 | ENTRY (strchr) |
| 32 | |
| 33 | pushl %edi /* Save callee-safe registers used here. */ |
| 34 | cfi_adjust_cfa_offset (4) |
| 35 | cfi_rel_offset (edi, 0) |
| 36 | movl STR(%esp), %eax |
| 37 | movl CHR(%esp), %edx |
| 38 | |
| 39 | /* At the moment %edx contains C. What we need for the |
| 40 | algorithm is C in all bytes of the dword. Avoid |
| 41 | operations on 16 bit words because these require an |
| 42 | prefix byte (and one more cycle). */ |
| 43 | movb %dl, %dh /* now it is 0|0|c|c */ |
| 44 | movl %edx, %ecx |
| 45 | shll $16, %edx /* now it is c|c|0|0 */ |
| 46 | movw %cx, %dx /* and finally c|c|c|c */ |
| 47 | |
| 48 | /* Before we start with the main loop we process single bytes |
| 49 | until the source pointer is aligned. This has two reasons: |
| 50 | 1. aligned 32-bit memory access is faster |
| 51 | and (more important) |
| 52 | 2. we process in the main loop 32 bit in one step although |
| 53 | we don't know the end of the string. But accessing at |
| 54 | 4-byte alignment guarantees that we never access illegal |
| 55 | memory if this would not also be done by the trivial |
| 56 | implementation (this is because all processor inherent |
| 57 | boundaries are multiples of 4. */ |
| 58 | |
| 59 | testb $3, %al /* correctly aligned ? */ |
| 60 | jz L(11) /* yes => begin loop */ |
| 61 | movb (%eax), %cl /* load byte in question (we need it twice) */ |
| 62 | cmpb %cl, %dl /* compare byte */ |
| 63 | je L(6) /* target found => return */ |
| 64 | testb %cl, %cl /* is NUL? */ |
| 65 | jz L(2) /* yes => return NULL */ |
| 66 | incl %eax /* increment pointer */ |
| 67 | |
| 68 | testb $3, %al /* correctly aligned ? */ |
| 69 | jz L(11) /* yes => begin loop */ |
| 70 | movb (%eax), %cl /* load byte in question (we need it twice) */ |
| 71 | cmpb %cl, %dl /* compare byte */ |
| 72 | je L(6) /* target found => return */ |
| 73 | testb %cl, %cl /* is NUL? */ |
| 74 | jz L(2) /* yes => return NULL */ |
| 75 | incl %eax /* increment pointer */ |
| 76 | |
| 77 | testb $3, %al /* correctly aligned ? */ |
| 78 | jz L(11) /* yes => begin loop */ |
| 79 | movb (%eax), %cl /* load byte in question (we need it twice) */ |
| 80 | cmpb %cl, %dl /* compare byte */ |
| 81 | je L(6) /* target found => return */ |
| 82 | testb %cl, %cl /* is NUL? */ |
| 83 | jz L(2) /* yes => return NULL */ |
| 84 | incl %eax /* increment pointer */ |
| 85 | |
| 86 | /* No we have reached alignment. */ |
| 87 | jmp L(11) /* begin loop */ |
| 88 | |
| 89 | /* We exit the loop if adding MAGIC_BITS to LONGWORD fails to |
| 90 | change any of the hole bits of LONGWORD. |
| 91 | |
| 92 | 1) Is this safe? Will it catch all the zero bytes? |
| 93 | Suppose there is a byte with all zeros. Any carry bits |
| 94 | propagating from its left will fall into the hole at its |
| 95 | least significant bit and stop. Since there will be no |
| 96 | carry from its most significant bit, the LSB of the |
| 97 | byte to the left will be unchanged, and the zero will be |
| 98 | detected. |
| 99 | |
| 100 | 2) Is this worthwhile? Will it ignore everything except |
| 101 | zero bytes? Suppose every byte of LONGWORD has a bit set |
| 102 | somewhere. There will be a carry into bit 8. If bit 8 |
| 103 | is set, this will carry into bit 16. If bit 8 is clear, |
| 104 | one of bits 9-15 must be set, so there will be a carry |
| 105 | into bit 16. Similarly, there will be a carry into bit |
| 106 | 24. If one of bits 24-31 is set, there will be a carry |
| 107 | into bit 32 (=carry flag), so all of the hole bits will |
| 108 | be changed. |
| 109 | |
| 110 | 3) But wait! Aren't we looking for C, not zero? |
| 111 | Good point. So what we do is XOR LONGWORD with a longword, |
| 112 | each of whose bytes is C. This turns each byte that is C |
| 113 | into a zero. */ |
| 114 | |
| 115 | /* Each round the main loop processes 16 bytes. */ |
| 116 | |
| 117 | ALIGN(4) |
| 118 | |
| 119 | L(1): addl $16, %eax /* adjust pointer for whole round */ |
| 120 | |
| 121 | L(11): movl (%eax), %ecx /* get word (= 4 bytes) in question */ |
| 122 | xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c |
| 123 | are now 0 */ |
| 124 | movl $0xfefefeff, %edi /* magic value */ |
| 125 | addl %ecx, %edi /* add the magic value to the word. We get |
| 126 | carry bits reported for each byte which |
| 127 | is *not* C */ |
| 128 | |
| 129 | /* According to the algorithm we had to reverse the effect of the |
| 130 | XOR first and then test the overflow bits. But because the |
| 131 | following XOR would destroy the carry flag and it would (in a |
| 132 | representation with more than 32 bits) not alter then last |
| 133 | overflow, we can now test this condition. If no carry is signaled |
| 134 | no overflow must have occurred in the last byte => it was 0. */ |
| 135 | jnc L(7) |
| 136 | |
| 137 | /* We are only interested in carry bits that change due to the |
| 138 | previous add, so remove original bits */ |
| 139 | xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */ |
| 140 | |
| 141 | /* Now test for the other three overflow bits. */ |
| 142 | orl $0xfefefeff, %edi /* set all non-carry bits */ |
| 143 | incl %edi /* add 1: if one carry bit was *not* set |
| 144 | the addition will not result in 0. */ |
| 145 | |
| 146 | /* If at least one byte of the word is C we don't get 0 in %edi. */ |
| 147 | jnz L(7) /* found it => return pointer */ |
| 148 | |
| 149 | /* Now we made sure the dword does not contain the character we are |
| 150 | looking for. But because we deal with strings we have to check |
| 151 | for the end of string before testing the next dword. */ |
| 152 | |
| 153 | xorl %edx, %ecx /* restore original dword without reload */ |
| 154 | movl $0xfefefeff, %edi /* magic value */ |
| 155 | addl %ecx, %edi /* add the magic value to the word. We get |
| 156 | carry bits reported for each byte which |
| 157 | is *not* 0 */ |
| 158 | jnc L(2) /* highest byte is NUL => return NULL */ |
| 159 | xorl %ecx, %edi /* (word+magic)^word */ |
| 160 | orl $0xfefefeff, %edi /* set all non-carry bits */ |
| 161 | incl %edi /* add 1: if one carry bit was *not* set |
| 162 | the addition will not result in 0. */ |
| 163 | jnz L(2) /* found NUL => return NULL */ |
| 164 | |
| 165 | movl 4(%eax), %ecx /* get word (= 4 bytes) in question */ |
| 166 | xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c |
| 167 | are now 0 */ |
| 168 | movl $0xfefefeff, %edi /* magic value */ |
| 169 | addl %ecx, %edi /* add the magic value to the word. We get |
| 170 | carry bits reported for each byte which |
| 171 | is *not* C */ |
| 172 | jnc L(71) /* highest byte is C => return pointer */ |
| 173 | xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */ |
| 174 | orl $0xfefefeff, %edi /* set all non-carry bits */ |
| 175 | incl %edi /* add 1: if one carry bit was *not* set |
| 176 | the addition will not result in 0. */ |
| 177 | jnz L(71) /* found it => return pointer */ |
| 178 | xorl %edx, %ecx /* restore original dword without reload */ |
| 179 | movl $0xfefefeff, %edi /* magic value */ |
| 180 | addl %ecx, %edi /* add the magic value to the word. We get |
| 181 | carry bits reported for each byte which |
| 182 | is *not* 0 */ |
| 183 | jnc L(2) /* highest byte is NUL => return NULL */ |
| 184 | xorl %ecx, %edi /* (word+magic)^word */ |
| 185 | orl $0xfefefeff, %edi /* set all non-carry bits */ |
| 186 | incl %edi /* add 1: if one carry bit was *not* set |
| 187 | the addition will not result in 0. */ |
| 188 | jnz L(2) /* found NUL => return NULL */ |
| 189 | |
| 190 | movl 8(%eax), %ecx /* get word (= 4 bytes) in question */ |
| 191 | xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c |
| 192 | are now 0 */ |
| 193 | movl $0xfefefeff, %edi /* magic value */ |
| 194 | addl %ecx, %edi /* add the magic value to the word. We get |
| 195 | carry bits reported for each byte which |
| 196 | is *not* C */ |
| 197 | jnc L(72) /* highest byte is C => return pointer */ |
| 198 | xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */ |
| 199 | orl $0xfefefeff, %edi /* set all non-carry bits */ |
| 200 | incl %edi /* add 1: if one carry bit was *not* set |
| 201 | the addition will not result in 0. */ |
| 202 | jnz L(72) /* found it => return pointer */ |
| 203 | xorl %edx, %ecx /* restore original dword without reload */ |
| 204 | movl $0xfefefeff, %edi /* magic value */ |
| 205 | addl %ecx, %edi /* add the magic value to the word. We get |
| 206 | carry bits reported for each byte which |
| 207 | is *not* 0 */ |
| 208 | jnc L(2) /* highest byte is NUL => return NULL */ |
| 209 | xorl %ecx, %edi /* (word+magic)^word */ |
| 210 | orl $0xfefefeff, %edi /* set all non-carry bits */ |
| 211 | incl %edi /* add 1: if one carry bit was *not* set |
| 212 | the addition will not result in 0. */ |
| 213 | jnz L(2) /* found NUL => return NULL */ |
| 214 | |
| 215 | movl 12(%eax), %ecx /* get word (= 4 bytes) in question */ |
| 216 | xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c |
| 217 | are now 0 */ |
| 218 | movl $0xfefefeff, %edi /* magic value */ |
| 219 | addl %ecx, %edi /* add the magic value to the word. We get |
| 220 | carry bits reported for each byte which |
| 221 | is *not* C */ |
| 222 | jnc L(73) /* highest byte is C => return pointer */ |
| 223 | xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */ |
| 224 | orl $0xfefefeff, %edi /* set all non-carry bits */ |
| 225 | incl %edi /* add 1: if one carry bit was *not* set |
| 226 | the addition will not result in 0. */ |
| 227 | jnz L(73) /* found it => return pointer */ |
| 228 | xorl %edx, %ecx /* restore original dword without reload */ |
| 229 | movl $0xfefefeff, %edi /* magic value */ |
| 230 | addl %ecx, %edi /* add the magic value to the word. We get |
| 231 | carry bits reported for each byte which |
| 232 | is *not* 0 */ |
| 233 | jnc L(2) /* highest byte is NUL => return NULL */ |
| 234 | xorl %ecx, %edi /* (word+magic)^word */ |
| 235 | orl $0xfefefeff, %edi /* set all non-carry bits */ |
| 236 | incl %edi /* add 1: if one carry bit was *not* set |
| 237 | the addition will not result in 0. */ |
| 238 | jz L(1) /* no NUL found => restart loop */ |
| 239 | |
| 240 | L(2): /* Return NULL. */ |
| 241 | xorl %eax, %eax |
| 242 | popl %edi /* restore saved register content */ |
| 243 | cfi_adjust_cfa_offset (-4) |
| 244 | cfi_restore (edi) |
| 245 | |
| 246 | ret |
| 247 | |
| 248 | cfi_adjust_cfa_offset (4) |
| 249 | cfi_rel_offset (edi, 0) |
| 250 | L(73): addl $4, %eax /* adjust pointer */ |
| 251 | L(72): addl $4, %eax |
| 252 | L(71): addl $4, %eax |
| 253 | |
| 254 | /* We now scan for the byte in which the character was matched. |
| 255 | But we have to take care of the case that a NUL char is |
| 256 | found before this in the dword. Note that we XORed %ecx |
| 257 | with the byte we're looking for, therefore the tests below look |
| 258 | reversed. */ |
| 259 | |
| 260 | L(7): testb %cl, %cl /* is first byte C? */ |
| 261 | jz L(6) /* yes => return pointer */ |
| 262 | cmpb %dl, %cl /* is first byte NUL? */ |
| 263 | je L(2) /* yes => return NULL */ |
| 264 | incl %eax /* it's not in the first byte */ |
| 265 | |
| 266 | testb %ch, %ch /* is second byte C? */ |
| 267 | jz L(6) /* yes => return pointer */ |
| 268 | cmpb %dl, %ch /* is second byte NUL? */ |
| 269 | je L(2) /* yes => return NULL? */ |
| 270 | incl %eax /* it's not in the second byte */ |
| 271 | |
| 272 | shrl $16, %ecx /* make upper byte accessible */ |
| 273 | testb %cl, %cl /* is third byte C? */ |
| 274 | jz L(6) /* yes => return pointer */ |
| 275 | cmpb %dl, %cl /* is third byte NUL? */ |
| 276 | je L(2) /* yes => return NULL */ |
| 277 | |
| 278 | /* It must be in the fourth byte and it cannot be NUL. */ |
| 279 | incl %eax |
| 280 | |
| 281 | L(6): |
| 282 | popl %edi /* restore saved register content */ |
| 283 | cfi_adjust_cfa_offset (-4) |
| 284 | cfi_restore (edi) |
| 285 | |
| 286 | ret |
| 287 | END (strchr) |
| 288 | |
| 289 | weak_alias (strchr, index) |
| 290 | libc_hidden_builtin_def (strchr) |