]>
Commit | Line | Data |
---|---|---|
3f02f778 | 1 | /* strchrnul (str, chr) -- Return pointer to first occurrence of CHR in STR |
c4563d2d UD |
2 | or the final NUL byte. |
3 | For Intel 80x86, x>=3. | |
6d7e8eda | 4 | Copyright (C) 1994-2023 Free Software Foundation, Inc. |
c4563d2d | 5 | This file is part of the GNU C Library. |
c4563d2d UD |
6 | |
7 | The GNU C Library is free software; you can redistribute it and/or | |
41bdb6e2 AJ |
8 | modify it under the terms of the GNU Lesser General Public |
9 | License as published by the Free Software Foundation; either | |
10 | version 2.1 of the License, or (at your option) any later version. | |
c4563d2d UD |
11 | |
12 | The GNU C Library is distributed in the hope that it will be useful, | |
13 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
14 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
41bdb6e2 | 15 | Lesser General Public License for more details. |
c4563d2d | 16 | |
41bdb6e2 | 17 | You should have received a copy of the GNU Lesser General Public |
59ba27a6 | 18 | License along with the GNU C Library; if not, see |
5a82c748 | 19 | <https://www.gnu.org/licenses/>. */ |
c4563d2d UD |
20 | |
21 | #include <sysdep.h> | |
22 | #include "asm-syntax.h" | |
23 | ||
2366713d | 24 | #define PARMS 4+4 /* space for 1 saved reg */ |
3f02f778 | 25 | #define RTN PARMS |
2366713d JM |
26 | #define STR RTN |
27 | #define CHR STR+4 | |
c4563d2d UD |
28 | |
29 | .text | |
2366713d | 30 | ENTRY (__strchrnul) |
3f02f778 | 31 | |
c4563d2d | 32 | pushl %edi /* Save callee-safe registers used here. */ |
1ad9da69 UD |
33 | cfi_adjust_cfa_offset (4) |
34 | cfi_rel_offset (edi, 0) | |
c4563d2d | 35 | |
3f02f778 GM |
36 | movl STR(%esp), %eax |
37 | movl CHR(%esp), %edx | |
c4563d2d | 38 | |
3f02f778 GM |
39 | /* At the moment %edx contains CHR. What we need for the |
40 | algorithm is CHR in all bytes of the dword. Avoid | |
c4563d2d UD |
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 | ||
a7123f0e | 59 | testb $3, %al /* correctly aligned ? */ |
c4563d2d UD |
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(6) /* yes => return NULL */ | |
66 | incl %eax /* increment pointer */ | |
67 | ||
a7123f0e | 68 | testb $3, %al /* correctly aligned ? */ |
c4563d2d UD |
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(6) /* yes => return NULL */ | |
75 | incl %eax /* increment pointer */ | |
76 | ||
a7123f0e | 77 | testb $3, %al /* correctly aligned ? */ |
c4563d2d UD |
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(6) /* 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 | ||
3f02f778 | 110 | 3) But wait! Aren't we looking for CHR, not zero? |
c4563d2d | 111 | Good point. So what we do is XOR LONGWORD with a longword, |
3f02f778 | 112 | each of whose bytes is CHR. This turns each byte that is CHR |
c4563d2d UD |
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 | |
3f02f778 | 127 | is *not* CHR */ |
c4563d2d UD |
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 | ||
3f02f778 | 146 | /* If at least one byte of the word is CHR we don't get 0 in %edi. */ |
c4563d2d UD |
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 */ | |
d8b29eea | 158 | jnc L(7) /* highest byte is NUL => return NULL */ |
c4563d2d UD |
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. */ | |
d8b29eea | 163 | jnz L(7) /* found NUL => return NULL */ |
c4563d2d UD |
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 | |
3f02f778 GM |
171 | is *not* CHR */ |
172 | jnc L(71) /* highest byte is CHR => return pointer */ | |
c4563d2d UD |
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(71) /* 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(71) /* 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 | |
3f02f778 GM |
196 | is *not* CHR */ |
197 | jnc L(72) /* highest byte is CHR => return pointer */ | |
c4563d2d UD |
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(72) /* 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(72) /* 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 | |
3f02f778 GM |
221 | is *not* CHR */ |
222 | jnc L(73) /* highest byte is CHR => return pointer */ | |
c4563d2d UD |
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(73) /* 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(73): addl $4, %eax /* adjust pointer */ | |
241 | L(72): addl $4, %eax | |
242 | L(71): addl $4, %eax | |
243 | ||
244 | /* We now scan for the byte in which the character was matched. | |
245 | But we have to take care of the case that a NUL char is | |
246 | found before this in the dword. */ | |
247 | ||
3f02f778 | 248 | L(7): testb %cl, %cl /* is first byte CHR? */ |
c4563d2d UD |
249 | jz L(6) /* yes => return pointer */ |
250 | cmpb %dl, %cl /* is first byte NUL? */ | |
251 | je L(6) /* yes => return NULL */ | |
252 | incl %eax /* it's not in the first byte */ | |
253 | ||
3f02f778 | 254 | testb %ch, %ch /* is second byte CHR? */ |
c4563d2d UD |
255 | jz L(6) /* yes => return pointer */ |
256 | cmpb %dl, %ch /* is second byte NUL? */ | |
257 | je L(6) /* yes => return NULL? */ | |
258 | incl %eax /* it's not in the second byte */ | |
259 | ||
260 | shrl $16, %ecx /* make upper byte accessible */ | |
3f02f778 | 261 | testb %cl, %cl /* is third byte CHR? */ |
c4563d2d UD |
262 | jz L(6) /* yes => return pointer */ |
263 | cmpb %dl, %cl /* is third byte NUL? */ | |
264 | je L(6) /* yes => return NULL */ | |
265 | ||
266 | /* It must be in the fourth byte and it cannot be NUL. */ | |
267 | incl %eax | |
268 | ||
92945b52 | 269 | L(6): popl %edi /* restore saved register content */ |
1ad9da69 UD |
270 | cfi_adjust_cfa_offset (-4) |
271 | cfi_restore (edi) | |
c4563d2d | 272 | |
2366713d JM |
273 | ret |
274 | END (__strchrnul) | |
c4563d2d | 275 | |
7ea51012 | 276 | libc_hidden_def (__strchrnul) |
2366713d | 277 | weak_alias (__strchrnul, strchrnul) |