lh | 9ed821d | 2023-04-07 01:36:19 -0700 | [diff] [blame^] | 1 | /* rawmemchr (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 | Optimised a little by Alan Modra <Alan@SPRI.Levels.UniSA.Edu.Au> |
| 7 | This version is developed using the same algorithm as the fast C |
| 8 | version which carries the following introduction: |
| 9 | Based on strlen implementation by Torbjorn Granlund (tege@sics.se), |
| 10 | with help from Dan Sahlin (dan@sics.se) and |
| 11 | commentary by Jim Blandy (jimb@ai.mit.edu); |
| 12 | adaptation to memchr suggested by Dick Karpinski (dick@cca.ucsf.edu), |
| 13 | and implemented by Roland McGrath (roland@ai.mit.edu). |
| 14 | |
| 15 | The GNU C Library is free software; you can redistribute it and/or |
| 16 | modify it under the terms of the GNU Lesser General Public |
| 17 | License as published by the Free Software Foundation; either |
| 18 | version 2.1 of the License, or (at your option) any later version. |
| 19 | |
| 20 | The GNU C Library is distributed in the hope that it will be useful, |
| 21 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 22 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 23 | Lesser General Public License for more details. |
| 24 | |
| 25 | You should have received a copy of the GNU Lesser General Public |
| 26 | License along with the GNU C Library; if not, see |
| 27 | <http://www.gnu.org/licenses/>. */ |
| 28 | |
| 29 | #include <sysdep.h> |
| 30 | #include "asm-syntax.h" |
| 31 | |
| 32 | #define PARMS 4+4 /* space for 1 saved reg */ |
| 33 | #define RTN PARMS |
| 34 | #define STR RTN |
| 35 | #define CHR STR+4 |
| 36 | |
| 37 | .text |
| 38 | ENTRY (__rawmemchr) |
| 39 | |
| 40 | /* Save callee-safe register used in this function. */ |
| 41 | pushl %edi |
| 42 | cfi_adjust_cfa_offset (4) |
| 43 | cfi_rel_offset (edi, 0) |
| 44 | |
| 45 | /* Load parameters into registers. */ |
| 46 | movl STR(%esp), %eax |
| 47 | movl CHR(%esp), %edx |
| 48 | |
| 49 | /* At the moment %edx contains C. What we need for the |
| 50 | algorithm is C in all bytes of the dword. Avoid |
| 51 | operations on 16 bit words because these require an |
| 52 | prefix byte (and one more cycle). */ |
| 53 | movb %dl, %dh /* Now it is 0|0|c|c */ |
| 54 | movl %edx, %ecx |
| 55 | shll $16, %edx /* Now c|c|0|0 */ |
| 56 | movw %cx, %dx /* And finally c|c|c|c */ |
| 57 | |
| 58 | /* Better performance can be achieved if the word (32 |
| 59 | bit) memory access is aligned on a four-byte-boundary. |
| 60 | So process first bytes one by one until boundary is |
| 61 | reached. Don't use a loop for better performance. */ |
| 62 | |
| 63 | testb $3, %al /* correctly aligned ? */ |
| 64 | je L(1) /* yes => begin loop */ |
| 65 | cmpb %dl, (%eax) /* compare byte */ |
| 66 | je L(9) /* target found => return */ |
| 67 | incl %eax /* increment source pointer */ |
| 68 | |
| 69 | testb $3, %al /* correctly aligned ? */ |
| 70 | je L(1) /* yes => begin loop */ |
| 71 | cmpb %dl, (%eax) /* compare byte */ |
| 72 | je L(9) /* target found => return */ |
| 73 | incl %eax /* increment source pointer */ |
| 74 | |
| 75 | testb $3, %al /* correctly aligned ? */ |
| 76 | je L(1) /* yes => begin loop */ |
| 77 | cmpb %dl, (%eax) /* compare byte */ |
| 78 | je L(9) /* target found => return */ |
| 79 | incl %eax /* increment source pointer */ |
| 80 | |
| 81 | /* We exit the loop if adding MAGIC_BITS to LONGWORD fails to |
| 82 | change any of the hole bits of LONGWORD. |
| 83 | |
| 84 | 1) Is this safe? Will it catch all the zero bytes? |
| 85 | Suppose there is a byte with all zeros. Any carry bits |
| 86 | propagating from its left will fall into the hole at its |
| 87 | least significant bit and stop. Since there will be no |
| 88 | carry from its most significant bit, the LSB of the |
| 89 | byte to the left will be unchanged, and the zero will be |
| 90 | detected. |
| 91 | |
| 92 | 2) Is this worthwhile? Will it ignore everything except |
| 93 | zero bytes? Suppose every byte of LONGWORD has a bit set |
| 94 | somewhere. There will be a carry into bit 8. If bit 8 |
| 95 | is set, this will carry into bit 16. If bit 8 is clear, |
| 96 | one of bits 9-15 must be set, so there will be a carry |
| 97 | into bit 16. Similarly, there will be a carry into bit |
| 98 | 24. If one of bits 24-31 is set, there will be a carry |
| 99 | into bit 32 (=carry flag), so all of the hole bits will |
| 100 | be changed. |
| 101 | |
| 102 | 3) But wait! Aren't we looking for C, not zero? |
| 103 | Good point. So what we do is XOR LONGWORD with a longword, |
| 104 | each of whose bytes is C. This turns each byte that is C |
| 105 | into a zero. */ |
| 106 | |
| 107 | |
| 108 | /* Each round the main loop processes 16 bytes. */ |
| 109 | ALIGN (4) |
| 110 | |
| 111 | L(1): movl (%eax), %ecx /* get word (= 4 bytes) in question */ |
| 112 | movl $0xfefefeff, %edi /* magic value */ |
| 113 | xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c |
| 114 | are now 0 */ |
| 115 | addl %ecx, %edi /* add the magic value to the word. We get |
| 116 | carry bits reported for each byte which |
| 117 | is *not* 0 */ |
| 118 | |
| 119 | /* According to the algorithm we had to reverse the effect of the |
| 120 | XOR first and then test the overflow bits. But because the |
| 121 | following XOR would destroy the carry flag and it would (in a |
| 122 | representation with more than 32 bits) not alter then last |
| 123 | overflow, we can now test this condition. If no carry is signaled |
| 124 | no overflow must have occurred in the last byte => it was 0. */ |
| 125 | jnc L(8) |
| 126 | |
| 127 | /* We are only interested in carry bits that change due to the |
| 128 | previous add, so remove original bits */ |
| 129 | xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */ |
| 130 | |
| 131 | /* Now test for the other three overflow bits. */ |
| 132 | orl $0xfefefeff, %edi /* set all non-carry bits */ |
| 133 | incl %edi /* add 1: if one carry bit was *not* set |
| 134 | the addition will not result in 0. */ |
| 135 | |
| 136 | /* If at least one byte of the word is C we don't get 0 in %edi. */ |
| 137 | jnz L(8) /* found it => return pointer */ |
| 138 | |
| 139 | /* This process is unfolded four times for better performance. |
| 140 | we don't increment the source pointer each time. Instead we |
| 141 | use offsets and increment by 16 in each run of the loop. But |
| 142 | before probing for the matching byte we need some extra code |
| 143 | (following LL(13) below). Even the len can be compared with |
| 144 | constants instead of decrementing each time. */ |
| 145 | |
| 146 | movl 4(%eax), %ecx /* get word (= 4 bytes) in question */ |
| 147 | movl $0xfefefeff, %edi /* magic value */ |
| 148 | xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c |
| 149 | are now 0 */ |
| 150 | addl %ecx, %edi /* add the magic value to the word. We get |
| 151 | carry bits reported for each byte which |
| 152 | is *not* 0 */ |
| 153 | jnc L(7) /* highest byte is C => return pointer */ |
| 154 | xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */ |
| 155 | orl $0xfefefeff, %edi /* set all non-carry bits */ |
| 156 | incl %edi /* add 1: if one carry bit was *not* set |
| 157 | the addition will not result in 0. */ |
| 158 | jnz L(7) /* found it => return pointer */ |
| 159 | |
| 160 | movl 8(%eax), %ecx /* get word (= 4 bytes) in question */ |
| 161 | movl $0xfefefeff, %edi /* magic value */ |
| 162 | xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c |
| 163 | are now 0 */ |
| 164 | addl %ecx, %edi /* add the magic value to the word. We get |
| 165 | carry bits reported for each byte which |
| 166 | is *not* 0 */ |
| 167 | jnc L(6) /* highest byte is C => return pointer */ |
| 168 | xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */ |
| 169 | orl $0xfefefeff, %edi /* set all non-carry bits */ |
| 170 | incl %edi /* add 1: if one carry bit was *not* set |
| 171 | the addition will not result in 0. */ |
| 172 | jnz L(6) /* found it => return pointer */ |
| 173 | |
| 174 | movl 12(%eax), %ecx /* get word (= 4 bytes) in question */ |
| 175 | movl $0xfefefeff, %edi /* magic value */ |
| 176 | xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c |
| 177 | are now 0 */ |
| 178 | addl %ecx, %edi /* add the magic value to the word. We get |
| 179 | carry bits reported for each byte which |
| 180 | is *not* 0 */ |
| 181 | jnc L(5) /* highest byte is C => return pointer */ |
| 182 | xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */ |
| 183 | orl $0xfefefeff, %edi /* set all non-carry bits */ |
| 184 | incl %edi /* add 1: if one carry bit was *not* set |
| 185 | the addition will not result in 0. */ |
| 186 | jnz L(5) /* found it => return pointer */ |
| 187 | |
| 188 | /* Adjust both counters for a full round, i.e. 16 bytes. */ |
| 189 | addl $16, %eax |
| 190 | jmp L(1) |
| 191 | /* add missing source pointer increments */ |
| 192 | L(5): addl $4, %eax |
| 193 | L(6): addl $4, %eax |
| 194 | L(7): addl $4, %eax |
| 195 | |
| 196 | /* Test for the matching byte in the word. %ecx contains a NUL |
| 197 | char in the byte which originally was the byte we are looking |
| 198 | at. */ |
| 199 | L(8): testb %cl, %cl /* test first byte in dword */ |
| 200 | jz L(9) /* if zero => return pointer */ |
| 201 | incl %eax /* increment source pointer */ |
| 202 | |
| 203 | testb %ch, %ch /* test second byte in dword */ |
| 204 | jz L(9) /* if zero => return pointer */ |
| 205 | incl %eax /* increment source pointer */ |
| 206 | |
| 207 | testl $0xff0000, %ecx /* test third byte in dword */ |
| 208 | jz L(9) /* if zero => return pointer */ |
| 209 | incl %eax /* increment source pointer */ |
| 210 | |
| 211 | /* No further test needed we we know it is one of the four bytes. */ |
| 212 | |
| 213 | L(9): |
| 214 | popl %edi /* pop saved register */ |
| 215 | cfi_adjust_cfa_offset (-4) |
| 216 | cfi_restore (edi) |
| 217 | |
| 218 | ret |
| 219 | END (__rawmemchr) |
| 220 | |
| 221 | libc_hidden_def (__rawmemchr) |
| 222 | weak_alias (__rawmemchr, rawmemchr) |