1 /* rawmemchr (str, ch) -- Return pointer to first occurrence of CH in STR.
3 Copyright (C) 1994, 95, 96, 97, 98, 99 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>
8 This version is developed using the same algorithm as the fast C
9 version which carries the following introduction:
11 Based on strlen implementation by Torbjorn Granlund (tege@sics.se),
12 with help from Dan Sahlin (dan@sics.se) and
13 commentary by Jim Blandy (jimb@ai.mit.edu);
14 adaptation to memchr suggested by Dick Karpinski (dick@cca.ucsf.edu),
15 and implemented by Roland McGrath (roland@ai.mit.edu).
17 The GNU C Library is free software; you can redistribute it and/or
18 modify it under the terms of the GNU Library General Public License as
19 published by the Free Software Foundation; either version 2 of the
20 License, or (at your option) any later version.
22 The GNU C Library is distributed in the hope that it will be useful,
23 but WITHOUT ANY WARRANTY; without even the implied warranty of
24 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
25 Library General Public License for more details.
27 You should have received a copy of the GNU Library General Public
28 License along with the GNU C Library; see the file COPYING.LIB. If not,
29 write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
30 Boston, MA 02111-1307, USA. */
33 #include "asm-syntax.h"
43 /* Save callee-safe register used in this function. */
46 /* Load parameters into registers. */
47 movl 8(%esp), %eax /* str: pointer to memory block. */
48 movl 12(%esp), %edx /* c: byte we are looking for. */
50 /* At the moment %edx contains C. What we need for the
51 algorithm is C in all bytes of the dword. Avoid
52 operations on 16 bit words because these require an
53 prefix byte (and one more cycle). */
54 movb %dl, %dh /* Now it is 0|0|c|c */
56 shll $16, %edx /* Now c|c|0|0 */
57 movw %cx, %dx /* And finally c|c|c|c */
59 /* Better performance can be achieved if the word (32
60 bit) memory access is aligned on a four-byte-boundary.
61 So process first bytes one by one until boundary is
62 reached. Don't use a loop for better performance. */
64 testb $3, %al /* correctly aligned ? */
65 je L(1) /* yes => begin loop */
66 cmpb %dl, (%eax) /* compare byte */
67 je L(9) /* target found => return */
68 incl %eax /* increment source pointer */
70 testb $3, %al /* correctly aligned ? */
71 je L(1) /* yes => begin loop */
72 cmpb %dl, (%eax) /* compare byte */
73 je L(9) /* target found => return */
74 incl %eax /* increment source pointer */
76 testb $3, %al /* correctly aligned ? */
77 je L(1) /* yes => begin loop */
78 cmpb %dl, (%eax) /* compare byte */
79 je L(9) /* target found => return */
80 incl %eax /* increment source pointer */
82 /* We exit the loop if adding MAGIC_BITS to LONGWORD fails to
83 change any of the hole bits of LONGWORD.
85 1) Is this safe? Will it catch all the zero bytes?
86 Suppose there is a byte with all zeros. Any carry bits
87 propagating from its left will fall into the hole at its
88 least significant bit and stop. Since there will be no
89 carry from its most significant bit, the LSB of the
90 byte to the left will be unchanged, and the zero will be
93 2) Is this worthwhile? Will it ignore everything except
94 zero bytes? Suppose every byte of LONGWORD has a bit set
95 somewhere. There will be a carry into bit 8. If bit 8
96 is set, this will carry into bit 16. If bit 8 is clear,
97 one of bits 9-15 must be set, so there will be a carry
98 into bit 16. Similarly, there will be a carry into bit
99 24. If one of bits 24-31 is set, there will be a carry
100 into bit 32 (=carry flag), so all of the hole bits will
103 3) But wait! Aren't we looking for C, not zero?
104 Good point. So what we do is XOR LONGWORD with a longword,
105 each of whose bytes is C. This turns each byte that is C
109 /* Each round the main loop processes 16 bytes. */
112 L(1): movl (%eax), %ecx /* get word (= 4 bytes) in question */
113 movl $0xfefefeff, %edi /* magic value */
114 xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c
116 addl %ecx, %edi /* add the magic value to the word. We get
117 carry bits reported for each byte which
120 /* According to the algorithm we had to reverse the effect of the
121 XOR first and then test the overflow bits. But because the
122 following XOR would destroy the carry flag and it would (in a
123 representation with more than 32 bits) not alter then last
124 overflow, we can now test this condition. If no carry is signaled
125 no overflow must have occurred in the last byte => it was 0. */
128 /* We are only interested in carry bits that change due to the
129 previous add, so remove original bits */
130 xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */
132 /* Now test for the other three overflow bits. */
133 orl $0xfefefeff, %edi /* set all non-carry bits */
134 incl %edi /* add 1: if one carry bit was *not* set
135 the addition will not result in 0. */
137 /* If at least one byte of the word is C we don't get 0 in %edi. */
138 jnz L(8) /* found it => return pointer */
140 /* This process is unfolded four times for better performance.
141 we don't increment the source pointer each time. Instead we
142 use offsets and increment by 16 in each run of the loop. But
143 before probing for the matching byte we need some extra code
144 (following LL(13) below). Even the len can be compared with
145 constants instead of decrementing each time. */
147 movl 4(%eax), %ecx /* get word (= 4 bytes) in question */
148 movl $0xfefefeff, %edi /* magic value */
149 xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c
151 addl %ecx, %edi /* add the magic value to the word. We get
152 carry bits reported for each byte which
154 jnc L(7) /* highest byte is C => return pointer */
155 xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */
156 orl $0xfefefeff, %edi /* set all non-carry bits */
157 incl %edi /* add 1: if one carry bit was *not* set
158 the addition will not result in 0. */
159 jnz L(7) /* found it => return pointer */
161 movl 8(%eax), %ecx /* get word (= 4 bytes) in question */
162 movl $0xfefefeff, %edi /* magic value */
163 xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c
165 addl %ecx, %edi /* add the magic value to the word. We get
166 carry bits reported for each byte which
168 jnc L(6) /* highest byte is C => return pointer */
169 xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */
170 orl $0xfefefeff, %edi /* set all non-carry bits */
171 incl %edi /* add 1: if one carry bit was *not* set
172 the addition will not result in 0. */
173 jnz L(6) /* found it => return pointer */
175 movl 12(%eax), %ecx /* get word (= 4 bytes) in question */
176 movl $0xfefefeff, %edi /* magic value */
177 xorl %edx, %ecx /* XOR with word c|c|c|c => bytes of str == c
179 addl %ecx, %edi /* add the magic value to the word. We get
180 carry bits reported for each byte which
182 jnc L(5) /* highest byte is C => return pointer */
183 xorl %ecx, %edi /* ((word^charmask)+magic)^(word^charmask) */
184 orl $0xfefefeff, %edi /* set all non-carry bits */
185 incl %edi /* add 1: if one carry bit was *not* set
186 the addition will not result in 0. */
187 jnz L(5) /* found it => return pointer */
189 /* Adjust both counters for a full round, i.e. 16 bytes. */
192 /* add missing source pointer increments */
197 /* Test for the matching byte in the word. %ecx contains a NUL
198 char in the byte which originally was the byte we are looking
200 L(8): testb %cl, %cl /* test first byte in dword */
201 jz L(9) /* if zero => return pointer */
202 incl %eax /* increment source pointer */
204 testb %ch, %ch /* test second byte in dword */
205 jz L(9) /* if zero => return pointer */
206 incl %eax /* increment source pointer */
208 testl $0xff0000, %ecx /* test third byte in dword */
209 jz L(9) /* if zero => return pointer */
210 incl %eax /* increment source pointer */
212 /* No further test needed we we know it is one of the four bytes. */
214 L(9): popl %edi /* pop saved register */
218 weak_alias (__rawmemchr, rawmemchr)