1 | /* Optimized strchr implementation for PowerPC64/POWER7 using cmpb insn. |
2 | Copyright (C) 2010-2024 Free Software Foundation, Inc. |
3 | This file is part of the GNU C Library. |
4 | |
5 | The GNU C Library is free software; you can redistribute it and/or |
6 | modify it under the terms of the GNU Lesser General Public |
7 | License as published by the Free Software Foundation; either |
8 | version 2.1 of the License, or (at your option) any later version. |
9 | |
10 | The GNU C Library is distributed in the hope that it will be useful, |
11 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
13 | Lesser General Public License for more details. |
14 | |
15 | You should have received a copy of the GNU Lesser General Public |
16 | License along with the GNU C Library; if not, see |
17 | <https://www.gnu.org/licenses/>. */ |
18 | |
19 | #include <sysdep.h> |
20 | |
21 | #ifndef STRCHR |
22 | # define STRCHR strchr |
23 | #endif |
24 | |
25 | /* int [r3] strchr (char *s [r3], int c [r4]) */ |
26 | .machine power7 |
27 | ENTRY_TOCLESS (STRCHR) |
28 | CALL_MCOUNT 2 |
29 | dcbt 0,r3 |
30 | clrrdi r8,r3,3 /* Align the address to doubleword boundary. */ |
31 | cmpdi cr7,r4,0 |
32 | ld r12,0(r8) /* Load doubleword from memory. */ |
33 | li r0,0 /* Doubleword with null chars to use |
34 | with cmpb. */ |
35 | |
36 | rlwinm r6,r3,3,26,28 /* Calculate padding. */ |
37 | |
38 | beq cr7,L(null_match) |
39 | |
40 | /* Replicate byte to doubleword. */ |
41 | insrdi r4,r4,8,48 |
42 | insrdi r4,r4,16,32 |
43 | insrdi r4,r4,32,0 |
44 | |
45 | /* Now r4 has a doubleword of c bytes and r0 has |
46 | a doubleword of null bytes. */ |
47 | |
48 | cmpb r10,r12,r4 /* Compare each byte against c byte. */ |
49 | cmpb r11,r12,r0 /* Compare each byte against null byte. */ |
50 | |
51 | /* Move the doublewords left and right to discard the bits that are |
52 | not part of the string and bring them back as zeros. */ |
53 | #ifdef __LITTLE_ENDIAN__ |
54 | srd r10,r10,r6 |
55 | srd r11,r11,r6 |
56 | sld r10,r10,r6 |
57 | sld r11,r11,r6 |
58 | #else |
59 | sld r10,r10,r6 |
60 | sld r11,r11,r6 |
61 | srd r10,r10,r6 |
62 | srd r11,r11,r6 |
63 | #endif |
64 | or r5,r10,r11 /* OR the results to speed things up. */ |
65 | cmpdi cr7,r5,0 /* If r5 == 0, no c or null bytes |
66 | have been found. */ |
67 | bne cr7,L(done) |
68 | |
69 | mtcrf 0x01,r8 |
70 | |
71 | /* Are we now aligned to a doubleword boundary? If so, skip to |
72 | the main loop. Otherwise, go through the alignment code. */ |
73 | |
74 | bt 28,L(loop) |
75 | |
76 | /* Handle WORD2 of pair. */ |
77 | ldu r12,8(r8) |
78 | cmpb r10,r12,r4 |
79 | cmpb r11,r12,r0 |
80 | or r5,r10,r11 |
81 | cmpdi cr7,r5,0 |
82 | bne cr7,L(done) |
83 | b L(loop) /* We branch here (rather than falling through) |
84 | to skip the nops due to heavy alignment |
85 | of the loop below. */ |
86 | |
87 | .p2align 5 |
88 | L(loop): |
89 | /* Load two doublewords, compare and merge in a |
90 | single register for speed. This is an attempt |
91 | to speed up the null-checking process for bigger strings. */ |
92 | ld r12,8(r8) |
93 | ldu r9,16(r8) |
94 | cmpb r10,r12,r4 |
95 | cmpb r11,r12,r0 |
96 | cmpb r6,r9,r4 |
97 | cmpb r7,r9,r0 |
98 | or r12,r10,r11 |
99 | or r9,r6,r7 |
100 | or r5,r12,r9 |
101 | cmpdi cr7,r5,0 |
102 | beq cr7,L(loop) |
103 | |
104 | /* OK, one (or both) of the doublewords contains a c/null byte. Check |
105 | the first doubleword and decrement the address in case the first |
106 | doubleword really contains a c/null byte. */ |
107 | |
108 | cmpdi cr6,r12,0 |
109 | addi r8,r8,-8 |
110 | bne cr6,L(done) |
111 | |
112 | /* The c/null byte must be in the second doubleword. Adjust the |
113 | address again and move the result of cmpb to r10 so we can calculate |
114 | the pointer. */ |
115 | |
116 | mr r10,r6 |
117 | mr r11,r7 |
118 | addi r8,r8,8 |
119 | |
120 | /* r10/r11 have the output of the cmpb instructions, that is, |
121 | 0xff in the same position as the c/null byte in the original |
122 | doubleword from the string. Use that to calculate the pointer. */ |
123 | L(done): |
124 | #ifdef __LITTLE_ENDIAN__ |
125 | addi r3,r10,-1 |
126 | andc r3,r3,r10 |
127 | popcntd r0,r3 |
128 | addi r4,r11,-1 |
129 | andc r4,r4,r11 |
130 | cmpld cr7,r3,r4 |
131 | bgt cr7,L(no_match) |
132 | #else |
133 | cntlzd r0,r10 /* Count leading zeros before c matches. */ |
134 | cmpld cr7,r11,r10 |
135 | bgt cr7,L(no_match) |
136 | #endif |
137 | srdi r0,r0,3 /* Convert leading zeros to bytes. */ |
138 | add r3,r8,r0 /* Return address of the matching c byte |
139 | or null in case c was not found. */ |
140 | blr |
141 | |
142 | .align 4 |
143 | L(no_match): |
144 | li r3,0 |
145 | blr |
146 | |
147 | /* We are here because strchr was called with a null byte. */ |
148 | .align 4 |
149 | L(null_match): |
150 | /* r0 has a doubleword of null bytes. */ |
151 | |
152 | cmpb r5,r12,r0 /* Compare each byte against null bytes. */ |
153 | |
154 | /* Move the doublewords left and right to discard the bits that are |
155 | not part of the string and bring them back as zeros. */ |
156 | #ifdef __LITTLE_ENDIAN__ |
157 | srd r5,r5,r6 |
158 | sld r5,r5,r6 |
159 | #else |
160 | sld r5,r5,r6 |
161 | srd r5,r5,r6 |
162 | #endif |
163 | cmpdi cr7,r5,0 /* If r10 == 0, no c or null bytes |
164 | have been found. */ |
165 | bne cr7,L(done_null) |
166 | |
167 | mtcrf 0x01,r8 |
168 | |
169 | /* Are we now aligned to a quadword boundary? If so, skip to |
170 | the main loop. Otherwise, go through the alignment code. */ |
171 | |
172 | bt 28,L(loop_null) |
173 | |
174 | /* Handle WORD2 of pair. */ |
175 | ldu r12,8(r8) |
176 | cmpb r5,r12,r0 |
177 | cmpdi cr7,r5,0 |
178 | bne cr7,L(done_null) |
179 | b L(loop_null) /* We branch here (rather than falling through) |
180 | to skip the nops due to heavy alignment |
181 | of the loop below. */ |
182 | |
183 | /* Main loop to look for the end of the string. Since it's a |
184 | small loop (< 8 instructions), align it to 32-bytes. */ |
185 | .p2align 5 |
186 | L(loop_null): |
187 | /* Load two doublewords, compare and merge in a |
188 | single register for speed. This is an attempt |
189 | to speed up the null-checking process for bigger strings. */ |
190 | ld r12,8(r8) |
191 | ldu r11,16(r8) |
192 | cmpb r5,r12,r0 |
193 | cmpb r10,r11,r0 |
194 | or r6,r5,r10 |
195 | cmpdi cr7,r6,0 |
196 | beq cr7,L(loop_null) |
197 | |
198 | /* OK, one (or both) of the doublewords contains a null byte. Check |
199 | the first doubleword and decrement the address in case the first |
200 | doubleword really contains a null byte. */ |
201 | |
202 | cmpdi cr6,r5,0 |
203 | addi r8,r8,-8 |
204 | bne cr6,L(done_null) |
205 | |
206 | /* The null byte must be in the second doubleword. Adjust the address |
207 | again and move the result of cmpb to r10 so we can calculate the |
208 | pointer. */ |
209 | |
210 | mr r5,r10 |
211 | addi r8,r8,8 |
212 | |
213 | /* r5 has the output of the cmpb instruction, that is, it contains |
214 | 0xff in the same position as the null byte in the original |
215 | doubleword from the string. Use that to calculate the pointer. */ |
216 | L(done_null): |
217 | #ifdef __LITTLE_ENDIAN__ |
218 | addi r0,r5,-1 |
219 | andc r0,r0,r5 |
220 | popcntd r0,r0 |
221 | #else |
222 | cntlzd r0,r5 /* Count leading zeros before the match. */ |
223 | #endif |
224 | srdi r0,r0,3 /* Convert leading zeros to bytes. */ |
225 | add r3,r8,r0 /* Return address of the matching null byte. */ |
226 | blr |
227 | END (STRCHR) |
228 | weak_alias (strchr, index) |
229 | libc_hidden_builtin_def (strchr) |
230 | |