1 | /* Optimized strrchr implementation for PowerPC64/POWER7 using cmpb insn. |
2 | Copyright (C) 2014-2024 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 | /* int [r3] strrchr (char *s [r3], int c [r4]) */ |
22 | |
23 | #ifndef STRRCHR |
24 | # define STRRCHR strrchr |
25 | #endif |
26 | |
27 | .machine power7 |
28 | ENTRY_TOCLESS (STRRCHR) |
29 | CALL_MCOUNT 2 |
30 | dcbt 0,r3 |
31 | clrrdi r8,r3,3 /* Align the address to doubleword boundary. */ |
32 | cmpdi cr7,r4,0 |
33 | ld r12,0(r8) /* Load doubleword from memory. */ |
34 | li r9,0 /* used to store last occurrence */ |
35 | li r0,0 /* Doubleword with null chars to use |
36 | with cmpb. */ |
37 | |
38 | rlwinm r6,r3,3,26,28 /* Calculate padding. */ |
39 | |
40 | beq cr7,L(null_match) |
41 | |
42 | /* Replicate byte to doubleword. */ |
43 | insrdi r4,r4,8,48 |
44 | insrdi r4,r4,16,32 |
45 | insrdi r4,r4,32,0 |
46 | |
47 | /* r4 is changed now ,if its passed as more chars |
48 | check for null again */ |
49 | cmpdi cr7,r4,0 |
50 | beq cr7,L(null_match) |
51 | /* Now r4 has a doubleword of c bytes and r0 has |
52 | a doubleword of null bytes. */ |
53 | |
54 | cmpb r10,r12,r4 /* Compare each byte against c byte. */ |
55 | cmpb r11,r12,r0 /* Compare each byte against null byte. */ |
56 | |
57 | /* Move the doublewords left and right to discard the bits that are |
58 | not part of the string and bring them back as zeros. */ |
59 | #ifdef __LITTLE_ENDIAN__ |
60 | srd r10,r10,r6 |
61 | srd r11,r11,r6 |
62 | sld r10,r10,r6 |
63 | sld r11,r11,r6 |
64 | #else |
65 | sld r10,r10,r6 |
66 | sld r11,r11,r6 |
67 | srd r10,r10,r6 |
68 | srd r11,r11,r6 |
69 | #endif |
70 | or r5,r10,r11 /* OR the results to speed things up. */ |
71 | cmpdi cr7,r5,0 /* If r5 == 0, no c or null bytes |
72 | have been found. */ |
73 | bne cr7,L(done) |
74 | |
75 | L(align): |
76 | mtcrf 0x01,r8 |
77 | |
78 | /* Are we now aligned to a doubleword boundary? If so, skip to |
79 | the main loop. Otherwise, go through the alignment code. */ |
80 | |
81 | bt 28,L(loop) |
82 | |
83 | /* Handle WORD2 of pair. */ |
84 | ldu r12,8(r8) |
85 | cmpb r10,r12,r4 |
86 | cmpb r11,r12,r0 |
87 | or r5,r10,r11 |
88 | cmpdi cr7,r5,0 |
89 | bne cr7,L(done) |
90 | b L(loop) /* We branch here (rather than falling through) |
91 | to skip the nops due to heavy alignment |
92 | of the loop below. */ |
93 | .p2align 5 |
94 | L(loop): |
95 | /* Load two doublewords, compare and merge in a |
96 | single register for speed. This is an attempt |
97 | to speed up the null-checking process for bigger strings. */ |
98 | ld r12,8(r8) |
99 | ldu r7,16(r8) |
100 | cmpb r10,r12,r4 |
101 | cmpb r11,r12,r0 |
102 | cmpb r6,r7,r4 |
103 | cmpb r7,r7,r0 |
104 | or r12,r10,r11 |
105 | or r5,r6,r7 |
106 | or r5,r12,r5 |
107 | cmpdi cr7,r5,0 |
108 | beq cr7,L(loop) |
109 | |
110 | /* OK, one (or both) of the doublewords contains a c/null byte. Check |
111 | the first doubleword and decrement the address in case the first |
112 | doubleword really contains a c/null byte. */ |
113 | cmpdi cr6,r12,0 |
114 | addi r8,r8,-8 |
115 | bne cr6,L(done) |
116 | |
117 | /* The c/null byte must be in the second doubleword. Adjust the |
118 | address again and move the result of cmpb to r10 so we can calculate |
119 | the pointer. */ |
120 | |
121 | mr r10,r6 |
122 | mr r11,r7 |
123 | addi r8,r8,8 |
124 | |
125 | /* r10/r11 have the output of the cmpb instructions, that is, |
126 | 0xff in the same position as the c/null byte in the original |
127 | doubleword from the string. Use that to calculate the pointer. */ |
128 | |
129 | L(done): |
130 | /* if there are more than one 0xff in r11, find the first pos of ff |
131 | in r11 and fill r10 with 0 from that position */ |
132 | cmpdi cr7,r11,0 |
133 | beq cr7,L(no_null) |
134 | #ifdef __LITTLE_ENDIAN__ |
135 | addi r3,r11,-1 |
136 | andc r3,r3,r11 |
137 | popcntd r0,r3 |
138 | #else |
139 | cntlzd r0,r11 |
140 | #endif |
141 | subfic r0,r0,63 |
142 | li r6,-1 |
143 | #ifdef __LITTLE_ENDIAN__ |
144 | srd r0,r6,r0 |
145 | #else |
146 | sld r0,r6,r0 |
147 | #endif |
148 | and r10,r0,r10 |
149 | L(no_null): |
150 | #ifdef __LITTLE_ENDIAN__ |
151 | cntlzd r0,r10 /* Count leading zeros before c matches. */ |
152 | addi r3,r10,-1 |
153 | andc r3,r3,r10 |
154 | addi r10,r11,-1 |
155 | andc r10,r10,r11 |
156 | cmpld cr7,r3,r10 |
157 | bgt cr7,L(no_match) |
158 | #else |
159 | addi r3,r10,-1 /* Count trailing zeros before c matches. */ |
160 | andc r3,r3,r10 |
161 | popcntd r0,r3 |
162 | cmpld cr7,r11,r10 |
163 | bgt cr7,L(no_match) |
164 | #endif |
165 | srdi r0,r0,3 /* Convert trailing zeros to bytes. */ |
166 | subfic r0,r0,7 |
167 | add r9,r8,r0 /* Return address of the matching c byte |
168 | or null in case c was not found. */ |
169 | li r0,0 |
170 | cmpdi cr7,r11,0 /* If r11 == 0, no null's have been found. */ |
171 | beq cr7,L(align) |
172 | |
173 | .align 4 |
174 | L(no_match): |
175 | mr r3,r9 |
176 | blr |
177 | |
178 | /* We are here because strrchr was called with a null byte. */ |
179 | .align 4 |
180 | L(null_match): |
181 | /* r0 has a doubleword of null bytes. */ |
182 | |
183 | cmpb r5,r12,r0 /* Compare each byte against null bytes. */ |
184 | |
185 | /* Move the doublewords left and right to discard the bits that are |
186 | not part of the string and bring them back as zeros. */ |
187 | #ifdef __LITTLE_ENDIAN__ |
188 | srd r5,r5,r6 |
189 | sld r5,r5,r6 |
190 | #else |
191 | sld r5,r5,r6 |
192 | srd r5,r5,r6 |
193 | #endif |
194 | cmpdi cr7,r5,0 /* If r10 == 0, no c or null bytes |
195 | have been found. */ |
196 | bne cr7,L(done_null) |
197 | |
198 | mtcrf 0x01,r8 |
199 | |
200 | /* Are we now aligned to a quadword boundary? If so, skip to |
201 | the main loop. Otherwise, go through the alignment code. */ |
202 | |
203 | bt 28,L(loop_null) |
204 | |
205 | /* Handle WORD2 of pair. */ |
206 | ldu r12,8(r8) |
207 | cmpb r5,r12,r0 |
208 | cmpdi cr7,r5,0 |
209 | bne cr7,L(done_null) |
210 | b L(loop_null) /* We branch here (rather than falling through) |
211 | to skip the nops due to heavy alignment |
212 | of the loop below. */ |
213 | |
214 | /* Main loop to look for the end of the string. Since it's a |
215 | small loop (< 8 instructions), align it to 32-bytes. */ |
216 | .p2align 5 |
217 | L(loop_null): |
218 | /* Load two doublewords, compare and merge in a |
219 | single register for speed. This is an attempt |
220 | to speed up the null-checking process for bigger strings. */ |
221 | ld r12,8(r8) |
222 | ldu r11,16(r8) |
223 | cmpb r5,r12,r0 |
224 | cmpb r10,r11,r0 |
225 | or r6,r5,r10 |
226 | cmpdi cr7,r6,0 |
227 | beq cr7,L(loop_null) |
228 | |
229 | /* OK, one (or both) of the doublewords contains a null byte. Check |
230 | the first doubleword and decrement the address in case the first |
231 | doubleword really contains a null byte. */ |
232 | |
233 | cmpdi cr6,r5,0 |
234 | addi r8,r8,-8 |
235 | bne cr6,L(done_null) |
236 | |
237 | /* The null byte must be in the second doubleword. Adjust the address |
238 | again and move the result of cmpb to r10 so we can calculate the |
239 | pointer. */ |
240 | |
241 | mr r5,r10 |
242 | addi r8,r8,8 |
243 | |
244 | /* r5 has the output of the cmpb instruction, that is, it contains |
245 | 0xff in the same position as the null byte in the original |
246 | doubleword from the string. Use that to calculate the pointer. */ |
247 | L(done_null): |
248 | #ifdef __LITTLE_ENDIAN__ |
249 | addi r0,r5,-1 |
250 | andc r0,r0,r5 |
251 | popcntd r0,r0 |
252 | #else |
253 | cntlzd r0,r5 /* Count leading zeros before the match. */ |
254 | #endif |
255 | srdi r0,r0,3 /* Convert trailing zeros to bytes. */ |
256 | add r3,r8,r0 /* Return address of the matching null byte. */ |
257 | blr |
258 | END (STRRCHR) |
259 | weak_alias (strrchr, rindex) |
260 | libc_hidden_builtin_def (strrchr) |
261 | |