1 | /* Vector optimized 32/64 bit S/390 version of wcsncat. |
2 | Copyright (C) 2015-2022 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 <ifunc-wcsncat.h> |
20 | #if HAVE_WCSNCAT_Z13 |
21 | |
22 | # include "sysdep.h" |
23 | # include "asm-syntax.h" |
24 | |
25 | .text |
26 | |
27 | /* wchar_t * wcsncat (wchar_t *dest, const wchar_t *src, size_t n) |
28 | Concatenate two strings - at most n characters of src. |
29 | |
30 | Register usage: |
31 | -r0=saved dest pointer for return |
32 | -r1=tmp |
33 | -r2=dest |
34 | -r3=src |
35 | -r4=n |
36 | -r5=current_len |
37 | -r6=tmp |
38 | -r7=tmp |
39 | -v16=part of src |
40 | -v17=index of zero |
41 | -v18=part of src |
42 | -v31=register save area for r6, r7 |
43 | */ |
44 | ENTRY(WCSNCAT_Z13) |
45 | .machine "z13" |
46 | .machinemode "zarch_nohighgprs" |
47 | |
48 | # if !defined __s390x__ |
49 | llgfr %r4,%r4 |
50 | # endif /* !defined __s390x__ */ |
51 | |
52 | clgfi %r4,0 |
53 | ber %r14 /* Nothing to do, if n == 0. */ |
54 | |
55 | vlbb %v16,0(%r2),6 /* Load s until next 4k-byte boundary. */ |
56 | lcbb %r1,0(%r2),6 /* Get bytes to 4k-byte boundary or 16. */ |
57 | |
58 | /* If either src or dest is not 4byte aligned, use __wcsncat_c. */ |
59 | tmll %r2,3 /* Test if s is 4-byte aligned? */ |
60 | jne .Lfallback /* And use common-code variant if not. */ |
61 | tmll %r3,3 /* Test if src is 4-byte aligned? */ |
62 | jne .Lfallback /* And use common-code variant if not. */ |
63 | |
64 | lgr %r0,%r2 /* Save destination pointer for return. */ |
65 | vlvgp %v31,%r6,%r7 /* Save registers. */ |
66 | |
67 | /* WCSLEN |
68 | %r1 = loaded bytes (tmp) |
69 | %r6 = zero byte index (tmp) |
70 | %r2 = dst |
71 | */ |
72 | vfenezf %v16,%v16,%v16 /* Find element not equal with zero search. */ |
73 | vlgvb %r5,%v16,7 /* Load zero index or 16 if not found. */ |
74 | clrjl %r5,%r1,.Llen_end /* Found zero within loaded bytes, end. */ |
75 | |
76 | /* Align s to 16 byte. */ |
77 | risbgn %r1,%r2,60,128+63,0 /* %r3 = bits 60-63 of %r2 'and' 15. */ |
78 | lghi %r5,16 /* current_len = 16. */ |
79 | slr %r5,%r1 /* Compute bytes to 16bytes boundary. */ |
80 | |
81 | /* Find zero in 16byte aligned loop. */ |
82 | .Llen_loop: |
83 | vl %v16,0(%r5,%r2) /* Load s. */ |
84 | vfenezfs %v16,%v16,%v16 /* Find element not equal with zero search. */ |
85 | je .Llen_found /* Jump away if zero was found. */ |
86 | vl %v16,16(%r5,%r2) |
87 | vfenezfs %v16,%v16,%v16 |
88 | je .Llen_found16 |
89 | vl %v16,32(%r5,%r2) |
90 | vfenezfs %v16,%v16,%v16 |
91 | je .Llen_found32 |
92 | vl %v16,48(%r5,%r2) |
93 | vfenezfs %v16,%v16,%v16 |
94 | je .Llen_found48 |
95 | |
96 | aghi %r5,64 |
97 | j .Llen_loop /* No zero -> loop. */ |
98 | |
99 | .Llen_found48: |
100 | aghi %r5,16 |
101 | .Llen_found32: |
102 | aghi %r5,16 |
103 | .Llen_found16: |
104 | aghi %r5,16 |
105 | .Llen_found: |
106 | vlgvb %r1,%v16,7 /* Load byte index of zero. */ |
107 | algr %r5,%r1 |
108 | |
109 | .Llen_end: |
110 | /* WCSNCPY |
111 | %r1 = zero byte index (tmp) |
112 | %r6 = loaded bytes (tmp) |
113 | %r3 = curr src pointer |
114 | %r2 = curr dst pointer |
115 | %r7 = border, tmp |
116 | */ |
117 | la %r2,0(%r5,%r2) /* strcpy at end of dst-string. */ |
118 | |
119 | vlbb %v16,0(%r3),6 /* Load s until next 4k-byte boundary. */ |
120 | lcbb %r6,0(%r3),6 /* Get bytes to 4k-byte boundary or 16. */ |
121 | llgfr %r6,%r6 /* Convert 32bit to 64bit. */ |
122 | |
123 | lghi %r5,0 /* current_len = 0. */ |
124 | |
125 | /* Check range of maxlen and convert to byte-count. */ |
126 | # ifdef __s390x__ |
127 | tmhh %r4,49152 /* Test bit 0 or 1 of maxlen. */ |
128 | lghi %r1,-4 /* Max byte-count is 18446744073709551612. */ |
129 | # else |
130 | tmlh %r4,49152 /* Test bit 0 or 1 of maxlen. */ |
131 | llilf %r1,4294967292 /* Max byte-count is 4294967292. */ |
132 | # endif /* !__s390x__ */ |
133 | sllg %r4,%r4,2 /* Convert character-count to byte-count. */ |
134 | locgrne %r4,%r1 /* Use max byte-count, if bit 0/1 was one. */ |
135 | |
136 | clgrjle %r4,%r6,.Lcpy_remaining_v16 /* If n <= loaded-bytes |
137 | -> process remaining. */ |
138 | |
139 | /* n > loaded-byte-count. */ |
140 | vfenezf %v17,%v16,%v16 /* Find element not equal with zero search. */ |
141 | vlgvb %r1,%v17,7 /* Load zero index or 16 if not found. */ |
142 | clrjl %r1,%r6,.Lcpy_found_v16_store /* Found zero within loaded bytes, |
143 | copy and return. */ |
144 | |
145 | /* Align s to 16 byte. */ |
146 | risbgn %r1,%r3,60,128+63,0 /* %r3 = bits 60-63 of %r2 'and' 15. */ |
147 | lghi %r5,15 /* current_len = 15. */ |
148 | slr %r5,%r1 /* Compute highest index to 16byte boundary. * |
149 | |
150 | /* Zero not found and maxlen > loaded-byte-count. */ |
151 | vstl %v16,%r5,0(%r2) /* Copy loaded characters - no zero. */ |
152 | ahi %r5,1 /* Start loop at next character. */ |
153 | |
154 | /* |
155 | Now we are 16byte aligned, so we can load a full vreg |
156 | without page fault. |
157 | */ |
158 | lgr %r1,%r5 /* If %r5 + 64 < maxlen? -> loop64. */ |
159 | aghi %r1,64 |
160 | clgrjl %r1,%r4,.Lcpy_loop64 |
161 | |
162 | vl %v16,0(%r5,%r3) /* Load s. */ |
163 | clgijl %r4,17,.Lcpy_remaining_v16 /* If n <=16, |
164 | process remaining bytes. */ |
165 | .Lcpy_lt64: |
166 | lgr %r7,%r4 |
167 | slgfi %r7,16 /* border_len = n - 16. */ |
168 | |
169 | clgrjhe %r5,%r7,.Lcpy_remaining_v16 |
170 | vfenezfs %v17,%v16,%v16 /* Find element not equal with zero search. */ |
171 | je .Lcpy_found_v16 /* Jump away if zero was found. */ |
172 | vl %v18,16(%r5,%r3) /* Load next part of s. */ |
173 | vst %v16,0(%r5,%r2) /* Save previous part without zero to dst. */ |
174 | aghi %r5,16 |
175 | |
176 | clgrjhe %r5,%r7,.Lcpy_remaining_v18 |
177 | vfenezfs %v17,%v18,%v18 |
178 | je .Lcpy_found_v18 |
179 | vl %v16,16(%r5,%r3) |
180 | vst %v18,0(%r5,%r2) |
181 | aghi %r5,16 |
182 | |
183 | clgrjhe %r5,%r7,.Lcpy_remaining_v16 |
184 | vfenezfs %v17,%v16,%v16 |
185 | je .Lcpy_found_v16 |
186 | vl %v18,16(%r5,%r3) |
187 | vst %v16,0(%r5,%r2) |
188 | aghi %r5,16 |
189 | |
190 | .Lcpy_remaining_v18: |
191 | vlr %v16,%v18 |
192 | .Lcpy_remaining_v16: |
193 | /* v16 contains the remaining bytes [1...16]. |
194 | Store remaining bytes and append string-termination. */ |
195 | vfenezf %v17,%v16,%v16 /* Find element not equal with zero search. */ |
196 | slgrk %r7,%r4,%r5 /* Remaining bytes = maxlen - current_len. */ |
197 | aghi %r7,-1 /* vstl needs highest index. */ |
198 | vlgvb %r1,%v17,7 /* Load zero index or 16 if not found. */ |
199 | la %r2,0(%r5,%r2) /* vstl has no index register. */ |
200 | /* Zero-index within remaining-bytes, store up to zero and end. */ |
201 | clgrjle %r1,%r7,.Lcpy_found_v16_store |
202 | vstl %v16,%r7,0(%r2) /* Store remaining bytes. */ |
203 | lghi %r1,0 |
204 | st %r1,1(%r7,%r2) /* Store string-null-termination beyond n. */ |
205 | .Lcpy_end: |
206 | /* Restore saved registers. */ |
207 | vlgvg %r6,%v31,0 |
208 | vlgvg %r7,%v31,1 |
209 | lgr %r2,%r0 /* Load saved dest-ptr. */ |
210 | br %r14 |
211 | |
212 | .Lcpy_found_v16_32: |
213 | aghi %r5,32 |
214 | j .Lcpy_found_v16 |
215 | .Lcpy_found_v18_48: |
216 | aghi %r5,32 |
217 | .Lcpy_found_v18_16: |
218 | aghi %r5,16 |
219 | .Lcpy_found_v18: |
220 | vlr %v16,%v18 |
221 | .Lcpy_found_v16: |
222 | /* v16 contains a zero. Store remaining bytes to zero. current_len |
223 | has not reached border, thus checking for n is not needed! */ |
224 | vlgvb %r1,%v17,7 /* Load byte index of zero. */ |
225 | la %r2,0(%r5,%r2) |
226 | .Lcpy_found_v16_store: |
227 | aghi %r1,3 /* Also copy remaining bytes of zero. */ |
228 | vstl %v16,%r1,0(%r2) /* Copy characters including zero. */ |
229 | j .Lcpy_end |
230 | |
231 | /* Find zero in 16byte aligned loop. */ |
232 | .Lcpy_loop2: |
233 | vl %v16,16(%r5,%r3) |
234 | vst %v18,0(%r5,%r2) |
235 | aghi %r5,16 |
236 | |
237 | .Lcpy_loop64: |
238 | vl %v16,0(%r5,%r3) |
239 | vfenezfs %v17,%v16,%v16 /* Find element not equal with zero search. */ |
240 | je .Lcpy_found_v16 /* Jump away if zero was found. */ |
241 | vl %v18,16(%r5,%r3) /* Load next part of s. */ |
242 | vst %v16,0(%r5,%r2) /* Save previous part without zero to dst. */ |
243 | vfenezfs %v17,%v18,%v18 |
244 | je .Lcpy_found_v18_16 |
245 | vl %v16,32(%r5,%r3) |
246 | vst %v18,16(%r5,%r2) |
247 | vfenezfs %v17,%v16,%v16 |
248 | je .Lcpy_found_v16_32 |
249 | vl %v18,48(%r5,%r3) |
250 | vst %v16,32(%r5,%r2) |
251 | vfenezfs %v17,%v18,%v18 |
252 | je .Lcpy_found_v18_48 |
253 | vst %v18,48(%r5,%r2) |
254 | |
255 | aghi %r5,64 |
256 | lgr %r1,%r5 /* If %r5 + 64 < maxlen? -> loop64. */ |
257 | aghi %r1,64 |
258 | clgrjl %r1,%r4,.Lcpy_loop64 |
259 | |
260 | vl %v16,0(%r5,%r3) /* Load s. */ |
261 | j .Lcpy_lt64 |
262 | |
263 | .Lfallback: |
264 | jg WCSNCAT_C |
265 | END(WCSNCAT_Z13) |
266 | |
267 | # if ! HAVE_WCSNCAT_IFUNC |
268 | strong_alias (WCSNCAT_Z13, wcsncat) |
269 | # endif |
270 | #endif |
271 | |