| blob | cd528907d6c3daa15c82f65b863e507f41b78cc4 |
1 /* Vector optimized 32/64 bit S/390 version of wcscspn.
2 Copyright (C) 2015-2017 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
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.
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.
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 <http://www.gnu.org/licenses/>. */
19 #if defined HAVE_S390_VX_ASM_SUPPORT && IS_IN (libc)
21 # include "sysdep.h"
22 # include "asm-syntax.h"
24 .text
26 /* size_t wcscspn (const wchar_t *s, const wchar_t * reject)
27 The wcscspn() function calculates the length of the initial segment
28 of s which consists entirely of characters not in reject.
30 This method checks the length of reject string. If it fits entirely
31 in one vector register, a fast algorithm is used, which does not need
32 to check multiple parts of accept-string. Otherwise a slower full
33 check of accept-string is used.
35 register overview:
36 r3: pointer to start of reject-string
37 r2: pointer to start of search-string
38 r0: loaded byte count of vlbb search-string
39 r4: found byte index
40 r1: current return len
41 v16: search-string
42 v17: reject-string
43 v18: temp-vreg
45 ONLY FOR SLOW:
46 v19: first reject-string
47 v20: zero for preparing acc-vector
48 v21: global mask; 1 indicates a match between
49 search-string-vreg and any reject-character
50 v22: current mask; 1 indicates a match between
51 search-string-vreg and any reject-character in current acc-vreg
52 v30, v31: for re-/storing registers r6, r8, r9
53 r5: current len of reject-string
54 r6: zero-index in search-string or 16 if no zero
55 or min(zero-index, loaded byte count)
56 r8: >0, if former reject-string-part contains a zero,
57 otherwise =0;
58 r9: loaded byte count of vlbb reject-string
59 */
60 ENTRY(__wcscspn_vx)
61 .machine "z13"
62 .machinemode "zarch_nohighgprs"
64 tmll %r2,3 /* Test if s is 4-byte aligned? */
65 jne .Lfallback /* And use common-code variant if not. */
67 /*
68 Check if reject-string fits in one vreg:
69 ----------------------------------------
70 */
71 vlbb %v17,0(%r3),0 /* Load reject. */
72 lcbb %r0,0(%r3),0
73 jo .Lcheck_onbb /* Special case if reject
74 lays on block-boundary. */
76 .Lcheck_notonbb:
77 lghi %r1,0 /* Zero out current len. */
78 vistrfs %v17,%v17 /* Fill with zeros after first zero. */
79 je .Lfast /* Zero found -> reject fits in one vreg. */
80 j .Lslow /* No zero -> reject exceeds one vreg. */
83 .Lcheck_onbb:
84 /* Reject lays on block-boundary. */
85 nill %r0,65532 /* Recognize only fully loaded characters. */
86 je .Lcheck_onbb2 /* Reload vr, if we loaded no full wchar_t. */
87 vfenezf %v18,%v17,%v17 /* Search zero in loaded reject bytes. */
88 vlgvb %r4,%v18,7 /* Get index of zero or 16 if not found. */
89 clrjl %r4,%r0,.Lcheck_notonbb /* Zero index < loaded bytes count ->
90 Reject fits in one vreg;
91 Fill with zeros and proceed
92 with FAST. */
93 .Lcheck_onbb2:
94 vl %v17,0(%r3) /* Load reject, which exceeds loaded bytes. */
95 j .Lcheck_notonbb /* Check if reject fits in one vreg. */
98 /*
99 Search s for reject in one vreg
100 -------------------------------
101 */
102 .Lfast:
103 /* Complete reject-string in v17 and remaining bytes are zero. */
105 vlbb %v16,0(%r2),6 /* Load s until next 4k-byte boundary. */
106 lcbb %r0,0(%r2),6 /* Get bytes to 4k-byte boundary or 16. */
108 vfaezfs %v18,%v16,%v17,0 /* Find first element in v16
109 unequal to any in v17
110 or first zero element. */
111 vlgvb %r4,%v18,7 /* Load byte index of found element. */
112 clrjl %r4,%r0,.Lfast_loop_found2 /* If found index is within loaded
113 bytes, return with found element
114 index (=equal count). */
116 /* Align s to 16 byte. */
117 risbgn %r4,%r2,60,128+63,0 /* %r3 = bits 60-63 of %r2 'and' 15. */
118 lghi %r1,16 /* current_len = 16. */
119 slr %r1,%r4 /* Compute bytes to 16bytes boundary. */
121 /* Process s in 16byte aligned loop. */
122 .Lfast_loop:
123 vl %v16,0(%r1,%r2) /* Load search-string. */
124 vfaezfs %v18,%v16,%v17,0 /* Find first element in v16 equal to any
125 in v17 or first zero element. */
126 jno .Lfast_loop_found
128 vl %v16,16(%r1,%r2)
129 vfaezfs %v18,%v16,%v17,0
130 jno .Lfast_loop_found16
132 vl %v16,32(%r1,%r2)
133 vfaezfs %v18,%v16,%v17,0
134 jno .Lfast_loop_found32
136 vl %v16,48(%r1,%r2)
137 vfaezfs %v18,%v16,%v17,0
138 jno .Lfast_loop_found48
140 aghi %r1,64
141 j .Lfast_loop /* Loop if no element was unequal to reject
142 and not zero. */
144 /* Found equal or zero element. */
145 .Lfast_loop_found48:
146 aghi %r1,16
147 .Lfast_loop_found32:
148 aghi %r1,16
149 .Lfast_loop_found16:
150 aghi %r1,16
151 .Lfast_loop_found:
152 vlgvb %r4,%v18,7 /* Load byte index of found element or zero. */
153 .Lfast_loop_found2:
154 algrk %r2,%r1,%r4 /* Add found index to current len. */
155 srlg %r2,%r2,2 /* Convert byte-count to character-count. */
156 br %r14
160 /*
161 Search s for reject in multiple vregs
162 -------------------------------------
163 */
164 .Lslow:
165 /* Save registers. */
166 vlvgg %v30,%r6,0
167 vlvgp %v31,%r8,%r9
169 /* Reject in v17 without zero. */
170 vlr %v19,%v17 /* Save first acc-part for a fast reload. */
171 vzero %v20 /* Zero for preparing acc-vector. */
172 vone %v24 /* One for checking result of former
173 string-part. */
175 /* Align s to 16 byte. */
176 risbg %r4,%r2,60,128+63,0 /* Test if s is aligned and
177 %r4 = bits 60-63 'and' 15. */
178 je .Lslow_loop_str /* If s is aligned, loop aligned. */
179 lghi %r0,15
180 slr %r0,%r4 /* Compute highest index to load (15-x). */
181 vll %v16,%r0,0(%r2) /* Load up to 16byte boundary (vll needs
182 highest index, remaining bytes are 0). */
183 ahi %r0,1 /* Work with loaded byte count. */
184 vzero %v21 /* Zero out global mask. */
185 lghi %r5,0 /* Set current len of reject-string to zero. */
186 vfenezf %v18,%v16,%v16 /* Find zero in current string-part. */
187 lghi %r8,0 /* There is no zero in first reject-part. */
188 vlgvb %r6,%v18,7 /* Load byte index of zero or 16 if no zero. */
189 clije %r6,0,.Lslow_end /* If first element is zero -> return 0. */
190 clr %r0,%r6 /* cc==1 if loaded byte count < zero-index. */
191 locrl %r6,%r0 /* Load on cc==1; zero-index = lbc. */
192 j .Lslow_loop_acc
195 /* Process s in 16byte aligned loop. */
196 .Lslow_next_str:
197 /* Check results of former processed str-part. */
198 vfeef %v18,%v21,%v24 /* Find first equal match in global mask
199 (ones in element). */
200 vlgvb %r4,%v18,7 /* Get index of first one (=equal) or 16. */
201 /* Equal-index < min(zero-index, loaded byte count)
202 -> Return pointer to equal element. */
203 clrjl %r4,%r6,.Lslow_index_found
204 /* Zero-index < loaded byte count
205 -> Former str-part was last str-part
206 -> Return null */
207 clrjl %r6,%r0,.Lslow_end_not_found
209 /* All elements are zero (=no match) -> proceed with next str-part. */
210 vlr %v17,%v19 /* Load first part of reject (no zero). */
211 algfr %r1,%r0 /* Add loaded byte count to current len. */
213 .Lslow_loop_str:
214 vl %v16,0(%r1,%r2) /* Load search-string. */
215 lghi %r0,16 /* Loaded byte count is 16. */
216 vzero %v21 /* Zero out global mask. */
217 lghi %r5,0 /* Set current len of reject to zero. */
218 vfenezf %v18,%v16,%v16 /* Find zero in current string-part. */
219 lghi %r8,0 /* There is no zero in first reject-part. */
220 vlgvb %r6,%v18,7 /* Load byte index of zero or 16 if no zero. */
221 clije %r6,0,.Lslow_end /* If first element is zero (end of string)
222 -> Return current length. */
224 .Lslow_loop_acc:
225 vfaef %v22,%v16,%v17,4 /* Create matching-mask (1 in mask ->
226 Character matches any rejected character in
227 this reject-string-part) IN=0, RT=1. */
228 vlgvf %r4,%v22,0 /* Get result of first element. */
229 /* First element is equal to any rejected characters?
230 (All other parts of reject cannot lead to a match before this one)
231 -> Return current len, which is pointing to this element. */
232 clijh %r4,0,.Lslow_end
233 vo %v21,%v21,%v22 /* Global-mask = global-|matching-mask. */
234 /* Proceed with next acc until end of acc is reached. */
237 .Lslow_next_acc:
238 clijh %r8,0,.Lslow_next_str /* There was a zero in last reject-part
239 -> Add found index to current len
240 and end. */
241 vlbb %v17,16(%r5,%r3),6 /* Load next reject part. */
242 aghi %r5,16 /* Increment current len of reject-string. */
243 lcbb %r9,0(%r5,%r3),6 /* Get loaded byte count of reject-string. */
244 jo .Lslow_next_acc_onbb /* Jump away if reject-string is
245 on block-boundary. */
246 .Lslow_next_acc_notonbb:
247 vistrfs %v17,%v17 /* Fill with zeros after first zero. */
248 jo .Lslow_loop_acc /* No zero found -> no preparation needed. */
250 .Lslow_next_acc_prepare_zero:
251 /* Zero in reject-part: fill zeros with first-reject-character. */
252 vlgvf %r8,%v17,0 /* Load first element of reject-part. */
253 clije %r8,0,.Lslow_next_str /* Process next str-part if first
254 character in this part of reject
255 is a zero. */
256 /* r8>0 -> zero found in this acc-part. */
257 vrepf %v18,%v17,0 /* Replicate first char accross all chars. */
258 vceqf %v22,%v20,%v17 /* Create a mask (v22) of null chars
259 by comparing with 0 (v20). */
260 vsel %v17,%v18,%v17,%v22 /* Replace null chars with first char. */
261 j .Lslow_loop_acc /* Reject-string part is prepared. */
263 .Lslow_next_acc_onbb:
264 nill %r9,65532 /* Recognize only fully loaded characters. */
265 je .Lslow_next_acc_onbb2 /* Reload vr, if no full wchar_t
266 loaded. */
267 vfenezf %v18,%v17,%v17 /* Find zero in loaded bytes of reject part. */
268 vlgvb %r8,%v18,7 /* Load byte index of zero. */
269 clrjl %r8,%r9,.Lslow_next_acc_notonbb /* Found a zero in loaded bytes
270 -> Prepare vreg. */
271 .Lslow_next_acc_onbb2:
272 vl %v17,0(%r5,%r3) /* Load over boundary ... */
273 lghi %r8,0 /* r8=0 -> no zero in this part of acc,
274 check for zero is in jump-target. */
275 j .Lslow_next_acc_notonbb /* ... and search for zero in
276 fully loaded vreg again. */
278 .Lslow_end_not_found:
279 algfr %r1,%r6 /* Add zero-index to current len. */
280 j .Lslow_end
281 .Lslow_index_found:
282 algfr %r1,%r4 /* Add found index of char to current len. */
283 .Lslow_end:
284 srlg %r2,%r1,2 /* Convert byte-count to character-count. */
285 /* Restore registers. */
286 vlgvg %r6,%v30,0
287 vlgvg %r8,%v31,0
288 vlgvg %r9,%v31,1
289 br %r14
290 .Lfallback:
291 jg __wcscspn_c
292 END(__wcscspn_vx)
293 #endif /* HAVE_S390_VX_ASM_SUPPORT && IS_IN (libc) */