1 /* { dg-additional-options "-msve-vector-bits=scalable" } */
2 /* { dg-final { check-function-bodies "**" "" "-DCHECK_ASM" } } */
3
4 #include "test_sve_acle.h"
5 #include <stdbool.h>
6
7 /*
8 ** cnt_b8_32:
9 ** cntp x0, p0, p1\.b
10 ** ret
11 */
12 TEST_PTEST (cnt_b8_32, uint32_t,
13 x0 = svcntp_b8 (p0, p1));
14
15 /*
16 ** cnt_b8_64:
17 ** cntp x0, p0, p1\.b
18 ** ret
19 */
20 TEST_PTEST (cnt_b8_64, uint64_t,
21 x0 = svcntp_b8 (p0, p1));
22
23 /*
24 ** inc_b8_32_general_x0:
25 ** cntp x([0-9]+), p0, p1\.b
26 ** add w0, (w0, w\1|w\1, w0)
27 ** ret
28 */
29 TEST_PTEST (inc_b8_32_general_x0, uint32_t,
30 x0 += svcntp_b8 (p0, p1));
31
32 /*
33 ** inc_b8_32_general_x1:
34 ** cntp x([0-9]+), p0, p1\.b
35 ** add w0, (w1, w\1|w\1, w1)
36 ** ret
37 */
38 TEST_PTEST (inc_b8_32_general_x1, uint32_t,
39 x0 = x1 + svcntp_b8 (p0, p1));
40
41 /*
42 ** inc_b8_32_ptrue_x0:
43 ** incp x0, p1\.b
44 ** ret
45 */
46 TEST_PTEST (inc_b8_32_ptrue_x0, uint32_t,
47 x0 += svcntp_b8 (svptrue_b8 (), p1));
48
49 /*
50 ** inc_b8_32_ptrue_x1:
51 ** mov w0, w1
52 ** incp x0, p1\.b
53 ** ret
54 */
55 TEST_PTEST (inc_b8_32_ptrue_x1, uint32_t,
56 x0 = x1 + svcntp_b8 (svptrue_b8 (), p1));
57
58 /*
59 ** inc_b8_64_general_x0:
60 ** cntp (x[0-9]+), p0, p1\.b
61 ** add x0, (x0, \1|\1, x0)
62 ** ret
63 */
64 TEST_PTEST (inc_b8_64_general_x0, uint64_t,
65 x0 += svcntp_b8 (p0, p1));
66
67 /*
68 ** inc_b8_64_general_x1:
69 ** cntp (x[0-9]+), p0, p1\.b
70 ** add x0, (x1, \1|\1, x1)
71 ** ret
72 */
73 TEST_PTEST (inc_b8_64_general_x1, uint64_t,
74 x0 = x1 + svcntp_b8 (p0, p1));
75
76 /*
77 ** inc_b8_64_ptrue_x0:
78 ** incp x0, p1\.b
79 ** ret
80 */
81 TEST_PTEST (inc_b8_64_ptrue_x0, uint64_t,
82 x0 += svcntp_b8 (svptrue_b8 (), p1));
83
84 /*
85 ** inc_b8_64_ptrue_x1:
86 ** mov x0, x1
87 ** incp x0, p1\.b
88 ** ret
89 */
90 TEST_PTEST (inc_b8_64_ptrue_x1, uint64_t,
91 x0 = x1 + svcntp_b8 (svptrue_b8 (), p1));
92
93 /*
94 ** dec_b8_32_general_x0:
95 ** cntp x([0-9]+), p0, p1\.b
96 ** sub w0, w0, w\1
97 ** ret
98 */
99 TEST_PTEST (dec_b8_32_general_x0, uint32_t,
100 x0 -= svcntp_b8 (p0, p1));
101
102 /*
103 ** dec_b8_32_general_x1:
104 ** cntp x([0-9]+), p0, p1\.b
105 ** sub w0, w1, w\1
106 ** ret
107 */
108 TEST_PTEST (dec_b8_32_general_x1, uint32_t,
109 x0 = x1 - svcntp_b8 (p0, p1));
110
111 /*
112 ** dec_b8_32_ptrue_x0:
113 ** decp x0, p1\.b
114 ** ret
115 */
116 TEST_PTEST (dec_b8_32_ptrue_x0, uint32_t,
117 x0 -= svcntp_b8 (svptrue_b8 (), p1));
118
119 /*
120 ** dec_b8_32_ptrue_x1:
121 ** mov w0, w1
122 ** decp x0, p1\.b
123 ** ret
124 */
125 TEST_PTEST (dec_b8_32_ptrue_x1, uint32_t,
126 x0 = x1 - svcntp_b8 (svptrue_b8 (), p1));
127
128 /*
129 ** dec_b8_64_general_x0:
130 ** cntp (x[0-9]+), p0, p1\.b
131 ** sub x0, x0, \1
132 ** ret
133 */
134 TEST_PTEST (dec_b8_64_general_x0, uint64_t,
135 x0 -= svcntp_b8 (p0, p1));
136
137 /*
138 ** dec_b8_64_general_x1:
139 ** cntp (x[0-9]+), p0, p1\.b
140 ** sub x0, x1, \1
141 ** ret
142 */
143 TEST_PTEST (dec_b8_64_general_x1, uint64_t,
144 x0 = x1 - svcntp_b8 (p0, p1));
145
146 /*
147 ** dec_b8_64_ptrue_x0:
148 ** decp x0, p1\.b
149 ** ret
150 */
151 TEST_PTEST (dec_b8_64_ptrue_x0, uint64_t,
152 x0 -= svcntp_b8 (svptrue_b8 (), p1));
153
154 /*
155 ** dec_b8_64_ptrue_x1:
156 ** mov x0, x1
157 ** decp x0, p1\.b
158 ** ret
159 */
160 TEST_PTEST (dec_b8_64_ptrue_x1, uint64_t,
161 x0 = x1 - svcntp_b8 (svptrue_b8 (), p1));
162
163 /*
164 ** inc_b8_s8_general_z0:
165 ** cntp x([0-9]+), p0, p1\.b
166 ** mov (z[0-9]+\.b), w\1
167 ** add z0\.b, (z0\.b, \2|\2, z0\.b)
168 ** ret
169 */
170 TEST_UNIFORM_Z (inc_b8_s8_general_z0, svint8_t,
171 z0 = svadd_n_s8_x (svptrue_b8 (), z0, svcntp_b8 (p0, p1)),
172 z0 = svadd_x (svptrue_b8 (), z0, svcntp_b8 (p0, p1)));
173
174 /*
175 ** inc_b8_s8_general_z1:
176 ** cntp x([0-9]+), p0, p1\.b
177 ** mov (z[0-9]+\.b), w\1
178 ** add z0\.b, (z1\.b, \2|\2, z1\.b)
179 ** ret
180 */
181 TEST_UNIFORM_Z (inc_b8_s8_general_z1, svint8_t,
182 z0 = svadd_n_s8_x (svptrue_b8 (), z1, svcntp_b8 (p0, p1)),
183 z0 = svadd_x (svptrue_b8 (), z1, svcntp_b8 (p0, p1)));
184
185 /*
186 ** inc_b8_s8_ptrue_z0:
187 ** ptrue (p[0-7])\.b, all
188 ** cntp x([0-9]+), \1, p0\.b
189 ** mov (z[0-9]+\.b), w\2
190 ** add z0\.b, (z0\.b, \3|\3, z0\.b)
191 ** ret
192 */
193 TEST_UNIFORM_Z (inc_b8_s8_ptrue_z0, svint8_t,
194 z0 = svadd_n_s8_x (svptrue_b8 (), z0, svcntp_b8 (svptrue_b8 (), p0)),
195 z0 = svadd_x (svptrue_b8 (), z0, svcntp_b8 (svptrue_b8 (), p0)));
196
197 /*
198 ** inc_b8_s8_ptrue_z1:
199 ** ptrue (p[0-7])\.b, all
200 ** cntp x([0-9]+), \1, p0\.b
201 ** mov (z[0-9]+\.b), w\2
202 ** add z0\.b, (z1\.b, \3|\3, z1\.b)
203 ** ret
204 */
205 TEST_UNIFORM_Z (inc_b8_s8_ptrue_z1, svint8_t,
206 z0 = svadd_n_s8_x (svptrue_b8 (), z1, svcntp_b8 (svptrue_b8 (), p0)),
207 z0 = svadd_x (svptrue_b8 (), z1, svcntp_b8 (svptrue_b8 (), p0)));
208
209 /*
210 ** dec_b8_s8_general_z0:
211 ** cntp x([0-9]+), p0, p1\.b
212 ** mov (z[0-9]+\.b), w\1
213 ** sub z0\.b, z0\.b, \2
214 ** ret
215 */
216 TEST_UNIFORM_Z (dec_b8_s8_general_z0, svint8_t,
217 z0 = svsub_n_s8_x (svptrue_b8 (), z0, svcntp_b8 (p0, p1)),
218 z0 = svsub_x (svptrue_b8 (), z0, svcntp_b8 (p0, p1)));
219
220 /*
221 ** dec_b8_s8_general_z1:
222 ** cntp x([0-9]+), p0, p1\.b
223 ** mov (z[0-9]+\.b), w\1
224 ** sub z0\.b, z1\.b, \2
225 ** ret
226 */
227 TEST_UNIFORM_Z (dec_b8_s8_general_z1, svint8_t,
228 z0 = svsub_n_s8_x (svptrue_b8 (), z1, svcntp_b8 (p0, p1)),
229 z0 = svsub_x (svptrue_b8 (), z1, svcntp_b8 (p0, p1)));
230
231 /*
232 ** dec_b8_s8_ptrue_z0:
233 ** ptrue (p[0-7])\.b, all
234 ** cntp x([0-9]+), \1, p0\.b
235 ** mov (z[0-9]+\.b), w\2
236 ** sub z0\.b, z0\.b, \3
237 ** ret
238 */
239 TEST_UNIFORM_Z (dec_b8_s8_ptrue_z0, svint8_t,
240 z0 = svsub_n_s8_x (svptrue_b8 (), z0, svcntp_b8 (svptrue_b8 (), p0)),
241 z0 = svsub_x (svptrue_b8 (), z0, svcntp_b8 (svptrue_b8 (), p0)));
242
243 /*
244 ** dec_b8_s8_ptrue_z1:
245 ** ptrue (p[0-7])\.b, all
246 ** cntp x([0-9]+), \1, p0\.b
247 ** mov (z[0-9]+\.b), w\2
248 ** sub z0\.b, z1\.b, \3
249 ** ret
250 */
251 TEST_UNIFORM_Z (dec_b8_s8_ptrue_z1, svint8_t,
252 z0 = svsub_n_s8_x (svptrue_b8 (), z1, svcntp_b8 (svptrue_b8 (), p0)),
253 z0 = svsub_x (svptrue_b8 (), z1, svcntp_b8 (svptrue_b8 (), p0)));