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_b64_32:
9 ** cntp x0, p0, p1\.d
10 ** ret
11 */
12 TEST_PTEST (cnt_b64_32, uint32_t,
13 x0 = svcntp_b64 (p0, p1));
14
15 /*
16 ** cnt_b64_64:
17 ** cntp x0, p0, p1\.d
18 ** ret
19 */
20 TEST_PTEST (cnt_b64_64, uint64_t,
21 x0 = svcntp_b64 (p0, p1));
22
23 /*
24 ** inc_b64_32_general_x0:
25 ** cntp x([0-9]+), p0, p1\.d
26 ** add w0, (w0, w\1|w\1, w0)
27 ** ret
28 */
29 TEST_PTEST (inc_b64_32_general_x0, uint32_t,
30 x0 += svcntp_b64 (p0, p1));
31
32 /*
33 ** inc_b64_32_general_x1:
34 ** cntp x([0-9]+), p0, p1\.d
35 ** add w0, (w1, w\1|w\1, w1)
36 ** ret
37 */
38 TEST_PTEST (inc_b64_32_general_x1, uint32_t,
39 x0 = x1 + svcntp_b64 (p0, p1));
40
41 /*
42 ** inc_b64_32_ptrue_x0:
43 ** incp x0, p1\.d
44 ** ret
45 */
46 TEST_PTEST (inc_b64_32_ptrue_x0, uint32_t,
47 x0 += svcntp_b64 (svptrue_b64 (), p1));
48
49 /*
50 ** inc_b64_32_ptrue_x1:
51 ** mov w0, w1
52 ** incp x0, p1\.d
53 ** ret
54 */
55 TEST_PTEST (inc_b64_32_ptrue_x1, uint32_t,
56 x0 = x1 + svcntp_b64 (svptrue_b64 (), p1));
57
58 /*
59 ** inc_b64_64_general_x0:
60 ** cntp (x[0-9]+), p0, p1\.d
61 ** add x0, (x0, \1|\1, x0)
62 ** ret
63 */
64 TEST_PTEST (inc_b64_64_general_x0, uint64_t,
65 x0 += svcntp_b64 (p0, p1));
66
67 /*
68 ** inc_b64_64_general_x1:
69 ** cntp (x[0-9]+), p0, p1\.d
70 ** add x0, (x1, \1|\1, x1)
71 ** ret
72 */
73 TEST_PTEST (inc_b64_64_general_x1, uint64_t,
74 x0 = x1 + svcntp_b64 (p0, p1));
75
76 /*
77 ** inc_b64_64_ptrue_x0:
78 ** incp x0, p1\.d
79 ** ret
80 */
81 TEST_PTEST (inc_b64_64_ptrue_x0, uint64_t,
82 x0 += svcntp_b64 (svptrue_b64 (), p1));
83
84 /*
85 ** inc_b64_64_ptrue_x1:
86 ** mov x0, x1
87 ** incp x0, p1\.d
88 ** ret
89 */
90 TEST_PTEST (inc_b64_64_ptrue_x1, uint64_t,
91 x0 = x1 + svcntp_b64 (svptrue_b64 (), p1));
92
93 /*
94 ** dec_b64_32_general_x0:
95 ** cntp x([0-9]+), p0, p1\.d
96 ** sub w0, w0, w\1
97 ** ret
98 */
99 TEST_PTEST (dec_b64_32_general_x0, uint32_t,
100 x0 -= svcntp_b64 (p0, p1));
101
102 /*
103 ** dec_b64_32_general_x1:
104 ** cntp x([0-9]+), p0, p1\.d
105 ** sub w0, w1, w\1
106 ** ret
107 */
108 TEST_PTEST (dec_b64_32_general_x1, uint32_t,
109 x0 = x1 - svcntp_b64 (p0, p1));
110
111 /*
112 ** dec_b64_32_ptrue_x0:
113 ** decp x0, p1\.d
114 ** ret
115 */
116 TEST_PTEST (dec_b64_32_ptrue_x0, uint32_t,
117 x0 -= svcntp_b64 (svptrue_b64 (), p1));
118
119 /*
120 ** dec_b64_32_ptrue_x1:
121 ** mov w0, w1
122 ** decp x0, p1\.d
123 ** ret
124 */
125 TEST_PTEST (dec_b64_32_ptrue_x1, uint32_t,
126 x0 = x1 - svcntp_b64 (svptrue_b64 (), p1));
127
128 /*
129 ** dec_b64_64_general_x0:
130 ** cntp (x[0-9]+), p0, p1\.d
131 ** sub x0, x0, \1
132 ** ret
133 */
134 TEST_PTEST (dec_b64_64_general_x0, uint64_t,
135 x0 -= svcntp_b64 (p0, p1));
136
137 /*
138 ** dec_b64_64_general_x1:
139 ** cntp (x[0-9]+), p0, p1\.d
140 ** sub x0, x1, \1
141 ** ret
142 */
143 TEST_PTEST (dec_b64_64_general_x1, uint64_t,
144 x0 = x1 - svcntp_b64 (p0, p1));
145
146 /*
147 ** dec_b64_64_ptrue_x0:
148 ** decp x0, p1\.d
149 ** ret
150 */
151 TEST_PTEST (dec_b64_64_ptrue_x0, uint64_t,
152 x0 -= svcntp_b64 (svptrue_b64 (), p1));
153
154 /*
155 ** dec_b64_64_ptrue_x1:
156 ** mov x0, x1
157 ** decp x0, p1\.d
158 ** ret
159 */
160 TEST_PTEST (dec_b64_64_ptrue_x1, uint64_t,
161 x0 = x1 - svcntp_b64 (svptrue_b64 (), p1));
162
163 /*
164 ** inc_b64_u64_general_z0:
165 ** cntp (x[0-9]+), p0, p1\.d
166 ** mov (z[0-9]+\.d), \1
167 ** add z0\.d, (z0\.d, \2|\2, z0\.d)
168 ** ret
169 */
170 TEST_UNIFORM_Z (inc_b64_u64_general_z0, svuint64_t,
171 z0 = svadd_n_u64_x (svptrue_b64 (), z0, svcntp_b64 (p0, p1)),
172 z0 = svadd_x (svptrue_b64 (), z0, svcntp_b64 (p0, p1)));
173
174 /*
175 ** inc_b64_u64_general_z1:
176 ** cntp (x[0-9]+), p0, p1\.d
177 ** mov (z[0-9]+\.d), \1
178 ** add z0\.d, (z1\.d, \2|\2, z1\.d)
179 ** ret
180 */
181 TEST_UNIFORM_Z (inc_b64_u64_general_z1, svuint64_t,
182 z0 = svadd_n_u64_x (svptrue_b64 (), z1, svcntp_b64 (p0, p1)),
183 z0 = svadd_x (svptrue_b64 (), z1, svcntp_b64 (p0, p1)));
184
185 /*
186 ** inc_b64_u64_ptrue_z0:
187 ** incp z0\.d, p0
188 ** ret
189 */
190 TEST_UNIFORM_Z (inc_b64_u64_ptrue_z0, svuint64_t,
191 z0 = svadd_n_u64_x (svptrue_b64 (), z0, svcntp_b64 (svptrue_b64 (), p0)),
192 z0 = svadd_x (svptrue_b64 (), z0, svcntp_b64 (svptrue_b64 (), p0)));
193
194 /*
195 ** inc_b64_u64_ptrue_z1:
196 ** movprfx z0, z1
197 ** incp z0\.d, p0
198 ** ret
199 */
200 TEST_UNIFORM_Z (inc_b64_u64_ptrue_z1, svuint64_t,
201 z0 = svadd_n_u64_x (svptrue_b64 (), z1, svcntp_b64 (svptrue_b64 (), p0)),
202 z0 = svadd_x (svptrue_b64 (), z1, svcntp_b64 (svptrue_b64 (), p0)));
203
204 /*
205 ** dec_b64_u64_general_z0:
206 ** cntp (x[0-9]+), p0, p1\.d
207 ** mov (z[0-9]+\.d), \1
208 ** sub z0\.d, z0\.d, \2
209 ** ret
210 */
211 TEST_UNIFORM_Z (dec_b64_u64_general_z0, svuint64_t,
212 z0 = svsub_n_u64_x (svptrue_b64 (), z0, svcntp_b64 (p0, p1)),
213 z0 = svsub_x (svptrue_b64 (), z0, svcntp_b64 (p0, p1)));
214
215 /*
216 ** dec_b64_u64_general_z1:
217 ** cntp (x[0-9]+), p0, p1\.d
218 ** mov (z[0-9]+\.d), \1
219 ** sub z0\.d, z1\.d, \2
220 ** ret
221 */
222 TEST_UNIFORM_Z (dec_b64_u64_general_z1, svuint64_t,
223 z0 = svsub_n_u64_x (svptrue_b64 (), z1, svcntp_b64 (p0, p1)),
224 z0 = svsub_x (svptrue_b64 (), z1, svcntp_b64 (p0, p1)));
225
226 /*
227 ** dec_b64_u64_ptrue_z0:
228 ** decp z0\.d, p0
229 ** ret
230 */
231 TEST_UNIFORM_Z (dec_b64_u64_ptrue_z0, svuint64_t,
232 z0 = svsub_n_u64_x (svptrue_b64 (), z0, svcntp_b64 (svptrue_b64 (), p0)),
233 z0 = svsub_x (svptrue_b64 (), z0, svcntp_b64 (svptrue_b64 (), p0)));
234
235 /*
236 ** dec_b64_u64_ptrue_z1:
237 ** movprfx z0, z1
238 ** decp z0\.d, p0
239 ** ret
240 */
241 TEST_UNIFORM_Z (dec_b64_u64_ptrue_z1, svuint64_t,
242 z0 = svsub_n_u64_x (svptrue_b64 (), z1, svcntp_b64 (svptrue_b64 (), p0)),
243 z0 = svsub_x (svptrue_b64 (), z1, svcntp_b64 (svptrue_b64 (), p0)));