1 /* { dg-final { check-function-bodies "**" "" "-DCHECK_ASM" } } */
2
3 #include "test_sve_acle.h"
4
5 /*
6 ** cdot_lane_0_0_s64_tied1:
7 ** cdot z0\.d, z4\.h, z5\.h\[0\], #0
8 ** ret
9 */
10 TEST_DUAL_Z (cdot_lane_0_0_s64_tied1, svint64_t, svint16_t,
11 z0 = svcdot_lane_s64 (z0, z4, z5, 0, 0),
12 z0 = svcdot_lane (z0, z4, z5, 0, 0))
13
14 /*
15 ** cdot_lane_0_0_s64_tied2:
16 ** mov (z[0-9]+)\.d, z0\.d
17 ** movprfx z0, z4
18 ** cdot z0\.d, \1\.h, z1\.h\[0\], #0
19 ** ret
20 */
21 TEST_DUAL_Z_REV (cdot_lane_0_0_s64_tied2, svint64_t, svint16_t,
22 z0_res = svcdot_lane_s64 (z4, z0, z1, 0, 0),
23 z0_res = svcdot_lane (z4, z0, z1, 0, 0))
24
25 /*
26 ** cdot_lane_0_0_s64_tied3:
27 ** mov (z[0-9]+)\.d, z0\.d
28 ** movprfx z0, z4
29 ** cdot z0\.d, z1\.h, \1\.h\[0\], #0
30 ** ret
31 */
32 TEST_DUAL_Z_REV (cdot_lane_0_0_s64_tied3, svint64_t, svint16_t,
33 z0_res = svcdot_lane_s64 (z4, z1, z0, 0, 0),
34 z0_res = svcdot_lane (z4, z1, z0, 0, 0))
35
36 /*
37 ** cdot_lane_0_0_s64_untied:
38 ** movprfx z0, z1
39 ** cdot z0\.d, z4\.h, z5\.h\[0\], #0
40 ** ret
41 */
42 TEST_DUAL_Z (cdot_lane_0_0_s64_untied, svint64_t, svint16_t,
43 z0 = svcdot_lane_s64 (z1, z4, z5, 0, 0),
44 z0 = svcdot_lane (z1, z4, z5, 0, 0))
45
46 /*
47 ** cdot_lane_0_90_s64_tied1:
48 ** cdot z0\.d, z4\.h, z5\.h\[0\], #90
49 ** ret
50 */
51 TEST_DUAL_Z (cdot_lane_0_90_s64_tied1, svint64_t, svint16_t,
52 z0 = svcdot_lane_s64 (z0, z4, z5, 0, 90),
53 z0 = svcdot_lane (z0, z4, z5, 0, 90))
54
55 /*
56 ** cdot_lane_0_90_s64_tied2:
57 ** mov (z[0-9]+)\.d, z0\.d
58 ** movprfx z0, z4
59 ** cdot z0\.d, \1\.h, z1\.h\[0\], #90
60 ** ret
61 */
62 TEST_DUAL_Z_REV (cdot_lane_0_90_s64_tied2, svint64_t, svint16_t,
63 z0_res = svcdot_lane_s64 (z4, z0, z1, 0, 90),
64 z0_res = svcdot_lane (z4, z0, z1, 0, 90))
65
66 /*
67 ** cdot_lane_0_90_s64_tied3:
68 ** mov (z[0-9]+)\.d, z0\.d
69 ** movprfx z0, z4
70 ** cdot z0\.d, z1\.h, \1\.h\[0\], #90
71 ** ret
72 */
73 TEST_DUAL_Z_REV (cdot_lane_0_90_s64_tied3, svint64_t, svint16_t,
74 z0_res = svcdot_lane_s64 (z4, z1, z0, 0, 90),
75 z0_res = svcdot_lane (z4, z1, z0, 0, 90))
76
77 /*
78 ** cdot_lane_0_90_s64_untied:
79 ** movprfx z0, z1
80 ** cdot z0\.d, z4\.h, z5\.h\[0\], #90
81 ** ret
82 */
83 TEST_DUAL_Z (cdot_lane_0_90_s64_untied, svint64_t, svint16_t,
84 z0 = svcdot_lane_s64 (z1, z4, z5, 0, 90),
85 z0 = svcdot_lane (z1, z4, z5, 0, 90))
86
87 /*
88 ** cdot_lane_0_180_s64_tied1:
89 ** cdot z0\.d, z4\.h, z5\.h\[0\], #180
90 ** ret
91 */
92 TEST_DUAL_Z (cdot_lane_0_180_s64_tied1, svint64_t, svint16_t,
93 z0 = svcdot_lane_s64 (z0, z4, z5, 0, 180),
94 z0 = svcdot_lane (z0, z4, z5, 0, 180))
95
96 /*
97 ** cdot_lane_0_180_s64_tied2:
98 ** mov (z[0-9]+)\.d, z0\.d
99 ** movprfx z0, z4
100 ** cdot z0\.d, \1\.h, z1\.h\[0\], #180
101 ** ret
102 */
103 TEST_DUAL_Z_REV (cdot_lane_0_180_s64_tied2, svint64_t, svint16_t,
104 z0_res = svcdot_lane_s64 (z4, z0, z1, 0, 180),
105 z0_res = svcdot_lane (z4, z0, z1, 0, 180))
106
107 /*
108 ** cdot_lane_0_180_s64_tied3:
109 ** mov (z[0-9]+)\.d, z0\.d
110 ** movprfx z0, z4
111 ** cdot z0\.d, z1\.h, \1\.h\[0\], #180
112 ** ret
113 */
114 TEST_DUAL_Z_REV (cdot_lane_0_180_s64_tied3, svint64_t, svint16_t,
115 z0_res = svcdot_lane_s64 (z4, z1, z0, 0, 180),
116 z0_res = svcdot_lane (z4, z1, z0, 0, 180))
117
118 /*
119 ** cdot_lane_0_180_s64_untied:
120 ** movprfx z0, z1
121 ** cdot z0\.d, z4\.h, z5\.h\[0\], #180
122 ** ret
123 */
124 TEST_DUAL_Z (cdot_lane_0_180_s64_untied, svint64_t, svint16_t,
125 z0 = svcdot_lane_s64 (z1, z4, z5, 0, 180),
126 z0 = svcdot_lane (z1, z4, z5, 0, 180))
127
128 /*
129 ** cdot_lane_0_270_s64_tied1:
130 ** cdot z0\.d, z4\.h, z5\.h\[0\], #270
131 ** ret
132 */
133 TEST_DUAL_Z (cdot_lane_0_270_s64_tied1, svint64_t, svint16_t,
134 z0 = svcdot_lane_s64 (z0, z4, z5, 0, 270),
135 z0 = svcdot_lane (z0, z4, z5, 0, 270))
136
137 /*
138 ** cdot_lane_0_270_s64_tied2:
139 ** mov (z[0-9]+)\.d, z0\.d
140 ** movprfx z0, z4
141 ** cdot z0\.d, \1\.h, z1\.h\[0\], #270
142 ** ret
143 */
144 TEST_DUAL_Z_REV (cdot_lane_0_270_s64_tied2, svint64_t, svint16_t,
145 z0_res = svcdot_lane_s64 (z4, z0, z1, 0, 270),
146 z0_res = svcdot_lane (z4, z0, z1, 0, 270))
147
148 /*
149 ** cdot_lane_0_270_s64_tied3:
150 ** mov (z[0-9]+)\.d, z0\.d
151 ** movprfx z0, z4
152 ** cdot z0\.d, z1\.h, \1\.h\[0\], #270
153 ** ret
154 */
155 TEST_DUAL_Z_REV (cdot_lane_0_270_s64_tied3, svint64_t, svint16_t,
156 z0_res = svcdot_lane_s64 (z4, z1, z0, 0, 270),
157 z0_res = svcdot_lane (z4, z1, z0, 0, 270))
158
159 /*
160 ** cdot_lane_0_270_s64_untied:
161 ** movprfx z0, z1
162 ** cdot z0\.d, z4\.h, z5\.h\[0\], #270
163 ** ret
164 */
165 TEST_DUAL_Z (cdot_lane_0_270_s64_untied, svint64_t, svint16_t,
166 z0 = svcdot_lane_s64 (z1, z4, z5, 0, 270),
167 z0 = svcdot_lane (z1, z4, z5, 0, 270))
168
169 /*
170 ** cdot_lane_z15_s64:
171 ** str d15, \[sp, -16\]!
172 ** cdot z0\.d, z1\.h, z15\.h\[1\], #0
173 ** ldr d15, \[sp\], 16
174 ** ret
175 */
176 TEST_DUAL_LANE_REG (cdot_lane_z15_s64, svint64_t, svint16_t, z15,
177 z0 = svcdot_lane_s64 (z0, z1, z15, 1, 0),
178 z0 = svcdot_lane (z0, z1, z15, 1, 0))
179
180 /*
181 ** cdot_lane_z16_s64:
182 ** mov (z[0-9]|z1[0-5])\.d, z16\.d
183 ** cdot z0\.d, z1\.h, \1\.h\[1\], #0
184 ** ret
185 */
186 TEST_DUAL_LANE_REG (cdot_lane_z16_s64, svint64_t, svint16_t, z16,
187 z0 = svcdot_lane_s64 (z0, z1, z16, 1, 0),
188 z0 = svcdot_lane (z0, z1, z16, 1, 0))