152
|
1 /* PR target/92658 */
|
|
2 /* { dg-do compile } */
|
|
3 /* { dg-options "-O2 -ftree-vectorize -msse4.1" } */
|
|
4
|
|
5 typedef unsigned char v16qi __attribute__((vector_size (16)));
|
|
6 typedef unsigned short v8hi __attribute__((vector_size (16)));
|
|
7 typedef unsigned int v4si __attribute__((vector_size (16)));
|
|
8 typedef unsigned long long v2di __attribute__((vector_size (16)));
|
|
9
|
|
10 void
|
|
11 foo_u8_u16 (v8hi * dst, v16qi * __restrict src)
|
|
12 {
|
|
13 unsigned short tem[8];
|
|
14 tem[0] = (*src)[0];
|
|
15 tem[1] = (*src)[1];
|
|
16 tem[2] = (*src)[2];
|
|
17 tem[3] = (*src)[3];
|
|
18 tem[4] = (*src)[4];
|
|
19 tem[5] = (*src)[5];
|
|
20 tem[6] = (*src)[6];
|
|
21 tem[7] = (*src)[7];
|
|
22 dst[0] = *(v8hi *) tem;
|
|
23 }
|
|
24
|
|
25 void
|
|
26 bar_u8_u16 (v8hi * dst, v16qi src)
|
|
27 {
|
|
28 unsigned short tem[8];
|
|
29 tem[0] = src[0];
|
|
30 tem[1] = src[1];
|
|
31 tem[2] = src[2];
|
|
32 tem[3] = src[3];
|
|
33 tem[4] = src[4];
|
|
34 tem[5] = src[5];
|
|
35 tem[6] = src[6];
|
|
36 tem[7] = src[7];
|
|
37 dst[0] = *(v8hi *) tem;
|
|
38 }
|
|
39
|
|
40 /* { dg-final { scan-assembler-times "pmovzxbw" 2 } } */
|
|
41
|
|
42 void
|
|
43 foo_u8_u32 (v4si * dst, v16qi * __restrict src)
|
|
44 {
|
|
45 unsigned int tem[4];
|
|
46 tem[0] = (*src)[0];
|
|
47 tem[1] = (*src)[1];
|
|
48 tem[2] = (*src)[2];
|
|
49 tem[3] = (*src)[3];
|
|
50 dst[0] = *(v4si *) tem;
|
|
51 }
|
|
52
|
|
53 void
|
|
54 bar_u8_u32 (v4si * dst, v16qi src)
|
|
55 {
|
|
56 unsigned int tem[4];
|
|
57 tem[0] = src[0];
|
|
58 tem[1] = src[1];
|
|
59 tem[2] = src[2];
|
|
60 tem[3] = src[3];
|
|
61 dst[0] = *(v4si *) tem;
|
|
62 }
|
|
63
|
|
64 /* { dg-final { scan-assembler-times "pmovzxbd" 2 { xfail *-*-* } } } */
|
|
65
|
|
66 void
|
|
67 foo_u8_u64 (v2di * dst, v16qi * __restrict src)
|
|
68 {
|
|
69 unsigned long long tem[2];
|
|
70 tem[0] = (*src)[0];
|
|
71 tem[1] = (*src)[1];
|
|
72 dst[0] = *(v2di *) tem;
|
|
73 }
|
|
74
|
|
75 void
|
|
76 bar_u8_u64 (v2di * dst, v16qi src)
|
|
77 {
|
|
78 unsigned long long tem[2];
|
|
79 tem[0] = src[0];
|
|
80 tem[1] = src[1];
|
|
81 dst[0] = *(v2di *) tem;
|
|
82 }
|
|
83
|
|
84 /* { dg-final { scan-assembler-times "pmovzxbq" 2 { xfail *-*-* } } } */
|
|
85
|
|
86 void
|
|
87 foo_u16_u32 (v4si * dst, v8hi * __restrict src)
|
|
88 {
|
|
89 unsigned int tem[4];
|
|
90 tem[0] = (*src)[0];
|
|
91 tem[1] = (*src)[1];
|
|
92 tem[2] = (*src)[2];
|
|
93 tem[3] = (*src)[3];
|
|
94 dst[0] = *(v4si *) tem;
|
|
95 }
|
|
96
|
|
97 void
|
|
98 bar_u16_u32 (v4si * dst, v8hi src)
|
|
99 {
|
|
100 unsigned int tem[4];
|
|
101 tem[0] = src[0];
|
|
102 tem[1] = src[1];
|
|
103 tem[2] = src[2];
|
|
104 tem[3] = src[3];
|
|
105 dst[0] = *(v4si *) tem;
|
|
106 }
|
|
107
|
|
108 /* { dg-final { scan-assembler-times "pmovzxwd" 2 } } */
|
|
109
|
|
110 void
|
|
111 foo_u16_u64 (v2di * dst, v8hi * __restrict src)
|
|
112 {
|
|
113 unsigned long long tem[2];
|
|
114 tem[0] = (*src)[0];
|
|
115 tem[1] = (*src)[1];
|
|
116 dst[0] = *(v2di *) tem;
|
|
117 }
|
|
118
|
|
119 void
|
|
120 bar_u16_u64 (v2di * dst, v8hi src)
|
|
121 {
|
|
122 unsigned long long tem[2];
|
|
123 tem[0] = src[0];
|
|
124 tem[1] = src[1];
|
|
125 dst[0] = *(v2di *) tem;
|
|
126 }
|
|
127
|
|
128 /* { dg-final { scan-assembler-times "pmovzxwq" 2 { xfail *-*-* } } } */
|
|
129
|
|
130 void
|
|
131 foo_u32_u64 (v2di * dst, v4si * __restrict src)
|
|
132 {
|
|
133 unsigned long long tem[2];
|
|
134 tem[0] = (*src)[0];
|
|
135 tem[1] = (*src)[1];
|
|
136 dst[0] = *(v2di *) tem;
|
|
137 }
|
|
138
|
|
139 void
|
|
140 bar_u32_u64 (v2di * dst, v4si src)
|
|
141 {
|
|
142 unsigned long long tem[2];
|
|
143 tem[0] = src[0];
|
|
144 tem[1] = src[1];
|
|
145 dst[0] = *(v2di *) tem;
|
|
146 }
|
|
147
|
|
148 /* { dg-final { scan-assembler-times "pmovzxdq" 2 } } */
|