|
111
|
1 /* PR tree-optimization/71488 */
|
|
|
2 /* { dg-require-effective-target vect_int } */
|
|
|
3 /* { dg-require-effective-target vect_pack_trunc } */
|
|
|
4 /* { dg-additional-options "-msse4" { target sse4_runtime } } */
|
|
|
5
|
|
|
6 #include "tree-vect.h"
|
|
|
7
|
|
|
8 int i1, i2;
|
|
|
9
|
|
|
10 void __attribute__((noclone,noinline))
|
|
|
11 fn1 (int * __restrict__ p1, int * __restrict__ p2, int * __restrict__ p3, int size)
|
|
|
12 {
|
|
|
13 int i;
|
|
|
14
|
|
|
15 for (i = 0; i < size; i++)
|
|
|
16 p1[i] = ((p2[i] == 0) > (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
17 }
|
|
|
18
|
|
|
19 void __attribute__((noclone,noinline))
|
|
|
20 fn2 (int * __restrict__ p1, int * __restrict__ p2, short * __restrict__ p3, int size)
|
|
|
21 {
|
|
|
22 int i;
|
|
|
23
|
|
|
24 for (i = 0; i < size; i++)
|
|
|
25 p1[i] = ((p2[i] == 0) > (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
26 }
|
|
|
27
|
|
|
28 void __attribute__((noclone,noinline))
|
|
|
29 fn3 (int * __restrict__ p1, int * __restrict__ p2, long long * __restrict__ p3, int size)
|
|
|
30 {
|
|
|
31 int i;
|
|
|
32
|
|
|
33 for (i = 0; i < size; i++)
|
|
|
34 p1[i] = ((p2[i] == 0) > (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
35 }
|
|
|
36
|
|
|
37 void __attribute__((noclone,noinline))
|
|
|
38 fn4 (int * __restrict__ p1, int * __restrict__ p2, int * __restrict__ p3, int size)
|
|
|
39 {
|
|
|
40 int i;
|
|
|
41
|
|
|
42 for (i = 0; i < size; i++)
|
|
|
43 p1[i] = ((p2[i] == 0) >= (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
44 }
|
|
|
45
|
|
|
46 void __attribute__((noclone,noinline))
|
|
|
47 fn5 (int * __restrict__ p1, int * __restrict__ p2, short * __restrict__ p3, int size)
|
|
|
48 {
|
|
|
49 int i;
|
|
|
50
|
|
|
51 for (i = 0; i < size; i++)
|
|
|
52 p1[i] = ((p2[i] == 0) >= (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
53 }
|
|
|
54
|
|
|
55 void __attribute__((noclone,noinline))
|
|
|
56 fn6 (int * __restrict__ p1, int * __restrict__ p2, long long * __restrict__ p3, int size)
|
|
|
57 {
|
|
|
58 int i;
|
|
|
59
|
|
|
60 for (i = 0; i < size; i++)
|
|
|
61 p1[i] = ((p2[i] == 0) >= (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
62 }
|
|
|
63
|
|
|
64 void __attribute__((noclone,noinline))
|
|
|
65 fn7 (int * __restrict__ p1, int * __restrict__ p2, int * __restrict__ p3, int size)
|
|
|
66 {
|
|
|
67 int i;
|
|
|
68
|
|
|
69 for (i = 0; i < size; i++)
|
|
|
70 p1[i] = ((p2[i] == 0) < (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
71 }
|
|
|
72
|
|
|
73 void __attribute__((noclone,noinline))
|
|
|
74 fn8 (int * __restrict__ p1, int * __restrict__ p2, short * __restrict__ p3, int size)
|
|
|
75 {
|
|
|
76 int i;
|
|
|
77
|
|
|
78 for (i = 0; i < size; i++)
|
|
|
79 p1[i] = ((p2[i] == 0) < (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
80 }
|
|
|
81
|
|
|
82 void __attribute__((noclone,noinline))
|
|
|
83 fn9 (int * __restrict__ p1, int * __restrict__ p2, long long * __restrict__ p3, int size)
|
|
|
84 {
|
|
|
85 int i;
|
|
|
86
|
|
|
87 for (i = 0; i < size; i++)
|
|
|
88 p1[i] = ((p2[i] == 0) < (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
89 }
|
|
|
90
|
|
|
91 void __attribute__((noclone,noinline))
|
|
|
92 fn10 (int * __restrict__ p1, int * __restrict__ p2, int * __restrict__ p3, int size)
|
|
|
93 {
|
|
|
94 int i;
|
|
|
95
|
|
|
96 for (i = 0; i < size; i++)
|
|
|
97 p1[i] = ((p2[i] == 0) <= (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
98 }
|
|
|
99
|
|
|
100 void __attribute__((noclone,noinline))
|
|
|
101 fn11 (int * __restrict__ p1, int * __restrict__ p2, short * __restrict__ p3, int size)
|
|
|
102 {
|
|
|
103 int i;
|
|
|
104
|
|
|
105 for (i = 0; i < size; i++)
|
|
|
106 p1[i] = ((p2[i] == 0) <= (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
107 }
|
|
|
108
|
|
|
109 void __attribute__((noclone,noinline))
|
|
|
110 fn12 (int * __restrict__ p1, int * __restrict__ p2, long long * __restrict__ p3, int size)
|
|
|
111 {
|
|
|
112 int i;
|
|
|
113
|
|
|
114 for (i = 0; i < size; i++)
|
|
|
115 p1[i] = ((p2[i] == 0) <= (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
116 }
|
|
|
117
|
|
|
118 void __attribute__((noclone,noinline))
|
|
|
119 fn13 (int * __restrict__ p1, int * __restrict__ p2, int * __restrict__ p3, int size)
|
|
|
120 {
|
|
|
121 int i;
|
|
|
122
|
|
|
123 for (i = 0; i < size; i++)
|
|
|
124 p1[i] = ((p2[i] == 0) == (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
125 }
|
|
|
126
|
|
|
127 void __attribute__((noclone,noinline))
|
|
|
128 fn14 (int * __restrict__ p1, int * __restrict__ p2, short * __restrict__ p3, int size)
|
|
|
129 {
|
|
|
130 int i;
|
|
|
131
|
|
|
132 for (i = 0; i < size; i++)
|
|
|
133 p1[i] = ((p2[i] == 0) == (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
134 }
|
|
|
135
|
|
|
136 void __attribute__((noclone,noinline))
|
|
|
137 fn15 (int * __restrict__ p1, int * __restrict__ p2, long long * __restrict__ p3, int size)
|
|
|
138 {
|
|
|
139 int i;
|
|
|
140
|
|
|
141 for (i = 0; i < size; i++)
|
|
|
142 p1[i] = ((p2[i] == 0) == (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
143 }
|
|
|
144
|
|
|
145 void __attribute__((noclone,noinline))
|
|
|
146 fn16 (int * __restrict__ p1, int * __restrict__ p2, int * __restrict__ p3, int size)
|
|
|
147 {
|
|
|
148 int i;
|
|
|
149
|
|
|
150 for (i = 0; i < size; i++)
|
|
|
151 p1[i] = ((p2[i] == 0) != (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
152 }
|
|
|
153
|
|
|
154 void __attribute__((noclone,noinline))
|
|
|
155 fn17 (int * __restrict__ p1, int * __restrict__ p2, short * __restrict__ p3, int size)
|
|
|
156 {
|
|
|
157 int i;
|
|
|
158
|
|
|
159 for (i = 0; i < size; i++)
|
|
|
160 p1[i] = ((p2[i] == 0) != (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
161 }
|
|
|
162
|
|
|
163 void __attribute__((noclone,noinline))
|
|
|
164 fn18 (int * __restrict__ p1, int * __restrict__ p2, long long * __restrict__ p3, int size)
|
|
|
165 {
|
|
|
166 int i;
|
|
|
167
|
|
|
168 for (i = 0; i < size; i++)
|
|
|
169 p1[i] = ((p2[i] == 0) != (unsigned)(p3[i] == 0)) + (p2[i] == 0);
|
|
|
170 }
|
|
|
171
|
|
|
172 int eq (int i1, int i2) { return i1 == i2; }
|
|
|
173 int ne (int i1, int i2) { return i1 != i2; }
|
|
|
174 int lt (int i1, int i2) { return i1 < i2; }
|
|
|
175 int le (int i1, int i2) { return i1 <= i2; }
|
|
|
176 int gt (int i1, int i2) { return i1 > i2; }
|
|
|
177 int ge (int i1, int i2) { return i1 >= i2; }
|
|
|
178
|
|
|
179 typedef int (*cmp_fn)(int, int);
|
|
|
180
|
|
|
181 void
|
|
|
182 check (int *p, cmp_fn fn)
|
|
|
183 {
|
|
|
184 int i;
|
|
|
185
|
|
|
186 for (i = 0; i < 32; i++)
|
|
|
187 {
|
|
|
188 int t1 = ((i % 4) > 1) == 0;
|
|
|
189 int t2 = (i % 2) == 0;
|
|
|
190 int res = fn (t1, t2) + t1;
|
|
|
191 if (p[i] != res)
|
|
|
192 __builtin_abort ();
|
|
|
193 }
|
|
|
194 }
|
|
|
195
|
|
|
196 int
|
|
|
197 main (int argc, char **argv)
|
|
|
198 {
|
|
|
199 int i1[32], i2[32], res[32];
|
|
|
200 short s2[32];
|
|
|
201 long long l2[32];
|
|
|
202 int i;
|
|
|
203
|
|
|
204 check_vect ();
|
|
|
205
|
|
|
206 for (i = 0; i < 32; i++)
|
|
|
207 {
|
|
|
208 l2[i] = i2[i] = s2[i] = i % 2;
|
|
|
209 i1[i] = (i % 4) > 1;
|
|
|
210 asm ("":::"memory");
|
|
|
211 }
|
|
|
212
|
|
|
213 fn1 (res, i1, i2, 32);
|
|
|
214 check (res, gt);
|
|
|
215 fn2 (res, i1, s2, 32);
|
|
|
216 check (res, gt);
|
|
|
217 fn3 (res, i1, l2, 32);
|
|
|
218 check (res, gt);
|
|
|
219
|
|
|
220 fn4 (res, i1, i2, 32);
|
|
|
221 check (res, ge);
|
|
|
222 fn5 (res, i1, s2, 32);
|
|
|
223 check (res, ge);
|
|
|
224 fn6 (res, i1, l2, 32);
|
|
|
225 check (res, ge);
|
|
|
226
|
|
|
227 fn7 (res, i1, i2, 32);
|
|
|
228 check (res, lt);
|
|
|
229 fn8 (res, i1, s2, 32);
|
|
|
230 check (res, lt);
|
|
|
231 fn9 (res, i1, l2, 32);
|
|
|
232 check (res, lt);
|
|
|
233
|
|
|
234 fn10 (res, i1, i2, 32);
|
|
|
235 check (res, le);
|
|
|
236 fn11 (res, i1, s2, 32);
|
|
|
237 check (res, le);
|
|
|
238 fn12 (res, i1, l2, 32);
|
|
|
239 check (res, le);
|
|
|
240
|
|
|
241 fn13 (res, i1, i2, 32);
|
|
|
242 check (res, eq);
|
|
|
243 fn14 (res, i1, s2, 32);
|
|
|
244 check (res, eq);
|
|
|
245 fn15 (res, i1, l2, 32);
|
|
|
246 check (res, eq);
|
|
|
247
|
|
|
248 fn16 (res, i1, i2, 32);
|
|
|
249 check (res, ne);
|
|
|
250 fn17 (res, i1, s2, 32);
|
|
|
251 check (res, ne);
|
|
|
252 fn18 (res, i1, l2, 32);
|
|
|
253 check (res, ne);
|
|
|
254 }
|
|
|
255
|
|
|
256 /* { dg-final { scan-tree-dump-times "VECTORIZED" 18 "vect" { target sse4_runtime } } } */
|
