1 /* { dg-do run } */
2 /* { dg-options "-O3 -floop-unroll-and-jam -fno-tree-loop-im --param unroll-jam-min-percent=0 -fdump-tree-unrolljam-details" } */
3 /* { dg-additional-options "--param max-completely-peel-times=16" { target { s390*-*-* } } } */
4 /* { dg-require-effective-target int32plus } */
5
6 #include <stdio.h>
7 extern unsigned int a[];
8 extern unsigned int b[];
9 extern unsigned int aa[][1024];
10 unsigned int checksum;
11 void checkaa(void)
12 {
13 unsigned sum = 1;
14 unsigned long i, j;
15 for (i = 0; i < 1024; i++) {
16 for (j = 0; j < 16; j++) {
17 sum += aa[j][i]*31+47;
18 }
19 }
20 checksum = checksum * 27 + sum;
21 //printf(" %d\n", sum);
22 }
23
24 void checkb(void)
25 {
26 unsigned sum = 1;
27 unsigned long i, j;
28 for (i = 0; i < 1024; i++) {
29 sum += b[i]*31+47;
30 }
31 checksum = checksum * 27 + sum;
32 //printf(" %d\n", sum);
33 }
34
35 #define TEST(name, body, test) \
36 static void __attribute__((noinline,noclone)) name (unsigned long n, unsigned long m) \
37 { \
38 unsigned i, j; \
39 for (i = 1; i < m; i++) { \
40 for (j = 1; j < n; j++) { \
41 body; \
42 } \
43 } \
44 test; \
45 } \
46 static void __attribute__((noinline,noclone,optimize("O1"))) name ## noopt (unsigned long n, unsigned long m) \
47 { \
48 unsigned long i, j; \
49 for (i = 1; i < m; i++) { \
50 for (j = 1; j < n; j++) { \
51 body; \
52 } \
53 } \
54 test; \
55 }
56 TEST(foo1, aa[i+1][j+1]=aa[i][j] * aa[i][j] / 2, checkaa()) //ok, -1,-1
57 TEST(foo2, aa[i][j+1]=3*aa[i+1][j], checkaa()) //notok, 1,-1
58 TEST(foo3, aa[i+1][j-1]=aa[i][j] * aa[i][j] / 2, checkaa()) //notok, -1,1
59 TEST(foo4, aa[i][j] = aa[i-1][j+1] * aa[i-1][j+1] / 2, checkaa()) //notok, -1,1
60 TEST(foo5, aa[i][j] = aa[i+1][j+1] * aa[i+1][j+1] / 2, checkaa()) //ok, 1,1
61 TEST(foo6, aa[i][j] = aa[i+1][j] * aa[i+1][j] / 2, checkaa()) //ok, -1,0
62 TEST(foo61, aa[i][0] = aa[i+1][0] * aa[i+1][0] / 2, checkaa()) //notok, -1,0
63 TEST(foo62, aa[i][j/2] = aa[i+1][j/2] * aa[i+1][j/2] / 2, checkaa()) //notok, not affine
64 TEST(foo63, aa[i][j%2] = aa[i+1][j%2] * aa[i+1][j%2] / 2, checkaa()) //notok, not affine
65 TEST(foo7, aa[i+1][j] = aa[i][j] * aa[i][j] / 2, checkaa()) //ok, 1,0
66 TEST(foo9, b[j] = 3*b[j+1] + 1, checkb()) //notok, 0,-1
67 TEST(foo10, b[j] = 3*b[j] + 1, checkb()) //ok, 0,0
68 extern int f;
69 TEST(foo11, f = b[i-1] = 1 + 3* b[i+1], checkb()) //ok, 2,0 but must reduce unroll factor to 2, (it would be incorrect with unroll-by-3, which the profitability would suggest)
70
71 /* foo8 should work as well, but currently doesn't because the distance
72 vectors we compute are too pessimistic. We compute
73 (0,1), (1,1) and (1,-1)
74 and the last one causes us to lose. */
75 TEST(foo8, b[j+1] = 3*b[j] + 1, checkb()) //ok, 0,1
76
77 int f;
78 unsigned int a[1024];
79 unsigned int b[1024];
80 unsigned int aa[16][1024];
81 void init(void)
82 {
83 unsigned long i,j;
84 for (i = 0; i < 1024; i++) {
85 for (j = 0; j < 16; j++) {
86 aa[j][i] = ((j+1)*2+i+1) % 17;
87 }
88 a[i] = ((i+1)*31) % 19;
89 b[i] = ((i+1)*47) % 23;
90 }
91 checksum = 1;
92 }
93
94 #define RUN(name) \
95 printf(" %s\n", #name); \
96 init();for(i=0;i<4;i++)name##noopt(32,8); checka = checksum; \
97 init();for(i=0;i<4;i++)name(32,8); \
98 if (checka != checksum) fail = 1; \
99 printf("%sok %s\n", checka != checksum ? "NOT " : "", #name);
100
101 int main()
102 {
103 int fail = 0;
104 int i;
105 unsigned checka;
106 RUN(foo1);
107 RUN(foo2);
108 RUN(foo3);
109 RUN(foo4);
110 RUN(foo5);
111 RUN(foo6);
112 RUN(foo61);
113 RUN(foo62);
114 RUN(foo63);
115 RUN(foo7);
116 RUN(foo8);
117 RUN(foo9);
118 RUN(foo10);
119 RUN(foo11);
120 if (fail)
121 __builtin_abort();
122 return fail;
123 }
124
125 /* Six loops should be unroll-jammed (actually seven, but see above). */
126 /* { dg-final { scan-tree-dump-times "applying unroll and jam" 6 "unrolljam" } } */