1 /* PR rtl-optimization/28982. Function foo() does the equivalent of:
2
3 float tmp_results[NVARS];
4 for (int i = 0; i < NVARS; i++)
5 {
6 int inc = incs[i];
7 float *ptr = ptrs[i], result = 0;
8 for (int j = 0; j < n; j++)
9 result += *ptr, ptr += inc;
10 tmp_results[i] = result;
11 }
12 memcpy (results, tmp_results, sizeof (results));
13
14 but without the outermost loop. The idea is to create high register
15 pressure and ensure that some INC and PTR variables are spilled.
16
17 On ARM targets, sequences like "result += *ptr, ptr += inc" can
18 usually be implemented using (mem (post_modify ...)), and we do
19 indeed create such MEMs before reload for this testcase. However,
20 (post_modify ...) is not a valid address for coprocessor loads, so
21 for -mfloat-abi=softfp, reload reloads the POST_MODIFY into a base
22 register. GCC did not deal correctly with cases where the base and
23 index of the POST_MODIFY are themselves reloaded. */
24 #define NITER 4
25 #define NVARS 20
26 #define MULTI(X) \
27 X( 0), X( 1), X( 2), X( 3), X( 4), X( 5), X( 6), X( 7), X( 8), X( 9), \
28 X(10), X(11), X(12), X(13), X(14), X(15), X(16), X(17), X(18), X(19)
29
30 #define DECLAREI(INDEX) inc##INDEX = incs[INDEX]
31 #define DECLAREF(INDEX) *ptr##INDEX = ptrs[INDEX], result##INDEX = 0
32 #define LOOP(INDEX) result##INDEX += *ptr##INDEX, ptr##INDEX += inc##INDEX
33 #define COPYOUT(INDEX) results[INDEX] = result##INDEX
34
35 float *ptrs[NVARS];
36 float results[NVARS];
37 int incs[NVARS];
38
39 void __attribute__((noinline))
foo(int n)40 foo (int n)
41 {
42 int MULTI (DECLAREI);
43 float MULTI (DECLAREF);
44 while (n--)
45 MULTI (LOOP);
46 MULTI (COPYOUT);
47 }
48
49 float input[NITER * NVARS];
50
51 int
main(void)52 main (void)
53 {
54 int i;
55
56 for (i = 0; i < NVARS; i++)
57 ptrs[i] = input + i, incs[i] = i;
58 for (i = 0; i < NITER * NVARS; i++)
59 input[i] = i;
60 foo (NITER);
61 for (i = 0; i < NVARS; i++)
62 if (results[i] != i * NITER * (NITER + 1) / 2)
63 return 1;
64 return 0;
65 }
66