blob: 176830580a9e216d02ba84d0904a9713ee881a73 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
|
/* { dg-do compile { target { { i?86-*-* x86_64-*-* } && ia32 } } } */
/* { dg-require-effective-target sse2 } */
/* { dg-options "-O2 -fprefetch-loop-arrays -march=athlon -msse2 -mfpmath=sse --param simultaneous-prefetches=100 --param min-insn-to-prefetch-ratio=6 -fdump-tree-aprefetch-details" } */
#define N 1000
#define K 900
double a[N][N];
double test(void)
{
unsigned i, j;
double sum = 0;
/* Here, we should use non-temporal prefetch instruction. */
for (i = 0; i < K; i++)
for (j = 0; j < K; j++)
sum += a[i][j];
/* Here, we should not use non-temporal prefetch instruction, since the
value of a[i+10][j] is reused in L2 cache. */
for (i = 0; i < K; i++)
for (j = 0; j < K; j++)
sum += a[i][j] * a[i + 10][j];
/* Here, we should use non-temporal prefetch instruction, since the
value of a[i+100][j] is too far to be reused in L2 cache. */
for (i = 0; i < K; i++)
for (j = 0; j < K; j++)
sum += a[i][j] * a[i + 100][j];
/* Here, temporal prefetches should be used, since the volume of the
memory accesses is smaller than L2 cache. */
for (i = 0; i < 100; i++)
for (j = 0; j < 100; j++)
sum += a[i][j] * a[i + 100][j];
/* Temporal prefetches should be used here (even though the accesses to
a[j][i] are independent, the same cache line is almost always hit
every N iterations). */
for (i = 0; i < N; i++)
for (j = 0; j < N; j++)
sum += a[j][i];
return sum;
}
/* { dg-final { scan-tree-dump-times "Issued prefetch" 5 "aprefetch" } } */
/* { dg-final { scan-tree-dump-times "Issued nontemporal prefetch" 3 "aprefetch" } } */
/* { dg-final { scan-assembler-times "prefetcht" 5 } } */
/* { dg-final { scan-assembler-times "prefetchnta" 3 } } */
/* { dg-final { cleanup-tree-dump "aprefetch" } } */
|
