| 1 |
298 |
jeremybenn |
/* { dg-do compile { target i?86-*-* x86_64-*-* } } */
|
| 2 |
|
|
/* { dg-require-effective-target ilp32 } */
|
| 3 |
|
|
/* { dg-require-effective-target sse2 } */
|
| 4 |
|
|
/* { dg-skip-if "" { i?86-*-* x86_64-*-* } { "-march=*" } { "-march=athlon" } } */
|
| 5 |
|
|
/* { dg-options "-O2 -fprefetch-loop-arrays -march=athlon -msse2 -mfpmath=sse --param simultaneous-prefetches=100 --param max-unrolled-insns=1 -fdump-tree-aprefetch-details -fdump-tree-optimized" } */
|
| 6 |
|
|
|
| 7 |
|
|
#define K 1000000
|
| 8 |
|
|
int a[K], b[K];
|
| 9 |
|
|
|
| 10 |
|
|
void test(int *p)
|
| 11 |
|
|
{
|
| 12 |
|
|
unsigned i;
|
| 13 |
|
|
|
| 14 |
|
|
/* Nontemporal store should be used for a. */
|
| 15 |
|
|
for (i = 0; i < K; i++)
|
| 16 |
|
|
a[i] = 0;
|
| 17 |
|
|
|
| 18 |
|
|
/* Nontemporal store should be used for a, nontemporal prefetch for b. */
|
| 19 |
|
|
for (i = 0; i < K; i++)
|
| 20 |
|
|
a[i] = b[i];
|
| 21 |
|
|
|
| 22 |
|
|
/* Nontemporal store should not be used here (only write and read temporal
|
| 23 |
|
|
prefetches). */
|
| 24 |
|
|
for (i = 0; i < K - 10000; i++)
|
| 25 |
|
|
a[i + 10000] = a[i];
|
| 26 |
|
|
|
| 27 |
|
|
/* Nontemporal store should not be used here (only write and read nontemporal
|
| 28 |
|
|
prefetches). */
|
| 29 |
|
|
for (i = 0; i < K - 100000; i++)
|
| 30 |
|
|
a[i + 100000] = a[i];
|
| 31 |
|
|
|
| 32 |
|
|
/* Nontemporal store should be used neither for a nor for p, as we do not know
|
| 33 |
|
|
whether they alias or not. */
|
| 34 |
|
|
for (i = 0; i < K; i++)
|
| 35 |
|
|
{
|
| 36 |
|
|
a[i] = 0;
|
| 37 |
|
|
*p++ = 1;
|
| 38 |
|
|
}
|
| 39 |
|
|
|
| 40 |
|
|
/* Nontemporal store should not be used for a, as we do not know whether its
|
| 41 |
|
|
value will be reused or not. */
|
| 42 |
|
|
for (i = 0; i < 1000; i++)
|
| 43 |
|
|
a[i] = 0;
|
| 44 |
|
|
}
|
| 45 |
|
|
|
| 46 |
|
|
/* { dg-final { scan-tree-dump-times "Issued prefetch" 5 "aprefetch" } } */
|
| 47 |
|
|
/* { dg-final { scan-tree-dump-times "Issued nontemporal prefetch" 3 "aprefetch" } } */
|
| 48 |
|
|
/* { dg-final { scan-tree-dump-times "nontemporal store" 2 "aprefetch" } } */
|
| 49 |
|
|
|
| 50 |
|
|
/* { dg-final { scan-tree-dump-times "builtin_prefetch" 8 "optimized" } } */
|
| 51 |
|
|
/* { dg-final { scan-tree-dump-times "=\\{nt\\}" 2 "optimized" } } */
|
| 52 |
|
|
/* { dg-final { scan-tree-dump-times "__builtin_ia32_mfence" 2 "optimized" } } */
|
| 53 |
|
|
|
| 54 |
|
|
/* { dg-final { scan-assembler-times "prefetchw" 5 } } */
|
| 55 |
|
|
/* { dg-final { scan-assembler-times "prefetcht" 1 } } */
|
| 56 |
|
|
/* { dg-final { scan-assembler-times "prefetchnta" 2 } } */
|
| 57 |
|
|
/* { dg-final { scan-assembler-times "movnti" 2 } } */
|
| 58 |
|
|
/* { dg-final { scan-assembler-times "mfence" 2 } } */
|
| 59 |
|
|
|
| 60 |
|
|
/* { dg-final { cleanup-tree-dump "aprefetch" } } */
|
| 61 |
|
|
/* { dg-final { cleanup-tree-dump "optimized" } } */
|
