| 1 |
/* { dg-do compile } */ |
| 2 |
/* { dg-options "-mdejagnu-cpu=power8 -mvsx -O3 " } */ |
| 3 |
/* { dg-require-effective-target powerpc_vsx } */ |
| 4 |
/* { dg-final { scan-assembler-times "lxvd2x" 2 } } */ |
| 5 |
/* { dg-final { scan-assembler-times "stxvd2x" 1 } } */ |
| 6 |
/* { dg-final { scan-assembler-times "xxpermdi" 3 } } */ |
| 7 |
|
| 8 |
/* Verify that swap optimization works correctly for a VSX direct splat. |
| 9 |
The three xxpermdi's that are generated correspond to two splats |
| 10 |
and the __builtin_vsx_xxpermdi. */ |
| 11 |
|
| 12 |
int printf (const char *__restrict __format, ...); |
| 13 |
typedef double __m128d __attribute__ ((__vector_size__ (16), __may_alias__)); |
| 14 |
|
| 15 |
double s1[] = {2134.3343, 6678.346}; |
| 16 |
double s2[] = {41124.234, 6678.346}; |
| 17 |
long long dd[] = {1, 2}, d[2]; |
| 18 |
union{long long l[2]; double d[2];} e; |
| 19 |
|
| 20 |
void |
| 21 |
foo () |
| 22 |
{ |
| 23 |
__m128d source1, source2, dest; |
| 24 |
__m128d a, b, c; |
| 25 |
|
| 26 |
e.d[1] = s1[1]; |
| 27 |
e.l[0] = !__builtin_isunordered(s1[0], s2[0]) |
| 28 |
&& s1[0] == s2[0] ? -1 : 0; |
| 29 |
source1 = __builtin_vec_vsx_ld (0, s1); |
| 30 |
source2 = __builtin_vec_vsx_ld (0, s2); |
| 31 |
a = __builtin_vec_splat (source1, 0); |
| 32 |
b = __builtin_vec_splat (source2, 0); |
| 33 |
c = (__m128d)__builtin_vec_cmpeq (a, b); |
| 34 |
dest = __builtin_vsx_xxpermdi (source1, c, 1); |
| 35 |
*(__m128d *)d = dest; |
| 36 |
} |