| 1 |
/* { dg-do compile } */ |
| 2 |
/* { dg-options "-O2 -ftree-vectorize -msse4 -mtune=generic" } */ |
| 3 |
|
| 4 |
/* We need SSE4 so the backend recognizes a { 0, 5, 2, 7 } constant |
| 5 |
permutation as supported as the vectorizer wants to generate |
| 6 |
|
| 7 |
vect__6.10_24 = vect__3.6_20 - vect__5.9_23; |
| 8 |
vect__6.11_25 = vect__3.6_20 + vect__5.9_23; |
| 9 |
_26 = VEC_PERM_EXPR <vect__6.10_24, vect__6.11_25, { 0, 5, 2, 7 }>; |
| 10 |
|
| 11 |
See also the ??? comment about using and/andn/or in expand_vec_perm_blend |
| 12 |
for non-SSE4 targets. */ |
| 13 |
|
| 14 |
void testf (float * __restrict__ p, float * __restrict q) |
| 15 |
{ |
| 16 |
p[0] = p[0] - q[0]; |
| 17 |
p[1] = p[1] + q[1]; |
| 18 |
p[2] = p[2] - q[2]; |
| 19 |
p[3] = p[3] + q[3]; |
| 20 |
} |
| 21 |
|
| 22 |
/* { dg-final { scan-assembler "addsubps" } } */ |