| 1 |
/* Test whether using target specific options, we can generate SSE2 code on |
| 2 |
32-bit, which does not generate SSE2 by default, but still generate 387 code |
| 3 |
for a function that doesn't use attribute((option)). */ |
| 4 |
/* { dg-do compile } */ |
| 5 |
/* { dg-require-effective-target ia32 } */ |
| 6 |
/* { dg-options "-O3 -ftree-vectorize -mno-sse" } */ |
| 7 |
|
| 8 |
#ifndef SIZE |
| 9 |
#define SIZE 1024 |
| 10 |
#endif |
| 11 |
|
| 12 |
float a[SIZE] __attribute__((__aligned__(16))); |
| 13 |
float b[SIZE] __attribute__((__aligned__(16))); |
| 14 |
float c[SIZE] __attribute__((__aligned__(16))); |
| 15 |
|
| 16 |
void sse_addnums (void) __attribute__ ((__target__ ("sse2"))); |
| 17 |
|
| 18 |
void |
| 19 |
sse_addnums (void) |
| 20 |
{ |
| 21 |
int i = 0; |
| 22 |
for (; i < SIZE; ++i) |
| 23 |
a[i] = b[i] + c[i]; |
| 24 |
} |
| 25 |
|
| 26 |
void |
| 27 |
i387_subnums (void) |
| 28 |
{ |
| 29 |
int i = 0; |
| 30 |
for (; i < SIZE; ++i) |
| 31 |
a[i] = b[i] - c[i]; |
| 32 |
} |
| 33 |
|
| 34 |
/* { dg-final { scan-assembler "addps\[ \t\]" } } */ |
| 35 |
/* { dg-final { scan-assembler "fsubs\[ \t\]" } } */ |