| 1 |
/* { dg-do compile } */ |
| 2 |
/* { dg-options "-O3 -march=armv8.2-a+nofp16" } */ |
| 3 |
/* { dg-skip-if "" { *-*-* } { "-mcpu=*" } { "" } } */ |
| 4 |
|
| 5 |
#pragma GCC target ("arch=armv8.2-a+nofp16") |
| 6 |
|
| 7 |
_Float16 |
| 8 |
foo_v8 (_Float16 x, _Float16 y, unsigned int *eval) |
| 9 |
{ |
| 10 |
*eval = __FLT_EVAL_METHOD__; |
| 11 |
return x * x + y; |
| 12 |
} |
| 13 |
|
| 14 |
__fp16 |
| 15 |
bar_v8 (__fp16 x, __fp16 y, unsigned int *eval) |
| 16 |
{ |
| 17 |
*eval = __FLT_EVAL_METHOD__; |
| 18 |
return x * x + y; |
| 19 |
} |
| 20 |
|
| 21 |
#pragma GCC target ("arch=armv8.2-a+fp16") |
| 22 |
|
| 23 |
_Float16 |
| 24 |
foo_v82 (_Float16 x, _Float16 y, unsigned int *eval) |
| 25 |
{ |
| 26 |
*eval = __FLT_EVAL_METHOD__; |
| 27 |
return x * x + y; |
| 28 |
} |
| 29 |
|
| 30 |
__fp16 |
| 31 |
bar_v82 (__fp16 x, __fp16 y, unsigned int *eval) |
| 32 |
{ |
| 33 |
*eval = __FLT_EVAL_METHOD__; |
| 34 |
return x * x + y; |
| 35 |
} |
| 36 |
|
| 37 |
/* Test that we merge to FMA operations. This indicates that we are not |
| 38 |
making extraneous conversions between modes. */ |
| 39 |
|
| 40 |
/* Three FMA operations in 32-bit precision, from foo_v8, bar_v8, bar_v82. */ |
| 41 |
/* { dg-final { scan-assembler-times "fmadd\ts\[0-9\]\+" 3 } } */ |
| 42 |
|
| 43 |
/* One FMA operation in 16-bit precision, from foo_v82. */ |
| 44 |
/* { dg-final { scan-assembler-times "fmadd\th\[0-9\]\+" 1 } } */ |
| 45 |
|
| 46 |
/* Test that we are resetting the __FLT_EVAL_METHOD__. */ |
| 47 |
/* { dg-final { scan-assembler-times "mov\tw\[0-9\]\+, 16" 2 } } */ |
| 48 |
/* { dg-final { scan-assembler-times "str\twzr" 2 } } */ |