| 1 |
/* PR target/99195. */ |
| 2 |
/* Check that we take advantage of 64-bit Advanced SIMD operations clearing |
| 3 |
the top half of the vector register and no explicit zeroing instructions |
| 4 |
are emitted. */ |
| 5 |
/* { dg-do compile } */ |
| 6 |
/* { dg-options "-O -march=armv8.2-a+dotprod" } */ |
| 7 |
|
| 8 |
#include <arm_neon.h> |
| 9 |
|
| 10 |
#define DOTPROD(OT,AT,IT1,IT2,OP,S) \ |
| 11 |
OT \ |
| 12 |
foo_##OP##_##S (AT a, IT1 b, IT2 c) \ |
| 13 |
{ \ |
| 14 |
AT zeros = vcreate_##S (0); \ |
| 15 |
return vcombine_##S (v##OP##_##S (a, b, c), zeros); \ |
| 16 |
} |
| 17 |
|
| 18 |
#define DOTPROD_IDX(OT,AT,IT1,IT2,OP,S) \ |
| 19 |
OT \ |
| 20 |
foo_##OP##_##S (AT a, IT1 b, IT2 c) \ |
| 21 |
{ \ |
| 22 |
AT zeros = vcreate_##S (0); \ |
| 23 |
return vcombine_##S (v##OP##_##S (a, b, c, 1), zeros); \ |
| 24 |
} |
| 25 |
|
| 26 |
DOTPROD (int32x4_t, int32x2_t, int8x8_t, int8x8_t, dot, s32) |
| 27 |
DOTPROD (uint32x4_t, uint32x2_t, uint8x8_t, uint8x8_t, dot, u32) |
| 28 |
DOTPROD_IDX (int32x4_t, int32x2_t, int8x8_t, int8x8_t, dot_lane, s32) |
| 29 |
DOTPROD_IDX (uint32x4_t, uint32x2_t, uint8x8_t, uint8x8_t, dot_lane, u32) |
| 30 |
|
| 31 |
#pragma GCC target ("+i8mm") |
| 32 |
DOTPROD (int32x4_t, int32x2_t, uint8x8_t, int8x8_t, usdot, s32) |
| 33 |
DOTPROD_IDX (int32x4_t, int32x2_t, uint8x8_t, int8x8_t, usdot_lane, s32) |
| 34 |
DOTPROD_IDX (int32x4_t, int32x2_t, int8x8_t, uint8x8_t, sudot_lane, s32) |
| 35 |
|
| 36 |
/* { dg-final { scan-assembler-not {\tfmov\t} } } */ |
| 37 |
/* { dg-final { scan-assembler-not {\tmov\t} } } */ |
| 38 |
|