1  /* { dg-do compile } */
       2  /* { dg-options "-O3" } */
       3  
       4  #include <arm_neon.h>
       5  
       6  #define TEST_MULL_VEC(name, rettype, intype, ts, rs) \
       7    rettype test_ ## name ## _ ## ts (intype a, intype b, intype c) \
       8  	{ \
       9  		rettype t0 = name ## _ ## ts (vget_high_ ## ts (a), \
      10  					      vget_high_ ## ts (c)); \
      11  		rettype t1 = name ## _ ## ts (vget_high_ ## ts (b), \
      12  					      vget_high_ ## ts (c)); \
      13  		return vqaddq ## _ ## rs (t0, t1); \
      14  	}
      15  
      16  TEST_MULL_VEC (vmull, int16x8_t, int8x16_t, s8, s16)
      17  TEST_MULL_VEC (vmull, uint16x8_t, uint8x16_t, u8, u16)
      18  TEST_MULL_VEC (vmull, int32x4_t, int16x8_t, s16, s32)
      19  TEST_MULL_VEC (vmull, uint32x4_t, uint16x8_t, u16, u32)
      20  TEST_MULL_VEC (vmull, int64x2_t, int32x4_t, s32, s64)
      21  TEST_MULL_VEC (vmull, uint64x2_t, uint32x4_t, u32, u64)
      22  
      23  TEST_MULL_VEC (vqdmull, int32x4_t, int16x8_t, s16, s32)
      24  TEST_MULL_VEC (vqdmull, int64x2_t, int32x4_t, s32, s64)
      25  
      26  #define TEST_MULL_N(name, rettype, intype, ts, rs) \
      27    rettype test_ ## name ## _ ## ts (intype a, intype b, intype c) \
      28  	{ \
      29  		rettype t0 = name ## _ ## ts (vget_high_ ## ts (a), b[1]); \
      30  		rettype t1 = name ## _ ## ts (vget_high_ ## ts (a), c[1]); \
      31  		return vqaddq ## _ ## rs (t0, t1); \
      32  	}
      33  
      34  TEST_MULL_N (vmull_n, int32x4_t, int16x8_t, s16, s32)
      35  TEST_MULL_N (vmull_n, uint32x4_t, uint16x8_t, u16, u32)
      36  TEST_MULL_N (vmull_n, int64x2_t, int32x4_t, s32, s64)
      37  TEST_MULL_N (vmull_n, uint64x2_t, uint32x4_t, u32, u64)
      38  
      39  TEST_MULL_N (vqdmull_n, int32x4_t, int16x8_t, s16, s32)
      40  TEST_MULL_N (vqdmull_n, int64x2_t, int32x4_t, s32, s64)
      41  
      42  #define TEST_MLXL_VEC(name, rettype, intype, ts) \
      43    rettype test_ ## name ## _ ## ts (rettype acc, intype a, intype b, \
      44  				    intype c) \
      45  	{ \
      46  		acc = name ## _ ## ts (acc, vget_high_ ## ts (a), \
      47  					    vget_high_ ## ts (b)); \
      48  		return name ## _ ## ts (acc, vget_high_ ## ts (a), \
      49  					     vget_high_ ## ts (c)); \
      50  	}
      51  
      52  TEST_MLXL_VEC (vmlal, int16x8_t, int8x16_t, s8)
      53  TEST_MLXL_VEC (vmlal, uint16x8_t, uint8x16_t, u8)
      54  TEST_MLXL_VEC (vmlal, int32x4_t, int16x8_t, s16)
      55  TEST_MLXL_VEC (vmlal, uint32x4_t, uint16x8_t, u16)
      56  
      57  TEST_MLXL_VEC (vmlsl, int16x8_t, int8x16_t, s8)
      58  TEST_MLXL_VEC (vmlsl, uint16x8_t, uint8x16_t, u8)
      59  TEST_MLXL_VEC (vmlsl, int32x4_t, int16x8_t, s16)
      60  TEST_MLXL_VEC (vmlsl, uint32x4_t, uint16x8_t, u16)
      61  
      62  #define TEST_MLXL_N(name, rettype, intype, ts) \
      63    rettype test_ ## name ## _ ## ts (rettype acc, intype a, intype b) \
      64  	{ \
      65  		acc = name ## _ ## ts (acc, vget_high_ ## ts (a), b[1]); \
      66  		return name ## _ ## ts (acc, vget_high_ ## ts (a), b[1]); \
      67  	}
      68  
      69  TEST_MLXL_N (vmlal_n, int32x4_t, int16x8_t, s16)
      70  TEST_MLXL_N (vmlal_n, uint32x4_t, uint16x8_t, u16)
      71  TEST_MLXL_N (vmlal_n, int64x2_t, int32x4_t, s32)
      72  TEST_MLXL_N (vmlal_n, uint64x2_t, uint32x4_t, u32)
      73  
      74  TEST_MLXL_N (vmlsl_n, int32x4_t, int16x8_t, s16)
      75  TEST_MLXL_N (vmlsl_n, uint32x4_t, uint16x8_t, u16)
      76  TEST_MLXL_N (vmlsl_n, int64x2_t, int32x4_t, s32)
      77  TEST_MLXL_N (vmlsl_n, uint64x2_t, uint32x4_t, u32)
      78  
      79  TEST_MLXL_N (vqdmlal_n, int32x4_t, int16x8_t, s16)
      80  TEST_MLXL_N (vqdmlal_n, int64x2_t, int32x4_t, s32)
      81  
      82  TEST_MLXL_N (vqdmlsl_n, int32x4_t, int16x8_t, s16)
      83  TEST_MLXL_N (vqdmlsl_n, int64x2_t, int32x4_t, s32)
      84  
      85  /* { dg-final { scan-assembler-not "dup\\t" } } */