1  /* { dg-do compile } */
       2  /* { dg-options "-O3" } */
       3  
       4  #include <arm_neon.h>
       5  
       6  #define TEST_MUL_UNIFORM(name, q, vectype, ts) \
       7    vectype test_ ## name ## q ## _ ## ts (vectype a, vectype b, vectype c) \
       8  	{ \
       9  		vectype t0 = name ## q ## _n_ ## ts (a, c[1]); \
      10  		vectype t1 = name ## q ## _n_ ## ts (b, c[1]); \
      11  		return vmul ## q ## _ ## ts (t0, t1); \
      12  	}
      13  
      14  TEST_MUL_UNIFORM (vmul, , int16x4_t, s16)
      15  TEST_MUL_UNIFORM (vmul, , uint16x4_t, u16)
      16  TEST_MUL_UNIFORM (vmul, , int32x2_t, s32)
      17  TEST_MUL_UNIFORM (vmul, , uint32x2_t, u32)
      18  TEST_MUL_UNIFORM (vmul, , float32x2_t, f32)
      19  TEST_MUL_UNIFORM (vmul, q, int16x8_t, s16)
      20  TEST_MUL_UNIFORM (vmul, q, uint16x8_t, u16)
      21  TEST_MUL_UNIFORM (vmul, q, int32x4_t, s32)
      22  TEST_MUL_UNIFORM (vmul, q, uint32x4_t, u32)
      23  TEST_MUL_UNIFORM (vmul, q, float32x4_t, f32)
      24  TEST_MUL_UNIFORM (vmul, q, float64x2_t, f64)
      25  
      26  #define TEST_MLX_UNIFORM(name, q, vectype, ts) \
      27    vectype test_ ## name ## q ## _ ## ts (vectype acc, vectype a, vectype b) \
      28  	{ \
      29  		acc = name ## q ## _n_ ## ts (acc, a, b[1]); \
      30  		return name ## q ## _n_ ## ts (acc, a, b[1]); \
      31  	}
      32  
      33  TEST_MLX_UNIFORM (vmla, , int16x4_t, s16)
      34  TEST_MLX_UNIFORM (vmla, , uint16x4_t, u16)
      35  TEST_MLX_UNIFORM (vmla, , int32x2_t, s32)
      36  TEST_MLX_UNIFORM (vmla, , uint32x2_t, u32)
      37  TEST_MLX_UNIFORM (vmla, , float32x2_t, f32)
      38  TEST_MLX_UNIFORM (vmla, q, int16x8_t, s16)
      39  TEST_MLX_UNIFORM (vmla, q, uint16x8_t, u16)
      40  TEST_MLX_UNIFORM (vmla, q, int32x4_t, s32)
      41  TEST_MLX_UNIFORM (vmla, q, uint32x4_t, u32)
      42  TEST_MLX_UNIFORM (vmla, q, float32x4_t, f32)
      43  
      44  TEST_MLX_UNIFORM (vmls, , int16x4_t, s16)
      45  TEST_MLX_UNIFORM (vmls, , uint16x4_t, u16)
      46  TEST_MLX_UNIFORM (vmls, , int32x2_t, s32)
      47  TEST_MLX_UNIFORM (vmls, , uint32x2_t, u32)
      48  TEST_MLX_UNIFORM (vmls, , float32x2_t, f32)
      49  TEST_MLX_UNIFORM (vmls, q, int16x8_t, s16)
      50  TEST_MLX_UNIFORM (vmls, q, uint16x8_t, u16)
      51  TEST_MLX_UNIFORM (vmls, q, int32x4_t, s32)
      52  TEST_MLX_UNIFORM (vmls, q, uint32x4_t, u32)
      53  TEST_MLX_UNIFORM (vmls, q, float32x4_t, f32)
      54  
      55  #define TEST_MUL_LONG(name, rettype, intype, ts, rs) \
      56    rettype test_ ## name ## ts (intype a, intype b, intype c) \
      57  	{ \
      58  		rettype t0 = name ## ts (a, c[1]); \
      59  		rettype t1 = name ## ts (b, c[1]); \
      60  		return vqaddq ## _ ## rs (t0, t1); \
      61  	}
      62  
      63  TEST_MUL_LONG (vmull_n_, int32x4_t, int16x4_t, s16, s32)
      64  TEST_MUL_LONG (vmull_n_, uint32x4_t, uint16x4_t, u16, u32)
      65  TEST_MUL_LONG (vmull_n_, int64x2_t, int32x2_t, s32, s64)
      66  TEST_MUL_LONG (vmull_n_, uint64x2_t, uint32x2_t, u32, u64)
      67  
      68  TEST_MUL_LONG (vqdmull_n_, int32x4_t, int16x4_t, s16, s32)
      69  TEST_MUL_LONG (vqdmull_n_, int64x2_t, int32x2_t, s32, s64)
      70  
      71  #define TEST_MLX_LONG(name, rettype, intype, ts, rs) \
      72    rettype test_ ## name ## _ ## ts (rettype acc, intype a, intype b) \
      73  	{ \
      74  		acc = name ## ts (acc, a, b[1]); \
      75  		return name ## ts (acc, a, b[1]); \
      76  	}
      77  
      78  TEST_MLX_LONG (vmlal_n_, int32x4_t, int16x4_t, s16, s32)
      79  TEST_MLX_LONG (vmlal_n_, uint32x4_t, uint16x4_t, u16, u32)
      80  TEST_MLX_LONG (vmlal_n_, int64x2_t, int32x2_t, s32, s64)
      81  TEST_MLX_LONG (vmlal_n_, uint64x2_t, uint32x2_t, u32, u64)
      82  
      83  TEST_MLX_LONG (vmlsl_n_, int32x4_t, int16x4_t, s16, s32)
      84  TEST_MLX_LONG (vmlsl_n_, uint32x4_t, uint16x4_t, u16, u32)
      85  TEST_MLX_LONG (vmlsl_n_, int64x2_t, int32x2_t, s32, s64)
      86  TEST_MLX_LONG (vmlsl_n_, uint64x2_t, uint32x2_t, u32, u64)
      87  
      88  TEST_MLX_LONG (vqdmlal_n_, int32x4_t, int16x4_t, s16, s32)
      89  TEST_MLX_LONG (vqdmlal_n_, int64x2_t, int32x2_t, s32, s64)
      90  
      91  TEST_MLX_LONG (vqdmlsl_n_, int32x4_t, int16x4_t, s16, s32)
      92  TEST_MLX_LONG (vqdmlsl_n_, int64x2_t, int32x2_t, s32, s64)
      93  
      94  /* { dg-final { scan-assembler-not "dup\\t" } } */