1  /* { dg-do assemble { target { aarch64*-*-* } } } */
       2  /* { dg-require-effective-target arm_v8_2a_i8mm_ok } */
       3  /* { dg-add-options arm_v8_2a_i8mm }  */
       4  /* { dg-additional-options "-save-temps" } */
       5  /* { dg-final { check-function-bodies "**" "" {-O[^0]} } } */
       6  /* { dg-skip-if "" { *-*-* } { "-fno-fat-lto-objects" } } */
       7  
       8  #include <arm_neon.h>
       9  
      10  /* Unsigned-Signed Dot Product instructions.  */
      11  
      12  /*
      13  **ufoo:
      14  **	usdot	v0\.2s, v1\.8b, v2\.8b
      15  **	ret
      16  */
      17  int32x2_t ufoo (int32x2_t r, uint8x8_t x, int8x8_t y)
      18  {
      19    return vusdot_s32 (r, x, y);
      20  }
      21  
      22  /*
      23  **ufooq:
      24  **	usdot	v0\.4s, v1\.16b, v2\.16b
      25  **	ret
      26  */
      27  int32x4_t ufooq (int32x4_t r, uint8x16_t x, int8x16_t y)
      28  {
      29    return vusdotq_s32 (r, x, y);
      30  }
      31  
      32  /*
      33  **ufoo_lane:
      34  **	usdot	v0\.2s, v1\.8b, v2\.4b\[0\]
      35  **	ret
      36  */
      37  int32x2_t ufoo_lane (int32x2_t r, uint8x8_t x, int8x8_t y)
      38  {
      39    return vusdot_lane_s32 (r, x, y, 0);
      40  }
      41  
      42  /*
      43  **ufoo_laneq:
      44  **	usdot	v0\.2s, v1\.8b, v2\.4b\[2\]
      45  **	ret
      46  */
      47  int32x2_t ufoo_laneq (int32x2_t r, uint8x8_t x, int8x16_t y)
      48  {
      49    return vusdot_laneq_s32 (r, x, y, 2);
      50  }
      51  
      52  /*
      53  **ufooq_lane:
      54  **	usdot	v0\.4s, v1\.16b, v2\.4b\[1\]
      55  **	ret
      56  */
      57  int32x4_t ufooq_lane (int32x4_t r, uint8x16_t x, int8x8_t y)
      58  {
      59    return vusdotq_lane_s32 (r, x, y, 1);
      60  }
      61  
      62  /*
      63  **ufooq_laneq:
      64  **	usdot	v0\.4s, v1\.16b, v2\.4b\[3\]
      65  **	ret
      66  */
      67  int32x4_t ufooq_laneq (int32x4_t r, uint8x16_t x, int8x16_t y)
      68  {
      69    return vusdotq_laneq_s32 (r, x, y, 3);
      70  }
      71  
      72  
      73  /* Signed-Unsigned Dot Product instructions.  */
      74  
      75  /*
      76  **sfoo_lane:
      77  **	sudot	v0\.2s, v1\.8b, v2\.4b\[0\]
      78  **	ret
      79  */
      80  int32x2_t sfoo_lane (int32x2_t r, int8x8_t x, uint8x8_t y)
      81  {
      82    return vsudot_lane_s32 (r, x, y, 0);
      83  }
      84  
      85  /*
      86  **sfoo_laneq:
      87  **	sudot	v0\.2s, v1\.8b, v2\.4b\[2\]
      88  **	ret
      89  */
      90  int32x2_t sfoo_laneq (int32x2_t r, int8x8_t x, uint8x16_t y)
      91  {
      92    return vsudot_laneq_s32 (r, x, y, 2);
      93  }
      94  
      95  /*
      96  **sfooq_lane:
      97  **	sudot	v0\.4s, v1\.16b, v2\.4b\[1\]
      98  **	ret
      99  */
     100  int32x4_t sfooq_lane (int32x4_t r, int8x16_t x, uint8x8_t y)
     101  {
     102    return vsudotq_lane_s32 (r, x, y, 1);
     103  }
     104  
     105  /*
     106  **sfooq_laneq:
     107  **	sudot	v0\.4s, v1\.16b, v2\.4b\[3\]
     108  **	ret
     109  */
     110  int32x4_t sfooq_laneq (int32x4_t r, int8x16_t x, uint8x16_t y)
     111  {
     112    return vsudotq_laneq_s32 (r, x, y, 3);
     113  }
     114  
     115  /*
     116  **ufoo_untied:
     117  **	mov	v0\.8b, v1\.8b
     118  **	usdot	v0\.2s, v2\.8b, v3\.8b
     119  **	ret
     120  */
     121  int32x2_t ufoo_untied (int32x2_t unused, int32x2_t r, uint8x8_t x, int8x8_t y)
     122  {
     123    return vusdot_s32 (r, x, y);
     124  }
     125  
     126  /*
     127  **ufooq_laneq_untied:
     128  **	mov	v0\.16b, v1\.16b
     129  **	usdot	v0\.4s, v2\.16b, v3\.4b\[3\]
     130  **	ret
     131  */
     132  int32x4_t ufooq_laneq_untied (int32x2_t unused, int32x4_t r, uint8x16_t x, int8x16_t y)
     133  {
     134    return vusdotq_laneq_s32 (r, x, y, 3);
     135  }
     136