1  /* { dg-do compile } */
       2  /* { dg-options "-O2 -ftree-vectorize" } */
       3  
       4  #include <stdint.h>
       5  
       6  #define VEC_PERM(TYPE1, TYPE2)					\
       7  void __attribute__ ((noinline, noclone))			\
       8  vec_slp_##TYPE1##_##TYPE2 (TYPE1 *restrict a,			\
       9  			   TYPE2 *restrict b, int n)		\
      10  {								\
      11    for (int i = 0; i < n; ++i)					\
      12      {								\
      13        a[i * 2] += 1;						\
      14        a[i * 2 + 1] += 2;					\
      15        b[i * 2] += 3;						\
      16        b[i * 2 + 1] += 4;					\
      17      }								\
      18  }
      19  
      20  #define TEST_ALL(T)				\
      21    T (int8_t, uint16_t)				\
      22    T (uint8_t, int16_t)				\
      23    T (int16_t, uint32_t)				\
      24    T (uint16_t, int32_t)				\
      25    T (int32_t, double)				\
      26    T (uint32_t, int64_t)				\
      27    T (float, uint64_t)
      28  
      29  TEST_ALL (VEC_PERM)
      30  
      31  /* The loop should be fully-masked.  The load XFAILs for fixed-length
      32     SVE account for extra loads from the constant pool.  */
      33  /* { dg-final { scan-assembler-times {\tld1b\t} 2 { xfail { aarch64_sve && { ! vect_variable_length } } } } }*/
      34  /* { dg-final { scan-assembler-times {\tst1b\t} 2 } } */
      35  /* { dg-final { scan-assembler-times {\tld1h\t} 6 { xfail { aarch64_sve && { ! vect_variable_length } } } } } */
      36  /* { dg-final { scan-assembler-times {\tst1h\t} 6 } } */
      37  /* { dg-final { scan-assembler-times {\tld1w\t} 7 { xfail { aarch64_sve && { ! vect_variable_length } } } } } */
      38  /* { dg-final { scan-assembler-times {\tst1w\t} 7 } } */
      39  /* { dg-final { scan-assembler-times {\tld1d\t} 6 { xfail { aarch64_sve && { ! vect_variable_length } } } } } */
      40  /* { dg-final { scan-assembler-times {\tst1d\t} 6 } } */
      41  /* { dg-final { scan-assembler-not {\tldr} } } */
      42  /* { dg-final { scan-assembler-not {\tstr} } } */
      43  
      44  /* We should use WHILEs for the accesses to "a" and unpacks for the accesses
      45     to "b".  */
      46  /* { dg-final { scan-assembler-times {\twhilelo\tp[0-7]\.b} 4 } } */
      47  /* { dg-final { scan-assembler-times {\twhilelo\tp[0-7]\.h} 4 } } */
      48  /* { dg-final { scan-assembler-times {\twhilelo\tp[0-7]\.s} 6 } } */
      49  /* { dg-final { scan-assembler-not {\twhilelo\tp[0-7]\.d} } } */
      50  /* { dg-final { scan-assembler-times {\tpunpklo\tp[0-7]\.h} 7 } } */
      51  /* { dg-final { scan-assembler-times {\tpunpkhi\tp[0-7]\.h} 7 } } */
      52  
      53  /* { dg-final { scan-assembler-not {\tuqdec} } } */