1  /* { dg-do compile } */
       2  /* { dg-options "-O2 -ftree-vectorize -msve-vector-bits=scalable" } */
       3  
       4  #include <stdint.h>
       5  
       6  #define VEC_PERM(TYPE)						\
       7  void __attribute__ ((noinline, noclone))			\
       8  vec_slp_##TYPE (TYPE *restrict a, TYPE *restrict b, int n)	\
       9  {								\
      10    for (int i = 0; i < n; ++i)					\
      11      {								\
      12        a[i] += 1;						\
      13        b[i * 4] += 2;						\
      14        b[i * 4 + 1] += 3;					\
      15        b[i * 4 + 2] += 4;					\
      16        b[i * 4 + 3] += 5;					\
      17      }								\
      18  }
      19  
      20  #define TEST_ALL(T)				\
      21    T (int8_t)					\
      22    T (uint8_t)					\
      23    T (int16_t)					\
      24    T (uint16_t)					\
      25    T (int32_t)					\
      26    T (uint32_t)					\
      27    T (int64_t)					\
      28    T (uint64_t)					\
      29    T (float)					\
      30    T (double)
      31  
      32  TEST_ALL (VEC_PERM)
      33  
      34  /* The loop should be fully-masked.  */
      35  /* { dg-final { scan-assembler-times {\tld1b\t} 10 } } */
      36  /* { dg-final { scan-assembler-times {\tst1b\t} 10 } } */
      37  /* { dg-final { scan-assembler-times {\tld1h\t} 10 } } */
      38  /* { dg-final { scan-assembler-times {\tst1h\t} 10 } } */
      39  /* { dg-final { scan-assembler-times {\tld1w\t} 15 } } */
      40  /* { dg-final { scan-assembler-times {\tst1w\t} 15 } } */
      41  /* { dg-final { scan-assembler-times {\tld1d\t} 15 } } */
      42  /* { dg-final { scan-assembler-times {\tst1d\t} 15 } } */
      43  /* { dg-final { scan-assembler-not {\tldr} } } */
      44  /* { dg-final { scan-assembler-not {\tstr} } } */
      45  
      46  /* We should use WHILEs for all accesses.  */
      47  /* { dg-final { scan-assembler-times {\twhilelo\tp[0-7]\.b} 20 } } */
      48  /* { dg-final { scan-assembler-times {\twhilelo\tp[0-7]\.h} 20 } } */
      49  /* { dg-final { scan-assembler-times {\twhilelo\tp[0-7]\.s} 30 } } */
      50  /* { dg-final { scan-assembler-times {\twhilelo\tp[0-7]\.d} 30 } } */
      51  
      52  /* 6 for the 8-bit types and 2 for the 16-bit types.  */
      53  /* { dg-final { scan-assembler-times {\tuqdecb\t} 8 } } */
      54  /* 4 for the 16-bit types and 3 for the 32-bit types.  */
      55  /* { dg-final { scan-assembler-times {\tuqdech\t} 7 } } */
      56  /* 6 for the 32-bit types and 3 for the 64-bit types.  */
      57  /* { dg-final { scan-assembler-times {\tuqdecw\t} 9 } } */
      58  /* { dg-final { scan-assembler-times {\tuqdecd\t} 6 } } */