1  /* { dg-do compile } */
       2  /* { dg-options "-O2 -ftree-vectorize -msve-vector-bits=scalable -ffast-math -fno-vect-cost-model" } */
       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    TYPE x0 = b[0];						\
      11    TYPE x1 = b[1];						\
      12    TYPE x2 = b[2];						\
      13    for (int i = 0; i < n; ++i)					\
      14      {								\
      15        x0 += a[i * 3];						\
      16        x1 += a[i * 3 + 1];					\
      17        x2 += a[i * 3 + 2];					\
      18      }								\
      19    b[0] = x0;							\
      20    b[1] = x1;							\
      21    b[2] = x2;							\
      22  }
      23  
      24  #define TEST_ALL(T)				\
      25    T (int8_t)					\
      26    T (uint8_t)					\
      27    T (int16_t)					\
      28    T (uint16_t)					\
      29    T (int32_t)					\
      30    T (uint32_t)					\
      31    T (int64_t)					\
      32    T (uint64_t)					\
      33    T (_Float16)					\
      34    T (float)					\
      35    T (double)
      36  
      37  TEST_ALL (VEC_PERM)
      38  
      39  /* These loops can't use SLP.  */
      40  /* { dg-final { scan-assembler-not {\tld1b\t} } } */
      41  /* { dg-final { scan-assembler-not {\tld1h\t} } } */
      42  /* { dg-final { scan-assembler-not {\tld1w\t} } } */
      43  /* { dg-final { scan-assembler-not {\tld1d\t} } } */
      44  /* { dg-final { scan-assembler {\tld3b\t} } } */
      45  /* { dg-final { scan-assembler {\tld3h\t} } } */
      46  /* { dg-final { scan-assembler {\tld3w\t} } } */
      47  /* { dg-final { scan-assembler {\tld3d\t} } } */
      48  
      49  /* { dg-final { scan-assembler-not {\tuqdec} } } */