gcc/testsuite/gcc.target/aarch64/sve/slp_12.c

   1 /* { dg-do compile } */
   2 /* { dg-options "-O2 -ftree-vectorize -msve-vector-bits=scalable" } */
   3
   4 #include <stdint.h>
   5
   6 #define N1 (19 * 2)
   7
   8 #define VEC_PERM(TYPE)                                          \
   9 void __attribute__ ((noinline, noclone))                        \
  10 vec_slp_##TYPE (TYPE *restrict a, TYPE *restrict b)             \
  11 {                                                               \
  12   for (int i = 0; i < N1; ++i)                                  \
  13     {                                                           \
  14       a[i] += 1;                                                \
  15       b[i * 4] += 2;                                            \
  16       b[i * 4 + 1] += 3;                                        \
  17       b[i * 4 + 2] += 4;                                        \
  18       b[i * 4 + 3] += 5;                                        \
  19     }                                                           \
  20 }
  21
  22 #define TEST_ALL(T)                             \
  23   T (int8_t)                                    \
  24   T (uint8_t)                                   \
  25   T (int16_t)                                   \
  26   T (uint16_t)                                  \
  27   T (int32_t)                                   \
  28   T (uint32_t)                                  \
  29   T (int64_t)                                   \
  30   T (uint64_t)                                  \
  31   T (float)                                     \
  32   T (double)
  33
  34 TEST_ALL (VEC_PERM)
  35
  36 /* The loop should be fully-masked.  */
  37 /* { dg-final { scan-assembler-times {\tld1b\t} 10 } } */
  38 /* { dg-final { scan-assembler-times {\tst1b\t} 10 } } */
  39 /* { dg-final { scan-assembler-times {\tld1h\t} 10 } } */
  40 /* { dg-final { scan-assembler-times {\tst1h\t} 10 } } */
  41 /* { dg-final { scan-assembler-times {\tld1w\t} 15 } } */
  42 /* { dg-final { scan-assembler-times {\tst1w\t} 15 } } */
  43 /* { dg-final { scan-assembler-times {\tld1d\t} 15 } } */
  44 /* { dg-final { scan-assembler-times {\tst1d\t} 15 } } */
  45 /* { dg-final { scan-assembler-not {\tldr} } } */
  46 /* { dg-final { scan-assembler-not {\tstr} } } */
  47
  48 /* We should use WHILEs for all accesses.  */
  49 /* { dg-final { scan-assembler-times {\twhilelo\tp[0-7]\.b} 20 } } */
  50 /* { dg-final { scan-assembler-times {\twhilelo\tp[0-7]\.h} 20 } } */
  51 /* { dg-final { scan-assembler-times {\twhilelo\tp[0-7]\.s} 30 } } */
  52 /* { dg-final { scan-assembler-times {\twhilelo\tp[0-7]\.d} 30 } } */
  53
  54 /* 6 for the 8-bit types and 2 for the 16-bit types.  */
  55 /* { dg-final { scan-assembler-times {\tuqdecb\t} 8 } } */
  56 /* 4 for the 16-bit types and 3 for the 32-bit types.  */
  57 /* { dg-final { scan-assembler-times {\tuqdech\t} 7 } } */
  58 /* 6 for the 32-bit types and 3 for the 64-bit types.  */
  59 /* { dg-final { scan-assembler-times {\tuqdecw\t} 9 } } */
  60 /* { dg-final { scan-assembler-times {\tuqdecd\t} 6 } } */