[Bug tree-optimization/92645] Hand written vector code is 450 times slower when compiled with GCC compared to Clang

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92645

--- Comment #20 from Richard Biener <rguenth at gcc dot gnu.org> ---
Small C testcase for one of the patterns we miss to optimize/vectorize:

void foo (char * __restrict src, short * __restrict dest)
{
  union {
      __int128_t i;
      char v[16];
  } u;
  __builtin_memcpy (&u.i, src, 16);
  dest[0] = u.v[0];
  dest[1] = u.v[1];
  dest[2] = u.v[2];
  dest[3] = u.v[3];
  dest[4] = u.v[4];
  dest[5] = u.v[5];
  dest[6] = u.v[6];
  dest[7] = u.v[7];
  dest[8] = u.v[8];
  dest[9] = u.v[9];
  dest[10] = u.v[10];
  dest[11] = u.v[11];
  dest[12] = u.v[12];
  dest[13] = u.v[13];
  dest[14] = u.v[14];
  dest[15] = u.v[15];
}

presents itself as

  _19 = MEM <__int128 unsigned> [(char * {ref-all})src_18(D)];
  _37 = (char) _19;
  _1 = (short int) _37;
  *dest_20(D) = _1;
  _38 = BIT_FIELD_REF <_19, 8, 8>;
  _2 = (short int) _38;
  MEM[(short int *)dest_20(D) + 2B] = _2;
  _39 = BIT_FIELD_REF <_19, 8, 16>;
  _3 = (short int) _39;
  MEM[(short int *)dest_20(D) + 4B] = _3;
...
  _16 = (short int) _52;
  MEM[(short int *)dest_20(D) + 30B] = _16;
  return;

where SLP vectorization is confused about (char) _19 vs. BIT_FIELD_REF
but also wouldn't handle BIT_FIELD_REFs.  It neither vectorizes the
store to a store from a CTOR which forwprop could then pattern-match.