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.
