https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109721
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Huh, we indeed vectorize count_averages (). But I have no idea how the
offending rev should have changed this ...
Ah. Before we had
gcc.dg/tree-ssa/predcom-2.c:27:55: missed: not vectorized: relevant stmt not
supported: _9 = _8 & 65535;
but now we let this slip through with
gcc.dg/tree-ssa/predcom-2.c:26:17: note:
proceeding using word mode.
but we do not actually use word_mode. Later veclower decomposes the
bitwise vector and:
- vect__9.38_33 = vect__8.37_31 & { 65535, 65535 };
+ _26 = BIT_FIELD_REF <vect__8.37_31, 32, 0>;
+ _19 = _26 & 65535;
+ _23 = BIT_FIELD_REF <vect__8.37_31, 32, 32>;
+ _10 = _23 & 65535;
+ vect__9.38_33 = {_19, _10};
Btw, when writing the patch I noticed we let SImode AND through when
target_support_p even though it isn't word_mode and I didn't want to
change that but had to catch the case where SImode PLUS is supported
(well, obviously...) but emulated vectors rely on it being word_mode.
I'm going to amend this more.
But maybe this arch is also missing bit operation vector support for
some unknown reason.