https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97770
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Component|target |tree-optimization
--- Comment #14 from Richard Biener <rguenth at gcc dot gnu.org> ---
So we vectorize to
_18 = .POPCOUNT (vect__5.7_22);
_17 = .POPCOUNT (vect__5.7_21);
vect__6.8_16 = VEC_PACK_TRUNC_EXPR <_18, _17>;
_6 = 0;
_7 = dest_13(D) + _2;
vect__8.9_10 = [vec_unpack_lo_expr] vect__6.8_16;
vect__8.9_9 = [vec_unpack_hi_expr] vect__6.8_16;
_8 = (long long int) _6;
which is exactly the issue that in the scalar code we have a 'int' producing
popcount with long long argument but the vector IFN produces a result of the
same width as the argument. So the vectorizer compensates for that
(VEC_PACK_TRUNC_EXPR) and then vectorizes the widening that's in the scalar
code (vec_unpack_{lo,hi}_expr). The fix for this and for the missing
byte and word variants is to add a pattern to tree-vect-patterns.c for this
case matching it to the .POPCOUNT internal function. That possibly applies
to other bitops, too, like parity, ctz, ffs, etc. There's quite some
_widen helpers in the pattern recog code so I'm not sure how complicated
it is to match
(long)popcountl(long)
and
(short)popcount((int)short)
Richard may have a good idea since he did the last "big" surgery there.