https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109964
--- Comment #4 from rsandifo at gcc dot gnu.org <rsandifo at gcc dot gnu.org> --- (In reply to Richard Biener from comment #3) > So the bug in the vectorizer is that it does > > t.ii:14:5: note: can narrow to signed:16 without loss of precision: _31 = > 1 >> _30; > t.ii:14:5: note: only the low 16 bits of _30 are significant > > while that's correct that's not the proper constraint to check for in > vect_recog_over_widening_pattern I think. That has code to deal with > overflow for plus/minus/mult but no defense against shifts > (that's also not a vect_truncatable_operation_p, so maybe it should simply > check that) Like you say, vect_truncatable_operation_p already checks for operations for which truncation is distributive. But the function is supposed to handle other cases too. E.g. I think the current code is correct for division. But yeah, right shifts are an awkward case, because the defined range of the second operand is much smaller than the range of the type, and yet the defined range depends on the range of the type. Sorry for not thinking about that at the time.
