https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98211
--- Comment #8 from rsandifo at gcc dot gnu.org <rsandifo at gcc dot gnu.org> --- (In reply to Richard Biener from comment #7) > Hmm, OK, so besides the incomplete bool pattern matching the issue seems to > be that while we reject the problematic conversion in > vectorizable_conversion it slips through via vectorizable_assignment because > it does > > /* Conversion between boolean types of different sizes is > a simple assignment in case their vectypes are same > boolean vectors. */ > && (!VECTOR_BOOLEAN_TYPE_P (vectype) > || !VECTOR_BOOLEAN_TYPE_P (vectype_in))) > > as opposed to vectorizable_conversions > > if (VECTOR_BOOLEAN_TYPE_P (vectype_out) > && !VECTOR_BOOLEAN_TYPE_P (vectype_in)) > > that was added by g:2dab46d5fc9f95de16bd9bf0f219be5e64324d1f without a > testcase > or PR reference so it's difficult to tell what it was supposed to allow. Agree that looks odd… > Now, > for the case in question the conversion would have slipped though anyway > since > the only difference in the types is the sign and that one is BOOLEAN_TYPE > and the other INTEGER_TYPE. So the exception above seems to intent to allow > conversions with different precisions (note we now have precision 1 for all > vector bools). > > So I'm going to just copy the vectorizable_conversion condition into > vectorizable_assignment as well. Shouldn't it instead be: VECTOR_BOOLEAN_TYPE_P (vectype) != VECTOR_BOOLEAN_TYPE_P (vectype_in) as is used in some other places (sometimes with ^ instead of !=)? AFAIK using VCE would break in both directions.
