Richard Biener <richard.guent...@gmail.com> writes: > On Thu, Jul 8, 2021 at 2:44 PM Richard Sandiford via Gcc-patches > <gcc-patches@gcc.gnu.org> wrote: >> >> Vector reduction accumulators can differ in signedness from the >> final scalar result. The conversions to handle that case were >> distributed through vect_create_epilog_for_reduction; this patch >> does the conversion up-front instead. > > But is that still correct? The conversions should be unsigned -> signed, > that is, we've performed the reduction in unsigned because we associated > the originally undefined overflow signed reduction. But the final > reduction of the vector lanes in the epilogue still needs to be done > unsigned. > > So it's just not obvious that the patch preserves this - if it does then > the patch is OK.
We ended up covering most of this in the later 6/10 thread, but just to follow up here for the record, in case anyone looks at the list archives: In that scenario, the phis are created with the signed type and then (like you say) the reduction happens in the unsigned type. These conversions are from the signed type to the unsigned type ready for the reduction. All later code either performed the conversion itself or (in the case of some of the cond reductions) required the phi and reduction vectypes to be the same. I've pushed the series now -- thanks for the reviews. Richard
