> >> It looks like we need more slideups than before for certain cases? > >> Is > >> that only for LMUL>1? Why is that? Before we could just slideup one > >> element while we now have two zero-padded vectors that are being > >> combined? If so, can we exclude those cases (everything <= > >> REGMODE_NATURAL_SIZE)? > > > > No, that's not a regression. > > On trunk now (https://godbolt.org/z/bjE5b7qPP) there are spills; > > with this patch, it turns out to have no spills and 3 slideups. > > What I meant here was just a hint for future work, ideally it should > > only need 1 slideup and no spills, but that's not this patch's job. > > I meant for LMUL > 1 like https://godbolt.org/z/b31vf8Wf4 > To me it looked like we spill twice as much with the patch. Or did I > mix something up? >
Hi Robin, You didn't mix anything up, I just checked the asm, and it's on me. I only checked the LMUL=1 case on eyes before and forgot to scan for spills in the testcase, so when regtest showed all PASS, I just assumed all LMUL were gone..... My bad! I'll follow your suggestion to exclude LMUL>1. Patch v2 on flight. -- Regards, Zhongyao
