Hi Robin, Thanks for your comment. I think your point is correct, especially the part about SEWmin. I will revise this patch again.
On Wed, Feb 5, 2025 at 4:18 PM Robin Dapp <rdapp....@gmail.com> wrote: > > Hi Robin, > > Sorry, I should have simplified the problem by presenting it in terms of > > Zve32x, because Zve32f implies Zve32x. > > As the specification states, the requirement is to support LMUL ≥ > SEW/ELEN. > > Regarding the implementation, > > But the spec requirement mentions SEW_min not SEW? > > "In general, the requirement is to support LMUL ≥ SEWMIN/ELEN, where > SEWMIN is > the narrowest supported SEW value and ELEN is the widest supported SEW > value" > > Further it states: > "For standard vector extensions with ELEN=32, fractional LMULs of 1/2 and > 1/4 > must be supported." > > > I followed this rule to fix the problem. > > In this link: https://godbolt.org/z/j59oTW371, there is a vsetivli > > zero,2,e32,mf2,ta,ma. > > Here, SEW=32, and Zve32x has ELEN=32, which makes LMUL=1/2 illegal. > > > > According to the rule LMUL ≥ SEW/ELEN => LMUL ≥ 32 / 32 => LMUL ≥ 1. > > As you're specifying VLEN=128 (with zvl128b) we enable the respective > modes, > i.e. V2SI and V4SI. With VLEN=32 those wouldn't be available. > RVVMF2SI already has the requirement TARGET_MIN_VLEN > 32 so wouldn't be > chosen > either. > > If LMUL <= 1 were illegal for all zve32 we couldn't vectorize anything that > doesn't fit a full vector? That can't be correct and would severely limit > us. > > I think the only necessary change is to make sure we're not emitting > LMUL = 1/8 for ELEN=32/zve32. That should be a far less invasive change. > > >> In particular, how would the same LMUL for AVL=2 and AVL=4 and the same > data > >> type be correct? > > > > That's right. The case just allocates more space, but storing 2 and 4 > > elements remains the same. > > Even if a V2SI with LMUL=1 on VLEN=128 doesn't lead to a SIGILL right away > it would surely modify the overlap constraints and such. To me that > doesn't > look right. > > -- > Regards > Robin > >
