On Thu, Sep 29, 2022 at 12:17 PM Richard Sandiford <richard.sandif...@arm.com> wrote: > > Andrew Stubbs <a...@codesourcery.com> writes: > > On 29/09/2022 10:24, Richard Sandiford wrote: > >> Otherwise: > >> > >> operand0[0] = operand1 < operand2; > >> for (i = 1; i < operand3; i++) > >> operand0[i] = operand0[i - 1] && (operand1 + i < operand2); > >> > >> looks like a "length and mask" operation, which IIUC is also what > >> RVV wanted? (Wasn't at the Cauldron, so not entirely sure.) > >> > >> Perhaps the difference is that in this case the length must be constant. > >> (Or is that true for RVV as well?) > > > > I too saw that presentation and I have compared notes with Juzhe before > > posting this. > > > > As he has posted, what they want is different because their config > > register has an explicit length field whereas GCN just uses a mask to > > limit the length (more like AArch64, I think). > > > > The RVV solution uses different logic in the gimple IR; this proposal is > > indistinguishable from the status quo at that point. > > Hmm, OK. (And thanks to Juzhe for the summary.) > > I can't think of any existing examples of optabs that have a variable > number of operands. But maybe this is a good reason to change that. > > Having to add what amounts to a vector type descriptor to make up for > the lack of mode information seems like a bit of a hack. But it's > possibly a hack that we'll need to do again (for other patterns), > if we keep representing multiple distinct vector/predicate types > using the same integer mode. I guess this establishes a way of > coping with the situation in general. > > So personally I'm OK with the patch, if Richi agrees.
It's not the nicest way of carrying the information but short of inventing new modes I can't see something better (well, another optab). I see the GCN backend expects a constant in operand 3 but the docs don't specify the operand has to be a CONST_INT, can you adjust them accordingly? Otherwise I'm fine with it. It might even prove useful for x86. Richard. > Richard
