On Tue, 30 May 2023, Richard Sandiford wrote:
> My understanding was that we went into this knowing that the IVs
> would defeat SCEV analysis. Apparently that wasn't a problem for RVV,
> but it's not surprising that it is a problem in general.
>
> This isn't just about SELECT_VL though. We use the same type of IV
> for cases what aren't going to use SELECT_VL.
>
> Richard Biener <rguent...@suse.de> writes:
> > On Tue, 30 May 2023, Kewen.Lin wrote:
> >
> >> on 2023/5/30 17:26, juzhe.zh...@rivai.ai wrote:
> >> > Ok.
> >> >
> >> > It seems that for this conditions:
> >> >
> >> > + /* If we're vectorizing a loop that uses length "controls" and
> >> > + can iterate more than once, we apply decrementing IV approach
> >> > + in loop control. */
> >> > + if (LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)
> >> > + && !LOOP_VINFO_LENS (loop_vinfo).is_empty ()
> >> > + && LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo) == 0
> >> > + && !(LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
> >> > + && known_le (LOOP_VINFO_INT_NITERS (loop_vinfo),
> >> > + LOOP_VINFO_VECT_FACTOR (loop_vinfo))))
> >> > + LOOP_VINFO_USING_DECREMENTING_IV_P (loop_vinfo) = true;
> >> >
> >> >
> >> > I should add direct_supportted_p (SELECT_VL...) to this is that right?
> >>
> >> I guess no, with this condition any targets without SELECT_VL are unable
> >> to leverage the new decrement scheme for lengths, as your reply in PR109971
> >> you didn't meant to disable it. IIUC, what Richi suggested is to introduce
> >> one new IV just like the previous one which has non-variable step, then
> >> it's
> >> SCEV-ed and some analysis based on it can do a good job.
> >
> > No, I said the current scheme does sth along
> >
> > do {
> > remain -= MIN (vf, remain);
> > } while (remain != 0);
> >
> > and I suggest to instead do
> >
> > do {
> > old_remain = remain;
> > len = MIN (vf, remain);
> > remain -= vf;
> > } while (old_remain >= vf);
> >
> > basically since only the last iteration will have len < vf we can
> > ignore that remain -= vf will underflow there if we appropriately
> > rewrite the exit test to use the pre-decrement value.
>
> Yeah, agree that should work.
Btw, it's still on my TOOD list (unless somebody beats me...) to
rewrite the vectorizer code gen to do all loop control and conditions
on a decrementing "remaining scalar iters" IV.
> But how easy would it be to extend SCEV analysis, via a pattern match?
> The evolution of the IV phi wrt the inner loop is still a normal SCEV.
No, the IV isn't a normal SCEV, the final value is different.
I think pattern matching this in niter analysis could work though.
Richard.
