Re: [PATCH] tree-optimization/113552 - fix num_call accounting in simd clone vectorization

[Richard Biener <rguenther@suse.de>]

On Tue, 23 Jan 2024, Jakub Jelinek wrote:

> On Tue, Jan 23, 2024 at 01:03:46PM +0100, Richard Biener wrote:
> > On Tue, 23 Jan 2024, Jakub Jelinek wrote:
> > 
> > > On Tue, Jan 23, 2024 at 12:56:52PM +0100, Richard Biener wrote:
> > > > The following avoids using exact_log2 on the number of SIMD clone calls
> > > > to be emitted when vectorizing calls since that can easily be not
> > > > a power of two in which case it will return -1.  For different simd
> > > > clones the number of calls will differ by a multiply with a power of two
> > > > only so using floor_log2 is good enough here.
> > > > 
> > > > Bootstrap and regtest running on x86_64-unknown-linux-gnu.
> > > > 
> > > > 	PR tree-optimization/113552
> > > > 	* tree-vect-stmts.cc (vectorizable_simd_clone_call): Use
> > > > 	floor_log2 instead of exact_log2 on the number of calls.
> > > 
> > > Is there any target which supports non-power-of-two simdlen?
> > > If not, perhaps we should add !pow2p_hwi (num_calls) to the continue;
> > > condition a few lines earlier?
> > 
> > Is non-power-of-two simdlen a thing?  Note there's nothing wrong
> > with non-power-of-two num_calls, with VF == 4 and a group size
> > of 3 you get 12 lanes and either 3 (simdlen == 4) or 6 (simdlen == 2)
> > calls.
> > 
> > Iff non-power-of-two simdlen is a thing then we could bias
> > by + num_calls (no idea why we use *_log2 in the first place, but it
> > was that way since the beginning).
> 
> Ah, with SLP I can understand it doesn't have to be a power of two,
> the original
> +       if (n->simdclone->simdlen
> +           < (unsigned) LOOP_VINFO_VECT_FACTOR (loop_vinfo))
> +         this_badness += (exact_log2 (LOOP_VINFO_VECT_FACTOR (loop_vinfo))
> +                          - exact_log2 (n->simdclone->simdlen)) * 1024;
> was written for loop vectorization only and I think correctly assumed
> power of 2 loop vectorization factors as well as simdlens.

Note even without SLP we could vectorize a group size of 3 with
interleaving.  But yes, VF for interleaving is a power of two and
so is simdlen so the above should have worked.

> I admit I don't remember why log2 rather than the count has been used,
> the first version of the patch is
> https://gcc.gnu.org/pipermail/gcc-patches/2013-November/374728.html
> But if we keep using log2, perhaps better ceil_log2 because num_calls of
> 3 is certainly more expensive than 2.

I think only the relative accounting to the other this_badness
increments matter since as said, at least with power-of-two simdlen
we'd get different {ceil,floor}_log2 values for different simdlen.

It's never going to be 3 vs 2 but 3 * 2^n vs. 3 * 2^m so floor or
ceil doesn't matter.  In fact we could have just using
some inverse of exact_log2 (n->simdclone->simdlen).  That is,
it's only simdlen that's varying in this addend to this_badness.

For the exact_log2 case the behavior is unchanged and we now only
get a sensible number for the others now.  Maybe log2 was for the
fear of overflow or over-accounting compared to the other increments.
When overflow is not an issue we could also use
floor_log2 (num_calls * 4096) to be more "precise".  I don't know
why we have all these different weights and why they are the way
they are - but it's a lot of apples and oranges we put together
in a magic number to compare ;)

I prefer to do a minimal change now, but both floor and ceil work
for me, so if you prefer I can change (but as said, it would only
matter with the mixing with the other cost factors, and in unknown
ways since the desire is not spelled out).

Richard.