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

[Jakub Jelinek <jakub@redhat.com>]

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.

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.

	Jakub