[Bug tree-optimization/111820] [13/14 Regression] Compiler time hog in the vectorizer with `-O3 -fno-tree-vrp`

["rguenther at suse dot de" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111820

--- Comment #8 from rguenther at suse dot de <rguenther at suse dot de> ---
On Tue, 17 Oct 2023, crazylht at gmail dot com wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111820
> 
> --- Comment #7 from Hongtao.liu <crazylht at gmail dot com> ---
> (In reply to rguenther@suse.de from comment #6)
> > On Mon, 16 Oct 2023, crazylht at gmail dot com wrote:
> > 
> > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111820
> > > 
> > > --- Comment #5 from Hongtao.liu <crazylht at gmail dot com> ---
> > > (In reply to Richard Biener from comment #3)
> > > >         for (unsigned i = 0; i != skipn - 1; i++)
> > > >           begin = wi::mul (begin, wi::to_wide (step_expr));
> > > > 
> > > > (gdb) p skipn
> > > > $5 = 4294967292
> > > > 
> > > > niters is 4294967292 in vect_update_ivs_after_vectorizer.  Maybe the loop
> > > > should terminate when begin is zero.  But I wonder why we pass in 'niters'
> > > Here, it want to calculate begin * pow (step_expr, skipn), yes we can just skip
> > > the loop when begin is 0.
> > 
> > I mean terminate it when the multiplication overflowed to zero.
> for pow (3, skipn), it will never overflowed to zero.
> To solve this problem once and for all, I'm leaning towards setting a threshold
> in vect_can_peel_nonlinear_iv_p for vect_step_op_mul,if step_expr is not
> exact_log2() and niter > TYPE_PRECISION (step_expr) we give up on doing
> vectorization.

Hm, yeah - that's probably best.

> > 
> > As for the MASK_ thing the skip is to be interpreted negative (we
> > should either not use a 'tree' here or make it have the correct type
> > maybe).  Can we even handle this here?  It would need to be
> > a division, no?
> > 
> > So I think we need to disable non-linear IV or masked peeling for
> > niter/aligment?  But I wonder how we run into this with plain -O3.
> I think we already disabled negative niters_skip in
> vect_can_peel_nonlinear_iv_p.
> 
> 416  /* Also doens't support peel for neg when niter is variable.
> 1417     ??? generate something like niter_expr & 1 ? init_expr : -init_expr? 
> */
> 1418  niters_skip = LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo);
> 1419  if ((niters_skip != NULL_TREE
> 1420       && TREE_CODE (niters_skip) != INTEGER_CST)

But we end up here with niters_skip being INTEGER_CST and ..

> 1421      || (!vect_use_loop_mask_for_alignment_p (loop_vinfo)

possibly vect_use_loop_mask_for_alignment_p.  Note
LOOP_VINFO_PEELING_FOR_ALIGNMENT < 0 simply means the amount of
peeling is unknown.

But I wonder how we run into this on x86 without enabling
loop masking ...

> 1422          && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) < 0))
> 1423    {
> 1424      if (dump_enabled_p ())
> 1425        dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
> 1426                         "Peeling for alignement is not supported"
> 1427                         " for nonlinear induction when niters_skip"
> 1428                         " is not constant.\n");
> 1429      return false;
> 1430    }