RE: [PATCH 8/19]middle-end: updated niters analysis to handle multiple exits.

[Tamar Christina <Tamar.Christina@arm.com>]

> -----Original Message-----
> From: Richard Biener <rguenther@suse.de>
> Sent: Thursday, July 13, 2023 12:49 PM
> To: Tamar Christina <Tamar.Christina@arm.com>
> Cc: gcc-patches@gcc.gnu.org; nd <nd@arm.com>; jlaw@ventanamicro.com
> Subject: Re: [PATCH 8/19]middle-end: updated niters analysis to handle
> multiple exits.
> 
> On Wed, 28 Jun 2023, Tamar Christina wrote:
> 
> > Hi All,
> >
> > For early break vectorization we have to update niters analysis to
> > record and analyze all exits of the loop, and so all conds.
> >
> > The niters of the loop is still determined by the main/natural exit of
> > the loop as this is the O(n) bounds.  For now we don't do much with
> > the secondary conds, but their assumptions can be used to generate
> versioning checks later.
> >
> > Bootstrapped Regtested on aarch64-none-linux-gnu and no issues.
> >
> > Ok for master?
> 
> I probably confused vec_init_exit_info in the previous patch - that said, I'm
> missing a clear function that determines the natural exit of the original (if-
> converted) scalar loop.  As vec_init_exit_info seems to (re-)compute that I'll
> comment on it here.

Ah was wondering if you'd seen it 😊

> 
> +  /* The main IV is to be determined by the block that's the first
> reachable
> +     block from the latch.  We cannot rely on the order the loop analysis
> +     returns and we don't have any SCEV analysis on the loop.  */
> + auto_vec <edge> workset;  workset.safe_push (loop_latch_edge (loop));
> + hash_set <edge> visited;
> +
> +  while (!workset.is_empty ())
> +    {
> +      edge e = workset.pop ();
> +      if (visited.contains (e))
> +       continue;
> +
> +      bool found_p = false;
> +      for (edge ex : e->src->succs)
> +       {
> +         if (exits.contains (ex))
> +           {
> +             found_p = true;
> +             e = ex;
> +             break;
> +           }
> +       }
> +
> +      if (found_p)
> +       {
> +         loop->vec_loop_iv = e;
> +         for (edge ex : exits)
> +           if (e != ex)
> +             loop->vec_loop_alt_exits.safe_push (ex);
> +         return;
> +       }
> +      else
> +       {
> +         for (edge ex : e->src->preds)
> +           workset.safe_insert (0, ex);
> +       }
> +      visited.add (e);
> +    }
> 
> So this greedily follows edges from the latch and takes the first exit.  Why's
> that better than simply choosing the first?
> 
> I'd have done
> 
>  auto_vec<edge> exits = get_loop_exit_edges (loop);  for (e : exits)
>    {
>      if (vect_get_loop_niters (...))
>        {
>          if no assumptions use that edge, if assumptions continue
>          searching, maybe ther's an edge w/o assumptions
>        }
>    }
>  use (first) exit with assumptions
> 
> we probably want to know 'may_be_zero' as well and prefer an edge without
> that.  So eventually call number_of_iterations_exit_assumptions
> directly and look for the best niter_desc and pass that to vect_get_loop_niters
> (or re-do the work).
> 
> As said for "copying" the exit to the loop copies use the block mapping.
> 

The issue is with the scalar loops, where we have no SCEV data and also no
SSA mapping data (from what I can tell, the map was cleared in ifcvt itself).

So for this to work with SCEV, we'd have to start analyzing the loop coming out of
LOOP_VINFO_SCALAR_LOOP as well unless I'm missing something?

Regards,
Tamar