RE: [PATCH 12/21]middle-end: Add remaining changes to peeling and vectorizer to support early breaks

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

> -----Original Message-----
> From: Richard Biener <rguenther@suse.de>
> Sent: Wednesday, December 6, 2023 8:32 AM
> To: Tamar Christina <Tamar.Christina@arm.com>
> Cc: gcc-patches@gcc.gnu.org; nd <nd@arm.com>; jlaw@ventanamicro.com
> Subject: Re: [PATCH 12/21]middle-end: Add remaining changes to peeling and
> vectorizer to support early breaks
> 
> On Mon, 6 Nov 2023, Tamar Christina wrote:
> 
> > Hi All,
> >
> > This finishes wiring that didn't fit in any of the other patches.
> > Essentially just adding related changes so peeling for early break works.
> >
> > Bootstrapped Regtested on aarch64-none-linux-gnu and no issues.
> >
> > Ok for master?
> >
> > Thanks,
> > Tamar
> >
> > gcc/ChangeLog:
> >
> > 	* tree-vect-loop-manip.cc (vect_set_loop_condition_normal,
> > 	vect_do_peeling): Support early breaks.
> > 	* tree-vect-loop.cc (vect_need_peeling_or_partial_vectors_p): Likewise.
> > 	* tree-vectorizer.cc (pass_vectorize::execute): Check all exits.
> >
> > --- inline copy of patch --
> > diff --git a/gcc/tree-vect-loop-manip.cc b/gcc/tree-vect-loop-manip.cc
> > index
> eef2bb50c1505f5cf802d5d80300affc2cbe69f6..9c1405d79fd8fe8689007df3b7
> 605b7a3d3ecdd7 100644
> > --- a/gcc/tree-vect-loop-manip.cc
> > +++ b/gcc/tree-vect-loop-manip.cc
> > @@ -1187,7 +1187,7 @@ vect_set_loop_condition_partial_vectors_avx512
> (class loop *loop,
> >     loop handles exactly VF scalars per iteration.  */
> >
> >  static gcond *
> > -vect_set_loop_condition_normal (loop_vec_info /* loop_vinfo */, edge
> exit_edge,
> > +vect_set_loop_condition_normal (loop_vec_info loop_vinfo, edge exit_edge,
> >  				class loop *loop, tree niters, tree step,
> >  				tree final_iv, bool niters_maybe_zero,
> >  				gimple_stmt_iterator loop_cond_gsi)
> > @@ -1296,7 +1296,8 @@ vect_set_loop_condition_normal (loop_vec_info /*
> loop_vinfo */, edge exit_edge,
> >    gsi_insert_before (&loop_cond_gsi, cond_stmt, GSI_SAME_STMT);
> >
> >    /* Record the number of latch iterations.  */
> > -  if (limit == niters)
> > +  if (limit == niters
> > +      || LOOP_VINFO_EARLY_BREAKS (loop_vinfo))
> >      /* Case A: the loop iterates NITERS times.  Subtract one to get the
> >         latch count.  */
> >      loop->nb_iterations = fold_build2 (MINUS_EXPR, niters_type, niters,
> > @@ -3242,6 +3243,16 @@ vect_do_peeling (loop_vec_info loop_vinfo, tree
> niters, tree nitersm1,
> >      bound_epilog += vf - 1;
> >    if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo))
> >      bound_epilog += 1;
> > +
> > +  /* For early breaks the scalar loop needs to execute at most VF times
> > +     to find the element that caused the break.  */
> > +  if (LOOP_VINFO_EARLY_BREAKS (loop_vinfo))
> > +    {
> > +      bound_epilog = vf;
> > +      /* Force a scalar epilogue as we can't vectorize the index finding.  */
> > +      vect_epilogues = false;
> 
> This is originally initialized with
> 
>   bool vect_epilogues = loop_vinfo->epilogue_vinfos.length () > 0;
> 
> so I think we should avoid filling that with LOOP_VINFO_EARLY_BREAKS
> rather than fixing up after the fact?  That is in vect_analyze_loop
> adjust
> 
>   /* Only vectorize epilogues if PARAM_VECT_EPILOGUES_NOMASK is
>      enabled, SIMDUID is not set, it is the innermost loop and we have
>      either already found the loop's SIMDLEN or there was no SIMDLEN to
>      begin with.
>      TODO: Enable epilogue vectorization for loops with SIMDUID set.  */
>   bool vect_epilogues = (!simdlen
>                          && loop->inner == NULL
>                          && param_vect_epilogues_nomask
>                          && LOOP_VINFO_PEELING_FOR_NITER
> (first_loop_vinfo)
>                          && !loop->simduid);
> 
> and add !LOOP_VINFO_EARLY_BREAKS?
> 
> > +    }
> > +
> >    bool epilog_peeling = maybe_ne (bound_epilog, 0U);
> >    poly_uint64 bound_scalar = bound_epilog;
> >
> > @@ -3376,14 +3387,23 @@ vect_do_peeling (loop_vec_info loop_vinfo, tree
> niters, tree nitersm1,
> >  				  bound_prolog + bound_epilog)
> >  		      : (!LOOP_REQUIRES_VERSIONING (loop_vinfo)
> >  			 || vect_epilogues));
> > +
> > +  /* We only support early break vectorization on known bounds at this time.
> > +     This means that if the vector loop can't be entered then we won't generate
> > +     it at all.  So for now force skip_vector off because the additional control
> > +     flow messes with the BB exits and we've already analyzed them.  */
> > + skip_vector = skip_vector && !LOOP_VINFO_EARLY_BREAKS (loop_vinfo);
> > +
> 
>   bool skip_vector = (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
>                       ? maybe_lt (LOOP_VINFO_INT_NITERS (loop_vinfo),
>                                   bound_prolog + bound_epilog)
>                       : (!LOOP_REQUIRES_VERSIONING (loop_vinfo)
>                          || vect_epilogues));
> 
> to me that looks like
> 
>   gcc_assert (!skip_vector || !LOOP_VINFO_EARLY_BREAKS (loop_vinfo));
> 
> should work?  You are basically relying on cost modeling rejecting
> vectorization that doesn't enter the vector loop.
> 
> >    /* Epilog loop must be executed if the number of iterations for epilog
> >       loop is known at compile time, otherwise we need to add a check at
> >       the end of vector loop and skip to the end of epilog loop.  */
> >    bool skip_epilog = (prolog_peeling < 0
> >  		      || !LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
> >  		      || !vf.is_constant ());
> > -  /* PEELING_FOR_GAPS is special because epilog loop must be executed.  */
> > -  if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo))
> > +  /* PEELING_FOR_GAPS and peeling for early breaks are special because epilog
> > +     loop must be executed.  */
> > +  if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)
> > +      || LOOP_VINFO_EARLY_BREAKS (loop_vinfo))
> >      skip_epilog = false;
> >
> >    class loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo);
> > diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> > index
> 55d6aee3d29151e6b528f6fdde15c693e5bdd847..51a054c5b035ac80dfbbf3b5
> ba2f6da82fda91f6 100644
> > --- a/gcc/tree-vect-loop.cc
> > +++ b/gcc/tree-vect-loop.cc
> > @@ -1236,6 +1236,14 @@ vect_need_peeling_or_partial_vectors_p
> (loop_vec_info loop_vinfo)
> >      th = LOOP_VINFO_COST_MODEL_THRESHOLD
> (LOOP_VINFO_ORIG_LOOP_INFO
> >  					  (loop_vinfo));
> >
> > +  /* When we have multiple exits and VF is unknown, we must require partial
> > +     vectors because the loop bounds is not a minimum but a maximum.  That is
> to
> > +     say we cannot unpredicate the main loop unless we peel or use partial
> > +     vectors in the epilogue.  */
> > +  if (LOOP_VINFO_EARLY_BREAKS (loop_vinfo)
> > +      && !LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant ())
> > +    return true;
> > +
> 
> I don't quite understand the !.is_constant (), early breaks always
> require peeling?

It's mostly to force the use of an unpredicated main loop, since this
forces LOOP_VINFO_PEELING_FOR_NITER.  I can alternatively not
set it and then mask the individual comparisons in
vectorizable_early_break.  Both work equally,  I guess that may be
better since the assumption is that the break is not taken.

Any preference?

Thanks,
Tamar

> 
> >    if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
> >        && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) >= 0)
> >      {
> > @@ -3149,7 +3157,8 @@ start_over:
> >
> >    /* If an epilogue loop is required make sure we can create one.  */
> >    if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)
> > -      || LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo))
> > +      || LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo)
> > +      || LOOP_VINFO_EARLY_BREAKS (loop_vinfo))
> >      {
> >        if (dump_enabled_p ())
> >          dump_printf_loc (MSG_NOTE, vect_location, "epilog loop required\n");
> > diff --git a/gcc/tree-vectorizer.cc b/gcc/tree-vectorizer.cc
> > index
> d97e2b54c25ac60378935392aa7b73476efed74b..8b495fc7ca137109fc0097c7
> 7837233c660e032a 100644
> > --- a/gcc/tree-vectorizer.cc
> > +++ b/gcc/tree-vectorizer.cc
> > @@ -1381,7 +1381,9 @@ pass_vectorize::execute (function *fun)
> >  	 predicates that need to be shared for optimal predicate usage.
> >  	 However reassoc will re-order them and prevent CSE from working
> >  	 as it should.  CSE only the loop body, not the entry.  */
> > -      bitmap_set_bit (exit_bbs, single_exit (loop)->dest->index);
> > +      auto_vec<edge> exits = get_loop_exit_edges (loop);
> > +      for (edge exit : exits)
> > +	bitmap_set_bit (exit_bbs, exit->dest->index);
> >
> >        edge entry = EDGE_PRED (loop_preheader_edge (loop)->src, 0);
> >        do_rpo_vn (fun, entry, exit_bbs);
> 
> Otherwise OK.
> 
> Richard.