[PATCH 7/21]middle-end: update IV update code to support early breaks and arbitrary exits
Tamar Christina
Tamar.Christina@arm.com
Thu Nov 16 11:08:27 GMT 2023
> -----Original Message-----
> From: Richard Biener <rguenther@suse.de>
> Sent: Thursday, November 16, 2023 10:40 AM
> To: Tamar Christina <Tamar.Christina@arm.com>
> Cc: gcc-patches@gcc.gnu.org; nd <nd@arm.com>; jlaw@ventanamicro.com
> Subject: RE: [PATCH 7/21]middle-end: update IV update code to support early
> breaks and arbitrary exits
>
> On Wed, 15 Nov 2023, Tamar Christina wrote:
>
> > > -----Original Message-----
> > > From: Richard Biener <rguenther@suse.de>
> > > Sent: Wednesday, November 15, 2023 1:23 PM
> > > To: Tamar Christina <Tamar.Christina@arm.com>
> > > Cc: gcc-patches@gcc.gnu.org; nd <nd@arm.com>;
> jlaw@ventanamicro.com
> > > Subject: RE: [PATCH 7/21]middle-end: update IV update code to
> > > support early breaks and arbitrary exits
> > >
> > > On Wed, 15 Nov 2023, Tamar Christina wrote:
> > >
> > > > > -----Original Message-----
> > > > > From: Richard Biener <rguenther@suse.de>
> > > > > Sent: Wednesday, November 15, 2023 1:01 PM
> > > > > To: Tamar Christina <Tamar.Christina@arm.com>
> > > > > Cc: gcc-patches@gcc.gnu.org; nd <nd@arm.com>;
> > > jlaw@ventanamicro.com
> > > > > Subject: RE: [PATCH 7/21]middle-end: update IV update code to
> > > > > support early breaks and arbitrary exits
> > > > >
> > > > > On Wed, 15 Nov 2023, Tamar Christina wrote:
> > > > >
> > > > > > Patch updated to latest trunk:
> > > > > >
> > > > > > Hi All,
> > > > > >
> > > > > > This changes the PHI node updates to support early breaks.
> > > > > > It has to support both the case where the loop's exit matches
> > > > > > the normal loop exit and one where the early exit is "inverted", i.e.
> > > > > > it's an early
> > > > > exit edge.
> > > > > >
> > > > > > In the latter case we must always restart the loop for VF iterations.
> > > > > > For an early exit the reason is obvious, but there are cases
> > > > > > where the "normal" exit is located before the early one. This
> > > > > > exit then does a check on ivtmp resulting in us leaving the
> > > > > > loop since it thinks we're
> > > done.
> > > > > >
> > > > > > In these case we may still have side-effects to perform so we
> > > > > > also go to the scalar loop.
> > > > > >
> > > > > > For the "normal" exit niters has already been adjusted for
> > > > > > peeling, for the early exits we must find out how many
> > > > > > iterations we actually did. So we have to recalculate the new position
> for each exit.
> > > > > >
> > > > > > Thanks,
> > > > > > Tamar
> > > > > >
> > > > > > gcc/ChangeLog:
> > > > > >
> > > > > > * tree-vect-loop-manip.cc (vect_set_loop_condition_normal):
> > > > > > Hide
> > > > > unused.
> > > > > > (vect_update_ivs_after_vectorizer): Support early break.
> > > > > > (vect_do_peeling): Use it.
> > > > > >
> > > > > > --- inline copy of patch ---
> > > > > >
> > > > > > diff --git a/gcc/tree-vect-loop-manip.cc
> > > > > > b/gcc/tree-vect-loop-manip.cc index
> > > > > >
> > > > >
> > >
> d3fa8699271c4d7f404d648a38a95beabeabc99a..e1d210ab4617c894dab3
> > > > > d2654cf1
> > > > > > c842baac58f5 100644
> > > > > > --- a/gcc/tree-vect-loop-manip.cc
> > > > > > +++ b/gcc/tree-vect-loop-manip.cc
> > > > > > @@ -1200,7 +1200,7 @@
> > > > > > 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)
> @@ -
> > > > > 1412,7 +1412,7 @@
> > > > > > vect_set_loop_condition (class loop *loop, edge loop_e,
> > > > > > loop_vec_info
> > > > > loop_vinfo
> > > > > > When this happens we need to flip the understanding of
> > > > > > main and
> > > other
> > > > > > exits by peeling and IV updates. */
> > > > > >
> > > > > > -bool inline
> > > > > > +bool
> > > > > > vect_is_loop_exit_latch_pred (edge loop_exit, class loop *loop) {
> > > > > > return single_pred (loop->latch) == loop_exit->src; @@
> > > > > > -2142,6
> > > > > > +2142,7 @@ vect_can_advance_ivs_p (loop_vec_info loop_vinfo)
> > > > > > Input:
> > > > > > - LOOP - a loop that is going to be vectorized. The last few
> iterations
> > > > > > of LOOP were peeled.
> > > > > > + - VF - The chosen vectorization factor for LOOP.
> > > > > > - NITERS - the number of iterations that LOOP executes (before it is
> > > > > > vectorized). i.e, the number of times the ivs should be
> bumped.
> > > > > > - UPDATE_E - a successor edge of LOOP->exit that is on
> > > > > > the
> > > > > > (only) path
> > > > >
> > > > > the comment on this is now a bit misleading, can you try to
> > > > > update it and/or move the comment bits to the docs on EARLY_EXIT?
> > > > >
> > > > > > @@ -2152,6 +2153,9 @@ vect_can_advance_ivs_p (loop_vec_info
> > > > > loop_vinfo)
> > > > > > The phi args associated with the edge UPDATE_E in the bb
> > > > > > UPDATE_E->dest are updated accordingly.
> > > > > >
> > > > > > + - restart_loop - Indicates whether the scalar loop needs
> > > > > > + to restart the
> > > > >
> > > > > params are ALL_CAPS
> > > > >
> > > > > > + iteration count where the vector loop began.
> > > > > > +
> > > > > > Assumption 1: Like the rest of the vectorizer, this function assumes
> > > > > > a single loop exit that has a single predecessor.
> > > > > >
> > > > > > @@ -2169,18 +2173,22 @@ vect_can_advance_ivs_p (loop_vec_info
> > > > > loop_vinfo)
> > > > > > */
> > > > > >
> > > > > > static void
> > > > > > -vect_update_ivs_after_vectorizer (loop_vec_info loop_vinfo,
> > > > > > - tree niters, edge update_e)
> > > > > > +vect_update_ivs_after_vectorizer (loop_vec_info loop_vinfo,
> > > > > > +poly_uint64 vf,
> > > > >
> > > > > LOOP_VINFO_VECT_FACTOR?
> > > > >
> > > > > > + tree niters, edge update_e, bool
> > > > > restart_loop)
> > > > >
> > > > > I think 'bool early_exit' is better here? I wonder if we have an "early"
> > > > > exit after the main exit we are probably sure there are no
> > > > > side-effects to re- execute and could avoid this restarting?
> > > >
> > > > Side effects yes, but the actual check may not have been performed yet.
> > > > If you remember
> > > >
> https://gist.github.com/Mistuke/66f14fe5c1be32b91ce149bd9b8bb35f
> > > > There in the clz loop through the "main" exit you still have to
> > > > see if that iteration did not contain the entry. This is because
> > > > the loop counter is incremented before you iterate.
> > > >
> > > > >
> > > > > > {
> > > > > > gphi_iterator gsi, gsi1;
> > > > > > class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
> > > > > > basic_block update_bb = update_e->dest;
> > > > > > -
> > > > > > - basic_block exit_bb = LOOP_VINFO_IV_EXIT
> > > > > > (loop_vinfo)->dest;
> > > > > > -
> > > > > > - /* Make sure there exists a single-predecessor exit bb: */
> > > > > > - gcc_assert (single_pred_p (exit_bb));
> > > > > > - gcc_assert (single_succ_edge (exit_bb) == update_e);
> > > > > > + bool inversed_iv
> > > > > > + = !vect_is_loop_exit_latch_pred (LOOP_VINFO_IV_EXIT
> (loop_vinfo),
> > > > > > + LOOP_VINFO_LOOP
> (loop_vinfo));
> > > > > > + bool needs_interm_block = LOOP_VINFO_EARLY_BREAKS
> (loop_vinfo)
> > > > > > + && flow_bb_inside_loop_p (loop,
> update_e->src);
> > > > > > + edge loop_e = LOOP_VINFO_IV_EXIT (loop_vinfo);
> > > > > > + gcond *cond = get_loop_exit_condition (loop_e);
> > > > > > + basic_block exit_bb = loop_e->dest;
> > > > > > + basic_block iv_block = NULL;
> > > > > > + gimple_stmt_iterator last_gsi = gsi_last_bb (exit_bb);
> > > > > >
> > > > > > for (gsi = gsi_start_phis (loop->header), gsi1 =
> > > > > > gsi_start_phis
> > > (update_bb);
> > > > > > !gsi_end_p (gsi) && !gsi_end_p (gsi1); @@ -2190,7
> > > > > > +2198,6 @@ vect_update_ivs_after_vectorizer (loop_vec_info
> loop_vinfo,
> > > > > > tree step_expr, off;
> > > > > > tree type;
> > > > > > tree var, ni, ni_name;
> > > > > > - gimple_stmt_iterator last_gsi;
> > > > > >
> > > > > > gphi *phi = gsi.phi ();
> > > > > > gphi *phi1 = gsi1.phi (); @@ -2222,11 +2229,52 @@
> > > > > > vect_update_ivs_after_vectorizer
> > > > > (loop_vec_info loop_vinfo,
> > > > > > enum vect_induction_op_type induction_type
> > > > > > = STMT_VINFO_LOOP_PHI_EVOLUTION_TYPE (phi_info);
> > > > > >
> > > > > > - if (induction_type == vect_step_op_add)
> > > > > > + tree iv_var = PHI_ARG_DEF_FROM_EDGE (phi,
> > > > > > + loop_latch_edge
> > > (loop));
> > > > > > + /* create_iv always places it on the LHS. Alternatively we can set a
> > > > > > + property during create_iv to identify it. */
> > > > > > + bool ivtemp = gimple_cond_lhs (cond) == iv_var;
> > > > > > + if (restart_loop && ivtemp)
> > > > > > {
> > > > > > + type = TREE_TYPE (gimple_phi_result (phi));
> > > > > > + ni = build_int_cst (type, vf);
> > > > > > + if (inversed_iv)
> > > > > > + ni = fold_build2 (MINUS_EXPR, type, ni,
> > > > > > + fold_convert (type, step_expr));
> > > > > > + }
> > > > > > + else if (induction_type == vect_step_op_add)
> > > > > > + {
> > > > > > +
> > > > > > tree stype = TREE_TYPE (step_expr);
> > > > > > - off = fold_build2 (MULT_EXPR, stype,
> > > > > > - fold_convert (stype, niters), step_expr);
> > > > > > +
> > > > > > + /* Early exits always use last iter value not niters. */
> > > > > > + if (restart_loop)
> > > > > > + {
> > > > > > + /* Live statements in the non-main exit shouldn't be
> adjusted. We
> > > > > > + normally didn't have this problem with a single exit as
> live
> > > > > > + values would be in the exit block. However when
> dealing with
> > > > > > + multiple exits all exits are redirected to the merge
> block
> > > > > > + and we restart the iteration. */
> > > > >
> > > > > Hmm, I fail to see how this works - we're either using the value
> > > > > to continue the induction or not, independent of STMT_VINFO_LIVE_P.
> > > >
> > > > That becomes clear in the patch to update live reductions.
> > > > Essentially any live Reductions inside an alternative exit will
> > > > reduce to the first element rather than the last and use that as
> > > > the seed for the
> > > scalar loop.
> > >
> > > Hum. Reductions are vectorized as N separate reductions. I don't
> > > think you can simply change the reduction between the lanes to "skip"
> > > part of the vector iteration. But you can use the value of the
> > > vector from before the vector iteration - the loop header PHI
> > > result, and fully reduce that to get at the proper value.
> >
> > That's what It's supposed to be doing though. The reason live
> > operations are skipped here is that if we don't we'll re-adjust the IV
> > even though the value will already be correct after vectorization.
> >
> > Remember that this code only gets so far for IV PHI nodes.
> >
> > The loop phi header result itself can be live, i.e. see testcases
> > vect-early-break_70.c to vect-early-break_75.c
> >
> > you have i_15 = PHI <i_14 (6), 1(2)>
> >
> > we use i_15 in the early exit. This should not be adjusted because
> > when it's vectorized the value at 0[lane 0] is already correct. This
> > is why for any PHI inside the early exits it uses the value 0[0] instead of
> N[lane_max].
> >
> > Perhaps I'm missing something here?
>
> OK, so I refreshed my mind of what vect_update_ivs_after_vectorizer does.
>
> I still do not understand the (complexity of the) patch. Basically the function
> computes the new value of the IV "from scratch" based on the number of
> scalar iterations of the vector loop, the 'niter'
> argument. I would have expected that for the early exits we either pass in a
> different 'niter' or alternatively a 'niter_adjustment'.
But for an early exit there's no static value for adjusted niter, since you don't know
which iteration you exited from. Unlike the normal exit when you know if you get
there you've done all possible iterations.
So you must compute the scalar iteration count on the exit itself.
>
> It seems your change handles different kinds of inductions differently.
> Specifically
>
> bool ivtemp = gimple_cond_lhs (cond) == iv_var;
> if (restart_loop && ivtemp)
> {
> type = TREE_TYPE (gimple_phi_result (phi));
> ni = build_int_cst (type, vf);
> if (inversed_iv)
> ni = fold_build2 (MINUS_EXPR, type, ni,
> fold_convert (type, step_expr));
> }
>
> it looks like for the exit test IV we use either 'VF' or 'VF - step'
> as the new value. That seems to be very odd special casing for unknown
> reasons. And while you adjust vec_step_op_add, you don't adjust
> vect_peel_nonlinear_iv_init (maybe not supported - better assert here).
The VF case is for a normal "non-inverted" loop, where if you take an early exit
you know that you have to do at most VF iterations. The VF - step is to account
for the inverted loop control flow where you exit after adjusting the IV already
by + step.
Peeling doesn't matter here, since you know you were able to do a vector iteration
so it's safe to do VF iterations. So having peeled doesn't affect the remaining
iters count.
>
> Also the vec_step_op_add case will keep the original scalar IV live even when it
> is a vectorized induction. The code recomputing the value from scratch avoids
> this.
>
> /* For non-main exit create an intermediat edge to get any updated iv
> calculations. */
> if (needs_interm_block
> && !iv_block
> && (!gimple_seq_empty_p (stmts) || !gimple_seq_empty_p
> (new_stmts)))
> {
> iv_block = split_edge (update_e);
> update_e = single_succ_edge (update_e->dest);
> last_gsi = gsi_last_bb (iv_block);
> }
>
> this is also odd, can we adjust the API instead? I suppose this is because your
> computation uses the original loop IV, if you based the computation off the
> initial value only this might not be necessary?
No, on the main exit the code updates the value in the loop header and puts the
Calculation in the merge block. This works because it only needs to consume PHI
nodes in the merge block and things like niters are adjusted in the guard block.
For an early exit, we don't have a guard block, only the merge block. We have to
update the PHI nodes in that block, but can't do so since you can't produce a value
and consume it in a PHI node in the same BB. So we need to create the block to put
the values in for use in the merge block. Because there's no "guard" block for early
exits.
The API can be adjusted by always creating the empty block either during peeling.
That would prevent us from having to do anything special here. Would that work
better? Or I can do it in the loop that iterates over the exits to before the call
to vect_update_ivs_after_vectorizer, which I think might be more consistent.
>
> That said, I wonder why we cannot simply pass in an adjusted niter which
> would be niters_vector_mult_vf - vf and be done with that?
>
We can ofcourse not have this and recompute it from niters itself, however this does
affect the epilog code layout. Particularly knowing the static number if iterations left
causes it to usually unroll the loop and share some of the computations. i.e. the scalar
code is often more efficient.
The computation would be niters_vector_mult_vf - iters_done * vf, since the value put
Here is the remaining iteration count. It's static for early exits.
But can do whatever you prefer here. Let me know what you prefer for the above.
Thanks,
Tamar
> Thanks,
> Richard.
>
>
> > Regards,
> > Tamar
> > >
> > > > It has to do this since you have to perform the side effects for
> > > > the non-matching elements still.
> > > >
> > > > Regards,
> > > > Tamar
> > > >
> > > > >
> > > > > > + if (STMT_VINFO_LIVE_P (phi_info))
> > > > > > + continue;
> > > > > > +
> > > > > > + /* For early break the final loop IV is:
> > > > > > + init + (final - init) * vf which takes into account peeling
> > > > > > + values and non-single steps. The main exit can use
> niters
> > > > > > + since if you exit from the main exit you've done all
> vector
> > > > > > + iterations. For an early exit we don't know when we
> exit
> > > > > > +so
> > > > > we
> > > > > > + must re-calculate this on the exit. */
> > > > > > + tree start_expr = gimple_phi_result (phi);
> > > > > > + off = fold_build2 (MINUS_EXPR, stype,
> > > > > > + fold_convert (stype, start_expr),
> > > > > > + fold_convert (stype, init_expr));
> > > > > > + /* Now adjust for VF to get the final iteration value. */
> > > > > > + off = fold_build2 (MULT_EXPR, stype, off,
> > > > > > + build_int_cst (stype, vf));
> > > > > > + }
> > > > > > + else
> > > > > > + off = fold_build2 (MULT_EXPR, stype,
> > > > > > + fold_convert (stype, niters), step_expr);
> > > > > > +
> > > > > > if (POINTER_TYPE_P (type))
> > > > > > ni = fold_build_pointer_plus (init_expr, off);
> > > > > > else
> > > > > > @@ -2238,6 +2286,8 @@ vect_update_ivs_after_vectorizer
> > > > > > (loop_vec_info
> > > > > loop_vinfo,
> > > > > > /* Don't bother call vect_peel_nonlinear_iv_init. */
> > > > > > else if (induction_type == vect_step_op_neg)
> > > > > > ni = init_expr;
> > > > > > + else if (restart_loop)
> > > > > > + continue;
> > > > >
> > > > > This looks all a bit complicated - why wouldn't we simply always
> > > > > use the PHI result when 'restart_loop'? Isn't that the correct
> > > > > old start value in
> > > all cases?
> > > > >
> > > > > > else
> > > > > > ni = vect_peel_nonlinear_iv_init (&stmts, init_expr,
> > > > > > niters, step_expr,
> > > > > > @@ -2245,9 +2295,20 @@ vect_update_ivs_after_vectorizer
> > > > > (loop_vec_info
> > > > > > loop_vinfo,
> > > > > >
> > > > > > var = create_tmp_var (type, "tmp");
> > > > > >
> > > > > > - last_gsi = gsi_last_bb (exit_bb);
> > > > > > gimple_seq new_stmts = NULL;
> > > > > > ni_name = force_gimple_operand (ni, &new_stmts, false,
> > > > > > var);
> > > > > > +
> > > > > > + /* For non-main exit create an intermediat edge to get any
> updated iv
> > > > > > + calculations. */
> > > > > > + if (needs_interm_block
> > > > > > + && !iv_block
> > > > > > + && (!gimple_seq_empty_p (stmts) || !gimple_seq_empty_p
> > > > > (new_stmts)))
> > > > > > + {
> > > > > > + iv_block = split_edge (update_e);
> > > > > > + update_e = single_succ_edge (update_e->dest);
> > > > > > + last_gsi = gsi_last_bb (iv_block);
> > > > > > + }
> > > > > > +
> > > > > > /* Exit_bb shouldn't be empty. */
> > > > > > if (!gsi_end_p (last_gsi))
> > > > > > {
> > > > > > @@ -3342,8 +3403,26 @@ vect_do_peeling (loop_vec_info
> > > > > > loop_vinfo, tree
> > > > > niters, tree nitersm1,
> > > > > > niters_vector_mult_vf steps. */
> > > > > > gcc_checking_assert (vect_can_advance_ivs_p (loop_vinfo));
> > > > > > update_e = skip_vector ? e : loop_preheader_edge (epilog);
> > > > > > - vect_update_ivs_after_vectorizer (loop_vinfo,
> niters_vector_mult_vf,
> > > > > > - update_e);
> > > > > > + if (LOOP_VINFO_EARLY_BREAKS (loop_vinfo))
> > > > > > + update_e = single_succ_edge (e->dest);
> > > > > > + bool inversed_iv
> > > > > > + = !vect_is_loop_exit_latch_pred (LOOP_VINFO_IV_EXIT
> (loop_vinfo),
> > > > > > + LOOP_VINFO_LOOP
> (loop_vinfo));
> > > > >
> > > > > You are computing this here and in vect_update_ivs_after_vectorizer?
> > > > >
> > > > > > +
> > > > > > + /* Update the main exit first. */
> > > > > > + vect_update_ivs_after_vectorizer (loop_vinfo, vf,
> > > niters_vector_mult_vf,
> > > > > > + update_e, inversed_iv);
> > > > > > +
> > > > > > + /* And then update the early exits. */
> > > > > > + for (auto exit : get_loop_exit_edges (loop))
> > > > > > + {
> > > > > > + if (exit == LOOP_VINFO_IV_EXIT (loop_vinfo))
> > > > > > + continue;
> > > > > > +
> > > > > > + vect_update_ivs_after_vectorizer (loop_vinfo, vf,
> > > > > > + niters_vector_mult_vf,
> > > > > > + exit, true);
> > > > >
> > > > > ... why does the same not work here? Wouldn't the proper
> > > > > condition be !dominated_by_p (CDI_DOMINATORS, exit->src,
> > > > > LOOP_VINFO_IV_EXIT
> > > > > (loop_vinfo)->src) or similar? That is, whether the exit is at
> > > > > or after the main IV exit? (consider having two)
> > > > >
> > > > > > + }
> > > > > >
> > > > > > if (skip_epilog)
> > > > > > {
> > > > > >
> > > >
> > >
> > > --
> > > Richard Biener <rguenther@suse.de>
> > > SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461
> > > Nuernberg, Germany;
> > > GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG
> > > Nuernberg)
> >
>
> --
> Richard Biener <rguenther@suse.de>
> SUSE Software Solutions Germany GmbH,
> Frankenstrasse 146, 90461 Nuernberg, Germany;
> GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG
> Nuernberg)
More information about the Gcc-patches
mailing list