diff mbox

[GCC8,31/33] Set range information for niter bound of vectorized loop

Message ID CAHFci2-GopkQO8ouwqddrROiFGbMvNYXne=FatLsw+fDFcg1ww@mail.gmail.com
State New
Headers show

Commit Message

Bin.Cheng May 15, 2017, 3:58 p.m. UTC 
On Thu, May 11, 2017 at 12:02 PM, Richard Biener
<richard.guenther@gmail.com> wrote:
> On Tue, Apr 18, 2017 at 12:54 PM, Bin Cheng <Bin.Cheng@arm.com> wrote
>> Hi,
>> Based on vect_peeling algorithm, we know for sure that vectorized loop must iterates at least once.
>> This patch sets range information for niter bounds of vectorized loop.  This helps niter analysis,
>> so iv elimination too.
>> Is it OK?
>
>        niters_vector = force_gimple_operand (niters_vector, &stmts, true, var);
>        gsi_insert_seq_on_edge_immediate (pe, stmts);
> +      /* Peeling algorithm guarantees that vector loop bound is at least ONE,
> +        we set range information to make niters analyzer's life easier.  */
> +      if (TREE_CODE (niters_vector) == SSA_NAME)
> +       set_range_info (niters_vector, VR_RANGE, build_int_cst (type, 1),
> +                       fold_build2 (RSHIFT_EXPR, type,
> +                                    TYPE_MAX_VALUE (type), log_vf));
>
> if all of niters_vector folds to an original SSA name then
> niters_vector after gimplification
> is not a new SSA name and thus you can't set range-info on it.
>
> Likewise for the other case where LOOP_VINFO_NITERS is just an SSA name.
Hi,
This is updated patch.  It checks whether the result ssa name is newly
created tmp and only sets range information if so.

Is it OK?

Thanks,
bin

2017-04-11  Bin Cheng  <bin.cheng@arm.com>

    * tree-vectorizer.h (vect_build_loop_niters): New parameter.
    * tree-vect-loop-manip.c (vect_build_loop_niters): New parameter.
    Set true to new parameter if new ssa variable is defined.
    (vect_gen_vector_loop_niters): Refactor.  Set range information
    for the new vector loop bound variable.
    (vect_do_peeling): Ditto.

>
> Richard.
>
>> Thanks,
>> bin
>> 2017-04-11  Bin Cheng  <bin.cheng@arm.com>
>>
>>         * tree-vect-loop-manip.c (vect_gen_vector_loop_niters): Refactor.
>>         Set range information for vector loop bound variable.
>>         (vect_do_peeling): Ditto.

Comments

Richard Biener May 19, 2017, 12:51 p.m. UTC | #1 
On Mon, May 15, 2017 at 5:58 PM, Bin.Cheng <amker.cheng@gmail.com> wrote:
> On Thu, May 11, 2017 at 12:02 PM, Richard Biener
> <richard.guenther@gmail.com> wrote:
>> On Tue, Apr 18, 2017 at 12:54 PM, Bin Cheng <Bin.Cheng@arm.com> wrote
>>> Hi,
>>> Based on vect_peeling algorithm, we know for sure that vectorized loop must iterates at least once.
>>> This patch sets range information for niter bounds of vectorized loop.  This helps niter analysis,
>>> so iv elimination too.
>>> Is it OK?
>>
>>        niters_vector = force_gimple_operand (niters_vector, &stmts, true, var);
>>        gsi_insert_seq_on_edge_immediate (pe, stmts);
>> +      /* Peeling algorithm guarantees that vector loop bound is at least ONE,
>> +        we set range information to make niters analyzer's life easier.  */
>> +      if (TREE_CODE (niters_vector) == SSA_NAME)
>> +       set_range_info (niters_vector, VR_RANGE, build_int_cst (type, 1),
>> +                       fold_build2 (RSHIFT_EXPR, type,
>> +                                    TYPE_MAX_VALUE (type), log_vf));
>>
>> if all of niters_vector folds to an original SSA name then
>> niters_vector after gimplification
>> is not a new SSA name and thus you can't set range-info on it.
>>
>> Likewise for the other case where LOOP_VINFO_NITERS is just an SSA name.
> Hi,
> This is updated patch.  It checks whether the result ssa name is newly
> created tmp and only sets range information if so.
>
> Is it OK?

A better way to check whether the SSA name is new is to see if 'stmts'
filled by force_gimple_operand is non-empty.

Richard.

> Thanks,
> bin
>
> 2017-04-11  Bin Cheng  <bin.cheng@arm.com>
>
>     * tree-vectorizer.h (vect_build_loop_niters): New parameter.
>     * tree-vect-loop-manip.c (vect_build_loop_niters): New parameter.
>     Set true to new parameter if new ssa variable is defined.
>     (vect_gen_vector_loop_niters): Refactor.  Set range information
>     for the new vector loop bound variable.
>     (vect_do_peeling): Ditto.
>
>>
>> Richard.
>>
>>> Thanks,
>>> bin
>>> 2017-04-11  Bin Cheng  <bin.cheng@arm.com>
>>>
>>>         * tree-vect-loop-manip.c (vect_gen_vector_loop_niters): Refactor.
>>>         Set range information for vector loop bound variable.
>>>         (vect_do_peeling): Ditto.
diff mbox

Patch

From fbf6867a1f8d8fed94ce53861ccefb0d704e9f96 Mon Sep 17 00:00:00 2001
From: Bin Cheng <binche01@e108451-lin.cambridge.arm.com>
Date: Thu, 11 May 2017 17:26:57 +0100
Subject: [PATCH 5/9] range_info-for-vect_loop-niters-20170225.txt

---
 gcc/tree-vect-loop-manip.c | 52 +++++++++++++++++++++++++++++-----------------
 gcc/tree-vectorizer.h      |  2 +-
 2 files changed, 34 insertions(+), 20 deletions(-)

diff --git a/gcc/tree-vect-loop-manip.c b/gcc/tree-vect-loop-manip.c
index f48336b..52cd2bb 100644
--- a/gcc/tree-vect-loop-manip.c
+++ b/gcc/tree-vect-loop-manip.c
@@ -1095,10 +1095,11 @@  vect_update_inits_of_drs (loop_vec_info loop_vinfo, tree niters)
 
 
 /* This function builds ni_name = number of iterations.  Statements
-   are emitted on the loop preheader edge.  */
+   are emitted on the loop preheader edge.  If NEW_VAR_P is not NULL, set
+   it to TRUE if new ssa_var is generated.  */
 
 tree
-vect_build_loop_niters (loop_vec_info loop_vinfo)
+vect_build_loop_niters (loop_vec_info loop_vinfo, bool *new_var_p)
 {
   tree ni = unshare_expr (LOOP_VINFO_NITERS (loop_vinfo));
   if (TREE_CODE (ni) == INTEGER_CST)
@@ -1114,6 +1115,10 @@  vect_build_loop_niters (loop_vec_info loop_vinfo)
       if (stmts)
 	gsi_insert_seq_on_edge_immediate (pe, stmts);
 
+      if (new_var_p != NULL)
+	*new_var_p = (TREE_CODE (ni_name) == SSA_NAME
+		      && SSA_NAME_VAR (ni_name) == var);
+
       return ni_name;
     }
 }
@@ -1177,22 +1182,21 @@  vect_gen_vector_loop_niters (loop_vec_info loop_vinfo, tree niters,
 			     tree *niters_vector_ptr, bool niters_no_overflow)
 {
   tree ni_minus_gap, var;
-  tree niters_vector;
+  tree niters_vector, type = TREE_TYPE (niters);
   int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
   edge pe = loop_preheader_edge (LOOP_VINFO_LOOP (loop_vinfo));
-  tree log_vf = build_int_cst (TREE_TYPE (niters), exact_log2 (vf));
+  tree log_vf = build_int_cst (type, exact_log2 (vf));
 
   /* If epilogue loop is required because of data accesses with gaps, we
      subtract one iteration from the total number of iterations here for
      correct calculation of RATIO.  */
   if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo))
     {
-      ni_minus_gap = fold_build2 (MINUS_EXPR, TREE_TYPE (niters),
-				  niters,
-				  build_one_cst (TREE_TYPE (niters)));
+      ni_minus_gap = fold_build2 (MINUS_EXPR, type, niters,
+				  build_one_cst (type));
       if (!is_gimple_val (ni_minus_gap))
 	{
-	  var = create_tmp_var (TREE_TYPE (niters), "ni_gap");
+	  var = create_tmp_var (type, "ni_gap");
 	  gimple *stmts = NULL;
 	  ni_minus_gap = force_gimple_operand (ni_minus_gap, &stmts,
 					       true, var);
@@ -1208,25 +1212,29 @@  vect_gen_vector_loop_niters (loop_vec_info loop_vinfo, tree niters,
      (niters - vf) >> log2(vf) + 1 by using the fact that we know ratio
      will be at least one.  */
   if (niters_no_overflow)
-    niters_vector = fold_build2 (RSHIFT_EXPR, TREE_TYPE (niters),
-				 ni_minus_gap, log_vf);
+    niters_vector = fold_build2 (RSHIFT_EXPR, type, ni_minus_gap, log_vf);
   else
     niters_vector
-      = fold_build2 (PLUS_EXPR, TREE_TYPE (niters),
-		     fold_build2 (RSHIFT_EXPR, TREE_TYPE (niters),
-				  fold_build2 (MINUS_EXPR, TREE_TYPE (niters),
-					       ni_minus_gap,
-					       build_int_cst
-						 (TREE_TYPE (niters), vf)),
+      = fold_build2 (PLUS_EXPR, type,
+		     fold_build2 (RSHIFT_EXPR, type,
+				  fold_build2 (MINUS_EXPR, type, ni_minus_gap,
+					       build_int_cst (type, vf)),
 				  log_vf),
-		     build_int_cst (TREE_TYPE (niters), 1));
+		     build_int_cst (type, 1));
 
   if (!is_gimple_val (niters_vector))
     {
-      var = create_tmp_var (TREE_TYPE (niters), "bnd");
+      var = create_tmp_var (type, "bnd");
       gimple *stmts = NULL;
       niters_vector = force_gimple_operand (niters_vector, &stmts, true, var);
       gsi_insert_seq_on_edge_immediate (pe, stmts);
+      /* Peeling algorithm guarantees that vector loop bound is at least ONE,
+	 we set range information to make niters analyzer's life easier.  */
+      if (TREE_CODE (niters_vector) == SSA_NAME
+	  && SSA_NAME_VAR (niters_vector) == var)
+	set_range_info (niters_vector, VR_RANGE, build_int_cst (type, 1),
+			fold_build2 (RSHIFT_EXPR, type,
+				     TYPE_MAX_VALUE (type), log_vf));
     }
   *niters_vector_ptr = niters_vector;
 
@@ -1771,7 +1779,13 @@  vect_do_peeling (loop_vec_info loop_vinfo, tree niters, tree nitersm1,
       LOOP_VINFO_NITERSM1 (loop_vinfo)
 	= fold_build2 (MINUS_EXPR, type,
 		       LOOP_VINFO_NITERSM1 (loop_vinfo), niters_prolog);
-      niters = vect_build_loop_niters (loop_vinfo);
+      bool new_var_p = false;
+      niters = vect_build_loop_niters (loop_vinfo, &new_var_p);
+      /* It's guaranteed that vector loop bound before vectorization is at
+	 least VF, so set range information for newly generated var.  */
+      if (TREE_CODE (niters) == SSA_NAME && new_var_p)
+	set_range_info (niters, VR_RANGE,
+			build_int_cst (type, vf), TYPE_MAX_VALUE (type));
 
       /* Prolog iterates at most bound_prolog times, latch iterates at
 	 most bound_prolog - 1 times.  */
diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
index 12bb904..f72a1d5 100644
--- a/gcc/tree-vectorizer.h
+++ b/gcc/tree-vectorizer.h
@@ -1168,7 +1168,7 @@  extern gimple *vect_force_simple_reduction (loop_vec_info, gimple *, bool,
 					    bool *, bool);
 /* Drive for loop analysis stage.  */
 extern loop_vec_info vect_analyze_loop (struct loop *, loop_vec_info);
-extern tree vect_build_loop_niters (loop_vec_info);
+extern tree vect_build_loop_niters (loop_vec_info, bool * = NULL);
 extern void vect_gen_vector_loop_niters (loop_vec_info, tree, tree *, bool);
 /* Drive for loop transformation stage.  */
 extern struct loop *vect_transform_loop (loop_vec_info);
-- 
1.9.1