public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [patch][graphite] Remove sese_adjust_liveout_phis
@ 2010-06-12  8:17 Sebastian Pop
  2010-06-12 12:03 ` Sebastian Pop
  2010-06-14 18:20 ` Sebastian Pop
  0 siblings, 2 replies; 5+ messages in thread
From: Sebastian Pop @ 2010-06-12  8:17 UTC (permalink / raw)
  To: GCC Patches, gcc-graphite

[-- Attachment #1: Type: text/plain, Size: 717 bytes --]

Hi,

with the attached patches we no longer need to use the rename_map to
substitute the arguments of the phi nodes created after the code
generated by Graphite.  I committed these patches to the graphite
branch, and I will commit them to trunk once they pass the usual
graphite branch tests.

The aim is to replace the rename_map with the map that CLooG exposes
for each user statement: associating to a loop level an expression.
The expand_* functions will be replaced by the code generation from a
chrec that is obtained from the original scev by chrec_applying the
loop level -> expression map on each varying loop.  This will then
allow us to remove the call to the IV canonicalization from Graphite.

Sebastian

[-- Attachment #2: 0001-Launch-dotty-in-background.patch --]
[-- Type: text/x-diff, Size: 2068 bytes --]

From e747edcf6bd61fe4eae201b70181cba5ce43edbf Mon Sep 17 00:00:00 2001
From: Sebastian Pop <sebpop@gmail.com>
Date: Fri, 11 Jun 2010 18:42:21 -0500
Subject: [PATCH 1/4] Launch dotty in background.

2010-06-12  Sebastian Pop  <sebastian.pop@amd.com>

	* graphite-dependences.c (dot_deps): Make system call to dotty run
	in background.
	(dot_deps_stmt): Same.
	* graphite-poly.c (dot_lst): Same.
---
 gcc/ChangeLog.graphite     |    7 +++++++
 gcc/graphite-dependences.c |    4 ++--
 gcc/graphite-poly.c        |    2 +-
 3 files changed, 10 insertions(+), 3 deletions(-)

diff --git a/gcc/ChangeLog.graphite b/gcc/ChangeLog.graphite
index 6df816b..72bb7f5 100644
--- a/gcc/ChangeLog.graphite
+++ b/gcc/ChangeLog.graphite
@@ -1,3 +1,10 @@
+2010-06-12  Sebastian Pop  <sebastian.pop@amd.com>
+
+	* graphite-dependences.c (dot_deps): Make system call to dotty run
+	in background.
+	(dot_deps_stmt): Same.
+	* graphite-poly.c (dot_lst): Same.
+
 2010-06-11  Sebastian Pop  <sebastian.pop@amd.com>
 
 	* graphite-clast-to-gimple.c (gcc_type_for_interval): Do not pass
diff --git a/gcc/graphite-dependences.c b/gcc/graphite-dependences.c
index 41824e4..91743f1 100644
--- a/gcc/graphite-dependences.c
+++ b/gcc/graphite-dependences.c
@@ -963,7 +963,7 @@ dot_deps (scop_p scop)
   dot_deps_1 (stream, scop);
   fclose (stream);
 
-  system ("dotty /tmp/scopdeps.dot");
+  system ("dotty /tmp/scopdeps.dot &");
 #else
   dot_deps_1 (stderr, scop);
 #endif
@@ -983,7 +983,7 @@ dot_deps_stmt (scop_p scop)
   dot_deps_stmt_1 (stream, scop);
   fclose (stream);
 
-  system ("dotty /tmp/scopdeps.dot");
+  system ("dotty /tmp/scopdeps.dot &");
 #else
   dot_deps_stmt_1 (stderr, scop);
 #endif
diff --git a/gcc/graphite-poly.c b/gcc/graphite-poly.c
index 24a47f0..28430ed 100644
--- a/gcc/graphite-poly.c
+++ b/gcc/graphite-poly.c
@@ -1320,7 +1320,7 @@ dot_lst (lst_p lst)
   fputs ("}\n\n", stream);
   fclose (stream);
 
-  system ("dotty /tmp/lst.dot");
+  system ("dotty /tmp/lst.dot &");
 #else
   fputs ("digraph all {\n", stderr);
   dot_lst_1 (stderr, lst);
-- 
1.7.0.4


[-- Attachment #3: 0002-Detect-commutative-reductions-in-all-the-scops-befor.patch --]
[-- Type: text/x-diff, Size: 5719 bytes --]

From 9c6db20feae3d4cb96966f0877c19d96ff4c5c66 Mon Sep 17 00:00:00 2001
From: Sebastian Pop <sebpop@gmail.com>
Date: Fri, 11 Jun 2010 19:37:07 -0500
Subject: [PATCH 2/4] Detect commutative reductions in all the scops before rewriting out of SSA all the other scalar dependences.

2010-06-12  Sebastian Pop  <sebastian.pop@amd.com>

	* graphite-sese-to-poly.c (build_scop_bbs): Not static anymore.
	(rewrite_reductions_out_of_ssa): Same.
	(rewrite_commutative_reductions_out_of_ssa): Same.
	(build_poly_scop): Do not call these functions.
	* graphite-sese-to-poly.h (build_poly_scop): Declared.
	(rewrite_reductions_out_of_ssa): Declared.
	(rewrite_commutative_reductions_out_of_ssa): Declared.
	* graphite.c (graphite_transform_loops): Call on every scop
	rewrite_commutative_reductions_out_of_ssa before calling
	rewrite_reductions_out_of_ssa and build_scop_bbs.
---
 gcc/ChangeLog.graphite      |   13 +++++++++++++
 gcc/graphite-sese-to-poly.c |   16 +++++-----------
 gcc/graphite-sese-to-poly.h |    3 +++
 gcc/graphite.c              |   14 ++++++++++++++
 4 files changed, 35 insertions(+), 11 deletions(-)

diff --git a/gcc/ChangeLog.graphite b/gcc/ChangeLog.graphite
index 72bb7f5..a87c69d 100644
--- a/gcc/ChangeLog.graphite
+++ b/gcc/ChangeLog.graphite
@@ -1,5 +1,18 @@
 2010-06-12  Sebastian Pop  <sebastian.pop@amd.com>
 
+	* graphite-sese-to-poly.c (build_scop_bbs): Not static anymore.
+	(rewrite_reductions_out_of_ssa): Same.
+	(rewrite_commutative_reductions_out_of_ssa): Same.
+	(build_poly_scop): Do not call these functions.
+	* graphite-sese-to-poly.h (build_poly_scop): Declared.
+	(rewrite_reductions_out_of_ssa): Declared.
+	(rewrite_commutative_reductions_out_of_ssa): Declared.
+	* graphite.c (graphite_transform_loops): Call on every scop
+	rewrite_commutative_reductions_out_of_ssa before calling
+	rewrite_reductions_out_of_ssa and build_scop_bbs.
+
+2010-06-12  Sebastian Pop  <sebastian.pop@amd.com>
+
 	* graphite-dependences.c (dot_deps): Make system call to dotty run
 	in background.
 	(dot_deps_stmt): Same.
diff --git a/gcc/graphite-sese-to-poly.c b/gcc/graphite-sese-to-poly.c
index 41e2c9f..66c0282 100644
--- a/gcc/graphite-sese-to-poly.c
+++ b/gcc/graphite-sese-to-poly.c
@@ -462,7 +462,7 @@ build_scop_bbs_1 (scop_p scop, sbitmap visited, basic_block bb, sbitmap reductio
 
 /* Gather the basic blocks belonging to the SCOP.  */
 
-static void
+void
 build_scop_bbs (scop_p scop, sbitmap reductions)
 {
   sbitmap visited = sbitmap_alloc (last_basic_block);
@@ -2398,7 +2398,7 @@ rewrite_cross_bb_scalar_deps (sese region, gimple_stmt_iterator *gsi)
 
 /* Rewrite out of SSA all the reduction phi nodes of SCOP.  */
 
-static void
+void
 rewrite_reductions_out_of_ssa (scop_p scop)
 {
   basic_block bb;
@@ -2577,7 +2577,7 @@ detect_commutative_reduction_arg (tree lhs, gimple stmt, tree arg,
 }
 
 /* Detect commutative and associative scalar reductions starting at
-   the STMT.  Return the phi node of the reduction cycle, or NULL.  */
+   STMT.  Return the phi node of the reduction cycle, or NULL.  */
 
 static gimple
 detect_commutative_reduction_assign (gimple stmt, VEC (gimple, heap) **in,
@@ -2665,7 +2665,7 @@ initial_value_for_loop_phi (gimple phi)
 }
 
 /* Detect commutative and associative scalar reductions starting at
-   the loop closed phi node CLOSE_PHI.  Return the phi node of the
+   the loop closed phi node STMT.  Return the phi node of the
    reduction cycle, or NULL.  */
 
 static gimple
@@ -2873,7 +2873,7 @@ rewrite_commutative_reductions_out_of_ssa_loop (loop_p loop,
 
 /* Rewrites all the commutative reductions from SCOP out of SSA.  */
 
-static void
+void
 rewrite_commutative_reductions_out_of_ssa (sese region, sbitmap reductions)
 {
   loop_iterator li;
@@ -2972,14 +2972,8 @@ void
 build_poly_scop (scop_p scop)
 {
   sese region = SCOP_REGION (scop);
-  sbitmap reductions = sbitmap_alloc (last_basic_block * 2);
   graphite_dim_t max_dim;
 
-  sbitmap_zero (reductions);
-  rewrite_commutative_reductions_out_of_ssa (region, reductions);
-  rewrite_reductions_out_of_ssa (scop);
-  build_scop_bbs (scop, reductions);
-  sbitmap_free (reductions);
 
   /* FIXME: This restriction is needed to avoid a problem in CLooG.
      Once CLooG is fixed, remove this guard.  Anyways, it makes no
diff --git a/gcc/graphite-sese-to-poly.h b/gcc/graphite-sese-to-poly.h
index ccf36dd..2e0cc08 100644
--- a/gcc/graphite-sese-to-poly.h
+++ b/gcc/graphite-sese-to-poly.h
@@ -30,5 +30,8 @@ struct base_alias_pair
 
 void build_poly_scop (scop_p);
 void check_poly_representation (scop_p);
+void rewrite_commutative_reductions_out_of_ssa (sese, sbitmap);
+void rewrite_reductions_out_of_ssa (scop_p);
+void build_scop_bbs (scop_p, sbitmap);
 
 #endif
diff --git a/gcc/graphite.c b/gcc/graphite.c
index 347ebc8..5f28aeb 100644
--- a/gcc/graphite.c
+++ b/gcc/graphite.c
@@ -254,6 +254,7 @@ graphite_transform_loops (void)
   bool need_cfg_cleanup_p = false;
   VEC (scop_p, heap) *scops = NULL;
   htab_t bb_pbb_mapping;
+  sbitmap reductions;
 
   if (!graphite_initialize ())
     return;
@@ -267,6 +268,19 @@ graphite_transform_loops (void)
     }
 
   bb_pbb_mapping = htab_create (10, bb_pbb_map_hash, eq_bb_pbb_map, free);
+  reductions = sbitmap_alloc (last_basic_block * 2);
+  sbitmap_zero (reductions);
+
+  for (i = 0; VEC_iterate (scop_p, scops, i, scop); i++)
+    rewrite_commutative_reductions_out_of_ssa (SCOP_REGION (scop), reductions);
+
+  for (i = 0; VEC_iterate (scop_p, scops, i, scop); i++)
+    {
+      rewrite_reductions_out_of_ssa (scop);
+      build_scop_bbs (scop, reductions);
+    }
+
+  sbitmap_free (reductions);
 
   for (i = 0; VEC_iterate (scop_p, scops, i, scop); i++)
     if (dbg_cnt (graphite_scop))
-- 
1.7.0.4


[-- Attachment #4: 0003-Remove-insert_copyout-and-insert_copyin.patch --]
[-- Type: text/x-diff, Size: 6148 bytes --]

From d36ba67360a736e061133c52e84aeaef7b1347eb Mon Sep 17 00:00:00 2001
From: Sebastian Pop <sebpop@gmail.com>
Date: Fri, 11 Jun 2010 19:43:13 -0500
Subject: [PATCH 3/4] Remove insert_copyout and insert_copyin.

2010-06-12  Sebastian Pop  <sebastian.pop@amd.com>

	* graphite-sese-to-poly.c (insert_out_of_ssa_copy): Pass an extra
	argument for the place after which to insert the out of SSA copy.
	(rewrite_close_phi_out_of_ssa): Update calls to insert_out_of_ssa_copy.
	(rewrite_phi_out_of_ssa): Same.
	(rewrite_cross_bb_scalar_deps): Same.
	(insert_copyout): Removed.
	(insert_copyin): Removed.
	(translate_scalar_reduction_to_array): Call insert_out_of_ssa_copy and
	insert_out_of_ssa_copy_on_edge instead of insert_copyout and
	insert_copyin.
---
 gcc/ChangeLog.graphite      |   13 ++++++++++
 gcc/graphite-sese-to-poly.c |   53 ++++++++++++------------------------------
 2 files changed, 28 insertions(+), 38 deletions(-)

diff --git a/gcc/ChangeLog.graphite b/gcc/ChangeLog.graphite
index a87c69d..888a77d 100644
--- a/gcc/ChangeLog.graphite
+++ b/gcc/ChangeLog.graphite
@@ -1,5 +1,18 @@
 2010-06-12  Sebastian Pop  <sebastian.pop@amd.com>
 
+	* graphite-sese-to-poly.c (insert_out_of_ssa_copy): Pass an extra
+	argument for the place after which to insert the out of SSA copy.
+	(rewrite_close_phi_out_of_ssa): Update calls to insert_out_of_ssa_copy.
+	(rewrite_phi_out_of_ssa): Same.
+	(rewrite_cross_bb_scalar_deps): Same.
+	(insert_copyout): Removed.
+	(insert_copyin): Removed.
+	(translate_scalar_reduction_to_array): Call insert_out_of_ssa_copy and
+	insert_out_of_ssa_copy_on_edge instead of insert_copyout and
+	insert_copyin.
+
+2010-06-12  Sebastian Pop  <sebastian.pop@amd.com>
+
 	* graphite-sese-to-poly.c (build_scop_bbs): Not static anymore.
 	(rewrite_reductions_out_of_ssa): Same.
 	(rewrite_commutative_reductions_out_of_ssa): Same.
diff --git a/gcc/graphite-sese-to-poly.c b/gcc/graphite-sese-to-poly.c
index 66c0282..7bb7bb8 100644
--- a/gcc/graphite-sese-to-poly.c
+++ b/gcc/graphite-sese-to-poly.c
@@ -2116,10 +2116,10 @@ gsi_for_phi_node (gimple stmt)
   return psi;
 }
 
-/* Insert the assignment "RES := VAR" just after the definition of VAR.  */
+/* Insert the assignment "RES := VAR" just after AFTER_STMT.  */
 
 static void
-insert_out_of_ssa_copy (tree res, tree var)
+insert_out_of_ssa_copy (tree res, tree var, gimple after_stmt)
 {
   gimple stmt;
   gimple_seq stmts;
@@ -2133,15 +2133,14 @@ insert_out_of_ssa_copy (tree res, tree var)
   si = gsi_last (stmts);
   gsi_insert_after (&si, stmt, GSI_NEW_STMT);
 
-  stmt = SSA_NAME_DEF_STMT (var);
-  if (gimple_code (stmt) == GIMPLE_PHI)
+  if (gimple_code (after_stmt) == GIMPLE_PHI)
     {
-      gsi = gsi_after_labels (gimple_bb (stmt));
+      gsi = gsi_after_labels (gimple_bb (after_stmt));
       gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
     }
   else
     {
-      gsi = gsi_for_stmt (stmt);
+      gsi = gsi_for_stmt (after_stmt);
       gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT);
     }
 }
@@ -2217,7 +2216,7 @@ rewrite_close_phi_out_of_ssa (gimple_stmt_iterator *psi)
 
   if (TREE_CODE (arg) == SSA_NAME
       && !SSA_NAME_IS_DEFAULT_DEF (arg))
-    insert_out_of_ssa_copy (zero_dim_array, arg);
+    insert_out_of_ssa_copy (zero_dim_array, arg, SSA_NAME_DEF_STMT (arg));
   else
     insert_out_of_ssa_copy_on_edge (single_pred_edge (gimple_bb (phi)),
 				    zero_dim_array, arg);
@@ -2238,7 +2237,7 @@ rewrite_phi_out_of_ssa (gimple_stmt_iterator *psi)
   basic_block bb = gimple_bb (phi);
   tree res = gimple_phi_result (phi);
   tree var = SSA_NAME_VAR (res);
-  tree zero_dim_array = create_zero_dim_array (var, "General_Reduction");
+  tree zero_dim_array = create_zero_dim_array (var, "phi_out_of_ssa");
   gimple_stmt_iterator gsi;
   gimple stmt;
   gimple_seq stmts;
@@ -2294,7 +2293,7 @@ rewrite_phi_out_of_ssa (gimple_stmt_iterator *psi)
 				     gimple_bb (SSA_NAME_DEF_STMT (arg)))
 	      || flow_bb_inside_loop_p (loop_outer (bb->loop_father),
 					gimple_bb (SSA_NAME_DEF_STMT (arg)))))
-	insert_out_of_ssa_copy (zero_dim_array, arg);
+	insert_out_of_ssa_copy (zero_dim_array, arg, SSA_NAME_DEF_STMT (arg));
       else
 	insert_out_of_ssa_copy_on_edge (gimple_phi_arg_edge (phi, i),
 					zero_dim_array, arg);
@@ -2388,7 +2387,8 @@ rewrite_cross_bb_scalar_deps (sese region, gimple_stmt_iterator *gsi)
 	  {
 	    zero_dim_array = create_zero_dim_array
 	      (SSA_NAME_VAR (def), "Cross_BB_scalar_dependence");
-	    insert_out_of_ssa_copy (zero_dim_array, def);
+	    insert_out_of_ssa_copy (zero_dim_array, def,
+				    SSA_NAME_DEF_STMT (def));
 	    gsi_next (gsi);
 	  }
 
@@ -2727,32 +2727,6 @@ translate_scalar_reduction_to_array_for_stmt (tree red, gimple stmt,
   gsi_insert_after (&insert_gsi, assign, GSI_SAME_STMT);
 }
 
-/* Insert the assignment "result (CLOSE_PHI) = RED".  */
-
-static void
-insert_copyout (tree red, gimple close_phi)
-{
-  tree res = gimple_phi_result (close_phi);
-  basic_block bb = gimple_bb (close_phi);
-  gimple_stmt_iterator insert_gsi = gsi_after_labels (bb);
-  gimple assign = gimple_build_assign (res, red);
-
-  gsi_insert_before (&insert_gsi, assign, GSI_SAME_STMT);
-}
-
-/* Insert the assignment "RED = initial_value (LOOP_PHI)".  */
-
-static void
-insert_copyin (tree red, gimple loop_phi)
-{
-  gimple_seq stmts;
-  tree init = initial_value_for_loop_phi (loop_phi);
-  tree expr = build2 (MODIFY_EXPR, TREE_TYPE (init), red, init);
-
-  force_gimple_operand (expr, &stmts, true, NULL);
-  gsi_insert_seq_on_edge (edge_initial_value_for_loop_phi (loop_phi), stmts);
-}
-
 /* Removes the PHI node and resets all the debug stmts that are using
    the PHI_RESULT.  */
 
@@ -2828,8 +2802,11 @@ translate_scalar_reduction_to_array (VEC (gimple, heap) *in,
 
       if (i == VEC_length (gimple, in) - 1)
 	{
-	  insert_copyout (red, close_phi);
-	  insert_copyin (red, loop_phi);
+	  insert_out_of_ssa_copy (gimple_phi_result (close_phi), red,
+				  close_phi);
+	  insert_out_of_ssa_copy_on_edge
+	    (edge_initial_value_for_loop_phi (loop_phi),
+	     red, initial_value_for_loop_phi (loop_phi));
 	}
 
       remove_phi (loop_phi);
-- 
1.7.0.4


[-- Attachment #5: 0004-Also-rewrite-out-of-SSA-scalar-dependences-going-out.patch --]
[-- Type: text/x-diff, Size: 8911 bytes --]

From e1e77ccc827024af17e5afc9baf5d8f044815052 Mon Sep 17 00:00:00 2001
From: Sebastian Pop <sebpop@gmail.com>
Date: Fri, 11 Jun 2010 19:58:39 -0500
Subject: [PATCH 4/4] Also rewrite out of SSA scalar dependences going outside the SCoP region.

2010-06-12  Sebastian Pop  <sebastian.pop@amd.com>

	* graphite-clast-to-gimple.c (gloog): Remove call to
	sese_adjust_liveout_phis.
	* graphite-sese-to-poly.c (scev_analyzable_p): When scev returns an
	SSA_NAME, allow it to be handled by rewrite_cross_bb_scalar_deps.
	(rewrite_cross_bb_scalar_deps): Handle GIMPLE_PHI nodes: call
	rewrite_phi_out_of_ssa.
	* sese.c (get_vdef_before_sese): Removed.
	(sese_adjust_vphi): Removed.
	(sese_adjust_liveout_phis): Removed.
	* sese.h (sese_adjust_liveout_phis): Removed.
---
 gcc/ChangeLog.graphite         |   13 +++
 gcc/graphite-clast-to-gimple.c |    4 -
 gcc/graphite-sese-to-poly.c    |   13 ++-
 gcc/sese.c                     |  159 ----------------------------------------
 gcc/sese.h                     |    1 -
 5 files changed, 22 insertions(+), 168 deletions(-)

diff --git a/gcc/ChangeLog.graphite b/gcc/ChangeLog.graphite
index 888a77d..d703048 100644
--- a/gcc/ChangeLog.graphite
+++ b/gcc/ChangeLog.graphite
@@ -1,5 +1,18 @@
 2010-06-12  Sebastian Pop  <sebastian.pop@amd.com>
 
+	* graphite-clast-to-gimple.c (gloog): Remove call to
+	sese_adjust_liveout_phis.
+	* graphite-sese-to-poly.c (scev_analyzable_p): When scev returns an
+	SSA_NAME, allow it to be handled by rewrite_cross_bb_scalar_deps.
+	(rewrite_cross_bb_scalar_deps): Handle GIMPLE_PHI nodes: call
+	rewrite_phi_out_of_ssa.
+	* sese.c (get_vdef_before_sese): Removed.
+	(sese_adjust_vphi): Removed.
+	(sese_adjust_liveout_phis): Removed.
+	* sese.h (sese_adjust_liveout_phis): Removed.
+
+2010-06-12  Sebastian Pop  <sebastian.pop@amd.com>
+
 	* graphite-sese-to-poly.c (insert_out_of_ssa_copy): Pass an extra
 	argument for the place after which to insert the out of SSA copy.
 	(rewrite_close_phi_out_of_ssa): Update calls to insert_out_of_ssa_copy.
diff --git a/gcc/graphite-clast-to-gimple.c b/gcc/graphite-clast-to-gimple.c
index a363f5d..01adba0 100644
--- a/gcc/graphite-clast-to-gimple.c
+++ b/gcc/graphite-clast-to-gimple.c
@@ -1578,10 +1578,6 @@ gloog (scop_p scop, VEC (scop_p, heap) *scops, htab_t bb_pbb_mapping)
 		   rename_map, &newivs, newivs_index,
 		   bb_pbb_mapping, 1, params_index);
   graphite_verify ();
-  sese_adjust_liveout_phis (region, rename_map,
-			    if_region->region->exit->src,
-			    if_region->false_region->exit,
-			    if_region->true_region->exit);
   scev_reset_htab ();
   rename_nb_iterations (rename_map);
 
diff --git a/gcc/graphite-sese-to-poly.c b/gcc/graphite-sese-to-poly.c
index 7bb7bb8..57967e5 100644
--- a/gcc/graphite-sese-to-poly.c
+++ b/gcc/graphite-sese-to-poly.c
@@ -2325,7 +2325,8 @@ scev_analyzable_p (tree def, sese region)
   loop_p loop = loop_containing_stmt (stmt);
   tree scev = scalar_evolution_in_region (region, loop, def);
 
-  return !chrec_contains_undetermined (scev);
+  return !chrec_contains_undetermined (scev)
+    && TREE_CODE (scev) != SSA_NAME;
 }
 
 /* Rewrite the scalar dependence of DEF used in USE_STMT with a memory
@@ -2379,9 +2380,13 @@ rewrite_cross_bb_scalar_deps (sese region, gimple_stmt_iterator *gsi)
   def_bb = gimple_bb (stmt);
 
   FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
-    if (def_bb != gimple_bb (use_stmt)
-	&& gimple_code (use_stmt) != GIMPLE_PHI
-	&& !is_gimple_debug (use_stmt))
+    if (gimple_code (use_stmt) == GIMPLE_PHI)
+      {
+	gimple_stmt_iterator si = gsi_for_stmt (use_stmt);
+	rewrite_phi_out_of_ssa (&si);
+      }
+    else if (def_bb != gimple_bb (use_stmt)
+	     && !is_gimple_debug (use_stmt))
       {
 	if (!zero_dim_array)
 	  {
diff --git a/gcc/sese.c b/gcc/sese.c
index 7e9bd91..051599e 100644
--- a/gcc/sese.c
+++ b/gcc/sese.c
@@ -394,100 +394,6 @@ sese_insert_phis_for_liveouts (sese region, basic_block bb,
   update_ssa (TODO_update_ssa);
 }
 
-/* Get the definition of NAME before the SESE.  Keep track of the
-   basic blocks that have been VISITED in a bitmap.  */
-
-static tree
-get_vdef_before_sese (sese region, tree name, sbitmap visited)
-{
-  unsigned i;
-  gimple stmt = SSA_NAME_DEF_STMT (name);
-  basic_block def_bb = gimple_bb (stmt);
-
-  if (!def_bb || !bb_in_sese_p (def_bb, region))
-    return name;
-
-  if (TEST_BIT (visited, def_bb->index))
-    return NULL_TREE;
-
-  SET_BIT (visited, def_bb->index);
-
-  switch (gimple_code (stmt))
-    {
-    case GIMPLE_PHI:
-      for (i = 0; i < gimple_phi_num_args (stmt); i++)
-	{
-	  tree arg = gimple_phi_arg_def (stmt, i);
-	  tree res;
-
-	  if (gimple_bb (SSA_NAME_DEF_STMT (arg))
-	      && def_bb->index == gimple_bb (SSA_NAME_DEF_STMT (arg))->index)
-	    continue;
-
-	  res = get_vdef_before_sese (region, arg, visited);
-	  if (res)
-	    return res;
-	}
-      return NULL_TREE;
-
-    case GIMPLE_ASSIGN:
-    case GIMPLE_CALL:
-      {
-	use_operand_p use_p = gimple_vuse_op (stmt);
-	tree use = USE_FROM_PTR (use_p);
-
-	if (def_bb->index == gimple_bb (SSA_NAME_DEF_STMT (use))->index)
-	  RESET_BIT (visited, def_bb->index);
-
-	return get_vdef_before_sese (region, use, visited);
-      }
-
-    default:
-      return NULL_TREE;
-    }
-}
-
-/* Adjust a virtual phi node PHI that is placed at the end of the
-   generated code for SCOP:
-
-   | if (1)
-   |   generated code from REGION;
-   | else
-   |   REGION;
-
-   The FALSE_E edge comes from the original code, TRUE_E edge comes
-   from the code generated for the SCOP.  */
-
-static void
-sese_adjust_vphi (sese region, gimple phi, edge true_e)
-{
-  unsigned i;
-
-  gcc_assert (gimple_phi_num_args (phi) == 2);
-
-  for (i = 0; i < gimple_phi_num_args (phi); i++)
-    if (gimple_phi_arg_edge (phi, i) == true_e)
-      {
-	tree true_arg, false_arg, before_scop_arg;
-	sbitmap visited;
-
-	true_arg = gimple_phi_arg_def (phi, i);
-	if (!SSA_NAME_IS_DEFAULT_DEF (true_arg))
-	  return;
-
-	false_arg = gimple_phi_arg_def (phi, i == 0 ? 1 : 0);
-	if (SSA_NAME_IS_DEFAULT_DEF (false_arg))
-	  return;
-
-	visited = sbitmap_alloc (last_basic_block);
-	sbitmap_zero (visited);
-	before_scop_arg = get_vdef_before_sese (region, false_arg, visited);
-	gcc_assert (before_scop_arg != NULL_TREE);
-	SET_PHI_ARG_DEF (phi, i, before_scop_arg);
-	sbitmap_free (visited);
-      }
-}
-
 /* Returns the expression associated to OLD_NAME in MAP.  */
 
 static tree
@@ -585,71 +491,6 @@ rename_sese_parameters (htab_t rename_map, sese region)
 		 rename_variables_in_expr (rename_map, p));
 }
 
-/* Adjusts the phi nodes in the block BB for variables defined in
-   SCOP_REGION and used outside the SCOP_REGION.  The code generation
-   moves SCOP_REGION in the else clause of an "if (1)" and generates
-   code in the then clause:
-
-   | if (1)
-   |   generated code from REGION;
-   | else
-   |   REGION;
-
-   To adjust the phi nodes after the condition, the RENAME_MAP is
-   used.  */
-
-void
-sese_adjust_liveout_phis (sese region, htab_t rename_map, basic_block bb,
-			  edge false_e, edge true_e)
-{
-  gimple_stmt_iterator si;
-
-  for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
-    {
-      unsigned i;
-      unsigned false_i = 0;
-      gimple phi = gsi_stmt (si);
-      tree res = gimple_phi_result (phi);
-
-      if (!is_gimple_reg (res))
-	{
-	  sese_adjust_vphi (region, phi, true_e);
-	  continue;
-	}
-
-      for (i = 0; i < gimple_phi_num_args (phi); i++)
-	if (gimple_phi_arg_edge (phi, i) == false_e)
-	  {
-	    false_i = i;
-	    break;
-	  }
-
-      for (i = 0; i < gimple_phi_num_args (phi); i++)
-	if (gimple_phi_arg_edge (phi, i) == true_e)
-	  {
-	    tree old_name = gimple_phi_arg_def (phi, false_i);
-	    tree expr = get_rename (rename_map, old_name);
-	    gimple_seq stmts;
-
-	    gcc_assert (old_name != expr);
-
-	    if (TREE_CODE (expr) != SSA_NAME
-		&& is_gimple_reg (old_name))
-	      {
-		tree type = TREE_TYPE (old_name);
-		tree var = create_tmp_var (type, "var");
-
-		expr = build2 (MODIFY_EXPR, type, var, expr);
-		expr = force_gimple_operand (expr, &stmts, true, NULL);
-		gsi_insert_seq_on_edge_immediate (true_e, stmts);
-	      }
-
-	    SET_PHI_ARG_DEF (phi, i, expr);
-	    set_rename (rename_map, old_name, res);
-	  }
-    }
-}
-
 /* Rename the SSA_NAMEs used in STMT and that appear in MAP.  */
 
 static void
diff --git a/gcc/sese.h b/gcc/sese.h
index 2b05ca8..4b2e5c2 100644
--- a/gcc/sese.h
+++ b/gcc/sese.h
@@ -55,7 +55,6 @@ typedef struct sese_s
 extern sese new_sese (edge, edge);
 extern void free_sese (sese);
 extern void sese_insert_phis_for_liveouts (sese, basic_block, edge, edge);
-extern void sese_adjust_liveout_phis (sese, htab_t, basic_block, edge, edge);
 extern void build_sese_loop_nests (sese);
 extern edge copy_bb_and_scalar_dependences (basic_block, sese, edge, htab_t);
 extern struct loop *outermost_loop_in_sese (sese, basic_block);
-- 
1.7.0.4


^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: [patch][graphite] Remove sese_adjust_liveout_phis
  2010-06-12  8:17 [patch][graphite] Remove sese_adjust_liveout_phis Sebastian Pop
@ 2010-06-12 12:03 ` Sebastian Pop
  2010-06-14 18:20 ` Sebastian Pop
  1 sibling, 0 replies; 5+ messages in thread
From: Sebastian Pop @ 2010-06-12 12:03 UTC (permalink / raw)
  To: GCC Patches, gcc-graphite

[-- Attachment #1: Type: text/plain, Size: 934 bytes --]

On Sat, Jun 12, 2010 at 02:53, Sebastian Pop <sebpop@gmail.com> wrote:
> Hi,
>
> with the attached patches we no longer need to use the rename_map to
> substitute the arguments of the phi nodes created after the code
> generated by Graphite.  I committed these patches to the graphite
> branch, and I will commit them to trunk once they pass the usual
> graphite branch tests.
>
> The aim is to replace the rename_map with the map that CLooG exposes
> for each user statement: associating to a loop level an expression.
> The expand_* functions will be replaced by the code generation from a
> chrec that is obtained from the original scev by chrec_applying the
> loop level -> expression map on each varying loop.  This will then
> allow us to remove the call to the IV canonicalization from Graphite.
>

The attached patch removes two uses of rename_map from gloog.
Committed to the graphite branch.

Sebastian

[-- Attachment #2: 0001-Remove-rename_nb_iterations-and-rename_sese_paramete.patch --]
[-- Type: text/x-diff, Size: 6278 bytes --]

From be9a6b386261452607f5440a46bbfb10700bf3d3 Mon Sep 17 00:00:00 2001
From: Sebastian Pop <sebpop@gmail.com>
Date: Sat, 12 Jun 2010 03:12:04 -0500
Subject: [PATCH] Remove rename_nb_iterations and rename_sese_parameters

2010-06-12  Sebastian Pop  <sebastian.pop@amd.com>

	* graphite-clast-to-gimple.c (gloog): Do not pass scops in parameter.
	Remove calls to rename_nb_iterations and rename_sese_parameters.
	* graphite-clast-to-gimple.h (gloog): Update declaration.
	* graphite.c (graphite_transform_loops): Update call to gloog.
	* sese.c (rename_variables_in_expr): Removed.
	(rename_nb_iterations): Removed.
	(rename_sese_parameters): Removed.
	* sese.h (rename_nb_iterations): Removed.
	(rename_sese_parameters): Removed.
---
 gcc/ChangeLog.graphite         |   12 ++++++++
 gcc/graphite-clast-to-gimple.c |    8 +-----
 gcc/graphite-clast-to-gimple.h |    2 +-
 gcc/graphite.c                 |    2 +-
 gcc/sese.c                     |   56 ----------------------------------------
 gcc/sese.h                     |    2 -
 6 files changed, 15 insertions(+), 67 deletions(-)

diff --git a/gcc/ChangeLog.graphite b/gcc/ChangeLog.graphite
index d703048..b4e2700 100644
--- a/gcc/ChangeLog.graphite
+++ b/gcc/ChangeLog.graphite
@@ -1,5 +1,17 @@
 2010-06-12  Sebastian Pop  <sebastian.pop@amd.com>
 
+	* graphite-clast-to-gimple.c (gloog): Do not pass scops in parameter.
+	Remove calls to rename_nb_iterations and rename_sese_parameters.
+	* graphite-clast-to-gimple.h (gloog): Update declaration.
+	* graphite.c (graphite_transform_loops): Update call to gloog.
+	* sese.c (rename_variables_in_expr): Removed.
+	(rename_nb_iterations): Removed.
+	(rename_sese_parameters): Removed.
+	* sese.h (rename_nb_iterations): Removed.
+	(rename_sese_parameters): Removed.
+
+2010-06-12  Sebastian Pop  <sebastian.pop@amd.com>
+
 	* graphite-clast-to-gimple.c (gloog): Remove call to
 	sese_adjust_liveout_phis.
 	* graphite-sese-to-poly.c (scev_analyzable_p): When scev returns an
diff --git a/gcc/graphite-clast-to-gimple.c b/gcc/graphite-clast-to-gimple.c
index 01adba0..54801ce 100644
--- a/gcc/graphite-clast-to-gimple.c
+++ b/gcc/graphite-clast-to-gimple.c
@@ -1531,7 +1531,7 @@ create_params_index (htab_t index_table, CloogProgram *prog) {
 */
 
 bool
-gloog (scop_p scop, VEC (scop_p, heap) *scops, htab_t bb_pbb_mapping)
+gloog (scop_p scop, htab_t bb_pbb_mapping)
 {
   VEC (tree, heap) *newivs = VEC_alloc (tree, heap, 10);
   loop_p context_loop;
@@ -1539,7 +1539,6 @@ gloog (scop_p scop, VEC (scop_p, heap) *scops, htab_t bb_pbb_mapping)
   ifsese if_region = NULL;
   htab_t rename_map, newivs_index, params_index;
   cloog_prog_clast pc;
-  int i;
 
   timevar_push (TV_GRAPHITE_CODE_GEN);
   gloog_error = false;
@@ -1579,11 +1578,6 @@ gloog (scop_p scop, VEC (scop_p, heap) *scops, htab_t bb_pbb_mapping)
 		   bb_pbb_mapping, 1, params_index);
   graphite_verify ();
   scev_reset_htab ();
-  rename_nb_iterations (rename_map);
-
-  for (i = 0; VEC_iterate (scop_p, scops, i, scop); i++)
-    rename_sese_parameters (rename_map, SCOP_REGION (scop));
-
   recompute_all_dominators ();
   graphite_verify ();
 
diff --git a/gcc/graphite-clast-to-gimple.h b/gcc/graphite-clast-to-gimple.h
index f6557f6..c90cfc0 100644
--- a/gcc/graphite-clast-to-gimple.h
+++ b/gcc/graphite-clast-to-gimple.h
@@ -36,7 +36,7 @@ typedef struct bb_pbb_def
   poly_bb_p pbb;
 }bb_pbb_def;
 
-extern bool gloog (scop_p, VEC (scop_p, heap) *, htab_t);
+extern bool gloog (scop_p, htab_t);
 extern cloog_prog_clast scop_to_clast (scop_p);
 extern void debug_clast_stmt (struct clast_stmt *);
 extern void print_clast_stmt (FILE *, struct clast_stmt *);
diff --git a/gcc/graphite.c b/gcc/graphite.c
index 5f28aeb..51bf800 100644
--- a/gcc/graphite.c
+++ b/gcc/graphite.c
@@ -289,7 +289,7 @@ graphite_transform_loops (void)
   for (i = 0; VEC_iterate (scop_p, scops, i, scop); i++)
     if (POLY_SCOP_P (scop)
 	&& apply_poly_transforms (scop)
-	&& gloog (scop, scops, bb_pbb_mapping))
+	&& gloog (scop, bb_pbb_mapping))
       need_cfg_cleanup_p = true;
 
   htab_delete (bb_pbb_mapping);
diff --git a/gcc/sese.c b/gcc/sese.c
index 051599e..361c9b6 100644
--- a/gcc/sese.c
+++ b/gcc/sese.c
@@ -435,62 +435,6 @@ set_rename (htab_t map, tree old_name, tree expr)
   *slot = new_rename_map_elt (old_name, expr);
 }
 
-/* Renames the expression T following the tuples (OLD_NAME, EXPR) in
-   the rename map M.  Returns the expression T after renaming.  */
-
-static tree
-rename_variables_in_expr (htab_t m, tree t)
-{
-  if (!t)
-    return t;
-
- if (TREE_CODE (t) == SSA_NAME)
-   return get_rename (m, t);
-
-  switch (TREE_CODE_LENGTH (TREE_CODE (t)))
-    {
-    case 3:
-      TREE_OPERAND (t, 2) = rename_variables_in_expr (m, TREE_OPERAND (t, 2));
-
-    case 2:
-      TREE_OPERAND (t, 1) = rename_variables_in_expr (m, TREE_OPERAND (t, 1));
-
-    case 1:
-      TREE_OPERAND (t, 0) = rename_variables_in_expr (m, TREE_OPERAND (t, 0));
-
-    default:
-      return t;
-    }
-}
-
-/* Renames all the loop->nb_iterations expressions following the
-   tuples (OLD_NAME, EXPR) in RENAME_MAP.  */
-
-void
-rename_nb_iterations (htab_t rename_map)
-{
-  loop_iterator li;
-  struct loop *loop;
-
-  FOR_EACH_LOOP (li, loop, 0)
-    loop->nb_iterations = rename_variables_in_expr (rename_map,
-						    loop->nb_iterations);
-}
-
-/* Renames all the parameters of SESE following the tuples (OLD_NAME,
-   EXPR) in RENAME_MAP.  */
-
-void
-rename_sese_parameters (htab_t rename_map, sese region)
-{
-  int i;
-  tree p;
-
-  for (i = 0; VEC_iterate (tree, SESE_PARAMS (region), i, p); i++)
-    VEC_replace (tree, SESE_PARAMS (region), i,
-		 rename_variables_in_expr (rename_map, p));
-}
-
 /* Rename the SSA_NAMEs used in STMT and that appear in MAP.  */
 
 static void
diff --git a/gcc/sese.h b/gcc/sese.h
index 4b2e5c2..d0d829f 100644
--- a/gcc/sese.h
+++ b/gcc/sese.h
@@ -262,8 +262,6 @@ extern void debug_rename_map (htab_t);
 extern hashval_t rename_map_elt_info (const void *);
 extern int eq_rename_map_elts (const void *, const void *);
 extern void set_rename (htab_t, tree, tree);
-extern void rename_nb_iterations (htab_t);
-extern void rename_sese_parameters (htab_t, sese);
 
 /* Constructs a new SCEV_INFO_STR structure for VAR and INSTANTIATED_BELOW.  */
 
-- 
1.7.0.4


^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: [patch][graphite] Remove sese_adjust_liveout_phis
  2010-06-12  8:17 [patch][graphite] Remove sese_adjust_liveout_phis Sebastian Pop
  2010-06-12 12:03 ` Sebastian Pop
@ 2010-06-14 18:20 ` Sebastian Pop
  2010-06-15  9:31   ` Richard Guenther
  1 sibling, 1 reply; 5+ messages in thread
From: Sebastian Pop @ 2010-06-14 18:20 UTC (permalink / raw)
  To: GCC Patches, gcc-graphite, Richard Guenther

Hi,

On Sat, Jun 12, 2010 at 02:53, Sebastian Pop <sebpop@gmail.com> wrote:
> Hi,
>
> with the attached patches we no longer need to use the rename_map to
> substitute the arguments of the phi nodes created after the code
> generated by Graphite.  I committed these patches to the graphite
> branch, and I will commit them to trunk once they pass the usual
> graphite branch tests.
>
> The aim is to replace the rename_map with the map that CLooG exposes
> for each user statement: associating to a loop level an expression.
> The expand_* functions will be replaced by the code generation from a
> chrec that is obtained from the original scev by chrec_applying the
> loop level -> expression map on each varying loop.  This will then
> allow us to remove the call to the IV canonicalization from Graphite.
>

One of these patches broke 464.h264ref: here is a reduced testcase

/* { dg-options "-O3 -fgraphite-identity -ffast-math" } */

typedef enum
{
  I_SLICE,
} SliceType;
typedef struct
{
  int type;
} ImageParameters;
extern ImageParameters *img;
int A[64], B[64], C[13][8][8], D[13][8][8];

void
foo (int q, int temp)
{
  int i, j, k;
  for(k=0; k<13; k++)
    for(j=0; j<8; j++)
      for(i=0; i<8; i++)
	{
	  if (img->type == I_SLICE)
	    C[k][j][i] = A[temp] << q;
	  D[k][j][i] = B[temp] << q;
	}
}

The problem is that img->type is moved out of the loop, making it a
parameter of the loop, and then this parameter is copied inside the
first loop:

p = img->type;
loop_1
  q = p
  loop_2
    loop_3
       if (q == 0) etc.

Now because we rewrite the scalar dependence into a data dependence,

p = img->type;
loop_1
  q = p
  cross_bb_dependence[0] = q
  loop_2
    loop_3
       if (cross_bb_dependence[0] == 0) etc.

scev is not smart enough to infer that cross_bb_dependence[0] is a
parameter defined outside the scop.

A possible fix for this is to run copy propagation before Graphite:

diff --git a/gcc/passes.c b/gcc/passes.c
index ad444fd..a646d19 100644
--- a/gcc/passes.c
+++ b/gcc/passes.c
@@ -897,6 +897,7 @@ init_optimization_passes (void)
 	  NEXT_PASS (pass_check_data_deps);
 	  NEXT_PASS (pass_loop_distribution);
 	  NEXT_PASS (pass_linear_transform);
+	  NEXT_PASS (pass_copy_prop);
 	  NEXT_PASS (pass_graphite_transforms);
 	    {
 	      struct opt_pass **p = &pass_graphite_transforms.pass.sub;


I don't know how to make the execution of copy_prop conditional to the
execution of Graphite.  Any suggestions?

Thanks,
Sebastian

^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: [patch][graphite] Remove sese_adjust_liveout_phis
  2010-06-14 18:20 ` Sebastian Pop
@ 2010-06-15  9:31   ` Richard Guenther
  2010-06-15 15:53     ` Sebastian Pop
  0 siblings, 1 reply; 5+ messages in thread
From: Richard Guenther @ 2010-06-15  9:31 UTC (permalink / raw)
  To: Sebastian Pop; +Cc: GCC Patches, gcc-graphite

[-- Attachment #1: Type: TEXT/PLAIN, Size: 2825 bytes --]

On Mon, 14 Jun 2010, Sebastian Pop wrote:

> Hi,
> 
> On Sat, Jun 12, 2010 at 02:53, Sebastian Pop <sebpop@gmail.com> wrote:
> > Hi,
> >
> > with the attached patches we no longer need to use the rename_map to
> > substitute the arguments of the phi nodes created after the code
> > generated by Graphite.  I committed these patches to the graphite
> > branch, and I will commit them to trunk once they pass the usual
> > graphite branch tests.
> >
> > The aim is to replace the rename_map with the map that CLooG exposes
> > for each user statement: associating to a loop level an expression.
> > The expand_* functions will be replaced by the code generation from a
> > chrec that is obtained from the original scev by chrec_applying the
> > loop level -> expression map on each varying loop.  This will then
> > allow us to remove the call to the IV canonicalization from Graphite.
> >
> 
> One of these patches broke 464.h264ref: here is a reduced testcase
> 
> /* { dg-options "-O3 -fgraphite-identity -ffast-math" } */
> 
> typedef enum
> {
>   I_SLICE,
> } SliceType;
> typedef struct
> {
>   int type;
> } ImageParameters;
> extern ImageParameters *img;
> int A[64], B[64], C[13][8][8], D[13][8][8];
> 
> void
> foo (int q, int temp)
> {
>   int i, j, k;
>   for(k=0; k<13; k++)
>     for(j=0; j<8; j++)
>       for(i=0; i<8; i++)
> 	{
> 	  if (img->type == I_SLICE)
> 	    C[k][j][i] = A[temp] << q;
> 	  D[k][j][i] = B[temp] << q;
> 	}
> }
> 
> The problem is that img->type is moved out of the loop, making it a
> parameter of the loop, and then this parameter is copied inside the
> first loop:
> 
> p = img->type;
> loop_1
>   q = p
>   loop_2
>     loop_3
>        if (q == 0) etc.
> 
> Now because we rewrite the scalar dependence into a data dependence,
> 
> p = img->type;
> loop_1
>   q = p
>   cross_bb_dependence[0] = q
>   loop_2
>     loop_3
>        if (cross_bb_dependence[0] == 0) etc.
> 
> scev is not smart enough to infer that cross_bb_dependence[0] is a
> parameter defined outside the scop.
> 
> A possible fix for this is to run copy propagation before Graphite:
> 
> diff --git a/gcc/passes.c b/gcc/passes.c
> index ad444fd..a646d19 100644
> --- a/gcc/passes.c
> +++ b/gcc/passes.c
> @@ -897,6 +897,7 @@ init_optimization_passes (void)
>  	  NEXT_PASS (pass_check_data_deps);
>  	  NEXT_PASS (pass_loop_distribution);
>  	  NEXT_PASS (pass_linear_transform);
> +	  NEXT_PASS (pass_copy_prop);
>  	  NEXT_PASS (pass_graphite_transforms);
>  	    {
>  	      struct opt_pass **p = &pass_graphite_transforms.pass.sub;
> 
> 
> I don't know how to make the execution of copy_prop conditional to the
> execution of Graphite.  Any suggestions?

Make pass_graphite_transforms a wrapper only (similar to 
pass_tree_loop), put the real graphite-transforms pass inside,
preceeded by copyprop.

Richard.

^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: [patch][graphite] Remove sese_adjust_liveout_phis
  2010-06-15  9:31   ` Richard Guenther
@ 2010-06-15 15:53     ` Sebastian Pop
  0 siblings, 0 replies; 5+ messages in thread
From: Sebastian Pop @ 2010-06-15 15:53 UTC (permalink / raw)
  To: Richard Guenther; +Cc: GCC Patches, gcc-graphite

[-- Attachment #1: Type: text/plain, Size: 347 bytes --]

On Tue, Jun 15, 2010 at 03:21, Richard Guenther <rguenther@suse.de> wrote:
> Make pass_graphite_transforms a wrapper only (similar to
> pass_tree_loop), put the real graphite-transforms pass inside,
> preceeded by copyprop.

Thanks for the recommendation.
Attached is the patch that I committed to the Graphite branch for further test.

Sebastian

[-- Attachment #2: 0001-Schedule-a-copy_prop-pass-before-graphite.patch --]
[-- Type: text/x-diff, Size: 4961 bytes --]

From 5bc5d64aa606ee3de31585856367485eed3d00bd Mon Sep 17 00:00:00 2001
From: Sebastian Pop <sebpop@gmail.com>
Date: Mon, 14 Jun 2010 12:23:00 -0500
Subject: [PATCH] Schedule a copy_prop pass before graphite.

2010-06-15  Sebastian Pop  <sebastian.pop@amd.com>

	* passes.c (init_optimization_passes): Add pass_graphite.
	Schedule a pass_copy_prop before pass_graphite_transforms.
	* timevar.def (TV_GRAPHITE): Declared.
	* tree-pass.h (pass_graphite): Declared.
	* tree-ssa-loop.c (pass_graphite): New.

	* gcc.dg/graphite/id-20.c: New.
---
 gcc/ChangeLog.graphite                |   10 ++++++++++
 gcc/passes.c                          |    7 +++++--
 gcc/testsuite/gcc.dg/graphite/id-20.c |   26 ++++++++++++++++++++++++++
 gcc/timevar.def                       |    1 +
 gcc/tree-pass.h                       |    1 +
 gcc/tree-ssa-loop.c                   |   19 +++++++++++++++++++
 6 files changed, 62 insertions(+), 2 deletions(-)
 create mode 100644 gcc/testsuite/gcc.dg/graphite/id-20.c

diff --git a/gcc/ChangeLog.graphite b/gcc/ChangeLog.graphite
index b4e2700..fa75c61 100644
--- a/gcc/ChangeLog.graphite
+++ b/gcc/ChangeLog.graphite
@@ -1,3 +1,13 @@
+2010-06-15  Sebastian Pop  <sebastian.pop@amd.com>
+
+	* passes.c (init_optimization_passes): Add pass_graphite.
+	Schedule a pass_copy_prop before pass_graphite_transforms.
+	* timevar.def (TV_GRAPHITE): Declared.
+	* tree-pass.h (pass_graphite): Declared.
+	* tree-ssa-loop.c (pass_graphite): New.
+
+	* gcc.dg/graphite/id-20.c: New.
+
 2010-06-12  Sebastian Pop  <sebastian.pop@amd.com>
 
 	* graphite-clast-to-gimple.c (gloog): Do not pass scops in parameter.
diff --git a/gcc/passes.c b/gcc/passes.c
index ad444fd..093d8e8 100644
--- a/gcc/passes.c
+++ b/gcc/passes.c
@@ -897,9 +897,12 @@ init_optimization_passes (void)
 	  NEXT_PASS (pass_check_data_deps);
 	  NEXT_PASS (pass_loop_distribution);
 	  NEXT_PASS (pass_linear_transform);
-	  NEXT_PASS (pass_graphite_transforms);
+	  NEXT_PASS (pass_copy_prop);
+	  NEXT_PASS (pass_graphite);
 	    {
-	      struct opt_pass **p = &pass_graphite_transforms.pass.sub;
+	      struct opt_pass **p = &pass_graphite.pass.sub;
+	      NEXT_PASS (pass_copy_prop);
+	      NEXT_PASS (pass_graphite_transforms);
 	      NEXT_PASS (pass_copy_prop);
 	      NEXT_PASS (pass_dce_loop);
 	      NEXT_PASS (pass_lim);
diff --git a/gcc/testsuite/gcc.dg/graphite/id-20.c b/gcc/testsuite/gcc.dg/graphite/id-20.c
new file mode 100644
index 0000000..7e0cf6d
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/graphite/id-20.c
@@ -0,0 +1,26 @@
+/* { dg-options "-O3 -fgraphite-identity -ffast-math" } */
+
+typedef enum
+{
+  I_SLICE,
+} SliceType;
+typedef struct
+{
+  int type;
+} ImageParameters;
+extern ImageParameters *img;
+int A[64], B[64], C[13][8][8], D[13][8][8];
+
+void
+foo (int q, int temp)
+{
+  int i, j, k;
+  for(k=0; k<13; k++)
+    for(j=0; j<8; j++)
+      for(i=0; i<8; i++)
+	{
+	  if (img->type == I_SLICE)
+	    C[k][j][i] = A[temp] << q;
+	  D[k][j][i] = B[temp] << q;
+	}
+}
diff --git a/gcc/timevar.def b/gcc/timevar.def
index a6ee4d2..9d56124 100644
--- a/gcc/timevar.def
+++ b/gcc/timevar.def
@@ -146,6 +146,7 @@ DEFTIMEVAR (TV_COMPLETE_UNROLL       , "complete unrolling")
 DEFTIMEVAR (TV_TREE_PARALLELIZE_LOOPS, "tree parallelize loops")
 DEFTIMEVAR (TV_TREE_VECTORIZATION    , "tree vectorization")
 DEFTIMEVAR (TV_TREE_SLP_VECTORIZATION, "tree slp vectorization")
+DEFTIMEVAR (TV_GRAPHITE              , "Graphite")
 DEFTIMEVAR (TV_GRAPHITE_TRANSFORMS   , "Graphite loop transforms")
 DEFTIMEVAR (TV_GRAPHITE_DATA_DEPS    , "Graphite data dep analysis")
 DEFTIMEVAR (TV_GRAPHITE_CODE_GEN     , "Graphite code generation")
diff --git a/gcc/tree-pass.h b/gcc/tree-pass.h
index b5971d5..9af12d9 100644
--- a/gcc/tree-pass.h
+++ b/gcc/tree-pass.h
@@ -372,6 +372,7 @@ extern struct gimple_opt_pass pass_iv_canon;
 extern struct gimple_opt_pass pass_scev_cprop;
 extern struct gimple_opt_pass pass_empty_loop;
 extern struct gimple_opt_pass pass_record_bounds;
+extern struct gimple_opt_pass pass_graphite;
 extern struct gimple_opt_pass pass_graphite_transforms;
 extern struct gimple_opt_pass pass_if_conversion;
 extern struct gimple_opt_pass pass_loop_distribution;
diff --git a/gcc/tree-ssa-loop.c b/gcc/tree-ssa-loop.c
index 344cfa8..006688a 100644
--- a/gcc/tree-ssa-loop.c
+++ b/gcc/tree-ssa-loop.c
@@ -309,6 +309,25 @@ gate_graphite_transforms (void)
   return flag_graphite != 0;
 }
 
+struct gimple_opt_pass pass_graphite =
+{
+ {
+  GIMPLE_PASS,
+  "graphite0",				/* name */
+  gate_graphite_transforms,		/* gate */
+  NULL,					/* execute */
+  NULL,					/* sub */
+  NULL,					/* next */
+  0,					/* static_pass_number */
+  TV_GRAPHITE,				/* tv_id */
+  PROP_cfg | PROP_ssa,			/* properties_required */
+  0,					/* properties_provided */
+  0,					/* properties_destroyed */
+  0,					/* todo_flags_start */
+  0					/* todo_flags_finish */
+ }
+};
+
 struct gimple_opt_pass pass_graphite_transforms =
 {
  {
-- 
1.7.0.4


^ permalink raw reply	[flat|nested] 5+ messages in thread

end of thread, other threads:[~2010-06-15 15:16 UTC | newest]

Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2010-06-12  8:17 [patch][graphite] Remove sese_adjust_liveout_phis Sebastian Pop
2010-06-12 12:03 ` Sebastian Pop
2010-06-14 18:20 ` Sebastian Pop
2010-06-15  9:31   ` Richard Guenther
2010-06-15 15:53     ` Sebastian Pop

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).