[tree-ssa] bootstrap comparison failure on ppc and ppc64

[tree-ssa] bootstrap comparison failure on ppc and ppc64

[Diego Novillo]

Jeff, Andrew,

One of these three patches is causing a comparison failure on ppc and
ppc64.  I'm going to try Andrew's fix from this morning to see if it has
any effect.


Diego.

--- gcc.prev/gcc/ChangeLog.tree-ssa     2003-07-28 00:04:32.000000000 -0400
+++ gcc/gcc/ChangeLog.tree-ssa  2003-07-29 00:05:38.000000000 -0400
@@ -1,3 +1,34 @@
+2003-07-28  Diego Novillo  <dnovillo@redhat.com>
+
+       * opts.c (decode_options): Disable must-alias optimization.
+
+2003-07-28  Andrew MacLeod  <amacleod@redhat.com>
+
+       * tree-ssa.c (struct _elim_graph): Add varray for constant copies.
+       (new_elim_graph): Initialize constant copy array..
+       (eliminate_build): Push constant copies onto stack instead of emitting.
+       (eliminate_phi): Emit any pending constant copies.
+
+2003-07-28  Jeff Law  <law@redhat.com>
+
+       * tree-ssa-dom.c (optimize_block): If a PHI has a single argument
+       that is a constant, then that creates a useful equivalence.
+       Propagate constant values into PHI nodes.
+
+       * tree-flow-inline.h (may_propagate_copy): Allow RHS to be a
+       constant.
+
+       * tree-dfa.c (compute_immediate_uses_for): Do not assume that
+       PHI arguments are SSA_NAMEs.
+       * tree-ssa-dce.c (process_worklist): Likewise.
+       * tree-ssa-copyprop.c (copyprop_phi): Likewise.  Use may_propagate_copy.
+       * tree-ssa-ccp.c (visit_phi_node): Do not assume that PHI arguments
+       are SSA_NAMEs.  Create a suitable value if a PHI argument is a
+       constant.
+
+       * tree-cfg.c (move_outgoing_edges): Correctly handle case where
+       an edge already exists from BB1 to BB2's successor.
+

Re: [tree-ssa] bootstrap comparison failure on ppc and ppc64

[Andrew MacLeod]

On Tue, 2003-07-29 at 07:52, Diego Novillo wrote:
> 
> Jeff, Andrew,
> 
> One of these three patches is causing a comparison failure on ppc and
> ppc64.  I'm going to try Andrew's fix from this morning to see if it has
> any effect.
> 
> 

I dont think any of these changes actually turn on the PHI as constant
code, so neither my patch nor most of these *ought* to affect anything
:-). Famous last words eh :-)

The likely things I see that affect code are you disabling must-alias
and the change to move_outgoing_edges.  

Im sure glad my big patch last night wasn't in this list :-)

Andrew

> Diego.
> 
> --- gcc.prev/gcc/ChangeLog.tree-ssa     2003-07-28 00:04:32.000000000 -0400
> +++ gcc/gcc/ChangeLog.tree-ssa  2003-07-29 00:05:38.000000000 -0400
> @@ -1,3 +1,34 @@
> +2003-07-28  Diego Novillo  <dnovillo@redhat.com>
> +
> +       * opts.c (decode_options): Disable must-alias optimization.
> +
> +2003-07-28  Andrew MacLeod  <amacleod@redhat.com>
> +
> +       * tree-ssa.c (struct _elim_graph): Add varray for constant copies.
> +       (new_elim_graph): Initialize constant copy array..
> +       (eliminate_build): Push constant copies onto stack instead of emitting.
> +       (eliminate_phi): Emit any pending constant copies.
> +
> +2003-07-28  Jeff Law  <law@redhat.com>
> +
> +       * tree-ssa-dom.c (optimize_block): If a PHI has a single argument
> +       that is a constant, then that creates a useful equivalence.
> +       Propagate constant values into PHI nodes.
> +
> +       * tree-flow-inline.h (may_propagate_copy): Allow RHS to be a
> +       constant.
> +
> +       * tree-dfa.c (compute_immediate_uses_for): Do not assume that
> +       PHI arguments are SSA_NAMEs.
> +       * tree-ssa-dce.c (process_worklist): Likewise.
> +       * tree-ssa-copyprop.c (copyprop_phi): Likewise.  Use may_propagate_copy.
> +       * tree-ssa-ccp.c (visit_phi_node): Do not assume that PHI arguments
> +       are SSA_NAMEs.  Create a suitable value if a PHI argument is a
> +       constant.
> +
> +       * tree-cfg.c (move_outgoing_edges): Correctly handle case where
> +       an edge already exists from BB1 to BB2's successor.
> +
> 

Re: [tree-ssa] bootstrap comparison failure on ppc and ppc64

[Diego Novillo]

On Tue, 2003-07-29 at 08:01, Andrew MacLeod wrote:

> Im sure glad my big patch last night wasn't in this list :-)
> 
Hmm, it's even worse now.  I just updated the local trees again.  We're
now getting an ICE building libgcc.  This includes your big patch from
last night.

Full log at
http://people.redhat.com/dnovillo/nightly-testing/tree-ssa-branch/ppc/log/20030729/

I'm on my way out now.  I'll take a closer look when I get into the
office.


Diego.



+2003-07-29  Andrew MacLeod  <amacleod@redhat.com>
+
+	* tree-ssa.c (insert_copy_on_edge): Only set used bit on DECL nodes.
+
+2003-07-29  Andrew MacLeod  <amacleod@redhat.com>
+
+	* common.opt (ftree-combine-temps): Add new option.
+	* flags.h (flag_tree_combine_temps): New flag.
+	* opts.c (decode_options): Initialize flag_tree_combine_temps.
+	(common_handle_option): Handle new flag.
+	* toplev.c (flag_tree_combine_temps): Declare.
+	(lang_independent_options f): Add tree-combine-temp.
+	* tree-ssa-live.c (var_union): When combining 2 root variables, choose 
+	the user variable over a temporary as the new variable.
+	(compact_var_map): Use renamed root_var routines.
+	(calculate_live_on_exit): Reformatting.
+	(tpa_init): Initialize a tpa object.
+	(tpa_remove_partition): Remove a partition from a tpa list.
+	(tpa_delete): Delete a tpa object.
+	(tpa_compact): Hide single elemenet lists.
+	(root_var_init): Split common part into tpa_init and rename.
+	(remove_root_var_partition, delete_root_va, dump_root_var): Delete.
+	(type_var_init): New. Initialize a type_var object.
+	(create_coalesce_list): New. Create a coalesce_list object.
+	(delete_coalesce_list): New. Free a coalesce list's memory.
+	(find_partition_pair): New. Find a coalesce pair in a coalesce list.
+	(add_coalesce): New. Add a coalesce between 2 partitions.
+	(sort_coalesce_list): New. Sort coalesce pairs by importance.
+	(pop_best_coalesce): New. Get best remaining pair to coalesce.
+	(add_conflicts_if_valid): Move from tree-ssa.c.
+	(build_tree_conflict_graph): Move from coalesce_ssa_name in tree-ssa.c.
+	Genericize to use tpa_p instead of root_var object. Don't add 
+	interferences between copies. Update coalesce list.
+	(coalesce_tpa_members): Move from coalesce_ssa_name in tree-ssa.c. Use
+	tpa_p instead of root_var. Use coalesce list if provided.
+	(dump_coalesce_list): New. Show debug info for a coalesce list.
+	(tpa_dump): Rename from dump_root_var and genericize to use tpa_p.
+	* tree-ssa-live.h (root_var_p): Rename structure type to tpa_p
+	(tpa_num_trees, tpa_tree, tpa_first_partition, tpa_next_partition, 
+	tpa_find_tree): New. Generic versions of existing root_var routines.
+	(tpa_decompact): New. Include single version lists.
+	(root_var_p): Declare as type tpa_p.
+	(root_var_num, root_var, root_var_first_partition, 
+	root_var_next_partition, root_var_dump, root_var_delete,
+	root_var_remove_partition, root_var_find , root_var_compac,
+	root_var_decompac): Rename and call generic versions.
+	(type_var_p): New. Use tpa_p structure for a type based association.
+	(type_var_num, type_var, type_var_first_partition, 
+	type_var_next_partition, type_var_dump, type_var_delete,
+	type_var_remove_partition, type_var_find, type_var_compact, 
+	type_var_decompact): New. Call generic versions of the routine.
+	(struct partition_pair_d): New. Represent a desired coalesce.
+	(struct coalesce_list_d): New. Organize lists of desired coalesces.
+	(NO_BEST_COALESCE): Define value.
+	* tree-ssa.c (set_if_valid): Remove.
+	(insert_copy_on_edge): Set variable as used when inserting a copy.
+	(add_conflicts_if_valid): Remove. Move to tree-ssa-live.c.
+	(print_exprs): New. Routine for commonly used output format.
+	(coalesce_abnormal_edges): New. Split from coalece_ssa_name. Force
+	partition coalesces across abnormal edges.
+	(coalesce_ssa_name): Split out build_tree_conflict_graph, 
+	coalesce_abnormal_edges, and coalesce_tpa_members. Return live
+	range info if required. Use renamed root_var routines.
+	(assign_vars): Use renamed root_var routines.
+	(replace_variable): Mark as inline.
+	(coalesce_vars): Coalesce variable memory storage.
+	(rewrite_out_of_ssa): Don't compact varmap anymore. Free live range
+	info if required. Call coalesce_vars if combining temps.
+
+2003-07-29  Andrew MacLeod  <amacleod@redhat.com>
+
+	* tree-cfg.c (handle_switch_fallthru): Use bsi_link_after if stmt is
+	in a basic block.
+
> Andrew
> 
> > Diego.
> > 
> > --- gcc.prev/gcc/ChangeLog.tree-ssa     2003-07-28 00:04:32.000000000 -0400
> > +++ gcc/gcc/ChangeLog.tree-ssa  2003-07-29 00:05:38.000000000 -0400
> > @@ -1,3 +1,34 @@
> > +2003-07-28  Diego Novillo  <dnovillo@redhat.com>
> > +
> > +       * opts.c (decode_options): Disable must-alias optimization.
> > +
> > +2003-07-28  Andrew MacLeod  <amacleod@redhat.com>
> > +
> > +       * tree-ssa.c (struct _elim_graph): Add varray for constant copies.
> > +       (new_elim_graph): Initialize constant copy array..
> > +       (eliminate_build): Push constant copies onto stack instead of emitting.
> > +       (eliminate_phi): Emit any pending constant copies.
> > +
> > +2003-07-28  Jeff Law  <law@redhat.com>
> > +
> > +       * tree-ssa-dom.c (optimize_block): If a PHI has a single argument
> > +       that is a constant, then that creates a useful equivalence.
> > +       Propagate constant values into PHI nodes.
> > +
> > +       * tree-flow-inline.h (may_propagate_copy): Allow RHS to be a
> > +       constant.
> > +
> > +       * tree-dfa.c (compute_immediate_uses_for): Do not assume that
> > +       PHI arguments are SSA_NAMEs.
> > +       * tree-ssa-dce.c (process_worklist): Likewise.
> > +       * tree-ssa-copyprop.c (copyprop_phi): Likewise.  Use may_propagate_copy.
> > +       * tree-ssa-ccp.c (visit_phi_node): Do not assume that PHI arguments
> > +       are SSA_NAMEs.  Create a suitable value if a PHI argument is a
> > +       constant.
> > +
> > +       * tree-cfg.c (move_outgoing_edges): Correctly handle case where
> > +       an edge already exists from BB1 to BB2's successor.
> > +
> > 
> 
> 

Re: [tree-ssa] bootstrap comparison failure on ppc and ppc64

[Diego Novillo]

Jeff,

Undoing your patch allowed me to bootstrap again on ppc32.  I'm not
bootstrapping on my lame ppc box anymore, so I think this is
reproduceable, but just in case, could you try bootstrapping on your ppc
box?

We're having spurious bootstrap problems on ppc which I think are
related to aliasing.  I wish I could nail it down, but it's proving
annoying to fix.


Thanks.  Diego.


diff -r -x CVS -dupN gcc.prev/gcc/tree-cfg.c gcc/gcc/tree-cfg.c
--- gcc.prev/gcc/tree-cfg.c	2003-07-26 00:02:42.000000000 -0400
+++ gcc/gcc/tree-cfg.c	2003-07-29 00:03:16.000000000 -0400
@@ -4458,19 +4458,33 @@ move_outgoing_edges (basic_block bb1, ba
       old_edge = bb2->succ;
       new_edge = make_edge (bb1, old_edge->dest, old_edge->flags);
 
-      /* Update PHI nodes at BB2's successor.  The arguments that used to
-	 come from BB2 now come from BB1.  */
-      for (phi = phi_nodes (old_edge->dest); phi; phi = TREE_CHAIN (phi))
+      /* If make_edge created a new edge, then we need to update the PHI
+	 node at BB2's successor.  The arguments that used to come from
+	 BB2 now come from BB1.
+
+	 If make_edge did not create a new edge, then we already had an
+	 edge from BB1 to BB2's successor.  In this case we want to
+	 remove the edge and remove its alternative from BB2's successor's
+	 PHI nodes, hence we use ssa_remove_edge.  */
+      if (new_edge)
 	{
-	  int i;
-	  for (i = 0; i < PHI_NUM_ARGS (phi); i++)
-	    if (PHI_ARG_EDGE (phi, i) == old_edge)
-	      PHI_ARG_EDGE (phi, i) = new_edge;
+	  for (phi = phi_nodes (old_edge->dest); phi; phi = TREE_CHAIN (phi))
+	    {
+	      int i;
+	      for (i = 0; i < PHI_NUM_ARGS (phi); i++)
+		if (PHI_ARG_EDGE (phi, i) == old_edge)
+		  PHI_ARG_EDGE (phi, i) = new_edge;
+	    }
+
+	  /* Note that we shouldn't call ssa_remove_edge here because we've
+	     already dealt with PHI nodes.  */
+	  remove_edge (old_edge);
+	}
+      else
+	{
+	  ssa_remove_edge (old_edge);
 	}
 
-      /* Note that we shouldn't call ssa_remove_edge here because we've
-	 already dealt with PHI nodes.  */
-      remove_edge (old_edge);
     }
 
   /* BB2's dominator children are now BB1's.  Also, remove BB2 as a
diff -r -x CVS -dupN gcc.prev/gcc/tree-dfa.c gcc/gcc/tree-dfa.c
--- gcc.prev/gcc/tree-dfa.c	2003-07-24 23:13:07.000000000 -0400
+++ gcc/gcc/tree-dfa.c	2003-07-29 00:03:16.000000000 -0400
@@ -1048,9 +1048,14 @@ compute_immediate_uses_for (tree stmt, i
 
       for (i = 0; i < PHI_NUM_ARGS (stmt); i++)
 	{
-	  tree imm_rdef_stmt = SSA_NAME_DEF_STMT (PHI_ARG_DEF (stmt, i));
-	  if (!IS_EMPTY_STMT (imm_rdef_stmt))
-	    add_immediate_use (imm_rdef_stmt, stmt);
+	  tree arg = PHI_ARG_DEF (stmt, i);
+
+	  if (TREE_CODE (arg) == SSA_NAME)
+	    { 
+	      tree imm_rdef_stmt = SSA_NAME_DEF_STMT (PHI_ARG_DEF (stmt, i));
+	      if (!IS_EMPTY_STMT (imm_rdef_stmt))
+		add_immediate_use (imm_rdef_stmt, stmt);
+	    }
 	}
       return;
     }
diff -r -x CVS -dupN gcc.prev/gcc/tree-flow-inline.h gcc/gcc/tree-flow-inline.h
--- gcc.prev/gcc/tree-flow-inline.h	2003-07-23 19:48:15.000000000 -0400
+++ gcc/gcc/tree-flow-inline.h	2003-07-29 00:03:16.000000000 -0400
@@ -540,7 +540,8 @@ static inline bool
 may_propagate_copy (tree dest, tree orig)
 {
   return (!SSA_NAME_OCCURS_IN_ABNORMAL_PHI (dest)
-	  && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (orig)
+	  && (TREE_CONSTANT (orig)
+	      || !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (orig))
 	  && !DECL_HARD_REGISTER (SSA_NAME_VAR (dest)));
 }
 
diff -r -x CVS -dupN gcc.prev/gcc/tree-ssa-ccp.c gcc/gcc/tree-ssa-ccp.c
--- gcc.prev/gcc/tree-ssa-ccp.c	2003-07-22 14:36:38.000000000 -0400
+++ gcc/gcc/tree-ssa-ccp.c	2003-07-29 00:03:16.000000000 -0400
@@ -398,7 +398,19 @@ visit_phi_node (tree phi)
 	if (e->flags & EDGE_EXECUTABLE)
 	  {
 	    tree rdef = PHI_ARG_DEF (phi, i);
-	    value *rdef_val = get_value (rdef);
+	    value *rdef_val;
+
+	    if (TREE_CONSTANT (rdef))
+	      {
+		value val;
+
+		val.lattice_val = CONSTANT;
+		val.const_val = rdef;
+		rdef_val = &val;
+	      }
+	    else
+	      rdef_val = get_value (rdef);
+
 	    phi_val = cp_lattice_meet (phi_val, *rdef_val);
 
 	    if (dump_file && (dump_flags & TDF_DETAILS))
diff -r -x CVS -dupN gcc.prev/gcc/tree-ssa-copyprop.c gcc/gcc/tree-ssa-copyprop.c
--- gcc.prev/gcc/tree-ssa-copyprop.c	2003-07-23 19:48:16.000000000 -0400
+++ gcc/gcc/tree-ssa-copyprop.c	2003-07-29 00:03:16.000000000 -0400
@@ -150,11 +150,13 @@ copyprop_phi (tree phi)
   for (i = 0; i < PHI_NUM_ARGS (phi); i++)
     {
       tree arg = PHI_ARG_DEF (phi, i);
-      tree orig = get_original (arg);
+      tree orig;
 
-      if (orig
-	  && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (arg)
-	  && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (orig))
+      if (TREE_CODE (arg) != SSA_NAME)
+	continue;
+
+      orig = get_original (arg);
+      if (orig && may_propagate_copy (arg, orig))
 	{
 	  if (dump_file && dump_flags & TDF_DETAILS)
 	    {
diff -r -x CVS -dupN gcc.prev/gcc/tree-ssa-dce.c gcc/gcc/tree-ssa-dce.c
--- gcc.prev/gcc/tree-ssa-dce.c	2003-07-21 22:50:15.000000000 -0400
+++ gcc/gcc/tree-ssa-dce.c	2003-07-29 00:03:17.000000000 -0400
@@ -357,7 +357,11 @@ process_worklist (void)
 	  /* All the statements feeding this PHI node's arguments are
 	     necessary.  */
 	  for (k = 0; k < PHI_NUM_ARGS (i); k++)
-	    mark_necessary (SSA_NAME_DEF_STMT (PHI_ARG_DEF (i, k)));
+	    {
+	      tree arg = PHI_ARG_DEF (i, k);
+	      if (TREE_CODE (arg) == SSA_NAME)
+		mark_necessary (SSA_NAME_DEF_STMT (PHI_ARG_DEF (i, k)));
+	    }
 
 	  /* Look at all the predecessors, and if this PHI is being fed
 	     from a conditional expression, mark that conditional
diff -r -x CVS -dupN gcc.prev/gcc/tree-ssa-dom.c gcc/gcc/tree-ssa-dom.c
--- gcc.prev/gcc/tree-ssa-dom.c	2003-07-26 00:02:42.000000000 -0400
+++ gcc/gcc/tree-ssa-dom.c	2003-07-29 00:03:17.000000000 -0400
@@ -261,7 +261,8 @@ optimize_block (basic_block bb, tree par
      creates an equivalence we can record into the const_and_copies table.  */
   for (phi = phi_nodes (bb); phi; phi = TREE_CHAIN (phi))
     if (PHI_NUM_ARGS (phi) == 1
-	&& TREE_CODE (PHI_ARG_DEF (phi, 0)) == SSA_NAME
+	&& (TREE_CODE (PHI_ARG_DEF (phi, 0)) == SSA_NAME
+	    || TREE_CONSTANT (PHI_ARG_DEF (phi, 0)))
 	&& may_propagate_copy (PHI_RESULT (phi), PHI_ARG_DEF (phi, 0)))
       set_value_for (PHI_RESULT (phi), PHI_ARG_DEF (phi, 0), const_and_copies);
 
@@ -286,12 +287,12 @@ optimize_block (basic_block bb, tree par
 		  if (! SSA_VAR_P (*orig_p))
 		    break;
 
-		  /* FIXME.  We should be able to propagate constants into
-		     PHI nodes in the not too distant future.  */
 		  new = get_value_for (*orig_p, const_and_copies);
+		  /* We want to allow copy propagation as well as constant
+		     propagation.  */
 		  if (new
-		      && TREE_CODE (new) == SSA_NAME
-		      && may_propagate_copy (new, *orig_p))
+		      && (TREE_CODE (new) == SSA_NAME || TREE_CONSTANT (new))
+		      && may_propagate_copy (*orig_p, new))
 		    *orig_p = new;
 		  break;
 		}

Re: [tree-ssa] bootstrap comparison failure on ppc and ppc64

[law]

In message <1059497409.3164.73.camel@frodo.toronto.redhat.com>, Diego Novillo w
rites:
 >Jeff,
 >
 >Undoing your patch allowed me to bootstrap again on ppc32.  I'm not
 >bootstrapping on my lame ppc box anymore, so I think this is
 >reproduceable, but just in case, could you try bootstrapping on your ppc
 >box?
 >
 >We're having spurious bootstrap problems on ppc which I think are
 >related to aliasing.  I wish I could nail it down, but it's proving
 >annoying to fix.
Did you ever fix the problem with getting different version #s and all
that stuff?  IF those underlying issues aren't fixed, then there's little
point in diving into these changes.  Fundamentally we've got a bigger
underlying problem.  

jeff

Re: [tree-ssa] bootstrap comparison failure on ppc and ppc64

[Diego Novillo]

On Tue, 2003-07-29 at 13:01, law@redhat.com wrote:

> Did you ever fix the problem with getting different version #s and all
> that stuff?  IF those underlying issues aren't fixed, then there's little
> point in diving into these changes.  Fundamentally we've got a bigger
> underlying problem.  
> 
I must have, because it's not happening anymore.

We do have a sneaky aliasing problem, that's why I'd like you to
bootstrap on your ppc box.  If you can't, that's fine, I'll try
something else.

I'm basically not sure whether your patch triggered some other bug or if
it's buggy.  I'm trying other combinations to see which hunk is the one
triggering the problem.


Thanks.  Diego.

Re: [tree-ssa] bootstrap comparison failure on ppc and ppc64

[law]

In message <1059498911.3164.97.camel@frodo.toronto.redhat.com>, Diego Novillo w
rites:
 >We do have a sneaky aliasing problem, that's why I'd like you to
 >bootstrap on your ppc box.  If you can't, that's fine, I'll try
 >something else.
It's running...  For the next N days....


Jeff