[PATCH] Time profiler - phase 2

[Martin LiÅka <marxin dot liska at gmail dot com>]

Hello,
   this patch introduces new function reordering feature that is based
on the patch I submitted few days ago.

The idea of a new flag (-fprofile-reorder-functions) is to reorder
functions according to the first execution in instrumented run. Of
course, the option operates only in LTO mode, where the compiler can
see all call graph nodes.

I've tested the patch on programs like Gimp, Inkscape, Firefox. There
are still some inaccuracies caused by weak symbols in C++ and has been
under investigation.

Thank you,
Martin

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index c566a85..1562098 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,15 @@
+2013-11-13	Martin Liska	<marxin.liska@gmail.com>
+						Jan Hubicka  <jh@suse.cz>
+
+	* cgraphunit.c (node_cmp): New function.
+	(expand_all_functions): Function ordering added.
+	* common.opt: New profile based function reordering flag introduced.
+	* coverage.c (get_coverage_counts): Wrong profile handled.
+	* ipa.c (cgraph_externally_visible_p): New late flag introduced.
+	* lto-partition.c: Support for time profile added.
+	* lto.c: Likewise.
+	* value-prof.c: Histogram instrumentation switch added.
+
 2013-11-13  Vladimir Makarov  <vmakarov@redhat.com>
 
 	PR rtl-optimization/59036
diff --git a/gcc/cgraphunit.c b/gcc/cgraphunit.c
index 4765e6a..7cdd9a4 100644
--- a/gcc/cgraphunit.c
+++ b/gcc/cgraphunit.c
@@ -1821,6 +1821,17 @@ expand_function (struct cgraph_node *node)
   ipa_remove_all_references (&node->ref_list);
 }
 
+/* Node comparer that is responsible for the order that corresponds
+   to time when a function was launched for the first time.  */
+
+static int
+node_cmp (const void *pa, const void *pb)
+{
+  const struct cgraph_node *a = *(const struct cgraph_node * const *) pa;
+  const struct cgraph_node *b = *(const struct cgraph_node * const *) pb;
+
+  return b->tp_first_run - a->tp_first_run;
+}
 
 /* Expand all functions that must be output.
 
@@ -1832,11 +1843,14 @@ expand_function (struct cgraph_node *node)
    to use subsections to make the output functions appear in top-down
    order).  */
 
+
 static void
 expand_all_functions (void)
 {
   struct cgraph_node *node;
   struct cgraph_node **order = XCNEWVEC (struct cgraph_node *, cgraph_n_nodes);
+
+  unsigned int expanded_func_count = 0, profiled_func_count = 0;
   int order_pos, new_order_pos = 0;
   int i;
 
@@ -1849,19 +1863,35 @@ expand_all_functions (void)
     if (order[i]->process)
       order[new_order_pos++] = order[i];
 
+  if (flag_profile_reorder_functions)
+    qsort (order, new_order_pos, sizeof (struct cgraph_node *), node_cmp);
+
   for (i = new_order_pos - 1; i >= 0; i--)
     {
       node = order[i];
+
       if (node->process)
 	{
+     expanded_func_count++;
+     if(node->tp_first_run)
+       profiled_func_count++;
+
 	  node->process = 0;
 	  expand_function (node);
 	}
     }
+
+    if (in_lto_p && dump_file)
+      fprintf (dump_file, "Expanded functions with time profile (%s):%u/%u\n",
+               main_input_filename, profiled_func_count, expanded_func_count);
+
+  if (cgraph_dump_file && flag_profile_reorder_functions && in_lto_p)
+    fprintf (cgraph_dump_file, "Expanded functions with time profile:%u/%u\n",
+             profiled_func_count, expanded_func_count);
+
   cgraph_process_new_functions ();
 
   free (order);
-
 }
 
 /* This is used to sort the node types by the cgraph order number.  */
@@ -2186,6 +2216,7 @@ compile (void)
 #endif
 
   cgraph_state = CGRAPH_STATE_EXPANSION;
+
   if (!flag_toplevel_reorder)
     output_in_order ();
   else
diff --git a/gcc/common.opt b/gcc/common.opt
index d5971df..85d5c74 100644
--- a/gcc/common.opt
+++ b/gcc/common.opt
@@ -1708,6 +1708,10 @@ fprofile-report
 Common Report Var(profile_report)
 Report on consistency of profile
 
+fprofile-reorder-functions
+Common Report Var(flag_profile_reorder_functions)
+Enable function reordering that improves code placement
+
 frandom-seed
 Common Var(common_deferred_options) Defer
 
diff --git a/gcc/ipa-profile.c b/gcc/ipa-profile.c
index 1260069..eff4b51 100644
--- a/gcc/ipa-profile.c
+++ b/gcc/ipa-profile.c
@@ -465,6 +465,7 @@ ipa_propagate_frequency (struct cgraph_node *node)
   if (d.maybe_unlikely_executed)
     {
       node->frequency = NODE_FREQUENCY_UNLIKELY_EXECUTED;
+      node->tp_first_run = 0;
       if (dump_file)
 	fprintf (dump_file, "Node %s promoted to unlikely executed.\n",
 		 cgraph_node_name (node));
diff --git a/gcc/ipa.c b/gcc/ipa.c
index a11b1c7..d92a332 100644
--- a/gcc/ipa.c
+++ b/gcc/ipa.c
@@ -761,10 +761,14 @@ cgraph_externally_visible_p (struct cgraph_node *node,
      This improves code quality and we know we will duplicate them at most twice
      (in the case that we are not using plugin and link with object file
       implementing same COMDAT)  */
-  if ((in_lto_p || whole_program)
-      && DECL_COMDAT (node->decl)
-      && comdat_can_be_unshared_p (node))
-    return false;
+  if ((in_lto_p || whole_program || profile_arc_flag)
+     && DECL_COMDAT (node->decl)
+     && comdat_can_be_unshared_p (node))
+    {
+      gcc_checking_assert (cgraph_function_body_availability (node)
+			   > AVAIL_OVERWRITABLE);
+      return false;
+    }
 
   /* When doing link time optimizations, hidden symbols become local.  */
   if (in_lto_p
@@ -932,7 +936,7 @@ function_and_variable_visibility (bool whole_program)
 	}
       gcc_assert ((!DECL_WEAK (node->decl)
 		  && !DECL_COMDAT (node->decl))
-      	          || TREE_PUBLIC (node->decl)
+		  || TREE_PUBLIC (node->decl)
 		  || node->weakref
 		  || DECL_EXTERNAL (node->decl));
       if (cgraph_externally_visible_p (node, whole_program))
@@ -949,7 +953,7 @@ function_and_variable_visibility (bool whole_program)
 	  && node->definition && !node->weakref
 	  && !DECL_EXTERNAL (node->decl))
 	{
-	  gcc_assert (whole_program || in_lto_p
+	  gcc_assert (whole_program || in_lto_p || profile_arc_flag
 		      || !TREE_PUBLIC (node->decl));
 	  node->unique_name = ((node->resolution == LDPR_PREVAILING_DEF_IRONLY
 				      || node->resolution == LDPR_PREVAILING_DEF_IRONLY_EXP)
diff --git a/gcc/lto/lto-partition.c b/gcc/lto/lto-partition.c
index 6a3d881..e4ca6df 100644
--- a/gcc/lto/lto-partition.c
+++ b/gcc/lto/lto-partition.c
@@ -280,9 +280,11 @@ add_symbol_to_partition (ltrans_partition part, symtab_node *node)
 
      Be lax about comdats; they may or may not be duplicated and we may
      end up in need to duplicate keyed comdat because it has unkeyed alias.  */
+
   gcc_assert (get_symbol_class (node) == SYMBOL_DUPLICATE
 	      || DECL_COMDAT (node->decl)
 	      || !symbol_partitioned_p (node));
+
   add_symbol_to_partition_1 (part, node);
 }
 
@@ -395,6 +397,20 @@ node_cmp (const void *pa, const void *pb)
 {
   const struct cgraph_node *a = *(const struct cgraph_node * const *) pa;
   const struct cgraph_node *b = *(const struct cgraph_node * const *) pb;
+
+  /* Profile reorder flag enables function reordering based on first execution
+     of a function. All functions with profile are placed in ascending
+     order at the beginning.  */
+
+  if (flag_profile_reorder_functions)
+  {
+    if (a->tp_first_run && b->tp_first_run)
+      return a->tp_first_run - b->tp_first_run;
+
+    if (a->tp_first_run || b->tp_first_run)
+      return b->tp_first_run - a->tp_first_run;
+  }
+
   return b->order - a->order;
 }
 
@@ -449,7 +465,7 @@ void
 lto_balanced_map (void)
 {
   int n_nodes = 0;
-  int n_varpool_nodes = 0, varpool_pos = 0, best_varpool_pos = 0;
+  int n_varpool_nodes = 0, varpool_pos = 0;
   struct cgraph_node **order = XNEWVEC (struct cgraph_node *, cgraph_max_uid);
   struct varpool_node **varpool_order = NULL;
   int i;
@@ -481,10 +497,13 @@ lto_balanced_map (void)
      get better about minimizing the function bounday, but until that
      things works smoother if we order in source order.  */
   qsort (order, n_nodes, sizeof (struct cgraph_node *), node_cmp);
+
+  if (cgraph_dump_file)
+    for(i = 0; i < n_nodes; i++)
+      fprintf (cgraph_dump_file, "Balanced map symbol order:%s:%u\n", cgraph_node_asm_name (order[i]), order[i]->tp_first_run);
+
   if (!flag_toplevel_reorder)
     {
-      qsort (order, n_nodes, sizeof (struct cgraph_node *), node_cmp);
-
       FOR_EACH_VARIABLE (vnode)
 	if (get_symbol_class (vnode) == SYMBOL_PARTITION)
 	  n_varpool_nodes++;
@@ -683,7 +702,6 @@ lto_balanced_map (void)
 	  best_i = i;
 	  best_n_nodes = lto_symtab_encoder_size (partition->encoder);
 	  best_total_size = total_size;
-	  best_varpool_pos = varpool_pos;
 	}
       if (cgraph_dump_file)
 	fprintf (cgraph_dump_file, "Step %i: added %s/%i, size %i, cost %i/%i "
@@ -701,7 +719,6 @@ lto_balanced_map (void)
 		fprintf (cgraph_dump_file, "Unwinding %i insertions to step %i\n",
 			 i - best_i, best_i);
 	      undo_partition (partition, best_n_nodes);
-	      varpool_pos = best_varpool_pos;
 	    }
 	  i = best_i;
  	  /* When we are finished, avoid creating empty partition.  */
@@ -849,7 +866,7 @@ may_need_named_section_p (lto_symtab_encoder_t encoder, symtab_node *node)
    of the same name in partition ENCODER (or in whole compilation unit if
    ENCODER is NULL) and if so, mangle the statics.  Always mangle all
    conflicting statics, so we reduce changes of silently miscompiling
-   asm statemnets referring to them by symbol name.  */
+   asm statements referring to them by symbol name.  */
 
 static void
 rename_statics (lto_symtab_encoder_t encoder, symtab_node *node)
diff --git a/gcc/lto/lto.c b/gcc/lto/lto.c
index 62856d0..2401c76 100644
--- a/gcc/lto/lto.c
+++ b/gcc/lto/lto.c
@@ -2453,9 +2453,12 @@ lto_wpa_write_files (void)
   /* Sort partitions by size so small ones are compiled last.
      FIXME: Even when not reordering we may want to output one list for parallel make
      and other for final link command.  */
-  ltrans_partitions.qsort (flag_toplevel_reorder
+
+  if (!flag_profile_reorder_functions || !flag_profile_use)
+    ltrans_partitions.qsort (flag_toplevel_reorder
 			   ? cmp_partitions_size
 			   : cmp_partitions_order);
+
   for (i = 0; i < n_sets; i++)
     {
       size_t len;
diff --git a/gcc/opts.c b/gcc/opts.c
index 3a939ac..d3a7735 100644
--- a/gcc/opts.c
+++ b/gcc/opts.c
@@ -1690,6 +1690,8 @@ common_handle_option (struct gcc_options *opts,
 	opts->x_flag_vect_cost_model = VECT_COST_MODEL_DYNAMIC;
       if (!opts_set->x_flag_tree_loop_distribute_patterns)
 	opts->x_flag_tree_loop_distribute_patterns = value;
+      if (!opts_set->x_flag_profile_reorder_functions)
+	opts->x_flag_profile_reorder_functions = value;
       /* Indirect call profiling should do all useful transformations
  	 speculative devirutalization does.  */
       if (!opts_set->x_flag_devirtualize_speculatively
@@ -1708,6 +1710,8 @@ common_handle_option (struct gcc_options *opts,
 	opts->x_flag_profile_values = value;
       if (!opts_set->x_flag_inline_functions)
 	opts->x_flag_inline_functions = value;
+      if (!opts_set->x_flag_profile_reorder_functions)
+	opts->x_flag_profile_reorder_functions = value;
       /* FIXME: Instrumentation we insert makes ipa-reference bitmaps
 	 quadratic.  Disable the pass until better memory representation
 	 is done.  */
diff --git a/gcc/varasm.c b/gcc/varasm.c
index 2226912..758367f 100644
--- a/gcc/varasm.c
+++ b/gcc/varasm.c
@@ -548,7 +548,14 @@ default_function_section (tree decl, enum node_frequency freq,
      unlikely executed (this happens especially with function splitting
      where we can split away unnecessary parts of static constructors.  */
   if (startup && freq != NODE_FREQUENCY_UNLIKELY_EXECUTED)
-    return get_named_text_section (decl, ".text.startup", NULL);
+    {
+      /* If we do have a profile or(and) LTO phase is executed, we do not need
+         these ELF section.  */
+      if (!in_lto_p || !flag_profile_values)
+        return get_named_text_section (decl, ".text.startup", NULL);
+      else
+        return NULL;
+    }
 
   /* Similarly for exit.  */
   if (exit && freq != NODE_FREQUENCY_UNLIKELY_EXECUTED)
@@ -560,7 +567,10 @@ default_function_section (tree decl, enum node_frequency freq,
       case NODE_FREQUENCY_UNLIKELY_EXECUTED:
 	return get_named_text_section (decl, ".text.unlikely", NULL);
       case NODE_FREQUENCY_HOT:
-	return get_named_text_section (decl, ".text.hot", NULL);
+        /* If we do have a profile or(and) LTO phase is executed, we do not need
+           these ELF section.  */
+        if (!in_lto_p || !flag_profile_values)
+          return get_named_text_section (decl, ".text.hot", NULL);
       default:
 	return NULL;
     }