Break LTO cgraph dump into more pieces.

Message ID 20180608145741.GA99335@kam.mff.cuni.cz
State New
Headers show
Series
  • Break LTO cgraph dump into more pieces.
Related show

Commit Message

Jan Hubicka June 8, 2018, 2:57 p.m.
Hi,
this patch splits LTO cgraph dump file into multiple files:
 1) cgraph which contains pretty much what cgraph dump has in normal compilation
 2) lto-link which contains symtab before symtab merging and decision of the
    LTO linker
 3) lto-decl-merge which dumps declaration merging
 4) lto-partition which contains partitioning decisions.

The main motivation is to make it easier to find relevant data.  Bit ugly is that
dumps are not quite chronological. Technically lto-link happens first, decl-merge next,
then all optimization passes are done and lto-partition is last.  I don't know how to
arrange it and I think i tis OK as it is.

Other bit non-standard thing is that I have added them as ipa dumps. It may also make
sense to declare them language dumps since LTO is techincally front-end but it seems
bit misleading to me. I think it makes more sense if they appear with -fdump-ipa-all.

Bootstrapped/regtested x86_64-linux, will commit later if there are no complains.

Honza

	* dumpfile.c (FIRST_ME_AUTO_NUMBERED_DUMP): Bump to 4.
	* lto-lang.c (lto_link_dump_id, decl_merge_dump_id, partition_dump_id):
	New global vars.
	(lto_register_dumps): New hook.
	(LANG_HOOKS_REGISTER_DUMPS): New.
	* lto-partition.c: Dump into dump_file instead of symtab->dump_file.
	* lto-symtab.c: Include lto.h; dump into dump_file instead of
	symtab->dump_file.
	(lto_symtab_merge_decls): Initialize dump file.
	* lto.c (read_cgraph_and_symbols): Initialize dump file.
	(do_whole_program_analysis): Likewise.

Patch

Index: dumpfile.c
===================================================================
--- dumpfile.c	(revision 261327)
+++ dumpfile.c	(working copy)
@@ -65,7 +65,7 @@ 
   DUMP_FILE_INFO (".gimple", "tree-gimple", DK_tree, 0),
   DUMP_FILE_INFO (".nested", "tree-nested", DK_tree, 0),
 #define FIRST_AUTO_NUMBERED_DUMP 1
-#define FIRST_ME_AUTO_NUMBERED_DUMP 3
+#define FIRST_ME_AUTO_NUMBERED_DUMP 4
 
   DUMP_FILE_INFO (NULL, "lang-all", DK_lang, 0),
   DUMP_FILE_INFO (NULL, "tree-all", DK_tree, 0),
Index: lto/lto-lang.c
===================================================================
--- lto/lto-lang.c	(revision 261327)
+++ lto/lto-lang.c	(working copy)
@@ -37,6 +37,9 @@ 
 #include "stringpool.h"
 #include "attribs.h"
 
+/* LTO specific dumps.  */
+int lto_link_dump_id, decl_merge_dump_id, partition_dump_id;
+
 static tree handle_noreturn_attribute (tree *, tree, tree, int, bool *);
 static tree handle_leaf_attribute (tree *, tree, tree, int, bool *);
 static tree handle_const_attribute (tree *, tree, tree, int, bool *);
@@ -1375,6 +1378,23 @@ 
   return true;
 }
 
+/* Register c++-specific dumps.  */
+
+void
+lto_register_dumps (gcc::dump_manager *dumps)
+{
+  lto_link_dump_id = dumps->dump_register
+    (".lto-link", "ipa-lto-link", "ipa-lto-link",
+     DK_ipa, OPTGROUP_NONE, false);
+  decl_merge_dump_id = dumps->dump_register
+    (".lto-decl-merge", "ipa-lto-decl-merge", "ipa-lto-decl-merge",
+     DK_ipa, OPTGROUP_NONE, false);
+  partition_dump_id = dumps->dump_register
+    (".lto-partition", "ipa-lto-partition", "ipa-lto-partition",
+     DK_ipa, OPTGROUP_NONE, false);
+}
+
+
 /* Initialize tree structures required by the LTO front end.  */
 
 static void lto_init_ts (void)
@@ -1390,6 +1410,8 @@ 
 #define LANG_HOOKS_COMPLAIN_WRONG_LANG_P lto_complain_wrong_lang_p
 #undef LANG_HOOKS_INIT_OPTIONS_STRUCT
 #define LANG_HOOKS_INIT_OPTIONS_STRUCT lto_init_options_struct
+#undef LANG_HOOKS_REGISTER_DUMPS
+#define LANG_HOOKS_REGISTER_DUMPS lto_register_dumps
 #undef LANG_HOOKS_HANDLE_OPTION
 #define LANG_HOOKS_HANDLE_OPTION lto_handle_option
 #undef LANG_HOOKS_POST_OPTIONS
Index: lto/lto-partition.c
===================================================================
--- lto/lto-partition.c	(revision 261327)
+++ lto/lto-partition.c	(working copy)
@@ -160,8 +160,8 @@ 
   if (symbol_partitioned_p (node))
     {
       node->in_other_partition = 1;
-      if (symtab->dump_file)
-	fprintf (symtab->dump_file,
+      if (dump_file)
+	fprintf (dump_file,
 		 "Symbol node %s now used in multiple partitions\n",
 		 node->name ());
     }
@@ -541,13 +541,13 @@ 
   order.qsort (node_cmp);
   noreorder.qsort (node_cmp);
 
-  if (symtab->dump_file)
+  if (dump_file)
     {
       for (unsigned i = 0; i < order.length (); i++)
-	fprintf (symtab->dump_file, "Balanced map symbol order:%s:%u\n",
+	fprintf (dump_file, "Balanced map symbol order:%s:%u\n",
 		 order[i]->name (), order[i]->tp_first_run);
       for (unsigned i = 0; i < noreorder.length (); i++)
-	fprintf (symtab->dump_file, "Balanced map symbol no_reorder:%s:%u\n",
+	fprintf (dump_file, "Balanced map symbol no_reorder:%s:%u\n",
 		 noreorder[i]->name (), noreorder[i]->tp_first_run);
     }
 
@@ -569,8 +569,8 @@ 
     partition_size = PARAM_VALUE (MIN_PARTITION_SIZE);
   npartitions = 1;
   partition = new_partition ("");
-  if (symtab->dump_file)
-    fprintf (symtab->dump_file, "Total unit size: %" PRId64 ", partition size: %" PRId64 "\n",
+  if (dump_file)
+    fprintf (dump_file, "Total unit size: %" PRId64 ", partition size: %" PRId64 "\n",
 	     total_size, partition_size);
 
   auto_vec<symtab_node *> next_nodes;
@@ -763,8 +763,8 @@ 
 	  best_n_nodes = lto_symtab_encoder_size (partition->encoder);
 	  best_varpool_pos = varpool_pos;
 	}
-      if (symtab->dump_file)
-	fprintf (symtab->dump_file, "Step %i: added %s/%i, size %i, "
+      if (dump_file)
+	fprintf (dump_file, "Step %i: added %s/%i, size %i, "
 		 "cost %" PRId64 "/%" PRId64 " "
 		 "best %" PRId64 "/%" PRId64", step %i\n", i,
 		 order[i]->name (), order[i]->order,
@@ -777,8 +777,8 @@ 
 	{
 	  if (best_i != i)
 	    {
-	      if (symtab->dump_file)
-		fprintf (symtab->dump_file, "Unwinding %i insertions to step %i\n",
+	      if (dump_file)
+		fprintf (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;
@@ -785,8 +785,8 @@ 
 	    }
 	  gcc_assert (best_size == partition->insns);
 	  i = best_i;
-	  if (symtab->dump_file)
-	    fprintf (symtab->dump_file,
+	  if (dump_file)
+	    fprintf (dump_file,
 		     "Partition insns: %i (want %" PRId64 ")\n",
 		     partition->insns, partition_size);
  	  /* When we are finished, avoid creating empty partition.  */
@@ -799,8 +799,8 @@ 
 	  last_visited_node = 0;
 	  cost = 0;
 
-	  if (symtab->dump_file)
-	    fprintf (symtab->dump_file, "New partition\n");
+	  if (dump_file)
+	    fprintf (dump_file, "New partition\n");
 	  best_n_nodes = 0;
 	  best_cost = -1;
 
@@ -812,8 +812,8 @@ 
 	    /* Watch for overflow.  */
 	    partition_size = INT_MAX / 16;
 
-	  if (symtab->dump_file)
-	    fprintf (symtab->dump_file,
+	  if (dump_file)
+	    fprintf (dump_file,
 		     "Total size: %" PRId64 " partition_size: %" PRId64 "\n",
 		     total_size, partition_size);
 	  if (partition_size < PARAM_VALUE (MIN_PARTITION_SIZE))
@@ -840,21 +840,21 @@ 
   gcc_assert (next_nodes.length () || npartitions != 1 || !best_cost || best_cost == -1);
   add_sorted_nodes (next_nodes, partition);
 
-  if (symtab->dump_file)
+  if (dump_file)
     {
-      fprintf (symtab->dump_file, "\nPartition sizes:\n");
+      fprintf (dump_file, "\nPartition sizes:\n");
       unsigned partitions = ltrans_partitions.length ();
 
       for (unsigned i = 0; i < partitions ; i++)
 	{
 	  ltrans_partition p = ltrans_partitions[i];
-	  fprintf (symtab->dump_file, "partition %d contains %d (%2.2f%%)"
+	  fprintf (dump_file, "partition %d contains %d (%2.2f%%)"
 		   " symbols and %d (%2.2f%%) insns\n", i, p->symbols,
 		   100.0 * p->symbols / order.length (), p->insns,
 		   100.0 * p->insns / original_total_size);
 	}
 
-      fprintf (symtab->dump_file, "\n");
+      fprintf (dump_file, "\n");
     }
 }
 
@@ -869,8 +869,8 @@ 
   if (node->lto_file_data
       && lto_get_decl_name_mapping (node->lto_file_data, name) != name)
     {
-      if (symtab->dump_file)
-	fprintf (symtab->dump_file,
+      if (dump_file)
+	fprintf (dump_file,
 		 "Not privatizing symbol name: %s. It privatized already.\n",
 		 name);
       return true;
@@ -881,8 +881,8 @@ 
      that are not really clones.  */
   if (node->unique_name)
     {
-      if (symtab->dump_file)
-	fprintf (symtab->dump_file,
+      if (dump_file)
+	fprintf (dump_file,
 		 "Not privatizing symbol name: %s. Has unique name.\n",
 		 name);
       return true;
@@ -972,8 +972,8 @@ 
 			     IDENTIFIER_POINTER
 			     (DECL_ASSEMBLER_NAME (decl)));
 
-  if (symtab->dump_file)
-    fprintf (symtab->dump_file,
+  if (dump_file)
+    fprintf (dump_file,
 	     "Privatizing symbol name: %s -> %s\n",
 	     name, IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl)));
 
@@ -1018,8 +1018,8 @@ 
   TREE_PUBLIC (node->decl) = 1;
   DECL_VISIBILITY (node->decl) = VISIBILITY_HIDDEN;
   DECL_VISIBILITY_SPECIFIED (node->decl) = true;
-  if (symtab->dump_file)
-    fprintf (symtab->dump_file,
+  if (dump_file)
+    fprintf (dump_file,
 	     "Promoting as hidden: %s (%s)\n", node->name (),
 	     IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (node->decl)));
 
@@ -1035,8 +1035,8 @@ 
 	  TREE_PUBLIC (alias->decl) = 1;
 	  DECL_VISIBILITY (alias->decl) = VISIBILITY_HIDDEN;
 	  DECL_VISIBILITY_SPECIFIED (alias->decl) = true;
-	  if (symtab->dump_file)
-	    fprintf (symtab->dump_file,
+	  if (dump_file)
+	    fprintf (dump_file,
 		     "Promoting alias as hidden: %s\n",
 		     IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (node->decl)));
 	}
@@ -1102,8 +1102,8 @@ 
   if (!s)
     return;
 
-  if (symtab->dump_file)
-    fprintf (symtab->dump_file,
+  if (dump_file)
+    fprintf (dump_file,
 	    "Renaming statics with asm name: %s\n", node->name ());
 
   /* Assign every symbol in the set that shares the same ASM name an unique
Index: lto/lto-symtab.c
===================================================================
--- lto/lto-symtab.c	(revision 261327)
+++ lto/lto-symtab.c	(working copy)
@@ -31,6 +31,7 @@ 
 #include "ipa-utils.h"
 #include "builtins.h"
 #include "alias.h"
+#include "lto.h"
 #include "lto-symtab.h"
 #include "stringpool.h"
 #include "attribs.h"
@@ -45,9 +46,9 @@ 
   struct cgraph_edge *e, *next;
   bool compatible_p;
 
-  if (symtab->dump_file)
+  if (dump_file)
     {
-      fprintf (symtab->dump_file, "Replacing cgraph node %s by %s"
+      fprintf (dump_file, "Replacing cgraph node %s by %s"
  	       " for symbol %s\n",
 	       node->dump_name (),
 	       prevailing_node->dump_name (),
@@ -536,8 +537,8 @@ 
 {
   if (TREE_CODE (prevailing) != TREE_CODE (decl))
     {
-      if (symtab->dump_file)
-	fprintf (symtab->dump_file, "Not merging decls; "
+      if (dump_file)
+	fprintf (dump_file, "Not merging decls; "
 		 "TREE_CODE mismatch\n");
       return false;
     }
@@ -547,8 +548,8 @@ 
     {
       if (DECL_BUILT_IN (prevailing) != DECL_BUILT_IN (decl))
 	{
-          if (symtab->dump_file)
-	    fprintf (symtab->dump_file, "Not merging decls; "
+	  if (dump_file)
+	    fprintf (dump_file, "Not merging decls; "
 		     "DECL_BUILT_IN mismatch\n");
 	  return false;
 	}
@@ -556,8 +557,8 @@ 
 	  && (DECL_BUILT_IN_CLASS (prevailing) != DECL_BUILT_IN_CLASS (decl)
 	      || DECL_FUNCTION_CODE (prevailing) != DECL_FUNCTION_CODE (decl)))
 	{
-          if (symtab->dump_file)
-	    fprintf (symtab->dump_file, "Not merging decls; "
+	  if (dump_file)
+	    fprintf (dump_file, "Not merging decls; "
 		     "DECL_BUILT_IN_CLASS or CODE mismatch\n");
 	  return false;
 	}
@@ -572,8 +573,8 @@ 
       if ((prev_attr == NULL) != (attr == NULL)
 	  || (prev_attr && !attribute_value_equal (prev_attr, attr)))
 	{
-          if (symtab->dump_file)
-	    fprintf (symtab->dump_file, "Not merging decls; "
+          if (dump_file)
+	    fprintf (dump_file, "Not merging decls; "
 		     "error attribute mismatch\n");
 	  return false;
 	}
@@ -583,8 +584,8 @@ 
       if ((prev_attr == NULL) != (attr == NULL)
 	  || (prev_attr && !attribute_value_equal (prev_attr, attr)))
 	{
-          if (symtab->dump_file)
-	    fprintf (symtab->dump_file, "Not merging decls; "
+          if (dump_file)
+	    fprintf (dump_file, "Not merging decls; "
 		     "warning attribute mismatch\n");
 	  return false;
 	}
@@ -593,8 +594,8 @@ 
       attr = lookup_attribute ("noreturn", DECL_ATTRIBUTES (decl));
       if ((prev_attr == NULL) != (attr == NULL))
 	{
-          if (symtab->dump_file)
-	    fprintf (symtab->dump_file, "Not merging decls; "
+          if (dump_file)
+	    fprintf (dump_file, "Not merging decls; "
 		     "noreturn attribute mismatch\n");
 	  return false;
 	}
@@ -753,13 +754,13 @@ 
   symtab_node *prevailing;
   bool diagnosed_p = false;
 
-  if (symtab->dump_file)
+  if (dump_file)
     {
-      fprintf (symtab->dump_file, "Merging nodes for %s. Candidates:\n",
+      fprintf (dump_file, "Merging nodes for %s. Candidates:\n",
 	       first->asm_name ());
       for (e = first; e; e = e->next_sharing_asm_name)
 	if (TREE_PUBLIC (e->decl))
-	  e->dump (symtab->dump_file);
+	  e->dump (dump_file);
     }
 
   /* Compute the symbol resolutions.  This is a no-op when using the
@@ -849,11 +850,11 @@ 
      mismatches.  */
   lto_symtab_merge_decls_2 (prevailing, diagnosed_p);
 
-  if (symtab->dump_file)
+  if (dump_file)
     {
-      fprintf (symtab->dump_file, "After resolution:\n");
+      fprintf (dump_file, "After resolution:\n");
       for (e = prevailing; e; e = e->next_sharing_asm_name)
-	e->dump (symtab->dump_file);
+	e->dump (dump_file);
     }
 }
 
@@ -864,6 +865,9 @@ 
 {
   symtab_node *node;
 
+  gcc_assert (!dump_file);
+  dump_file = dump_begin (decl_merge_dump_id, NULL);
+
   /* Populate assembler name hash.   */
   symtab->symtab_initialize_asm_name_hash ();
 
@@ -871,6 +875,10 @@ 
     if (!node->previous_sharing_asm_name
 	&& node->next_sharing_asm_name)
       lto_symtab_merge_decls_1 (node);
+
+  if (dump_file)
+    dump_end (decl_merge_dump_id, dump_file);
+  dump_file = NULL;
 }
 
 /* Helper to process the decl chain for the symbol table entry *SLOT.  */
Index: lto/lto.c
===================================================================
--- lto/lto.c	(revision 261327)
+++ lto/lto.c	(working copy)
@@ -2972,30 +2972,44 @@ 
       all_file_decl_data[i]->current_decl_state = NULL; 
     }
 
-  /* Finally merge the cgraph according to the decl merging decisions.  */
-  timevar_push (TV_IPA_LTO_CGRAPH_MERGE);
-  if (symtab->dump_file)
-    {
-      fprintf (symtab->dump_file, "Before merging:\n");
-      symtab->dump (symtab->dump_file);
-    }
   if (!flag_ltrans)
     {
+      /* Finally merge the cgraph according to the decl merging decisions.  */
+      timevar_push (TV_IPA_LTO_CGRAPH_MERGE);
+
+      gcc_assert (!dump_file);
+      dump_file = dump_begin (lto_link_dump_id, NULL);
+
+      if (dump_file)
+	{
+	  fprintf (dump_file, "Before merging:\n");
+	  symtab->dump (dump_file);
+	}
       lto_symtab_merge_symbols ();
       /* Removal of unreachable symbols is needed to make verify_symtab to pass;
 	 we are still having duplicated comdat groups containing local statics.
 	 We could also just remove them while merging.  */
       symtab->remove_unreachable_nodes (dump_file);
+      ggc_collect ();
+
+      if (dump_file)
+        dump_end (lto_link_dump_id, dump_file);
+      dump_file = NULL;
+      timevar_pop (TV_IPA_LTO_CGRAPH_MERGE);
     }
-  ggc_collect ();
   symtab->state = IPA_SSA;
-  /* FIXME: Technically all node removals happening here are useless, because
-     WPA should not stream them.  */
+  /* All node removals happening here are useless, because
+     WPA should not stream them. Still always perform remove_unreachable_nodes
+     because we may reshape clone tree, get rid of dead masters of inline
+     clones and remove symbol entries for read-only variables we keep around
+     only to be able to constant fold them.  */
   if (flag_ltrans)
-    symtab->remove_unreachable_nodes (dump_file);
+    {
+      if (symtab->dump_file)
+	 symtab->dump (symtab->dump_file);
+      symtab->remove_unreachable_nodes (symtab->dump_file);
+    }
 
-  timevar_pop (TV_IPA_LTO_CGRAPH_MERGE);
-
   /* Indicate that the cgraph is built and ready.  */
   symtab->function_flags_ready = true;
 
@@ -3152,19 +3166,19 @@ 
   if (seen_error ())
     return;
 
-  if (symtab->dump_file)
-    {
-      fprintf (symtab->dump_file, "Optimized ");
-      symtab->dump (symtab->dump_file);
-    }
-
-  symtab_node::checking_verify_symtab_nodes ();
-  bitmap_obstack_release (NULL);
-
   /* We are about to launch the final LTRANS phase, stop the WPA timer.  */
   timevar_pop (TV_WHOPR_WPA);
 
   timevar_push (TV_WHOPR_PARTITIONING);
+
+  gcc_assert (!dump_file);
+  dump_file = dump_begin (partition_dump_id, NULL);
+
+  if (dump_file)
+    symtab->dump (dump_file);
+
+  symtab_node::checking_verify_symtab_nodes ();
+  bitmap_obstack_release (NULL);
   if (flag_lto_partition == LTO_PARTITION_1TO1)
     lto_1_to_1_map ();
   else if (flag_lto_partition == LTO_PARTITION_MAX)
@@ -3192,6 +3206,9 @@ 
      to globals with hidden visibility because they are accessed from multiple
      partitions.  */
   lto_promote_cross_file_statics ();
+  if (dump_file)
+     dump_end (partition_dump_id, dump_file);
+  dump_file = NULL;
   timevar_pop (TV_WHOPR_PARTITIONING);
 
   timevar_stop (TV_PHASE_OPT_GEN);
Index: lto/lto.h
===================================================================
--- lto/lto.h	(revision 261327)
+++ lto/lto.h	(working copy)
@@ -51,6 +51,8 @@ 
 extern lto_file *lto_set_current_out_file (lto_file *file);
 extern lto_file *lto_get_current_out_file (void);
 
+extern int lto_link_dump_id, decl_merge_dump_id, partition_dump_id;
+
 /* Hash table entry to hold the start offset and length of an LTO
    section in a .o file.  */
 struct lto_section_slot