re PR rtl-optimization/38583 (huge test case makes register allocator run out of…

re PR rtl-optimization/38583 (huge test case makes register allocator run out of memory while constructing the conflict graph)

2009-01-05  Vladimir Makarov  <vmakarov@redhat.com>

	PR rtl-optimization/38583
	* params.h (IRA_MAX_CONFLICT_TABLE_SIZE): New macro.

	* params.def (ira-max-conflict-table-size): New.

	* doc/invoke.texi (ira-max-conflict-table-size): Decribe.
	
	* ira.h (ira_conflicts_p): New external definition.
	
	* ira-conflicts.c (build_conflict_bit_table): Do not build too big
	table.  Report this.  Return result of building.
	(ira_build_conflicts): Use ira_conflicts_p.  Check result of
	building conflict table.

	* ira-color.c (fast_allocation): Use num instead of
	ira_allocnos_num.
	(ira_color): Use ira_conflicts_p.
	
	* global.c: Include ira.h.
	(pseudo_for_reload_consideration_p, build_insn_chain): Use
	ira_conflicts_p.

	* Makefile.in (global.o): Add ira.h.
	
	* ira-build.c (mark_all_loops_for_removal,
	propagate_some_info_from_allocno): New.
	(remove_unnecessary_allocnos): Call
	propagate_some_info_from_allocno.
	(remove_low_level_allocnos): New.
	(remove_unnecessary_regions): Add parameter.  Call
	mark_all_loops_for_removal and remove_low_level_allocnos.  Pass
	parameter to remove_unnecessary_regions.
	(ira_build): Remove all regions but root if the conflict table was
	not built.  Update conflict hard regs for allocnos crossing calls.

	* ira.c (ira_conflicts_p): New global.
	(ira): Define and use ira_conflicts_p.

	* reload1.c (compute_use_by_pseudos, reload, count_pseudo,
	count_spilled_pseudo, find_reg, alter_reg, finish_spills,
	emit_input_reload_insns, delete_output_reload): Use
	ira_conflicts_p.

From-SVN: r143112

gcc/ChangeLog

+2009-01-05  Vladimir Makarov  <vmakarov@redhat.com>
+
+	PR rtl-optimization/38583
+	* params.h (IRA_MAX_CONFLICT_TABLE_SIZE): New macro.
+
+	* params.def (ira-max-conflict-table-size): New.
+
+	* doc/invoke.texi (ira-max-conflict-table-size): Decribe.
+	
+	* ira.h (ira_conflicts_p): New external definition.
+	
+	* ira-conflicts.c (build_conflict_bit_table): Do not build too big
+	table.  Report this.  Return result of building.
+	(ira_build_conflicts): Use ira_conflicts_p.  Check result of
+	building conflict table.
+
+	* ira-color.c (fast_allocation): Use num instead of
+	ira_allocnos_num.
+	(ira_color): Use ira_conflicts_p.
+	
+	* global.c: Include ira.h.
+	(pseudo_for_reload_consideration_p, build_insn_chain): Use
+	ira_conflicts_p.
+
+	* Makefile.in (global.o): Add ira.h.
+	
+	* ira-build.c (mark_all_loops_for_removal,
+	propagate_some_info_from_allocno): New.
+	(remove_unnecessary_allocnos): Call
+	propagate_some_info_from_allocno.
+	(remove_low_level_allocnos): New.
+	(remove_unnecessary_regions): Add parameter.  Call
+	mark_all_loops_for_removal and remove_low_level_allocnos.  Pass
+	parameter to remove_unnecessary_regions.
+	(ira_build): Remove all regions but root if the conflict table was
+	not built.  Update conflict hard regs for allocnos crossing calls.
+
+	* ira.c (ira_conflicts_p): New global.
+	(ira): Define and use ira_conflicts_p.
+
+	* reload1.c (compute_use_by_pseudos, reload, count_pseudo,
+	count_spilled_pseudo, find_reg, alter_reg, finish_spills,
+	emit_input_reload_insns, delete_output_reload): Use
+	ira_conflicts_p.
+	
 2009-01-06  Ben Elliston  <bje@au.ibm.com>
 
 	* gengtype-lex.l (YY_NO_INPUT): Define.

gcc/Makefile.in

@@ -2860,7 +2860,7 @@ bitmap.o : bitmap.c $(CONFIG_H) $(SYSTEM_H)  coretypes.h $(TM_H) $(RTL_H) \
 global.o : global.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) $(RTL_H) \
    $(FLAGS_H) reload.h $(FUNCTION_H) $(RECOG_H) $(REGS_H) hard-reg-set.h \
    insn-config.h output.h $(TOPLEV_H) $(TM_P_H) $(MACHMODE_H) tree-pass.h \
-   $(TIMEVAR_H) vecprim.h $(DF_H) $(DBGCNT_H) $(RA_H)
+   $(TIMEVAR_H) vecprim.h $(DF_H) $(DBGCNT_H) $(RA_H) ira.h
 ra-conflict.o : ra-conflict.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) $(RTL_H) \
    $(FLAGS_H) reload.h $(FUNCTION_H) $(RECOG_H) $(REGS_H) hard-reg-set.h \
    insn-config.h output.h $(TOPLEV_H) $(TM_P_H) $(MACHMODE_H) tree-pass.h \

gcc/doc/invoke.texi

@@ -7617,6 +7617,15 @@ given number of the most frequently executed loops will form regions
 for the regional register allocation.  The default value of the
 parameter is 100.
 
+@item ira-max-conflict-table-size
+Although IRA uses a sophisticated algorithm of compression conflict
+table, the table can be still big for huge functions.  If the conflict
+table for a function could be more than size in MB given by the
+parameter, the conflict table is not built and faster, simpler, and
+lower quality register allocation algorithm will be used.  The
+algorithm do not use pseudo-register conflicts.  The default value of
+the parameter is 2000.
+
 @end table
 @end table
 

gcc/global.c

@@ -42,6 +42,7 @@ along with GCC; see the file COPYING3.  If not see
 #include "vecprim.h"
 #include "dbgcnt.h"
 #include "ra.h"
+#include "ira.h"
 
 /* This pass of the compiler performs global register allocation.
    It assigns hard register numbers to all the pseudo registers
@@ -1394,7 +1395,7 @@ pseudo_for_reload_consideration_p (int regno)
   /* Consider spilled pseudos too for IRA because they still have a
      chance to get hard-registers in the reload when IRA is used.  */
   return (reg_renumber[regno] >= 0
-	  || (flag_ira && optimize && flag_ira_share_spill_slots));
+	  || (flag_ira && ira_conflicts_p && flag_ira_share_spill_slots));
 }
 
 /* Walk the insns of the current function and build reload_insn_chain,
@@ -1483,7 +1484,7 @@ build_insn_chain (void)
 
 		    if ((regno < FIRST_PSEUDO_REGISTER
 			 || reg_renumber[regno] >= 0
-			 || (flag_ira && optimize))
+			 || (flag_ira && ira_conflicts_p))
 			&& (!DF_REF_FLAGS_IS_SET (def, DF_REF_CONDITIONAL)))
 		      {
 			rtx reg = DF_REF_REG (def);

gcc/ira-color.c

@@ -3267,7 +3267,7 @@ fast_allocation (void)
 						  * ira_max_point);
   for (i = 0; i < ira_max_point; i++)
     CLEAR_HARD_REG_SET (used_hard_regs[i]);
-  qsort (sorted_allocnos, ira_allocnos_num, sizeof (ira_allocno_t), 
+  qsort (sorted_allocnos, num, sizeof (ira_allocno_t),
 	 allocno_priority_compare_func);
   for (i = 0; i < num; i++)
     {
@@ -3329,7 +3329,7 @@ ira_color (void)
       ALLOCNO_UPDATED_MEMORY_COST (a) = ALLOCNO_MEMORY_COST (a);
       ALLOCNO_UPDATED_COVER_CLASS_COST (a) = ALLOCNO_COVER_CLASS_COST (a);
     }
-  if (optimize)
+  if (ira_conflicts_p)
     color ();
   else
     fast_allocation ();

gcc/ira-conflicts.c

@@ -60,8 +60,10 @@ static IRA_INT_TYPE **conflicts;
 
 
 
-/* Build allocno conflict table by processing allocno live ranges.  */
-static void
+/* Build allocno conflict table by processing allocno live ranges.
+   Return true if the table was built.  The table is not built if it
+   is too big.  */
+static bool
 build_conflict_bit_table (void)
 {
   int i, num, id, allocated_words_num, conflict_bit_vec_words_num;
@@ -74,6 +76,26 @@ build_conflict_bit_table (void)
   int allocno_set_words;
 
   allocno_set_words = (ira_allocnos_num + IRA_INT_BITS - 1) / IRA_INT_BITS;
+  allocated_words_num = 0;
+  FOR_EACH_ALLOCNO (allocno, ai)
+    {
+      if (ALLOCNO_MAX (allocno) < ALLOCNO_MIN (allocno))
+	  continue;
+      conflict_bit_vec_words_num
+	= ((ALLOCNO_MAX (allocno) - ALLOCNO_MIN (allocno) + IRA_INT_BITS)
+	   / IRA_INT_BITS);
+      allocated_words_num += conflict_bit_vec_words_num;
+      if ((unsigned long long) allocated_words_num * sizeof (IRA_INT_TYPE)
+	  > (unsigned long long) IRA_MAX_CONFLICT_TABLE_SIZE * 1024 * 1024)
+	{
+	  if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL)
+	    fprintf
+	      (ira_dump_file,
+	       "+++Conflict table will be too big(>%dMB) -- don't use it\n",
+	       IRA_MAX_CONFLICT_TABLE_SIZE);
+	  return false;
+	}
+    }
   allocnos_live = sparseset_alloc (ira_allocnos_num);
   conflicts = (IRA_INT_TYPE **) ira_allocate (sizeof (IRA_INT_TYPE *)
 					      * ira_allocnos_num);
@@ -134,6 +156,7 @@ build_conflict_bit_table (void)
 	sparseset_clear_bit (allocnos_live, ALLOCNO_NUM (r->allocno));
     }
   sparseset_free (allocnos_live);
+  return true;
 }
 
 
@@ -743,29 +766,34 @@ ira_build_conflicts (void)
   ira_allocno_iterator ai;
   HARD_REG_SET temp_hard_reg_set;
 
-  if (optimize)
+  if (ira_conflicts_p)
     {
-      build_conflict_bit_table ();
-      build_conflicts ();
-      ira_traverse_loop_tree (true, ira_loop_tree_root, NULL, add_copies);
-      /* We need finished conflict table for the subsequent call.  */
-      if (flag_ira_region == IRA_REGION_ALL
-	  || flag_ira_region == IRA_REGION_MIXED)
-	propagate_copies ();
-      /* Now we can free memory for the conflict table (see function
-	 build_allocno_conflicts for details).  */
-      FOR_EACH_ALLOCNO (a, ai)
+      ira_conflicts_p = build_conflict_bit_table ();
+      if (ira_conflicts_p)
 	{
-	  if (ALLOCNO_CONFLICT_ALLOCNO_ARRAY (a) != conflicts[ALLOCNO_NUM (a)])
-	    ira_free (conflicts[ALLOCNO_NUM (a)]);
+	  build_conflicts ();
+	  ira_traverse_loop_tree (true, ira_loop_tree_root, NULL, add_copies);
+	  /* We need finished conflict table for the subsequent call.  */
+	  if (flag_ira_region == IRA_REGION_ALL
+	      || flag_ira_region == IRA_REGION_MIXED)
+	    propagate_copies ();
+	  /* Now we can free memory for the conflict table (see function
+	     build_allocno_conflicts for details).  */
+	  FOR_EACH_ALLOCNO (a, ai)
+	    {
+	      if (ALLOCNO_CONFLICT_ALLOCNO_ARRAY (a)
+		  != conflicts[ALLOCNO_NUM (a)])
+		ira_free (conflicts[ALLOCNO_NUM (a)]);
+	    }
+	  ira_free (conflicts);
 	}
-      ira_free (conflicts);
     }
   if (! CLASS_LIKELY_SPILLED_P (BASE_REG_CLASS))
     CLEAR_HARD_REG_SET (temp_hard_reg_set);
   else
     {
-      COPY_HARD_REG_SET (temp_hard_reg_set, reg_class_contents[BASE_REG_CLASS]);
+      COPY_HARD_REG_SET (temp_hard_reg_set,
+			 reg_class_contents[BASE_REG_CLASS]);
       AND_COMPL_HARD_REG_SET (temp_hard_reg_set, ira_no_alloc_regs);
       AND_HARD_REG_SET (temp_hard_reg_set, call_used_reg_set);
     }
@@ -796,6 +824,7 @@ ira_build_conflicts (void)
 	    }
 	}
     }
-  if (optimize && internal_flag_ira_verbose > 2 && ira_dump_file != NULL)
+  if (optimize && ira_conflicts_p
+      && internal_flag_ira_verbose > 2 && ira_dump_file != NULL)
     print_conflicts (ira_dump_file, false);
 }

gcc/ira.c

@@ -1827,6 +1827,10 @@ too_high_register_pressure_p (void)
 /* All natural loops.  */
 struct loops ira_loops;
 
+/* True if we have allocno conflicts.  It is false for non-optimized
+   mode or when the conflict table is too big.  */
+bool ira_conflicts_p;
+
 /* This is the main entry of IRA.  */
 static void
 ira (FILE *f)
@@ -1851,6 +1855,7 @@ ira (FILE *f)
       ira_dump_file = stderr;
     }
 
+  ira_conflicts_p = optimize > 0;
   setup_prohibited_mode_move_regs ();
 
   df_note_add_problem ();
@@ -1916,6 +1921,8 @@ ira (FILE *f)
   loops_p = ira_build (optimize
 		       && (flag_ira_region == IRA_REGION_ALL
 			   || flag_ira_region == IRA_REGION_MIXED));
+  
+  ira_assert (ira_conflicts_p || !loops_p);
 
   saved_flag_ira_share_spill_slots = flag_ira_share_spill_slots;
   if (too_high_register_pressure_p ())
@@ -1929,7 +1936,7 @@ ira (FILE *f)
       
   ira_emit (loops_p);
   
-  if (optimize)
+  if (ira_conflicts_p)
     {
       max_regno = max_reg_num ();
       
@@ -1963,7 +1970,7 @@ ira (FILE *f)
   calculate_allocation_cost ();
   
 #ifdef ENABLE_IRA_CHECKING
-  if (optimize)
+  if (ira_conflicts_p)
     check_allocation ();
 #endif
       
@@ -1992,7 +1999,7 @@ ira (FILE *f)
   allocate_initial_values (reg_equiv_memory_loc);
 
   overall_cost_before = ira_overall_cost;
-  if (optimize)
+  if (ira_conflicts_p)
     {
       fix_reg_equiv_init ();
       
@@ -2015,13 +2022,13 @@ ira (FILE *f)
   df_set_flags (DF_NO_INSN_RESCAN);
   build_insn_chain ();
 
-  reload_completed = !reload (get_insns (), optimize > 0);
+  reload_completed = !reload (get_insns (), ira_conflicts_p);
 
   timevar_pop (TV_RELOAD);
 
   timevar_push (TV_IRA);
 
-  if (optimize)
+  if (ira_conflicts_p)
     {
       ira_free (ira_spilled_reg_stack_slots);
       

gcc/ira.h

@@ -20,6 +20,10 @@ You should have received a copy of the GNU General Public License
 along with GCC; see the file COPYING3.  If not see
 <http://www.gnu.org/licenses/>.  */
 
+/* True if we have allocno conflicts.  It is false for non-optimized
+   mode or when the conflict table is too big.  */
+extern bool ira_conflicts_p;
+
 extern void ira_init_once (void);
 extern void ira_init (void);
 extern void ira_finish_once (void);

gcc/params.def

@@ -750,6 +750,11 @@ DEFPARAM (PARAM_IRA_MAX_LOOPS_NUM,
 	  "max loops number for regional RA",
 	  100, 0, 0)
 
+DEFPARAM (PARAM_IRA_MAX_CONFLICT_TABLE_SIZE,
+	  "ira-max-conflict-table-size",
+	  "max size of conflict table in MB",
+	  2000, 0, 0)
+
 /* Switch initialization conversion will refuse to create arrays that are
    bigger than this parameter times the number of switch branches.  */
 

gcc/params.h

@@ -169,6 +169,8 @@ typedef enum compiler_param
   PARAM_VALUE (PARAM_USE_CANONICAL_TYPES)
 #define IRA_MAX_LOOPS_NUM \
   PARAM_VALUE (PARAM_IRA_MAX_LOOPS_NUM)
+#define IRA_MAX_CONFLICT_TABLE_SIZE \
+  PARAM_VALUE (PARAM_IRA_MAX_CONFLICT_TABLE_SIZE)
 #define SWITCH_CONVERSION_BRANCH_RATIO \
   PARAM_VALUE (PARAM_SWITCH_CONVERSION_BRANCH_RATIO)
 #endif /* ! GCC_PARAMS_H */

gcc/reload1.c

@@ -557,7 +557,7 @@ compute_use_by_pseudos (HARD_REG_SET *to, regset from)
 	     which might still contain registers that have not
 	     actually been allocated since they have an
 	     equivalence.  */
-	  gcc_assert ((flag_ira && optimize) || reload_completed);
+	  gcc_assert ((flag_ira && ira_conflicts_p) || reload_completed);
 	}
       else
 	add_to_hard_reg_set (to, PSEUDO_REGNO_MODE (regno), r);
@@ -901,7 +901,7 @@ reload (rtx first, int global)
   for (n = 0, i = LAST_VIRTUAL_REGISTER + 1; i < max_regno; i++)
     temp_pseudo_reg_arr[n++] = i;
   
-  if (flag_ira && optimize)
+  if (flag_ira && ira_conflicts_p)
     /* Ask IRA to order pseudo-registers for better stack slot
        sharing.  */
     ira_sort_regnos_for_alter_reg (temp_pseudo_reg_arr, n, reg_max_ref_width);
@@ -1055,7 +1055,7 @@ reload (rtx first, int global)
 
       calculate_needs_all_insns (global);
 
-      if (! flag_ira || ! optimize)
+      if (! flag_ira || ! ira_conflicts_p)
 	/* Don't do it for IRA.  We need this info because we don't
 	   change live_throughout and dead_or_set for chains when IRA
 	   is used.  */
@@ -1614,7 +1614,7 @@ calculate_needs_all_insns (int global)
 				       reg_equiv_memory_loc
 				       [REGNO (SET_DEST (set))]))))
 		{
-		  if (flag_ira && optimize)
+		  if (flag_ira && ira_conflicts_p)
 		    /* Inform IRA about the insn deletion.  */
 		    ira_mark_memory_move_deletion (REGNO (SET_DEST (set)),
 						   REGNO (SET_SRC (set)));
@@ -1723,7 +1723,7 @@ count_pseudo (int reg)
       || REGNO_REG_SET_P (&spilled_pseudos, reg)
       /* Ignore spilled pseudo-registers which can be here only if IRA
 	 is used.  */
-      || (flag_ira && optimize && r < 0))
+      || (flag_ira && ira_conflicts_p && r < 0))
     return;
 
   SET_REGNO_REG_SET (&pseudos_counted, reg);
@@ -1804,7 +1804,7 @@ count_spilled_pseudo (int spilled, int spilled_nregs, int reg)
 
   /* Ignore spilled pseudo-registers which can be here only if IRA is
      used.  */
-  if ((flag_ira && optimize && r < 0)
+  if ((flag_ira && ira_conflicts_p && r < 0)
       || REGNO_REG_SET_P (&spilled_pseudos, reg)
       || spilled + spilled_nregs <= r || r + nregs <= spilled)
     return;
@@ -1876,7 +1876,7 @@ find_reg (struct insn_chain *chain, int order)
 	  if (! ok)
 	    continue;
 
-	  if (flag_ira && optimize)
+	  if (flag_ira && ira_conflicts_p)
 	    {
 	      /* Ask IRA to find a better pseudo-register for
 		 spilling.  */
@@ -2158,7 +2158,7 @@ alter_reg (int i, int from_reg, bool dont_share_p)
       unsigned int min_align = reg_max_ref_width[i] * BITS_PER_UNIT;
       int adjust = 0;
 
-      if (flag_ira && optimize)
+      if (flag_ira && ira_conflicts_p)
 	{
 	  /* Mark the spill for IRA.  */
 	  SET_REGNO_REG_SET (&spilled_pseudos, i);
@@ -2177,7 +2177,8 @@ alter_reg (int i, int from_reg, bool dont_share_p)
 	 enough inherent space and enough total space.
 	 Otherwise, we allocate a new slot, making sure that it has no less
 	 inherent space, and no less total space, then the previous slot.  */
-      else if (from_reg == -1 || (!dont_share_p && flag_ira && optimize))
+      else if (from_reg == -1
+	       || (!dont_share_p && flag_ira && ira_conflicts_p))
 	{
 	  rtx stack_slot;
 
@@ -2202,7 +2203,7 @@ alter_reg (int i, int from_reg, bool dont_share_p)
 				       adjust);
 	    }
 
-	  if (! dont_share_p && flag_ira && optimize)
+	  if (! dont_share_p && flag_ira && ira_conflicts_p)
 	    /* Inform IRA about allocation a new stack slot.  */
 	    ira_mark_new_stack_slot (stack_slot, i, total_size);
 	}
@@ -3904,7 +3905,7 @@ finish_spills (int global)
       spill_reg_order[i] = -1;
 
   EXECUTE_IF_SET_IN_REG_SET (&spilled_pseudos, FIRST_PSEUDO_REGISTER, i, rsi)
-    if (! flag_ira || ! optimize || reg_renumber[i] >= 0)
+    if (! flag_ira || ! ira_conflicts_p || reg_renumber[i] >= 0)
       {
 	/* Record the current hard register the pseudo is allocated to
 	   in pseudo_previous_regs so we avoid reallocating it to the
@@ -3914,7 +3915,7 @@ finish_spills (int global)
 	SET_HARD_REG_BIT (pseudo_previous_regs[i], reg_renumber[i]);
 	/* Mark it as no longer having a hard register home.  */
 	reg_renumber[i] = -1;
-	if (flag_ira && optimize)
+	if (flag_ira && ira_conflicts_p)
 	  /* Inform IRA about the change.  */
 	  ira_mark_allocation_change (i);
 	/* We will need to scan everything again.  */
@@ -3944,7 +3945,7 @@ finish_spills (int global)
 	    }
 	}
 
-      if (! flag_ira || ! optimize)
+      if (! flag_ira || ! ira_conflicts_p)
 	{
 	  /* Retry allocating the spilled pseudos.  For each reg,
 	     merge the various reg sets that indicate which hard regs
@@ -3996,7 +3997,7 @@ finish_spills (int global)
       HARD_REG_SET used_by_pseudos;
       HARD_REG_SET used_by_pseudos2;
 
-      if (! flag_ira || ! optimize)
+      if (! flag_ira || ! ira_conflicts_p)
 	{
 	  /* Don't do it for IRA because IRA and the reload still can
 	     assign hard registers to the spilled pseudos on next
@@ -6999,7 +7000,7 @@ emit_input_reload_insns (struct insn_chain *chain, struct reload *rl,
 		  && REG_N_SETS (REGNO (old)) == 1)
 		{
 		  reg_renumber[REGNO (old)] = REGNO (reloadreg);
-		  if (flag_ira && optimize)
+		  if (flag_ira && ira_conflicts_p)
 		    /* Inform IRA about the change.  */
 		    ira_mark_allocation_change (REGNO (old));
 		  alter_reg (REGNO (old), -1, false);
@@ -8540,7 +8541,7 @@ delete_output_reload (rtx insn, int j, int last_reload_reg, rtx new_reload_reg)
 
       /* For the debugging info, say the pseudo lives in this reload reg.  */
       reg_renumber[REGNO (reg)] = REGNO (new_reload_reg);
-      if (flag_ira && optimize)
+      if (flag_ira && ira_conflicts_p)
 	/* Inform IRA about the change.  */
 	ira_mark_allocation_change (REGNO (reg));
       alter_reg (REGNO (reg), -1, false);