This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

profile.c revamp for new flow.c



Hi
This patch rewrites profile.c to use new flow infrastructure.  This is required
to make it at least vaguely functional with new ia-32 backend, since originally
it missed all edges caused by fp conditional branches (represented as
compare-and-jump).

The advantage is killing of approx 30% of code and having count information
calculated in the flowgraph.  Disadvantage for now is, that basic block notion
is slightly diferent.  We don't have basic block boundary after each call, so
if single basic block contains multiple function calls and one of them don't
return, gcov output is incorrect for rest of the block.

I plan to address this later by adding subblocks, but since I don't have any
idea what are the long term plans about that code and it would add some
complexity and make the patch harder to review and yet I consider this as quite
marginal case, I think we can live with this for a moment.

I basically did the revamp to allow experimental testing of various branch
prediction heruistics and for this purpose it works well.  I need to use
-fprofile-arcs even when testing, since disabling -fprofile-arcs makes CFGs to
not match, but this is problem for the old code as well and needs to be
addressed later in case we really want profile driven compilation.

I can live with -fprofile-arcs for the branch prediction heruistics now, since
I've modified split_edge to propagate count information, so the information is
kept up to date after instrumenting the edges for profiling.

I've also renamed most of terminology from arc to edge except for the flag
name.

I've succesfully instrumented and profilled the gcc itself using this code
(the original one just didn't work).

The problem arises with ordering now.  I am now executing the code once
CFG is built in flow pass, but it is having it's own dump file I can't use.
So I am now dumping to the flow dump everything except the overall statistics.

The flow dump is getting quite large now :(
Would be OK to sent separate patch to split it into multiple dumps?
(.cfg, .pred, .life) probably

Honza

Mon May  8 16:55:49 CEST 2000  Jan Hubicka  <jh@suse.cz>
	* basic-block.h (struct edge_def): New field "count".
	(struct basic_block_def): Likewise.
	* flow.c (entry_exit_blocks): Add count fileds.
	(split_edge): Keep count information up-to-date.
	(dump_edge_info, dump_flow_info): Dump count fields.
	* profile.c: Include basic-block.h and obstack-h
	(struct adj_list, ARC_TARGET, ARC_SOURCE, ARC_COUNT,
	struct bb_info, label_to_bb, label_to_bb_size, bb_graph,
	ignore_next_note, return_label_execution_count): Kill.
	(obstack_chunk_alloc, obstack_chunk_free): New macros.
	(edge_info, bb_info): New structures.
	(EDGE_INFO, BB_INFO, GCOV_INDEX_TO_BB, BB_TO_GCOV_INDEX): New
	macros.
	(total_num_arcs, total_num_arcs_instrumented): Rename arc to edge.
	(init_arc, expand_spanning_tree, fill_spanning_tree, tablejump_entry_p):
	Kill.
	(init_edge_profiler): Rename from ...
	(output_arc_profiler): ... this function; do not emit instructions,
	don't handle CALL insn instrumenting.
	(union_groups, find_group): New.
	(instrument_edges):  Rewrite from instrument_arcs to use
	commit_edge_insertions infrastructure.
	(compute_branch_probabilities): Remove argument num_blocks;
	rewrite to use new flow infrastructure.
	(branch_prob): Rewrite to use new flow infrastructure.
	(find_splanning_tree): Likewise.
	* toplev.c (rest_of_compilation): Do edge profiling after CFG
	is built.



*** basic-block.h.old	Sun May  7 09:18:33 2000
--- basic-block.h	Mon May  8 10:10:36 2000
*************** typedef struct edge_def {
*** 129,134 ****
--- 129,136 ----
  
    int flags;			/* see EDGE_* below  */
    int probability;		/* biased by REG_BR_PROB_BASE */
+   int count;			/* Expected number of executions calculated
+ 				   in profile.c  */
  } *edge;
  
  #define EDGE_FALLTHRU		1
*************** typedef struct basic_block_def {
*** 164,169 ****
--- 166,173 ----
  
    /* The active eh region before head and after end.  */
    int eh_beg, eh_end;
+ 
+   int count;		/* Expected number of executions calculated in profile.c  */
  } *basic_block;
  
  /* Number of basic blocks in the current function.  */
*** flow.c.old	Sun May  7 21:49:53 2000
--- flow.c	Mon May  8 14:10:32 2000
*************** struct basic_block_def entry_exit_blocks
*** 194,200 ****
      NULL,			/* aux */
      ENTRY_BLOCK,		/* index */
      0,				/* loop_depth */
!     -1, -1			/* eh_beg, eh_end */
    },
    {
      NULL,			/* head */
--- 194,201 ----
      NULL,			/* aux */
      ENTRY_BLOCK,		/* index */
      0,				/* loop_depth */
!     -1, -1,			/* eh_beg, eh_end */
!     0				/* count */
    },
    {
      NULL,			/* head */
*************** struct basic_block_def entry_exit_blocks
*** 207,213 ****
      NULL,			/* aux */
      EXIT_BLOCK,			/* index */
      0,				/* loop_depth */
!     -1, -1			/* eh_beg, eh_end */
    }
  };
  
--- 208,215 ----
      NULL,			/* aux */
      EXIT_BLOCK,			/* index */
      0,				/* loop_depth */
!     -1, -1,			/* eh_beg, eh_end */
!     0				/* count */
    }
  };
  
*************** split_edge (edge_in)
*** 1367,1372 ****
--- 1369,1375 ----
    /* Wire them up.  */
    bb->pred = edge_in;
    bb->succ = edge_out;
+   bb->count = edge_in->count;
  
    edge_in->dest = bb;
    edge_in->flags &= ~EDGE_CRITICAL;
*************** split_edge (edge_in)
*** 1377,1382 ****
--- 1380,1386 ----
    edge_out->dest = old_succ;
    edge_out->flags = EDGE_FALLTHRU;
    edge_out->probability = REG_BR_PROB_BASE;
+   edge_out->count = edge_in->count;
  
    old_succ->pred = edge_out;
  
*************** insn_dead_p (pbi, x, call_ok, notes)
*** 3777,3783 ****
  	  /* Walk the set of memory locations we are currently tracking
  	     and see if one is an identical match to this memory location.
  	     If so, this memory write is dead (remember, we're walking
! 	     backwards from the end of the block to the start.  */
  	  temp = pbi->mem_set_list;
  	  while (temp)
  	    {
--- 3782,3788 ----
  	  /* Walk the set of memory locations we are currently tracking
  	     and see if one is an identical match to this memory location.
  	     If so, this memory write is dead (remember, we're walking
! 	     backwards from the end of the block to the start).  */
  	  temp = pbi->mem_set_list;
  	  while (temp)
  	    {
*************** dump_flow_info (file)
*** 5571,5578 ****
        register basic_block bb = BASIC_BLOCK (i);
        register edge e;
  
!       fprintf (file, "\nBasic block %d: first insn %d, last %d, loop_depth %d.\n",
! 	       i, INSN_UID (bb->head), INSN_UID (bb->end), bb->loop_depth);
  
        fprintf (file, "Predecessors: ");
        for (e = bb->pred; e ; e = e->pred_next)
--- 5576,5583 ----
        register basic_block bb = BASIC_BLOCK (i);
        register edge e;
  
!       fprintf (file, "\nBasic block %d: first insn %d, last %d, loop_depth %d, count %d.\n",
! 	       i, INSN_UID (bb->head), INSN_UID (bb->end), bb->loop_depth, bb->count);
  
        fprintf (file, "Predecessors: ");
        for (e = bb->pred; e ; e = e->pred_next)
*************** dump_edge_info (file, e, do_succ)
*** 5615,5620 ****
--- 5620,5628 ----
    else
      fprintf (file, " %d", side->index);
  
+   if (e->count)
+     fprintf (file, " count:%d", e->count);
+ 
    if (e->flags)
      {
        static const char * const bitnames[] = {
*************** dump_bb (bb, outf)
*** 5653,5660 ****
    rtx last;
    edge e;
  
!   fprintf (outf, ";; Basic block %d, loop depth %d",
! 	   bb->index, bb->loop_depth);
    if (bb->eh_beg != -1 || bb->eh_end != -1)
      fprintf (outf, ", eh regions %d/%d", bb->eh_beg, bb->eh_end);
    putc ('\n', outf);
--- 5661,5668 ----
    rtx last;
    edge e;
  
!   fprintf (outf, ";; Basic block %d, loop depth %d, count %d",
! 	   bb->index, bb->loop_depth, bb->count);
    if (bb->eh_beg != -1 || bb->eh_end != -1)
      fprintf (outf, ", eh regions %d/%d", bb->eh_beg, bb->eh_end);
    putc ('\n', outf);
*** profile.c.old	Sun May  7 14:41:25 2000
--- profile.c	Mon May  8 12:22:28 2000
*************** along with GNU CC; see the file COPYING.
*** 21,30 ****
  the Free Software Foundation, 59 Temple Place - Suite 330,
  Boston, MA 02111-1307, USA.  */
  
- /* ??? Really should not put insns inside of LIBCALL sequences, when putting
-    insns after a call, should look for the insn setting the retval, and
-    insert the insns after that one.  */
- 
  /* ??? Register allocation should use basic block execution counts to
     give preference to the most commonly executed blocks.  */
  
--- 21,26 ----
*************** Boston, MA 02111-1307, USA.  */
*** 36,46 ****
  /* ??? Should calculate branch probabilities before instrumenting code, since
     then we can use arc counts to help decide which arcs to instrument.  */
  
- /* ??? Rearrange code so that the most frequently executed arcs become from
-    one block to the next block (i.e. a fall through), move seldom executed
-    code outside of loops even at the expense of adding a few branches to
-    achieve this, see Dain Sample's UC Berkeley thesis.  */
- 
  #include "config.h"
  #include "system.h"
  #include "rtl.h"
--- 32,37 ----
*************** Boston, MA 02111-1307, USA.  */
*** 54,108 ****
  #include "gcov-io.h"
  #include "toplev.h"
  #include "ggc.h"
  
- /* One of these is dynamically created whenever we identify an arc in the
-    function.  */
- 
- struct adj_list
- {
-   int source;
-   int target;
-   int arc_count;
-   unsigned int count_valid : 1;
-   unsigned int on_tree : 1;
-   unsigned int fake : 1;
-   unsigned int fall_through : 1;
-   rtx branch_insn;
-   struct adj_list *pred_next;
-   struct adj_list *succ_next;
- };
- 
- #define ARC_TARGET(ARCPTR) (ARCPTR->target)
- #define ARC_SOURCE(ARCPTR) (ARCPTR->source)
- #define ARC_COUNT(ARCPTR)  (ARCPTR->arc_count)
- 
- /* Count the number of basic blocks, and create an array of these structures,
-    one for each bb in the function.  */
  
  struct bb_info
! {
!   struct adj_list *succ;
!   struct adj_list *pred;
!   int succ_count;
!   int pred_count;
!   int exec_count;
!   unsigned int count_valid : 1;
!   unsigned int on_tree : 1;
!   rtx first_insn;
! };
! 
! /* Indexed by label number, gives the basic block number containing that
!    label.  */
! 
! static int *label_to_bb;
! 
! /* Number of valid entries in the label_to_bb array.  */
! 
! static int label_to_bb_size;
! 
! /* Indexed by block index, holds the basic block graph.  */
! 
! static struct bb_info *bb_graph;
  
  /* Name and file pointer of the output file for the basic block graph.  */
  
--- 45,81 ----
  #include "gcov-io.h"
  #include "toplev.h"
  #include "ggc.h"
+ #include "basic-block.h"
+ #include "obstack.h"
+ #define obstack_chunk_alloc xmalloc
+ #define obstack_chunk_free free
  
  
+ /* Additional information about the edges we need.  */
+ struct edge_info
+   {
+     unsigned int count_valid:1;
+     unsigned int on_tree:1;
+     unsigned int ignore:1;
+   };
  struct bb_info
!   {
!     unsigned int count_valid:1;
!     int succ_count;
!     int pred_count;
!   };
! 
! #define EDGE_INFO(e)  ((struct edge_info *)(e)->aux)
! #define BB_INFO(b)  ((struct bb_info *)(b)->aux)
! 
! /* Keep all basic block indexes nonnegative in the gcov output.  Index 0
!    is used for entry block, last block exit block.   */
! #define GCOV_INDEX_TO_BB(i)  ((i) == 0 ? ENTRY_BLOCK_PTR		\
!     			      : (((i) == n_basic_blocks + 1)		\
! 			         ? EXIT_BLOCK_PTR : BASIC_BLOCK ((i)-1)))
! #define BB_TO_GCOV_INDEX(bb)  ((bb) == ENTRY_BLOCK_PTR ? 0 		\
!   			       : ((bb) == EXIT_BLOCK_PTR		\
! 				  ? n_basic_blocks + 1 : (bb)->index + 1))
  
  /* Name and file pointer of the output file for the basic block graph.  */
  
*************** static FILE *bb_file;
*** 121,147 ****
  
  static char *last_bb_file_name;
  
- /* Indicates whether the next line number note should be output to
-    bb_file or not.  Used to eliminate a redundant note after an
-    expanded inline function call.  */
- 
- static int ignore_next_note;
- 
  /* Used by final, for allocating the proper amount of storage for the
     instrumented arc execution counts.  */
  
! int count_instrumented_arcs;
! 
! /* Number of executions for the return label.  */
! 
! int return_label_execution_count;
  
  /* Collect statistics on the performance of this pass for the entire source
     file.  */
  
  static int total_num_blocks;
! static int total_num_arcs;
! static int total_num_arcs_instrumented;
  static int total_num_blocks_created;
  static int total_num_passes;
  static int total_num_times_called;
--- 94,110 ----
  
  static char *last_bb_file_name;
  
  /* Used by final, for allocating the proper amount of storage for the
     instrumented arc execution counts.  */
  
! int count_instrumented_edges;
  
  /* Collect statistics on the performance of this pass for the entire source
     file.  */
  
  static int total_num_blocks;
! static int total_num_edges;
! static int total_num_edges_instrumented;
  static int total_num_blocks_created;
  static int total_num_passes;
  static int total_num_times_called;
*************** static int total_num_never_executed;
*** 150,165 ****
  static int total_num_branches;
  
  /* Forward declarations.  */
! static void init_arc PARAMS ((struct adj_list *, int, int, rtx));
! static void find_spanning_tree PARAMS ((int));
! static void expand_spanning_tree PARAMS ((int));
! static void fill_spanning_tree PARAMS ((int));
! static void init_arc_profiler PARAMS ((void));
! static void output_arc_profiler PARAMS ((int, rtx));
! static void instrument_arcs PARAMS ((rtx, int, FILE *));
  static void output_gcov_string PARAMS ((const char *, long));
! static int tablejump_entry_p PARAMS ((rtx, rtx));
! static void compute_branch_probabilities PARAMS ((int, FILE *));
  
  #ifndef LONG_TYPE_SIZE
  #define LONG_TYPE_SIZE BITS_PER_WORD
--- 113,126 ----
  static int total_num_branches;
  
  /* Forward declarations.  */
! static void find_spanning_tree PARAMS ((struct edge_list *));
! static void init_edge_profiler PARAMS ((void));
! static rtx gen_edge_profiler PARAMS ((int));
! static void instrument_edges PARAMS ((struct edge_list *, FILE *));
  static void output_gcov_string PARAMS ((const char *, long));
! static void compute_branch_probabilities PARAMS ((FILE *));
! static basic_block find_group PARAMS ((basic_block));
! static void union_groups PARAMS ((basic_block, basic_block));
  
  #ifndef LONG_TYPE_SIZE
  #define LONG_TYPE_SIZE BITS_PER_WORD
*************** static void compute_branch_probabilities
*** 168,390 ****
  /* If non-zero, we need to output a constructor to set up the
     per-object-file data. */
  static int need_func_profiler = 0;
- 
  
! /* Add arc instrumentation code to the entire insn chain.
  
     F is the first insn of the chain.
     NUM_BLOCKS is the number of basic blocks found in F.
     DUMP_FILE, if nonzero, is an rtl dump file we can write to.  */
  
  static void
! instrument_arcs (f, num_blocks, dump_file)
!      rtx f;
!      int num_blocks;
       FILE *dump_file;
  {
!   register int i;
!   register struct adj_list *arcptr, *backptr;
!   int num_arcs = 0;
!   int num_instr_arcs = 0;
!   rtx insn;
! 
!   /* Instrument the program start.  */
!   /* Handle block 0 specially, since it will always be instrumented,
!      but it doesn't have a valid first_insn or branch_insn.  We must
!      put the instructions before the NOTE_INSN_FUNCTION_BEG note, so
!      that they don't clobber any of the parameters of the current
!      function.  */
!   for (insn = f; insn; insn = NEXT_INSN (insn))
!     if (GET_CODE (insn) == NOTE
! 	&& NOTE_LINE_NUMBER (insn) == NOTE_INSN_FUNCTION_BEG)
!       break;
!   insn = PREV_INSN (insn);
!   need_func_profiler = 1;
!   output_arc_profiler (total_num_arcs_instrumented + num_instr_arcs++, insn);
! 
!   for (i = 1; i < num_blocks; i++)
!     for (arcptr = bb_graph[i].succ; arcptr; arcptr = arcptr->succ_next)
!       if (! arcptr->on_tree)
! 	{
! 	  if (dump_file)
! 	    fprintf (dump_file, "Arc %d to %d instrumented\n", i,
! 		     ARC_TARGET (arcptr));
  
! 	  /* Check to see if this arc is the only exit from its source block,
! 	     or the only entrance to its target block.  In either case,
! 	     we don't need to create a new block to instrument the arc.  */
! 	  
! 	  if (bb_graph[i].succ == arcptr && arcptr->succ_next == 0)
! 	    {
! 	      /* Instrument the source block.  */
! 	      output_arc_profiler (total_num_arcs_instrumented
! 				   + num_instr_arcs++,
! 				   PREV_INSN (bb_graph[i].first_insn));
! 	    }
! 	  else if (arcptr == bb_graph[ARC_TARGET (arcptr)].pred
! 		   && arcptr->pred_next == 0)
! 	    {
! 	      /* Instrument the target block.  */
! 	      output_arc_profiler (total_num_arcs_instrumented
! 				   + num_instr_arcs++, 
! 				   PREV_INSN (bb_graph[ARC_TARGET (arcptr)].first_insn));
! 	    }
! 	  else if (arcptr->fall_through)
! 	    {
! 	      /* This is a fall-through; put the instrumentation code after
! 		 the branch that ends this block.  */
! 	      
! 	      for (backptr = bb_graph[i].succ; backptr;
! 		   backptr = backptr->succ_next)
! 		if (backptr != arcptr)
! 		  break;
! 	      
! 	      output_arc_profiler (total_num_arcs_instrumented
! 				   + num_instr_arcs++,
! 				   backptr->branch_insn);
! 	    }
! 	  else
  	    {
! 	      /* Must emit a new basic block to hold the arc counting code.  */
! 	      enum rtx_code code = GET_CODE (PATTERN (arcptr->branch_insn));
! 
! 	      if (code == SET)
! 		{
! 		  /* Create the new basic block right after the branch.
! 		     Invert the branch so that it jumps past the end of the new
! 		     block.  The new block will consist of the instrumentation
! 		     code, and a jump to the target of this arc.  */
! 		  int this_is_simplejump = simplejump_p (arcptr->branch_insn);
! 		  rtx new_label = gen_label_rtx ();
! 		  rtx old_label, set_src;
! 		  rtx after = arcptr->branch_insn;
! 		  
! 		  /* Simplejumps can't reach here.  */
! 		  if (this_is_simplejump)
! 		    abort ();
! 
! 		  /* We can't use JUMP_LABEL, because it won't be set if we
! 		     are compiling without optimization.  */
! 
! 		  set_src = SET_SRC (single_set (arcptr->branch_insn));
! 		  if (GET_CODE (set_src) == LABEL_REF)
! 		    old_label = set_src;
! 		  else if (GET_CODE (set_src) != IF_THEN_ELSE)
! 		    abort ();
! 		  else if (XEXP (set_src, 1) == pc_rtx)
! 		    old_label = XEXP (XEXP (set_src, 2), 0);
! 		  else
! 		    old_label = XEXP (XEXP (set_src, 1), 0);
! 
! 		  /* Set the JUMP_LABEL so that redirect_jump will work.  */
! 		  JUMP_LABEL (arcptr->branch_insn) = old_label;
! 
! 		  /* Add a use for OLD_LABEL that will be needed when we emit
! 		     the JUMP_INSN below.  If we don't do this here,
! 		     `invert_jump' might delete it for us.  We must add two
! 		     when not optimizing, because the NUSES is zero now,
! 		     but must be at least two to prevent the label from being
! 		     deleted.  */
! 		  LABEL_NUSES (old_label) += 2;
! 		  
! 		  /* Emit the insns for the new block in reverse order,
! 		     since that is most convenient.  */
! 
! 		  if (this_is_simplejump)
! 		    {
! 		      after = NEXT_INSN (arcptr->branch_insn);
! 		      if (! redirect_jump (arcptr->branch_insn, new_label))
! 			/* Don't know what to do if this branch won't
! 			   redirect.  */
! 			abort ();
! 		    }
! 		  else
! 		    {
! 		      if (! invert_jump (arcptr->branch_insn, new_label))
! 			/* Don't know what to do if this branch won't invert.  */
! 			abort ();
! 
! 		      emit_label_after (new_label, after);
! 		      LABEL_NUSES (new_label)++;
! 		    }
! 		  emit_barrier_after (after);
! 		  emit_jump_insn_after (gen_jump (old_label), after);
! 		  JUMP_LABEL (NEXT_INSN (after)) = old_label;
! 		  
! 		  /* Instrument the source arc.  */
! 		  output_arc_profiler (total_num_arcs_instrumented
! 				       + num_instr_arcs++,
! 				       after);
! 		  if (this_is_simplejump)
! 		    {
! 		      emit_label_after (new_label, after);
! 		      LABEL_NUSES (new_label)++;
! 		    }
! 		}
! 	      else if (code == ADDR_VEC || code == ADDR_DIFF_VEC)
! 		{
! 		  /* A table jump.  Create a new basic block immediately
! 		     after the table, by emitting a barrier, a label, a
! 		     counting note, and a jump to the old label.  Put the
! 		     new label in the table.  */
! 		  
! 		  rtx new_label = gen_label_rtx ();
! 		  rtx old_lref, new_lref;
! 		  int index;
! 		  
! 		  /* Must determine the old_label reference, do this
! 		     by counting the arcs after this one, which will
! 		     give the index of our label in the table.  */
! 		  
! 		  index = 0;
! 		  for (backptr = arcptr->succ_next; backptr;
! 		       backptr = backptr->succ_next)
! 		    index++;
! 		  
! 		  old_lref = XVECEXP (PATTERN (arcptr->branch_insn),
! 				      (code == ADDR_DIFF_VEC), index);
! 		  
! 		  /* Emit the insns for the new block in reverse order,
! 		     since that is most convenient.  */
! 		  emit_jump_insn_after (gen_jump (XEXP (old_lref, 0)),
! 					arcptr->branch_insn);
! 		  JUMP_LABEL (NEXT_INSN (arcptr->branch_insn))
! 		    = XEXP (old_lref, 0);
! 
! 		  /* Instrument the source arc.  */
! 		  output_arc_profiler (total_num_arcs_instrumented
! 				       + num_instr_arcs++,
! 				       arcptr->branch_insn);
! 
! 		  emit_label_after (new_label, arcptr->branch_insn);
! 		  LABEL_NUSES (NEXT_INSN (arcptr->branch_insn))++;
! 		  emit_barrier_after (arcptr->branch_insn);
! 		  
! 		  /* Fix up the table jump.  */
! 		  new_lref = gen_rtx_LABEL_REF (Pmode, new_label);
! 		  XVECEXP (PATTERN (arcptr->branch_insn),
! 			   (code == ADDR_DIFF_VEC), index) = new_lref;
! 		}
! 	      else
! 		abort ();
! 
! 	      num_arcs += 1;
  	      if (dump_file)
! 		fprintf (dump_file,
! 			 "Arc %d to %d needed new basic block\n", i,
! 			 ARC_TARGET (arcptr));
  	    }
  	}
!   
!   total_num_arcs_instrumented += num_instr_arcs;
!   count_instrumented_arcs = total_num_arcs_instrumented;
  
!   total_num_blocks_created += num_arcs;
    if (dump_file)
!     {
!       fprintf (dump_file, "%d arcs instrumented\n", num_instr_arcs);
!       fprintf (dump_file, "%d extra basic blocks created\n", num_arcs);
!     }
  }
  
  /* Output STRING to bb_file, surrounded by DELIMITER.  */
--- 129,183 ----
  /* If non-zero, we need to output a constructor to set up the
     per-object-file data. */
  static int need_func_profiler = 0;
  
! /* Add edge instrumentation code to the entire insn chain.
  
     F is the first insn of the chain.
     NUM_BLOCKS is the number of basic blocks found in F.
     DUMP_FILE, if nonzero, is an rtl dump file we can write to.  */
  
  static void
! instrument_edges (el, dump_file)
!      struct edge_list *el;
       FILE *dump_file;
  {
!   int i;
!   int num_instr_edges = 0;
!   int num_edges = NUM_EDGES (el);
!   remove_fake_edges ();
  
!   for (i = 0; i < n_basic_blocks + 2; i++)
!     {
!       basic_block bb = GCOV_INDEX_TO_BB (i);
!       edge e = bb->succ;
!       while (e)
! 	{
! 	  struct edge_info *inf = EDGE_INFO (e);
! 	  if (!inf->ignore && !inf->on_tree)
  	    {
! 	      if (e->flags & EDGE_ABNORMAL)
! 		abort();
  	      if (dump_file)
! 		fprintf (dump_file, "Edge %d to %d instrumented %s\n",
! 			 e->src->index, e->dest->index,
! 			 e->flags & EDGE_CRITICAL ? "(and split)" : "");
! 	      need_func_profiler = 1;
! 	      insert_insn_on_edge (
! 			 gen_edge_profiler (total_num_edges_instrumented
! 					    + num_instr_edges++), e);
  	    }
+ 	  e = e->succ_next;
  	}
!     }
! 
!   total_num_edges_instrumented += num_instr_edges;
!   count_instrumented_edges = total_num_edges_instrumented;
  
!   total_num_blocks_created += num_edges;
    if (dump_file)
!     fprintf (dump_file, "%d edges instrumented\n", num_instr_edges);
! 
!   commit_edge_insertions ();
  }
  
  /* Output STRING to bb_file, surrounded by DELIMITER.  */
*************** output_gcov_string (string, delimiter)
*** 418,505 ****
    __write_long (delimiter, bb_file, 4);
  }
  
- /* Return TRUE if this insn must be a tablejump entry insn.  This works for
-    the MIPS port, but may give false negatives for some targets.  */
- 
- static int
- tablejump_entry_p (insn, label)
-      rtx insn, label;
- {
-   rtx next = next_active_insn (insn);
-   enum rtx_code code = GET_CODE (PATTERN (next));
- 
-   if (code != ADDR_DIFF_VEC && code != ADDR_VEC)
-     return 0;
- 
-   if (PREV_INSN (next) == XEXP (label, 0))
-     return 1;
- 
-   return 0;
- }
  
  /* Compute the branch probabilities for the various branches.
     Annotate them accordingly.  */
  
  static void
! compute_branch_probabilities (num_blocks, dump_file)
!      int num_blocks;
       FILE *dump_file;
  {
    int i;
!   int bad_counts = 0;
!   int num_arcs;
    int changes;
    int passes;
!   int prob;
!   int total;
!   int num_branches;
!   int num_never_executed;
    int hist_br_prob[20];
!   struct adj_list *arcptr;
  
!   /* For each arc not on the spanning tree, set its execution count from
       the .da file.  */
  
    /* The first count in the .da file is the number of times that the function
       was entered.  This is the exec_count for block zero.  */
  
!   num_arcs = 0;
!   for (i = 0; i < num_blocks; i++)
!     for (arcptr = bb_graph[i].succ; arcptr; arcptr = arcptr->succ_next)
!       if (! arcptr->on_tree)
! 	{
! 	  num_arcs++;
! 	  if (da_file)
! 	    {
! 	      long value;
! 	      __read_long (&value, da_file, 8);
! 	      ARC_COUNT (arcptr) = value;
! 	    }
! 	  else
! 	    ARC_COUNT (arcptr) = 0;
! 	  arcptr->count_valid = 1;
! 	  bb_graph[i].succ_count--;
! 	  bb_graph[ARC_TARGET (arcptr)].pred_count--;
! 	}
  
    if (dump_file)
!     fprintf (dump_file, "%d arc counts read\n", num_arcs);
  
    /* For every block in the file,
!      - if every exit/entrance arc has a known count, then set the block count
!      - if the block count is known, and every exit/entrance arc but one has
!        a known execution count, then set the count of the remaining arc
! 
!      As arc counts are set, decrement the succ/pred count, but don't delete
!      the arc, that way we can easily tell when all arcs are known, or only
!      one arc is unknown.  */
  
    /* The order that the basic blocks are iterated through is important.
       Since the code that finds spanning trees starts with block 0, low numbered
!      arcs are put on the spanning tree in preference to high numbered arcs.
!      Hence, most instrumented arcs are at the end.  Graph solving works much
       faster if we propagate numbers from the end to the start.
!      
       This takes an average of slightly more than 3 passes.  */
  
    changes = 1;
--- 211,301 ----
    __write_long (delimiter, bb_file, 4);
  }
  
  
  /* Compute the branch probabilities for the various branches.
     Annotate them accordingly.  */
  
  static void
! compute_branch_probabilities (dump_file)
       FILE *dump_file;
  {
    int i;
!   int num_edges = 0;
    int changes;
    int passes;
!   struct obstack bb_infos;
    int hist_br_prob[20];
!   int num_never_executed;
!   int num_branches;
!   int bad_counts = 0;
  
!   obstack_init (&bb_infos);
! 
!   /* Attach extra info block to each bb.  */
!   for (i = 0; i < n_basic_blocks + 2; i++)
!     {
!       basic_block bb = GCOV_INDEX_TO_BB (i);
!       edge e;
!       bb->aux = obstack_alloc (&bb_infos, sizeof (struct bb_info));
!       memset (bb->aux, 0, sizeof (struct bb_info));
!       for (e = bb->succ; e; e = e->succ_next)
! 	if (!EDGE_INFO (e)->ignore)
! 	  BB_INFO (bb)->succ_count++;
!       for (e = bb->pred; e; e = e->pred_next)
! 	if (!EDGE_INFO (e)->ignore)
! 	  BB_INFO (bb)->pred_count++;
!     }
! 
!   /* Avoid us from predicting entry on exit nodes.  */
!   BB_INFO (EXIT_BLOCK_PTR)->succ_count=2;
!   BB_INFO (ENTRY_BLOCK_PTR)->pred_count=2;
! 
!   /* For each edge not on the spanning tree, set its execution count from
       the .da file.  */
  
    /* The first count in the .da file is the number of times that the function
       was entered.  This is the exec_count for block zero.  */
  
!   for (i = 0; i < n_basic_blocks + 2; i++)
!     {
!       basic_block bb = GCOV_INDEX_TO_BB (i);
!       edge e;
!       for (e = bb->succ; e; e = e->succ_next)
! 	if (!EDGE_INFO (e)->ignore && !EDGE_INFO (e)->on_tree)
! 	  {
! 	    num_edges++;
! 	    if (da_file)
! 	      {
! 		long value;
! 		__read_long (&value, da_file, 8);
! 		e->count = value;
! 	      }
! 	    else
! 	      e->count = 0;
! 	    EDGE_INFO (e)->count_valid = 1;
! 	    BB_INFO (bb)->succ_count--;
! 	    BB_INFO (e->dest)->pred_count--;
! 	  }
!     }
  
    if (dump_file)
!     fprintf (dump_file, "%d edge counts read\n", num_edges);
  
    /* For every block in the file,
!      - if every exit/entrance edge has a known count, then set the block count
!      - if the block count is known, and every exit/entrance edge but one has
!      a known execution count, then set the count of the remaining edge
! 
!      As edge counts are set, decrement the succ/pred count, but don't delete
!      the edge, that way we can easily tell when all edges are known, or only
!      one edge is unknown.  */
  
    /* The order that the basic blocks are iterated through is important.
       Since the code that finds spanning trees starts with block 0, low numbered
!      edges are put on the spanning tree in preference to high numbered edges.
!      Hence, most instrumented edges are at the end.  Graph solving works much
       faster if we propagate numbers from the end to the start.
! 
       This takes an average of slightly more than 3 passes.  */
  
    changes = 1;
*************** compute_branch_probabilities (num_blocks
*** 508,593 ****
      {
        passes++;
        changes = 0;
! 
!       for (i = num_blocks - 1; i >= 0; i--)
  	{
! 	  struct bb_info *binfo = &bb_graph[i];
! 	  if (! binfo->count_valid)
  	    {
! 	      if (binfo->succ_count == 0)
  		{
! 		  total = 0;
! 		  for (arcptr = binfo->succ; arcptr;
! 		       arcptr = arcptr->succ_next)
! 		    total += ARC_COUNT (arcptr);
! 		  binfo->exec_count = total;
! 		  binfo->count_valid = 1;
  		  changes = 1;
  		}
! 	      else if (binfo->pred_count == 0)
  		{
! 		  total = 0;
! 		  for (arcptr = binfo->pred; arcptr;
! 		       arcptr = arcptr->pred_next)
! 		    total += ARC_COUNT (arcptr);
! 		  binfo->exec_count = total;
! 		  binfo->count_valid = 1;
  		  changes = 1;
  		}
  	    }
! 	  if (binfo->count_valid)
  	    {
! 	      if (binfo->succ_count == 1)
  		{
! 		  total = 0;
  		  /* One of the counts will be invalid, but it is zero,
  		     so adding it in also doesn't hurt.  */
! 		  for (arcptr = binfo->succ; arcptr;
! 		       arcptr = arcptr->succ_next)
! 		    total += ARC_COUNT (arcptr);
! 		  /* Calculate count for remaining arc by conservation.  */
! 		  total = binfo->exec_count - total;
! 		  /* Search for the invalid arc, and set its count.  */
! 		  for (arcptr = binfo->succ; arcptr;
! 		       arcptr = arcptr->succ_next)
! 		    if (! arcptr->count_valid)
  		      break;
! 		  if (! arcptr)
  		    abort ();
! 		  arcptr->count_valid = 1;
! 		  ARC_COUNT (arcptr) = total;
! 		  binfo->succ_count--;
  		  
! 		  bb_graph[ARC_TARGET (arcptr)].pred_count--;
  		  changes = 1;
  		}
! 	      if (binfo->pred_count == 1)
  		{
! 		  total = 0;
  		  /* One of the counts will be invalid, but it is zero,
  		     so adding it in also doesn't hurt.  */
! 		  for (arcptr = binfo->pred; arcptr;
! 		       arcptr = arcptr->pred_next)
! 		    total += ARC_COUNT (arcptr);
! 		  /* Calculate count for remaining arc by conservation.  */
! 		  total = binfo->exec_count - total;
! 		  /* Search for the invalid arc, and set its count.  */
! 		  for (arcptr = binfo->pred; arcptr;
! 		       arcptr = arcptr->pred_next)
! 		    if (! arcptr->count_valid)
  		      break;
! 		  if (! arcptr)
  		    abort ();
! 		  arcptr->count_valid = 1;
! 		  ARC_COUNT (arcptr) = total;
! 		  binfo->pred_count--;
  		  
! 		  bb_graph[ARC_SOURCE (arcptr)].succ_count--;
  		  changes = 1;
  		}
  	    }
  	}
      }
  
    total_num_passes += passes;
    if (dump_file)
--- 304,399 ----
      {
        passes++;
        changes = 0;
!       for (i = n_basic_blocks + 1; i >= 0; i--)
  	{
! 	  basic_block bb = GCOV_INDEX_TO_BB (i);
! 	  struct bb_info *bi = BB_INFO (bb);
! 	  if (! bi->count_valid)
  	    {
! 	      if (bi->succ_count == 0)
  		{
! 		  edge e;
! 		  int total = 0;
! 
! 		  for (e = bb->succ; e; e = e->succ_next)
! 		    total += e->count;
! 		  bb->count = total;
! 		  bi->count_valid = 1;
  		  changes = 1;
  		}
! 	      else if (bi->pred_count == 0)
  		{
! 		  edge e;
! 		  int total = 0;
! 
! 		  for (e = bb->pred; e; e = e->pred_next)
! 		    total += e->count;
! 		  bb->count = total;
! 		  bi->count_valid = 1;
  		  changes = 1;
  		}
  	    }
! 	  if (bi->count_valid)
  	    {
! 	      if (bi->succ_count == 1)
  		{
! 		  edge e;
! 		  int total = 0;
! 
  		  /* One of the counts will be invalid, but it is zero,
  		     so adding it in also doesn't hurt.  */
! 		  for (e = bb->succ; e; e = e->succ_next)
! 		    total += e->count;
! 
! 		  /* Seedgeh for the invalid edge, and set its count.  */
! 		  for (e = bb->succ; e; e = e->succ_next)
! 		    if (! EDGE_INFO (e)->count_valid && ! EDGE_INFO (e)->ignore)
  		      break;
! 
! 		  /* Calculate count for remaining edge by conservation.  */
! 		  total = bb->count - total;
! 
! 		  if (! e)
  		    abort ();
! 		  EDGE_INFO (e)->count_valid = 1;
! 		  e->count = total;
! 		  bi->succ_count--;
  		  
! 		  BB_INFO (e->dest)->pred_count--;
  		  changes = 1;
  		}
! 	      if (bi->pred_count == 1)
  		{
! 		  edge e;
! 		  int total = 0;
! 
  		  /* One of the counts will be invalid, but it is zero,
  		     so adding it in also doesn't hurt.  */
! 		  for (e = bb->pred; e; e = e->pred_next)
! 		    total += e->count;
! 
! 		  /* Seedgeh for the invalid edge, and set its count.  */
! 		  for (e = bb->pred; e; e = e->pred_next)
! 		    if (! EDGE_INFO (e)->count_valid && ! EDGE_INFO (e)->ignore)
  		      break;
! 
! 		  /* Calculate count for remaining edge by conservation.  */
! 		  total = bb->count - total + e->count;
! 
! 		  if (! e)
  		    abort ();
! 		  EDGE_INFO (e)->count_valid = 1;
! 		  e->count = total;
! 		  bi->pred_count--;
  		  
! 		  BB_INFO (e->src)->succ_count--;
  		  changes = 1;
  		}
  	    }
  	}
      }
+   if (dump_file)
+     dump_flow_info (dump_file);
  
    total_num_passes += passes;
    if (dump_file)
*************** compute_branch_probabilities (num_blocks
*** 595,608 ****
  
    /* If the graph has been correctly solved, every block will have a
       succ and pred count of zero.  */
!   for (i = 0; i < num_blocks; i++)
      {
!       struct bb_info *binfo = &bb_graph[i];
!       if (binfo->succ_count || binfo->pred_count)
  	abort ();
      }
  
!   /* For every arc, calculate its branch probability and add a reg_note
       to the branch insn to indicate this.  */
  
    for (i = 0; i < 20; i++)
--- 401,414 ----
  
    /* If the graph has been correctly solved, every block will have a
       succ and pred count of zero.  */
!   for (i = 0; i < n_basic_blocks; i++)
      {
!       basic_block bb = BASIC_BLOCK (i);
!       if (BB_INFO (bb)->succ_count || BB_INFO (bb)->pred_count)
  	abort ();
      }
  
!   /* For every edge, calculate its branch probability and add a reg_note
       to the branch insn to indicate this.  */
  
    for (i = 0; i < 20; i++)
*************** compute_branch_probabilities (num_blocks
*** 610,642 ****
    num_never_executed = 0;
    num_branches = 0;
  
!   for (i = 0; i < num_blocks; i++)
      {
!       struct bb_info *binfo = &bb_graph[i];
! 
!       total = binfo->exec_count;
!       for (arcptr = binfo->succ; arcptr; arcptr = arcptr->succ_next)
  	{
! 	  if (arcptr->branch_insn)
  	    {
  	      /* This calculates the branch probability as an integer between
  		 0 and REG_BR_PROB_BASE, properly rounded to the nearest
  		 integer.  Perform the arithmetic in double to avoid
  		 overflowing the range of ints.  */
- 
  	      if (total == 0)
  		prob = -1;
  	      else
  		{
! 		  rtx pat = PATTERN (arcptr->branch_insn);
  		  
! 		  prob = (((double)ARC_COUNT (arcptr) * REG_BR_PROB_BASE)
  			  + (total >> 1)) / total;
  		  if (prob < 0 || prob > REG_BR_PROB_BASE)
  		    {
  		      if (dump_file)
  			fprintf (dump_file, "bad count: prob for %d-%d thought to be %d (forcibly normalized)\n",
! 				 ARC_SOURCE (arcptr), ARC_TARGET (arcptr),
  				 prob);
  
  		      bad_counts = 1;
--- 416,456 ----
    num_never_executed = 0;
    num_branches = 0;
  
!   for (i = 0; i < n_basic_blocks; i++)
      {
!       basic_block bb = BASIC_BLOCK (i);
!       edge e;
!       rtx insn;
!       int total;
!       rtx note;
! 
!       total = bb->count;
!       if (!total)
! 	continue;
!       for (e = bb->succ; e; e = e->succ_next)
  	{
! 	  if (GET_CODE (e->src->end) == JUMP_INSN && nontrivial_condjump_p (e->src->end)
! 	      && !EDGE_INFO (e)->ignore
! 	      && !(e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL | EDGE_FAKE)))
  	    {
+ 	      int prob;
  	      /* This calculates the branch probability as an integer between
  		 0 and REG_BR_PROB_BASE, properly rounded to the nearest
  		 integer.  Perform the arithmetic in double to avoid
  		 overflowing the range of ints.  */
  	      if (total == 0)
  		prob = -1;
  	      else
  		{
! 		  rtx set = pc_set (e->src->end);
  		  
! 		  prob = (((double)e->count * REG_BR_PROB_BASE)
  			  + (total >> 1)) / total;
  		  if (prob < 0 || prob > REG_BR_PROB_BASE)
  		    {
  		      if (dump_file)
  			fprintf (dump_file, "bad count: prob for %d-%d thought to be %d (forcibly normalized)\n",
! 				 e->src->index, e->dest->index,
  				 prob);
  
  		      bad_counts = 1;
*************** compute_branch_probabilities (num_blocks
*** 644,666 ****
  		    }
  		  
  		  /* Match up probability with JUMP pattern.  */
! 
! 		  if (GET_CODE (pat) == SET
! 		      && GET_CODE (SET_SRC (pat)) == IF_THEN_ELSE)
! 		    {
! 		      if (ARC_TARGET (arcptr) == ARC_SOURCE (arcptr) + 1)
! 			{
! 			  /* A fall through arc should never have a
! 			     branch insn.  */
! 			  abort ();
! 			}
! 		      else
! 			{
! 			  /* This is the arc for the taken branch.  */
! 			  if (GET_CODE (XEXP (SET_SRC (pat), 2)) != PC)
! 			    prob = REG_BR_PROB_BASE - prob;
! 			}
! 		    }
  		}
  	      
  	      if (prob == -1)
--- 458,465 ----
  		    }
  		  
  		  /* Match up probability with JUMP pattern.  */
! 		  if (e != condjump_branch_edge (bb))
! 		    prob = REG_BR_PROB_BASE - prob;
  		}
  	      
  	      if (prob == -1)
*************** compute_branch_probabilities (num_blocks
*** 674,710 ****
  		}
  	      num_branches++;
  	      
! 	      REG_NOTES (arcptr->branch_insn)
! 		= gen_rtx_EXPR_LIST (REG_BR_PROB, GEN_INT (prob),
! 				     REG_NOTES (arcptr->branch_insn));
  	    }
  	}
  
        /* Add a REG_EXEC_COUNT note to the first instruction of this block.  */
!       if (! binfo->first_insn 
! 	  || GET_RTX_CLASS (GET_CODE (binfo->first_insn)) != 'i')
! 	{
! 	  /* Block 0 is a fake block representing function entry, and does
! 	     not have a real first insn.  The second last block might not
! 	     begin with a real insn.  */
! 	  if (i == num_blocks - 1)
! 	    return_label_execution_count = total;
! 	  else if (i != 0 && i != num_blocks - 2)
! 	    abort ();
! 	}
!       else
! 	{
! 	  REG_NOTES (binfo->first_insn)
! 	    = gen_rtx_EXPR_LIST (REG_EXEC_COUNT, GEN_INT (total),
! 				 REG_NOTES (binfo->first_insn));
! 	  if (i == num_blocks - 1)
! 	    return_label_execution_count = total;
! 	}
      }
    
    /* This should never happen.  */
    if (bad_counts)
!     warning ("Arc profiling: some arc counts were bad.");
  
    if (dump_file)
      {
--- 473,506 ----
  		}
  	      num_branches++;
  	      
! 	      note = find_reg_note (e->src->end, REG_BR_PROB, 0);
! 	      /*  There may be already note put by some other pass, such
! 	          as builtin_expect expander.  */
! 	      if (note)
! 		XEXP (note, 0) = GEN_INT (prob);
! 	      else
! 		REG_NOTES (e->src->end)
! 		  = gen_rtx_EXPR_LIST (REG_BR_PROB, GEN_INT (prob),
! 				       REG_NOTES (e->src->end));
  	    }
  	}
  
        /* Add a REG_EXEC_COUNT note to the first instruction of this block.  */
!       insn = next_nonnote_insn (bb->head);
! 
!       if (GET_CODE (bb->head) == CODE_LABEL)
! 	insn = next_nonnote_insn (insn);
! 
!       /* Avoid crash on empty basic blocks.  */
!       if (insn && INSN_P (insn))
! 	REG_NOTES (insn)
! 	  = gen_rtx_EXPR_LIST (REG_EXEC_COUNT, GEN_INT (total),
! 			       REG_NOTES (insn));
      }
    
    /* This should never happen.  */
    if (bad_counts)
!     warning ("Arc profiling: some edge counts were bad.");
  
    if (dump_file)
      {
*************** compute_branch_probabilities (num_blocks
*** 722,739 ****
        for (i = 0; i < 20; i++)
  	total_hist_br_prob[i] += hist_br_prob[i];
      }
  }
  
  /* Instrument and/or analyze program behavior based on program flow graph.
     In either case, this function builds a flow graph for the function being
     compiled.  The flow graph is stored in BB_GRAPH.
  
!    When FLAG_PROFILE_ARCS is nonzero, this function instruments the arcs in
     the flow graph that are needed to reconstruct the dynamic behavior of the
     flow graph.
  
     When FLAG_BRANCH_PROBABILITIES is nonzero, this function reads auxiliary
!    information from a data file containing arc count information from previous
     executions of the function being compiled.  In this case, the flow graph is
     annotated with actual execution counts, which are later propagated into the
     rtl for optimization purposes.
--- 518,537 ----
        for (i = 0; i < 20; i++)
  	total_hist_br_prob[i] += hist_br_prob[i];
      }
+ 
+   obstack_free (&bb_infos, NULL);
  }
  
  /* Instrument and/or analyze program behavior based on program flow graph.
     In either case, this function builds a flow graph for the function being
     compiled.  The flow graph is stored in BB_GRAPH.
  
!    When FLAG_PROFILE_ARCS is nonzero, this function instruments the edges in
     the flow graph that are needed to reconstruct the dynamic behavior of the
     flow graph.
  
     When FLAG_BRANCH_PROBABILITIES is nonzero, this function reads auxiliary
!    information from a data file containing edge count information from previous
     executions of the function being compiled.  In this case, the flow graph is
     annotated with actual execution counts, which are later propagated into the
     rtl for optimization purposes.
*************** compute_branch_probabilities (num_blocks
*** 742,1428 ****
  
  void
  branch_prob (f, dump_file)
!      rtx f;
       FILE *dump_file;
  {
!   int i, num_blocks;
!   struct adj_list *arcptr;
!   int num_arcs;
! 
!   /* start of a function.  */
!   if (flag_test_coverage)
!     output_gcov_string (current_function_name, (long) -2);
  
!   /* Execute this only if doing arc profiling or branch probabilities.  */
    if (! profile_arc_flag && ! flag_branch_probabilities
        && ! flag_test_coverage)
      abort ();
  
-   total_num_times_called++;
- 
-   /* Create an array label_to_bb of ints of size max_label_num.  */
-   label_to_bb_size = max_label_num ();
-   label_to_bb = (int *) oballoc (label_to_bb_size * sizeof (int));
-   bzero ((char *) label_to_bb, label_to_bb_size * sizeof (int));
- 
-   /* Scan the insns in the function, count the number of basic blocks
-      present.  When a code label is passed, set label_to_bb[label] = bb
-      number.  */
- 
-   /* The first block found will be block 1, so that function entry can be
-      block 0.  */
- 
-   {
-     register RTX_CODE prev_code = JUMP_INSN;
-     register RTX_CODE code;
-     register rtx insn;
-     register int i;
-     int block_separator_emitted = 0;
- 
-     ignore_next_note = 0;
- 
-     for (insn = NEXT_INSN (f), i = 0; insn; insn = NEXT_INSN (insn))
-       {
- 	code = GET_CODE (insn);
- 
- 	if (code == BARRIER)
- 	  ;
- 	else if (code == CODE_LABEL)
- 	  /* This label is part of the next block, but we can't increment
- 	     block number yet since there might be multiple labels.  */
- 	  label_to_bb[CODE_LABEL_NUMBER (insn)] = i + 1;
- 	/* We make NOTE_INSN_SETJMP notes into a block of their own, so that
- 	   they can be the target of the fake arc for the setjmp call.
- 	   This avoids creating cycles of fake arcs, which would happen if
- 	   the block after the setjmp call contained a call insn.  */
- 	else if ((prev_code == JUMP_INSN || prev_code == CALL_INSN
- 		  || prev_code == CODE_LABEL || prev_code == BARRIER)
- 		 && (GET_RTX_CLASS (code) == 'i'
- 		     || (code == NOTE
- 			 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_SETJMP)))
- 	  {
- 	    i += 1;
- 
- 	    /* Emit the block separator if it hasn't already been emitted.  */
- 	    if (flag_test_coverage && ! block_separator_emitted)
- 	      {
- 		/* Output a zero to the .bb file to indicate that a new
- 		   block list is starting.  */
- 		__write_long (0, bb_file, 4);
- 	      }
- 	    block_separator_emitted = 0;
- 	  }
- 	/* If flag_test_coverage is true, then we must add an entry to the
- 	   .bb file for every note.  */
- 	else if (code == NOTE && flag_test_coverage)
- 	  {
- 	    /* Must ignore the line number notes that immediately follow the
- 	       end of an inline function to avoid counting it twice.  There
- 	       is a note before the call, and one after the call.  */
- 	    if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_REPEATED_LINE_NUMBER)
- 	      ignore_next_note = 1;
- 	    else if (NOTE_LINE_NUMBER (insn) > 0)
- 	      {
- 		if (ignore_next_note)
- 		  ignore_next_note = 0;
- 		else
- 		  {
- 		    /* Emit a block separator here to ensure that a NOTE
- 		       immediately following a JUMP_INSN or CALL_INSN will end
- 		       up in the right basic block list.  */
- 		    if ((prev_code == JUMP_INSN || prev_code == CALL_INSN
- 			 || prev_code == CODE_LABEL || prev_code == BARRIER)
- 			&& ! block_separator_emitted)
- 		      {
- 			/* Output a zero to the .bb file to indicate that
- 			   a new block list is starting.  */
- 			__write_long (0, bb_file, 4);
- 
- 			block_separator_emitted = 1;
- 		      }
- 		    
- 		    /* If this is a new source file, then output the file's
- 		       name to the .bb file.  */
- 		    if (! last_bb_file_name
- 			|| strcmp (NOTE_SOURCE_FILE (insn),
- 				   last_bb_file_name))
- 		      {
- 			if (last_bb_file_name)
- 			  free (last_bb_file_name);
- 			last_bb_file_name = xstrdup (NOTE_SOURCE_FILE (insn));
- 			output_gcov_string (NOTE_SOURCE_FILE (insn), (long)-1);
- 		      }
- 
- 		    /* Output the line number to the .bb file.  Must be done
- 		       after the output_bb_profile_data() call, and after the
- 		       file name is written, to ensure that it is correctly
- 		       handled by gcov.  */
- 		    __write_long (NOTE_LINE_NUMBER (insn), bb_file, 4);
- 		  }
- 	      }
- 	  }
- 
- 	if (code != NOTE)
- 	  prev_code = code;
- 	else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_SETJMP)
- 	  prev_code = CALL_INSN;
-       }
- 
-     /* Allocate last `normal' entry for bb_graph.  */
- 
-     /* The last insn was a jump, call, or label.  In that case we have
-        a block at the end of the function with no insns.  */
-     if (prev_code == JUMP_INSN || prev_code == CALL_INSN
- 	|| prev_code == CODE_LABEL || prev_code == BARRIER)
-       {
- 	i++;
- 
- 	/* Emit the block separator if it hasn't already been emitted.  */
- 	if (flag_test_coverage && ! block_separator_emitted)
- 	  {
- 	    /* Output a zero to the .bb file to indicate that a new
- 	       block list is starting.  */
- 	    __write_long (0, bb_file, 4);
- 	  }
-       }
- 
-     /* Create another block to stand for EXIT, and make all return insns, and
-        the last basic block point here.  Add one more to account for block
-        zero.  */
-     num_blocks = i + 2;
-   }
- 
-   total_num_blocks += num_blocks;
    if (dump_file)
!     fprintf (dump_file, "%d basic blocks\n", num_blocks);
! 
!   /* If we are only doing test coverage here, then return now.  */
!   if (! profile_arc_flag && ! flag_branch_probabilities)
!     return;
! 
!   /* Create and initialize the arrays that will hold bb_graph
!      and execution count info.  */
  
!   bb_graph = (struct bb_info *) xcalloc (num_blocks, 
! 					 sizeof (struct bb_info));
! 
!   {
!     /* Scan the insns again:
!        - at the entry to each basic block, increment the predecessor count
!        (and successor of previous block) if it is a fall through entry,
!        create adj_list entries for this and the previous block
!        - at each jump insn, increment predecessor/successor counts for
!        target/source basic blocks, add this insn to pred/succ lists.
! 
!        This also cannot be broken out as a separate subroutine
!        because it uses `alloca'.  */
! 
!     register RTX_CODE prev_code = JUMP_INSN;
!     register RTX_CODE code;
!     register rtx insn;
!     register int i;
!     int fall_through = 0;
!     struct adj_list *arcptr;
!     int dest = 0;
! 
!     /* Block 0 always falls through to block 1.  */
!     num_arcs = 0;
!     arcptr = (struct adj_list *) alloca (sizeof (struct adj_list));
!     init_arc (arcptr, 0, 1, 0);
!     arcptr->fall_through = 1;
!     num_arcs++;
! 
!     /* Add a fake fall through arc from the last block to block 0, to make the
!        graph complete.  */
!     arcptr = (struct adj_list *) alloca (sizeof (struct adj_list));
!     init_arc (arcptr, num_blocks - 1, 0, 0);
!     arcptr->fake = 1;
!     num_arcs++;
! 
!     /* Exit must be one node of the graph, and all exits from the function
!        must point there.  When see a return branch, must point the arc to the
!        exit node.  */
! 
!     /* Must start scan with second insn in function as above.  */
!     for (insn = NEXT_INSN (f), i = 0; insn; insn = NEXT_INSN (insn))
!       {
! 	code = GET_CODE (insn);
! 
! 	if (code == BARRIER)
! 	  fall_through = 0;
! 	else if (code == CODE_LABEL)
! 	  ;
! 	/* We make NOTE_INSN_SETJMP notes into a block of their own, so that
! 	   they can be the target of the fake arc for the setjmp call.
! 	   This avoids creating cycles of fake arcs, which would happen if
! 	   the block after the setjmp call ended with a call.  */
! 	else if ((prev_code == JUMP_INSN || prev_code == CALL_INSN
! 		  || prev_code == CODE_LABEL || prev_code == BARRIER)
! 		 && (GET_RTX_CLASS (code) == 'i'
! 		     || (code == NOTE
! 			 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_SETJMP)))
! 	  {
! 	    /* This is the first insn of the block.  */
! 	    i += 1;
! 	    if (fall_through)
! 	      {
! 		arcptr = (struct adj_list *) alloca (sizeof (struct adj_list));
! 		init_arc (arcptr, i - 1, i, 0);
! 		arcptr->fall_through = 1;
! 
! 		num_arcs++;
! 	      }
! 	    fall_through = 1;
! 	    bb_graph[i].first_insn = insn;
! 	  }
! 	else if (code == NOTE)
! 	  {;}
! 
! 	if (code == CALL_INSN)
! 	  {
! 	    /* In the normal case, the call returns, and this is just like
! 	       a branch fall through.  */
! 	    fall_through = 1;
! 
! 	    /* Setjmp may return more times than called, so to make the graph
! 	       solvable, add a fake arc from the function entrance to the
! 	       next block.
! 
! 	       All other functions may return fewer times than called (if
! 	       a descendent call longjmp or exit), so to make the graph
! 	       solvable, add a fake arc to the function exit from the
! 	       current block.
! 
! 	       Distinguish the cases by checking for a SETJUMP note.
! 	       A call_insn can be the last ins of a function, so must check
! 	       to see if next insn actually exists.  */
! 	    arcptr = (struct adj_list *) alloca (sizeof (struct adj_list));
! 	    if (NEXT_INSN (insn)
! 		&& GET_CODE (NEXT_INSN (insn)) == NOTE
! 		&& NOTE_LINE_NUMBER (NEXT_INSN (insn)) == NOTE_INSN_SETJMP)
! 	      init_arc (arcptr, 0, i+1, insn);
! 	    else
! 	      init_arc (arcptr, i, num_blocks-1, insn);
! 	    arcptr->fake = 1;
! 	    num_arcs++;
! 	  }
! 	else if (code == JUMP_INSN)
! 	  {
! 	    rtx tem, pattern = PATTERN (insn);
! 	    rtx tablejump = 0;
! 
! 	    /* If running without optimization, then jump label won't be valid,
! 	       so we must search for the destination label in that case.
! 	       We have to handle tablejumps and returns specially anyways, so
! 	       we don't check the JUMP_LABEL at all here.  */
! 
! 	    /* ??? This code should be rewritten.  We need a more elegant way
! 	       to find the LABEL_REF.  We need a more elegant way to
! 	       differentiate tablejump entries from computed gotos.
! 	       We should perhaps reuse code from flow to compute the CFG
! 	       instead of trying to compute it here.
! 
! 	       We can't use current_function_has_computed_jump, because that
! 	       is calculated later by flow.  We can't use computed_jump_p,
! 	       because that returns true for tablejump entry insns for some
! 	       targets, e.g. HPPA and MIPS.  */
! 
! 	    if (GET_CODE (pattern) == PARALLEL)
! 	      {
! 		/* This assumes that PARALLEL jumps with a USE are
! 		   tablejump entry jumps.  The same assumption can be found
! 		   in computed_jump_p.  */
! 		/* Make an arc from this jump to the label of the
! 		   jump table.  This will instrument the number of
! 		   times the switch statement is executed.  */
! 		if (GET_CODE (XVECEXP (pattern, 0, 1)) == USE)
! 		  {
! 		    tem = XEXP (XVECEXP (pattern, 0, 1), 0);
! 		    if (GET_CODE (tem) != LABEL_REF)
! 		      abort ();
! 		    dest = label_to_bb[CODE_LABEL_NUMBER (XEXP (tem, 0))];
! 		  }
! 		else if (GET_CODE (XVECEXP (pattern, 0, 0)) == SET
! 			 && SET_DEST (XVECEXP (pattern, 0, 0)) == pc_rtx)
! 		  {
! 		    tem = SET_SRC (XVECEXP (pattern, 0, 0));
! 		    if (GET_CODE (tem) == PLUS
! 			&& GET_CODE (XEXP (tem, 1)) == LABEL_REF)
! 		      {
! 			tem = XEXP (tem, 1);
! 			dest = label_to_bb [CODE_LABEL_NUMBER (XEXP (tem, 0))];
! 		      }
! 		  }
! 		else
! 		  abort ();
! 	      }
! 	    else if (GET_CODE (pattern) == ADDR_VEC
! 		     || GET_CODE (pattern) == ADDR_DIFF_VEC)
! 	      tablejump = pattern;
! 	    else if (GET_CODE (pattern) == RETURN)
! 	      dest = num_blocks - 1;
! 	    else if (GET_CODE (pattern) != SET)
! 	      abort ();
! 	    else if ((tem = SET_SRC (pattern))
! 		     && GET_CODE (tem) == LABEL_REF)
! 	      dest = label_to_bb[CODE_LABEL_NUMBER (XEXP (tem, 0))];
! 	    /* Recognize HPPA table jump entry.  This code is similar to
! 	       the code above in the PARALLEL case.  */
! 	    else if (GET_CODE (tem) == PLUS
! 		     && GET_CODE (XEXP (tem, 0)) == MEM
! 		     && GET_CODE (XEXP (XEXP (tem, 0), 0)) == PLUS
! 		     && GET_CODE (XEXP (XEXP (XEXP (tem, 0), 0), 0)) == PC
! 		     && GET_CODE (XEXP (tem, 1)) == LABEL_REF
! 		     && tablejump_entry_p (insn, XEXP (tem, 1)))
! 	      dest = label_to_bb[CODE_LABEL_NUMBER (XEXP (XEXP (tem, 1), 0))];
! 	    /* Recognize the MIPS table jump entry.  */
! 	    else if (GET_CODE (tem) == PLUS
! 		     && GET_CODE (XEXP (tem, 0)) == REG
! 		     && GET_CODE (XEXP (tem, 1)) == LABEL_REF
! 		     && tablejump_entry_p (insn, XEXP (tem, 1)))
! 	      dest = label_to_bb[CODE_LABEL_NUMBER (XEXP (XEXP (tem, 1), 0))];
! 	    else
! 	      {
! 		rtx label_ref;
! 
! 		/* Must be an IF_THEN_ELSE branch.  If it isn't, assume it
! 		   is a computed goto, which aren't supported yet.  */
! 		if (GET_CODE (tem) != IF_THEN_ELSE)
! 		  fatal ("-fprofile-arcs does not support computed gotos");
! 		if (XEXP (tem, 1) != pc_rtx)
! 		  label_ref = XEXP (tem, 1);
! 		else
! 		  label_ref = XEXP (tem, 2);
! 		dest = label_to_bb[CODE_LABEL_NUMBER (XEXP (label_ref, 0))];
! 	      }
  
! 	    if (tablejump)
! 	      {
! 		int diff_vec_p = GET_CODE (tablejump) == ADDR_DIFF_VEC;
! 		int len = XVECLEN (tablejump, diff_vec_p);
! 		int k;
! 
! 		for (k = 0; k < len; k++)
! 		  {
! 		    rtx tem = XEXP (XVECEXP (tablejump, diff_vec_p, k), 0);
! 		    dest = label_to_bb[CODE_LABEL_NUMBER (tem)];
  
! 		    arcptr = (struct adj_list *) alloca (sizeof(struct adj_list));
! 		    init_arc (arcptr, i, dest, insn);
  
! 		    num_arcs++;
! 		  }
! 	      }
! 	    else
! 	      {
! 		arcptr = (struct adj_list *) alloca (sizeof (struct adj_list));
! 		init_arc (arcptr, i, dest, insn);
  
! 		num_arcs++;
! 	      }
  
! 	    /* Determine whether or not this jump will fall through.
! 	       Unconditional jumps and returns are not always followed by
! 	       barriers.  */
! 	    pattern = PATTERN (insn);
! 	    if (GET_CODE (pattern) == PARALLEL
! 		|| GET_CODE (pattern) == RETURN)
! 	      fall_through = 0;
! 	    else if (GET_CODE (pattern) == ADDR_VEC
! 		     || GET_CODE (pattern) == ADDR_DIFF_VEC)
! 	      /* These aren't actually jump insns, but they never fall
! 		 through, so...  */
! 	      fall_through = 0;
! 	    else
! 	      {
! 		if (GET_CODE (pattern) != SET || SET_DEST (pattern) != pc_rtx)
! 		  abort ();
! 		if (GET_CODE (SET_SRC (pattern)) != IF_THEN_ELSE)
! 		  fall_through = 0;
! 	      }
! 	  }
  
! 	if (code != NOTE)
! 	  prev_code = code;
! 	else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_SETJMP)
! 	  {
! 	    /* Make a fake insn to tag our notes on.  */
! 	    bb_graph[i].first_insn = insn
! 	      = emit_insn_after (gen_rtx_USE (VOIDmode, stack_pointer_rtx),
! 				 insn);
! 	    prev_code = CALL_INSN;
! 	  }
!       }
  
!     /* If the code at the end of the function would give a new block, then
!        do the following.  */
  
!     if (prev_code == JUMP_INSN || prev_code == CALL_INSN
! 	|| prev_code == CODE_LABEL || prev_code == BARRIER)
!       {
! 	if (fall_through)
! 	  {
! 	    arcptr = (struct adj_list *) alloca (sizeof (struct adj_list));
! 	    init_arc (arcptr, i, i + 1, 0);
! 	    arcptr->fall_through = 1;
  
! 	    num_arcs++;
! 	  }
! 	  
! 	/* This may not be a real insn, but that should not cause a problem.  */
! 	bb_graph[i+1].first_insn = get_last_insn ();
!       }
! 
!     /* There is always a fake arc from the last block of the function
!        to the function exit block.  */
!     arcptr = (struct adj_list *) alloca (sizeof (struct adj_list));
!     init_arc (arcptr, num_blocks-2, num_blocks-1, 0);
!     arcptr->fake = 1;
!     num_arcs++;
!   }
  
!   total_num_arcs += num_arcs;
!   if (dump_file)
!     fprintf (dump_file, "%d arcs\n", num_arcs);
  
!   /* Create spanning tree from basic block graph, mark each arc that is
!      on the spanning tree.  */
  
!   /* To reduce the instrumentation cost, make two passes over the tree.
!      First, put as many must-split (crowded and fake) arcs on the tree as
!      possible, then on the second pass fill in the rest of the tree.
!      Note that the spanning tree is considered undirected, so that as many
!      must-split arcs as possible can be put on it.
! 
!      Fallthrough arcs which are crowded should not be chosen on the first
!      pass, since they do not require creating a new basic block.  These
!      arcs will have fall_through set.  */
  
!   find_spanning_tree (num_blocks);
  
    /* Create a .bbg file from which gcov can reconstruct the basic block
       graph.  First output the number of basic blocks, and then for every
!      arc output the source and target basic block numbers.
       NOTE: The format of this file must be compatible with gcov.  */
  
    if (flag_test_coverage)
      {
        int flag_bits;
  
!       __write_long (num_blocks, bbg_file, 4);
!       __write_long (num_arcs, bbg_file, 4);
  
!       for (i = 0; i < num_blocks; i++)
  	{
  	  long count = 0;
! 	  for (arcptr = bb_graph[i].succ; arcptr; arcptr = arcptr->succ_next)
! 	    count++;
  	  __write_long (count, bbg_file, 4);
  
! 	  for (arcptr = bb_graph[i].succ; arcptr; arcptr = arcptr->succ_next)
  	    {
! 	      flag_bits = 0;
! 	      if (arcptr->on_tree)
! 		flag_bits |= 0x1;
! 	      if (arcptr->fake)
! 		flag_bits |= 0x2;
! 	      if (arcptr->fall_through)
! 		flag_bits |= 0x4;
! 
! 	      __write_long (ARC_TARGET (arcptr), bbg_file, 4);
! 	      __write_long (flag_bits, bbg_file, 4);
  	    }
  	}
  
!       /* Emit a -1 to separate the list of all arcs from the list of
  	 loop back edges that follows.  */
        __write_long (-1, bbg_file, 4);
      }
  
!   /* For each arc not on the spanning tree, add counting code as rtl.  */
  
    if (profile_arc_flag)
      {
!       instrument_arcs (f, num_blocks, dump_file);
        allocate_reg_info (max_reg_num (), FALSE, FALSE);
      }
  
!   /* Execute the rest only if doing branch probabilities.  */
!   if (flag_branch_probabilities)
!     compute_branch_probabilities (num_blocks, dump_file);
! 
!   /* Clean up.  */
!   free (bb_graph);
  }
  
! /* Initialize a new arc.
!    ARCPTR is the empty adj_list this function fills in.
!    SOURCE is the block number of the source block.
!    TARGET is the block number of the target block.
!    INSN is the insn which transfers control from SOURCE to TARGET,
!    or zero if the transfer is implicit.  */
  
! static void
! init_arc (arcptr, source, target, insn)
!      struct adj_list *arcptr;
!      int source, target;
!      rtx insn;
  {
!   ARC_TARGET (arcptr) = target;
!   ARC_SOURCE (arcptr) = source;
! 
!   ARC_COUNT (arcptr) = 0;
!   arcptr->count_valid = 0;
!   arcptr->on_tree = 0;
!   arcptr->fake = 0;
!   arcptr->fall_through = 0;
!   arcptr->branch_insn = insn;
! 
!   arcptr->succ_next = bb_graph[source].succ;
!   bb_graph[source].succ = arcptr;
!   bb_graph[source].succ_count++;
! 
!   arcptr->pred_next = bb_graph[target].pred;
!   bb_graph[target].pred = arcptr;
!   bb_graph[target].pred_count++;
  }
  
! /* This function searches all of the arcs in the program flow graph, and puts
!    as many bad arcs as possible onto the spanning tree.  Bad arcs include
!    fake arcs (needed for setjmp(), longjmp(), exit()) which MUST be on the
!    spanning tree as they can't be instrumented.  Also, arcs which must be
!    split when instrumented should be part of the spanning tree if possible.  */
  
  static void
! find_spanning_tree (num_blocks)
!      int num_blocks;
  {
    int i;
!   struct adj_list *arcptr;
!   struct bb_info *binfo = &bb_graph[0];
! 
!   /* Fake arcs must be part of the spanning tree, and are always safe to put
!      on the spanning tree.  Fake arcs will either be a successor of node 0,
!      a predecessor of the last node, or from the last node to node 0.  */
! 
!   for (arcptr = bb_graph[0].succ; arcptr; arcptr = arcptr->succ_next)
!     if (arcptr->fake)
!       {
! 	/* Adding this arc should never cause a cycle.  This is a fatal 
! 	   error if it would.  */
! 	if (bb_graph[ARC_TARGET (arcptr)].on_tree && binfo->on_tree)
! 	  abort();
! 	else
! 	  {
! 	    arcptr->on_tree = 1;
! 	    bb_graph[ARC_TARGET (arcptr)].on_tree = 1;
! 	    binfo->on_tree = 1;
! 	  }
!       }
  
!   binfo = &bb_graph[num_blocks-1];
!   for (arcptr = binfo->pred; arcptr; arcptr = arcptr->pred_next)
!     if (arcptr->fake)
!       {
! 	/* Adding this arc should never cause a cycle.  This is a fatal 
! 	   error if it would.  */
! 	if (bb_graph[ARC_SOURCE (arcptr)].on_tree && binfo->on_tree)
! 	  abort();
! 	else
! 	  {
! 	    arcptr->on_tree = 1;
! 	    bb_graph[ARC_SOURCE (arcptr)].on_tree = 1;
! 	    binfo->on_tree = 1;
! 	  }
!       }
!   /* The only entrace to node zero is a fake arc.  */
!   bb_graph[0].pred->on_tree = 1;
!   
!   /* Arcs which are crowded at both the source and target should be put on
!      the spanning tree if possible, except for fall_throuch arcs which never
!      require adding a new block even if crowded, add arcs with the same source
!      and dest which must always be instrumented.  */
!   for (i = 0; i < num_blocks; i++)
!     {
!       binfo = &bb_graph[i];
! 
!       for (arcptr = binfo->succ; arcptr; arcptr = arcptr->succ_next)
! 	if (! ((binfo->succ == arcptr && arcptr->succ_next == 0)
! 	       || (bb_graph[ARC_TARGET (arcptr)].pred
! 		   && arcptr->pred_next == 0))
! 	    && ! arcptr->fall_through
! 	    && ARC_TARGET (arcptr) != i)
! 	  {
! 	    /* This is a crowded arc at both source and target.  Try to put
! 	       in on the spanning tree.  Can do this if either the source or
! 	       target block is not yet on the tree.  */
! 	    if (! bb_graph[ARC_TARGET (arcptr)].on_tree	|| ! binfo->on_tree)
! 	      {
! 		arcptr->on_tree = 1;
! 		bb_graph[ARC_TARGET (arcptr)].on_tree = 1;
! 		binfo->on_tree = 1;
! 	      }
! 	  }
!     }
  
!   /* Clear all of the basic block on_tree bits, so that we can use them to
!      create the spanning tree.  */
!   for (i = 0; i < num_blocks; i++)
!     bb_graph[i].on_tree = 0;
! 
!   /* Now fill in the spanning tree until every basic block is on it.
!      Don't put the 0 to 1 fall through arc on the tree, since it is 
!      always cheap to instrument, so start filling the tree from node 1.  */
! 
!   for (i = 1; i < num_blocks; i++)
!     for (arcptr = bb_graph[i].succ; arcptr; arcptr = arcptr->succ_next)
!       if (! arcptr->on_tree
! 	  && ! bb_graph[ARC_TARGET (arcptr)].on_tree)
! 	{
! 	  fill_spanning_tree (i);
! 	  break;
! 	}
! }
  
! /* Add arcs reached from BLOCK to the spanning tree if they are needed and
!    not already there.  */
  
! static void
! fill_spanning_tree (block)
!      int block;
! {
!   struct adj_list *arcptr;
!   
!   expand_spanning_tree (block);
  
!   for (arcptr = bb_graph[block].succ; arcptr; arcptr = arcptr->succ_next)
!     if (! arcptr->on_tree
! 	&& ! bb_graph[ARC_TARGET (arcptr)].on_tree)
!       {
! 	arcptr->on_tree = 1;
! 	fill_spanning_tree (ARC_TARGET (arcptr));
!       }
! }
  
! /* When first visit a block, must add all blocks that are already connected
!    to this block via tree arcs to the spanning tree.  */
  
! static void
! expand_spanning_tree (block)
!      int block;
! {
!   struct adj_list *arcptr;
  
!   bb_graph[block].on_tree = 1;
  
!   for (arcptr = bb_graph[block].succ; arcptr; arcptr = arcptr->succ_next)
!     if (arcptr->on_tree && ! bb_graph[ARC_TARGET (arcptr)].on_tree)
!       expand_spanning_tree (ARC_TARGET (arcptr));
!     
!   for (arcptr = bb_graph[block].pred;
!        arcptr; arcptr = arcptr->pred_next)
!     if (arcptr->on_tree && ! bb_graph[ARC_SOURCE (arcptr)].on_tree)
!       expand_spanning_tree (ARC_SOURCE (arcptr));
  }
  
  /* Perform file-level initialization for branch-prob processing.  */
--- 540,893 ----
  
  void
  branch_prob (f, dump_file)
!      rtx f ATTRIBUTE_UNUSED;
       FILE *dump_file;
  {
!   int i;
!   int num_edges;
!   /* Obstack used to allocate edge info structures.  */
!   struct obstack edge_infos;
!   struct edge_list *el;
  
!   /* Execute this only if doing edge profiling or branch probabilities.  */
    if (! profile_arc_flag && ! flag_branch_probabilities
        && ! flag_test_coverage)
      abort ();
  
    if (dump_file)
!     dump_flow_info (dump_file);
  
!   /* start of a function.  */
!   if (flag_test_coverage)
!     output_gcov_string (current_function_name, (long) -2);
  
!   total_num_times_called++;
  
!   obstack_init (&edge_infos);
  
!   /* We can't handle cyclic regions constructed using abnormal edges.  To avoid
!      these we replace every source of abnormal edge by an fake edge from entry node 
!      and every destination by fake edge to exit.  This keeps graph acyclic and our
!      calculation exact for all normal edges except for exit and entrance ones.
!    
!      We also add fake exit edges for each call and asm statement in the basic, since
!      it may nor return.  ??? This makes gcov results inexact for basic blocks with
!      multiple calls.  We will need to add notion of subblocks here.  */
  
!   for (i = 0; i < n_basic_blocks ; i++)
!     {
!       rtx insn;
!       int need_exit_edge = 0, need_entry_edge = 0;
!       int have_exit_edge = 0, have_entry_edge = 0;
!       basic_block bb = BASIC_BLOCK (i);
!       edge e;
!       for (e = bb->succ; e; e = e->succ_next)
! 	{
! 	  if ((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL))
! 	       && e->dest != EXIT_BLOCK_PTR)
! 	    need_exit_edge = 1;
! 	  if (e->dest == EXIT_BLOCK_PTR)
! 	    have_exit_edge = 1;
! 	}
!       for (e = bb->pred; e; e = e->pred_next)
! 	{
! 	  if ((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL))
! 	       && e->src != ENTRY_BLOCK_PTR)
! 	    need_entry_edge = 1;
! 	  if (e->src == ENTRY_BLOCK_PTR)
! 	    have_entry_edge = 1;
! 	}
!       for (insn = bb->head; insn != NEXT_INSN (bb->end); insn = NEXT_INSN (insn))
! 	{
! 	  rtx set;
! 	  if (GET_CODE (insn) == CALL_INSN && !CONST_CALL_P (insn))
! 	    need_exit_edge = 1;
! 	  if (GET_CODE (insn) == INSN)
! 	    {
! 	      set = PATTERN (insn);
! 	      if (GET_CODE (set) == PARALLEL)
! 	        set = XVECEXP (set, 0, 0);
! 	      if (GET_CODE (set) == ASM_OPERANDS && MEM_VOLATILE_P (set))
! 	        need_exit_edge = 1;
! 	      if (GET_CODE (set) == ASM_INPUT)
! 	        need_exit_edge = 1;
! 	    }
! 	}
!       if (need_exit_edge && !have_exit_edge)
! 	{
! 	  if (dump_file)
! 	    fprintf (dump_file, "Adding fake exit edge to bb %i\n", bb->index);
!           make_edge (NULL, bb, EXIT_BLOCK_PTR, EDGE_FAKE);
! 	}
!       if (need_entry_edge && !have_entry_edge)
! 	{
! 	  if (dump_file)
! 	    fprintf (dump_file, "Adding fake entry edge to bb %i\n", bb->index);
!           make_edge (NULL, ENTRY_BLOCK_PTR, bb, EDGE_FAKE);
! 	}
!     }
  
!   el=create_edge_list();
!   num_edges = NUM_EDGES (el);
  
!   for (i = 0 ; i < num_edges ; i++)
!     {
!       edge e = INDEX_EDGE (el, i);
!       e->count = 0;
!       e->aux = obstack_alloc (&edge_infos, sizeof (struct edge_info));
!       memset (e->aux, 0, sizeof (struct edge_info));
! 
!       /* Mark edges we've replaced by fake edges above as ignored.  */
!       if ((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL))
! 	  && e->src != ENTRY_BLOCK_PTR && e->dest != EXIT_BLOCK_PTR)
! 	EDGE_INFO (e)->ignore = 1;
!     }
  
!   total_num_blocks += n_basic_blocks + 2;
!   if (dump_file)
!     fprintf (dump_file, "%d basic blocks\n", n_basic_blocks);
  
!   total_num_edges += num_edges;
!   if (dump_file)
!     fprintf (dump_file, "%d edges\n", num_edges);
  
!   verify_flow_info();
!   /* Output line number information about each basic block for GCOV utility.  */
!   if (flag_test_coverage)
!     {
!       int i = 0;
!       for (i = 0 ; i < n_basic_blocks; i++)
!         {
! 	  basic_block bb = BASIC_BLOCK (i);
! 	  rtx insn = bb->head;
!           static int ignore_next_note = 0;
! 
! 	  /* We are looking for line number notes.  Search backward before
! 	     basic block to find correct ones.  */
! 	  insn = prev_nonnote_insn (insn);
! 	  if (!insn)
! 	    insn = get_insns();
! 	  else
! 	    insn = NEXT_INSN (insn);
  
! 	  /* Output a zero to the .bb file to indicate that a new
! 	     block list is starting.  */
! 	  __write_long (0, bb_file, 4);
  
! 	  while (insn != bb->end)
! 	    {
! 	      if (GET_CODE (insn) == NOTE)
! 		{
! 		  /* Must ignore the line number notes that immediately follow
! 		     the end of an inline function to avoid counting it twice.
! 		     There is a note before the call, and one after the call.
! 		   */
! 		  if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_REPEATED_LINE_NUMBER)
! 		    ignore_next_note = 1;
! 		  else if (NOTE_LINE_NUMBER (insn) > 0)
! 		    {
! 		      if (ignore_next_note)
! 			ignore_next_note = 0;
! 		      else
! 			{
! 			  /* If this is a new source file, then output the
! 			     file's name to the .bb file.  */
! 			  if (! last_bb_file_name
! 			      || strcmp (NOTE_SOURCE_FILE (insn),
! 					 last_bb_file_name))
! 			    {
! 			      if (last_bb_file_name)
! 				free (last_bb_file_name);
! 			      last_bb_file_name
! 				= xstrdup (NOTE_SOURCE_FILE (insn));
! 			      output_gcov_string (NOTE_SOURCE_FILE (insn),
! 						  (long)-1);
! 			    }
! 			  /* Output the line number to the .bb file.  Must be
! 			     done after the output_bb_profile_data() call, and
! 			     after the file name is written, to ensure that it
! 			     is correctly handled by gcov.  */
! 			  __write_long (NOTE_LINE_NUMBER (insn), bb_file, 4);
! 			}
! 		    }
! 		}
! 	      insn = NEXT_INSN (insn);
! 	    }
!         }
!       __write_long (0, bb_file, 4);
!     }
  
!   /* Create spanning tree from basic block graph, mark each edge that is
!      on the spanning tree.  We insert as many abnormal and critical edges
!      as possible to minimize number of edge splits necesary. */
  
!   find_spanning_tree (el);
  
    /* Create a .bbg file from which gcov can reconstruct the basic block
       graph.  First output the number of basic blocks, and then for every
!      edge output the source and target basic block numbers.
       NOTE: The format of this file must be compatible with gcov.  */
  
    if (flag_test_coverage)
      {
        int flag_bits;
  
!       /* The plus 2 stands for entry and exit block.  */
!       __write_long (n_basic_blocks + 2, bbg_file, 4);
!       __write_long (num_edges + 1, bbg_file, 4);
  
!       for (i = 0; i < n_basic_blocks + 1; i++)
  	{
+ 	  basic_block bb = GCOV_INDEX_TO_BB (i);
+ 	  edge e;
  	  long count = 0;
! 
! 	  for (e = bb->succ; e; e = e->succ_next)
! 	    if (!EDGE_INFO (e)->ignore)
! 	      count++;
  	  __write_long (count, bbg_file, 4);
  
! 	  for (e = bb->succ; e; e = e->succ_next)
  	    {
! 	      struct edge_info *i = EDGE_INFO (e);
! 	      if (!i->ignore)
! 		{
! 		  flag_bits = 0;
! 		  if (i->on_tree)
! 		    flag_bits |= 0x1;
! 		  if (e->flags & EDGE_ABNORMAL)
! 		    flag_bits |= 0x2;
! 		  if (e->flags & EDGE_FALLTHRU)
! 		    flag_bits |= 0x4;
! 
! 		  __write_long (BB_TO_GCOV_INDEX (e->dest), bbg_file, 4);
! 		  __write_long (flag_bits, bbg_file, 4);
! 	        }
  	    }
  	}
+       /* Emit fake loopback edge for EXIT block to maitain compatibility with
+          old gcov format.  */
+       __write_long (1, bbg_file, 4);
+       __write_long (0, bbg_file, 4);
+       __write_long (0x1, bbg_file, 4);
  
!       /* Emit a -1 to separate the list of all edges from the list of
  	 loop back edges that follows.  */
        __write_long (-1, bbg_file, 4);
      }
  
!   if (flag_branch_probabilities)
!     compute_branch_probabilities (dump_file);
! 
!   /* For each edge not on the spanning tree, add counting code as rtl.  */
  
    if (profile_arc_flag)
      {
!       instrument_edges (el, dump_file);
        allocate_reg_info (max_reg_num (), FALSE, FALSE);
      }
  
!   remove_fake_edges ();
!   obstack_free (&edge_infos, NULL);
!   free_edge_list (el);
  }
  
! /* Union find algorithm implementation for the basic blocks using
!    aux fields. */
  
! static basic_block
! find_group (bb)
!      basic_block bb;
  {
!   basic_block group = bb, bb1;
!   while (group->aux != group)
!     group = (basic_block) group->aux;
!   /* Compress path.  */
!   while (bb->aux != group)
!     {
!       bb1 = bb->aux;
!       bb->aux = group;
!       bb = bb1;
!     }
!   return group;
  }
  
! static void
! union_groups (bb1, bb2)
!      basic_block bb1, bb2;
! {
!   /* I am not having place for the rank field.  OK.  Lets go w/o it,
!      this code is unlikely going to be performance problem anyway.  */
!   if (find_group (bb1) == find_group(bb2))
!     abort();
!   bb1 = find_group (bb1);
!   bb1->aux = find_group (bb2);
! }
! 
! /* This function seedgehes all of the edges in the program flow graph, and puts
!    as many bad edges as possible onto the spanning tree.  Bad edges include
!    abnormals edges, wich can't be instrumented at the moment.  Since it is
!    possible for fake edges to form an cycle, we will have to develop some
!    better way in future.  Also put critical edges to the tree, since they
!    are expensive.  */
  
  static void
! find_spanning_tree (el)
!      struct edge_list *el;
  {
    int i;
!   int num_edges = NUM_EDGES (el);
  
!   /* We use aux field for standard union-find algorithm.  */
!   EXIT_BLOCK_PTR->aux = EXIT_BLOCK_PTR;
!   ENTRY_BLOCK_PTR->aux = ENTRY_BLOCK_PTR;
!   for (i = 0; i < n_basic_blocks; i++)
!     BASIC_BLOCK (i)->aux = BASIC_BLOCK (i);
  
!   /* Add fake edge exit to entry we can't instrument.  */
  
!   union_groups (EXIT_BLOCK_PTR, ENTRY_BLOCK_PTR);
  
!   /* First insert all abnormal edges to the tree unless they form an cycle.  */
  
!   for (i = 0; i < num_edges; i++)
!     {
!       edge e = INDEX_EDGE (el, i);
!       if ((e->flags & (EDGE_ABNORMAL | EDGE_ABNORMAL_CALL | EDGE_FAKE))
! 	  && !EDGE_INFO (e)->ignore
! 	  && (find_group (e->src) != find_group (e->dest)))
! 	{
! 	  EDGE_INFO (e)->on_tree = 1;
! 	  union_groups (e->src, e->dest);
! 	}
!     }
  
!   /* Now insert all critical edges to the tree unless they form an cycle.  */
  
!   for (i = 0; i < num_edges; i++)
!     {
!       edge e = INDEX_EDGE (el, i);
!       if ((e->flags & EDGE_CRITICAL)
! 	  && !EDGE_INFO (e)->ignore
! 	  && (find_group (e->src) != find_group (e->dest)))
! 	{
! 	  EDGE_INFO (e)->on_tree = 1;
! 	  union_groups (e->src, e->dest);
! 	}
!     }
  
!   /* And now the rest.  */
  
!   for (i = 0; i < num_edges; i++)
!     {
!       edge e = INDEX_EDGE (el, i);
!       if (find_group (e->src) != find_group (e->dest)
! 	  && !EDGE_INFO (e)->ignore)
! 	{
! 	  EDGE_INFO (e)->on_tree = 1;
! 	  union_groups (e->src, e->dest);
! 	}
!     }
  }
  
  /* Perform file-level initialization for branch-prob processing.  */
*************** init_branch_prob (filename)
*** 1470,1476 ****
  	warning ("file %s not found, execution counts assumed to be zero.",
  		 da_file_name);
  
!       /* The first word in the .da file gives the number of instrumented arcs,
  	 which is not needed for our purposes.  */
  
        if (da_file)
--- 935,941 ----
  	warning ("file %s not found, execution counts assumed to be zero.",
  		 da_file_name);
  
!       /* The first word in the .da file gives the number of instrumented edges,
  	 which is not needed for our purposes.  */
  
        if (da_file)
*************** init_branch_prob (filename)
*** 1478,1488 ****
      }
  
    if (profile_arc_flag)
!     init_arc_profiler ();
  
    total_num_blocks = 0;
!   total_num_arcs = 0;
!   total_num_arcs_instrumented = 0;
    total_num_blocks_created = 0;
    total_num_passes = 0;
    total_num_times_called = 0;
--- 943,953 ----
      }
  
    if (profile_arc_flag)
!     init_edge_profiler ();
  
    total_num_blocks = 0;
!   total_num_edges = 0;
!   total_num_edges_instrumented = 0;
    total_num_blocks_created = 0;
    total_num_passes = 0;
    total_num_times_called = 0;
*************** end_branch_prob (dump_file)
*** 1526,1534 ****
      {
        fprintf (dump_file, "\n");
        fprintf (dump_file, "Total number of blocks: %d\n", total_num_blocks);
!       fprintf (dump_file, "Total number of arcs: %d\n", total_num_arcs);
!       fprintf (dump_file, "Total number of instrumented arcs: %d\n",
! 	       total_num_arcs_instrumented);
        fprintf (dump_file, "Total number of blocks created: %d\n",
  	       total_num_blocks_created);
        fprintf (dump_file, "Total number of graph solution passes: %d\n",
--- 991,999 ----
      {
        fprintf (dump_file, "\n");
        fprintf (dump_file, "Total number of blocks: %d\n", total_num_blocks);
!       fprintf (dump_file, "Total number of edges: %d\n", total_num_edges);
!       fprintf (dump_file, "Total number of instrumented edges: %d\n",
! 	       total_num_edges_instrumented);
        fprintf (dump_file, "Total number of blocks created: %d\n",
  	       total_num_blocks_created);
        fprintf (dump_file, "Total number of graph solution passes: %d\n",
*************** end_branch_prob (dump_file)
*** 1552,1565 ****
      }
  }
  
! /* The label used by the arc profiling code.  */
  
  static rtx profiler_label;
  
  /* Initialize the profiler_label.  */
  
  static void
! init_arc_profiler ()
  {
    /* Generate and save a copy of this so it can be shared.  */
    char *name = ggc_alloc_string (NULL, 20);
--- 1017,1030 ----
      }
  }
  
! /* The label used by the edge profiling code.  */
  
  static rtx profiler_label;
  
  /* Initialize the profiler_label.  */
  
  static void
! init_edge_profiler ()
  {
    /* Generate and save a copy of this so it can be shared.  */
    char *name = ggc_alloc_string (NULL, 20);
*************** init_arc_profiler ()
*** 1568,1583 ****
    ggc_add_rtx_root (&profiler_label, 1);
  }
  
! /* Output instructions as RTL to increment the arc execution count.  */
  
! static void
! output_arc_profiler (arcno, insert_after)
!      int arcno;
!      rtx insert_after;
  {
    rtx profiler_target_addr
!     = (arcno ? plus_constant (profiler_label,
! 			      LONG_TYPE_SIZE / BITS_PER_UNIT * arcno)
         : profiler_label);
    enum machine_mode mode = mode_for_size (LONG_TYPE_SIZE, MODE_INT, 0);
    rtx profiler_reg = gen_reg_rtx (mode);
--- 1033,1047 ----
    ggc_add_rtx_root (&profiler_label, 1);
  }
  
! /* Output instructions as RTL to increment the edge execution count.  */
  
! static rtx
! gen_edge_profiler (edgeno)
!      int edgeno;
  {
    rtx profiler_target_addr
!     = (edgeno ? plus_constant (profiler_label,
! 			      LONG_TYPE_SIZE / BITS_PER_UNIT * edgeno)
         : profiler_label);
    enum machine_mode mode = mode_for_size (LONG_TYPE_SIZE, MODE_INT, 0);
    rtx profiler_reg = gen_reg_rtx (mode);
*************** output_arc_profiler (arcno, insert_after
*** 1585,1635 ****
    rtx mem_ref, add_ref;
    rtx sequence;
  
-   /* In this case, reload can use explicitly mentioned hard registers for
-      reloads.  It is not safe to output profiling code between a call
-      and the instruction that copies the result to a pseudo-reg.  This
-      is because reload may allocate one of the profiling code pseudo-regs
-      to the return value reg, thus clobbering the return value.  So we
-      must check for calls here, and emit the profiling code after the
-      instruction that uses the return value, if any.
- 
-      ??? The code here performs the same tests that reload does so hopefully
-      all the bases are covered.  */
- 
-   if (SMALL_REGISTER_CLASSES
-       && GET_CODE (insert_after) == CALL_INSN
-       && (GET_CODE (PATTERN (insert_after)) == SET
- 	  || (GET_CODE (PATTERN (insert_after)) == PARALLEL
- 	      && GET_CODE (XVECEXP (PATTERN (insert_after), 0, 0)) == SET)))
-     {
-       rtx return_reg;
-       rtx next_insert_after = next_nonnote_insn (insert_after);
- 
-       /* The first insn after the call may be a stack pop, skip it.  */
-       if (next_insert_after
- 	  && GET_CODE (next_insert_after) == INSN
- 	  && GET_CODE (PATTERN (next_insert_after)) == SET
- 	  && SET_DEST (PATTERN (next_insert_after)) == stack_pointer_rtx)
- 	next_insert_after = next_nonnote_insn (next_insert_after);
- 
-       if (next_insert_after
- 	  && GET_CODE (next_insert_after) == INSN)
- 	{
- 	  if (GET_CODE (PATTERN (insert_after)) == SET)
- 	    return_reg = SET_DEST (PATTERN (insert_after));
- 	  else
- 	    return_reg = SET_DEST (XVECEXP (PATTERN (insert_after), 0, 0));
- 
- 	  /* Now, NEXT_INSERT_AFTER may be an instruction that uses the
- 	     return value.  However, it could also be something else,
- 	     like a CODE_LABEL, so check that the code is INSN.  */
- 	  if (next_insert_after != 0
- 	      && GET_RTX_CLASS (GET_CODE (next_insert_after)) == 'i'
- 	      && reg_referenced_p (return_reg, PATTERN (next_insert_after)))
- 	    insert_after = next_insert_after;
- 	}
-     }
- 
    start_sequence ();
  
    emit_move_insn (address_reg, profiler_target_addr);
--- 1049,1054 ----
*************** output_arc_profiler (arcno, insert_after
*** 1645,1651 ****
  
    sequence = gen_sequence ();
    end_sequence ();
!   emit_insn_after (sequence, insert_after);
  }
  
  /* Output code for a constructor that will invoke __bb_init_func, if
--- 1064,1070 ----
  
    sequence = gen_sequence ();
    end_sequence ();
!   return sequence;
  }
  
  /* Output code for a constructor that will invoke __bb_init_func, if
*************** output_func_start_profiler ()
*** 1661,1667 ****
    int save_flag_inline_functions = flag_inline_functions;
  
    /* It's either already been output, or we don't need it because we're
!      not doing profile-arcs. */
    if (! need_func_profiler)
      return;
  
--- 1080,1086 ----
    int save_flag_inline_functions = flag_inline_functions;
  
    /* It's either already been output, or we don't need it because we're
!      not doing profile-edges. */
    if (! need_func_profiler)
      return;
  
*** toplev.c.old1	Sun Apr 30 19:13:05 2000
--- toplev.c	Sun May  7 22:24:05 2000
*************** rest_of_compilation (decl)
*** 3078,3097 ****
  
    cse_not_expected = 1;
  
-   if (profile_arc_flag || flag_test_coverage || flag_branch_probabilities)
-     {
-       timevar_push (TV_BRANCH_PROB);
-       open_dump_file (DFI_bp, decl);
- 
-       branch_prob (insns, rtl_dump_file);
- 
-       close_dump_file (DFI_bp, print_rtl, insns);
-       timevar_pop (TV_BRANCH_PROB);
- 
-       if (ggc_p)
- 	ggc_collect ();
-     }
- 
    regclass_init ();
  
    /* Print function header into flow dump now
--- 3100,3105 ----
*************** rest_of_compilation (decl)
*** 3105,3110 ****
--- 3113,3130 ----
    timevar_push (TV_FLOW);
    find_basic_blocks (insns, max_reg_num (), rtl_dump_file);
    cleanup_cfg (insns);
+ 
+   if (profile_arc_flag || flag_test_coverage || flag_branch_probabilities)
+     {
+       timevar_push (TV_BRANCH_PROB);
+ 
+       branch_prob (insns, rtl_dump_file);
+ 
+       timevar_pop (TV_BRANCH_PROB);
+ 
+       if (ggc_p)
+ 	ggc_collect ();
+     }
    /*structural_analysis();*/
    if (optimize)
      {

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]