[PATCH] Remove RTL branch prediction as a separate pass

[Steven Bosscher <stevenb dot gcc at gmail dot com>]

Hi,

With loop.c out of the way, we never have PROFILE_ABSENT
unless we're not predicting branches at all.  This means
that pass_branch_prob is dead.

I have verified this first by making eestimate_probability
abort when it is called, just to be sure.  Then I crafted
the patch below.  Bootstrapped and tested on x86_64-linux.

This should give us a small speedup because we don't have
to compute dominators and find loops anymore for no reason.

Is this patch OK for mainline?

Gr.
Steven

<date>  Steven Bosscher  <stevenb.gcc@gmail.com>

	* predict.c (estimate_probability): Remove.
	* profile.c (rest_of_handle_branch_prob, pass_branch_prob): Remove.
	* basic-block.h (estimate_probability): Remove prototype.
	* passes.c (finish_optimization_passes): Dump branch prediction or
	profiling heuristics to the .profile tree dump.
	(init_optimization_passes): Unqueue pass_branch_prob.

	* doc/invoke.texi (-fdump-rtl-bp): Remove.

Index: predict.c
===================================================================
--- predict.c	(revision 111977)
+++ predict.c	(working copy)
@@ -826,85 +826,6 @@ bb_estimate_probability_locally (basic_b
       }
 }
 
-/* Statically estimate the probability that a branch will be taken and 
produce
-   estimated profile.  When profile feedback is present never executed 
portions
-   of function gets estimated.  */
-
-void
-estimate_probability (struct loops *loops_info)
-{
-  basic_block bb;
-
-  connect_infinite_loops_to_exit ();
-  calculate_dominance_info (CDI_DOMINATORS);
-  calculate_dominance_info (CDI_POST_DOMINATORS);
-
-  predict_loops (loops_info, true);
-
-  iv_analysis_done ();
-
-  /* Attempt to predict conditional jumps using a number of heuristics.  */
-  FOR_EACH_BB (bb)
-    {
-      rtx last_insn = BB_END (bb);
-      edge e;
-      edge_iterator ei;
-
-      if (! can_predict_insn_p (last_insn))
-	continue;
-
-      FOR_EACH_EDGE (e, ei, bb->succs)
-	{
-	  /* Predict early returns to be probable, as we've already taken
-	     care for error returns and other are often used for fast paths
-	     trought function.  */
-	  if ((e->dest == EXIT_BLOCK_PTR
-	       || (single_succ_p (e->dest)
-		   && single_succ (e->dest) == EXIT_BLOCK_PTR))
-	       && !predicted_by_p (bb, PRED_NULL_RETURN)
-	       && !predicted_by_p (bb, PRED_CONST_RETURN)
-	       && !predicted_by_p (bb, PRED_NEGATIVE_RETURN)
-	       && !last_basic_block_p (e->dest))
-	    predict_edge_def (e, PRED_EARLY_RETURN, TAKEN);
-
-	  /* Look for block we are guarding (i.e. we dominate it,
-	     but it doesn't postdominate us).  */
-	  if (e->dest != EXIT_BLOCK_PTR && e->dest != bb
-	      && dominated_by_p (CDI_DOMINATORS, e->dest, e->src)
-	      && !dominated_by_p (CDI_POST_DOMINATORS, e->src, e->dest))
-	    {
-	      rtx insn;
-
-	      /* The call heuristic claims that a guarded function call
-		 is improbable.  This is because such calls are often used
-		 to signal exceptional situations such as printing error
-		 messages.  */
-	      for (insn = BB_HEAD (e->dest); insn != NEXT_INSN (BB_END (e->dest));
-		   insn = NEXT_INSN (insn))
-		if (CALL_P (insn)
-		    /* Constant and pure calls are hardly used to signalize
-		       something exceptional.  */
-		    && ! CONST_OR_PURE_CALL_P (insn))
-		  {
-		    predict_edge_def (e, PRED_CALL, NOT_TAKEN);
-		    break;
-		  }
-	    }
-	}
-      bb_estimate_probability_locally (bb);
-    }
-
-  /* Attach the combined probability to each conditional jump.  */
-  FOR_EACH_BB (bb)
-    combine_predictions_for_insn (BB_END (bb), bb);
-
-  remove_fake_edges ();
-  estimate_bb_frequencies (loops_info);
-  free_dominance_info (CDI_POST_DOMINATORS);
-  if (profile_status == PROFILE_ABSENT)
-    profile_status = PROFILE_GUESSED;
-}
-
 /* Set edge->probability for each successor edge of BB.  */
 void
 guess_outgoing_edge_probabilities (basic_block bb)
Index: profile.c
===================================================================
--- profile.c	(revision 111977)
+++ profile.c	(working copy)
@@ -1336,46 +1336,3 @@ tree_register_profile_hooks (void)
   profile_hooks = &tree_profile_hooks;
 }
 
-
-/* Do branch profiling and static profile estimation passes.  */
-static unsigned int
-rest_of_handle_branch_prob (void)
-{
-  struct loops loops;
-
-  /* Discover and record the loop depth at the head of each basic
-     block.  The loop infrastructure does the real job for us.  */
-  flow_loops_find (&loops);
-
-  if (dump_file)
-    flow_loops_dump (&loops, dump_file, NULL, 0);
-
-  /* Estimate using heuristics if no profiling info is available.  */
-  if (flag_guess_branch_prob
-      && profile_status == PROFILE_ABSENT)
-    estimate_probability (&loops);
-
-  flow_loops_free (&loops);
-  free_dominance_info (CDI_DOMINATORS);
-  return 0;
-}
-
-struct tree_opt_pass pass_branch_prob =
-{
-  "bp",                                 /* name */
-  NULL,                                 /* gate */   
-  rest_of_handle_branch_prob,           /* execute */       
-  NULL,                                 /* sub */
-  NULL,                                 /* next */
-  0,                                    /* static_pass_number */
-  TV_BRANCH_PROB,                       /* tv_id */
-  0,                                    /* properties_required */
-  0,                                    /* properties_provided */
-  0,                                    /* properties_destroyed */
-  0,                                    /* todo_flags_start */
-  TODO_dump_func,                       /* todo_flags_finish */
-  'b'                                   /* letter */
-};
-
-
-
Index: basic-block.h
===================================================================
--- basic-block.h	(revision 111977)
+++ basic-block.h	(working copy)
@@ -869,7 +869,6 @@ extern struct edge_list *pre_edge_rev_lc
 extern void compute_available (sbitmap *, sbitmap *, sbitmap *, sbitmap *);
 
 /* In predict.c */
-extern void estimate_probability (struct loops *);
 extern void expected_value_to_br_prob (void);
 extern bool maybe_hot_bb_p (basic_block);
 extern bool probably_cold_bb_p (basic_block);
Index: passes.c
===================================================================
--- passes.c	(revision 111977)
+++ passes.c	(working copy)
@@ -216,10 +216,10 @@ finish_optimization_passes (void)
   timevar_push (TV_DUMP);
   if (profile_arc_flag || flag_test_coverage || flag_branch_probabilities)
     {
-      dump_file = dump_begin (pass_branch_prob.static_pass_number, NULL);
+      dump_file = dump_begin (pass_profile.static_pass_number, NULL);
       end_branch_prob ();
       if (dump_file)
-	dump_end (pass_branch_prob.static_pass_number, dump_file);
+	dump_end (pass_profile.static_pass_number, dump_file);
     }
 
   if (optimize > 0)
@@ -635,7 +635,6 @@ init_optimization_passes (void)
   NEXT_PASS (pass_gcse);
   NEXT_PASS (pass_jump_bypass);
   NEXT_PASS (pass_cfg);
-  NEXT_PASS (pass_branch_prob);
   NEXT_PASS (pass_rtl_ifcvt);
   NEXT_PASS (pass_tracer);
   /* Perform loop optimizations.  It might be better to do them a bit
Index: doc/invoke.texi
===================================================================
--- doc/invoke.texi	(revision 111977)
+++ doc/invoke.texi	(working copy)
@@ -3712,12 +3712,6 @@ letters for use in @var{letters} and @va
 @opindex dA
 Annotate the assembler output with miscellaneous debugging information.
 
-@item -db
-@itemx -fdump-rtl-bp
-@opindex db
-@opindex fdump-rtl-bp
-Dump after computing branch probabilities, to @file{@var{file}.09.bp}.
-
 @item -dB
 @itemx -fdump-rtl-bbro
 @opindex dB