Patchwork Back to forward-scan for combine in forwprop

login
register
mail settings
Submitter Richard Guenther
Date June 10, 2011, 11:37 a.m.
Message ID <alpine.LNX.2.00.1106101330540.810@zhemvz.fhfr.qr>
Download mbox | patch
Permalink /patch/99884/
State New
Headers show

Comments

Richard Guenther - June 10, 2011, 11:37 a.m.
Dunno why I thought a backward scan was preferable (it would scan
inserted stmts), but it clearly is better to first visit the
defs of the uses of stmts we are trying to combine.  The following
patch uses some trick to make sure stmts inserted by the combining
are processed as well.

Bootstrapped and tested on x86_64-unknown-linux-gnu, applied to trunk.

Richard.

2011-06-10  Richard Guenther  <rguenther@suse.de>

	* tree-ssa-forwprop.c (ssa_forward_propagate_and_combine):
	Scan stmts forward when combining, visit inserted stmts when
	a stmt was changed.

Patch

Index: gcc/tree-ssa-forwprop.c
===================================================================
--- gcc/tree-ssa-forwprop.c	(revision 174841)
+++ gcc/tree-ssa-forwprop.c	(working copy)
@@ -2212,7 +2212,8 @@  ssa_forward_propagate_and_combine (void)
 
   FOR_EACH_BB (bb)
     {
-      gimple_stmt_iterator gsi;
+      gimple_stmt_iterator gsi, prev;
+      bool prev_initialized;
 
       /* Apply forward propagation to all stmts in the basic-block.
 	 Note we update GSI within the loop as necessary.  */
@@ -2304,7 +2305,8 @@  ssa_forward_propagate_and_combine (void)
 
       /* Combine stmts with the stmts defining their operands.
 	 Note we update GSI within the loop as necessary.  */
-      for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi);)
+      prev_initialized = false;
+      for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi);)
 	{
 	  gimple stmt = gsi_stmt (gsi);
 	  bool changed = false;
@@ -2386,9 +2388,24 @@  ssa_forward_propagate_and_combine (void)
 	    default:;
 	    }
 
-	  /* If the stmt changed try combining it again.  */
-	  if (!changed)
-	    gsi_prev (&gsi);
+	  if (changed)
+	    {
+	      /* If the stmt changed then re-visit it and the statements
+		 inserted before it.  */
+	      if (!prev_initialized)
+		gsi = gsi_start_bb (bb);
+	      else
+		{
+		  gsi = prev;
+		  gsi_next (&gsi);
+		}
+	    }
+	  else
+	    {
+	      prev = gsi;
+	      prev_initialized = true;
+	      gsi_next (&gsi);
+	    }
 	}
     }