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]
Other format: [Raw text]

[PATCH] PR23948 et al., take 2


This patch is a reissue of the tree-ssa-math-opts.c rewrite. The biggest changes are:

- I changed almost every occurrence of "divide" to "division".

- I include documentation about the algorithm and complexity at the top of the file.

- I check and assert that execute_cse_reciprocals_1's argument is a floating-point register.

- I dropped the first basic_block argument to insert_bb (which was not documented either), and improved the comments in that routine.

- I moved single_noncomplex_succ to tree-cfg.c.

- All the tests pass the new param correctly, if necessary.

Bootstrapped/regtested powerpc-apple-darwin8.3.0, ok for mainline?

Paolo
2006-01-07  Paolo Bonzini  <bonzini@gnu.org>

	PR tree-optimization/23109
	PR tree-optimization/23948
	PR tree-optimization/24123

	* Makefile.in (tree-ssa-math-opts.o): Adjust dependencies.
        * tree-cfg.c (single_noncomplex_succ): New.
        * tree-flow.h (single_noncomplex_succ): Declare it.
        * tree-ssa-math-opts.c (enum place_reciprocal): Remove.
        * tree-ssa-math-opts.c (enum place_reciprocal): Remove.
        (struct occurrence, occ_head, occ_pool, is_divide_by, compute_merit,
	insert_bb, register_division_in, insert_reciprocals,
	replace_reciprocal, free_bb): New.
        (execute_cse_reciprocals_1): Rewritten.
        (execute_cse_reciprocals): Adjust calls to execute_cse_reciprocals_1.
        Do not commit any edge insertion.  Always compute dominators and
        create the allocation pool.
        * params.def (PARAM_MIN_DIVISIONS_FOR_RECIP_MUL): New.
        * params.h (MIN_DIVISIONS_FOR_RECIP_MUL): New.
        * passes.c (init_optimization_passes): Run recip after tree loop
        optimizations.
        * doc/invoke.texi (--param): Add the min-divisions-for-recip-mul
        parameter.

2006-01-07  Paolo Bonzini  <bonzini@gnu.org>

	PR tree-optimization/23109
	PR tree-optimization/23948
	PR tree-optimization/24123

	* gcc.dg/tree-ssa/recip-3.c, gcc.dg/tree-ssa/recip-4.c,
	gcc.dg/tree-ssa/recip-5.c, gcc.dg/tree-ssa/recip-6.c,
	gcc.dg/tree-ssa/recip-7.c, gcc.dg/tree-ssa/pr23109.c,
	g++.dg/tree-ssa/pr23948.C: New testcases.
	* gcc.dg/tree-ssa/pr23234.c: Add --param min-divisions-for-recip-mul=2.

Index: Makefile.in
===================================================================
--- Makefile.in	(revision 109374)
+++ Makefile.in	(working copy)
@@ -1969,7 +1969,8 @@ tree-ssa-loop-im.o : tree-ssa-loop-im.c 
    $(TREE_DUMP_H) tree-pass.h $(FLAGS_H) real.h $(BASIC_BLOCK_H) \
    hard-reg-set.h
 tree-ssa-math-opts.o : tree-ssa-math-opts.c $(TREE_FLOW_H) $(CONFIG_H) \
-   $(SYSTEM_H) $(TREE_H) $(TIMEVAR_H) tree-pass.h $(TM_H) $(FLAGS_H)
+   $(SYSTEM_H) $(TREE_H) $(TIMEVAR_H) tree-pass.h $(TM_H) $(FLAGS_H) \
+   alloc-pool.h $(BASIC_BLOCK_H) $(PARAMS_H)
 tree-ssa-alias.o : tree-ssa-alias.c $(TREE_FLOW_H) $(CONFIG_H) $(SYSTEM_H) \
    $(RTL_H) $(TREE_H) $(TM_P_H) $(EXPR_H) $(GGC_H) tree-inline.h $(FLAGS_H) \
    function.h $(TIMEVAR_H) convert.h $(TM_H) coretypes.h langhooks.h \
Index: tree-ssa-math-opts.c
===================================================================
--- tree-ssa-math-opts.c	(revision 109374)
+++ tree-ssa-math-opts.c	(working copy)
@@ -35,7 +35,55 @@ Software Foundation, 51 Franklin Street,
 	z = z * rmodulus;
 
    We do this for loop invariant divisors, and with this pass whenever
-   we notice that a division has the same divisor multiple times.  */
+   we notice that a division has the same divisor multiple times.
+
+   Of course, like in PRE, we don't insert a division if a dominator
+   already has one.  However, this cannot be done as an extension of
+   PRE for several reasons.
+
+   First of all, with some experiments it was found out that the
+   transformation is not always useful if there are only two divisions
+   hy the same divisor.  This is probably because modern processors
+   can pipeline the divisions; on older, in-order processors it should
+   still be effective to optimize two divisions by the same number.
+   We make this a param, and it shall be called N in the remainder of
+   this comment.
+
+   Second, if trapping math is active, we have less freedom on where
+   to insert divisions: we can only do so in basic blocks that already
+   contain one.  (If divisions don't trap, instead, we can insert
+   divisions elsewhere, which will be in blocks that are common dominators
+   of those that have the division).
+
+   We really don't want to compute the reciprocal unless a division will
+   be found.  To do this, we won't insert the division in a basic block
+   that has less than N divisions *post-dominating* it.
+
+   The algorithm constructs a subset of the dominator tree, holding the
+   blocks containing the divisions and the common dominators to them,
+   and walk it twice.  The first walk is in post-order, and it annotates
+   each block with the number of divisions that post-dominate it: this
+   gives information on where divisions can be inserted profitably.
+   The second walk is in pre-order, and it inserts divisions as explained
+   above, and replaces divisions by multiplications.
+
+   In the best case, the cost of the pass is O(n_statements).  In the
+   worst-case, the cost is due to creating the dominator tree subset,
+   with a cost of O(n_basic_blocks ^ 2); however this can only happen
+   for n_statements / n_basic_blocks statements.  So, the amortized cost
+   of creating the dominator tree subset is O(n_basic_blocks) and the
+   worst-case cost of the pass is O(n_statements * n_basic_blocks).
+
+   More practically, the cost will be small because there are few
+   divisions, and they tend to be in the same basic block, so insert_bb
+   is called very few times.
+
+   If we did this using domwalk.c, an efficient implementation would have
+   to work on all the variables in a single pass, because we could not
+   work on just a subset of the dominator tree, as we do now, and the
+   cost would also be something like O(n_statements * n_basic_blocks).
+   The data structures would be more complex in order to work on all the
+   variables in a single pass.  */
 
 #include "config.h"
 #include "system.h"
@@ -47,142 +95,406 @@ Software Foundation, 51 Franklin Street,
 #include "real.h"
 #include "timevar.h"
 #include "tree-pass.h"
+#include "alloc-pool.h"
+#include "basic-block.h"
+#include "params.h"
 
-static bool
-gate_cse_reciprocals (void)
+
+/* This structure represents one basic block that either computes a
+   division, or is a common dominator for basic block that compute a
+   division.  */
+struct occurrence {
+  /* The basic block represented by this structure.  */
+  basic_block bb;
+
+  /* If non-NULL, the SSA_NAME holding the definition for a reciprocal
+     inserted in BB.  */
+  tree recip_def;
+
+  /* If non-NULL, the MODIFY_EXPR for a reciprocal computation that
+     was inserted in BB.  */
+  tree recip_def_stmt;
+
+  /* Pointer to a list of "struct occurrence"s for blocks dominated
+     by BB.  */
+  struct occurrence *children;
+
+  /* Pointer to the next "struct occurrence"s in the list of blocks
+     sharing a common dominator.  */
+  struct occurrence *next;
+
+  /* The number of divisions that are in BB before compute_merit.  The
+     number of divisions that are in BB or post-dominate it after
+     compute_merit.  */
+  int num_divisions;
+
+  /* True if the basic block has a division, false if it is a common
+     dominator for basic blocks that do.  If it is false and trapping
+     math is active, BB is not a candidate for inserting a reciprocal.  */
+  bool bb_has_division;
+};
+
+
+/* The instance of "struct occurrence" representing the highest
+   interesting block in the dominator tree.  */
+static struct occurrence *occ_head;
+
+/* Allocation pool for getting instances of "struct occurrence".  */
+static alloc_pool occ_pool;
+
+
+
+/* Allocate and return a new struct occurrence for basic block BB, and
+   whose children list is headed by CHILDREN.  */
+static struct occurrence *
+occ_new (basic_block bb, struct occurrence *children)
 {
-  return optimize && !optimize_size && flag_unsafe_math_optimizations;
+  struct occurrence *occ;
+
+  occ = bb->aux = pool_alloc (occ_pool);
+  memset (occ, 0, sizeof (struct occurrence));
+
+  occ->bb = bb;
+  occ->children = children;
+  return occ;
 }
 
-/* Where to put the statement computing a reciprocal.  */
-enum place_reciprocal
+
+/* Insert NEW_OCC into our subset of the dominator tree.  P_HEAD points to a
+   list of "struct occurrence"s, one per basic block, having IDOM as
+   their common dominator.
+
+   We try to insert NEW_OCC as deep as possible in the tree, and we also
+   insert any other block that is a common dominator for BB and one
+   block already in the tree.  */
+
+static void
+insert_bb (struct occurrence *new_occ, basic_block idom,
+	   struct occurrence **p_head)
 {
-  PR_BEFORE_BSI,	/* Put it using bsi_insert_before.  */
-  PR_AFTER_BSI,		/* Put it using bsi_insert_after.  */
-  PR_ON_ENTRY_EDGE	/* Put it on the edge between the entry
-			   and the first basic block.  */
-};
+  struct occurrence *occ, **p_occ;
+
+  for (p_occ = p_head; (occ = *p_occ) != NULL; )
+    {
+      basic_block bb = new_occ->bb, occ_bb = occ->bb;
+      basic_block dom = nearest_common_dominator (CDI_DOMINATORS, occ_bb, bb);
+      if (dom == bb)
+	{
+	  /* BB dominates OCC_BB.  OCC becomes NEW_OCC's child: remove OCC
+	     from its list.  */
+	  *p_occ = occ->next;
+	  occ->next = new_occ->children;
+	  new_occ->children = occ;
+
+	  /* Try the next block (it may as well be dominated by BB).  */
+	}
+
+      else if (dom == occ_bb)
+	{
+	  /* OCC_BB dominates BB.  Tail recurse to look deeper.  */
+	  insert_bb (new_occ, dom, &occ->children);
+	  return;
+	}
+
+      else if (dom != idom)
+	{
+	  gcc_assert (!dom->aux);
+
+	  /* There is a dominator between IDOM and BB, add it and make
+	     two children out of NEW_OCC and OCC.  First, remove OCC from
+	     its list and construct the new list of children.  */
+	  *p_occ = occ->next;
+	  new_occ->next = occ;
+	  occ->next = NULL;
+
+	  /* None of the previous blocks has DOM as a dominator: if we tail
+	     recursed, we would reexamine them uselessly.  Just create a new
+	     struct for DOM, and go on looking for blocks dominated by DOM.  */
+          new_occ = occ_new (dom, new_occ);
+	}
+
+      else
+	{
+	  /* Nothing special, go on with the next element.  */
+	  p_occ = &occ->next;
+	}
+    }
+
+  /* No place was found as a child of IDOM.  Make BB a sibling of IDOM.  */
+  new_occ->next = *p_head;
+  *p_head = new_occ;
+}
+
+/* Register that we found a division in BB.  */
+
+static inline void
+register_division_in (basic_block bb)
+{
+  struct occurrence *occ;
+
+  occ = (struct occurrence *) bb->aux;
+  if (!occ)
+    {
+      occ = occ_new (bb, NULL);
+      insert_bb (occ, ENTRY_BLOCK_PTR, &occ_head);
+    }
+
+  occ->bb_has_division = true;
+  occ->num_divisions++;
+}
+
+
+/* Compute the number of divisions that postdominate each block in OCC and
+   its children.  */
 
-/* Check if DEF's uses include more than one floating-point division,
-   and if so replace them by multiplications with the reciprocal.  Add
-   the statement computing the reciprocal according to WHERE.
-
-   Does not check the type of DEF, nor that DEF is a GIMPLE register.
-   This is done in the caller for speed, because otherwise this routine
-   would be called for every definition and phi node.  */
 static void
-execute_cse_reciprocals_1 (block_stmt_iterator *bsi, tree def,
-			   enum place_reciprocal where)
+compute_merit (struct occurrence *occ)
 {
-  use_operand_p use_p;
-  imm_use_iterator use_iter;
-  tree t, new_stmt, type;
-  int count = 0;
-  bool ok = !flag_trapping_math;
+  struct occurrence *occ_child;
+  basic_block dom = occ->bb;
 
-  /* Find uses.  */
-  FOR_EACH_IMM_USE_FAST (use_p, use_iter, def)
+  for (occ_child = occ->children; occ_child; occ_child = occ_child->next)
     {
-      tree use_stmt = USE_STMT (use_p);
-      if (TREE_CODE (use_stmt) == MODIFY_EXPR
-	  && TREE_CODE (TREE_OPERAND (use_stmt, 1)) == RDIV_EXPR
-	  && TREE_OPERAND (TREE_OPERAND (use_stmt, 1), 1) == def)
+      basic_block bb;
+      if (occ_child->children)
+        compute_merit (occ_child);
+
+      if (flag_exceptions)
+	bb = single_noncomplex_succ (dom);
+      else
+	bb = dom;
+
+      if (dominated_by_p (CDI_POST_DOMINATORS, bb, occ_child->bb))
+        occ->num_divisions += occ_child->num_divisions;
+    }
+}
+
+
+/* Return whether USE_STMT is a floating-point division by DEF.  */
+static inline bool
+is_division_by (tree use_stmt, tree def)
+{
+  return TREE_CODE (use_stmt) == MODIFY_EXPR
+	 && TREE_CODE (TREE_OPERAND (use_stmt, 1)) == RDIV_EXPR
+	 && TREE_OPERAND (TREE_OPERAND (use_stmt, 1), 1) == def;
+}
+
+/* Walk the subset of the dominator tree rooted at OCC, setting the
+   RECIP_DEF field to a definition of 1.0 / DEF that can be used in
+   the given basic block.  The field may be left NULL, of course,
+   if it is not possible or profitable to do the optimization.
+
+   DEF_BSI is an iterator pointing at the statement defining DEF.
+   If RECIP_DEF is set, a dominator already has a computation that can
+   be used.  */
+
+static void
+insert_reciprocals (block_stmt_iterator *def_bsi, struct occurrence *occ,
+		    tree def, tree recip_def)
+{
+  tree type, new_stmt;
+  block_stmt_iterator bsi;
+  struct occurrence *occ_child;
+
+  if (!recip_def
+      && (occ->bb_has_division || !flag_trapping_math)
+      && occ->num_divisions >= MIN_DIVISIONS_FOR_RECIP_MUL)
+    {
+      /* Make a variable with the replacement and substitute it.  */
+      type = TREE_TYPE (def);
+      recip_def = make_rename_temp (type, "reciptmp");
+      new_stmt = build2 (MODIFY_EXPR, void_type_node, recip_def,
+		         fold_build2 (RDIV_EXPR, type,
+				      build_real (type, dconst1), def));
+  
+  
+      if (occ->bb_has_division)
+        {
+          /* Case 1: insert before an existing division.  */
+          bsi = bsi_after_labels (occ->bb);
+          while (!bsi_end_p (bsi) && !is_division_by (bsi_stmt (bsi), def))
+	    bsi_next (&bsi);
+
+          bsi_insert_before (&bsi, new_stmt, BSI_SAME_STMT);
+        }
+      else if (def_bsi && occ->bb == def_bsi->bb)
         {
-          ++count;
-          /* Check if this use post-dominates the insertion point.  */
-          if (ok || dominated_by_p (CDI_POST_DOMINATORS, bsi->bb,
-				    bb_for_stmt (use_stmt)))
-	    ok = true;
+          /* Case 2: insert right after the definition.  Note that this will
+	     never happen if the definition statement can throw, because in
+	     that case the sole successor of the statement's basic block will
+	     dominate all the uses as well.  */
+          bsi_insert_after (def_bsi, new_stmt, BSI_NEW_STMT);
         }
-      if (count >= 2 && ok)
-        break;
+      else
+        {
+          /* Case 3: insert in a basic block not containing defs/uses.  */
+          bsi = bsi_after_labels (occ->bb);
+          bsi_insert_before (&bsi, new_stmt, BSI_SAME_STMT);
+        }
+
+      occ->recip_def_stmt = new_stmt;
     }
 
-  if (count < 2 || !ok)
-    return;
+  occ->recip_def = recip_def;
+  for (occ_child = occ->children; occ_child; occ_child = occ_child->next)
+    insert_reciprocals (def_bsi, occ_child, def, recip_def);
+}
+
+
+/* Replace the division at USE_P with a multiplication by the reciprocal, if
+   possible.  */
+
+static inline void
+replace_reciprocal (use_operand_p use_p)
+{
+  tree use_stmt = USE_STMT (use_p);
+  basic_block bb = bb_for_stmt (use_stmt);
+  struct occurrence *occ = (struct occurrence *) bb->aux;
 
-  /* Make a variable with the replacement and substitute it.  */
-  type = TREE_TYPE (def);
-  t = make_rename_temp (type, "reciptmp");
-  new_stmt = build2 (MODIFY_EXPR, void_type_node, t,
-		     fold_build2 (RDIV_EXPR, type, build_real (type, dconst1),
-				  def));
-
-  if (where == PR_BEFORE_BSI)
-    bsi_insert_before (bsi, new_stmt, BSI_SAME_STMT);
-  else if (where == PR_AFTER_BSI)
-    bsi_insert_after (bsi, new_stmt, BSI_NEW_STMT);
-  else if (where == PR_ON_ENTRY_EDGE)
-    bsi_insert_on_edge (single_succ_edge (ENTRY_BLOCK_PTR), new_stmt);
+  if (occ->recip_def && use_stmt != occ->recip_def_stmt)
+    {
+      TREE_SET_CODE (TREE_OPERAND (use_stmt, 1), MULT_EXPR);
+      SET_USE (use_p, occ->recip_def);
+      fold_stmt_inplace (use_stmt);
+      update_stmt (use_stmt);
+    }
+}
+
+
+/* Free OCC and return one more "struct occurrence" to be freed.  */
+
+static struct occurrence *
+free_bb (struct occurrence *occ)
+{
+  struct occurrence *child, *next;
+
+  /* First get the two pointers hanging off OCC.  */
+  next = occ->next;
+  child = occ->children;
+  occ->bb->aux = NULL;
+  pool_free (occ_pool, occ);
+
+  /* Now ensure that we don't recurse unless it is necessary.  */
+  if (!child)
+    return next;
   else
-    gcc_unreachable ();
+    {
+      while (next)
+	next = free_bb (next);
+
+      return child;
+    }
+}
+
+
+/* Look for floating-point divisions among DEF's uses, and try to
+   replace them by multiplications with the reciprocal.  Add
+   as many statements computing the reciprocal as needed.
+
+   DEF must be a GIMPLE register of a floating-point type.  */
 
-  FOR_EACH_IMM_USE_SAFE (use_p, use_iter, def)
+static void
+execute_cse_reciprocals_1 (block_stmt_iterator *def_bsi, tree def)
+{
+  use_operand_p use_p;
+  imm_use_iterator use_iter;
+  struct occurrence *occ;
+  int count = 0;
+
+  gcc_assert (FLOAT_TYPE_P (TREE_TYPE (def)) && is_gimple_reg (def));
+
+  FOR_EACH_IMM_USE_FAST (use_p, use_iter, def)
     {
       tree use_stmt = USE_STMT (use_p);
-      if (use_stmt != new_stmt
-	  && TREE_CODE (use_stmt) == MODIFY_EXPR
-	  && TREE_CODE (TREE_OPERAND (use_stmt, 1)) == RDIV_EXPR
-	  && TREE_OPERAND (TREE_OPERAND (use_stmt, 1), 1) == def)
+      if (is_division_by (use_stmt, def))
+	{
+	  register_division_in (bb_for_stmt (use_stmt));
+	  count++;
+	}
+    }
+  
+  /* Do the expensive part only if we can hope to optimize something.  */
+  if (count >= MIN_DIVISIONS_FOR_RECIP_MUL)
+    {
+      for (occ = occ_head; occ; occ = occ->next)
+	{
+	  compute_merit (occ);
+	  insert_reciprocals (def_bsi, occ, def, NULL);
+	}
+
+      FOR_EACH_IMM_USE_SAFE (use_p, use_iter, def)
 	{
-	  TREE_SET_CODE (TREE_OPERAND (use_stmt, 1), MULT_EXPR);
-	  SET_USE (use_p, t);
+	  tree use_stmt = USE_STMT (use_p);
+	  if (is_division_by (use_stmt, def))
+	    replace_reciprocal (use_p);
 	}
     }
+
+  for (occ = occ_head; occ; )
+    occ = free_bb (occ);
+
+  occ_head = NULL;
 }
 
+
+static bool
+gate_cse_reciprocals (void)
+{
+  return optimize && !optimize_size && flag_unsafe_math_optimizations;
+}
+
+
+/* Go through all the floating-point SSA_NAMEs, and call
+   execute_cse_reciprocals_1 on each of them.  */
 static void
 execute_cse_reciprocals (void)
 {
   basic_block bb;
   tree arg;
 
-  if (flag_trapping_math)
-    calculate_dominance_info (CDI_POST_DOMINATORS);
+  occ_pool = create_alloc_pool ("dominators for recip",
+				sizeof (struct occurrence),
+				n_basic_blocks / 3 + 1);
 
-  if (single_succ_p (ENTRY_BLOCK_PTR))
-    for (arg = DECL_ARGUMENTS (cfun->decl); arg; arg = TREE_CHAIN (arg))
-      if (default_def (arg))
-	{
-	  block_stmt_iterator bsi;
-	  bsi = bsi_start (single_succ (ENTRY_BLOCK_PTR));
-	  execute_cse_reciprocals_1 (&bsi, default_def (arg),
-				     PR_ON_ENTRY_EDGE);
-	}
+  calculate_dominance_info (CDI_DOMINATORS | CDI_POST_DOMINATORS);
+
+#ifdef ENABLE_CHECKING
+  FOR_EACH_BB (bb)
+    gcc_assert (!bb->aux);
+#endif
+
+  for (arg = DECL_ARGUMENTS (cfun->decl); arg; arg = TREE_CHAIN (arg))
+    if (default_def (arg)
+	&& FLOAT_TYPE_P (TREE_TYPE (arg))
+	&& is_gimple_reg (arg))
+      execute_cse_reciprocals_1 (NULL, default_def (arg));
 
   FOR_EACH_BB (bb)
     {
       block_stmt_iterator bsi;
       tree phi, def;
-      for (bsi = bsi_start (bb);
-	   !bsi_end_p (bsi) && TREE_CODE (bsi_stmt (bsi)) == LABEL_EXPR;
-	   bsi_next (&bsi))
-        ;
 
       for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
 	{
 	  def = PHI_RESULT (phi);
 	  if (FLOAT_TYPE_P (TREE_TYPE (def))
 	      && is_gimple_reg (def))
-	    execute_cse_reciprocals_1 (&bsi, def, PR_BEFORE_BSI);
+	    execute_cse_reciprocals_1 (NULL, def);
 	}
 
-      for (; !bsi_end_p (bsi); bsi_next (&bsi))
+      for (bsi = bsi_after_labels (bb); !bsi_end_p (bsi); bsi_next (&bsi))
         {
 	  tree stmt = bsi_stmt (bsi);
 	  if (TREE_CODE (stmt) == MODIFY_EXPR
 	      && (def = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_DEF)) != NULL
 	      && FLOAT_TYPE_P (TREE_TYPE (def))
 	      && TREE_CODE (def) == SSA_NAME)
-	    execute_cse_reciprocals_1 (&bsi, def, PR_AFTER_BSI);
+	    execute_cse_reciprocals_1 (&bsi, def);
 	}
     }
 
-  if (flag_trapping_math)
-    free_dominance_info (CDI_POST_DOMINATORS);
-  
-  if (single_succ_p (ENTRY_BLOCK_PTR))
-    bsi_commit_one_edge_insert (single_succ_edge (ENTRY_BLOCK_PTR), NULL);
+  free_dominance_info (CDI_DOMINATORS | CDI_POST_DOMINATORS);
+  free_alloc_pool (occ_pool);
 }
 
 struct tree_opt_pass pass_cse_reciprocals =
Index: params.def
===================================================================
--- params.def	(revision 109374)
+++ params.def	(working copy)
@@ -202,7 +202,7 @@ DEFPARAM(PARAM_MAX_GCSE_MEMORY,
 DEFPARAM(PARAM_MAX_GCSE_PASSES,
 	"max-gcse-passes",
 	"The maximum number of passes to make when doing GCSE",
-	1, 1, 0)
+	2, 1, 0)
 /* This is the threshold ratio when to perform partial redundancy
    elimination after reload. We perform partial redundancy elimination
    when the following holds:
@@ -552,6 +552,11 @@ DEFPARAM (PARAM_MAX_JUMP_THREAD_DUPLICAT
           "Maximum number of statements allowed in a block that needs to be duplicated when threading jumps",
 	  15, 0, 0)
    
+DEFPARAM (PARAM_MIN_DIVISIONS_FOR_RECIP_MUL,
+	  "min-divisions-for-recip-mul",
+	  "The minimum numbers of divisions with a common divisor, for which it is profitable to instead multiply by the reciprocal",
+	  3, 2, 0)
+
 /*
 Local variables:
 mode:c
Index: params.h
===================================================================
--- params.h	(revision 109374)
+++ params.h	(working copy)
@@ -145,4 +145,6 @@ typedef enum compiler_param
   PARAM_VALUE (PARAM_MIN_VIRTUAL_MAPPINGS)
 #define VIRTUAL_MAPPINGS_TO_SYMS_RATIO \
   PARAM_VALUE (PARAM_VIRTUAL_MAPPINGS_TO_SYMS_RATIO)
+#define MIN_DIVISIONS_FOR_RECIP_MUL \
+  PARAM_VALUE (PARAM_MIN_DIVISIONS_FOR_RECIP_MUL)
 #endif /* ! GCC_PARAMS_H */
Index: passes.c
===================================================================
--- passes.c	(revision 109374)
+++ passes.c	(working copy)
@@ -551,12 +529,12 @@ init_optimization_passes (void)
      we add may_alias right after fold builtins
      which can create arbitrary GIMPLE.  */
   NEXT_PASS (pass_may_alias);
-  NEXT_PASS (pass_cse_reciprocals);
   NEXT_PASS (pass_split_crit_edges);
   NEXT_PASS (pass_pre);
   NEXT_PASS (pass_may_alias);
   NEXT_PASS (pass_sink_code);
   NEXT_PASS (pass_tree_loop);
+  NEXT_PASS (pass_cse_reciprocals);
   NEXT_PASS (pass_reassoc);
   NEXT_PASS (pass_dominator);
 
Index: tree-cfg.c
===================================================================
--- tree-cfg.c	(revision 109374)
+++ tree-cfg.c	(working copy)
@@ -1389,6 +1389,30 @@ tree_merge_blocks (basic_block a, basic_
 }
 
 
+/* Return the one of two successors of BB that is not reachable by a
+   reached by a complex edge, if there is one.  Else, return BB.  We use
+   this in optimizations that use post-dominators for their heuristics,
+   to catch the cases in C++ where function calls are involved.  */
+    
+basic_block
+single_noncomplex_succ (basic_block bb)  
+{
+  edge e0, e1;
+  if (EDGE_COUNT (bb->succs) != 2)
+    return bb;
+   
+  e0 = EDGE_SUCC (bb, 0);
+  e1 = EDGE_SUCC (bb, 1);
+  if (e0->flags & EDGE_COMPLEX)
+    return e1->dest;
+  if (e1->flags & EDGE_COMPLEX)
+    return e0->dest;
+   
+  return bb;
+}       
+        
+
+
 /* Walk the function tree removing unnecessary statements.
 
      * Empty statement nodes are removed
Index: tree-flow.h
===================================================================
--- tree-flow.h	(revision 109374)
+++ tree-flow.h	(working copy)
@@ -487,6 +487,7 @@ extern bool is_ctrl_stmt (tree);
 extern bool is_ctrl_altering_stmt (tree);
 extern bool computed_goto_p (tree);
 extern bool simple_goto_p (tree);
+extern basic_block single_noncomplex_succ (basic_block bb);
 extern void tree_dump_bb (basic_block, FILE *, int);
 extern void debug_tree_bb (basic_block);
 extern basic_block debug_tree_bb_n (int);
Index: doc/invoke.texi
===================================================================
--- doc/invoke.texi	(revision 109374)
+++ doc/invoke.texi	(working copy)
@@ -6144,6 +6144,14 @@ protection when @option{-fstack-protecti
 @item max-jump-thread-duplication-stmts
 Maximum number of statements allowed in a block that needs to be
 duplicated when threading jumps.
+
+@item min-divisions-for-recip-mul
+When @option{-ffast-math} is in effect, GCC tries to optimize
+divisions by the same divisor, by turning them into multiplications by
+the reciprocal.  This parameter sets the minimum number of divisions
+that should be there for GCC to perform the optimization.  The default
+is 3.
+
 @end table
 @end table
 
Index: testsuite/g++.dg/tree-ssa/pr23948.C
===================================================================
RCS file: testsuite/g++.dg/tree-ssa/pr23948.C
diff -N testsuite/g++.dg/tree-ssa/pr23948.C
--- /dev/null	1 Jan 1970 00:00:00 -0000
+++ testsuite/g++.dg/tree-ssa/pr23948.C	12 Oct 2005 10:59:27 -0000
@@ -0,0 +1,18 @@
+/* { dg-options "-O1 -ffast-math -fdump-tree-recip --param min-divisions-for-recip-mul=2" } */
+/* { dg-do compile } */
+
+struct MIOFILE {
+  ~MIOFILE();
+};
+double potentially_runnable_resource_share();
+void f1(double);
+int make_scheduler_request(double a)
+{
+  MIOFILE mf;
+  double prrs = potentially_runnable_resource_share();
+  f1(a/prrs);
+  f1(1/prrs);
+}
+
+/* { dg-final { scan-tree-dump-times " / " 1 "recip" } } */
+/* { dg-final { cleanup-tree-dump "recip" } } */
Index: testsuite/gcc.dg/tree-ssa/pr23109.c
===================================================================
RCS file: testsuite/gcc.dg/tree-ssa/pr23109.c
diff -N testsuite/gcc.dg/tree-ssa/pr23109.c
--- /dev/null	1 Jan 1970 00:00:00 -0000
+++ testsuite/gcc.dg/tree-ssa/pr23109.c	12 Oct 2005 10:59:28 -0000
@@ -0,0 +1,34 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -funsafe-math-optimizations -fdump-tree-recip" } */
+
+double F[2] = { 0., 0. }, e = 0.;
+
+int main()
+{
+	int i;
+	double E, W, P, d;
+
+        /* make sure the program crashes on FP exception */
+        unsigned short int Mask;
+
+	W = 1.;
+	d = 2.*e;
+	E = 1. - d;
+
+	for( i=0; i < 2; i++ )
+		if( d > 0.01 )
+		{
+			P = ( W < E ) ? (W - E)/d : (E - W)/d;
+			F[i] += P;
+		}
+
+	return 0;
+}
+
+/* LIM only performs the transformation in the no-trapping-math case.  In
+   the future we will do it for trapping-math as well in recip, check that
+   this is not wrongly optimized.  */
+/* { dg-final { scan-tree-dump-not "reciptmp" "lim" } } */
+/* { dg-final { scan-tree-dump-not "reciptmp" "recip" } } */
+/* { dg-final { cleanup-tree-dump "recip" } } */
+
Index: testsuite/gcc.dg/tree-ssa/pr23234.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/testsuite/gcc.dg/tree-ssa/pr23234.c,v
retrieving revision 1.1
diff -p -u -u -r1.1 pr23234.c
--- testsuite/gcc.dg/tree-ssa/pr23234.c	9 Aug 2005 03:28:38 -0000	1.1
+++ testsuite/gcc.dg/tree-ssa/pr23234.c	12 Oct 2005 10:59:28 -0000
@@ -1,7 +1,7 @@
 /* The problem in this PR was mostly finding a suitable place to insert
    the reciprocals of the function arguments.  This test case tries to
    test three possible ways of how this may go wrong.  */
-/* { dg-options "-O2 -ffast-math" } */
+/* { dg-options "-O2 -ffast-math --param min-divisions-for-recip-mul=2" } */
 /* { dg-do compile } */
 
 /* The original test case.  */
Index: testsuite/gcc.dg/tree-ssa/recip-2.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/testsuite/gcc.dg/tree-ssa/recip-2.c,v
retrieving revision 1.1
diff -p -u -u -r1.1 recip-2.c
--- testsuite/gcc.dg/tree-ssa/recip-2.c	17 May 2005 09:55:44 -0000	1.1
+++ testsuite/gcc.dg/tree-ssa/recip-2.c	12 Oct 2005 10:59:28 -0000
@@ -1,5 +1,5 @@
 /* { dg-do compile } */
-/* { dg-options "-O1 -funsafe-math-optimizations -fdump-tree-recip" } */
+/* { dg-options "-O1 -funsafe-math-optimizations -fdump-tree-recip --param min-divisions-for-recip-mul=2" } */
 
 float e(float a, float b, float c, float d, float e, float f)
 {
Index: testsuite/gcc.dg/tree-ssa/recip-3.c
===================================================================
RCS file: testsuite/gcc.dg/tree-ssa/recip-3.c
diff -N testsuite/gcc.dg/tree-ssa/recip-3.c
--- /dev/null	1 Jan 1970 00:00:00 -0000
+++ testsuite/gcc.dg/tree-ssa/recip-3.c	12 Oct 2005 10:59:28 -0000
@@ -0,0 +1,27 @@
+/* { dg-do compile } */
+/* { dg-options "-O1 -fno-trapping-math -funsafe-math-optimizations -fdump-tree-recip" } */
+
+double F[2] = { 0.0, 0.0 }, e;
+
+/* In this case the optimization is interesting.  */
+float h ()
+{
+	int i;
+	double E, W, P, d;
+
+	W = 1.;
+	d = 2.*e;
+	E = 1. - d;
+
+	for( i=0; i < 2; i++ )
+		if( d > 0.01 )
+		{
+			P = ( W < E ) ? (W - E)/d : (E - W)/d;
+			F[i] += P;
+		}
+
+	F[0] += E / d;
+}
+
+/* { dg-final { scan-tree-dump-times " / " 1 "recip" } } */
+/* { dg-final { cleanup-tree-dump "recip" } } */
Index: testsuite/gcc.dg/tree-ssa/recip-4.c
===================================================================
RCS file: testsuite/gcc.dg/tree-ssa/recip-4.c
diff -N testsuite/gcc.dg/tree-ssa/recip-4.c
--- /dev/null	1 Jan 1970 00:00:00 -0000
+++ testsuite/gcc.dg/tree-ssa/recip-4.c	12 Oct 2005 10:59:28 -0000
@@ -0,0 +1,45 @@
+/* { dg-do compile } */
+/* { dg-options "-O1 -fno-trapping-math -funsafe-math-optimizations -fdump-tree-recip" } */
+
+/* based on the test case in pr23109 */
+
+double F[2] = { 0., 0. }, e = 0.;
+
+/* Nope, we cannot prove the optimization is worthwhile in this case.  */
+void f ()
+{
+	int i;
+	double E, W, P, d;
+
+	W = 1.;
+	d = 2.*e;
+	E = 1. - d;
+
+	if( d > 0.01 )
+	{
+		P = ( W < E ) ? (W - E)/d : (E - W)/d;
+		F[i] += P;
+	}
+}
+
+/* We also cannot prove the optimization is worthwhile in this case.  */
+float g ()
+{
+	int i;
+	double E, W, P, d;
+
+	W = 1.;
+	d = 2.*e;
+	E = 1. - d;
+
+	if( d > 0.01 )
+	{
+		P = ( W < E ) ? (W - E)/d : (E - W)/d;
+		F[i] += P;
+	}
+
+	return 1.0 / d;
+}
+
+/* { dg-final { scan-tree-dump-not "reciptmp" "recip" } } */
+/* { dg-final { cleanup-tree-dump "recip" } } */
Index: testsuite/gcc.dg/tree-ssa/recip-5.c
===================================================================
RCS file: testsuite/gcc.dg/tree-ssa/recip-5.c
diff -N testsuite/gcc.dg/tree-ssa/recip-5.c
--- /dev/null	1 Jan 1970 00:00:00 -0000
+++ testsuite/gcc.dg/tree-ssa/recip-5.c	12 Oct 2005 10:59:28 -0000
@@ -0,0 +1,30 @@
+/* { dg-options "-O1 -funsafe-math-optimizations -ftrapping-math -fdump-tree-recip -fdump-tree-optimized --param min-divisions-for-recip-mul=2" } */
+/* { dg-do compile } */
+
+/* Test the reciprocal optimizations together with trapping math.  */
+
+extern int g();
+
+double f(double y, double z, double w)
+{
+  double b, c, d, e;
+
+  if (g ())
+    /* inserts one division here */
+    b = 1 / y, c = z / y;
+  else
+    /* one division here */
+    b = 3 / y, c = w / y;
+
+  /* and one here, that should be removed afterwards */
+  d = b / y;
+  e = c / y;
+
+  return b + c + d + e;
+}
+
+/* { dg-final { scan-tree-dump-times " / " 3 "recip" } } */
+/* { dg-final { scan-tree-dump-times " / " 2 "optimized" { xfail *-*-* } } } */
+/* { dg-final { cleanup-tree-dump "recip" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
+
Index: testsuite/gcc.dg/tree-ssa/recip-6.c
===================================================================
RCS file: testsuite/gcc.dg/tree-ssa/recip-6.c
diff -N testsuite/gcc.dg/tree-ssa/recip-6.c
--- /dev/null	1 Jan 1970 00:00:00 -0000
+++ testsuite/gcc.dg/tree-ssa/recip-6.c	12 Oct 2005 10:59:28 -0000
@@ -0,0 +1,25 @@
+/* { dg-options "-O1 -funsafe-math-optimizations -fno-trapping-math -fdump-tree-recip --param min-divisions-for-recip-mul=2" } */
+/* { dg-do compile } */
+
+/* Test inserting in a block that does not contain a division.  */
+
+extern int g();
+
+double f(double y, double z, double w)
+{
+  double b, c, d, e;
+
+  if (g ())
+    b = 1 / y, c = z / y;
+  else
+    b = 3 / y, c = w / y;
+
+  d = b / y;
+  e = c / y;
+
+  return b + c + d + e;
+}
+
+/* { dg-final { scan-tree-dump-times " / " 1 "recip" } } */
+/* { dg-final { cleanup-tree-dump "recip" } } */
+
Index: testsuite/gcc.dg/tree-ssa/recip-7.c
===================================================================
RCS file: testsuite/gcc.dg/tree-ssa/recip-7.c
diff -N testsuite/gcc.dg/tree-ssa/recip-7.c
--- /dev/null	1 Jan 1970 00:00:00 -0000
+++ testsuite/gcc.dg/tree-ssa/recip-7.c	12 Oct 2005 10:59:28 -0000
@@ -0,0 +1,26 @@
+/* { dg-options "-O1 -funsafe-math-optimizations -fno-trapping-math -fdump-tree-recip --param min-divisions-for-recip-mul=2" } */
+/* { dg-do compile } */
+
+/* Test inserting in a block that does not contain a division.  */
+
+extern double h();
+
+double f(int x, double z, double w)
+{
+  double b, c, d, e;
+  double y = h ();
+
+  if (x)
+    b = 1 / y, c = z / y;
+  else
+    b = 3 / y, c = w / y;
+
+  d = b / y;
+  e = c / y;
+
+  return b + c + d + e;
+}
+
+/* { dg-final { scan-tree-dump-times " / " 1 "recip" } } */
+/* { dg-final { cleanup-tree-dump "recip" } } */
+

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