gimple-ssa-strength-reduction.c (struct incr_info_d): New struct.
authorBill Schmidt <wschmidt@linux.vnet.ibm.com>
Wed, 8 Aug 2012 01:35:22 +0000 (01:35 +0000)
committerWilliam Schmidt <wschmidt@gcc.gnu.org>
Wed, 8 Aug 2012 01:35:22 +0000 (01:35 +0000)
gcc:

2012-08-07  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>

* gimple-ssa-strength-reduction.c (struct incr_info_d): New struct.
(incr_vec): New static var.
(incr_vec_len): Likewise.
(address_arithmetic_p): Likewise.
(stmt_cost): Remove dead assignment.
(dump_incr_vec): New function.
(cand_abs_increment): Likewise.
(lazy_create_slsr_reg): Likewise.
(incr_vec_index): Likewise.
(count_candidates): Likewise.
(record_increment): Likewise.
(record_increments): Likewise.
(unreplaced_cand_in_tree): Likewise.
(optimize_cands_for_speed_p): Likewise.
(lowest_cost_path): Likewise.
(total_savings): Likewise.
(analyze_increments): Likewise.
(ncd_for_two_cands): Likewise.
(nearest_common_dominator_for_cands): Likewise.
(profitable_increment_p): Likewise.
(insert_initializers): Likewise.
(introduce_cast_before_cand): Likewise.
(replace_rhs_if_not_dup): Likewise.
(replace_one_candidate): Likewise.
(replace_profitable_candidates): Likewise.
(analyze_candidates_and_replace): Handle candidates with SSA-name
strides.

gcc/testsuite:

2012-08-07  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>

* gcc.dg/tree-ssa/slsr-5.c: New.
* gcc.dg/tree-ssa/slsr-6.c: New.
* gcc.dg/tree-ssa/slsr-7.c: New.
* gcc.dg/tree-ssa/slsr-8.c: New.
* gcc.dg/tree-ssa/slsr-9.c: New.
* gcc.dg/tree-ssa/slsr-10.c: New.
* gcc.dg/tree-ssa/slsr-11.c: New.
* gcc.dg/tree-ssa/slsr-12.c: New.
* gcc.dg/tree-ssa/slsr-13.c: New.
* gcc.dg/tree-ssa/slsr-14.c: New.
* gcc.dg/tree-ssa/slsr-15.c: New.
* gcc.dg/tree-ssa/slsr-16.c: New.
* gcc.dg/tree-ssa/slsr-17.c: New.
* gcc.dg/tree-ssa/slsr-18.c: New.
* gcc.dg/tree-ssa/slsr-19.c: New.
* gcc.dg/tree-ssa/slsr-20.c: New.
* gcc.dg/tree-ssa/slsr-21.c: New.
* gcc.dg/tree-ssa/slsr-22.c: New.
* gcc.dg/tree-ssa/slsr-23.c: New.
* gcc.dg/tree-ssa/slsr-24.c: New.
* gcc.dg/tree-ssa/slsr-25.c: New.
* gcc.dg/tree-ssa/slsr-26.c: New.
* gcc.dg/tree-ssa/slsr-30.c: New.
* gcc.dg/tree-ssa/slsr-31.c: New.

From-SVN: r190220

27 files changed:
gcc/ChangeLog
gcc/gimple-ssa-strength-reduction.c
gcc/testsuite/ChangeLog
gcc/testsuite/gcc.dg/tree-ssa/slsr-10.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-11.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-12.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-13.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-14.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-15.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-16.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-17.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-18.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-19.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-20.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-21.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-22.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-23.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-24.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-25.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-26.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-30.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-31.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-5.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-6.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-7.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-8.c [new file with mode: 0644]
gcc/testsuite/gcc.dg/tree-ssa/slsr-9.c [new file with mode: 0644]

index 1bffab7..33f692f 100644 (file)
@@ -1,3 +1,33 @@
+2012-08-07  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>
+
+       * gimple-ssa-strength-reduction.c (struct incr_info_d): New struct.
+       (incr_vec): New static var.
+       (incr_vec_len): Likewise.
+       (address_arithmetic_p): Likewise.
+       (stmt_cost): Remove dead assignment.
+       (dump_incr_vec): New function.
+       (cand_abs_increment): Likewise.
+       (lazy_create_slsr_reg): Likewise.
+       (incr_vec_index): Likewise.
+       (count_candidates): Likewise.
+       (record_increment): Likewise.
+       (record_increments): Likewise.
+       (unreplaced_cand_in_tree): Likewise.
+       (optimize_cands_for_speed_p): Likewise.
+       (lowest_cost_path): Likewise.
+       (total_savings): Likewise.
+       (analyze_increments): Likewise.
+       (ncd_for_two_cands): Likewise.
+       (nearest_common_dominator_for_cands): Likewise.
+       (profitable_increment_p): Likewise.
+       (insert_initializers): Likewise.
+       (introduce_cast_before_cand): Likewise.
+       (replace_rhs_if_not_dup): Likewise.
+       (replace_one_candidate): Likewise.
+       (replace_profitable_candidates): Likewise.
+       (analyze_candidates_and_replace): Handle candidates with SSA-name
+       strides.
+
 2012-08-07  Steven Bosscher  <steven@gcc.gnu.org>
 
        * cfg.c (debug_bb): Do not set TDF_BLOCKS.
index 6a46408..ccc361e 100644 (file)
@@ -30,8 +30,7 @@ along with GCC; see the file COPYING3.  If not see
    1) Explicit multiplies, known constant multipliers, no
       conditional increments. (complete)
    2) Explicit multiplies, unknown constant multipliers,
-      no conditional increments. (data gathering complete,
-      replacements pending)
+      no conditional increments. (complete)
    3) Implicit multiplies in addressing expressions. (complete)
    4) Explicit multiplies, conditional increments. (pending)
 
@@ -235,6 +234,41 @@ struct cand_chain_d
 typedef struct cand_chain_d cand_chain, *cand_chain_t;
 typedef const struct cand_chain_d *const_cand_chain_t;
 
+/* Information about a unique "increment" associated with candidates
+   having an SSA name for a stride.  An increment is the difference
+   between the index of the candidate and the index of its basis,
+   i.e., (i - i') as discussed in the module commentary.
+
+   When we are not going to generate address arithmetic we treat
+   increments that differ only in sign as the same, allowing sharing
+   of the cost of initializers.  The absolute value of the increment
+   is stored in the incr_info.  */
+
+struct incr_info_d
+{
+  /* The increment that relates a candidate to its basis.  */
+  double_int incr;
+
+  /* How many times the increment occurs in the candidate tree.  */
+  unsigned count;
+
+  /* Cost of replacing candidates using this increment.  Negative and
+     zero costs indicate replacement should be performed.  */
+  int cost;
+
+  /* If this increment is profitable but is not -1, 0, or 1, it requires
+     an initializer T_0 = stride * incr to be found or introduced in the
+     nearest common dominator of all candidates.  This field holds T_0
+     for subsequent use.  */
+  tree initializer;
+
+  /* If the initializer was found to already exist, this is the block
+     where it was found.  */
+  basic_block init_bb;
+};
+
+typedef struct incr_info_d incr_info, *incr_info_t;
+
 /* Candidates are maintained in a vector.  If candidate X dominates
    candidate Y, then X appears before Y in the vector; but the
    converse does not necessarily hold.  */
@@ -259,6 +293,16 @@ static htab_t base_cand_map;
 
 /* Obstack for candidate chains.  */
 static struct obstack chain_obstack;
+
+/* An array INCR_VEC of incr_infos is used during analysis of related
+   candidates having an SSA name for a stride.  INCR_VEC_LEN describes
+   its current length.  */
+static incr_info_t incr_vec;
+static unsigned incr_vec_len;
+
+/* For a chain of candidates with unknown stride, indicates whether or not
+   we must generate pointer arithmetic when replacing statements.  */
+static bool address_arithmetic_p;
 \f
 /* Produce a pointer to the IDX'th candidate in the candidate vector.  */
 
@@ -421,7 +465,6 @@ stmt_cost (gimple gs, bool speed)
     case PLUS_EXPR:
     case POINTER_PLUS_EXPR:
     case MINUS_EXPR:
-      rhs2 = gimple_assign_rhs2 (gs);
       return add_cost (speed, lhs_mode);
 
     case NEGATE_EXPR:
@@ -1407,6 +1450,30 @@ dump_cand_chains (void)
   htab_traverse_noresize (base_cand_map, base_cand_dump_callback, NULL);
   fputs ("\n", dump_file);
 }
+
+/* Dump the increment vector for debug.  */
+
+static void
+dump_incr_vec (void)
+{
+  if (dump_file && (dump_flags & TDF_DETAILS))
+    {
+      unsigned i;
+
+      fprintf (dump_file, "\nIncrement vector:\n\n");
+  
+      for (i = 0; i < incr_vec_len; i++)
+       {
+         fprintf (dump_file, "%3d  increment:   ", i);
+         dump_double_int (dump_file, incr_vec[i].incr, false);
+         fprintf (dump_file, "\n     count:       %d", incr_vec[i].count);
+         fprintf (dump_file, "\n     cost:        %d", incr_vec[i].cost);
+         fputs ("\n     initializer: ", dump_file);
+         print_generic_expr (dump_file, incr_vec[i].initializer, 0);
+         fputs ("\n\n", dump_file);
+       }
+    }
+}
 \f
 /* Recursive helper for unconditional_cands_with_known_stride_p.
    Returns TRUE iff C, its siblings, and its dependents are all
@@ -1508,6 +1575,32 @@ cand_increment (slsr_cand_t c)
   return double_int_sub (c->index, basis->index);
 }
 
+/* Calculate the increment required for candidate C relative to
+   its basis.  If we aren't going to generate pointer arithmetic
+   for this candidate, return the absolute value of that increment
+   instead.  */
+
+static inline double_int
+cand_abs_increment (slsr_cand_t c)
+{
+  double_int increment = cand_increment (c);
+
+  if (!address_arithmetic_p && double_int_negative_p (increment))
+    increment = double_int_neg (increment);
+
+  return increment;
+}
+
+/* If *VAR is NULL or is not of a compatible type with TYPE, create a
+   new temporary reg of type TYPE and store it in *VAR.  */
+
+static inline void
+lazy_create_slsr_reg (tree *var, tree type)
+{
+  if (!*var || !types_compatible_p (TREE_TYPE (*var), type))
+    *var = create_tmp_reg (type, "slsr");
+}
+
 /* Return TRUE iff candidate C has already been replaced under
    another interpretation.  */
 
@@ -1630,6 +1723,764 @@ replace_dependents (slsr_cand_t c)
     replace_dependents (lookup_cand (c->dependent));
 }
 \f
+/* Return the index in the increment vector of the given INCREMENT.  */
+
+static inline unsigned
+incr_vec_index (double_int increment)
+{
+  unsigned i;
+  
+  for (i = 0;
+       i < incr_vec_len && !double_int_equal_p (increment, incr_vec[i].incr);
+       i++)
+    ;
+
+  gcc_assert (i < incr_vec_len);
+  return i;
+}
+
+/* Count the number of candidates in the tree rooted at C that have
+   not already been replaced under other interpretations.  */
+
+static unsigned
+count_candidates (slsr_cand_t c)
+{
+  unsigned count = cand_already_replaced (c) ? 0 : 1;
+
+  if (c->sibling)
+    count += count_candidates (lookup_cand (c->sibling));
+
+  if (c->dependent)
+    count += count_candidates (lookup_cand (c->dependent));
+
+  return count;
+}
+
+/* Increase the count of INCREMENT by one in the increment vector.
+   INCREMENT is associated with candidate C.  If an initializer
+   T_0 = stride * I is provided by a candidate that dominates all
+   candidates with the same increment, also record T_0 for subsequent use.  */
+
+static void
+record_increment (slsr_cand_t c, double_int increment)
+{
+  bool found = false;
+  unsigned i;
+
+  /* Treat increments that differ only in sign as identical so as to
+     share initializers, unless we are generating pointer arithmetic.  */
+  if (!address_arithmetic_p && double_int_negative_p (increment))
+    increment = double_int_neg (increment);
+
+  for (i = 0; i < incr_vec_len; i++)
+    {
+      if (double_int_equal_p (incr_vec[i].incr, increment))
+       {
+         incr_vec[i].count++;
+         found = true;
+
+         /* If we previously recorded an initializer that doesn't
+            dominate this candidate, it's not going to be useful to
+            us after all.  */
+         if (incr_vec[i].initializer
+             && !dominated_by_p (CDI_DOMINATORS,
+                                 gimple_bb (c->cand_stmt),
+                                 incr_vec[i].init_bb))
+           {
+             incr_vec[i].initializer = NULL_TREE;
+             incr_vec[i].init_bb = NULL;
+           }
+         
+         break;
+       }
+    }
+
+  if (!found)
+    {
+      /* The first time we see an increment, create the entry for it.
+        If this is the root candidate which doesn't have a basis, set
+        the count to zero.  We're only processing it so it can possibly
+        provide an initializer for other candidates.  */
+      incr_vec[incr_vec_len].incr = increment;
+      incr_vec[incr_vec_len].count = c->basis ? 1 : 0;
+      incr_vec[incr_vec_len].cost = COST_INFINITE;
+      
+      /* Optimistically record the first occurrence of this increment
+        as providing an initializer (if it does); we will revise this
+        opinion later if it doesn't dominate all other occurrences.
+         Exception:  increments of -1, 0, 1 never need initializers.  */
+      if (c->kind == CAND_ADD
+         && double_int_equal_p (c->index, increment)
+         && (double_int_scmp (increment, double_int_one) > 0
+             || double_int_scmp (increment, double_int_minus_one) < 0))
+       {
+         tree t0;
+         tree rhs1 = gimple_assign_rhs1 (c->cand_stmt);
+         tree rhs2 = gimple_assign_rhs2 (c->cand_stmt);
+         if (operand_equal_p (rhs1, c->base_expr, 0))
+           t0 = rhs2;
+         else
+           t0 = rhs1;
+         if (SSA_NAME_DEF_STMT (t0) && gimple_bb (SSA_NAME_DEF_STMT (t0)))
+           {
+             incr_vec[incr_vec_len].initializer = t0;
+             incr_vec[incr_vec_len++].init_bb
+               = gimple_bb (SSA_NAME_DEF_STMT (t0));
+           }
+         else
+           {
+             incr_vec[incr_vec_len].initializer = NULL_TREE;
+             incr_vec[incr_vec_len++].init_bb = NULL;
+           }
+       }
+      else
+       {
+         incr_vec[incr_vec_len].initializer = NULL_TREE;
+         incr_vec[incr_vec_len++].init_bb = NULL;
+       }
+    }
+}
+
+/* Determine how many times each unique increment occurs in the set
+   of candidates rooted at C's parent, recording the data in the
+   increment vector.  For each unique increment I, if an initializer
+   T_0 = stride * I is provided by a candidate that dominates all
+   candidates with the same increment, also record T_0 for subsequent
+   use.  */
+
+static void
+record_increments (slsr_cand_t c)
+{
+  if (!cand_already_replaced (c))
+    record_increment (c, cand_increment (c));
+
+  if (c->sibling)
+    record_increments (lookup_cand (c->sibling));
+
+  if (c->dependent)
+    record_increments (lookup_cand (c->dependent));
+}
+
+/* Return the first candidate in the tree rooted at C that has not
+   already been replaced, favoring siblings over dependents.  */
+
+static slsr_cand_t
+unreplaced_cand_in_tree (slsr_cand_t c)
+{
+  if (!cand_already_replaced (c))
+    return c;
+
+  if (c->sibling)
+    {
+      slsr_cand_t sib = unreplaced_cand_in_tree (lookup_cand (c->sibling));
+      if (sib)
+       return sib;
+    }
+
+  if (c->dependent)
+    {
+      slsr_cand_t dep = unreplaced_cand_in_tree (lookup_cand (c->dependent));
+      if (dep)
+       return dep;
+    }
+
+  return NULL;
+}
+
+/* Return TRUE if the candidates in the tree rooted at C should be
+   optimized for speed, else FALSE.  We estimate this based on the block
+   containing the most dominant candidate in the tree that has not yet
+   been replaced.  */
+
+static bool
+optimize_cands_for_speed_p (slsr_cand_t c)
+{
+  slsr_cand_t c2 = unreplaced_cand_in_tree (c);
+  gcc_assert (c2);
+  return optimize_bb_for_speed_p (gimple_bb (c2->cand_stmt));
+}
+
+/* Add COST_IN to the lowest cost of any dependent path starting at
+   candidate C or any of its siblings, counting only candidates along
+   such paths with increment INCR.  Assume that replacing a candidate
+   reduces cost by REPL_SAVINGS.  Also account for savings from any
+   statements that would go dead.  */
+
+static int
+lowest_cost_path (int cost_in, int repl_savings, slsr_cand_t c, double_int incr)
+{
+  int local_cost, sib_cost;
+  double_int cand_incr = cand_abs_increment (c);
+
+  if (cand_already_replaced (c))
+    local_cost = cost_in;
+  else if (double_int_equal_p (incr, cand_incr))
+    local_cost = cost_in - repl_savings - c->dead_savings;
+  else
+    local_cost = cost_in - c->dead_savings;
+
+  if (c->dependent)
+    local_cost = lowest_cost_path (local_cost, repl_savings, 
+                                  lookup_cand (c->dependent), incr);
+
+  if (c->sibling)
+    {
+      sib_cost = lowest_cost_path (cost_in, repl_savings,
+                                  lookup_cand (c->sibling), incr);
+      local_cost = MIN (local_cost, sib_cost);
+    }
+
+  return local_cost;
+}
+
+/* Compute the total savings that would accrue from all replacements
+   in the candidate tree rooted at C, counting only candidates with
+   increment INCR.  Assume that replacing a candidate reduces cost
+   by REPL_SAVINGS.  Also account for savings from statements that
+   would go dead.  */
+
+static int
+total_savings (int repl_savings, slsr_cand_t c, double_int incr)
+{
+  int savings = 0;
+  double_int cand_incr = cand_abs_increment (c);
+
+  if (double_int_equal_p (incr, cand_incr)
+      && !cand_already_replaced (c))
+    savings += repl_savings + c->dead_savings;
+
+  if (c->dependent)
+    savings += total_savings (repl_savings, lookup_cand (c->dependent), incr);
+
+  if (c->sibling)
+    savings += total_savings (repl_savings, lookup_cand (c->sibling), incr);
+
+  return savings;
+}
+
+/* Use target-specific costs to determine and record which increments
+   in the current candidate tree are profitable to replace, assuming
+   MODE and SPEED.  FIRST_DEP is the first dependent of the root of
+   the candidate tree.
+
+   One slight limitation here is that we don't account for the possible
+   introduction of casts in some cases.  See replace_one_candidate for
+   the cases where these are introduced.  This should probably be cleaned
+   up sometime.  */
+
+static void
+analyze_increments (slsr_cand_t first_dep, enum machine_mode mode, bool speed)
+{
+  unsigned i;
+
+  for (i = 0; i < incr_vec_len; i++)
+    {
+      HOST_WIDE_INT incr = double_int_to_shwi (incr_vec[i].incr);
+
+      /* If somehow this increment is bigger than a HWI, we won't
+        be optimizing candidates that use it.  And if the increment
+        has a count of zero, nothing will be done with it.  */
+      if (!double_int_fits_in_shwi_p (incr_vec[i].incr)
+         || !incr_vec[i].count)
+       incr_vec[i].cost = COST_INFINITE;
+
+      /* Increments of 0, 1, and -1 are always profitable to replace,
+        because they always replace a multiply or add with an add or
+        copy, and may cause one or more existing instructions to go
+        dead.  Exception:  -1 can't be assumed to be profitable for
+        pointer addition.  */
+      else if (incr == 0
+              || incr == 1
+              || (incr == -1
+                  && (gimple_assign_rhs_code (first_dep->cand_stmt)
+                      != POINTER_PLUS_EXPR)))
+       incr_vec[i].cost = COST_NEUTRAL;
+      
+      /* For any other increment, if this is a multiply candidate, we
+        must introduce a temporary T and initialize it with
+        T_0 = stride * increment.  When optimizing for speed, walk the
+        candidate tree to calculate the best cost reduction along any
+        path; if it offsets the fixed cost of inserting the initializer,
+        replacing the increment is profitable.  When optimizing for
+         size, instead calculate the total cost reduction from replacing
+        all candidates with this increment.  */
+      else if (first_dep->kind == CAND_MULT)
+       {
+         int cost = mult_by_coeff_cost (incr, mode, speed);
+         int repl_savings = mul_cost (speed, mode) - add_cost (speed, mode);
+         if (speed)
+           cost = lowest_cost_path (cost, repl_savings, first_dep,
+                                    incr_vec[i].incr);
+         else
+           cost -= total_savings (repl_savings, first_dep, incr_vec[i].incr);
+
+         incr_vec[i].cost = cost;
+       }
+
+      /* If this is an add candidate, the initializer may already
+        exist, so only calculate the cost of the initializer if it
+        doesn't.  We are replacing one add with another here, so the
+        known replacement savings is zero.  We will account for removal
+        of dead instructions in lowest_cost_path or total_savings.  */
+      else
+       {
+         int cost = 0;
+         if (!incr_vec[i].initializer)
+           cost = mult_by_coeff_cost (incr, mode, speed);
+
+         if (speed)
+           cost = lowest_cost_path (cost, 0, first_dep, incr_vec[i].incr);
+         else
+           cost -= total_savings (0, first_dep, incr_vec[i].incr);
+
+         incr_vec[i].cost = cost;
+       }
+    }
+}
+
+/* Return the nearest common dominator of BB1 and BB2.  If the blocks
+   are identical, return the earlier of C1 and C2 in *WHERE.  Otherwise,
+   if the NCD matches BB1, return C1 in *WHERE; if the NCD matches BB2,
+   return C2 in *WHERE; and if the NCD matches neither, return NULL in
+   *WHERE.  Note: It is possible for one of C1 and C2 to be NULL.  */
+
+static basic_block
+ncd_for_two_cands (basic_block bb1, basic_block bb2,
+                  slsr_cand_t c1, slsr_cand_t c2, slsr_cand_t *where)
+{
+  basic_block ncd;
+
+  if (!bb1)
+    {
+      *where = c2;
+      return bb2;
+    }
+
+  if (!bb2)
+    {
+      *where = c1;
+      return bb1;
+    }
+
+  ncd = nearest_common_dominator (CDI_DOMINATORS, bb1, bb2);
+      
+  /* If both candidates are in the same block, the earlier
+     candidate wins.  */
+  if (bb1 == ncd && bb2 == ncd)
+    {
+      if (!c1 || (c2 && c2->cand_num < c1->cand_num))
+       *where = c2;
+      else
+       *where = c1;
+    }
+
+  /* Otherwise, if one of them produced a candidate in the
+     dominator, that one wins.  */
+  else if (bb1 == ncd)
+    *where = c1;
+
+  else if (bb2 == ncd)
+    *where = c2;
+
+  /* If neither matches the dominator, neither wins.  */
+  else
+    *where = NULL;
+
+  return ncd;
+}
+
+/* Consider all candidates in the tree rooted at C for which INCR
+   represents the required increment of C relative to its basis.
+   Find and return the basic block that most nearly dominates all
+   such candidates.  If the returned block contains one or more of
+   the candidates, return the earliest candidate in the block in
+   *WHERE.  */
+
+static basic_block
+nearest_common_dominator_for_cands (slsr_cand_t c, double_int incr,
+                                   slsr_cand_t *where)
+{
+  basic_block sib_ncd = NULL, dep_ncd = NULL, this_ncd = NULL, ncd;
+  slsr_cand_t sib_where = NULL, dep_where = NULL, this_where = NULL, new_where;
+  double_int cand_incr;
+
+  /* First find the NCD of all siblings and dependents.  */
+  if (c->sibling)
+    sib_ncd = nearest_common_dominator_for_cands (lookup_cand (c->sibling),
+                                                 incr, &sib_where);
+  if (c->dependent)
+    dep_ncd = nearest_common_dominator_for_cands (lookup_cand (c->dependent),
+                                                 incr, &dep_where);
+  if (!sib_ncd && !dep_ncd)
+    {
+      new_where = NULL;
+      ncd = NULL;
+    }
+  else if (sib_ncd && !dep_ncd)
+    {
+      new_where = sib_where;
+      ncd = sib_ncd;
+    }
+  else if (dep_ncd && !sib_ncd)
+    {
+      new_where = dep_where;
+      ncd = dep_ncd;
+    }
+  else
+    ncd = ncd_for_two_cands (sib_ncd, dep_ncd, sib_where,
+                            dep_where, &new_where);
+
+  /* If the candidate's increment doesn't match the one we're interested
+     in, then the result depends only on siblings and dependents.  */
+  cand_incr = cand_abs_increment (c);
+
+  if (!double_int_equal_p (cand_incr, incr) || cand_already_replaced (c))
+    {
+      *where = new_where;
+      return ncd;
+    }
+
+  /* Otherwise, compare this candidate with the result from all siblings
+     and dependents.  */
+  this_where = c;
+  this_ncd = gimple_bb (c->cand_stmt);
+  ncd = ncd_for_two_cands (ncd, this_ncd, new_where, this_where, where);
+
+  return ncd;
+}
+
+/* Return TRUE if the increment indexed by INDEX is profitable to replace.  */
+
+static inline bool
+profitable_increment_p (unsigned index)
+{
+  return (incr_vec[index].cost <= COST_NEUTRAL);
+}
+
+/* For each profitable increment in the increment vector not equal to
+   0 or 1 (or -1, for non-pointer arithmetic), find the nearest common
+   dominator of all statements in the candidate chain rooted at C
+   that require that increment, and insert an initializer
+   T_0 = stride * increment at that location.  Record T_0 with the
+   increment record.  */
+
+static void
+insert_initializers (slsr_cand_t c)
+{
+  unsigned i;
+  tree new_var = NULL_TREE;
+
+  for (i = 0; i < incr_vec_len; i++)
+    {
+      basic_block bb;
+      slsr_cand_t where = NULL;
+      gimple init_stmt;
+      tree stride_type, new_name, incr_tree;
+      double_int incr = incr_vec[i].incr;
+
+      if (!profitable_increment_p (i)
+         || double_int_one_p (incr)
+         || (double_int_minus_one_p (incr)
+             && gimple_assign_rhs_code (c->cand_stmt) != POINTER_PLUS_EXPR)
+         || double_int_zero_p (incr))
+       continue;
+
+      /* We may have already identified an existing initializer that
+        will suffice.  */
+      if (incr_vec[i].initializer)
+       {
+         if (dump_file && (dump_flags & TDF_DETAILS))
+           {
+             fputs ("Using existing initializer: ", dump_file);
+             print_gimple_stmt (dump_file,
+                                SSA_NAME_DEF_STMT (incr_vec[i].initializer),
+                                0, 0);
+           }
+         continue;
+       }
+
+      /* Find the block that most closely dominates all candidates
+        with this increment.  If there is at least one candidate in
+        that block, the earliest one will be returned in WHERE.  */
+      bb = nearest_common_dominator_for_cands (c, incr, &where);
+
+      /* Create a new SSA name to hold the initializer's value.  */
+      stride_type = TREE_TYPE (c->stride);
+      lazy_create_slsr_reg (&new_var, stride_type);
+      new_name = make_ssa_name (new_var, NULL);
+      incr_vec[i].initializer = new_name;
+
+      /* Create the initializer and insert it in the latest possible
+        dominating position.  */
+      incr_tree = double_int_to_tree (stride_type, incr);
+      init_stmt = gimple_build_assign_with_ops (MULT_EXPR, new_name,
+                                               c->stride, incr_tree);
+      if (where)
+       {
+         gimple_stmt_iterator gsi = gsi_for_stmt (where->cand_stmt);
+         gsi_insert_before (&gsi, init_stmt, GSI_SAME_STMT);
+         gimple_set_location (init_stmt, gimple_location (where->cand_stmt));
+       }
+      else
+       {
+         gimple_stmt_iterator gsi = gsi_last_bb (bb);
+         gimple basis_stmt = lookup_cand (c->basis)->cand_stmt;
+
+         if (!gsi_end_p (gsi) && is_ctrl_stmt (gsi_stmt (gsi)))
+           gsi_insert_before (&gsi, init_stmt, GSI_SAME_STMT);
+         else
+           gsi_insert_after (&gsi, init_stmt, GSI_SAME_STMT);
+
+         gimple_set_location (init_stmt, gimple_location (basis_stmt));
+       }
+
+      if (dump_file && (dump_flags & TDF_DETAILS))
+       {
+         fputs ("Inserting initializer: ", dump_file);
+         print_gimple_stmt (dump_file, init_stmt, 0, 0);
+       }
+    }
+}
+
+/* Create a NOP_EXPR that copies FROM_EXPR into a new SSA name of
+   type TO_TYPE, and insert it in front of the statement represented
+   by candidate C.  Use *NEW_VAR to create the new SSA name.  Return
+   the new SSA name.  */
+
+static tree
+introduce_cast_before_cand (slsr_cand_t c, tree to_type,
+                           tree from_expr, tree *new_var)
+{
+  tree cast_lhs;
+  gimple cast_stmt;
+  gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt);
+
+  lazy_create_slsr_reg (new_var, to_type);
+  cast_lhs = make_ssa_name (*new_var, NULL);
+  cast_stmt = gimple_build_assign_with_ops (NOP_EXPR, cast_lhs,
+                                           from_expr, NULL_TREE);
+  gimple_set_location (cast_stmt, gimple_location (c->cand_stmt));
+  gsi_insert_before (&gsi, cast_stmt, GSI_SAME_STMT);
+
+  if (dump_file && (dump_flags & TDF_DETAILS))
+    {
+      fputs ("  Inserting: ", dump_file);
+      print_gimple_stmt (dump_file, cast_stmt, 0, 0);
+    }
+
+  return cast_lhs;
+}
+
+/* Replace the RHS of the statement represented by candidate C with 
+   NEW_CODE, NEW_RHS1, and NEW_RHS2, provided that to do so doesn't
+   leave C unchanged or just interchange its operands.  The original
+   operation and operands are in OLD_CODE, OLD_RHS1, and OLD_RHS2.
+   If the replacement was made and we are doing a details dump,
+   return the revised statement, else NULL.  */
+
+static gimple
+replace_rhs_if_not_dup (enum tree_code new_code, tree new_rhs1, tree new_rhs2,
+                       enum tree_code old_code, tree old_rhs1, tree old_rhs2,
+                       slsr_cand_t c)
+{
+  if (new_code != old_code
+      || ((!operand_equal_p (new_rhs1, old_rhs1, 0)
+          || !operand_equal_p (new_rhs2, old_rhs2, 0))
+         && (!operand_equal_p (new_rhs1, old_rhs2, 0)
+             || !operand_equal_p (new_rhs2, old_rhs1, 0))))
+    {
+      gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt);
+      gimple_assign_set_rhs_with_ops (&gsi, new_code, new_rhs1, new_rhs2);
+      update_stmt (gsi_stmt (gsi));
+
+      if (dump_file && (dump_flags & TDF_DETAILS))
+       return gsi_stmt (gsi);
+    }
+
+  else if (dump_file && (dump_flags & TDF_DETAILS))
+    fputs ("  (duplicate, not actually replacing)\n", dump_file);
+
+  return NULL;
+}
+
+/* Strength-reduce the statement represented by candidate C by replacing
+   it with an equivalent addition or subtraction.  I is the index into
+   the increment vector identifying C's increment.  NEW_VAR is used to
+   create a new SSA name if a cast needs to be introduced.  BASIS_NAME
+   is the rhs1 to use in creating the add/subtract.  */
+
+static void
+replace_one_candidate (slsr_cand_t c, unsigned i, tree *new_var,
+                      tree basis_name)
+{
+  gimple stmt_to_print = NULL;
+  tree orig_rhs1, orig_rhs2;
+  tree rhs2;
+  enum tree_code orig_code, repl_code;
+  double_int cand_incr;
+
+  orig_code = gimple_assign_rhs_code (c->cand_stmt);
+  orig_rhs1 = gimple_assign_rhs1 (c->cand_stmt);
+  orig_rhs2 = gimple_assign_rhs2 (c->cand_stmt);
+  cand_incr = cand_increment (c);
+
+  if (dump_file && (dump_flags & TDF_DETAILS))
+    {
+      fputs ("Replacing: ", dump_file);
+      print_gimple_stmt (dump_file, c->cand_stmt, 0, 0);
+      stmt_to_print = c->cand_stmt;
+    }
+
+  if (address_arithmetic_p)
+    repl_code = POINTER_PLUS_EXPR;
+  else
+    repl_code = PLUS_EXPR;
+
+  /* If the increment has an initializer T_0, replace the candidate
+     statement with an add of the basis name and the initializer.  */
+  if (incr_vec[i].initializer)
+    {
+      tree init_type = TREE_TYPE (incr_vec[i].initializer);
+      tree orig_type = TREE_TYPE (orig_rhs2);
+
+      if (types_compatible_p (orig_type, init_type))
+       rhs2 = incr_vec[i].initializer;
+      else
+       rhs2 = introduce_cast_before_cand (c, orig_type,
+                                          incr_vec[i].initializer,
+                                          new_var);
+
+      if (!double_int_equal_p (incr_vec[i].incr, cand_incr))
+       {
+         gcc_assert (repl_code == PLUS_EXPR);
+         repl_code = MINUS_EXPR;
+       }
+
+      stmt_to_print = replace_rhs_if_not_dup (repl_code, basis_name, rhs2,
+                                             orig_code, orig_rhs1, orig_rhs2,
+                                             c);
+    }
+
+  /* Otherwise, the increment is one of -1, 0, and 1.  Replace
+     with a subtract of the stride from the basis name, a copy
+     from the basis name, or an add of the stride to the basis
+     name, respectively.  It may be necessary to introduce a
+     cast (or reuse an existing cast).  */
+  else if (double_int_one_p (cand_incr))
+    {
+      tree stride_type = TREE_TYPE (c->stride);
+      tree orig_type = TREE_TYPE (orig_rhs2);
+      
+      if (types_compatible_p (orig_type, stride_type))
+       rhs2 = c->stride;
+      else
+       rhs2 = introduce_cast_before_cand (c, orig_type, c->stride, new_var);
+      
+      stmt_to_print = replace_rhs_if_not_dup (repl_code, basis_name, rhs2,
+                                             orig_code, orig_rhs1, orig_rhs2,
+                                             c);
+    }
+
+  else if (double_int_minus_one_p (cand_incr))
+    {
+      tree stride_type = TREE_TYPE (c->stride);
+      tree orig_type = TREE_TYPE (orig_rhs2);
+      gcc_assert (repl_code != POINTER_PLUS_EXPR);
+      
+      if (types_compatible_p (orig_type, stride_type))
+       rhs2 = c->stride;
+      else
+       rhs2 = introduce_cast_before_cand (c, orig_type, c->stride, new_var);
+      
+      if (orig_code != MINUS_EXPR
+         || !operand_equal_p (basis_name, orig_rhs1, 0)
+         || !operand_equal_p (rhs2, orig_rhs2, 0))
+       {
+         gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt);
+         gimple_assign_set_rhs_with_ops (&gsi, MINUS_EXPR, basis_name, rhs2);
+         update_stmt (gsi_stmt (gsi));
+
+         if (dump_file && (dump_flags & TDF_DETAILS))
+           stmt_to_print = gsi_stmt (gsi);
+       }
+      else if (dump_file && (dump_flags & TDF_DETAILS))
+       fputs ("  (duplicate, not actually replacing)\n", dump_file);
+    }
+
+  else if (double_int_zero_p (cand_incr))
+    {
+      tree lhs = gimple_assign_lhs (c->cand_stmt);
+      tree lhs_type = TREE_TYPE (lhs);
+      tree basis_type = TREE_TYPE (basis_name);
+      
+      if (types_compatible_p (lhs_type, basis_type))
+       {
+         gimple copy_stmt = gimple_build_assign (lhs, basis_name);
+         gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt);
+         gimple_set_location (copy_stmt, gimple_location (c->cand_stmt));
+         gsi_replace (&gsi, copy_stmt, false);
+
+         if (dump_file && (dump_flags & TDF_DETAILS))
+           stmt_to_print = copy_stmt;
+       }
+      else
+       {
+         gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt);
+         gimple cast_stmt = gimple_build_assign_with_ops (NOP_EXPR, lhs,
+                                                          basis_name,
+                                                          NULL_TREE);
+         gimple_set_location (cast_stmt, gimple_location (c->cand_stmt));
+         gsi_replace (&gsi, cast_stmt, false);
+
+         if (dump_file && (dump_flags & TDF_DETAILS))
+           stmt_to_print = cast_stmt;
+       }
+    }
+  else
+    gcc_unreachable ();
+  
+  if (dump_file && (dump_flags & TDF_DETAILS) && stmt_to_print)
+    {
+      fputs ("With: ", dump_file);
+      print_gimple_stmt (dump_file, stmt_to_print, 0, 0);
+      fputs ("\n", dump_file);
+    }
+}
+
+/* For each candidate in the tree rooted at C, replace it with
+   an increment if such has been shown to be profitable.  */
+
+static void
+replace_profitable_candidates (slsr_cand_t c)
+{
+  if (!cand_already_replaced (c))
+    {
+      double_int increment = cand_abs_increment (c);
+      tree new_var = NULL;
+      enum tree_code orig_code = gimple_assign_rhs_code (c->cand_stmt);
+      unsigned i;
+
+      i = incr_vec_index (increment);
+
+      /* Only process profitable increments.  Nothing useful can be done
+        to a cast or copy.  */
+      if (profitable_increment_p (i) 
+         && orig_code != MODIFY_EXPR
+         && orig_code != NOP_EXPR)
+       {
+         slsr_cand_t basis = lookup_cand (c->basis);
+         tree basis_name = gimple_assign_lhs (basis->cand_stmt);
+         replace_one_candidate (c, i, &new_var, basis_name);
+       }
+    }
+
+  if (c->sibling)
+    replace_profitable_candidates (lookup_cand (c->sibling));
+
+  if (c->dependent)
+    replace_profitable_candidates (lookup_cand (c->dependent));
+}
+\f
 /* Analyze costs of related candidates in the candidate vector,
    and make beneficial replacements.  */
 
@@ -1670,11 +2521,46 @@ analyze_candidates_and_replace (void)
       else if (unconditional_cands_with_known_stride_p (c))
        replace_dependents (first_dep);
 
-      /* TODO:  When the stride is an SSA name, it may still be
-        profitable to replace some or all of the dependent
-        candidates, depending on whether the introduced increments
-        can be reused, or are less expensive to calculate than
-        the replaced statements.  */
+      /* When the stride is an SSA name, it may still be profitable
+        to replace some or all of the dependent candidates, depending
+        on whether the introduced increments can be reused, or are
+        less expensive to calculate than the replaced statements.  */
+      else
+       {
+         unsigned length;
+         enum machine_mode mode;
+         bool speed;
+
+         /* Determine whether we'll be generating pointer arithmetic
+            when replacing candidates.  */
+         address_arithmetic_p = (c->kind == CAND_ADD
+                                 && POINTER_TYPE_P (TREE_TYPE (c->base_expr)));
+
+         /* If all candidates have already been replaced under other
+            interpretations, nothing remains to be done.  */
+         length = count_candidates (c);
+         if (!length)
+           continue;
+
+         /* Construct an array of increments for this candidate chain.  */
+         incr_vec = XNEWVEC (incr_info, length);
+         incr_vec_len = 0;
+         record_increments (c);
+
+         /* Determine which increments are profitable to replace.  */
+         mode = TYPE_MODE (TREE_TYPE (gimple_assign_lhs (c->cand_stmt)));
+         speed = optimize_cands_for_speed_p (c);
+         analyze_increments (first_dep, mode, speed);
+
+         /* Insert initializers of the form T_0 = stride * increment
+            for use in profitable replacements.  */
+         insert_initializers (first_dep);
+         dump_incr_vec ();
+
+         /* Perform the replacements.  */
+         replace_profitable_candidates (first_dep);
+         free (incr_vec);
+       }
 
       /* TODO:  When conditional increments occur so that a 
         candidate is dependent upon a phi-basis, the cost of
index 61ac4e7..171a16e 100644 (file)
@@ -1,3 +1,30 @@
+2012-08-07  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>
+
+       * gcc.dg/tree-ssa/slsr-5.c: New.
+       * gcc.dg/tree-ssa/slsr-6.c: New.
+       * gcc.dg/tree-ssa/slsr-7.c: New.
+       * gcc.dg/tree-ssa/slsr-8.c: New.
+       * gcc.dg/tree-ssa/slsr-9.c: New.
+       * gcc.dg/tree-ssa/slsr-10.c: New.
+       * gcc.dg/tree-ssa/slsr-11.c: New.
+       * gcc.dg/tree-ssa/slsr-12.c: New.
+       * gcc.dg/tree-ssa/slsr-13.c: New.
+       * gcc.dg/tree-ssa/slsr-14.c: New.
+       * gcc.dg/tree-ssa/slsr-15.c: New.
+       * gcc.dg/tree-ssa/slsr-16.c: New.
+       * gcc.dg/tree-ssa/slsr-17.c: New.
+       * gcc.dg/tree-ssa/slsr-18.c: New.
+       * gcc.dg/tree-ssa/slsr-19.c: New.
+       * gcc.dg/tree-ssa/slsr-20.c: New.
+       * gcc.dg/tree-ssa/slsr-21.c: New.
+       * gcc.dg/tree-ssa/slsr-22.c: New.
+       * gcc.dg/tree-ssa/slsr-23.c: New.
+       * gcc.dg/tree-ssa/slsr-24.c: New.
+       * gcc.dg/tree-ssa/slsr-25.c: New.
+       * gcc.dg/tree-ssa/slsr-26.c: New.
+       * gcc.dg/tree-ssa/slsr-30.c: New.
+       * gcc.dg/tree-ssa/slsr-31.c: New.
+
 2012-08-07  Steven Bosscher  <steven@gcc.gnu.org>
 
        * gcc.dg/tree-prof/update-loopch.c: Ask for dump with blocks info.
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-10.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-10.c
new file mode 100644 (file)
index 0000000..f131dae
--- /dev/null
@@ -0,0 +1,23 @@
+/* Verify straight-line strength reduction for simple integer addition
+   with stride reversed on 1st and 3rd instances.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int s, int c)
+{
+  int a1, a2, a3, x1, x2, x3, x;
+
+  a1 = 2 * s;
+  x1 = a1 + c;
+  a2 = 4 * s;
+  x2 = c + a2;
+  a3 = 6 * s;
+  x3 = a3 + c;
+  x = x1 + x2 + x3;
+  return x;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* " 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-11.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-11.c
new file mode 100644 (file)
index 0000000..6f77d51
--- /dev/null
@@ -0,0 +1,24 @@
+/* Verify straight-line strength reduction for simple integer addition
+   with casts thrown in.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+long
+f (int s, long c)
+{
+  int a1, a2, a3;
+  long x1, x2, x3, x;
+
+  a1 = 2 * s;
+  x1 = c + a1;
+  a2 = 4 * s;
+  x2 = c + a2;
+  a3 = 6 * s;
+  x3 = c + a3;
+  x = x1 + x2 + x3;
+  return x;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* " 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-12.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-12.c
new file mode 100644 (file)
index 0000000..558b359
--- /dev/null
@@ -0,0 +1,30 @@
+/* Verify that no straight-line strength reduction occurs across sibling
+   blocks.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int s, int c)
+{
+  int a1, a2, a3, x1, x2, x3, x;
+
+  if (c > 0)
+    {
+      a1 = 2 * s;
+      x1 = c + a1;
+    }
+  else
+    {
+      a1 = 4 * s;
+      x1 = c + a1;
+    }
+
+  a2 = 6 * s;
+  x2 = c + a2;
+  x = x1 + x2;
+  return x;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* " 3 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-13.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-13.c
new file mode 100644 (file)
index 0000000..5cdfb9d
--- /dev/null
@@ -0,0 +1,25 @@
+/* x2 and x3 will be strength-reduced based on the same statement
+   but with different variables as the stride.  Note that they will
+   be strength-reduced by introducing an initializer 4*s which is
+   cheaper than 5*s; similar for 4*c and 5*c.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int s, int c)
+{
+  int a2, a3, x1, x2, x3, x;
+
+  x1 = c + s;
+  a2 = 5 * s;
+  x2 = c + a2;
+  a3 = 5 * c;
+  x3 = s + a3;
+  x = x1 + x2 + x3;
+  return x;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* 4" 2 "optimized" } } */
+/* { dg-final { scan-tree-dump-times " \\* 5" 0 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-14.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-14.c
new file mode 100644 (file)
index 0000000..1941345
--- /dev/null
@@ -0,0 +1,32 @@
+/* Straight-line strength reduction control flow variation.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int n, int c, int s)
+{
+  int a1, a2, x, x1, x2, x3, x4;
+
+  a1 = 2 * s;
+
+  if (n > 64)
+    {
+      x1 = c + a1;
+      a2 = 4 * s;
+      x2 = c + a2;
+      x = x1 + x2;
+    }
+  else
+    {
+      x3 = c + a1;
+      a2 = 4 * s;
+      x4 = c + a2;
+      x = x4 / x3;
+    }
+
+  return x;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* " 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-15.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-15.c
new file mode 100644 (file)
index 0000000..98d2df8
--- /dev/null
@@ -0,0 +1,27 @@
+/* Straight-line strength reduction control flow variation.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int 
+f (int n, int c, int s)
+{
+  int a, x1, x2, x3;
+
+  x1 = x2 = x3 = c;
+
+  if (n > 64)
+    {
+      a = 2 * s;
+      x1 = c + a;
+      a = 4 * s;
+      x2 = c + a;
+      a = 6 * s;
+      x3 = c + a;
+    }
+
+  return x1 + x2 + x3;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* " 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-16.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-16.c
new file mode 100644 (file)
index 0000000..5fe0175
--- /dev/null
@@ -0,0 +1,28 @@
+/* Straight-line strength reduction control flow variation.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int n, int c, int s)
+{
+  int a2, a3, a4, x1, x2, x3, x4;
+
+  x1 = c + s;
+  a2 = 3 * s;
+  x2 = c + a2;
+  x3 = x4 = c;
+
+  if (n > 64)
+    {
+      a3 = 5 * s;
+      x3 = c + a3;
+      a4 = 7 * s;
+      x4 = c + a4;
+    }
+
+  return x1 + x2 + x3 + x4;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* " 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-17.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-17.c
new file mode 100644 (file)
index 0000000..3da203c
--- /dev/null
@@ -0,0 +1,31 @@
+/* Straight-line strength reduction control flow variation with incr = 1.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int n, int c, int s)
+{
+  int a2, a3, a4, x1, x2, x3, x4;
+
+  x1 = c + s;
+  x2 = x3 = x4 = c;
+
+  if (n > 64)
+    {
+      a2 = 2 * s;
+      x2 = c + a2;
+      a3 = 3 * s;
+      x3 = c + a3;
+    }
+  else
+    {
+      a4 = 2 * s;
+      x4 = c + a4;
+    }
+
+  return x1 + x2 + x3 + x4;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* " 0 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-18.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-18.c
new file mode 100644 (file)
index 0000000..b8d368e
--- /dev/null
@@ -0,0 +1,32 @@
+/* Straight-line strength reduction control flow variation with incr = -1.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int n, int c, int s)
+{
+  int a1, a2, a3, a4, x1, x2, x3, x4;
+
+  a1 = 4 * s;
+  x1 = c + a1;
+  x2 = x3 = x4 = c;
+
+  if (n > 64)
+    {
+      a2 = 3 * s;
+      x2 = c + a2;
+      a3 = 2 * s;
+      x3 = c + a3;
+    }
+  else
+    {
+      a4 = 3 * s;
+      x4 = c + a4;
+    }
+
+  return x1 + x2 + x3 + x4;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* " 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-19.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-19.c
new file mode 100644 (file)
index 0000000..c92730f
--- /dev/null
@@ -0,0 +1,22 @@
+/* Verify straight-line strength reduction for multiply candidates
+   with stride in RHS1 position.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int c, int s)
+{
+  int x1, x2, y1, y2;
+
+  y1 = c + 2;
+  x1 = s * y1;
+  y2 = y1 + 2;
+  x2 = s * y2;
+  return x1 + x2;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* y" 1 "optimized" } } */
+/* { dg-final { scan-tree-dump-times " \\* 2" 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
+
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-20.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-20.c
new file mode 100644 (file)
index 0000000..cf3e1b3
--- /dev/null
@@ -0,0 +1,21 @@
+/* Verify straight-line strength reduction for multiply candidates
+   with stride in inconsistent positions.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int c, int s)
+{
+  int x1, x2, y1, y2;
+
+  y1 = c + 2;
+  x1 = y1 * s;
+  y2 = y1 + 2;
+  x2 = s * y2;
+  return x1 + x2;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* s" 1 "optimized" } } */
+/* { dg-final { scan-tree-dump-times " \\* 2" 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-21.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-21.c
new file mode 100644 (file)
index 0000000..c32afb2
--- /dev/null
@@ -0,0 +1,32 @@
+/* Verify straight-line strength reduction for multiply candidates
+   with variable stride and control flow.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int n, int x, int stride)
+{
+  int a, x1, x2, x3;
+
+  a = x * stride;
+
+  if (n > 64)
+    {
+      x1 = x + 3;
+      a += x1 * stride;
+      x2 = x1 + 3;
+      a += x2 * stride;
+    }
+  else
+    {
+      x3 = x + 3;
+      a += x3 * stride;
+    }
+
+  return a;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* stride" 1 "optimized" } } */
+/* { dg-final { scan-tree-dump-times " \\* 3" 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-22.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-22.c
new file mode 100644 (file)
index 0000000..6e56d9b
--- /dev/null
@@ -0,0 +1,29 @@
+/* Verify straight-line strength reduction for multiply candidates
+   with variable stride and control flow.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int n, int x, int stride)
+{
+  int a, x1, x2, x3;
+
+  a = x * stride;
+
+  if (n > 64)
+    {
+      x1 = x + 3;
+      a += x1 * stride;
+      x2 = x1 + 3;
+      a += x2 * stride;
+      x3 = x2 + 3;
+      a += x3 * stride;
+    }
+
+  return a;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* stride" 1 "optimized" } } */
+/* { dg-final { scan-tree-dump-times " \\* 3" 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-23.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-23.c
new file mode 100644 (file)
index 0000000..88bd087
--- /dev/null
@@ -0,0 +1,29 @@
+/* Verify straight-line strength reduction for multiply candidates
+   with variable stride and control flow.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int 
+f (int n, int x, int stride)
+{
+  int a, x1, x2, x3;
+
+  a = x * stride;
+  x1 = x + 3;
+  a += x1 * stride;
+
+  if (n > 64)
+    {
+      x2 = x1 + 3;
+      a += x2 * stride;
+      x3 = x2 + 3;
+      a += x3 * stride;
+    }
+
+  return a;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* stride" 1 "optimized" } } */
+/* { dg-final { scan-tree-dump-times " \\* 3" 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-24.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-24.c
new file mode 100644 (file)
index 0000000..04e093b
--- /dev/null
@@ -0,0 +1,31 @@
+/* Verify straight-line strength reduction for multiply candidates
+   with variable stride and control flow, increment = 1.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int n, int x, int stride)
+{
+  int a, x1, x2, x3;
+
+  a = x * stride;
+
+  if (n > 64)
+    {
+      x1 = x + 1;
+      a += x1 * stride;
+      x2 = x1 + 1;
+      a += x2 * stride;
+    }
+  else
+    {
+      x3 = x + 1;
+      a += x3 * stride;
+    }
+
+  return a;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* " 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-25.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-25.c
new file mode 100644 (file)
index 0000000..9e16e04
--- /dev/null
@@ -0,0 +1,31 @@
+/* Verify straight-line strength reduction for multiply candidates
+   with variable stride and control flow, increment = -1.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int 
+f (int n, int x, int stride)
+{
+  int a, x1, x2, x3;
+
+  a = x * stride;
+
+  if (n > 64)
+    {
+      x1 = x - 1;
+      a += x1 * stride;
+      x2 = x1 - 1;
+      a += x2 * stride;
+    }
+  else
+    {
+      x3 = x - 1;
+      a += x3 * stride;
+    }
+
+  return a;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* " 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-26.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-26.c
new file mode 100644 (file)
index 0000000..c333e00
--- /dev/null
@@ -0,0 +1,32 @@
+/* Verify straight-line strength reduction for multiply candidates
+   with variable stride and control flow, increment = -3.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int n, int x, int stride)
+{
+  int a, x1, x2, x3;
+
+  a = x * stride;
+
+  if (n > 64)
+    {
+      x1 = x - 3;
+      a += x1 * stride;
+      x2 = x1 - 3;
+      a += x2 * stride;
+    }
+  else
+    {
+      x3 = x - 3;
+      a += x3 * stride;
+    }
+
+  return a;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* stride" 1 "optimized" } } */
+/* { dg-final { scan-tree-dump-times " \\* 3" 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-30.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-30.c
new file mode 100644 (file)
index 0000000..fbd6897
--- /dev/null
@@ -0,0 +1,25 @@
+/* Verify straight-line strength reduction fails for simple integer addition
+   with casts thrown in when -fwrapv is used.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-dom2 -fwrapv" } */
+/* { dg-skip-if "" { ilp32 } { "-m32" } { "" } } */
+
+long
+f (int s, long c)
+{
+  int a1, a2, a3;
+  long x1, x2, x3, x;
+
+  a1 = 2 * s;
+  x1 = c + a1;
+  a2 = 4 * s;
+  x2 = c + a2;
+  a3 = 6 * s;
+  x3 = c + a3;
+  x = x1 + x2 + x3;
+  return x;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* " 3 "dom2" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-31.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-31.c
new file mode 100644 (file)
index 0000000..edf3129
--- /dev/null
@@ -0,0 +1,27 @@
+/* Verify straight-line strength reduction for add candidates in
+   which the stride is unknown and increments appear that differ
+   only in sign.  Verify the increments are shared.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int s, int c)
+{
+  int a1, a2, a3, a4, x1, x2, x3, x4, x;
+
+  a1 = 2 * s;
+  x1 = c + a1;
+  a2 = 4 * s;  /* incr = +2  */
+  x2 = c + a2;
+  a3 = 7 * s;
+  x3 = c + a3;
+  a4 = 5 * s;  /* incr = -2  */
+  x4 = c + a4;
+  x = x1 + x2 + x3 + x4;
+  return x;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* 2" 1 "optimized" } } */
+/* { dg-final { scan-tree-dump-times " \\* -2" 0 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-5.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-5.c
new file mode 100644 (file)
index 0000000..8612feb
--- /dev/null
@@ -0,0 +1,22 @@
+/* Verify straight-line strength reduction for simple add candidates.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int s, int c)
+{
+  int a1, a2, a3, x1, x2, x3, x;
+
+  a1 = 2 * s;
+  x1 = c + a1;
+  a2 = 4 * s;
+  x2 = c + a2;
+  a3 = 6 * s;
+  x3 = c + a3;
+  x = x1 + x2 + x3;
+  return x;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* " 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-6.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-6.c
new file mode 100644 (file)
index 0000000..56b1677
--- /dev/null
@@ -0,0 +1,25 @@
+/* Verify straight-line strength reduction for simple add candidates,
+   pointer version.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+void
+f (int s, char *c, char *x1, char *x2, char *x3)
+{
+  int a1, a2, a3;
+
+  a1 = 2 * s;
+  x1 = c + a1;
+  *x1 = 1;
+  a2 = 4 * s;
+  x2 = c + a2;
+  *x2 = 2;
+  a3 = 6 * s;
+  x3 = c + a3;
+  *x3 = 3;
+}
+
+/* There will be four ' * ' instances for the parms, one in the code.  */
+/* { dg-final { scan-tree-dump-times " \\* " 5 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-7.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-7.c
new file mode 100644 (file)
index 0000000..ace755f
--- /dev/null
@@ -0,0 +1,22 @@
+/* Verify straight-line strength reduction for simple integer subtraction.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int s, int c)
+{
+  int a1, a2, a3, x1, x2, x3, x;
+
+  a1 = 2 * s;
+  x1 = c - a1;
+  a2 = 4 * s;
+  x2 = c - a2;
+  a3 = 6 * s;
+  x3 = c - a3;
+  x = x1 + x2 + x3;
+  return x;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* " 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-8.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-8.c
new file mode 100644 (file)
index 0000000..b03c52c
--- /dev/null
@@ -0,0 +1,23 @@
+/* Verify straight-line strength reduction for simple pointer subtraction.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int*
+f (int s, int *c)
+{
+  int a1, a2, a3, *x1, *x2, *x3;
+
+  a1 = 2 * s;
+  x1 = c - a1;
+  a2 = 4 * s;
+  x2 = c - a2;
+  a3 = 6 * s;
+  x3 = c - a3;
+  return x1 ? x2 : x3;
+}
+
+/* There are 2 ' * ' instances in the decls (since "int * x3;" is
+   optimized out), 1 parm, 2 in the code.  */
+/* { dg-final { scan-tree-dump-times " \\* " 5 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/slsr-9.c b/gcc/testsuite/gcc.dg/tree-ssa/slsr-9.c
new file mode 100644 (file)
index 0000000..3a38873
--- /dev/null
@@ -0,0 +1,23 @@
+/* Verify straight-line strength reduction for simple integer addition
+   with stride reversed.  */
+
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized" } */
+
+int
+f (int s, int c)
+{
+  int a1, a2, a3, x1, x2, x3, x;
+
+  a1 = 2 * s;
+  x1 = a1 + c;
+  a2 = 4 * s;
+  x2 = a2 + c;
+  a3 = 6 * s;
+  x3 = a3 + c;
+  x = x1 + x2 + x3;
+  return x;
+}
+
+/* { dg-final { scan-tree-dump-times " \\* " 1 "optimized" } } */
+/* { dg-final { cleanup-tree-dump "optimized" } } */