--- a/gcc/tree-ssa-loop-prefetch.c	Tue May 25 18:58:51 2010 +0900
+++ b/gcc/tree-ssa-loop-prefetch.c	Tue Mar 22 17:18:12 2011 +0900
@@ -25,24 +25,27 @@
 #include "tm_p.h"
 #include "basic-block.h"
 #include "output.h"
-#include "diagnostic.h"
 #include "tree-pretty-print.h"
 #include "tree-flow.h"
 #include "tree-dump.h"
 #include "timevar.h"
 #include "cfgloop.h"
-#include "expr.h"
 #include "tree-pass.h"
 #include "insn-config.h"
 #include "recog.h"
 #include "hashtab.h"
 #include "tree-chrec.h"
 #include "tree-scalar-evolution.h"
-#include "toplev.h"
+#include "diagnostic-core.h"
 #include "params.h"
 #include "langhooks.h"
 #include "tree-inline.h"
 #include "tree-data-ref.h"
+
+
+/* FIXME: Needed for optabs, but this should all be moved to a TBD interface
+   between the GIMPLE and RTL worlds.  */
+#include "expr.h"
 #include "optabs.h"
 
 /* This pass inserts prefetch instructions to optimize cache usage during
@@ -76,7 +79,7 @@
        (2) has PREFETCH_MOD 64
        (3) has PREFETCH_MOD 4
        (4) has PREFETCH_MOD 1.  We do not set PREFETCH_BEFORE here, since
-           the cache line accessed by (4) is the same with probability only
+           the cache line accessed by (5) is the same with probability only
 	   7/32.
        (5) has PREFETCH_MOD 1 as well.
 
@@ -106,19 +109,29 @@
       prefetch instructions with guards in cases where 5) was not sufficient
       to satisfy the constraints?
 
-   The function is_loop_prefetching_profitable() implements a cost model
-   to determine if prefetching is profitable for a given loop. The cost
-   model has two heuristcs:
-   1. A heuristic that determines whether the given loop has enough CPU
-      ops that can be overlapped with cache missing memory ops.
-      If not, the loop won't benefit from prefetching. This is implemented
-      by requirung the ratio between the instruction count and the mem ref
-      count to be above a certain minimum.
-   2. A heuristic that disables prefetching in a loop with an unknown trip
-      count if the prefetching cost is above a certain limit. The relative
-      prefetching cost is estimated by taking the ratio between the
-      prefetch count and the total intruction count (this models the I-cache
-      cost).
+   A cost model is implemented to determine whether or not prefetching is
+   profitable for a given loop.  The cost model has three heuristics:
+
+   1. Function trip_count_to_ahead_ratio_too_small_p implements a
+      heuristic that determines whether or not the loop has too few
+      iterations (compared to ahead).  Prefetching is not likely to be
+      beneficial if the trip count to ahead ratio is below a certain
+      minimum.
+
+   2. Function mem_ref_count_reasonable_p implements a heuristic that
+      determines whether the given loop has enough CPU ops that can be
+      overlapped with cache missing memory ops.  If not, the loop
+      won't benefit from prefetching.  In the implementation,
+      prefetching is not considered beneficial if the ratio between
+      the instruction count and the mem ref count is below a certain
+      minimum.
+
+   3. Function insn_to_prefetch_ratio_too_small_p implements a
+      heuristic that disables prefetching in a loop if the prefetching
+      cost is above a certain limit.  The relative prefetching cost is
+      estimated by taking the ratio between the prefetch count and the
+      total intruction count (this models the I-cache cost).
+
    The limits used in these heuristics are defined as parameters with
    reasonable default values. Machine-specific default values will be
    added later.
@@ -225,13 +238,21 @@
 
 /* Do not generate a prefetch if the unroll factor is significantly less
    than what is required by the prefetch.  This is to avoid redundant
-   prefetches.  For example, if prefetch_mod is 16 and unroll_factor is
-   1, this means prefetching requires unrolling the loop 16 times, but
-   the loop is not going to be unrolled.  In this case (ratio = 16),
+   prefetches.  For example, when prefetch_mod is 16 and unroll_factor is
+   2, prefetching requires unrolling the loop 16 times, but
+   the loop is actually unrolled twice.  In this case (ratio = 8),
    prefetching is not likely to be beneficial.  */
 
 #ifndef PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO
-#define PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO 8
+#define PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO 4
+#endif
+
+/* Some of the prefetch computations have quadratic complexity.  We want to
+   avoid huge compile times and, therefore, want to limit the amount of
+   memory references per loop where we consider prefetching.  */
+
+#ifndef PREFETCH_MAX_MEM_REFS_PER_LOOP
+#define PREFETCH_MAX_MEM_REFS_PER_LOOP 200
 #endif
 
 /* The memory reference.  */
@@ -401,10 +422,6 @@
   HOST_WIDE_INT idelta = 0, imult = 1;
   affine_iv iv;
 
-  if (TREE_CODE (base) == MISALIGNED_INDIRECT_REF
-      || TREE_CODE (base) == ALIGN_INDIRECT_REF)
-    return false;
-
   if (!simple_iv (ar_data->loop, loop_containing_stmt (ar_data->stmt),
 		  *index, &iv, true))
     return false;
@@ -423,25 +440,24 @@
       ibase = build_int_cst (TREE_TYPE (ibase), 0);
     }
 
+  if (TREE_CODE (base) == ARRAY_REF)
+    {
+      stepsize = array_ref_element_size (base);
+      if (!cst_and_fits_in_hwi (stepsize))
+	return false;
+      imult = int_cst_value (stepsize);
+      step = fold_build2 (MULT_EXPR, sizetype,
+			  fold_convert (sizetype, step),
+			  fold_convert (sizetype, stepsize));
+      idelta *= imult;
+    }
+
   if (*ar_data->step == NULL_TREE)
     *ar_data->step = step;
   else
     *ar_data->step = fold_build2 (PLUS_EXPR, sizetype,
 				  fold_convert (sizetype, *ar_data->step),
 				  fold_convert (sizetype, step));
-  if (TREE_CODE (base) == ARRAY_REF)
-    {
-      stepsize = array_ref_element_size (base);
-      if (!cst_and_fits_in_hwi (stepsize))
-	return false;
-      imult = int_cst_value (stepsize);
-
-      *ar_data->step = fold_build2 (MULT_EXPR, sizetype,
-				    fold_convert (sizetype, *ar_data->step),
-				    fold_convert (sizetype, step));
-      idelta *= imult;
-    }
-
   *ar_data->delta += idelta;
   *index = ibase;
 
@@ -466,10 +482,18 @@
   *step = NULL_TREE;
   *delta = 0;
 
-  /* First strip off the component references.  Ignore bitfields.  */
-  if (TREE_CODE (ref) == COMPONENT_REF
-      && DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1)))
-    ref = TREE_OPERAND (ref, 0);
+  /* First strip off the component references.  Ignore bitfields.
+     Also strip off the real and imagine parts of a complex, so that
+     they can have the same base.  */
+  if (TREE_CODE (ref) == REALPART_EXPR
+      || TREE_CODE (ref) == IMAGPART_EXPR
+      || (TREE_CODE (ref) == COMPONENT_REF
+          && DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1))))
+    {
+      if (TREE_CODE (ref) == IMAGPART_EXPR)
+        *delta += int_size_in_bytes (TREE_TYPE (ref));
+      ref = TREE_OPERAND (ref, 0);
+    }
 
   *ref_p = ref;
 
@@ -511,6 +535,14 @@
   if (step == NULL_TREE)
     return false;
 
+  /* Stop if the address of BASE could not be taken.  */
+  if (may_be_nonaddressable_p (base))
+    return false;
+
+  /* Limit non-constant step prefetching only to the innermost loops.  */
+  if (!cst_and_fits_in_hwi (step) && loop->inner != NULL)
+    return false;
+
   /* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP
      are integer constants.  */
   agrp = find_or_create_group (refs, base, step);
@@ -634,22 +666,29 @@
 /* Given a CACHE_LINE_SIZE and two inductive memory references
    with a common STEP greater than CACHE_LINE_SIZE and an address
    difference DELTA, compute the probability that they will fall
-   in different cache lines.  DISTINCT_ITERS is the number of
-   distinct iterations after which the pattern repeats itself.
+   in different cache lines.  Return true if the computed miss rate
+   is not greater than the ACCEPTABLE_MISS_RATE.  DISTINCT_ITERS is the
+   number of distinct iterations after which the pattern repeats itself.
    ALIGN_UNIT is the unit of alignment in bytes.  */
 
-static int
-compute_miss_rate (unsigned HOST_WIDE_INT cache_line_size,
+static bool
+is_miss_rate_acceptable (unsigned HOST_WIDE_INT cache_line_size,
 		   HOST_WIDE_INT step, HOST_WIDE_INT delta,
 		   unsigned HOST_WIDE_INT distinct_iters,
 		   int align_unit)
 {
   unsigned align, iter;
-  int total_positions, miss_positions, miss_rate;
+  int total_positions, miss_positions, max_allowed_miss_positions;
   int address1, address2, cache_line1, cache_line2;
 
-  total_positions = 0;
+  /* It always misses if delta is greater than or equal to the cache
+     line size.  */
+  if (delta >= (HOST_WIDE_INT) cache_line_size)
+    return false;
+
   miss_positions = 0;
+  total_positions = (cache_line_size / align_unit) * distinct_iters;
+  max_allowed_miss_positions = (ACCEPTABLE_MISS_RATE * total_positions) / 1000;
 
   /* Iterate through all possible alignments of the first
      memory reference within its cache line.  */
@@ -662,12 +701,14 @@
 	address2 = address1 + delta;
 	cache_line1 = address1 / cache_line_size;
 	cache_line2 = address2 / cache_line_size;
-	total_positions += 1;
 	if (cache_line1 != cache_line2)
-	  miss_positions += 1;
+	  {
+	    miss_positions += 1;
+            if (miss_positions > max_allowed_miss_positions)
+	      return false;
+          }
       }
-  miss_rate = 1000 * miss_positions / total_positions;
-  return miss_rate;
+  return true;
 }
 
 /* Prune the prefetch candidate REF using the reuse with BY.
@@ -683,7 +724,6 @@
   HOST_WIDE_INT delta = delta_b - delta_r;
   HOST_WIDE_INT hit_from;
   unsigned HOST_WIDE_INT prefetch_before, prefetch_block;
-  int miss_rate;
   HOST_WIDE_INT reduced_step;
   unsigned HOST_WIDE_INT reduced_prefetch_block;
   tree ref_type;
@@ -782,9 +822,8 @@
   delta %= step;
   ref_type = TREE_TYPE (ref->mem);
   align_unit = TYPE_ALIGN (ref_type) / 8;
-  miss_rate = compute_miss_rate(prefetch_block, step, delta,
-				reduced_prefetch_block, align_unit);
-  if (miss_rate <= ACCEPTABLE_MISS_RATE)
+  if (is_miss_rate_acceptable (prefetch_block, step, delta,
+			       reduced_prefetch_block, align_unit))
     {
       /* Do not reduce prefetch_before if we meet beyond cache size.  */
       if (prefetch_before > L2_CACHE_SIZE_BYTES / PREFETCH_BLOCK)
@@ -798,9 +837,8 @@
   /* Try also the following iteration.  */
   prefetch_before++;
   delta = step - delta;
-  miss_rate = compute_miss_rate(prefetch_block, step, delta,
-				reduced_prefetch_block, align_unit);
-  if (miss_rate <= ACCEPTABLE_MISS_RATE)
+  if (is_miss_rate_acceptable (prefetch_block, step, delta,
+			       reduced_prefetch_block, align_unit))
     {
       if (prefetch_before < ref->prefetch_before)
 	ref->prefetch_before = prefetch_before;
@@ -988,18 +1026,40 @@
   return any;
 }
 
-/* Estimate the number of prefetches in the given GROUPS.  */
+/* Return TRUE if no prefetch is going to be generated in the given
+   GROUPS.  */
+
+static bool
+nothing_to_prefetch_p (struct mem_ref_group *groups)
+{
+  struct mem_ref *ref;
+
+  for (; groups; groups = groups->next)
+    for (ref = groups->refs; ref; ref = ref->next)
+      if (should_issue_prefetch_p (ref))
+	return false;
+
+  return true;
+}
+
+/* Estimate the number of prefetches in the given GROUPS.
+   UNROLL_FACTOR is the factor by which LOOP was unrolled.  */
 
 static int
-estimate_prefetch_count (struct mem_ref_group *groups)
+estimate_prefetch_count (struct mem_ref_group *groups, unsigned unroll_factor)
 {
   struct mem_ref *ref;
+  unsigned n_prefetches;
   int prefetch_count = 0;
 
   for (; groups; groups = groups->next)
     for (ref = groups->refs; ref; ref = ref->next)
       if (should_issue_prefetch_p (ref))
-	  prefetch_count++;
+	{
+	  n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
+			  / ref->prefetch_mod);
+	  prefetch_count += n_prefetches;
+	}
 
   return prefetch_count;
 }
@@ -1031,7 +1091,7 @@
   addr_base = force_gimple_operand_gsi (&bsi, unshare_expr (addr_base),
 					true, NULL, true, GSI_SAME_STMT);
   write_p = ref->write_p ? integer_one_node : integer_zero_node;
-  local = build_int_cst (integer_type_node, nontemporal ? 0 : 3);
+  local = nontemporal ? integer_zero_node : integer_three_node;
 
   for (ap = 0; ap < n_prefetches; ap++)
     {
@@ -1102,7 +1162,7 @@
   if (mode == BLKmode)
     return false;
 
-  code = optab_handler (storent_optab, mode)->insn_code;
+  code = optab_handler (storent_optab, mode);
   return code != CODE_FOR_nothing;
 }
 
@@ -1136,7 +1196,7 @@
   gimple_stmt_iterator bsi;
   unsigned i;
 
-  for (i = 0; VEC_iterate (edge, exits, i, exit); i++)
+  FOR_EACH_VEC_ELT (edge, exits, i, exit)
     {
       call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0);
 
@@ -1173,7 +1233,7 @@
       unsigned i;
       edge exit;
 
-      for (i = 0; VEC_iterate (edge, exits, i, exit); i++)
+      FOR_EACH_VEC_ELT (edge, exits, i, exit)
 	if ((exit->flags & EDGE_ABNORMAL)
 	    && exit->dest == EXIT_BLOCK_PTR)
 	  ret = false;
@@ -1392,7 +1452,7 @@
   strides = XCNEWVEC (HOST_WIDE_INT, n);
   access_fns = DR_ACCESS_FNS (dr);
 
-  for (i = 0; VEC_iterate (tree, access_fns, i, access_fn); i++)
+  FOR_EACH_VEC_ELT (tree, access_fns, i, access_fn)
     {
       /* Keep track of the reference corresponding to the subscript, so that we
 	 know its stride.  */
@@ -1502,7 +1562,8 @@
   for (gr = refs; gr; gr = gr->next)
     for (ref = gr->refs; ref; ref = ref->next)
       {
-	dr = create_data_ref (nest, ref->mem, ref->stmt, !ref->write_p);
+	dr = create_data_ref (nest, loop_containing_stmt (ref->stmt),
+			      ref->mem, ref->stmt, !ref->write_p);
 
 	if (dr)
 	  {
@@ -1514,7 +1575,7 @@
 	  no_other_refs = false;
       }
 
-  for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
+  FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr)
     {
       dist = self_reuse_distance (dr, loop_data_size, n, loop);
       ref = (struct mem_ref *) dr->aux;
@@ -1527,7 +1588,7 @@
 
   compute_all_dependences (datarefs, &dependences, vloops, true);
 
-  for (i = 0; VEC_iterate (ddr_p, dependences, i, dep); i++)
+  FOR_EACH_VEC_ELT (ddr_p, dependences, i, dep)
     {
       if (DDR_ARE_DEPENDENT (dep) == chrec_known)
 	continue;
@@ -1606,24 +1667,51 @@
     }
 }
 
-/* Do a cost-benefit analysis to determine if prefetching is profitable
-   for the current loop given the following parameters:
+/* Determine whether or not the trip count to ahead ratio is too small based
+   on prefitablility consideration.
    AHEAD: the iteration ahead distance,
-   EST_NITER: the estimated trip count,
+   EST_NITER: the estimated trip count.  */
+
+static bool
+trip_count_to_ahead_ratio_too_small_p (unsigned ahead, HOST_WIDE_INT est_niter)
+{
+  /* Assume trip count to ahead ratio is big enough if the trip count could not
+     be estimated at compile time.  */
+  if (est_niter < 0)
+    return false;
+
+  if (est_niter < (HOST_WIDE_INT) (TRIP_COUNT_TO_AHEAD_RATIO * ahead))
+    {
+      if (dump_file && (dump_flags & TDF_DETAILS))
+	fprintf (dump_file,
+		 "Not prefetching -- loop estimated to roll only %d times\n",
+		 (int) est_niter);
+      return true;
+    }
+
+  return false;
+}
+
+/* Determine whether or not the number of memory references in the loop is
+   reasonable based on the profitablity and compilation time considerations.
    NINSNS: estimated number of instructions in the loop,
-   PREFETCH_COUNT: an estimate of the number of prefetches
    MEM_REF_COUNT: total number of memory references in the loop.  */
 
 static bool
-is_loop_prefetching_profitable (unsigned ahead, HOST_WIDE_INT est_niter,
-				unsigned ninsns, unsigned prefetch_count,
-				unsigned mem_ref_count, unsigned unroll_factor)
+mem_ref_count_reasonable_p (unsigned ninsns, unsigned mem_ref_count)
 {
-  int insn_to_mem_ratio, insn_to_prefetch_ratio;
+  int insn_to_mem_ratio;
 
   if (mem_ref_count == 0)
     return false;
 
+  /* Miss rate computation (is_miss_rate_acceptable) and dependence analysis
+     (compute_all_dependences) have high costs based on quadratic complexity.
+     To avoid huge compilation time, we give up prefetching if mem_ref_count
+     is too large.  */
+  if (mem_ref_count > PREFETCH_MAX_MEM_REFS_PER_LOOP)
+    return false;
+
   /* Prefetching improves performance by overlapping cache missing
      memory accesses with CPU operations.  If the loop does not have
      enough CPU operations to overlap with memory operations, prefetching
@@ -1644,6 +1732,21 @@
       return false;
     }
 
+  return true;
+}
+
+/* Determine whether or not the instruction to prefetch ratio in the loop is
+   too small based on the profitablity consideration.
+   NINSNS: estimated number of instructions in the loop,
+   PREFETCH_COUNT: an estimate of the number of prefetches,
+   UNROLL_FACTOR:  the factor to unroll the loop if prefetching.  */
+
+static bool
+insn_to_prefetch_ratio_too_small_p (unsigned ninsns, unsigned prefetch_count,
+                                     unsigned unroll_factor)
+{
+  int insn_to_prefetch_ratio;
+
   /* Prefetching most likely causes performance degradation when the instruction
      to prefetch ratio is too small.  Too many prefetch instructions in a loop
      may reduce the I-cache performance.
@@ -1663,23 +1766,10 @@
         fprintf (dump_file,
 		 "Not prefetching -- instruction to prefetch ratio (%d) too small\n",
 		 insn_to_prefetch_ratio);
-      return false;
+      return true;
     }
 
-  /* Could not do further estimation if the trip count is unknown.  Just assume
-     prefetching is profitable. Too aggressive???  */
-  if (est_niter < 0)
-    return true;
-
-  if (est_niter < (HOST_WIDE_INT) (TRIP_COUNT_TO_AHEAD_RATIO * ahead))
-    {
-      if (dump_file && (dump_flags & TDF_DETAILS))
-	fprintf (dump_file,
-		 "Not prefetching -- loop estimated to roll only %d times\n",
-		 (int) est_niter);
-      return false;
-    }
-  return true;
+  return false;
 }
 
 
@@ -1704,29 +1794,48 @@
       return false;
     }
 
+  /* FIXME: the time should be weighted by the probabilities of the blocks in
+     the loop body.  */
+  time = tree_num_loop_insns (loop, &eni_time_weights);
+  if (time == 0)
+    return false;
+
+  ahead = (PREFETCH_LATENCY + time - 1) / time;
+  est_niter = estimated_loop_iterations_int (loop, false);
+
+  /* Prefetching is not likely to be profitable if the trip count to ahead
+     ratio is too small.  */
+  if (trip_count_to_ahead_ratio_too_small_p (ahead, est_niter))
+    return false;
+
+  ninsns = tree_num_loop_insns (loop, &eni_size_weights);
+
   /* Step 1: gather the memory references.  */
   refs = gather_memory_references (loop, &no_other_refs, &mem_ref_count);
 
+  /* Give up prefetching if the number of memory references in the
+     loop is not reasonable based on profitablity and compilation time
+     considerations.  */
+  if (!mem_ref_count_reasonable_p (ninsns, mem_ref_count))
+    goto fail;
+
   /* Step 2: estimate the reuse effects.  */
   prune_by_reuse (refs);
 
-  prefetch_count = estimate_prefetch_count (refs);
-  if (prefetch_count == 0)
+  if (nothing_to_prefetch_p (refs))
     goto fail;
 
   determine_loop_nest_reuse (loop, refs, no_other_refs);
 
-  /* Step 3: determine the ahead and unroll factor.  */
-
-  /* FIXME: the time should be weighted by the probabilities of the blocks in
-     the loop body.  */
-  time = tree_num_loop_insns (loop, &eni_time_weights);
-  ahead = (PREFETCH_LATENCY + time - 1) / time;
-  est_niter = estimated_loop_iterations_int (loop, false);
-
-  ninsns = tree_num_loop_insns (loop, &eni_size_weights);
+  /* Step 3: determine unroll factor.  */
   unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc,
 					   est_niter);
+
+  /* Estimate prefetch count for the unrolled loop.  */
+  prefetch_count = estimate_prefetch_count (refs, unroll_factor);
+  if (prefetch_count == 0)
+    goto fail;
+
   if (dump_file && (dump_flags & TDF_DETAILS))
     fprintf (dump_file, "Ahead %d, unroll factor %d, trip count "
 	     HOST_WIDE_INT_PRINT_DEC "\n"
@@ -1734,8 +1843,10 @@
 	     ahead, unroll_factor, est_niter,
 	     ninsns, mem_ref_count, prefetch_count);
 
-  if (!is_loop_prefetching_profitable (ahead, est_niter, ninsns, prefetch_count,
-				       mem_ref_count, unroll_factor))
+  /* Prefetching is not likely to be profitable if the instruction to prefetch
+     ratio is too small.  */
+  if (insn_to_prefetch_ratio_too_small_p (ninsns, prefetch_count,
+					  unroll_factor))
     goto fail;
 
   mark_nontemporal_stores (loop, refs);
@@ -1801,10 +1912,8 @@
 
   if (!built_in_decls[BUILT_IN_PREFETCH])
     {
-      tree type = build_function_type (void_type_node,
-				       tree_cons (NULL_TREE,
-						  const_ptr_type_node,
-						  NULL_TREE));
+      tree type = build_function_type_list (void_type_node,
+					    const_ptr_type_node, NULL_TREE);
       tree decl = add_builtin_function ("__builtin_prefetch", type,
 					BUILT_IN_PREFETCH, BUILT_IN_NORMAL,
 					NULL, NULL_TREE);