ipa: Multiple predicates for loop properties, with frequencies
authorMartin Jambor <mjambor@suse.cz>
Fri, 2 Oct 2020 16:41:35 +0000 (18:41 +0200)
committerMartin Jambor <mjambor@suse.cz>
Fri, 2 Oct 2020 16:41:35 +0000 (18:41 +0200)
This patch enhances the ability of IPA to reason under what conditions
loops in a function have known iteration counts or strides because it
replaces single predicates which currently hold conjunction of
predicates for all loops with vectors capable of holding multiple
predicates, each with a cumulative frequency of loops with the
property.

This second property is then used by IPA-CP to much more aggressively
boost its heuristic score for cloning opportunities which make
iteration counts or strides of frequent loops compile time constant.

gcc/ChangeLog:

2020-09-03  Martin Jambor  <mjambor@suse.cz>

* ipa-fnsummary.h (ipa_freqcounting_predicate): New type.
(ipa_fn_summary): Change the type of loop_iterations and loop_strides
to vectors of ipa_freqcounting_predicate.
(ipa_fn_summary::ipa_fn_summary): Construct the new vectors.
(ipa_call_estimates): New fields loops_with_known_iterations and
loops_with_known_strides.
* ipa-cp.c (hint_time_bonus): Multiply param_ipa_cp_loop_hint_bonus
with the expected frequencies of loops with known iteration count or
stride.
* ipa-fnsummary.c (add_freqcounting_predicate): New function.
(ipa_fn_summary::~ipa_fn_summary): Release the new vectors instead of
just two predicates.
(remap_hint_predicate_after_duplication): Replace with function
remap_freqcounting_preds_after_dup.
(ipa_fn_summary_t::duplicate): Use it or duplicate new vectors.
(ipa_dump_fn_summary): Dump the new vectors.
(analyze_function_body): Compute the loop property vectors.
(ipa_call_context::estimate_size_and_time): Calculate also
loops_with_known_iterations and loops_with_known_strides.  Adjusted
dumping accordinly.
(remap_hint_predicate): Replace with function
remap_freqcounting_predicate.
(ipa_merge_fn_summary_after_inlining): Use it.
(inline_read_section): Stream loopcounting vectors instead of two
simple predicates.
(ipa_fn_summary_write): Likewise.
* params.opt (ipa-max-loop-predicates): New parameter.
* doc/invoke.texi (ipa-max-loop-predicates): Document new param.

gcc/testsuite/ChangeLog:

2020-09-03  Martin Jambor  <mjambor@suse.cz>

* gcc.dg/ipa/ipcp-loophint-1.c: New test.

gcc/doc/invoke.texi
gcc/ipa-cp.c
gcc/ipa-fnsummary.c
gcc/ipa-fnsummary.h
gcc/params.opt
gcc/testsuite/gcc.dg/ipa/ipcp-loophint-1.c [new file with mode: 0644]

index a5ecb1b49372f8029d5bf0625b0e0c762f1951ae..f623467b7637af3db67a4834bff883b4e75e1475 100644 (file)
@@ -13453,6 +13453,10 @@ of iterations of a loop known, it adds a bonus of
 @option{ipa-cp-loop-hint-bonus} to the profitability score of
 the candidate.
 
+@item ipa-max-loop-predicates
+The maximum number of different predicates IPA will use to describe when
+loops in a function have known properties.
+
 @item ipa-max-aa-steps
 During its analysis of function bodies, IPA-CP employs alias analysis
 in order to track values pointed to by function parameters.  In order
index 77c84a6ed5dfbd07435e2aba30087a7756a65b2d..f6320c787de3eddb1f3dad34904fe30ad24ea3ed 100644 (file)
@@ -3205,6 +3205,15 @@ hint_time_bonus (cgraph_node *node, const ipa_call_estimates &estimates)
   ipa_hints hints = estimates.hints;
   if (hints & (INLINE_HINT_loop_iterations | INLINE_HINT_loop_stride))
     result += opt_for_fn (node->decl, param_ipa_cp_loop_hint_bonus);
+
+  sreal bonus_for_one = opt_for_fn (node->decl, param_ipa_cp_loop_hint_bonus);
+
+  if (hints & INLINE_HINT_loop_iterations)
+    result += (estimates.loops_with_known_iterations * bonus_for_one).to_int ();
+
+  if (hints & INLINE_HINT_loop_stride)
+    result += (estimates.loops_with_known_strides * bonus_for_one).to_int ();
+
   return result;
 }
 
index f89f1583b6b225ba30527f54d7c89d124d7453b8..771f432ebecac7308e268665da5735748434c3d7 100644 (file)
@@ -310,6 +310,36 @@ set_hint_predicate (predicate **p, predicate new_predicate)
     }
 }
 
+/* Find if NEW_PREDICATE is already in V and if so, increment its freq.
+   Otherwise add a new item to the vector with this predicate and frerq equal
+   to add_freq, unless the number of predicates would exceed MAX_NUM_PREDICATES
+   in which case the function does nothing.  */
+
+static void
+add_freqcounting_predicate (vec<ipa_freqcounting_predicate, va_gc> **v,
+                           const predicate &new_predicate, sreal add_freq,
+                           unsigned max_num_predicates)
+{
+  if (new_predicate == false || new_predicate == true)
+    return;
+  ipa_freqcounting_predicate *f;
+  for (int i = 0; vec_safe_iterate (*v, i, &f); i++)
+    if (new_predicate == f->predicate)
+      {
+       f->freq += add_freq;
+       return;
+      }
+  if (vec_safe_length (*v) >= max_num_predicates)
+    /* Too many different predicates to account for.  */
+    return;
+
+  ipa_freqcounting_predicate fcp;
+  fcp.predicate = NULL;
+  set_hint_predicate (&fcp.predicate, new_predicate);
+  fcp.freq = add_freq;
+  vec_safe_push (*v, fcp);
+  return;
+}
 
 /* Compute what conditions may or may not hold given information about
    parameters.  RET_CLAUSE returns truths that may hold in a specialized copy,
@@ -710,13 +740,17 @@ ipa_call_summary::~ipa_call_summary ()
 
 ipa_fn_summary::~ipa_fn_summary ()
 {
-  if (loop_iterations)
-    edge_predicate_pool.remove (loop_iterations);
-  if (loop_stride)
-    edge_predicate_pool.remove (loop_stride);
+  unsigned len = vec_safe_length (loop_iterations);
+  for (unsigned i = 0; i < len; i++)
+    edge_predicate_pool.remove ((*loop_iterations)[i].predicate);
+  len = vec_safe_length (loop_strides);
+  for (unsigned i = 0; i < len; i++)
+    edge_predicate_pool.remove ((*loop_strides)[i].predicate);
   vec_free (conds);
   vec_free (size_time_table);
   vec_free (call_size_time_table);
+  vec_free (loop_iterations);
+  vec_free (loop_strides);
 }
 
 void
@@ -729,24 +763,33 @@ ipa_fn_summary_t::remove_callees (cgraph_node *node)
     ipa_call_summaries->remove (e);
 }
 
-/* Same as remap_predicate_after_duplication but handle hint predicate *P.
-   Additionally care about allocating new memory slot for updated predicate
-   and set it to NULL when it becomes true or false (and thus uninteresting).
- */
+/* Duplicate predicates in loop hint vector, allocating memory for them and
+   remove and deallocate any uninteresting (true or false) ones.  Return the
+   result.  */
 
-static void
-remap_hint_predicate_after_duplication (predicate **p,
-                                       clause_t possible_truths)
+static vec<ipa_freqcounting_predicate, va_gc> *
+remap_freqcounting_preds_after_dup (vec<ipa_freqcounting_predicate, va_gc> *v,
+                                   clause_t possible_truths)
 {
-  predicate new_predicate;
+  if (vec_safe_length (v) == 0)
+    return NULL;
 
-  if (!*p)
-    return;
+  vec<ipa_freqcounting_predicate, va_gc> *res = v->copy ();
+  int len = res->length();
+  for (int i = len - 1; i >= 0; i--)
+    {
+      predicate new_predicate
+       = (*res)[i].predicate->remap_after_duplication (possible_truths);
+      /* We do not want to free previous predicate; it is used by node
+        origin.  */
+      (*res)[i].predicate = NULL;
+      set_hint_predicate (&(*res)[i].predicate, new_predicate);
+
+      if (!(*res)[i].predicate)
+       res->unordered_remove (i);
+    }
 
-  new_predicate = (*p)->remap_after_duplication (possible_truths);
-  /* We do not want to free previous predicate; it is used by node origin.  */
-  *p = NULL;
-  set_hint_predicate (p, new_predicate);
+  return res;
 }
 
 
@@ -859,9 +902,11 @@ ipa_fn_summary_t::duplicate (cgraph_node *src,
            optimized_out_size += es->call_stmt_size * ipa_fn_summary::size_scale;
          edge_set_predicate (edge, &new_predicate);
        }
-      remap_hint_predicate_after_duplication (&info->loop_iterations,
+      info->loop_iterations
+       = remap_freqcounting_preds_after_dup (info->loop_iterations,
                                              possible_truths);
-      remap_hint_predicate_after_duplication (&info->loop_stride,
+      info->loop_strides
+       = remap_freqcounting_preds_after_dup (info->loop_strides,
                                              possible_truths);
 
       /* If inliner or someone after inliner will ever start producing
@@ -873,17 +918,21 @@ ipa_fn_summary_t::duplicate (cgraph_node *src,
   else
     {
       info->size_time_table = vec_safe_copy (info->size_time_table);
-      if (info->loop_iterations)
+      info->loop_iterations = vec_safe_copy (info->loop_iterations);
+      info->loop_strides = vec_safe_copy (info->loop_strides);
+
+      ipa_freqcounting_predicate *f;
+      for (int i = 0; vec_safe_iterate (info->loop_iterations, i, &f); i++)
        {
-         predicate p = *info->loop_iterations;
-         info->loop_iterations = NULL;
-         set_hint_predicate (&info->loop_iterations, p);
+         predicate p = *f->predicate;
+         f->predicate = NULL;
+         set_hint_predicate (&f->predicate, p);
        }
-      if (info->loop_stride)
+      for (int i = 0; vec_safe_iterate (info->loop_strides, i, &f); i++)
        {
-         predicate p = *info->loop_stride;
-         info->loop_stride = NULL;
-         set_hint_predicate (&info->loop_stride, p);
+         predicate p = *f->predicate;
+         f->predicate = NULL;
+         set_hint_predicate (&f->predicate, p);
        }
     }
   if (!dst->inlined_to)
@@ -1045,15 +1094,28 @@ ipa_dump_fn_summary (FILE *f, struct cgraph_node *node)
                }
              fprintf (f, "\n");
            }
-         if (s->loop_iterations)
+         ipa_freqcounting_predicate *fcp;
+         bool first_fcp = true;
+         for (int i = 0; vec_safe_iterate (s->loop_iterations, i, &fcp); i++)
            {
-             fprintf (f, "  loop iterations:");
-             s->loop_iterations->dump (f, s->conds);
+             if (first_fcp)
+               {
+                 fprintf (f, "  loop iterations:");
+                 first_fcp = false;
+               }
+             fprintf (f, "  %3.2f for ", fcp->freq.to_double ());
+             fcp->predicate->dump (f, s->conds);
            }
-         if (s->loop_stride)
+         first_fcp = true;
+         for (int i = 0; vec_safe_iterate (s->loop_strides, i, &fcp); i++)
            {
-             fprintf (f, "  loop stride:");
-             s->loop_stride->dump (f, s->conds);
+             if (first_fcp)
+               {
+                 fprintf (f, "  loop strides:");
+                 first_fcp = false;
+               }
+             fprintf (f, "  %3.2f for :", fcp->freq.to_double ());
+             fcp->predicate->dump (f, s->conds);
            }
          fprintf (f, "  calls:\n");
          dump_ipa_call_summary (f, 4, node, s);
@@ -2543,12 +2605,13 @@ analyze_function_body (struct cgraph_node *node, bool early)
 
   if (fbi.info)
     compute_bb_predicates (&fbi, node, info, params_summary);
+  const profile_count entry_count = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count;
   order = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
   nblocks = pre_and_rev_post_order_compute (NULL, order, false);
   for (n = 0; n < nblocks; n++)
     {
       bb = BASIC_BLOCK_FOR_FN (cfun, order[n]);
-      freq = bb->count.to_sreal_scale (ENTRY_BLOCK_PTR_FOR_FN (cfun)->count);
+      freq = bb->count.to_sreal_scale (entry_count);
       if (clobber_only_eh_bb_p (bb))
        {
          if (dump_file && (dump_flags & TDF_DETAILS))
@@ -2790,23 +2853,28 @@ analyze_function_body (struct cgraph_node *node, bool early)
 
   if (nonconstant_names.exists () && !early)
     {
+      ipa_fn_summary *s = ipa_fn_summaries->get (node);
       class loop *loop;
-      predicate loop_iterations = true;
-      predicate loop_stride = true;
+      unsigned max_loop_predicates = opt_for_fn (node->decl,
+                                                param_ipa_max_loop_predicates);
 
       if (dump_file && (dump_flags & TDF_DETAILS))
        flow_loops_dump (dump_file, NULL, 0);
       scev_initialize ();
       FOR_EACH_LOOP (loop, 0)
        {
+         predicate loop_iterations = true;
+         sreal header_freq;
          edge ex;
          unsigned int j;
          class tree_niter_desc niter_desc;
-         if (loop->header->aux)
-           bb_predicate = *(predicate *) loop->header->aux;
-         else
-           bb_predicate = false;
+         if (!loop->header->aux)
+           continue;
 
+         profile_count phdr_count = loop_preheader_edge (loop)->count ();
+         sreal phdr_freq = phdr_count.to_sreal_scale (entry_count);
+
+         bb_predicate = *(predicate *) loop->header->aux;
          auto_vec<edge> exits = get_loop_exit_edges (loop);
          FOR_EACH_VEC_ELT (exits, j, ex)
            if (number_of_iterations_exit (loop, ex, &niter_desc, false)
@@ -2821,10 +2889,10 @@ analyze_function_body (struct cgraph_node *node, bool early)
                will_be_nonconstant = bb_predicate & will_be_nonconstant;
              if (will_be_nonconstant != true
                  && will_be_nonconstant != false)
-               /* This is slightly inprecise.  We may want to represent each
-                  loop with independent predicate.  */
                loop_iterations &= will_be_nonconstant;
            }
+         add_freqcounting_predicate (&s->loop_iterations, loop_iterations,
+                                     phdr_freq, max_loop_predicates);
        }
 
       /* To avoid quadratic behavior we analyze stride predicates only
@@ -2833,14 +2901,17 @@ analyze_function_body (struct cgraph_node *node, bool early)
       for (loop = loops_for_fn (cfun)->tree_root->inner;
           loop != NULL; loop = loop->next)
        {
+         predicate loop_stride = true;
          basic_block *body = get_loop_body (loop);
+         profile_count phdr_count = loop_preheader_edge (loop)->count ();
+         sreal phdr_freq = phdr_count.to_sreal_scale (entry_count);
          for (unsigned i = 0; i < loop->num_nodes; i++)
            {
              gimple_stmt_iterator gsi;
-             if (body[i]->aux)
-               bb_predicate = *(predicate *) body[i]->aux;
-             else
-               bb_predicate = false;
+             if (!body[i]->aux)
+               continue;
+
+             bb_predicate = *(predicate *) body[i]->aux;
              for (gsi = gsi_start_bb (body[i]); !gsi_end_p (gsi);
                   gsi_next (&gsi))
                {
@@ -2869,16 +2940,13 @@ analyze_function_body (struct cgraph_node *node, bool early)
                    will_be_nonconstant = bb_predicate & will_be_nonconstant;
                  if (will_be_nonconstant != true
                      && will_be_nonconstant != false)
-                   /* This is slightly inprecise.  We may want to represent
-                      each loop with independent predicate.  */
                    loop_stride = loop_stride & will_be_nonconstant;
                }
            }
+         add_freqcounting_predicate (&s->loop_strides, loop_stride,
+                                     phdr_freq, max_loop_predicates);
          free (body);
        }
-      ipa_fn_summary *s = ipa_fn_summaries->get (node);
-      set_hint_predicate (&s->loop_iterations, loop_iterations);
-      set_hint_predicate (&s->loop_stride, loop_stride);
       scev_finalize ();
     }
   FOR_ALL_BB_FN (bb, my_function)
@@ -3551,6 +3619,8 @@ ipa_call_context::estimate_size_and_time (ipa_call_estimates *estimates,
   sreal time = 0;
   int min_size = 0;
   ipa_hints hints = 0;
+  sreal loops_with_known_iterations = 0;
+  sreal loops_with_known_strides = 0;
   int i;
 
   if (dump_file && (dump_flags & TDF_DETAILS))
@@ -3643,16 +3713,27 @@ ipa_call_context::estimate_size_and_time (ipa_call_estimates *estimates,
 
   if (est_hints)
     {
-      if (info->loop_iterations
-         && !info->loop_iterations->evaluate (m_possible_truths))
-       hints |= INLINE_HINT_loop_iterations;
-      if (info->loop_stride
-         && !info->loop_stride->evaluate (m_possible_truths))
-       hints |= INLINE_HINT_loop_stride;
       if (info->scc_no)
        hints |= INLINE_HINT_in_scc;
       if (DECL_DECLARED_INLINE_P (m_node->decl))
        hints |= INLINE_HINT_declared_inline;
+
+      ipa_freqcounting_predicate *fcp;
+      for (i = 0; vec_safe_iterate (info->loop_iterations, i, &fcp); i++)
+       if (!fcp->predicate->evaluate (m_possible_truths))
+         {
+           hints |= INLINE_HINT_loop_iterations;
+           loops_with_known_iterations += fcp->freq;
+         }
+      estimates->loops_with_known_iterations = loops_with_known_iterations;
+
+      for (i = 0; vec_safe_iterate (info->loop_strides, i, &fcp); i++)
+       if (!fcp->predicate->evaluate (m_possible_truths))
+         {
+           hints |= INLINE_HINT_loop_stride;
+           loops_with_known_strides += fcp->freq;
+         }
+      estimates->loops_with_known_strides = loops_with_known_strides;
     }
 
   size = RDIV (size, ipa_fn_summary::size_scale);
@@ -3660,12 +3741,15 @@ ipa_call_context::estimate_size_and_time (ipa_call_estimates *estimates,
 
   if (dump_file && (dump_flags & TDF_DETAILS))
     {
+      fprintf (dump_file, "\n   size:%i", (int) size);
       if (est_times)
-       fprintf (dump_file, "\n   size:%i time:%f nonspec time:%f\n",
-                (int) size, time.to_double (),
-                nonspecialized_time.to_double ());
-      else
-       fprintf (dump_file, "\n   size:%i (time not estimated)\n", (int) size);
+       fprintf (dump_file, " time:%f nonspec time:%f",
+                time.to_double (), nonspecialized_time.to_double ());
+      if (est_hints)
+       fprintf (dump_file, " loops with known iterations:%f "
+                "known strides:%f", loops_with_known_iterations.to_double (),
+                loops_with_known_strides.to_double ());
+      fprintf (dump_file, "\n");
     }
   if (est_times)
     {
@@ -3865,32 +3949,29 @@ remap_edge_summaries (struct cgraph_edge *inlined_edge,
     }
 }
 
-/* Same as remap_predicate, but set result into hint *HINT.  */
+/* Run remap_after_inlining on each predicate in V.  */
 
 static void
-remap_hint_predicate (class ipa_fn_summary *info,
-                     class ipa_node_params *params_summary,
-                     class ipa_fn_summary *callee_info,
-                     predicate **hint,
-                     vec<int> operand_map,
-                     vec<int> offset_map,
-                     clause_t possible_truths,
-                     predicate *toplev_predicate)
-{
-  predicate p;
+remap_freqcounting_predicate (class ipa_fn_summary *info,
+                             class ipa_node_params *params_summary,
+                             class ipa_fn_summary *callee_info,
+                             vec<ipa_freqcounting_predicate, va_gc> *v,
+                             vec<int> operand_map,
+                             vec<int> offset_map,
+                             clause_t possible_truths,
+                             predicate *toplev_predicate)
 
-  if (!*hint)
-    return;
-  p = (*hint)->remap_after_inlining
-                        (info, params_summary, callee_info,
-                         operand_map, offset_map,
-                         possible_truths, *toplev_predicate);
-  if (p != false && p != true)
+{
+  ipa_freqcounting_predicate *fcp;
+  for (int i = 0; vec_safe_iterate (v, i, &fcp); i++)
     {
-      if (!*hint)
-       set_hint_predicate (hint, p);
-      else
-       **hint &= p;
+      predicate p
+       = fcp->predicate->remap_after_inlining (info, params_summary,
+                                               callee_info, operand_map,
+                                               offset_map, possible_truths,
+                                               *toplev_predicate);
+      if (p != false && p != true)
+       *fcp->predicate &= p;
     }
 }
 
@@ -3998,12 +4079,12 @@ ipa_merge_fn_summary_after_inlining (struct cgraph_edge *edge)
   remap_edge_summaries (edge, edge->callee, info, params_summary,
                        callee_info, operand_map,
                        offset_map, clause, &toplev_predicate);
-  remap_hint_predicate (info, params_summary, callee_info,
-                       &callee_info->loop_iterations,
-                       operand_map, offset_map, clause, &toplev_predicate);
-  remap_hint_predicate (info, params_summary, callee_info,
-                       &callee_info->loop_stride,
-                       operand_map, offset_map, clause, &toplev_predicate);
+  remap_freqcounting_predicate (info, params_summary, callee_info,
+                               info->loop_iterations, operand_map,
+                               offset_map, clause, &toplev_predicate);
+  remap_freqcounting_predicate (info, params_summary, callee_info,
+                               info->loop_strides, operand_map,
+                               offset_map, clause, &toplev_predicate);
 
   HOST_WIDE_INT stack_frame_offset = ipa_get_stack_frame_offset (edge->callee);
   HOST_WIDE_INT peak = stack_frame_offset + callee_info->estimated_stack_size;
@@ -4334,12 +4415,34 @@ inline_read_section (struct lto_file_decl_data *file_data, const char *data,
            info->size_time_table->quick_push (e);
        }
 
-      p.stream_in (&ib);
-      if (info)
-        set_hint_predicate (&info->loop_iterations, p);
-      p.stream_in (&ib);
-      if (info)
-        set_hint_predicate (&info->loop_stride, p);
+      count2 = streamer_read_uhwi (&ib);
+      for (j = 0; j < count2; j++)
+       {
+         p.stream_in (&ib);
+         sreal fcp_freq = sreal::stream_in (&ib);
+         if (info)
+           {
+             ipa_freqcounting_predicate fcp;
+             fcp.predicate = NULL;
+             set_hint_predicate (&fcp.predicate, p);
+             fcp.freq = fcp_freq;
+             vec_safe_push (info->loop_iterations, fcp);
+           }
+       }
+      count2 = streamer_read_uhwi (&ib);
+      for (j = 0; j < count2; j++)
+       {
+         p.stream_in (&ib);
+         sreal fcp_freq = sreal::stream_in (&ib);
+         if (info)
+           {
+             ipa_freqcounting_predicate fcp;
+             fcp.predicate = NULL;
+             set_hint_predicate (&fcp.predicate, p);
+             fcp.freq = fcp_freq;
+             vec_safe_push (info->loop_strides, fcp);
+           }
+       }
       for (e = node->callees; e; e = e->next_callee)
        read_ipa_call_summary (&ib, e, info != NULL);
       for (e = node->indirect_calls; e; e = e->next_callee)
@@ -4502,14 +4605,19 @@ ipa_fn_summary_write (void)
              e->exec_predicate.stream_out (ob);
              e->nonconst_predicate.stream_out (ob);
            }
-         if (info->loop_iterations)
-           info->loop_iterations->stream_out (ob);
-         else
-           streamer_write_uhwi (ob, 0);
-         if (info->loop_stride)
-           info->loop_stride->stream_out (ob);
-         else
-           streamer_write_uhwi (ob, 0);
+         ipa_freqcounting_predicate *fcp;
+         streamer_write_uhwi (ob, vec_safe_length (info->loop_iterations));
+         for (i = 0; vec_safe_iterate (info->loop_iterations, i, &fcp); i++)
+           {
+             fcp->predicate->stream_out (ob);
+             fcp->freq.stream_out (ob);
+           }
+         streamer_write_uhwi (ob, vec_safe_length (info->loop_strides));
+         for (i = 0; vec_safe_iterate (info->loop_strides, i, &fcp); i++)
+           {
+             fcp->predicate->stream_out (ob);
+             fcp->freq.stream_out (ob);
+           }
          for (edge = cnode->callees; edge; edge = edge->next_callee)
            write_ipa_call_summary (ob, edge);
          for (edge = cnode->indirect_calls; edge; edge = edge->next_callee)
index ccb6b432f0b9b53cca05a24eb1560057daa0622d..f4dd5b85ab9045ff10b86e0e25ecbe6b636e0a6e 100644 (file)
@@ -101,6 +101,19 @@ public:
   }
 };
 
+/* Structure to capture how frequently some interesting events occur given a
+   particular predicate.  The structure is used to estimate how often we
+   encounter loops with known iteration count or stride in various
+   contexts.  */
+
+struct GTY(()) ipa_freqcounting_predicate
+{
+  /* The described event happens with this frequency... */
+  sreal freq;
+  /* ...when this predicate evaluates to false. */
+  class predicate * GTY((skip)) predicate;
+};
+
 /* Function inlining information.  */
 class GTY(()) ipa_fn_summary
 {
@@ -112,8 +125,9 @@ public:
       inlinable (false), single_caller (false),
       fp_expressions (false), estimated_stack_size (false),
       time (0), conds (NULL),
-      size_time_table (NULL), call_size_time_table (NULL), loop_iterations (NULL),
-      loop_stride (NULL), growth (0), scc_no (0)
+      size_time_table (NULL), call_size_time_table (NULL),
+      loop_iterations (NULL), loop_strides (NULL),
+      growth (0), scc_no (0)
   {
   }
 
@@ -125,7 +139,7 @@ public:
     estimated_stack_size (s.estimated_stack_size),
     time (s.time), conds (s.conds), size_time_table (s.size_time_table),
     call_size_time_table (NULL),
-    loop_iterations (s.loop_iterations), loop_stride (s.loop_stride),
+    loop_iterations (s.loop_iterations), loop_strides (s.loop_strides),
     growth (s.growth), scc_no (s.scc_no)
   {}
 
@@ -164,12 +178,10 @@ public:
   vec<size_time_entry, va_gc> *size_time_table;
   vec<size_time_entry, va_gc> *call_size_time_table;
 
-  /* Predicate on when some loop in the function becomes to have known
-     bounds.   */
-  predicate * GTY((skip)) loop_iterations;
-  /* Predicate on when some loop in the function becomes to have known
-     stride.   */
-  predicate * GTY((skip)) loop_stride;
+  /* Predicates on when some loops in the function can have known bounds.  */
+  vec<ipa_freqcounting_predicate, va_gc> *loop_iterations;
+  /* Predicates on when some loops in the function can have known strides.  */
+  vec<ipa_freqcounting_predicate, va_gc> *loop_strides;
   /* Estimated growth for inlining all copies of the function before start
      of small functions inlining.
      This value will get out of date as the callers are duplicated, but
@@ -308,6 +320,14 @@ struct ipa_call_estimates
 
   /* Further discovered reasons why to inline or specialize the give calls.  */
   ipa_hints hints;
+
+  /* Frequency how often a loop with known number of iterations is encountered.
+     Calculated with hints.  */
+  sreal loops_with_known_iterations;
+
+  /* Frequency how often a loop with known strides is encountered.  Calculated
+     with hints.  */
+  sreal loops_with_known_strides;
 };
 
 class ipa_cached_call_context;
index a232161ac05929e094065ae49f0108d92657242b..c0b94bbb96cd8e28afa8418348d74daa22d7e5be 100644 (file)
@@ -230,6 +230,10 @@ Maximum number of aggregate content items for a parameter in jump functions and
 Common Joined UInteger Var(param_ipa_max_param_expr_ops) Init(10) Param Optimization
 Maximum number of operations in a parameter expression that can be handled by IPA analysis.
 
+-param=ipa-max-loop-predicates=
+Common Joined UInteger Var(param_ipa_max_loop_predicates) Init(16) Param Optimization
+Maximum number of different predicates used to track properties of loops in IPA analysis.
+
 -param=ipa-max-switch-predicate-bounds=
 Common Joined UInteger Var(param_ipa_max_switch_predicate_bounds) Init(5) Param Optimization
 Maximal number of boundary endpoints of case ranges of switch statement used during IPA function summary generation.
diff --git a/gcc/testsuite/gcc.dg/ipa/ipcp-loophint-1.c b/gcc/testsuite/gcc.dg/ipa/ipcp-loophint-1.c
new file mode 100644 (file)
index 0000000..6d049af
--- /dev/null
@@ -0,0 +1,29 @@
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-ipa-cp-details"  } */
+
+extern int *o, *p, *q, *r;
+
+#define FUNCTIONS fa(), fb(), fc(), fd(), fe(), ff(), fg()
+
+extern void FUNCTIONS;
+
+void foo (int c)
+{
+  FUNCTIONS;
+  FUNCTIONS;
+  for (int i = 0; i < 100; i++)
+    {
+      for (int j = 0; j < c; j++)
+       o[i] = p[i] + q[i] * r[i];
+    }
+  FUNCTIONS;
+  FUNCTIONS;
+}
+
+void bar()
+{
+  foo (8);
+  p[4]++;
+}
+
+/* { dg-final { scan-ipa-dump {with known iterations:[1-9]} "cp"  } } */