Make inliner loop hints more agressive

This patch makes loop inline hints more agressive.  If we know iteration
count or stride, we currently assume improvement in time relative to
preheader count.  I changed it to header count, since this knowledge
is supposed to likely help unrolling and vectorizing which brings
benefits relative to that.

	* ipa-fnsummary.cc (analyze_function_body): For loop
	heuristics use header count instead of preheader count.

diff --git a/gcc/ipa-fnsummary.cc b/gcc/ipa-fnsummary.cc
index 4c062fe8a0e286b07f98817d84d1b4d9671a34c9..48343a7696808f509ad25263ac5b5a66f7fe492d 100644 (file)
--- a/gcc/ipa-fnsummary.cc
+++ b/gcc/ipa-fnsummary.cc
@@ -3236,8 +3236,8 @@ analyze_function_body (struct cgraph_node *node, bool early)
          if (!loop->header->aux)
            continue;
 
-         profile_count phdr_count = loop_preheader_edge (loop)->count ();
-         sreal phdr_freq = phdr_count.to_sreal_scale (entry_count);
+         profile_count hdr_count = loop->header->count;
+         sreal hdr_freq = hdr_count.to_sreal_scale (entry_count);
 
          bb_predicate = *(ipa_predicate *)loop->header->aux;
          auto_vec<edge> exits = get_loop_exit_edges (loop);
@@ -3257,7 +3257,7 @@ analyze_function_body (struct cgraph_node *node, bool early)
                loop_iterations &= will_be_nonconstant;
            }
          add_freqcounting_predicate (&s->loop_iterations, loop_iterations,
-                                     phdr_freq, max_loop_predicates);
+                                     hdr_freq, max_loop_predicates);
        }
 
       /* To avoid quadratic behavior we analyze stride predicates only
@@ -3268,8 +3268,8 @@ analyze_function_body (struct cgraph_node *node, bool early)
        {
          ipa_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);
+         profile_count hdr_count = loop->header->count;
+         sreal hdr_freq = hdr_count.to_sreal_scale (entry_count);
          for (unsigned i = 0; i < loop->num_nodes; i++)
            {
              gimple_stmt_iterator gsi;
@@ -3309,7 +3309,7 @@ analyze_function_body (struct cgraph_node *node, bool early)
                }
            }
          add_freqcounting_predicate (&s->loop_strides, loop_stride,
-                                     phdr_freq, max_loop_predicates);
+                                     hdr_freq, max_loop_predicates);
          free (body);
        }
       scev_finalize ();