[Bug tree-optimization/103278] [12 Regression] Recent change to cddce inhibits switch optimization

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103278

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|rguenth at gcc dot gnu.org         |unassigned at gcc dot gnu.org
             Status|RESOLVED                    |NEW
         Resolution|FIXED                       |---

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
So for

rl78-elf
  if-to-switch-5

we fail the JT because it's not enabled for the target.  w/o the CD-DCE change
we build a bit-test from

;; Canonical GIMPLE case clusters: 34 39 44 46 58 59 60 62 92

which we then expand into a new switch without any bit tests(?).
After the CD-DCE change we have improved(?) clusters:

;; Canonical GIMPLE case clusters: 34 39 44 46 58-60 62 92

and thus the output is no longer(?) beneficial.

I wonder if the merging of clusters done as part of the sorting is
counter-productive since it distorts the profitability?  If we check against
the
size of the unmerged clusters like with the following the testcase
will pass again.  I also notice we don't take advantage of the
merging and create

  switch (c_3(D)) <default: <L13> [INV], case 34: <L12> [INV], case 39: <L12>
[INV], case 44: <L12> [INV], case 46: <L12> [INV], case 58: <L12> [INV], case
59: <L12> [INV], case 60: <L12> [INV], case 62: <L12> [INV], case 92: <L12>
[INV]>

instead of one 'case 58-60:'

diff --git a/gcc/gimple-if-to-switch.cc b/gcc/gimple-if-to-switch.cc
index 16fabef7ca0..b64a1915e15 100644
--- a/gcc/gimple-if-to-switch.cc
+++ b/gcc/gimple-if-to-switch.cc
@@ -240,7 +240,7 @@ if_chain::is_beneficial ()

   vec<cluster *> output
     = jump_table_cluster::find_jump_tables (filtered_clusters);
-  bool r = output.length () < filtered_clusters.length ();
+  bool r = output.length () < clusters.length ();
   if (r)
     {
       dump_clusters (&output, "JT can be built");
@@ -251,7 +251,7 @@ if_chain::is_beneficial ()
     output.release ();

   output = bit_test_cluster::find_bit_tests (filtered_clusters);
-  r = output.length () < filtered_clusters.length ();
+  r = output.length () < clusters.length ();
   if (r)
     dump_clusters (&output, "BT can be built");


Martin, can you please look into this?