final_cleanup: Call make_forwarders_with_degenerate_phis [PR46555]

This fixes a regression introduced with r16-5258-g1d8e2d51e5c5cb.
With GCC 12+, we would not merge forwarders (with phis, vops included),
this meant that after the last cddce, degenerate phis would stay not
merged which allowed for better expansion. Now after my patch, the forwarder
block would be removed and get worse expansion. This fixes the problem
by creating the forwarder blocks in "optimized" and no other cleanupcfg
is called afterwards.

Oh this also fixes the problem at -O1 which was missed because the agressive
version of dce was not done at -O1.

Bootstrapped and tested on x86_64-linux-gnu.

	PR tree-optimization/46555
gcc/ChangeLog:

	* tree-cfgcleanup.cc (execute_cleanup_cfg_post_optimizing):
	Don't set todo to include cleanupcfg; do it manually.
	Call make_forwarders_with_degenerate_phis if optimizing.

gcc/testsuite/ChangeLog:

	* gcc.dg/tree-ssa/pr46555.c: New test.

Signed-off-by: Andrew Pinski <andrew.pinski@oss.qualcomm.com>

diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr46555.c b/gcc/testsuite/gcc.dg/tree-ssa/pr46555.c
new file mode 100644 (file)
index 0000000..d4de7c2
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/pr46555.c
@@ -0,0 +1,28 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-tree-optimized-details -fdump-rtl-pro_and_epilogue" } */
+/* PR tree-optimization/46555 */
+/* Here should not remove the forwarder block (or rather recreate it and not
+   remove it again). This improves expansion to RTL as there is one less copy
+   (or constant formation) in some cases. In this case we also get the ability
+   to shrink wrap the function.  */
+
+int h(void);
+int f(int a, int b, int c)
+{
+  if (a)
+    return 2;
+  h();
+  if (b)
+    return 2;
+  h();
+  if (c)
+    return 2;
+  h();
+  return 4;
+}
+
+/* { dg-final { scan-tree-dump-times "New forwarder block for edge" 1 "optimized" } } */
+/* Make sure we only have a PHI with 2 arguments here, 2 and 4.  */
+/* { dg-final { scan-tree-dump "PHI <2..., 4...>|PHI <4..., 2...>" "optimized" } } */
+/* Make sure we can shrink wrap the function now too. */
+/* { dg-final { scan-rtl-dump "Performing shrink-wrapping" "pro_and_epilogue" { target { { { i?86-*-* x86_64-*-* } && { ! ia32 } } || { powerpc*-*-* aarch64*-*-* riscv*-*-* arm*-*-* }  } } } } */

diff --git a/gcc/tree-cfgcleanup.cc b/gcc/tree-cfgcleanup.cc
index 872ded3d15ecbc70f19ea115466782972e2f8d41..093fde93da6d436a0db459fb85a4a61b30305d32 100644 (file)
--- a/gcc/tree-cfgcleanup.cc
+++ b/gcc/tree-cfgcleanup.cc
@@ -1415,8 +1415,13 @@ execute_cleanup_cfg_post_optimizing (void)
     }
   maybe_remove_unreachable_handlers ();
   cleanup_dead_labels ();
-  if (group_case_labels ())
-    todo |= TODO_cleanup_cfg;
+  if (group_case_labels () && cleanup_tree_cfg ())
+    todo |= TODO_update_ssa;
+
+  /* When optimizing undo the merging of forwarder blocks
+     that phis for better out of ssa expansion.  */
+  if (optimize)
+    make_forwarders_with_degenerate_phis (cfun);
 
   basic_block bb = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun));
   gimple_stmt_iterator gsi = gsi_start_nondebug_after_labels_bb (bb);