diff mbox series

tree-optimization/109237 - last_stmt is possibly slow

Message ID 20230322123021.35D7E385735E@sourceware.org
State New
Headers
Series tree-optimization/109237 - last_stmt is possibly slow |

Commit Message

Richard Biener March 22, 2023, 12:29 p.m. UTC 
  Most uses of last_stmt are interested in control transfer stmts
and for the testcase gimple_purge_dead_eh_edges shows up in
the profile.  But last_stmt looks past trailing debug stmts but
those would be rejected by GIMPLEs verify_flow_info.  The following
adds possible_ctrl_stmt besides last_stmt which does not look
past trailing debug stmts and adjusts gimple_purge_dead_eh_edges.

I've put checking code into possible_ctrl_stmt that it will not
miss a control statement if the real last stmt is a debug stmt.

The alternative would be to change last_stmt, explicitely introducing
last_nondebug_stmt.  I remember we chickened out and made last_stmt
conservative here but not anticipating the compile-time issues this
creates.  I count 227 last_stmt and 12 last_and_only_stmt uses.

Bootstrapped and tested on x86_64-unknown-linux-gnu.

Any opinions?  I probably lean towards s/last_stmt/last_nondebug_stmt/
in next stage1 and then adding last_stmt and changing some
uses back - through for maintainance that's going to be a
nightmare (or maybe not, a "wrong" last_stmt should be safe to
backport and a last_nondebug_stmt will fail to build).

Richard.

	PR tree-optimization/109237
	* tree-cfg.h (possible_ctrl_stmt): New function returning
	the last stmt not skipping debug stmts.
	(gimple_purge_dead_eh_edges): Use it.
---
 gcc/tree-cfg.cc | 22 +++++++++++++++++++++-
 gcc/tree-cfg.h  |  1 +
 2 files changed, 22 insertions(+), 1 deletion(-)
diff mbox series

Patch

diff --git a/gcc/tree-cfg.cc b/gcc/tree-cfg.cc
index a9fcc7fd050..e167596209b 100644
--- a/gcc/tree-cfg.cc
+++ b/gcc/tree-cfg.cc
@@ -2947,6 +2947,26 @@  first_non_label_stmt (basic_block bb)
   return !gsi_end_p (i) ? gsi_stmt (i) : NULL;
 }
 
+/* Return the last statement of a basic block BB that's possibly control
+   altering.  Compared to last_stmt this will return a debug stmt if that
+   is the last stmt.  */
+
+gimple *
+possible_ctrl_stmt (basic_block bb)
+{
+  gimple_stmt_iterator i = gsi_last_bb (bb);
+  if (gsi_end_p (i))
+    return NULL;
+  if (flag_checking && is_gimple_debug (gsi_stmt (i)))
+    {
+      /* Verify that if the real last stmt is a debug stmt the
+	 last non-debug stmt isn't control altering.  */
+      gimple *last = last_stmt (bb);
+      gcc_assert (!last || !stmt_ends_bb_p (last));
+    }
+  return gsi_stmt (i);
+}
+
 /* Return the last statement in basic block BB.  */
 
 gimple *
@@ -8990,7 +9010,7 @@  gimple_purge_dead_eh_edges (basic_block bb)
   bool changed = false;
   edge e;
   edge_iterator ei;
-  gimple *stmt = last_stmt (bb);
+  gimple *stmt = possible_ctrl_stmt (bb);
 
   if (stmt && stmt_can_throw_internal (cfun, stmt))
     return false;
diff --git a/gcc/tree-cfg.h b/gcc/tree-cfg.h
index 9b56a68fe9d..7c6a9a4f16b 100644
--- a/gcc/tree-cfg.h
+++ b/gcc/tree-cfg.h
@@ -61,6 +61,7 @@  extern bool assert_unreachable_fallthru_edge_p (edge);
 extern void delete_tree_cfg_annotations (function *);
 extern gphi *get_virtual_phi (basic_block);
 extern gimple *first_stmt (basic_block);
+extern gimple *possible_ctrl_stmt (basic_block);
 extern gimple *last_stmt (basic_block);
 extern gimple *last_and_only_stmt (basic_block);
 extern bool verify_gimple_in_seq (gimple_seq, bool = true);