[PATCH] Move jump threading before reload

[PATCH] Move jump threading before reload

[Ilya Leoshkevich]

Bootstrapped and regtested on x86_64-redhat-linux, s390x-redhat-linux
and ppc64le-redhat-linux.  The offending patch is in gcc-9_1_0-release
and gcc-9_2_0-release - do I need to backport this fix to gcc-9-branch?


r266734 has introduced a new instance of jump threading pass in order to
take advantage of opportunities that combine opens up.  It was perceived
back then that it was beneficial to delay it after reload, since that
might produce even more such opportunities.

Unfortunately jump threading interferes with hot/cold partitioning.  In
the code from PR92007, it converts the following

  +-------------------------- 2/HOT ------------------------+
  |                                                         |
  v                                                         v
3/HOT --> 5/HOT --> 8/HOT --> 11/COLD --> 6/HOT --EH--> 16/HOT
            |                               ^
            |                               |
            +-------------------------------+

into the following:

  +---------------------- 2/HOT ------------------+
  |                                               |
  v                                               v
3/HOT --> 8/HOT --> 11/COLD --> 6/COLD --EH--> 16/HOT

This makes hot bb 6 dominated by cold bb 11, and because of this
fixup_partitions makes bb 6 cold as well, which in turn makes EH edge
6->16 a crossing one.  Not only can't we have crossing EH edges, we are
also not allowed to introduce new crossing edges after reload in
general, since it might require extra registers on some targets.

Therefore, move the jump threading pass between combine and hot/cold
partitioning.  Building SPEC 2006 and SPEC 2017 with the old and the new
code indicates that:

* When doing jump threading right after reload, 3889 edges are threaded.
* When doing jump threading right after combine, 3918 edges are
  threaded.

This means this change will not introduce performance regressions.

gcc/ChangeLog:

2019-10-17  Ilya Leoshkevich  <iii@linux.ibm.com>

	PR rtl-optimization/92007
	* cfgcleanup.c (thread_jump): Add an assertion that we don't
	call it after reload if hot/cold partitioning has been done.
	(class pass_postreload_jump): Rename to
	pass_jump_after_combine.
	(make_pass_postreload_jump): Rename to
	make_pass_jump_after_combine.
	* passes.def(pass_postreload_jump): Move before reload, rename
	to pass_jump_after_combine.
	* tree-pass.h (make_pass_postreload_jump): Rename to
	make_pass_jump_after_combine.

gcc/testsuite/ChangeLog:

2019-10-17  Ilya Leoshkevich  <iii@linux.ibm.com>

	PR rtl-optimization/92007
	* g++.dg/opt/pr92007.C: New test (from Arseny Solokha).
---
 gcc/cfgcleanup.c                   | 22 +++++++++++---------
 gcc/passes.def                     |  2 +-
 gcc/testsuite/g++.dg/opt/pr92007.C | 32 ++++++++++++++++++++++++++++++
 gcc/tree-pass.h                    |  2 +-
 4 files changed, 47 insertions(+), 11 deletions(-)
 create mode 100644 gcc/testsuite/g++.dg/opt/pr92007.C

diff --git a/gcc/cfgcleanup.c b/gcc/cfgcleanup.c
index ced7e0a4283..835f7d79ea4 100644
--- a/gcc/cfgcleanup.c
+++ b/gcc/cfgcleanup.c
@@ -259,6 +259,10 @@ thread_jump (edge e, basic_block b)
   bool failed = false;
   reg_set_iterator rsi;
 
+  /* Jump threading may cause fixup_partitions to introduce new crossing edges,
+     which is not allowed after reload.  */
+  gcc_checking_assert (!reload_completed || !crtl->has_bb_partition);
+
   if (b->flags & BB_NONTHREADABLE_BLOCK)
     return NULL;
 
@@ -3280,10 +3284,10 @@ make_pass_jump (gcc::context *ctxt)
 \f
 namespace {
 
-const pass_data pass_data_postreload_jump =
+const pass_data pass_data_jump_after_combine =
 {
   RTL_PASS, /* type */
-  "postreload_jump", /* name */
+  "jump_after_combine", /* name */
   OPTGROUP_NONE, /* optinfo_flags */
   TV_JUMP, /* tv_id */
   0, /* properties_required */
@@ -3293,20 +3297,20 @@ const pass_data pass_data_postreload_jump =
   0, /* todo_flags_finish */
 };
 
-class pass_postreload_jump : public rtl_opt_pass
+class pass_jump_after_combine : public rtl_opt_pass
 {
 public:
-  pass_postreload_jump (gcc::context *ctxt)
-    : rtl_opt_pass (pass_data_postreload_jump, ctxt)
+  pass_jump_after_combine (gcc::context *ctxt)
+    : rtl_opt_pass (pass_data_jump_after_combine, ctxt)
   {}
 
   /* opt_pass methods: */
   virtual unsigned int execute (function *);
 
-}; // class pass_postreload_jump
+}; // class pass_jump_after_combine
 
 unsigned int
-pass_postreload_jump::execute (function *)
+pass_jump_after_combine::execute (function *)
 {
   cleanup_cfg (flag_thread_jumps ? CLEANUP_THREADING : 0);
   return 0;
@@ -3315,9 +3319,9 @@ pass_postreload_jump::execute (function *)
 } // anon namespace
 
 rtl_opt_pass *
-make_pass_postreload_jump (gcc::context *ctxt)
+make_pass_jump_after_combine (gcc::context *ctxt)
 {
-  return new pass_postreload_jump (ctxt);
+  return new pass_jump_after_combine (ctxt);
 }
 
 namespace {
diff --git a/gcc/passes.def b/gcc/passes.def
index 8999ceec636..798a391bd35 100644
--- a/gcc/passes.def
+++ b/gcc/passes.def
@@ -439,6 +439,7 @@ along with GCC; see the file COPYING3.  If not see
       NEXT_PASS (pass_ud_rtl_dce);
       NEXT_PASS (pass_combine);
       NEXT_PASS (pass_if_after_combine);
+      NEXT_PASS (pass_jump_after_combine);
       NEXT_PASS (pass_partition_blocks);
       NEXT_PASS (pass_outof_cfg_layout_mode);
       NEXT_PASS (pass_split_all_insns);
@@ -455,7 +456,6 @@ along with GCC; see the file COPYING3.  If not see
       NEXT_PASS (pass_reload);
       NEXT_PASS (pass_postreload);
       PUSH_INSERT_PASSES_WITHIN (pass_postreload)
-	  NEXT_PASS (pass_postreload_jump);
 	  NEXT_PASS (pass_postreload_cse);
 	  NEXT_PASS (pass_gcse2);
 	  NEXT_PASS (pass_split_after_reload);
diff --git a/gcc/testsuite/g++.dg/opt/pr92007.C b/gcc/testsuite/g++.dg/opt/pr92007.C
new file mode 100644
index 00000000000..9434cc929dd
--- /dev/null
+++ b/gcc/testsuite/g++.dg/opt/pr92007.C
@@ -0,0 +1,32 @@
+// PR rtl-optimization/92007
+// { dg-do compile }
+// { dg-options "-O2 -fno-tree-dominator-opts -fno-tree-forwprop --param max-cse-insns=0 -Wno-return-type -std=gnu++98 -freorder-blocks-and-partition" }
+
+void
+sb (int *);
+
+class d4 {
+public:
+  ~d4();
+  void gb ();
+  int op () { return no; }
+  int wl () { return tf; }
+  bool ee () try { gb (); } catch (...) { return false; }
+  bool b1 () { return (tf == no) ? false : ee (); }
+
+private:
+  int no, tf;
+};
+
+void
+hs (int *v9)
+{
+  d4 p6;
+
+  p6.gb ();
+  if (p6.op () > p6.wl ())
+    {
+      p6.b1 ();
+      sb (v9);
+    }
+}
diff --git a/gcc/tree-pass.h b/gcc/tree-pass.h
index 85b1c828f3a..a987661530e 100644
--- a/gcc/tree-pass.h
+++ b/gcc/tree-pass.h
@@ -564,6 +564,7 @@ extern rtl_opt_pass *make_pass_stack_ptr_mod (gcc::context *ctxt);
 extern rtl_opt_pass *make_pass_initialize_regs (gcc::context *ctxt);
 extern rtl_opt_pass *make_pass_combine (gcc::context *ctxt);
 extern rtl_opt_pass *make_pass_if_after_combine (gcc::context *ctxt);
+extern rtl_opt_pass *make_pass_jump_after_combine (gcc::context *ctxt);
 extern rtl_opt_pass *make_pass_ree (gcc::context *ctxt);
 extern rtl_opt_pass *make_pass_partition_blocks (gcc::context *ctxt);
 extern rtl_opt_pass *make_pass_match_asm_constraints (gcc::context *ctxt);
@@ -581,7 +582,6 @@ extern rtl_opt_pass *make_pass_clean_state (gcc::context *ctxt);
 extern rtl_opt_pass *make_pass_branch_prob (gcc::context *ctxt);
 extern rtl_opt_pass *make_pass_value_profile_transformations (gcc::context
 							      *ctxt);
-extern rtl_opt_pass *make_pass_postreload_jump (gcc::context *ctxt);
 extern rtl_opt_pass *make_pass_postreload_cse (gcc::context *ctxt);
 extern rtl_opt_pass *make_pass_gcse2 (gcc::context *ctxt);
 extern rtl_opt_pass *make_pass_split_after_reload (gcc::context *ctxt);
-- 
2.23.0

Re: [PATCH] Move jump threading before reload

[Segher Boessenkool]

On Fri, Oct 18, 2019 at 11:06:45AM +0200, Ilya Leoshkevich wrote:
> Bootstrapped and regtested on x86_64-redhat-linux, s390x-redhat-linux
> and ppc64le-redhat-linux.  The offending patch is in gcc-9_1_0-release
> and gcc-9_2_0-release - do I need to backport this fix to gcc-9-branch?

It is a regression on 9 (or so I assume), so yes please.

> 	PR rtl-optimization/92007
> 	* cfgcleanup.c (thread_jump): Add an assertion that we don't
> 	call it after reload if hot/cold partitioning has been done.
> 	(class pass_postreload_jump): Rename to
> 	pass_jump_after_combine.

This fits on one line just fine.

> 	(make_pass_postreload_jump): Rename to
> 	make_pass_jump_after_combine.
> 	* passes.def(pass_postreload_jump): Move before reload, rename

Space before (.

> --- a/gcc/cfgcleanup.c
> +++ b/gcc/cfgcleanup.c
> @@ -259,6 +259,10 @@ thread_jump (edge e, basic_block b)
>    bool failed = false;
>    reg_set_iterator rsi;
>  
> +  /* Jump threading may cause fixup_partitions to introduce new crossing edges,
> +     which is not allowed after reload.  */
> +  gcc_checking_assert (!reload_completed || !crtl->has_bb_partition);

Thanks for the assert, that will help prevent people from running into
this again.

The patch looks fine to me, but I'm not a global reviewer :-)


Segher

Re: [PATCH] Move jump threading before reload

[Jeff Law]

On 10/18/19 3:06 AM, Ilya Leoshkevich wrote:
> Bootstrapped and regtested on x86_64-redhat-linux, s390x-redhat-linux
> and ppc64le-redhat-linux.  The offending patch is in gcc-9_1_0-release
> and gcc-9_2_0-release - do I need to backport this fix to gcc-9-branch?
> 
> 
> r266734 has introduced a new instance of jump threading pass in order to
> take advantage of opportunities that combine opens up.  It was perceived
> back then that it was beneficial to delay it after reload, since that
> might produce even more such opportunities.
> 
> Unfortunately jump threading interferes with hot/cold partitioning.  In
> the code from PR92007, it converts the following
> 
>   +-------------------------- 2/HOT ------------------------+
>   |                                                         |
>   v                                                         v
> 3/HOT --> 5/HOT --> 8/HOT --> 11/COLD --> 6/HOT --EH--> 16/HOT
>             |                               ^
>             |                               |
>             +-------------------------------+
> 
> into the following:
> 
>   +---------------------- 2/HOT ------------------+
>   |                                               |
>   v                                               v
> 3/HOT --> 8/HOT --> 11/COLD --> 6/COLD --EH--> 16/HOT
> 
> This makes hot bb 6 dominated by cold bb 11, and because of this
> fixup_partitions makes bb 6 cold as well, which in turn makes EH edge
> 6->16 a crossing one.  Not only can't we have crossing EH edges, we are
> also not allowed to introduce new crossing edges after reload in
> general, since it might require extra registers on some targets.
> 
> Therefore, move the jump threading pass between combine and hot/cold
> partitioning.  Building SPEC 2006 and SPEC 2017 with the old and the new
> code indicates that:
> 
> * When doing jump threading right after reload, 3889 edges are threaded.
> * When doing jump threading right after combine, 3918 edges are
>   threaded.
> 
> This means this change will not introduce performance regressions.
> 
> gcc/ChangeLog:
> 
> 2019-10-17  Ilya Leoshkevich  <iii@linux.ibm.com>
> 
> 	PR rtl-optimization/92007
> 	* cfgcleanup.c (thread_jump): Add an assertion that we don't
> 	call it after reload if hot/cold partitioning has been done.
> 	(class pass_postreload_jump): Rename to
> 	pass_jump_after_combine.
> 	(make_pass_postreload_jump): Rename to
> 	make_pass_jump_after_combine.
> 	* passes.def(pass_postreload_jump): Move before reload, rename
> 	to pass_jump_after_combine.
> 	* tree-pass.h (make_pass_postreload_jump): Rename to
> 	make_pass_jump_after_combine.
> 
> gcc/testsuite/ChangeLog:
> 
> 2019-10-17  Ilya Leoshkevich  <iii@linux.ibm.com>
> 
> 	PR rtl-optimization/92007
> 	* g++.dg/opt/pr92007.C: New test (from Arseny Solokha).
OK for trunk and gcc-9 branch.

jeff