diff mbox

Pass manager: add support for termination of pass list

Message ID 56263B07.1010900@suse.cz
State New
Headers show

Commit Message

Martin Liška Oct. 20, 2015, 1 p.m. UTC 
Hello.

As part of upcoming merge of HSA branch, we would like to have possibility to terminate
pass manager after execution of the HSA generation pass. The HSA back-end is implemented
as a tree pass that directly emits HSAIL from gimple tree representation. The pass operates
on clones created by HSA IPA pass and the pass manager should stop execution of further
RTL passes.

Suggested patch survives bootstrap and regression tests on x86_64-linux-pc.

What do you think about it?
Thanks,
Martin

Comments

Richard Biener Oct. 20, 2015, 1:39 p.m. UTC | #1 
On Tue, Oct 20, 2015 at 3:00 PM, Martin Liška <mliska@suse.cz> wrote:
> Hello.
>
> As part of upcoming merge of HSA branch, we would like to have possibility to terminate
> pass manager after execution of the HSA generation pass. The HSA back-end is implemented
> as a tree pass that directly emits HSAIL from gimple tree representation. The pass operates
> on clones created by HSA IPA pass and the pass manager should stop execution of further
> RTL passes.
>
> Suggested patch survives bootstrap and regression tests on x86_64-linux-pc.
>
> What do you think about it?

Are you sure it works this way?

Btw, you will miss executing of all the cleanup passes that will
eventually free memory
associated with the function.  So I'd rather support a
TODO_discard_function which
should basically release the body from the cgraph.

Richard.

> Thanks,
> Martin
Martin Liška Oct. 21, 2015, 9:19 a.m. UTC | #2 
On 10/20/2015 03:39 PM, Richard Biener wrote:
> On Tue, Oct 20, 2015 at 3:00 PM, Martin Liška <mliska@suse.cz> wrote:
>> Hello.
>>
>> As part of upcoming merge of HSA branch, we would like to have possibility to terminate
>> pass manager after execution of the HSA generation pass. The HSA back-end is implemented
>> as a tree pass that directly emits HSAIL from gimple tree representation. The pass operates
>> on clones created by HSA IPA pass and the pass manager should stop execution of further
>> RTL passes.
>>
>> Suggested patch survives bootstrap and regression tests on x86_64-linux-pc.
>>
>> What do you think about it?
> 
> Are you sure it works this way?
> 
> Btw, you will miss executing of all the cleanup passes that will
> eventually free memory
> associated with the function.  So I'd rather support a
> TODO_discard_function which
> should basically release the body from the cgraph.

Hi.

Agree with you that I should execute all TODOs, which can be easily done.
However, if I just try to introduce the suggested TODO and handle it properly
by calling cgraph_node::release_body, then for instance fn->gimple_df, fn->cfg are
released and I hit ICEs on many places.

Stopping the pass manager looks necessary, or do I miss something?

Thanks,
Martin

> 
> Richard.
> 
>> Thanks,
>> Martin
Richard Biener Oct. 21, 2015, 9:59 a.m. UTC | #3 
On Wed, Oct 21, 2015 at 11:19 AM, Martin Liška <mliska@suse.cz> wrote:
> On 10/20/2015 03:39 PM, Richard Biener wrote:
>> On Tue, Oct 20, 2015 at 3:00 PM, Martin Liška <mliska@suse.cz> wrote:
>>> Hello.
>>>
>>> As part of upcoming merge of HSA branch, we would like to have possibility to terminate
>>> pass manager after execution of the HSA generation pass. The HSA back-end is implemented
>>> as a tree pass that directly emits HSAIL from gimple tree representation. The pass operates
>>> on clones created by HSA IPA pass and the pass manager should stop execution of further
>>> RTL passes.
>>>
>>> Suggested patch survives bootstrap and regression tests on x86_64-linux-pc.
>>>
>>> What do you think about it?
>>
>> Are you sure it works this way?
>>
>> Btw, you will miss executing of all the cleanup passes that will
>> eventually free memory
>> associated with the function.  So I'd rather support a
>> TODO_discard_function which
>> should basically release the body from the cgraph.
>
> Hi.
>
> Agree with you that I should execute all TODOs, which can be easily done.
> However, if I just try to introduce the suggested TODO and handle it properly
> by calling cgraph_node::release_body, then for instance fn->gimple_df, fn->cfg are
> released and I hit ICEs on many places.
>
> Stopping the pass manager looks necessary, or do I miss something?

"Stopping the pass manager" is necessary after TODO_discard_function, yes.
But that may be simply done via a has_body () check then?

> Thanks,
> Martin
>
>>
>> Richard.
>>
>>> Thanks,
>>> Martin
>
diff mbox

Patch

From fc60561c3ac09188fb61dac61b1cc2422061fc1d Mon Sep 17 00:00:00 2001
From: marxin <mliska@suse.cz>
Date: Mon, 19 Oct 2015 23:26:54 +0200
Subject: [PATCH] Add support for termination of pass manager.

gcc/ChangeLog:

2015-10-19  Martin Liska  <mliska@suse.cz>

	* passes.c (execute_one_pass): Add new argument called exit.
	(execute_pass_list_1): Terminate pass manager if a pass
	requests termination.
	(execute_ipa_pass_list): Likewise.
	* tree-pass.h: Introduce new TODO_stop_pass_execution.
---
 gcc/passes.c    | 24 ++++++++++++++++++++----
 gcc/tree-pass.h |  3 +++
 2 files changed, 23 insertions(+), 4 deletions(-)

diff --git a/gcc/passes.c b/gcc/passes.c
index 6ef6d2e..1199ae3 100644
--- a/gcc/passes.c
+++ b/gcc/passes.c
@@ -2279,10 +2279,11 @@  override_gate_status (opt_pass *pass, tree func, bool gate_status)
 }
 
 
-/* Execute PASS. */
+/* Execute PASS.  If the PASS requests to stop after its execution, EXIT
+   value is set to true.  */
 
 bool
-execute_one_pass (opt_pass *pass)
+execute_one_pass (opt_pass *pass, bool *exit)
 {
   unsigned int todo_after = 0;
 
@@ -2387,18 +2388,28 @@  execute_one_pass (opt_pass *pass)
   if (!((todo_after | pass->todo_flags_finish) & TODO_do_not_ggc_collect))
     ggc_collect ();
 
+  /* If finish TODO flags contain TODO_stop_pass_execution, set exit = true.  */
+  if (todo_after & TODO_stop_pass_execution)
+    *exit = true;
+
   return true;
 }
 
 static void
 execute_pass_list_1 (opt_pass *pass)
 {
+  bool stop_pass_execution = false;
+
   do
     {
       gcc_assert (pass->type == GIMPLE_PASS
 		  || pass->type == RTL_PASS);
-      if (execute_one_pass (pass) && pass->sub)
+      if (execute_one_pass (pass, &stop_pass_execution) && pass->sub)
         execute_pass_list_1 (pass->sub);
+
+      if (stop_pass_execution)
+	return;
+
       pass = pass->next;
     }
   while (pass);
@@ -2739,12 +2750,14 @@  ipa_read_optimization_summaries (void)
 void
 execute_ipa_pass_list (opt_pass *pass)
 {
+  bool stop_pass_execution = false;
+
   do
     {
       gcc_assert (!current_function_decl);
       gcc_assert (!cfun);
       gcc_assert (pass->type == SIMPLE_IPA_PASS || pass->type == IPA_PASS);
-      if (execute_one_pass (pass) && pass->sub)
+      if (execute_one_pass (pass, &stop_pass_execution) && pass->sub)
 	{
 	  if (pass->sub->type == GIMPLE_PASS)
 	    {
@@ -2763,6 +2776,9 @@  execute_ipa_pass_list (opt_pass *pass)
       gcc_assert (!current_function_decl);
       symtab->process_new_functions ();
       pass = pass->next;
+
+      if (stop_pass_execution)
+	return;
     }
   while (pass);
 }
diff --git a/gcc/tree-pass.h b/gcc/tree-pass.h
index c37e4b2..a481bac 100644
--- a/gcc/tree-pass.h
+++ b/gcc/tree-pass.h
@@ -300,6 +300,9 @@  protected:
 /* Rebuild the callgraph edges.  */
 #define TODO_rebuild_cgraph_edges       (1 << 22)
 
+/* Stop pass manager after execution of a pass.  */
+#define TODO_stop_pass_execution	(1 << 23)
+
 /* Internally used in execute_function_todo().  */
 #define TODO_update_ssa_any		\
     (TODO_update_ssa			\
-- 
2.6.0