Fix hang in inline_small_functions (PR ipa/84425)

Message ID 20180216224944.GG5867@tucnak
State New
Headers show
Series
  • Fix hang in inline_small_functions (PR ipa/84425)
Related show

Commit Message

Jakub Jelinek Feb. 16, 2018, 10:49 p.m.
Hi!

When debugging a wrong-code in miniruby on aarch64, I've noticed a hang
in inline_small_functions when adding __attribute__((optimize(0)))
to one of the functions.

I don't know what exactly the code is doing, but it feels like a typo for
me, at least I'd expect the intent is for n2 to walk the whole chain,
rather than just the first 2 elements of it and then hope something changes
the ((struct ipa_dfs_info *) node->aux)->next_cycle to something different,
when the loop body is
                  if (opt_for_fn (n2->decl, optimize))
                    {
                      struct ipa_fn_summary *info2 = ipa_fn_summaries->get (n2);
                      if (info2->scc_no)
                        break;
                      info2->scc_no = id;
                    }
and so for some n2's isn't doing anything and at least the method name get
doesn't suggest it would update something.
While the regression on this testcase is recent, I think the bug is latent
since 2012 when this code has been introduced.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2018-02-16  Jakub Jelinek  <jakub@redhat.com>

	PR ipa/84425
	* ipa-inline.c (inline_small_functions): Fix a typo.

	* gcc.c-torture/compile/pr84425.c: New test.


	Jakub

Comments

Jeff Law Feb. 16, 2018, 11:03 p.m. | #1
On 02/16/2018 03:49 PM, Jakub Jelinek wrote:
> Hi!

> 

> When debugging a wrong-code in miniruby on aarch64, I've noticed a hang

> in inline_small_functions when adding __attribute__((optimize(0)))

> to one of the functions.

> 

> I don't know what exactly the code is doing, but it feels like a typo for

> me, at least I'd expect the intent is for n2 to walk the whole chain,

> rather than just the first 2 elements of it and then hope something changes

> the ((struct ipa_dfs_info *) node->aux)->next_cycle to something different,

> when the loop body is

>                   if (opt_for_fn (n2->decl, optimize))

>                     {

>                       struct ipa_fn_summary *info2 = ipa_fn_summaries->get (n2);

>                       if (info2->scc_no)

>                         break;

>                       info2->scc_no = id;

>                     }

> and so for some n2's isn't doing anything and at least the method name get

> doesn't suggest it would update something.

> While the regression on this testcase is recent, I think the bug is latent

> since 2012 when this code has been introduced.

> 

> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

> 

> 2018-02-16  Jakub Jelinek  <jakub@redhat.com>

> 

> 	PR ipa/84425

> 	* ipa-inline.c (inline_small_functions): Fix a typo.

> 

> 	* gcc.c-torture/compile/pr84425.c: New test.

Looks like a typo to me too.

OK.
jeff

Patch

--- gcc/ipa-inline.c.jj	2018-01-30 16:57:25.265095696 +0100
+++ gcc/ipa-inline.c	2018-02-16 20:53:54.803512238 +0100
@@ -1782,7 +1782,7 @@  inline_small_functions (void)
 		struct cgraph_node *n2;
 		int id = dfs->scc_no + 1;
 		for (n2 = node; n2;
-		     n2 = ((struct ipa_dfs_info *) node->aux)->next_cycle)
+		     n2 = ((struct ipa_dfs_info *) n2->aux)->next_cycle)
 		  if (opt_for_fn (n2->decl, optimize))
 		    {
 		      struct ipa_fn_summary *info2 = ipa_fn_summaries->get (n2);
--- gcc/testsuite/gcc.c-torture/compile/pr84425.c.jj	2018-02-16 20:54:46.785534252 +0100
+++ gcc/testsuite/gcc.c-torture/compile/pr84425.c	2018-02-16 20:55:00.628540119 +0100
@@ -0,0 +1,17 @@ 
+/* PR ipa/84425 */
+
+void bar (int);
+
+void
+foo (int x)
+{
+  if (x < 5)
+    bar (x);
+}
+
+__attribute__((optimize(0))) void
+bar (int x)
+{
+  if (x > 10)
+    foo (x);
+}