This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[committed] Fix segfault in free_growth_caches (PR jit/80954)


r248336 added these lines to free_growth_caches

  if (edge_removal_hook_holder)
    symtab->remove_edge_removal_hook (edge_removal_hook_holder);

which broke the JIT; attempts to compile more than one time
within a single process segfault here:

  305	symbol_table::remove_edge_removal_hook (cgraph_edge_hook_list *entry)
  306	{
  307	  cgraph_edge_hook_list **ptr = &m_first_edge_removal_hook;
  308
  309	  while (*ptr != entry)
>>310	    ptr = &(*ptr)->next;
  311	  *ptr = entry->next;
  312	  free (entry);
  313	}

  (gdb) p ptr
  $3 = (cgraph_edge_hook_list **) 0x10
  (gdb) bt
  #0  symbol_table::remove_edge_removal_hook (this=0x7fffeef83200, entry=0x664c90) at ../../src/gcc/cgraph.c:310
  #1  0x00007ffff6a03e2c in free_growth_caches () at ../../src/gcc/ipa-inline-analysis.c:113
  #2  0x00007ffff73e7bd9 in inline_small_functions () at ../../src/gcc/ipa-inline.c:2059
  #3  ipa_inline () at ../../src/gcc/ipa-inline.c:2417
  [...]

The root cause is that initialize_growth_caches lazily adds
edge_removal_hook_holder to the symtab:

  99	  if (!edge_removal_hook_holder)
  100	    edge_removal_hook_holder =
  101	      symtab->add_edge_removal_hook (&inline_edge_removal_hook, NULL);

but free_growth_caches removes it without NULL-ing it:

  112	  if (edge_removal_hook_holder)
  113	    symtab->remove_edge_removal_hook (edge_removal_hook_holder);

Hence on the second call to free_growth_caches, it attempts to remove
the edge_removal_hook_holder from the 1st iteration's
initialize_growth_caches, which isn't present in the 2nd iteration's
symtab.

Hence the edge_removal_hook_holder isn't present, and
symtab::remove_edge_removal_hook, relying on it as a sentinel value,
reads through ((cgraph_edge_hook_list *)NULL)->next.

This patch fixes the segfault by resetting it to NULL when removing it.

Successfully bootstrapped&regrtested on x86_64-pc-linux-gnu.
Restores jit.sum from:

  # of expected passes		2892
  # of unexpected failures	61
  # of unresolved testcases	1

to:

  # of expected passes		3202

Committed to trunk as r248841, under the "obvious" rule.

gcc/ChangeLog:
	PR jit/80954
	* ipa-inline-analysis.c (free_growth_caches): Set
	edge_removal_hook_holder to NULL after removing it.
---
 gcc/ipa-inline-analysis.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/gcc/ipa-inline-analysis.c b/gcc/ipa-inline-analysis.c
index f562ca5..9f7b2a1 100644
--- a/gcc/ipa-inline-analysis.c
+++ b/gcc/ipa-inline-analysis.c
@@ -110,7 +110,10 @@ void
 free_growth_caches (void)
 {
   if (edge_removal_hook_holder)
-    symtab->remove_edge_removal_hook (edge_removal_hook_holder);
+    {
+      symtab->remove_edge_removal_hook (edge_removal_hook_holder);
+      edge_removal_hook_holder = NULL;
+    }
   edge_growth_cache.release ();
 }
 
-- 
1.8.5.3


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]