[patch] Fix ICE during LTRANS on big Ada application

[patch] Fix ICE during LTRANS on big Ada application

[Eric Botcazou]

[-- Attachment #1: Type: text/plain, Size: 2517 bytes --]

Hi,

we have run into an ICE in LTO mode on a big Ada application with our 4.9-
based compiler, whose error message is:

xxxxxxxxxxxxxx.constprop.15627/1865799 (xxxxxxxxxxxxxxxxxxxxxxx.constprop) 
@0x7ffff6835bd0
  Type: function definition analyzed
  Visibility: used_from_other_partition public visibility_specified 
visibility:hidden artificial
  References: __gnat_personality_v0/19663 (addr)__gnat_others_value/19664 
(addr)
  Referring: 
  Read from file: xxxxxx.ltrans31.o
  Availability: local
  First run: 0
  Function flags: local
  Called by: 
yyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyy.constprop.15352/1865524 
zzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzz.constprop.15343/1865515 
  Calls: 
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx.constprop.15628/1865800 
(1.00 per call) 
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx.part.62.lto_priv.21947/1557884 
(can throw external) __gnat_end_handler/19660 (can throw external) 
__builtin_eh_pointer/20253 __gnat_begin_handler/19662 __gnat_end_handler/19660 
(can throw external) 
lto1: fatal error: xxxxxx.ltrans31.o: section 
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx.constprop.15627 is missing

So LTRANS complains that a .constprop function is missing on disk, which 
doesn't really make sense since it's a clone so it is not streamed.

The problem is that a clone function has it clone_of field zeroed during clone 
materialization by cgraph_node::remove:

	  /* We are removing node with clones.  This makes clones inconsistent,
	     but assume they will be removed subsequently and just keep clone
	     tree intact.  This can happen in unreachable function removal since
	     we remove unreachable functions in random order, not by bottom-up
	     walk of clone trees.  */
	  for (n = clones; n; n = next)
	    {
	       next = n->next_sibling_clone;
	       n->next_sibling_clone = NULL;
	       n->prev_sibling_clone = NULL;
	       n->clone_of = NULL;
	    }

because it's the clone of an inlined_to node that gets discarded because it's 
inlined into another clone(!).  So the first clone function cannot (and is 
never) materialized and the compiler ICEs later when its body is requested.

It seems to be that the root cause is clone_inlined_nodes overriding a master 
clone that has non-inline clones, hence the attached patch.

Tested on x86_64-suse-linux, OK for the mainline?


2014-10-06  Eric Botcazou  <ebotcazou@adacore.com>

	* ipa-inline-transform.c (can_remove_node_now_p_1): Return false for a
	master clone that has non-inline clones.


-- 
Eric Botcazou

[-- Attachment #2: p.diff --]
[-- Type: text/x-patch, Size: 789 bytes --]

Index: ipa-inline-transform.c
===================================================================
--- ipa-inline-transform.c	(revision 215843)
+++ ipa-inline-transform.c	(working copy)
@@ -82,6 +82,13 @@ update_noncloned_frequencies (struct cgr
 static bool
 can_remove_node_now_p_1 (struct cgraph_node *node)
 {
+  /* We cannot remove a master clone that has non-inline clones until after
+     these clones are materialized.  */
+  if (!node->clone_of)
+    for (struct cgraph_node *n = node->clones; n; n = n->next_sibling_clone)
+      if (n->decl != node->decl)
+	return false;
+
   /* FIXME: When address is taken of DECL_EXTERNAL function we still
      can remove its offline copy, but we would need to keep unanalyzed node in
      the callgraph so references can point to it.  */

Re: [patch] Fix ICE during LTRANS on big Ada application

[Eric Botcazou]

Jan, any opinion on the following patch?

> 2014-10-06  Eric Botcazou  <ebotcazou@adacore.com>
> 
> 	* ipa-inline-transform.c (can_remove_node_now_p_1): Return false for a
> 	master clone that has non-inline clones.

https://gcc.gnu.org/ml/gcc-patches/2014-10/msg00436.html

Thanks in advance.

-- 
Eric Botcazou

Re: [patch] Fix ICE during LTRANS on big Ada application

[Jan Hubicka]

> Jan, any opinion on the following patch?
> 
> > 2014-10-06  Eric Botcazou  <ebotcazou@adacore.com>
> > 
> > 	* ipa-inline-transform.c (can_remove_node_now_p_1): Return false for a
> > 	master clone that has non-inline clones.
> 
> https://gcc.gnu.org/ml/gcc-patches/2014-10/msg00436.html
> 
> Thanks in advance.

Yes, this looks OK. Inliner coul also take care to turn the master clone into unanalyzed
node as remove_unreachable_nodes would, but I do not think it is worth the effort.
Please put the loop later in the function so it does not slow things down unnecesarily
(there are often many inline clones)

Honza
> 
> -- 
> Eric Botcazou

Re: [patch] Fix ICE during LTRANS on big Ada application

[Eric Botcazou]

> Yes, this looks OK. Inliner coul also take care to turn the master clone
> into unanalyzed node as remove_unreachable_nodes would, but I do not think
> it is worth the effort. Please put the loop later in the function so it
> does not slow things down unnecesarily (there are often many inline clones)

If you're concerned about efficiency, I can move it into clone_inlined_nodes 
itself instead:

  if (duplicate)
    {
      /* We may eliminate the need for out-of-line copy to be output.
	 In that case just go ahead and re-use it.  This is not just an
	 memory optimization.  Making offline copy of fuction disappear
	 from the program will improve future decisions on inlining.  */
      if (!e->callee->callers->next_caller
	  /* Recursive inlining never wants the master clone to
	     be overwritten.  */
	  && update_original
	  && can_remove_node_now_p (e->callee, e))

i.e. I can set update_original to false just before the above test, that would
probably be sufficient to fix the bug.

-- 
Eric Botcazou

Re: [patch] Fix ICE during LTRANS on big Ada application

[Eric Botcazou]

[-- Attachment #1: Type: text/plain, Size: 641 bytes --]

> Yes, this looks OK. Inliner coul also take care to turn the master clone
> into unanalyzed node as remove_unreachable_nodes would, but I do not think
> it is worth the effort. Please put the loop later in the function so it
> does not slow things down unnecesarily (there are often many inline clones)

Here's what I have installed on the mainline after re-testing, this defers the 
check as much as possible.


2014-10-17  Eric Botcazou  <ebotcazou@adacore.com>

	* ipa-inline-transform.c (master_clone_with_noninline_clones_p): New.
	(clone_inlined_nodes): Do not overwrite the clone if above predicate
	returns true.


-- 
Eric Botcazou

[-- Attachment #2: p.diff --]
[-- Type: text/x-patch, Size: 1323 bytes --]

Index: ipa-inline-transform.c
===================================================================
--- ipa-inline-transform.c	(revision 216252)
+++ ipa-inline-transform.c	(working copy)
@@ -122,6 +122,20 @@ can_remove_node_now_p (struct cgraph_nod
   return true;
 }
 
+/* Return true if NODE is a master clone with non-inline clones.  */
+
+static bool
+master_clone_with_noninline_clones_p (struct cgraph_node *node)
+{
+  if (node->clone_of)
+    return false;
+
+  for (struct cgraph_node *n = node->clones; n; n = n->next_sibling_clone)
+    if (n->decl != node->decl)
+      return true;
+
+  return false;
+}
 
 /* E is expected to be an edge being inlined.  Clone destination node of
    the edge and redirect it to the new clone.
@@ -155,7 +169,10 @@ clone_inlined_nodes (struct cgraph_edge
 	  /* Recursive inlining never wants the master clone to
 	     be overwritten.  */
 	  && update_original
-	  && can_remove_node_now_p (e->callee, e))
+	  && can_remove_node_now_p (e->callee, e)
+	  /* We cannot overwrite a master clone with non-inline clones
+	     until after these clones are materialized.  */
+	  && !master_clone_with_noninline_clones_p (e->callee))
 	{
 	  /* TODO: When callee is in a comdat group, we could remove all of it,
 	     including all inline clones inlined into it.  That would however