[Bug tree-optimization/49772] New: [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[Bug tree-optimization/49772] New: [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[zsojka at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772

           Summary: [4.7 Regression] ICE: in ipa_pta_execute, at
                    tree-ssa-structalias.c:6790 with -fipa-pta
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: zsojka@seznam.cz
                CC: rguenth@gcc.gnu.org
              Host: x86_64-pc-linux-gnu
            Target: x86_64-pc-linux-gnu


Created attachment 24784
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24784
reduced testcase (from g++.dg/tree-ssa/pr42337.C)

Compiler output:
$ gcc -O -fipa-pta testcase.C 
testcase.C: In member function 'int S< <template-parameter-1-1> >::foo() [with
<template-parameter-1-1> = int]':
testcase.C:4:39: error: inlining failed in call to always_inline 'int S<
<template-parameter-1-1> >::foo() [with <template-parameter-1-1> = int]':
recursive inlining
testcase.C:7:7: error: called from here
testcase.C: At global scope:
testcase.C:16:1: internal compiler error: in ipa_pta_execute, at
tree-ssa-structalias.c:6790
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.

Compiler output - 4.6:
$ gcc -O -fipa-pta testcase.C 
testcase.C: In function 'void foo()':
testcase.C:4:39: sorry, unimplemented: inlining failed in call to 'int S<
<template-parameter-1-1> >::foo() [with <template-parameter-1-1> = int]':
recursive inlining
testcase.C:7:7: sorry, unimplemented: called from here

Tested revisions:
r176370 - crash
4.6 r176280 - sorry

[Bug tree-optimization/49772] [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[zsojka at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772

Zdenek Sojka <zsojka at seznam dot cz> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #24784|0                           |1
        is obsolete|                            |

--- Comment #1 from Zdenek Sojka <zsojka at seznam dot cz> 2011-07-17 21:06:08 UTC ---
Created attachment 24785
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24785
reduced testcase (from g++.dg/tree-ssa/pr42337.C)

[Bug tree-optimization/49772] [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2011.07.18 08:27:57
                 CC|                            |hubicka at gcc dot gnu.org
   Target Milestone|---                         |4.7.0
     Ever Confirmed|0                           |1

--- Comment #2 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-07-18 08:27:57 UTC ---
Honza, now that we error instead of sorry (which immediately aborts
compilation),
should we fixup the cgraph to point to non-inline variants?

[Bug tree-optimization/49772] [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[hubicka at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772

--- Comment #3 from Jan Hubicka <hubicka at gcc dot gnu.org> 2011-08-03 19:57:06 UTC ---
Hmm, we should never make the cgrpah point to inline variant in this case, so
rest of compilation should go smoothly after the error is output. I don't think
it is what is confusing ipa-pta, especially now when it is run after inlining?

Honza

[Bug tree-optimization/49772] [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[rguenther at suse dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772

--- Comment #4 from rguenther at suse dot de <rguenther at suse dot de> 2011-08-04 09:20:21 UTC ---
On Wed, 3 Aug 2011, hubicka at gcc dot gnu.org wrote:

> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772
> 
> --- Comment #3 from Jan Hubicka <hubicka at gcc dot gnu.org> 2011-08-03 19:57:06 UTC ---
> Hmm, we should never make the cgrpah point to inline variant in this case, so
> rest of compilation should go smoothly after the error is output. I don't think
> it is what is confusing ipa-pta, especially now when it is run after inlining?

What happens is that in expand_call_inline we fail the inlining via
an error, but at this point of course the edge points to the inline
clone and that stays so.

I suppose when expand_call_inline fails we have to re-direct the
edge to the original function (huh, eventually make sure it is not
optimized away as well).

Or, finally, _never_ fail that late

Richard.

[Bug tree-optimization/49772] [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[hubicka at ucw dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772

--- Comment #5 from Jan Hubicka <hubicka at ucw dot cz> 2011-08-04 18:32:27 UTC ---
> What happens is that in expand_call_inline we fail the inlining via
> an error, but at this point of course the edge points to the inline
> clone and that stays so.
> 
> I suppose when expand_call_inline fails we have to re-direct the
> edge to the original function (huh, eventually make sure it is not
> optimized away as well).

Well, the problem is that the original function don't have to exist at this
point,
just an inline clone with posisbly constants propagated into it and thus some
edges removed.  Undoing this is nontrivial/impossible.
> 
> Or, finally, _never_ fail that late

Yep, we need to avoid failing late.
Honza

[Bug tree-optimization/49772] [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P1
             Status|NEW                         |ASSIGNED
         AssignedTo|unassigned at gcc dot       |rguenth at gcc dot gnu.org
                   |gnu.org                     |

--- Comment #6 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-10-27 09:44:40 UTC ---
Hmmmm, I'll take this one.

[Bug tree-optimization/49772] [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED

--- Comment #7 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-12-06 10:37:53 UTC ---
We no longer ICE.

[Bug tree-optimization/49772] [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[zsojka at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772

Zdenek Sojka <zsojka at seznam dot cz> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |REOPENED
         Resolution|FIXED                       |

--- Comment #8 from Zdenek Sojka <zsojka at seznam dot cz> 2011-12-07 05:35:48 UTC ---
The original testcase still fails:
$ gcc -O -fipa-pta g++.dg/tree-ssa/pr42337.C  
g++.dg/tree-ssa/pr42337.C: In member function 'VALUE SFVT<KEY, VALUE>::IS(const
SFVT<KEY, VALUE>&) const [with KEY = long long unsigned int; VALUE = double;
SFVT<KEY, VALUE> = SFVT<long long unsigned int, double>]':
g++.dg/tree-ssa/pr42337.C:117:40: error: inlining failed in call to
always_inline 'VALUE SFVT<KEY, VALUE>::IS(const SFVT<KEY, VALUE>&) const [with
KEY = long long unsigned int; VALUE = double; SFVT<KEY, VALUE> = SFVT<long long
unsigned int, double>]': recursive inlining
g++.dg/tree-ssa/pr42337.C:119:26: error: called from here
g++.dg/tree-ssa/pr42337.C: At global scope:
g++.dg/tree-ssa/pr42337.C:173:1: internal compiler error: in ipa_pta_execute,
at tree-ssa-structalias.c:6863
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.

[Bug tree-optimization/49772] [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[zsojka at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772

--- Comment #9 from Zdenek Sojka <zsojka at seznam dot cz> 2011-12-07 05:37:04 UTC ---
Created attachment 26014
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26014
autoreduced testcase

$ gcc -O -fipa-pta testcase-min14.ii 
testcase-min14.ii: In member function 'VALUE SFVT< <template-parameter-1-1>,
VALUE>::IS(SFVT< <template-parameter-1-1>, VALUE>) [with
<template-parameter-1-1> = unsigned int; VALUE = double; SFVT<
<template-parameter-1-1>, VALUE> = SFVT<unsigned int, double>]':
testcase-min14.ii:17:41: error: inlining failed in call to always_inline 'VALUE
SFVT< <template-parameter-1-1>, VALUE>::IS(SFVT< <template-parameter-1-1>,
VALUE>) [with <template-parameter-1-1> = unsigned int; VALUE = double; SFVT<
<template-parameter-1-1>, VALUE> = SFVT<unsigned int, double>]': recursive
inlining
testcase-min14.ii:21:2: error: called from here
testcase-min14.ii: At global scope:
testcase-min14.ii:41:1: internal compiler error: in ipa_pta_execute, at
tree-ssa-structalias.c:6863
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.

[Bug tree-optimization/49772] [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772

--- Comment #10 from Andrew Pinski <pinskia at gcc dot gnu.org> 2011-12-08 04:54:41 UTC ---
A better non specific to tree-ssa-structaliasing patch:
Index: cgraphunit.c
===================================================================
--- cgraphunit.c        (revision 182098)
+++ cgraphunit.c        (working copy)
@@ -1817,6 +1817,9 @@ cgraph_expand_function (struct cgraph_no
   /* Generate RTL for the body of DECL.  */
   tree_rest_of_compilation (decl);

+  if (seen_error ())
+    return;
+
   /* Make sure that BE didn't give up on compiling.  */
   gcc_assert (TREE_ASM_WRITTEN (decl));
   current_function_decl = NULL;
Index: passes.c
===================================================================
--- passes.c    (revision 182098)
+++ passes.c    (working copy)
@@ -2074,7 +2074,7 @@ execute_one_pass (struct opt_pass *pass)
     timevar_push (pass->tv_id);

   /* Do it!  */
-  if (pass->execute)
+  if (pass->execute && !seen_error ())
     {
       todo_after = pass->execute ();
       do_per_function (clear_last_verified, NULL);

[Bug tree-optimization/49772] [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[rguenther at suse dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772

--- Comment #11 from rguenther at suse dot de <rguenther at suse dot de> 2011-12-08 08:57:22 UTC ---
On Thu, 8 Dec 2011, pinskia at gcc dot gnu.org wrote:

> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772
> 
> --- Comment #10 from Andrew Pinski <pinskia at gcc dot gnu.org> 2011-12-08 04:54:41 UTC ---
> A better non specific to tree-ssa-structaliasing patch:
> Index: cgraphunit.c
> ===================================================================
> --- cgraphunit.c        (revision 182098)
> +++ cgraphunit.c        (working copy)
> @@ -1817,6 +1817,9 @@ cgraph_expand_function (struct cgraph_no
>    /* Generate RTL for the body of DECL.  */
>    tree_rest_of_compilation (decl);
> 
> +  if (seen_error ())
> +    return;
> +
>    /* Make sure that BE didn't give up on compiling.  */
>    gcc_assert (TREE_ASM_WRITTEN (decl));
>    current_function_decl = NULL;
> Index: passes.c
> ===================================================================
> --- passes.c    (revision 182098)
> +++ passes.c    (working copy)
> @@ -2074,7 +2074,7 @@ execute_one_pass (struct opt_pass *pass)
>      timevar_push (pass->tv_id);
> 
>    /* Do it!  */
> -  if (pass->execute)
> +  if (pass->execute && !seen_error ())
>      {
>        todo_after = pass->execute ();
>        do_per_function (clear_last_verified, NULL);

We don't want late warnings in random functions to be suppressed
by earlier errors.

[Bug tree-optimization/49772] [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772

--- Comment #12 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-12-08 11:39:21 UTC ---
It's a complete mystery to me how this cgraph node cloning and edge code
works.  We create the inline clone for a completely different edge
(that _is_ decided to be inlined) than the one we later fail on
(that one is _not_ marked inline and does _not_ have an inline clone callee).

Thus, we shouldn't arrive at this situation at all.  Instead we seem to fail
to inline a call that we should inline (and decided to inline) because of:

/* Expand calls to inline functions in the body of FN.  */

unsigned int
optimize_inline_calls (tree fn)
{
...
  /* There is no point in performing inlining if errors have already
     occurred -- and we might crash if we try to inline invalid
     code.  */
  if (seen_error ())
    return 0;

which leaves us with such state.

So this is yet another case of overeager fail to complete our task.
I suppose we can remove that late fail now that we bail out way earlier
at proper IPA "scope".

The ICE is fixed with removing the above out.

[Bug tree-optimization/49772] [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772

--- Comment #13 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-12-08 13:23:59 UTC ---
Author: rguenth
Date: Thu Dec  8 13:23:54 2011
New Revision: 182110

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=182110
Log:
2011-12-08  Richard Guenther  <rguenther@suse.de>

    PR tree-optimization/49772
    * tree-inline.c (optimize_inline_calls): Remove bail out
    on errors.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/tree-inline.c

[Bug tree-optimization/49772] [4.7 Regression] ICE: in ipa_pta_execute, at tree-ssa-structalias.c:6790 with -fipa-pta

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49772

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|REOPENED                    |RESOLVED
         Resolution|                            |FIXED

--- Comment #14 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-12-08 13:25:50 UTC ---
Fixed.