ipa-cp: Fix assert triggering with -fno-toplevel-reorder (PR 106260)

Message ID ri6o7xsena9.fsf@suse.cz
State New, archived
Headers
Series ipa-cp: Fix assert triggering with -fno-toplevel-reorder (PR 106260) |

Commit Message

Martin Jambor July 13, 2022, 9:05 p.m. UTC
  Hi,

with -fno-toplevel-reorder (and -fwhole-program), there apparently can
be local functions without any callers.  This is something that IPA-CP
does not like because its propagation verifier checks that local
functions do not end up with TOP in their lattices.  Therefore there
is an assert checking that all call-less unreachable functions have
been removed, which triggers in PR 106260 with these two options.

This patch detects the situation and marks the lattices as variable,
thus avoiding both the assert trigger and the verification failure.

Bootstrapped and tested on x86_64-linux.  OK for master and then all
active release branches?

Thanks,

Martin


gcc/ChangeLog:

2022-07-13  Martin Jambor  <mjambor@suse.cz>

	PR ipa/106260
	* ipa-cp.cc (initialize_node_lattices): Replace assert that there are
	callers with handling that situation when -fno-toplevel_reorder.

gcc/testsuite/ChangeLog:

2022-07-13  Martin Jambor  <mjambor@suse.cz>

	PR ipa/106260
	* g++.dg/ipa/pr106260.C: New test.
---
 gcc/ipa-cp.cc                       |  6 ++-
 gcc/testsuite/g++.dg/ipa/pr106260.C | 64 +++++++++++++++++++++++++++++
 2 files changed, 69 insertions(+), 1 deletion(-)
 create mode 100644 gcc/testsuite/g++.dg/ipa/pr106260.C
  

Comments

Richard Biener July 14, 2022, 6:53 a.m. UTC | #1
On Wed, Jul 13, 2022 at 11:06 PM Martin Jambor <mjambor@suse.cz> wrote:
>
> Hi,
>
> with -fno-toplevel-reorder (and -fwhole-program), there apparently can
> be local functions without any callers.

Did you check why?  Can't we fix that?

>  This is something that IPA-CP
> does not like because its propagation verifier checks that local
> functions do not end up with TOP in their lattices.  Therefore there
> is an assert checking that all call-less unreachable functions have
> been removed, which triggers in PR 106260 with these two options.
>
> This patch detects the situation and marks the lattices as variable,
> thus avoiding both the assert trigger and the verification failure.
>
> Bootstrapped and tested on x86_64-linux.  OK for master and then all
> active release branches?
>
> Thanks,
>
> Martin
>
>
> gcc/ChangeLog:
>
> 2022-07-13  Martin Jambor  <mjambor@suse.cz>
>
>         PR ipa/106260
>         * ipa-cp.cc (initialize_node_lattices): Replace assert that there are
>         callers with handling that situation when -fno-toplevel_reorder.
>
> gcc/testsuite/ChangeLog:
>
> 2022-07-13  Martin Jambor  <mjambor@suse.cz>
>
>         PR ipa/106260
>         * g++.dg/ipa/pr106260.C: New test.
> ---
>  gcc/ipa-cp.cc                       |  6 ++-
>  gcc/testsuite/g++.dg/ipa/pr106260.C | 64 +++++++++++++++++++++++++++++
>  2 files changed, 69 insertions(+), 1 deletion(-)
>  create mode 100644 gcc/testsuite/g++.dg/ipa/pr106260.C
>
> diff --git a/gcc/ipa-cp.cc b/gcc/ipa-cp.cc
> index 543a9334e2c..f699a8dadc0 100644
> --- a/gcc/ipa-cp.cc
> +++ b/gcc/ipa-cp.cc
> @@ -1286,10 +1286,14 @@ initialize_node_lattices (struct cgraph_node *node)
>        int caller_count = 0;
>        node->call_for_symbol_thunks_and_aliases (count_callers, &caller_count,
>                                                 true);
> -      gcc_checking_assert (caller_count > 0);
>        if (caller_count == 1)
>         node->call_for_symbol_thunks_and_aliases (set_single_call_flag,
>                                                   NULL, true);
> +      else if (caller_count == 0)
> +       {
> +         gcc_checking_assert (!opt_for_fn (node->decl, flag_toplevel_reorder));
> +         variable = true;
> +       }
>      }
>    else
>      {
> diff --git a/gcc/testsuite/g++.dg/ipa/pr106260.C b/gcc/testsuite/g++.dg/ipa/pr106260.C
> new file mode 100644
> index 00000000000..bd3b6e0af79
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/ipa/pr106260.C
> @@ -0,0 +1,64 @@
> +// { dg-do compile }
> +// { dg-options "-O2 -std=gnu++14 -fwhole-program -fno-unit-at-a-time" }
> +
> +struct A;
> +template <class T>
> +struct Q { Q (T); };
> +template<typename T, class D>
> +struct U {
> +  ~U () { m1 (nullptr); }
> +  D m2 ();
> +  T *u;
> +  void m1 (T *) { m2 () (u); }
> +};
> +struct F { F (int *); };
> +template <class, class T = F>
> +using W = Q<T>;
> +int a, b;
> +void fn1 (void *);
> +template <class T>
> +void
> +fn2 (T *x)
> +{
> +  if (x)
> +    x->~T();
> +  fn1 (x);
> +}
> +template <typename T>
> +struct C {
> +  void operator() (T *x) { fn2 (x); }
> +};
> +struct D;
> +template <typename T, typename D = C<T> >
> +using V = U<T, D>;
> +struct A {
> +  A (int *);
> +};
> +struct S;
> +struct G {
> +  V<S> m3 ();
> +};
> +struct S {
> +  int e;
> +  virtual ~S () {}
> +};
> +template<typename T>
> +struct H {
> +  H (int, T x, int) : h(x) {}
> +  G g;
> +  void m4 () { g.m3 (); }
> +  T h;
> +};
> +struct I {
> +  I(A, W<D>);
> +};
> +void
> +test ()
> +{
> +  A c (&b);
> +  W<D> d (&b);
> +  I e (c, d);
> +  H<I> f (0, e, a);
> +  f.m4 ();
> +}
> +
> --
> 2.36.1
>
  
Martin Jambor July 14, 2022, 9:38 a.m. UTC | #2
Hi,

On Thu, Jul 14 2022, Richard Biener wrote:
> On Wed, Jul 13, 2022 at 11:06 PM Martin Jambor <mjambor@suse.cz> wrote:
>>
>> Hi,
>>
>> with -fno-toplevel-reorder (and -fwhole-program), there apparently can
>> be local functions without any callers.
>
> Did you check why?  Can't we fix that?

no, I have not checked much.

Our manual description of -fno-toplevel-reorder says that:

  "When this option is used, unreferenced static variables are not
   removed."

In the beginning, static (because of -fwhole-program) variables _ZTV1S
and __gxx_personality_v0 do refer to the destructor in question, my
hypothesis is that the dead function removal code somehow has to be
careful in case these have been already output, even when these two
variables are no longer in the symbol table when IPA-CP runs.

But I did not look deeper, hoping that Honza would correct me if I am
totally off.  Perhaps the bug is that the function should not be local
even with -fwhole-program.  The semantics of this combination of options
is a bit unclear to me.

Martin


>
>>  This is something that IPA-CP
>> does not like because its propagation verifier checks that local
>> functions do not end up with TOP in their lattices.  Therefore there
>> is an assert checking that all call-less unreachable functions have
>> been removed, which triggers in PR 106260 with these two options.
>>
>> This patch detects the situation and marks the lattices as variable,
>> thus avoiding both the assert trigger and the verification failure.
>>
>> Bootstrapped and tested on x86_64-linux.  OK for master and then all
>> active release branches?
>>
>> Thanks,
>>
>> Martin
>>
>>
>> gcc/ChangeLog:
>>
>> 2022-07-13  Martin Jambor  <mjambor@suse.cz>
>>
>>         PR ipa/106260
>>         * ipa-cp.cc (initialize_node_lattices): Replace assert that there are
>>         callers with handling that situation when -fno-toplevel_reorder.
>>
>> gcc/testsuite/ChangeLog:
>>
>> 2022-07-13  Martin Jambor  <mjambor@suse.cz>
>>
>>         PR ipa/106260
>>         * g++.dg/ipa/pr106260.C: New test.
>> ---
>>  gcc/ipa-cp.cc                       |  6 ++-
>>  gcc/testsuite/g++.dg/ipa/pr106260.C | 64 +++++++++++++++++++++++++++++
>>  2 files changed, 69 insertions(+), 1 deletion(-)
>>  create mode 100644 gcc/testsuite/g++.dg/ipa/pr106260.C
>>
>> diff --git a/gcc/ipa-cp.cc b/gcc/ipa-cp.cc
>> index 543a9334e2c..f699a8dadc0 100644
>> --- a/gcc/ipa-cp.cc
>> +++ b/gcc/ipa-cp.cc
>> @@ -1286,10 +1286,14 @@ initialize_node_lattices (struct cgraph_node *node)
>>        int caller_count = 0;
>>        node->call_for_symbol_thunks_and_aliases (count_callers, &caller_count,
>>                                                 true);
>> -      gcc_checking_assert (caller_count > 0);
>>        if (caller_count == 1)
>>         node->call_for_symbol_thunks_and_aliases (set_single_call_flag,
>>                                                   NULL, true);
>> +      else if (caller_count == 0)
>> +       {
>> +         gcc_checking_assert (!opt_for_fn (node->decl, flag_toplevel_reorder));
>> +         variable = true;
>> +       }
>>      }
>>    else
>>      {
>> diff --git a/gcc/testsuite/g++.dg/ipa/pr106260.C b/gcc/testsuite/g++.dg/ipa/pr106260.C
>> new file mode 100644
>> index 00000000000..bd3b6e0af79
>> --- /dev/null
>> +++ b/gcc/testsuite/g++.dg/ipa/pr106260.C
>> @@ -0,0 +1,64 @@
>> +// { dg-do compile }
>> +// { dg-options "-O2 -std=gnu++14 -fwhole-program -fno-unit-at-a-time" }
>> +
>> +struct A;
>> +template <class T>
>> +struct Q { Q (T); };
>> +template<typename T, class D>
>> +struct U {
>> +  ~U () { m1 (nullptr); }
>> +  D m2 ();
>> +  T *u;
>> +  void m1 (T *) { m2 () (u); }
>> +};
>> +struct F { F (int *); };
>> +template <class, class T = F>
>> +using W = Q<T>;
>> +int a, b;
>> +void fn1 (void *);
>> +template <class T>
>> +void
>> +fn2 (T *x)
>> +{
>> +  if (x)
>> +    x->~T();
>> +  fn1 (x);
>> +}
>> +template <typename T>
>> +struct C {
>> +  void operator() (T *x) { fn2 (x); }
>> +};
>> +struct D;
>> +template <typename T, typename D = C<T> >
>> +using V = U<T, D>;
>> +struct A {
>> +  A (int *);
>> +};
>> +struct S;
>> +struct G {
>> +  V<S> m3 ();
>> +};
>> +struct S {
>> +  int e;
>> +  virtual ~S () {}
>> +};
>> +template<typename T>
>> +struct H {
>> +  H (int, T x, int) : h(x) {}
>> +  G g;
>> +  void m4 () { g.m3 (); }
>> +  T h;
>> +};
>> +struct I {
>> +  I(A, W<D>);
>> +};
>> +void
>> +test ()
>> +{
>> +  A c (&b);
>> +  W<D> d (&b);
>> +  I e (c, d);
>> +  H<I> f (0, e, a);
>> +  f.m4 ();
>> +}
>> +
>> --
>> 2.36.1
>>
  
Jan Hubicka July 18, 2022, 10:23 a.m. UTC | #3
> Hi,
> 
> with -fno-toplevel-reorder (and -fwhole-program), there apparently can
> be local functions without any callers.  This is something that IPA-CP

If there is possibility to trigger a local function without callers, I
think one can also make two local functions calling each other but with
no other callers.

So I think you need to wait until dataflow solution stabilizes and then
then find such isolated cases and do something sensible (as of not
crashing since the code is dead anyway).

-fno-toplevel-reorder was kind of meant to get close to what -fno-unit-at-a-time
did to make legacy code (kernel) happy.  Without unit-at-a-time we did
not remove local functions unless we decided to inline them since they
were output before we had chance to realize that they were unused.

Since we document that only about static variables, perhaps we could
simply start removing functions to avoid these side cases?

Honza

> does not like because its propagation verifier checks that local
> functions do not end up with TOP in their lattices.  Therefore there
> is an assert checking that all call-less unreachable functions have
> been removed, which triggers in PR 106260 with these two options.
> 
> This patch detects the situation and marks the lattices as variable,
> thus avoiding both the assert trigger and the verification failure.
> 
> Bootstrapped and tested on x86_64-linux.  OK for master and then all
> active release branches?
> 
> Thanks,
> 
> Martin
> 
> 
> gcc/ChangeLog:
> 
> 2022-07-13  Martin Jambor  <mjambor@suse.cz>
> 
> 	PR ipa/106260
> 	* ipa-cp.cc (initialize_node_lattices): Replace assert that there are
> 	callers with handling that situation when -fno-toplevel_reorder.
> 
> gcc/testsuite/ChangeLog:
> 
> 2022-07-13  Martin Jambor  <mjambor@suse.cz>
> 
> 	PR ipa/106260
> 	* g++.dg/ipa/pr106260.C: New test.
> ---
>  gcc/ipa-cp.cc                       |  6 ++-
>  gcc/testsuite/g++.dg/ipa/pr106260.C | 64 +++++++++++++++++++++++++++++
>  2 files changed, 69 insertions(+), 1 deletion(-)
>  create mode 100644 gcc/testsuite/g++.dg/ipa/pr106260.C
> 
> diff --git a/gcc/ipa-cp.cc b/gcc/ipa-cp.cc
> index 543a9334e2c..f699a8dadc0 100644
> --- a/gcc/ipa-cp.cc
> +++ b/gcc/ipa-cp.cc
> @@ -1286,10 +1286,14 @@ initialize_node_lattices (struct cgraph_node *node)
>        int caller_count = 0;
>        node->call_for_symbol_thunks_and_aliases (count_callers, &caller_count,
>  						true);
> -      gcc_checking_assert (caller_count > 0);
>        if (caller_count == 1)
>  	node->call_for_symbol_thunks_and_aliases (set_single_call_flag,
>  						  NULL, true);
> +      else if (caller_count == 0)
> +	{
> +	  gcc_checking_assert (!opt_for_fn (node->decl, flag_toplevel_reorder));
> +	  variable = true;
> +	}
>      }
>    else
>      {
> diff --git a/gcc/testsuite/g++.dg/ipa/pr106260.C b/gcc/testsuite/g++.dg/ipa/pr106260.C
> new file mode 100644
> index 00000000000..bd3b6e0af79
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/ipa/pr106260.C
> @@ -0,0 +1,64 @@
> +// { dg-do compile }
> +// { dg-options "-O2 -std=gnu++14 -fwhole-program -fno-unit-at-a-time" }
> +
> +struct A;
> +template <class T>
> +struct Q { Q (T); };
> +template<typename T, class D>
> +struct U {
> +  ~U () { m1 (nullptr); }
> +  D m2 ();
> +  T *u;
> +  void m1 (T *) { m2 () (u); }
> +};
> +struct F { F (int *); };
> +template <class, class T = F>
> +using W = Q<T>;
> +int a, b;
> +void fn1 (void *);
> +template <class T>
> +void
> +fn2 (T *x)
> +{
> +  if (x)
> +    x->~T();
> +  fn1 (x);
> +}
> +template <typename T>
> +struct C {
> +  void operator() (T *x) { fn2 (x); }
> +};
> +struct D;
> +template <typename T, typename D = C<T> >
> +using V = U<T, D>;
> +struct A {
> +  A (int *);
> +};
> +struct S;
> +struct G {
> +  V<S> m3 ();
> +};
> +struct S {
> +  int e;
> +  virtual ~S () {}
> +};
> +template<typename T>
> +struct H {
> +  H (int, T x, int) : h(x) {}
> +  G g;
> +  void m4 () { g.m3 (); }
> +  T h;
> +};
> +struct I {
> +  I(A, W<D>);
> +};
> +void
> +test ()
> +{
> +  A c (&b);
> +  W<D> d (&b);
> +  I e (c, d);
> +  H<I> f (0, e, a);
> +  f.m4 ();
> +}
> +
> -- 
> 2.36.1
>
  
Martin Jambor July 18, 2022, 4:10 p.m. UTC | #4
Hi,

On Mon, Jul 18 2022, Jan Hubicka wrote:
>> Hi,
>> 
>> with -fno-toplevel-reorder (and -fwhole-program), there apparently can
>> be local functions without any callers.  This is something that IPA-CP
>
> If there is possibility to trigger a local function without callers, I
> think one can also make two local functions calling each other but with
> no other callers.

I am not sure I understand.  The above can happen only with
-fno-toplevel-reorder, right?

>
> So I think you need to wait until dataflow solution stabilizes and
> then then find such isolated cases and do something sensible (as of
> not crashing since the code is dead anyway).

After dataflow stabilizes, it crashes only because a verifier does not
like the result.  Perhaps the solution is simply not verify nodes built
with -fno-toplevel-reorder?  But the assert was useful in the past in
the general case so I'd like to keep it.

I can now see that the proposed patch could also pessimize the lattices
of functions that the dead functions call a bit, although arguably it
might be the right thing to do, since we cannot say anything about these
calls.  But I do not care much.

>
> -fno-toplevel-reorder was kind of meant to get close to what -fno-unit-at-a-time
> did to make legacy code (kernel) happy.  Without unit-at-a-time we did
> not remove local functions unless we decided to inline them since they
> were output before we had chance to realize that they were unused.

I understand that it is legacy stuff but actually I wonder what the
option means for IPA in general.  Is IPA run after we might have output
something already?  (If not then why do we see any differences?  If yes,
can all IPA optimizations work, especially with -fwhole-program?)

>
> Since we document that only about static variables, perhaps we could
> simply start removing functions to avoid these side cases?

Well, the funny thing is that we eventually do remove the function, just
at IPA-CP stage it is there and not even local yet (I have not noticed
this last week).  Removing it earlier would definitely make life simpler
for me :-)

Anyway, I'm happy to work in whichever direction you prefer, just now I
am not sure which one it is.

Thanks,

Martin

>
>> does not like because its propagation verifier checks that local
>> functions do not end up with TOP in their lattices.  Therefore there
>> is an assert checking that all call-less unreachable functions have
>> been removed, which triggers in PR 106260 with these two options.
>> 
>> This patch detects the situation and marks the lattices as variable,
>> thus avoiding both the assert trigger and the verification failure.
>> 
>> Bootstrapped and tested on x86_64-linux.  OK for master and then all
>> active release branches?
>> 
>> Thanks,
>> 
>> Martin
>> 
>> 
>> gcc/ChangeLog:
>> 
>> 2022-07-13  Martin Jambor  <mjambor@suse.cz>
>> 
>> 	PR ipa/106260
>> 	* ipa-cp.cc (initialize_node_lattices): Replace assert that there are
>> 	callers with handling that situation when -fno-toplevel_reorder.
>>
  

Patch

diff --git a/gcc/ipa-cp.cc b/gcc/ipa-cp.cc
index 543a9334e2c..f699a8dadc0 100644
--- a/gcc/ipa-cp.cc
+++ b/gcc/ipa-cp.cc
@@ -1286,10 +1286,14 @@  initialize_node_lattices (struct cgraph_node *node)
       int caller_count = 0;
       node->call_for_symbol_thunks_and_aliases (count_callers, &caller_count,
 						true);
-      gcc_checking_assert (caller_count > 0);
       if (caller_count == 1)
 	node->call_for_symbol_thunks_and_aliases (set_single_call_flag,
 						  NULL, true);
+      else if (caller_count == 0)
+	{
+	  gcc_checking_assert (!opt_for_fn (node->decl, flag_toplevel_reorder));
+	  variable = true;
+	}
     }
   else
     {
diff --git a/gcc/testsuite/g++.dg/ipa/pr106260.C b/gcc/testsuite/g++.dg/ipa/pr106260.C
new file mode 100644
index 00000000000..bd3b6e0af79
--- /dev/null
+++ b/gcc/testsuite/g++.dg/ipa/pr106260.C
@@ -0,0 +1,64 @@ 
+// { dg-do compile }
+// { dg-options "-O2 -std=gnu++14 -fwhole-program -fno-unit-at-a-time" }
+
+struct A;
+template <class T>
+struct Q { Q (T); };
+template<typename T, class D>
+struct U {
+  ~U () { m1 (nullptr); }
+  D m2 ();
+  T *u;
+  void m1 (T *) { m2 () (u); }
+};
+struct F { F (int *); };
+template <class, class T = F>
+using W = Q<T>;
+int a, b;
+void fn1 (void *);
+template <class T>
+void
+fn2 (T *x)
+{
+  if (x)
+    x->~T();
+  fn1 (x);
+}
+template <typename T>
+struct C {
+  void operator() (T *x) { fn2 (x); }
+};
+struct D;
+template <typename T, typename D = C<T> >
+using V = U<T, D>;
+struct A {
+  A (int *);
+};
+struct S;
+struct G {
+  V<S> m3 ();
+};
+struct S {
+  int e;
+  virtual ~S () {}
+};
+template<typename T>
+struct H {
+  H (int, T x, int) : h(x) {}
+  G g;
+  void m4 () { g.m3 (); }
+  T h;
+};
+struct I {
+  I(A, W<D>);
+};
+void
+test ()
+{
+  A c (&b);
+  W<D> d (&b);
+  I e (c, d);
+  H<I> f (0, e, a);
+  f.m4 ();
+}
+