Re: [PATCH] ipa: Careful processing ANCESTOR jump functions and NULL pointers (PR 103083)

[Jan Hubicka via Gcc-patches]

> IPA_JF_ANCESTOR jump functions are constructed also when the formal
> parameter of the caller is first checked whether it is NULL and left
> as it is if it is NULL, to accommodate C++ casts to an ancestor class.
> 
> The jump function type was invented for devirtualization and IPA-CP
> propagation of tree constants is also careful to apply it only to
> existing DECLs(*) but as PR 103083 shows, the part propagating "known
> bits" was not careful about this, which can lead to miscompilations.
> 
> This patch introduces a flag to the ancestor jump functions which
> tells whether a NULL-check was elided when creating it and makes the
> bits propagation behave accordingly, masking any bits otherwise would
> be known to be one.  This should safely preserve alignment info, which
> is the primary ifnormation that we keep in bits for pointers.
> 
> (*) There still may remain problems when a DECL resides on address
> zero (with -fno-delete-null-pointer-checks ...I hope it cannot happen
> otherwise).  I am looking into that now but I think it will be easier
> for everyone if I do so in a follow-up patch.
> 
> gcc/ChangeLog:
> 
> 2022-02-11  Martin Jambor  
> 
>   PR ipa/103083
>   * ipa-prop.h (ipa_ancestor_jf_data): New flag keep_null;
>   (ipa_get_jf_ancestor_keep_null): New function.
>   * ipa-prop.c (ipa_set_ancestor_jf): Initialize keep_null field of the
>   ancestor function.
>   (compute_complex_assign_jump_func): Pass false to keep_null
>   parameter of ipa_set_ancestor_jf.
>   (compute_complex_ancestor_jump_func): Pass true to keep_null
>   parameter of ipa_set_ancestor_jf.
>   (update_jump_functions_after_inlining): Carry over keep_null from the
>   original ancestor jump-function or merge them.
>   (ipa_write_jump_function): Stream keep_null flag.
>   (ipa_read_jump_function): Likewise.
>   (ipa_print_node_jump_functions_for_edge): Print the new flag.
>   * ipa-cp.c (class ipcp_bits_lattice): Make various getters const.  New
>   member function known_nonzero_p.
>   (ipcp_bits_lattice::known_nonzero_p): New.
>   (ipcp_bits_lattice::meet_with_1): New parameter drop_all_ones,
>   observe it.
>   (ipcp_bits_lattice::meet_with): Likewise.
>   (propagate_bits_across_jump_function): Simplify.  Pass true in
>   drop_all_ones when it is necessary.
>   (propagate_aggs_across_jump_function): Take care of keep_null
>   flag.
>   (ipa_get_jf_ancestor_result): Propagate NULL accross keep_null
>   jump functions.
> 
> gcc/testsuite/ChangeLog:
> 
> 2021-11-25  Martin Jambor  
> 
>   * gcc.dg/ipa/pr103083-1.c: New test.
>   * gcc.dg/ipa/pr103083-2.c: Likewise.

OK,
thanks!
Honza

Re: [PATCH] ipa: Careful processing ANCESTOR jump functions and NULL pointers (PR 103083)

[Martin Jambor]

Hello Honza,

On Mon, Dec 13 2021, Jan Hubicka wrote:
>> >>> +  || (only_for_nonzero && 
>> >>> !src_lats->bits_lattice.known_nonzero_p ()))
>> >>> +{
>> >>> +  if (jfunc->bits)
>> >>> +return dest_lattice->meet_with (jfunc->bits->value,
>> >>> +jfunc->bits->mask, 
>> >>> precision);
>> >>> +  else
>> >>> +return dest_lattice->set_to_bottom ();
>> >>> +}
>> >> I do not think you need to set to bottom here. For pointers, we
>> >> primarily track alignment and NULL is aligned - all you need to do is to
>> >> make sure that we do not believe that some bits are 1.
>> >
>> > I am not sure I understand, the testcase you wrote has all bits as zeros
>> > and still miscompiles?  It is primarily used for alignment but not only
>> > for that.
>
> Maybe I misunderstand the code.  But if you have only_for_nonzero and
> you do not know htat src lattice is non-zero you will get
>  - if src is 0, then dest is 0
>  - if src is non-zero then dest is src+offset
>
> So all trialing bits that are known to be 0 in src and offset are known
> to be 0 in the result.  But I do not see where th eoffset is mixed in?
>

I went back and tried to figure out again what you meant and I hope it
was the following patch, which does not drop the lattice to bottom for
ANCESTORs but instead masks any bits that would be considered known
ones.  It is indeed clever and I did not see the possibility when
writing the patch.

The following has passed bootstrap, LTO bootstrap and testing on
x86-64.  OK for trunk (and then to be backported to 11 and 10)?

Thanks,

Martin


IPA_JF_ANCESTOR jump functions are constructed also when the formal
parameter of the caller is first checked whether it is NULL and left
as it is if it is NULL, to accommodate C++ casts to an ancestor class.

The jump function type was invented for devirtualization and IPA-CP
propagation of tree constants is also careful to apply it only to
existing DECLs(*) but as PR 103083 shows, the part propagating "known
bits" was not careful about this, which can lead to miscompilations.

This patch introduces a flag to the ancestor jump functions which
tells whether a NULL-check was elided when creating it and makes the
bits propagation behave accordingly, masking any bits otherwise would
be known to be one.  This should safely preserve alignment info, which
is the primary ifnormation that we keep in bits for pointers.

(*) There still may remain problems when a DECL resides on address
zero (with -fno-delete-null-pointer-checks ...I hope it cannot happen
otherwise).  I am looking into that now but I think it will be easier
for everyone if I do so in a follow-up patch.

gcc/ChangeLog:

2022-02-11  Martin Jambor  

PR ipa/103083
* ipa-prop.h (ipa_ancestor_jf_data): New flag keep_null;
(ipa_get_jf_ancestor_keep_null): New function.
* ipa-prop.c (ipa_set_ancestor_jf): Initialize keep_null field of the
ancestor function.
(compute_complex_assign_jump_func): Pass false to keep_null
parameter of ipa_set_ancestor_jf.
(compute_complex_ancestor_jump_func): Pass true to keep_null
parameter of ipa_set_ancestor_jf.
(update_jump_functions_after_inlining): Carry over keep_null from the
original ancestor jump-function or merge them.
(ipa_write_jump_function): Stream keep_null flag.
(ipa_read_jump_function): Likewise.
(ipa_print_node_jump_functions_for_edge): Print the new flag.
* ipa-cp.c (class ipcp_bits_lattice): Make various getters const.  New
member function known_nonzero_p.
(ipcp_bits_lattice::known_nonzero_p): New.
(ipcp_bits_lattice::meet_with_1): New parameter drop_all_ones,
observe it.
(ipcp_bits_lattice::meet_with): Likewise.
(propagate_bits_across_jump_function): Simplify.  Pass true in
drop_all_ones when it is necessary.
(propagate_aggs_across_jump_function): Take care of keep_null
flag.
(ipa_get_jf_ancestor_result): Propagate NULL accross keep_null
jump functions.

gcc/testsuite/ChangeLog:

2021-11-25  Martin Jambor  

* gcc.dg/ipa/pr103083-1.c: New test.
* gcc.dg/ipa/pr103083-2.c: Likewise.
---
 gcc/ipa-cp.cc | 75 ++-
 gcc/ipa-prop.cc   | 20 +--
 gcc/ipa-prop.h| 13 +
 gcc/testsuite/gcc.dg/ipa/pr103083-1.c | 28 ++
 gcc/testsuite/gcc.dg/ipa/pr103083-2.c | 30 +++
 5 files changed, 137 insertions(+), 29 deletions(-)
 create mode 100644 gcc/testsuite/gcc.dg/ipa/pr103083-1.c
 create mode 100644 gcc/testsuite/gcc.dg/ipa/pr103083-2.c

diff --git a/gcc/ipa-cp.cc b/gcc/ipa-cp.cc
index 453e9c93cc3..93a54ae83fc 100644
--- a/gcc/ipa-cp.cc
+++ b/gcc/ipa-cp.cc
@@ -306,17 +306,18 @@ public:
 class ipcp_bits_lattice
 {
 public:
-  bool bottom_p () { return m_lattice_val ==

Re: [PATCH] ipa: Careful processing ANCESTOR jump functions and NULL pointers (PR 103083)

[Martin Jambor]

On Mon, Dec 13 2021, Jan Hubicka wrote:
>> >>> +  || (only_for_nonzero && 
>> >>> !src_lats->bits_lattice.known_nonzero_p ()))
>> >>> +{
>> >>> +  if (jfunc->bits)
>> >>> +return dest_lattice->meet_with (jfunc->bits->value,
>> >>> +jfunc->bits->mask, 
>> >>> precision);
>> >>> +  else
>> >>> +return dest_lattice->set_to_bottom ();
>> >>> +}
>> >> I do not think you need to set to bottom here. For pointers, we
>> >> primarily track alignment and NULL is aligned - all you need to do is to
>> >> make sure that we do not believe that some bits are 1.
>> >
>> > I am not sure I understand, the testcase you wrote has all bits as zeros
>> > and still miscompiles?  It is primarily used for alignment but not only
>> > for that.
>
> Maybe I misunderstand the code.  But if you have only_for_nonzero and
> you do not know htat src lattice is non-zero you will get
>  - if src is 0, then dest is 0
>  - if src is non-zero then dest is src+offset
>

or all known bits of src are zero but then there are unknown masked-out
bits with unknown value and in that case we just don't know.  The patch
does src+offset when there is a known non-zero bit but we can only set
dest to zero if all bits are known.  But as you pointed out, the
constant part of IPA-CP should catch this case, and now does, so I
decided not to handle it here.

If you prefer, I can add a case if both m_value and m_mask or zero
(well, could the latter in theory only need "precision" zeros)? and if
so, simply meet the destination lattice with the source one.

Is that what you are asking for?

Thanks,

Martin

Re: [PATCH] ipa: Careful processing ANCESTOR jump functions and NULL pointers (PR 103083)

[Jan Hubicka via Gcc-patches]

> >>> +   || (only_for_nonzero && !src_lats->bits_lattice.known_nonzero_p ()))
> >>> + {
> >>> +   if (jfunc->bits)
> >>> + return dest_lattice->meet_with (jfunc->bits->value,
> >>> + jfunc->bits->mask, precision);
> >>> +   else
> >>> + return dest_lattice->set_to_bottom ();
> >>> + }
> >> I do not think you need to set to bottom here. For pointers, we
> >> primarily track alignment and NULL is aligned - all you need to do is to
> >> make sure that we do not believe that some bits are 1.
> >
> > I am not sure I understand, the testcase you wrote has all bits as zeros
> > and still miscompiles?  It is primarily used for alignment but not only
> > for that.

Maybe I misunderstand the code.  But if you have only_for_nonzero and
you do not know htat src lattice is non-zero you will get
 - if src is 0, then dest is 0
 - if src is non-zero then dest is src+offset

So all trialing bits that are known to be 0 in src and offset are known
to be 0 in the result.  But I do not see where th eoffset is mixed in?

Honza

Re: [PATCH] ipa: Careful processing ANCESTOR jump functions and NULL pointers (PR 103083)

[Martin Jambor]

Hello,

I would like to ping the patch below.

Martin

On Mon, Nov 29 2021, Martin Jambor wrote:
> Hi,
>
> On Sat, Nov 27 2021, Jan Hubicka wrote:
>>> Hi,
>>> 
>>> IPA_JF_ANCESTOR jump functions are constructed also when the formal
>>> parameter of the caller is first checked whether it is NULL and left
>>> as it is if it is NULL, to accommodate C++ casts to an ancestor class.
>>> 
>>> The jump function type was invented for devirtualization and IPA-CP
>>> propagation of tree constants is also careful to apply it only to
>>> existing DECLs(*) but as PR 103083 shows, the part propagating "known
>>> bits" was not careful about this, which can lead to miscompilations.
>>> 
>>> This patch introduces a flag to the ancestor jump functions which
>>> tells whether a NULL-check was elided when creating it and makes the
>>> bits propagation behave accordingly.
>>> 
>>> (*) There still may remain problems when a DECL resides on address
>>> zero (with -fno-delete-null-pointer-checks ...I hope it cannot happen
>>> otherwise).  I am looking into that now but I think it will be easier
>>> for everyone if I do so in a follow-up patch.
>>
>> I guess so.  Thinking about it bit more after our discussion yesterday
>> I think it really matters where the ADDR_EXPR was created (since in
>> funciton with -fno-delete-null-pointer-checks it might be equal to 0)
>> and we need to somehow keep track of it.
>>
>> One observation is that it is unsafe to constant propagate
>> ADDR_EXPR with -fno-delete-null-pointer-checks to function with
>> -fdelete-null-pointer-checks, so we may just simply stop propagating
>> known ADDR_EXPRs on when caller is -fno... and call is -f
>
> Makes sense.
>
>>> +/* Return true if any of the known bits are non-zero.  */
>>> +bool
>>> +ipcp_bits_lattice::known_nonzero_p () const
>>> +{
>>> +  if (!constant_p ())
>>> +return false;
>>> +  return !wi::eq_p (wi::bit_and (wi::bit_not (m_mask), m_value), 0);
>> There is also ne_p
>
> Changed.
>
>>> @@ -2374,6 +2384,7 @@ propagate_bits_across_jump_function (cgraph_edge *cs, 
>>> int idx,
>>>tree operand = NULL_TREE;
>>>enum tree_code code;
>>>unsigned src_idx;
>>> +  bool only_for_nonzero = false;
>>>  
>>>if (jfunc->type == IPA_JF_PASS_THROUGH)
>>> {
>>> @@ -2386,7 +2397,9 @@ propagate_bits_across_jump_function (cgraph_edge *cs, 
>>> int idx,
>>> {
>>>   code = POINTER_PLUS_EXPR;
>>>   src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
>>> - unsigned HOST_WIDE_INT offset = ipa_get_jf_ancestor_offset (jfunc) / 
>>> BITS_PER_UNIT;
>>> + unsigned HOST_WIDE_INT offset
>>> +   = ipa_get_jf_ancestor_offset (jfunc) / BITS_PER_UNIT;
>> I still think the offset should be in bytes (and probably poly_int64).
>> There is get_addr_base_and_unit_offset
>
> I agree about bytes, not sure about poly_ints, but could be persuaded.
> But probably not as a part of this patch?
>
>>
>>> + only_for_nonzero = (ipa_get_jf_ancestor_keep_null (jfunc) || !offset);
>>>   operand = build_int_cstu (size_type_node, offset);
>>> }
>>>  
>>> @@ -2404,16 +2417,18 @@ propagate_bits_across_jump_function (cgraph_edge 
>>> *cs, int idx,
>>>  and we store it in jump function during analysis stage.  */
>>>  
>>>if (src_lats->bits_lattice.bottom_p ()
>>> - && jfunc->bits)
>>> -   return dest_lattice->meet_with (jfunc->bits->value, jfunc->bits->mask,
>>> -   precision);
>>> + || (only_for_nonzero && !src_lats->bits_lattice.known_nonzero_p ()))
>>> +   {
>>> + if (jfunc->bits)
>>> +   return dest_lattice->meet_with (jfunc->bits->value,
>>> +   jfunc->bits->mask, precision);
>>> + else
>>> +   return dest_lattice->set_to_bottom ();
>>> +   }
>> I do not think you need to set to bottom here. For pointers, we
>> primarily track alignment and NULL is aligned - all you need to do is to
>> make sure that we do not believe that some bits are 1.
>
> I am not sure I understand, the testcase you wrote has all bits as zeros
> and still miscompiles?  It is primarily used for alignment but not only
> for that.
>
>>> @@ -3327,7 +3332,8 @@ update_jump_functions_after_inlining (struct 
>>> cgraph_edge *cs,
>>> ipa_set_ancestor_jf (dst,
>>>  ipa_get_jf_ancestor_offset (src),
>>>  ipa_get_jf_ancestor_formal_id (src),
>>> -agg_p);
>>> +agg_p,
>>> +ipa_get_jf_ancestor_keep_null (src));
>>> break;
>> Somewhere here you need to consider the case where you have two accessor
>> jump functions to combine together and merge the keep_null flags...
>
> Oops, I missed that, fixed.
>
>>
>> Also with the flags, I think you want to modify ipa-cp so NULL pointers
>> gets propagated acroess the ancestor jump functions.
>
> OK, added, along with a new

Re: [PATCH] ipa: Careful processing ANCESTOR jump functions and NULL pointers (PR 103083)

[Martin Jambor]

Hi,

On Sat, Nov 27 2021, Jan Hubicka wrote:
>> Hi,
>> 
>> IPA_JF_ANCESTOR jump functions are constructed also when the formal
>> parameter of the caller is first checked whether it is NULL and left
>> as it is if it is NULL, to accommodate C++ casts to an ancestor class.
>> 
>> The jump function type was invented for devirtualization and IPA-CP
>> propagation of tree constants is also careful to apply it only to
>> existing DECLs(*) but as PR 103083 shows, the part propagating "known
>> bits" was not careful about this, which can lead to miscompilations.
>> 
>> This patch introduces a flag to the ancestor jump functions which
>> tells whether a NULL-check was elided when creating it and makes the
>> bits propagation behave accordingly.
>> 
>> (*) There still may remain problems when a DECL resides on address
>> zero (with -fno-delete-null-pointer-checks ...I hope it cannot happen
>> otherwise).  I am looking into that now but I think it will be easier
>> for everyone if I do so in a follow-up patch.
>
> I guess so.  Thinking about it bit more after our discussion yesterday
> I think it really matters where the ADDR_EXPR was created (since in
> funciton with -fno-delete-null-pointer-checks it might be equal to 0)
> and we need to somehow keep track of it.
>
> One observation is that it is unsafe to constant propagate
> ADDR_EXPR with -fno-delete-null-pointer-checks to function with
> -fdelete-null-pointer-checks, so we may just simply stop propagating
> known ADDR_EXPRs on when caller is -fno... and call is -f

Makes sense.

>> +/* Return true if any of the known bits are non-zero.  */
>> +bool
>> +ipcp_bits_lattice::known_nonzero_p () const
>> +{
>> +  if (!constant_p ())
>> +return false;
>> +  return !wi::eq_p (wi::bit_and (wi::bit_not (m_mask), m_value), 0);
> There is also ne_p

Changed.

>> @@ -2374,6 +2384,7 @@ propagate_bits_across_jump_function (cgraph_edge *cs, 
>> int idx,
>>tree operand = NULL_TREE;
>>enum tree_code code;
>>unsigned src_idx;
>> +  bool only_for_nonzero = false;
>>  
>>if (jfunc->type == IPA_JF_PASS_THROUGH)
>>  {
>> @@ -2386,7 +2397,9 @@ propagate_bits_across_jump_function (cgraph_edge *cs, 
>> int idx,
>>  {
>>code = POINTER_PLUS_EXPR;
>>src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
>> -  unsigned HOST_WIDE_INT offset = ipa_get_jf_ancestor_offset (jfunc) / 
>> BITS_PER_UNIT;
>> +  unsigned HOST_WIDE_INT offset
>> += ipa_get_jf_ancestor_offset (jfunc) / BITS_PER_UNIT;
> I still think the offset should be in bytes (and probably poly_int64).
> There is get_addr_base_and_unit_offset

I agree about bytes, not sure about poly_ints, but could be persuaded.
But probably not as a part of this patch?

>
>> +  only_for_nonzero = (ipa_get_jf_ancestor_keep_null (jfunc) || !offset);
>>operand = build_int_cstu (size_type_node, offset);
>>  }
>>  
>> @@ -2404,16 +2417,18 @@ propagate_bits_across_jump_function (cgraph_edge 
>> *cs, int idx,
>>   and we store it in jump function during analysis stage.  */
>>  
>>if (src_lats->bits_lattice.bottom_p ()
>> -  && jfunc->bits)
>> -return dest_lattice->meet_with (jfunc->bits->value, jfunc->bits->mask,
>> -precision);
>> +  || (only_for_nonzero && !src_lats->bits_lattice.known_nonzero_p ()))
>> +{
>> +  if (jfunc->bits)
>> +return dest_lattice->meet_with (jfunc->bits->value,
>> +jfunc->bits->mask, precision);
>> +  else
>> +return dest_lattice->set_to_bottom ();
>> +}
> I do not think you need to set to bottom here. For pointers, we
> primarily track alignment and NULL is aligned - all you need to do is to
> make sure that we do not believe that some bits are 1.

I am not sure I understand, the testcase you wrote has all bits as zeros
and still miscompiles?  It is primarily used for alignment but not only
for that.

>> @@ -3327,7 +3332,8 @@ update_jump_functions_after_inlining (struct 
>> cgraph_edge *cs,
>>  ipa_set_ancestor_jf (dst,
>>   ipa_get_jf_ancestor_offset (src),
>>   ipa_get_jf_ancestor_formal_id (src),
>> - agg_p);
>> + agg_p,
>> + ipa_get_jf_ancestor_keep_null (src));
>>  break;
> Somewhere here you need to consider the case where you have two accessor
> jump functions to combine together and merge the keep_null flags...

Oops, I missed that, fixed.

>
> Also with the flags, I think you want to modify ipa-cp so NULL pointers
> gets propagated acroess the ancestor jump functions.

OK, added, along with a new testcase.

> Moreover ipa-modref is using ancestor jf to update its summary.  Here I
> think with -fno-delete-null-pointer-checks it is safe to pdate also with
> the flag set, since we know the me

Re: [PATCH] ipa: Careful processing ANCESTOR jump functions and NULL pointers (PR 103083)

[Jan Hubicka via Gcc-patches]

> Hi,
> 
> IPA_JF_ANCESTOR jump functions are constructed also when the formal
> parameter of the caller is first checked whether it is NULL and left
> as it is if it is NULL, to accommodate C++ casts to an ancestor class.
> 
> The jump function type was invented for devirtualization and IPA-CP
> propagation of tree constants is also careful to apply it only to
> existing DECLs(*) but as PR 103083 shows, the part propagating "known
> bits" was not careful about this, which can lead to miscompilations.
> 
> This patch introduces a flag to the ancestor jump functions which
> tells whether a NULL-check was elided when creating it and makes the
> bits propagation behave accordingly.
> 
> (*) There still may remain problems when a DECL resides on address
> zero (with -fno-delete-null-pointer-checks ...I hope it cannot happen
> otherwise).  I am looking into that now but I think it will be easier
> for everyone if I do so in a follow-up patch.

I guess so.  Thinking about it bit more after our discussion yesterday
I think it really matters where the ADDR_EXPR was created (since in
funciton with -fno-delete-null-pointer-checks it might be equal to 0)
and we need to somehow keep track of it.

One observation is that it is unsafe to constant propagate
ADDR_EXPR with -fno-delete-null-pointer-checks to function with
-fdelete-null-pointer-checks, so we may just simply stop propagating
known ADDR_EXPRs on when caller is -fno... and call is -f
> +/* Return true if any of the known bits are non-zero.  */
> +bool
> +ipcp_bits_lattice::known_nonzero_p () const
> +{
> +  if (!constant_p ())
> +return false;
> +  return !wi::eq_p (wi::bit_and (wi::bit_not (m_mask), m_value), 0);
There is also ne_p
> @@ -2374,6 +2384,7 @@ propagate_bits_across_jump_function (cgraph_edge *cs, 
> int idx,
>tree operand = NULL_TREE;
>enum tree_code code;
>unsigned src_idx;
> +  bool only_for_nonzero = false;
>  
>if (jfunc->type == IPA_JF_PASS_THROUGH)
>   {
> @@ -2386,7 +2397,9 @@ propagate_bits_across_jump_function (cgraph_edge *cs, 
> int idx,
>   {
> code = POINTER_PLUS_EXPR;
> src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
> -   unsigned HOST_WIDE_INT offset = ipa_get_jf_ancestor_offset (jfunc) / 
> BITS_PER_UNIT;
> +   unsigned HOST_WIDE_INT offset
> + = ipa_get_jf_ancestor_offset (jfunc) / BITS_PER_UNIT;
I still think the offset should be in bytes (and probably poly_int64).
There is get_addr_base_and_unit_offset

> +   only_for_nonzero = (ipa_get_jf_ancestor_keep_null (jfunc) || !offset);
> operand = build_int_cstu (size_type_node, offset);
>   }
>  
> @@ -2404,16 +2417,18 @@ propagate_bits_across_jump_function (cgraph_edge *cs, 
> int idx,
>and we store it in jump function during analysis stage.  */
>  
>if (src_lats->bits_lattice.bottom_p ()
> -   && jfunc->bits)
> - return dest_lattice->meet_with (jfunc->bits->value, jfunc->bits->mask,
> - precision);
> +   || (only_for_nonzero && !src_lats->bits_lattice.known_nonzero_p ()))
> + {
> +   if (jfunc->bits)
> + return dest_lattice->meet_with (jfunc->bits->value,
> + jfunc->bits->mask, precision);
> +   else
> + return dest_lattice->set_to_bottom ();
> + }
I do not think you need to set to bottom here. For pointers, we
primarily track alignment and NULL is aligned - all you need to do is to
make sure that we do not believe that some bits are 1.
> @@ -3327,7 +3332,8 @@ update_jump_functions_after_inlining (struct 
> cgraph_edge *cs,
>   ipa_set_ancestor_jf (dst,
>ipa_get_jf_ancestor_offset (src),
>ipa_get_jf_ancestor_formal_id (src),
> -  agg_p);
> +  agg_p,
> +  ipa_get_jf_ancestor_keep_null (src));
>   break;
Somewhere here you need to consider the case where you have two accessor
jump functions to combine together and merge the keep_null flags...

Also with the flags, I think you want to modify ipa-cp so NULL pointers
gets propagated acroess the ancestor jump functions.
Moreover ipa-modref is using ancestor jf to update its summary.  Here I
think with -fno-delete-null-pointer-checks it is safe to pdate also with
the flag set, since we know the memory access is invalid otherwise.

Honza
>   default:
> @@ -4758,6 +4764,7 @@ ipa_write_jump_function (struct output_block *ob,
>streamer_write_uhwi (ob, jump_func->value.ancestor.formal_id);
>bp = bitpack_create (ob->main_stream);
>bp_pack_value (&bp, jump_func->value.ancestor.agg_preserved, 1);
> +  bp_pack_value (&bp, jump_func->value.ancestor.keep_null, 1);
>streamer_write_bitpack (&bp);
>break;
>  default:
> @@ -4883,7 +4890,9 @@ ipa_

[PATCH] ipa: Careful processing ANCESTOR jump functions and NULL pointers (PR 103083)

[Martin Jambor]

Hi,

IPA_JF_ANCESTOR jump functions are constructed also when the formal
parameter of the caller is first checked whether it is NULL and left
as it is if it is NULL, to accommodate C++ casts to an ancestor class.

The jump function type was invented for devirtualization and IPA-CP
propagation of tree constants is also careful to apply it only to
existing DECLs(*) but as PR 103083 shows, the part propagating "known
bits" was not careful about this, which can lead to miscompilations.

This patch introduces a flag to the ancestor jump functions which
tells whether a NULL-check was elided when creating it and makes the
bits propagation behave accordingly.

(*) There still may remain problems when a DECL resides on address
zero (with -fno-delete-null-pointer-checks ...I hope it cannot happen
otherwise).  I am looking into that now but I think it will be easier
for everyone if I do so in a follow-up patch.

Bootstrapped, LTO-bootstrapped and tested on x86_64-linux.  OK for
master?

Thanks,

Martin


gcc/ChangeLog:

2021-11-25  Martin Jambor  

PR ipa/103083
* ipa-prop.h (ipa_ancestor_jf_data): New flag keep_null;
(ipa_get_jf_ancestor_keep_null): New function.
* ipa-prop.c (ipa_set_ancestor_jf): Initialize keep_null field of the
ancestor function.
(compute_complex_assign_jump_func): Pass false to keep_null
parameter of ipa_set_ancestor_jf.
(compute_complex_ancestor_jump_func): Pass true to keep_null
parameter of ipa_set_ancestor_jf.
(update_jump_functions_after_inlining): Carry over keep_null from the
original ancestor jump-function.
(ipa_write_jump_function): Stream keep_null flag.
(ipa_read_jump_function): Likewise.
(ipa_print_node_jump_functions_for_edge): Print the new flag.
* ipa-cp.c (class ipcp_bits_lattice): Make various getters const.  New
member function known_nonzero_p.
(ipcp_bits_lattice::known_nonzero_p): New.
(propagate_bits_across_jump_function): Only process ancestor functions
when safe.  Remove extraneous condition handling ancestor jump
functions.
(propagate_aggs_across_jump_function): Take care of keep_null flag.

gcc/testsuite/ChangeLog:

2021-11-25  Martin Jambor  

* gcc.dg/ipa/pr103083.c: New test.
---
 gcc/ipa-cp.c| 39 -
 gcc/ipa-prop.c  | 19 ++
 gcc/ipa-prop.h  | 13 ++
 gcc/testsuite/gcc.dg/ipa/pr103083.c | 28 +
 4 files changed, 82 insertions(+), 17 deletions(-)
 create mode 100644 gcc/testsuite/gcc.dg/ipa/pr103083.c

diff --git a/gcc/ipa-cp.c b/gcc/ipa-cp.c
index 703541d15cc..7e94bd78b98 100644
--- a/gcc/ipa-cp.c
+++ b/gcc/ipa-cp.c
@@ -306,14 +306,15 @@ public:
 class ipcp_bits_lattice
 {
 public:
-  bool bottom_p () { return m_lattice_val == IPA_BITS_VARYING; }
-  bool top_p () { return m_lattice_val == IPA_BITS_UNDEFINED; }
-  bool constant_p () { return m_lattice_val == IPA_BITS_CONSTANT; }
+  bool bottom_p () const { return m_lattice_val == IPA_BITS_VARYING; }
+  bool top_p () const { return m_lattice_val == IPA_BITS_UNDEFINED; }
+  bool constant_p () const { return m_lattice_val == IPA_BITS_CONSTANT; }
   bool set_to_bottom ();
   bool set_to_constant (widest_int, widest_int);
+  bool known_nonzero_p () const;
 
-  widest_int get_value () { return m_value; }
-  widest_int get_mask () { return m_mask; }
+  widest_int get_value () const { return m_value; }
+  widest_int get_mask () const { return m_mask; }
 
   bool meet_with (ipcp_bits_lattice& other, unsigned, signop,
  enum tree_code, tree);
@@ -1081,6 +1082,15 @@ ipcp_bits_lattice::set_to_constant (widest_int value, 
widest_int mask)
   return true;
 }
 
+/* Return true if any of the known bits are non-zero.  */
+bool
+ipcp_bits_lattice::known_nonzero_p () const
+{
+  if (!constant_p ())
+return false;
+  return !wi::eq_p (wi::bit_and (wi::bit_not (m_mask), m_value), 0);
+}
+
 /* Convert operand to value, mask form.  */
 
 void
@@ -2374,6 +2384,7 @@ propagate_bits_across_jump_function (cgraph_edge *cs, int 
idx,
   tree operand = NULL_TREE;
   enum tree_code code;
   unsigned src_idx;
+  bool only_for_nonzero = false;
 
   if (jfunc->type == IPA_JF_PASS_THROUGH)
{
@@ -2386,7 +2397,9 @@ propagate_bits_across_jump_function (cgraph_edge *cs, int 
idx,
{
  code = POINTER_PLUS_EXPR;
  src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
- unsigned HOST_WIDE_INT offset = ipa_get_jf_ancestor_offset (jfunc) / 
BITS_PER_UNIT;
+ unsigned HOST_WIDE_INT offset
+   = ipa_get_jf_ancestor_offset (jfunc) / BITS_PER_UNIT;
+ only_for_nonzero = (ipa_get_jf_ancestor_keep_null (jfunc) || !offset);
  operand = build_int_cstu (size_type_node, offset);
}
 
@@ -2404,16 +2417,18 @@ propagate_bits_across_jump_function (cgraph_edge *cs, 
int