Hi again,
On Wed, Oct 14 2026, Martin Jambor wrote:
> Hi,
>
> On Mon, Sep 21 2026, Josef Melcr wrote:
>> The jump function logic for regular callback edges does not apply to
>> internal callback edges, as it relies on the existence of a
>> callback-carrying edge for the remapping. In order to make ipa-cp act
>> on assume bodies, we need to remap the arguments before the analysis is
>> done. This patch adds the remapping logic and also adds an early return
>> to ipa_polymorphic_call_context::get_dynamic_type, which tries to get
>> the called function from the call statement with gimple_call_fn.
>> This of course doesn't work for internal calls.
>>
>> gcc/ChangeLog:
>>
>> * ipa-polymorphic-call.cc
>> (ipa_polymorphic_call_context::get_dynamic_type): Exclude
>> IFN_ASSUME internal calls.
>> * ipa-prop.cc
>> (ipa_compute_jump_functions_for_edge): Remap arguments for
>> callback edges.
>>
>> Signed-off-by: Josef Melcr <[email protected]>
>> ---
>> gcc/ipa-polymorphic-call.cc | 7 +++++++
>> gcc/ipa-prop.cc | 14 ++++++++++++--
>> 2 files changed, 19 insertions(+), 2 deletions(-)
>>
>> diff --git a/gcc/ipa-polymorphic-call.cc b/gcc/ipa-polymorphic-call.cc
>> index addb52561c7..12d9c9f878a 100644
>> --- a/gcc/ipa-polymorphic-call.cc
>> +++ b/gcc/ipa-polymorphic-call.cc
>> @@ -41,6 +41,8 @@ along with GCC; see the file COPYING3. If not see
>> #include "alloc-pool.h"
>> #include "symbol-summary.h"
>> #include "symtab-thunks.h"
>> +#include "attr-callback.h"
>> +#include "callback-info.h"
>>
>> /* Return true when TYPE contains an polymorphic type and thus is
>> interesting
>> for devirtualization machinery. */
>> @@ -1584,6 +1586,11 @@ ipa_polymorphic_call_context::get_dynamic_type (tree
>> instance,
>> if (!instance)
>> return false;
>>
>> + /* Bail to avoid segfaults for IFN_ASSUME internal calls. They are not
>> + polymorphic anyway. */
>> + if (callback_ifn_assume_call_p (call))
>> + return false;
>
> This will not return if call is an assume using an outlined function
> that has also somehow has a call which is not callback (because of
> ICF?). callback_ifn_assume_call_p seems to be written with some other
> use in mind and perhaps needs to be split into two slightly different
> predicates?
>
Ah, scratch that, I mixed up callback_ifn_assume_call_p and
callback_ifn_assume_call_with_edge_p, the above is technically OK, of
course.
But now I think the check should be probably open-coded as
if (is_gimple_call (stmt)
&& gimple_call_internal_p (stmt, IFN_ASSUME))
so that readers don't need to wonder about callbacks to comprehend
devirtualization code.
Martin
>> +
>> if (otr_type)
>> otr_type = TYPE_MAIN_VARIANT (otr_type);
>>
>> diff --git a/gcc/ipa-prop.cc b/gcc/ipa-prop.cc
>> index 3f9f45b02c8..bc31e8e3399 100644
>> --- a/gcc/ipa-prop.cc
>> +++ b/gcc/ipa-prop.cc
>> @@ -2492,7 +2492,8 @@ ipa_compute_jump_functions_for_edge (struct
>> ipa_func_body_info *fbi,
>> ipa_node_params *info = ipa_node_params_sum->get (cs->caller);
>> ipa_edge_args *args = ipa_edge_args_sum->get_create (cs);
>> gcall *call = cs->call_stmt;
>> - int n, arg_num = gimple_call_num_args (call);
>> + callback_info *ci = callback_info_sum->get (cs);
>> + int n, arg_num = callback_call_num_args_summary (call, ci);
>> bool useful_context = false;
>>
>> if (arg_num == 0 || args->jump_functions)
>> @@ -2508,7 +2509,16 @@ ipa_compute_jump_functions_for_edge (struct
>> ipa_func_body_info *fbi,
>> for (n = 0; n < arg_num; n++)
>> {
>> struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, n);
>> - tree arg = gimple_call_arg (call, n);
>> + tree arg;
>> + if (cs->callback)
>> + {
>> + int remapped_idx = ci->arg_mapping[n];
>> + if (remapped_idx == ARG_MAPPING_UNKNOWN_IDX)
>> + continue;
>> + arg = gimple_call_arg (call, remapped_idx);
>> + }
>> + else
>> + arg = gimple_call_arg (call, n);
>
> This feels a bit strange because before this patch series, the callback
> edges were created in this function and their jump functions populated
> differently by copying over what we calculated for the call to the
> dispatcher.
>
> I understand that callback edges for assumes have to be created
> elsewhere, can we add an assert to the if block above so that readers of
> the code know this is why it is that we handle things differently?
>
> Thanks,
>
> Martin
>
>
>
>> tree param_type = ipa_get_callee_param_type (cs, n);
>> if (flag_devirtualize && POINTER_TYPE_P (TREE_TYPE (arg)))
>> {
>> --
>> 2.55.0
