Re: [PATCH v2 1/1] gdb: Preserve IFUNC marker when finding inferior functions

[Muhammad Kamran <muhammad.kamran@arm.com>]

Hi Simon,

On 26/06/2026 15:20, Simon Marchi wrote:
> 
> 
> On 2026-06-26 05:24, Muhammad Kamran wrote:
>> Hi Simon,
>>
>> On 25/06/2026 21:05, Simon Marchi wrote:
>>>
>>>> @@ -356,6 +361,39 @@ proc misc_tests {resolver_attr resolver_debug final_debug} {
>>>>        }
>>>>    }
>>>>    +# Test that GDB resolves a GNU IFUNC minimal symbol when it uses
>>>> +# find_function_in_inferior to make an internal inferior call.  String
>>>> +# literals are copied into the inferior with a call to malloc, so a
>>>> +# no-debug IFUNC malloc exercises the minimal-symbol fallback.
>>>> +
>>>> +proc_with_prefix test_inferior_call {} {
>>>> +    global srcdir subdir
>>>> +    global infcall_file infcall_src
>>>> +    global infcall_malloc_file infcall_malloc_src
>>>> +
>>>> +    set executable $infcall_file
>>>> +    set binfile [standard_output_file $executable]
>>>> +    set malloc_obj [standard_output_file ${infcall_malloc_file}.o]
>>>> +
>>>> +    if { [gdb_compile ${srcdir}/${subdir}/${infcall_malloc_src} \
>>>> +          $malloc_obj object {}] != ""
>>>> +     || [gdb_compile [list ${srcdir}/${subdir}/${infcall_src} \
>>>> +                  $malloc_obj] \
>>>> +         $binfile executable {debug}] != "" } {
>>>> +    untested "failed to compile inferior call testcase"
>>>> +    return
>>>> +    }
>>>
>>> I think that the test should also cover the case where we do have debug
>>> info.  However, the gdb.base/gnu-ifunc.exp test is already written in
>>> a such a way that it tests all imaginable combinations:
>>>
>>> # Test all the combinations of:
>>> #
>>> # - An ifunc resolver with the same name as the ifunc symbol vs an
>>> #   ifunc resolver with a different name as the ifunc symbol.
>>> #
>>> # - ifunc resolver compiled with and without debug info.  This ensures
>>> #   that GDB understands that a function not a regular function by
>>> #   looking at the STT_GNU_IFUNC type in the elf symbols.  DWARF has
>>> #   no way to express the STT_GNU_IFUNC type.
>>> #
>>> # - ifunc target function (resolved) compiled with and without debug
>>> #   info.
>>> foreach_with_prefix resolver_attr {0 1} {
>>>       foreach_with_prefix resolver_debug {0 1} {
>>>      foreach_with_prefix final_debug {0 1} {
>>>          if { [build $resolver_attr $resolver_debug $final_debug] != 0 } {
>>>          misc_tests $resolver_attr $resolver_debug $final_debug
>>>          set-break $resolver_attr $resolver_debug $final_debug
>>>          }
>>>      }
>>>       }
>>> }
>>>
>>> Could you somehow hook the new infcall tests into that, so that we also
>>> test infcalls in all imaginable situations.
>>>
>>
>> I've extended the test to run through the existing
>> resolver_attr/resolver_debug/final_debug matrix.  With only the test
>> change, the new inferior-call test fails in all eight combinations on
>> AArch64, so the issue is wider than the original no-debug minimal-symbol
>> case.  I'll work on the patch accordingly and post a new version once I
>> have the patch ready.
>>
>>>> diff --git a/gdb/valops.c b/gdb/valops.c
>>>> index ab6fd5079e1..7d305871efc 100644
>>>> --- a/gdb/valops.c
>>>> +++ b/gdb/valops.c
>>>> @@ -133,11 +133,15 @@ find_function_in_inferior (const char *name, struct objfile **objf_p)
>>>>          struct gdbarch *gdbarch = objfile->arch ();
>>>>            struct type *type;
>>>> +      struct type *resolved_type;
>>>>          CORE_ADDR maddr;
>>>>          type = lookup_pointer_type (builtin_type (gdbarch)->builtin_char);
>>>>          type = lookup_function_type (type);
>>>>          type = lookup_pointer_type (type);
>>>> -      maddr = msymbol.value_address ();
>>>> +      resolved_type = find_minsym_type_and_address (msymbol.minsym, objfile,
>>>> +                            &maddr);
>>>> +      if (resolved_type->is_gnu_ifunc ())
>>>> +        type->target_type ()->set_is_gnu_ifunc (true);
>>>
>>> Calling find_minsym_type_and_address just to know if the minsym is an
>>> ifunc seems rather heavyweight for nothing.  Can't we just check that
>>> the minsym type is mst_text_gnu_ifunc?
>>>

I posted a v3 using the direct minimal-symbol type check you suggested.
That is better here because find_function_in_inferior only needs to
preserve the IFUNC marker on the synthetic function type; it does not
need to redo minimal-symbol address/type resolution.

The descriptor case is already handled later in the normal inferior-call
path.  find_function_in_inferior creates a pointer-to-function value,
and find_function_addr handles such values by calling
gdbarch_convert_from_func_ptr_addr before IFUNC resolution.

Thanks,
Kamran

>>
>> I used find_minsym_type_and_address because it keeps
>> find_function_in_inferior consistent with normal minimal-symbol
>> evaluation.  It also handles function-descriptor symbols:
>> mst_data_gnu_ifunc may have a descriptor address as its value, and
>> find_minsym_type_and_address can convert that to the code address and
>> reclassify it as mst_text_gnu_ifunc.  A direct mst_text_gnu_ifunc check
>> would miss that case, and value_address () could be the descriptor
>> address rather than the callable address.
>> For reference: gdb/minsyms.c:1594 has:
>>    /* The minimal symbol might point to a function descriptor;
>>       resolve it to the actual code address instead.  */
>>
>> If you still prefer avoiding find_minsym_type_and_address here, I think
>> we would need a helper that shares the same descriptor/address handling,
>> rather than checking only mst_text_gnu_ifunc.
> 
> If the descriptor to actual address translation is really needed, then
> find_minsym_type_and_address sounds fine.
> 



> Simon