On Tue, 2025-01-14 at 13:14 +0000, Giuliano Procida wrote:
> Tested with pahole v1.24.
>
> The following code, when compiled with either GCC 14 or Clang 16 or
> several other versions, and post processed with pahole -J, results in
> the private definition of foo in bar being exposed as part of the type
> of baz in BTF - which is definitely wrong.
>
> struct foo;
>
> int bar() {
> struct foo {
> int x;
> };
> struct foo j = { 0 };
> struct foo k = j;
> return j.x == k.x;
> }
>
> int baz(struct foo* j) {
> return j != 0;
> }
>
> Note that we had exactly the same bug in STG and I resolved it by
> pushing an anonymous scope before processing the function child DIEs
> (we deal with C++ scopes). You could perhaps push / pop a flag which
> inhibits type definition processing.
>
> It's not possible to just ignore type DIEs inside functions because
> (at least) Clang sometimes emits references outside the function to
> such DIEs.
>
> Regards,
> Giuliano.
>
Hi Giuliano,
Thank you for the report, indeed, the BTF generated for your example is:
[1] STRUCT 'foo' size=4 vlen=1
'x' type_id=2 bits_offset=0
[2] INT 'int' size=4 bits_offset=0 nr_bits=32 encoding=SIGNED
[3] PTR '(anon)' type_id=1
[4] FUNC_PROTO '(anon)' ret_type_id=2 vlen=0
[5] FUNC 'bar' type_id=4 linkage=static
[6] FUNC_PROTO '(anon)' ret_type_id=2 vlen=1
'j' type_id=3 // <-------------- this is wrong
[7] FUNC 'baz' type_id=6 linkage=static
Thanks,
Eduard
