https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102586
--- Comment #10 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
So, on
struct C0 {};
struct C1 {};
struct C2 : C1, virtual C0 {};
struct C3 : virtual C2, C1 { virtual int foo () { return 1; } };
struct C6 { char c; };
struct C7 : virtual C6, virtual C3, C1 { virtual int foo () { return 4; } };
struct C8 : C7 { virtual int foo () { return 5; } };
__attribute__((noipa)) void
foo (C8 *q)
{
__builtin_clear_padding (q);
}
C8 c8;
C7 c7;
void
bar ()
{
c8.c = 42;
c7.c = 43;
}
distilled from the testcase in the patch, c8.c is at 9 bytes into c8, while
c7.c is at 8 bytes into c7, so that is likely the reason why the code considers
both byte at offset 8 and 9 as non-padding - the first FIELD_DECL in C8 is
one with C7 type (32 byte long like C8) but the field size is 9 bytes, so that
the __builtin_clear_padding code thinks the first 9 bytes are non-padding,
virtual table pointer inside of C2, then empty type C1 at off 8, then C6 at off
8 with the c 1-byte field, then C3 at offset 16 with 9 byte size but all this
is already outside of the size of FIELD_DECL with C7 type.
But C8 has a FIELD_DECL with C6 type at off 9...
