Re: size_t :0 doesn't always work with llvm-16

[Kui-Feng Lee <sinquersw@gmail.com>]

On 10/28/23 18:12, Andrii Nakryiko wrote:
> On Sat, Oct 28, 2023 at 5:52 PM Kui-Feng Lee <sinquersw@gmail.com> wrote:
>>
>> Recently, while running the test_maps, I encountered an error
>> message. Upon further investigation, I discovered that llvm-16 behaves
>> inconsistently when it comes to clearing partially initialized local
>> variables.
>>
>> We appends a 'size_t :0' at the end of many types to prevent dirty
>> bytes at the end of struct types.  libbpf also check dirty padding
>> bytes for CORE. It works most of time with gcc and llvm.  However, I
>> have discovered that it is not always work with llvm. The C
>> specification does not guarantee this either. Nonetheless, we
>> primarily rely on gcc and llvm. In most cases, on x86_64 platforms,
>> llvm utilizes memset() to clear a partially initialized variable of a
>> struct type.
>>
> 
> Yes, which is why using LIBBPF_OPTS() is a good idea. And which is why
> I submitted [0] to fix this in our test_maps tests. I'll resend this
> fix outside of BPF token series.
> 
>    [0] https://patchwork.kernel.org/project/netdevbpf/patch/20231016180220.3866105-14-andrii@kernel.org/

Great to hear that!
Thanks!