On Tue, Feb 18, 2025 at 09:58:44AM +0100, Paolo Bonzini wrote:
> Date: Tue, 18 Feb 2025 09:58:44 +0100
> From: Paolo Bonzini <pbonz...@redhat.com>
> Subject: Re: [PATCH] rust: Prefer link_with over link_whole
>
> On 2/18/25 09:37, Zhao Liu wrote:
> > The commit fccb744f41c6 ("gdbstub: Try unlinking the unix socket before
> > binding") causes the compilation of rust-qemu-api-integration to fail,
> > because rust-qemu-api-integration uses link_whole which meets the
> > duplicate symbol linker error.
> >
> > Though it's not the issue of link_whole used by Rust side, there's no
> > need to use link_whole.
> >
> > Use link_with, which may also bring some benefits, such as faster
> > linking or smaller output files.
>
> link_with, if I remember correctly, drops the constructors. Using it in
> rust_devices_ss therefore prevents the devices from being registered with
> QOM.
Interesting, I test with link_with and HPET QOM type can be registered.
I find this is because `module_init` macro adds `#[used]` for `LOAD_MODULE`,
which prevents linker dropping QOM type.
Without `#[used]`, both link_with and link_whole will optimize
`LOAD_MODULE`... So here `#[used]` is the key to getting everything to work
properly.
> The real issue is lack of support for "objects" in Meson for Rust
> executables. A patch to fix that is under review, after which it will be
> possible to build Rust executables using the same "dependency" objects as C
> code.
HMM, but with link_with, everything seems to compile fine?
Thanks,
Zhao
