Re: [musl] [RFC v2 0/2] xtensa FDPIC port

[Rich Felker <dalias@libc.org>]

On Thu, Mar 28, 2024 at 01:03:17PM -0700, Max Filippov wrote:
> Hello,
> 
> this is the second RFC version of the xtensa FDPIC port for musl.
> There are two patches: the first adds the port and will likely not
> change in the future versions. The second adds two bits of
> xtensa-specific GOT initialization for compatibility with the current
> binutils/gcc and will go away once binutils and gcc are changed to do
> local references differently.
> 
> The following binutils and gcc branches can be used to build the
> xtensa-linux-muslfdpic toolchain:
> 
>   https://github.com/jcmvbkbc/binutils-gdb-xtensa xtensa-2.42-fdpic-musl
>   https://github.com/jcmvbkbc/gcc-xtensa xtensa-14-9655-fdpic-musl
> 
> I've tested this version with libc-test using QEMU linux-user and full
> system emulation. The results in the linux-user are the following:
> 
>   FAIL src/functional/dlopen.exe [status 1]
>   FAIL src/functional/pthread_robust-static.exe [timed out]
>   FAIL src/functional/pthread_robust.exe [timed out]
>   FAIL src/functional/strptime-static.exe [status 1]
>   FAIL src/functional/strptime.exe [status 1]
>   FAIL src/math/acoshl.exe [status 1]
>   FAIL src/math/asinhl.exe [status 1]
>   FAIL src/math/erfcl.exe [status 1]
>   FAIL src/math/fma.exe [status 1]
>   FAIL src/math/fmal.exe [status 1]
>   FAIL src/math/lgammal.exe [status 1]
>   FAIL src/math/tgammal.exe [status 1]
>   FAIL src/regression/pthread-robust-detach-static.exe [status 1]
>   FAIL src/regression/pthread-robust-detach.exe [status 1]
> 
> functional/dlopen fails with the
>   src/functional/dlopen.c:39: dlsym main failed: (null)
> There's no failure in the dlsym call, but the pointers don't match.

Is this something related to canonical function descriptors? Is it
musl's fault or a bug in the tooling? I suspect the latter.

> functional/ and regression/ pthread-related failures are expected
> because the robust list functions are not available in the linux-user
> mode. These particular tests pass in full system emulation.
> 
> math tests fail with ULP differences.
> 
> I have also added the following changes to the abi tests to fix the
> build:
> 
> diff --git a/src/api/sys_sem.c b/src/api/sys_sem.c
> index a473cad0a2aa..bd4df9a4fe70 100644
> --- a/src/api/sys_sem.c
> +++ b/src/api/sys_sem.c
> @@ -18,7 +18,11 @@ C(SETALL)
>  {
>  struct semid_ds x;
>  F(struct ipc_perm,sem_perm)
> +#ifdef __xtensa__
> +F(unsigned long, sem_nsems)
> +#else
>  F(unsigned short, sem_nsems)
> +#endif

This is invalid and indicates an error in the port; the whole point of
the tests is to catch that so it can be fixed. POSIX requires
sem_nsems to have type unsigned short. I believe Linux has this wrong
on some (most? all?) archs and we just fix it by replacing the rest
with padding.

Rich