Re: [PATCH] Introduce ELF_INITFINI for all architectures

[Florian Weimer <fweimer at redhat dot com>]

On 08/08/2018 09:49 PM, Palmer Dabbelt wrote:
> On Wed, 08 Aug 2018 07:42:27 PDT (-0700), fweimer@redhat.com wrote:
> > On 06/23/2018 11:45 PM, Florian Weimer wrote:
> > > 2018-06-23  Florian Weimer<fweimer@redhat.com>
> > >
> > >     Introduce ELF_INITFINI, set everywhere except on RISC-V.
> > >     * sysdeps/init_array/crti.S: Move to ...
> > >     * sysdepes/generic/crti.S: here.  Report an error if ELF_INITFINI.
> > >     * sysdeps/init_array/crtn.S: Move to ...
> > >     * sysdeps/generic/crtn.S: here.  Report an error if ELF_INITFINI.
> > >     * csu/elf-init.c: Check ELF_INITFINI instead of NO_INITFINI.
> > >     * gmon/gmon-start.c [!ELF_INITFINI] (GMON_START_ARRAY_SECTION):
> > >     Define.
> > >     * elf/dl-fini.c (_dl_fini): Check for ELF_INITFINI before using
> > >     DT_FINI.
> > >     * elf/dl-init.c (call_init): Check for ELF_INITFINI before using
> > >     DT_INIT.
> > >     * nptl/pt-crti.S [ELF_INITFINI]: Use .init_array instead of
> > >     PREINIT_FUNCTION.
> > >     * sysdeps/generic/platform-params.h: New file.
> > >     * sysdeps/init_array/elf-init.c: Remove file.
> > >     * sysdeps/init_array/gmon-start.c: Likewise.
> > >     * sysdeps/init_array/pt-crti.S: Likewise.
> > >     * sysdeps/riscv/Implies (init_array): Remove.
> > >     * sysdeps/riscv/platform-params.h: New file.
> >
> > Ping?
> >
> >    <https://sourceware.org/ml/libc-alpha/2018-07/msg00051.html>
>
> I'm happy with the RISC-V aspects, which I understand to be functionally 
> a NOP. I like avoiding the ABI bloat.
>
> For some reason I thought someone mentioned some references to 
> _init/_fini floating around our port, but I can't find any so maybe I'm 
> just going crazy? If they're there then we should probably remove them, 
> as they don't actually do anything.

The patch removes the DT_INIT/DT_FINI execution logic from the RISC-V 
port because ELF_INITFINI && l->l_info[DT_FINI] != NULL and the similar 
condition for DT_INIT are always false.  Before the patch, the code is 
still there, but dormant because the GNU toolchain will always generate 
NULL pointers there.

Thanks,
Florian