This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] elf: Add tests with a local IFUNC resolver [BZ #23937]
On 20/02/2019 13:29, Florian Weimer wrote:
> * H. J. Lu:
>
>> On Wed, Feb 20, 2019 at 4:53 AM Florian Weimer <fweimer@redhat.com> wrote:
>>>
>>> The existing tests all use global symbols (but with different
>>> visibility). Local symbols could be treated differently by the
>>> compiler and linker (as was the case on POWER ELFv2, causing
>>> bug 23937), and we did not have test coverage for this.
>>>
>>> Tested on x86-64 and POWER ELFv2 little-endian, with and without
>>> --disable-multi-arch. On POWER, the test cases elf/ifuncmain9,
>>> elf/ifuncmain9pic, elf/ifuncmain9pie reproduce bug 23937 with older
>>> binutils.
>>>
>>> 2019-02-20 Florian Weimer <fweimer@redhat.com>
>>>
>>> [BZ #23937]
>>> elf: Add test with a local IFUNC resolver.
>>> * elf/ifuncmain9.c: New file.
>>> * elf/ifuncmain9pic.c: Likewise.
>>> * elf/ifuncmain9picstatic.c: Likewise.
>>> * elf/ifuncmain9pie.c: Likewise.
>>> * elf/ifuncmain9static.c: Likewise.
>>> * elf/Makefile [multi-arch] (tests-ifuncstatic): Add
>>> ifuncmain9static, ifuncmain9picstatic.
>>> * elf/Makefile [multi-arch && build-shared] (tests-internal):
>>> Add ifuncmain9, ifuncmain9pic.
>>> * elf/Makefile [multi-arch && build-shared && have-fpie]
>>> (ifunc-pie-tests): Add ifuncmain9pie.
>>> (CFLAGS-ifuncmain9pic.c): Add $(pic-ccflag).
>>> (CFLAGS-ifuncmain9picstatic.c): Likewise.
>>> (CFLAGS-ifuncmain9pie.c): Add $(pie-ccflag).
>>
>> It isn't caused by your patch. I am wondering why we don't run IFUNC
>> tests with --disable-multi-arch. Multi-arch requires IFUNC. Does IFUNC
>> require multi-arch?
>
> Hah. I wondered the same thing. Yes, we can run IFUNC tests with
> --disable-multi-arch if the toolchain supports IFUNCs. For correctness,
> --disable-multi-arch must not remove IFUNC support from the loader.
>
> I do not know what the exact rules for IFUNC support in the loader are.
> Logically, support cannot depend on whether the toolchain supports
> IFUNCs because it's a property of the ABI. Calling IFUNCs in the loader
> should not require toolchain support, only internal use in glibc and
> tests need it.
gcc rejects __attribute__((ifunc( ))) if the compiler
was not configured with --enable-gnu-indirect-function
(which i think is not default on some targets e.g. arm)
(i think gcc should just handle the attribute and
emit the gas directive if there is gas support anyway,
and the configure flag should only affect target libs
in gcc that use ifunc like libatomic)