Re: [12/16][binutils][AARCH64]Add relocation support for large memory model. [BFD]Create GOT section for TLSLE_MOVW_TPREL_G(1, 1_NC, 2).

[Jiong Wang <jiong dot wang at foss dot arm dot com>]

On 04/12/15 09:25, Christophe Lyon wrote:
> On 3 December 2015 at 16:19, Jiong Wang <jiong.wang@foss.arm.com> wrote:
> > On 03/12/15 14:57, Jiong Wang wrote:
> > > On 03/12/15 14:50, Renlin Li wrote:
> > > > On 03/12/15 14:23, Jiong Wang wrote:
> > > > > On 08/09/15 19:00, Renlin Li wrote:
> > > > > > Hi all,
> > > > > >
> > > > > > This patch fixes a bug, got section is not generated for
> > > > > > TLSLE_MOVW_TPREL_G1,TLSLE_MOVW_TPREL_G1_NC, TLSLE_MOVW_TPREL_G2 previously.
> > > > > >
> > > > > > LD regression test Okay. Okay to commit?
> > > > > >
> > > > > > Regards,
> > > > > > Renlin
> > > > > >
> > > > > > bfd/ChangeLog:
> > > > > >
> > > > > > 2015-09-08  Renlin Li <renlin.li@arm.com>
> > > > > >
> > > > > >      * elfnn-aarch64.c (elfNN_aarch64_check_relocs): Create GOT section
> > > > > >          for TLSLE_MOVW_TPREL_G(1, 1_NC, 2) relocation.
> > > > >
> > > > > This looks wrong to me, and will cause inconsistent inside AArch64
> > > > > backend.
> > > > >
> > > > > All TLS Local Executable relocation types don't need any GOT entry, we
> > > > > shouldn't listed them here.
> > > >
> > > > Yes, you are right. I am wrong there.
> > > >
> > > > > Looks like this patch was trying to fix the following two testcase
> > > > > failures:
> > > > >
> > > > > Running
> > > > > /work/GNU-TOOLCHAIN/binutils-gdb/ld/testsuite/ld-aarch64/aarch64-elf.exp ...
> > > > > FAIL: ld-aarch64/tls-relax-large-gd-le
> > > > > FAIL: ld-aarch64/tls-relax-large-desc-le
> > > > >
> > > > > The correct fix is to add "-shared" to the test option, then the symbol
> > > > > _GLOBAL_OFFSET_TABLE_ will be defined, always.
> > > >
> > > > -shared could make the test cases pass, but is it required? The test case
> > > > itself should be legal even without "-shared".
> > >
> > > For that two particular testcases, the most important thing is to make
> > > sure the tls relaxations are tested, either "-shared" or not doesn't matter.
> >
> > Oh, for these GD->LE testcases, shouldn't be shared, otherwise the
> > relaxation will not happen.
> I guess I'm missing something obvious: these 2 testcases PASS for me
> (target aarch64-linux-gnu).
> In which configuration do you see them failing?

I mean they will fail after we revert this patch.

This patch was trying to fix the failing in an incorrect way.

> > > As you have referenced _GLOBAL_OFFSET_TABLE_, you need to specify -shared,
> > > or you can just remove the reference, they are not related what's tested.
> > >
> > > > Regards,
> > > > Renlin Li