On Thu, Feb 15, 2024 at 10:02 AM Adhemerval Zanella Netto adhemerval.zanella@linaro.org wrote:
On 15/02/24 09:13, H.J. Lu wrote:
On Thu, Feb 15, 2024 at 4:12 AM H.J. Lu hjl.tools@gmail.com wrote:
On Thu, Feb 15, 2024 at 3:49 AM Adhemerval Zanella Netto adhemerval.zanella@linaro.org wrote:
On 15/02/24 08:47, H.J. Lu wrote:
On Thu, Feb 15, 2024 at 12:01 AM Maxim Kuvyrkov maxim.kuvyrkov@linaro.org wrote:
> On Feb 15, 2024, at 03:54, H.J. Lu hjl.tools@gmail.com wrote: > > FAIL: elf/tst-gnu2-tls2 > > indicates that your _dl_tlsdesc_dynamic may not preserve all caller-saved > registers. Please find out how the test fails.
Hi H.J.,
See below.
... > FAIL: 1 regressions > > regressions.sum: > === glibc tests === > > Running glibc:elf ... > FAIL: elf/tst-gnu2-tls2 > > > You can find the failure logs in *.log.1.xz files in > - https://ci.linaro.org/job/tcwg_glibc_check--master-arm-precommit/1460/artifa...
tests.log.1.xz contains output of failed tests -- https://ci.linaro.org/job/tcwg_glibc_check--master-arm-precommit/1460/artifa... === FAIL: elf/tst-gnu2-tls2 original exit status 1 open tst-gnu2-tls2mod0.so open tst-gnu2-tls2mod1.so open tst-gnu2-tls2mod2.so close tst-gnu2-tls2mod0.so close tst-gnu2-tls2mod1.so open tst-gnu2-tls2mod0.so open tst-gnu2-tls2mod1.so Didn't expect signal from child: got `Segmentation fault' ===
Let me know if you need any help investigating this.
I don't have access to aarch64 machine which can build glibc. Please configure glibc with --enable-hardcoded-path-in-tests and run elf/tst-gnu2-tls2 under GDB to find out what is going on.
I will check this out H.J.
Does your _dl_tlsdesc_dynamic save ALL caller-saved registers, except for the return value register? I saw
stp x5, x6, [sp, #16*1] stp x7, x8, [sp, #16*2] stp x9, x10, [sp, #16*3] stp x11, x12, [sp, #16*4] stp x13, x14, [sp, #16*5] stp x15, x16, [sp, #16*6] stp x17, x18, [sp, #16*7]Do you need to save x1 to x4?
-- H.J.
If your processor is also impacted, please add it to
Hi H.J.
The issues is not really on aarch64, but rather arm 32 bits. And it is seems to be a real one on arm _dl_tlsdesc_dynamic implementation, that fail to save/restore r12 that is used by gcc as a scratch register.
I have added 3 more fixes on top on your patches [1]. First one is a small fix for the -mtls-dialect=gnu2 configure test that fail when -mtp=soft is used (used by default arm-linux-gnueabihf cross compiler produced by build-many-glibcs.py).
Second is the arm fix for BZ 31372 regression. However, I am not sure if this suffice, since similar to others ABIs, arm also support vector extensions (VFP, VFP3, and neon). I think we will eventually need to do something similar to what you did for x86 and provided either multiple _dl_tlsdesc_dynamic or handle the vector register save/restore using hwcap feature check (or we can also eventually just remove the slow patch and get over this whole save/restore vector extensions).
This is not an issue now on arm32 because gnu2 is not default and I don't think gcc will flip the switch in near future.
The last patch enables TLS descriptor tests on aarch64 as well, since it uses a different name of gnu2. I think RISC-V will use the same naming as aarch64, so it would make easier to enable such tests on this abi as well.
[1] https://github.com/zatrazz/glibc/commits/azanella/tls-descriptor-fixes-arm/
Great. I added Arm to
https://sourceware.org/bugzilla/show_bug.cgi?id=31372
Thanks.