[PATCH v4] i386: Update ___tls_get_addr to preserve vector registers

[H.J. Lu]

On Tue, Jun 17, 2025 at 10:04 PM Adhemerval Zanella Netto
<adhemerval.zanella@linaro.org> wrote:
>
>
>
> On 17/06/25 09:52, Adhemerval Zanella Netto wrote:
> >
> >
> > On 17/06/25 09:33, H.J. Lu wrote:
> >> On Tue, Jun 17, 2025 at 7:46 PM Florian Weimer <fweimer@redhat.com> wrote:
> >>>
> >>> * H. J. Lu:
> >>>
> >>>> diff --git a/elf/tst-tls23.c b/elf/tst-tls23.c
> >>>> new file mode 100644
> >>>> index 0000000000..afe594c067
> >>>> --- /dev/null
> >>>> +++ b/elf/tst-tls23.c
> >>>> @@ -0,0 +1,106 @@
> >>>> +/* Test that __tls_get_addr preserves caller-saved registers.
> >>>> +   Copyright (C) 2025 Free Software Foundation, Inc.
> >>>> +   This file is part of the GNU C Library.
> >>>> +
> >>>> +   The GNU C Library is free software; you can redistribute it and/or
> >>>> +   modify it under the terms of the GNU Lesser General Public
> >>>> +   License as published by the Free Software Foundation; either
> >>>> +   version 2.1 of the License, or (at your option) any later version.
> >>>> +
> >>>> +   The GNU C Library is distributed in the hope that it will be useful,
> >>>> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> >>>> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> >>>> +   Lesser General Public License for more details.
> >>>> +
> >>>> +   You should have received a copy of the GNU Lesser General Public
> >>>> +   License along with the GNU C Library; if not, see
> >>>> +   <http://www.gnu.org/licenses/>.  */
> >>>> +
> >>>> +#include <stdio.h>
> >>>> +#include <stdlib.h>
> >>>> +#include <string.h>
> >>>> +#include <dlfcn.h>
> >>>> +#include <pthread.h>
> >>>> +#include <support/xdlfcn.h>
> >>>> +#include <support/xthread.h>
> >>>> +#include <support/check.h>
> >>>> +#include <support/test-driver.h>
> >>>> +#include <tst-tls23.h>
> >>>> +
> >>>> +#ifndef IS_SUPPORTED
> >>>> +# define IS_SUPPORTED() true
> >>>> +#endif
> >>>> +
> >>>> +/* An architecture can define it to clobber caller-saved registers in
> >>>> +   malloc below to verify that __tls_get_addr won't change caller-saved
> >>>> +   registers.  */
> >>>> +#ifndef PREPARE_MALLOC
> >>>> +# define PREPARE_MALLOC()
> >>>> +#endif
> >>>> +
> >>>> +extern void * __libc_malloc (size_t);
> >>>> +
> >>>> +size_t malloc_counter = 0;
> >>>> +
> >>>> +void *
> >>>> +malloc (size_t n)
> >>>> +{
> >>>> +  PREPARE_MALLOC ();
> >>>> +  malloc_counter++;
> >>>> +  return __libc_malloc (n);
> >>>> +}
> >>>> +
> >>>> +static void *mod;
> >>>> +static const char *modname = "tst-tls23-mod.so";
> >>>> +
> >>>> +static void
> >>>> +open_mod (void)
> >>>> +{
> >>>> +  mod = xdlopen (modname, RTLD_LAZY);
> >>>> +  printf ("open %s\n", modname);
> >>>> +}
> >>>> +
> >>>> +static void
> >>>> +close_mod (void)
> >>>> +{
> >>>> +  xdlclose (mod);
> >>>> +  mod = NULL;
> >>>> +  printf ("close %s\n", modname);
> >>>> +}
> >>>> +
> >>>> +static void
> >>>> +access_mod (const char *sym)
> >>>> +{
> >>>> +  struct tls var = { -4, -4, -4, -4 };
> >>>> +  struct tls *(*f) (struct tls *) = xdlsym (mod, sym);
> >>>> +  /* Check that our malloc is called.  */
> >>>> +  malloc_counter = 0;
> >>>> +  struct tls *p = f (&var);
> >>>> +  TEST_VERIFY (malloc_counter != 0);
> >>>> +  printf ("access %s: %s() = %p\n", modname, sym, p);
> >>>> +  TEST_VERIFY_EXIT (memcmp (p, &var, sizeof (var)) == 0);
> >>>> +  ++(p->a);
> >>>> +}
> >>>> +
> >>>> +static void *
> >>>> +start (void *arg)
> >>>> +{
> >>>> +  access_mod ("apply_tls");
> >>>> +  return arg;
> >>>> +}
> >>>> +
> >>>> +static int
> >>>> +do_test (void)
> >>>> +{
> >>>> +  if (!IS_SUPPORTED ())
> >>>> +    return EXIT_UNSUPPORTED;
> >>>> +
> >>>> +  open_mod ();
> >>>> +  pthread_t t = xpthread_create (NULL, start, NULL);
> >>>> +  xpthread_join (t);
> >>>> +  close_mod ();
> >>>> +
> >>>> +  return 0;
> >>>> +}
> >>>> +
> >>>> +#include <support/test-driver.c>
> >>>
> >>> This test fails on ppc64le because malloc_counter is not increased:
> >>>
> >>> =====FAIL: elf/tst-tls23.out=====
> >>> open tst-tls23-mod.so
> >>> error: tst-tls23.c:79: not true: malloc_counter != 0
> >>> access tst-tls23-mod.so: apply_tls() = 0x7fff9f7ff980
> >>> close tst-tls23-mod.so
> >>> error: 1 test failures
> >>>
> >>> This may be because of the POWER-specific __tls_get_addr_opt
> >>> optimization.
> >>>
> >>> Increasing the size of TLS variables may force an allocation ppc64le as
> >>> well.
> >>
> >> Can you find out what size should be?
> >
> > I am checking on some different ABI to see if everything is ok and get a
> > proper number for powerpc.
> >
>
> I think instead of resorting to specific TLS variable size (which might
> need to adjust if/when we increase the default static TLS size), it would
> be better to just disable the optimization (since the idea is to always
> call __tls_get_addr):
>
> diff --git a/sysdeps/powerpc/Makefile b/sysdeps/powerpc/Makefile
> index 5e6cb07ce6..88b5c961fc 100644
> --- a/sysdeps/powerpc/Makefile
> +++ b/sysdeps/powerpc/Makefile
> @@ -28,6 +28,11 @@ tst-cache-ppc-static-dlopen-ENV = LD_LIBRARY_PATH=$(objpfx):$(common-objpfx):$(c
>  $(objpfx)tst-cache-ppc-static-dlopen.out: $(objpfx)mod-cache-ppc.so
>
>  $(objpfx)tst-cache-ppc: $(objpfx)mod-cache-ppc.so
> +
> +# The test checks if the __tls_get_addr does not clobber caller-saved
> +# register, so disable the powerpc specific optimization to force a
> +# __tls_get_addr call.
> +LDFLAGS-tst-tls23-mod.so = -Wl,--no-tls-get-addr-optimize
>  endif
>
>  ifneq (no,$(multi-arch))

The v5 patch was sent with this sysdeps/powerpc/Makefile change.

-- 
H.J.