diff mbox series

x86-64: Ignore r_addend for R_X86_64_GLOB_DAT/R_X86_64_JUMP_SLOT

Message ID 20220521022336.2464080-1-hjl.tools@gmail.com
State New
Headers show
Series x86-64: Ignore r_addend for R_X86_64_GLOB_DAT/R_X86_64_JUMP_SLOT | expand

Commit Message

H.J. Lu May 21, 2022, 2:23 a.m. UTC 
According to x86-64 psABI, r_addend should be ignored for R_X86_64_GLOB_DAT
and R_X86_64_JUMP_SLOT.
---
 sysdeps/x86_64/dl-machine.h | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

Comments

Fangrui Song May 22, 2022, 10:10 p.m. UTC | #1 
On 2022-05-20, H.J. Lu via Libc-alpha wrote:
>According to x86-64 psABI, r_addend should be ignored for R_X86_64_GLOB_DAT
>and R_X86_64_JUMP_SLOT.

Looks good. In musl, All GLOB_DAT and JUMP_SLOT ignore the addend.
Ignoring r_addend may or may not be profitable. Since GLOB and JUMP_SLOT
are handled the same way as the symbolic relocation R_X86_64_64, this
may not save code.

>---
> sysdeps/x86_64/dl-machine.h | 6 ++++--
> 1 file changed, 4 insertions(+), 2 deletions(-)
>
>diff --git a/sysdeps/x86_64/dl-machine.h b/sysdeps/x86_64/dl-machine.h
>index c70af7ab1e..7f607f6dff 100644
>--- a/sysdeps/x86_64/dl-machine.h
>+++ b/sysdeps/x86_64/dl-machine.h
>@@ -325,11 +325,13 @@ and creates an unsatisfiable circular dependency.\n",
> #  endif
> 	  /* Set to symbol size plus addend.  */
> 	  value = sym->st_size;
>+	  *reloc_addr = value + reloc->r_addend;
>+	  break;
> # endif
>-	  /* Fall through.  */
>+
> 	case R_X86_64_GLOB_DAT:
> 	case R_X86_64_JUMP_SLOT:
>-	  *reloc_addr = value + reloc->r_addend;
>+	  *reloc_addr = value;
> 	  break;
>
> 	case R_X86_64_DTPMOD64:
>-- 
>2.36.1
>
Florian Weimer May 25, 2022, 4:52 p.m. UTC | #2 
* H. J. Lu via Libc-alpha:

> According to x86-64 psABI, r_addend should be ignored for R_X86_64_GLOB_DAT
> and R_X86_64_JUMP_SLOT.
> ---
>  sysdeps/x86_64/dl-machine.h | 6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
>
> diff --git a/sysdeps/x86_64/dl-machine.h b/sysdeps/x86_64/dl-machine.h
> index c70af7ab1e..7f607f6dff 100644
> --- a/sysdeps/x86_64/dl-machine.h
> +++ b/sysdeps/x86_64/dl-machine.h
> @@ -325,11 +325,13 @@ and creates an unsatisfiable circular dependency.\n",
>  #  endif
>  	  /* Set to symbol size plus addend.  */
>  	  value = sym->st_size;
> +	  *reloc_addr = value + reloc->r_addend;
> +	  break;
>  # endif
> -	  /* Fall through.  */
> +
>  	case R_X86_64_GLOB_DAT:
>  	case R_X86_64_JUMP_SLOT:
> -	  *reloc_addr = value + reloc->r_addend;
> +	  *reloc_addr = value;
>  	  break;
>  
>  	case R_X86_64_DTPMOD64:

Context:

	{
# ifndef RTLD_BOOTSTRAP
#  ifdef __ILP32__
	case R_X86_64_SIZE64:
	  /* Set to symbol size plus addend.  */
	  *(Elf64_Addr *) (uintptr_t) reloc_addr
	    = (Elf64_Addr) sym->st_size + reloc->r_addend;
	  break;

	case R_X86_64_SIZE32:
#  else
	case R_X86_64_SIZE64:
#  endif
	  /* Set to symbol size plus addend.  */
	  value = sym->st_size;
# endif
	  /* Fall through.  */
	case R_X86_64_GLOB_DAT:
	case R_X86_64_JUMP_SLOT:
	  *reloc_addr = value + reloc->r_addend;
	  break;


The new version is:

	{
# ifndef RTLD_BOOTSTRAP
#  ifdef __ILP32__
	case R_X86_64_SIZE64:
	  /* Set to symbol size plus addend.  */
	  *(Elf64_Addr *) (uintptr_t) reloc_addr
	    = (Elf64_Addr) sym->st_size + reloc->r_addend;
	  break;

	case R_X86_64_SIZE32:
#  else
	case R_X86_64_SIZE64:
#  endif
	  /* Set to symbol size plus addend.  */
	  value = sym->st_size;
	  *reloc_addr = value + reloc->r_addend;
	  break;
# endif
	case R_X86_64_GLOB_DAT:
	case R_X86_64_JUMP_SLOT:
	  *reloc_addr = value;
	  break;

So that “break” is actually in the right place.

Are there binaries out there with a non-zero addend?  I'm a bit worried
that this change introduces hard-to-diagnose regressions.

Thanks,
Florian
H.J. Lu May 25, 2022, 5:49 p.m. UTC | #3 
On Wed, May 25, 2022 at 9:52 AM Florian Weimer <fweimer@redhat.com> wrote:
>
> * H. J. Lu via Libc-alpha:
>
> > According to x86-64 psABI, r_addend should be ignored for R_X86_64_GLOB_DAT
> > and R_X86_64_JUMP_SLOT.
> > ---
> >  sysdeps/x86_64/dl-machine.h | 6 ++++--
> >  1 file changed, 4 insertions(+), 2 deletions(-)
> >
> > diff --git a/sysdeps/x86_64/dl-machine.h b/sysdeps/x86_64/dl-machine.h
> > index c70af7ab1e..7f607f6dff 100644
> > --- a/sysdeps/x86_64/dl-machine.h
> > +++ b/sysdeps/x86_64/dl-machine.h
> > @@ -325,11 +325,13 @@ and creates an unsatisfiable circular dependency.\n",
> >  #  endif
> >         /* Set to symbol size plus addend.  */
> >         value = sym->st_size;
> > +       *reloc_addr = value + reloc->r_addend;
> > +       break;
> >  # endif
> > -       /* Fall through.  */
> > +
> >       case R_X86_64_GLOB_DAT:
> >       case R_X86_64_JUMP_SLOT:
> > -       *reloc_addr = value + reloc->r_addend;
> > +       *reloc_addr = value;
> >         break;
> >
> >       case R_X86_64_DTPMOD64:
>
> Context:
>
>         {
> # ifndef RTLD_BOOTSTRAP
> #  ifdef __ILP32__
>         case R_X86_64_SIZE64:
>           /* Set to symbol size plus addend.  */
>           *(Elf64_Addr *) (uintptr_t) reloc_addr
>             = (Elf64_Addr) sym->st_size + reloc->r_addend;
>           break;
>
>         case R_X86_64_SIZE32:
> #  else
>         case R_X86_64_SIZE64:
> #  endif
>           /* Set to symbol size plus addend.  */
>           value = sym->st_size;
> # endif
>           /* Fall through.  */
>         case R_X86_64_GLOB_DAT:
>         case R_X86_64_JUMP_SLOT:
>           *reloc_addr = value + reloc->r_addend;
>           break;
>
>
> The new version is:
>
>         {
> # ifndef RTLD_BOOTSTRAP
> #  ifdef __ILP32__
>         case R_X86_64_SIZE64:
>           /* Set to symbol size plus addend.  */
>           *(Elf64_Addr *) (uintptr_t) reloc_addr
>             = (Elf64_Addr) sym->st_size + reloc->r_addend;
>           break;
>
>         case R_X86_64_SIZE32:
> #  else
>         case R_X86_64_SIZE64:
> #  endif
>           /* Set to symbol size plus addend.  */
>           value = sym->st_size;
>           *reloc_addr = value + reloc->r_addend;
>           break;
> # endif
>         case R_X86_64_GLOB_DAT:
>         case R_X86_64_JUMP_SLOT:
>           *reloc_addr = value;
>           break;
>
> So that “break” is actually in the right place.
>
> Are there binaries out there with a non-zero addend?  I'm a bit worried

Linker always sets r_addend to 0 for R_X86_64_GLOB_DAT and
R_X86_64_JUMP_SLOT.

> that this change introduces hard-to-diagnose regressions.
>
> Thanks,
> Florian
>
Florian Weimer May 25, 2022, 6:19 p.m. UTC | #4 
* H. J. Lu:

> Linker always sets r_addend to 0 for R_X86_64_GLOB_DAT and
> R_X86_64_JUMP_SLOT.

Okay, so it's very unlikely that we see non-zero addends.
Maybe you can add this to the commit message?

Given that, I think the change should be safe.

Thanks,
Florian
Adhemerval Zanella Netto May 25, 2022, 6:28 p.m. UTC | #5 
On 25/05/2022 15:19, Florian Weimer via Libc-alpha wrote:
> * H. J. Lu:
> 
>> Linker always sets r_addend to 0 for R_X86_64_GLOB_DAT and
>> R_X86_64_JUMP_SLOT.
> 
> Okay, so it's very unlikely that we see non-zero addends.
> Maybe you can add this to the commit message?
> 
> Given that, I think the change should be safe.
> 
> Thanks,
> Florian
> 

Maybe assert that r_addend is always zero?
H.J. Lu May 25, 2022, 7:29 p.m. UTC | #6 
On Wed, May 25, 2022 at 11:28 AM Adhemerval Zanella via Libc-alpha
<libc-alpha@sourceware.org> wrote:
>
>
>
> On 25/05/2022 15:19, Florian Weimer via Libc-alpha wrote:
> > * H. J. Lu:
> >
> >> Linker always sets r_addend to 0 for R_X86_64_GLOB_DAT and
> >> R_X86_64_JUMP_SLOT.
> >
> > Okay, so it's very unlikely that we see non-zero addends.
> > Maybe you can add this to the commit message?
> >
> > Given that, I think the change should be safe.
> >
> > Thanks,
> > Florian
> >
>
> Maybe assert that r_addend is always zero?

I don't think it is necessary since r_addend has never been non-zero.
Fangrui Song May 25, 2022, 8:24 p.m. UTC | #7 
On 2022-05-25, H.J. Lu via Libc-alpha wrote:
>On Wed, May 25, 2022 at 11:28 AM Adhemerval Zanella via Libc-alpha
><libc-alpha@sourceware.org> wrote:
>>
>>
>>
>> On 25/05/2022 15:19, Florian Weimer via Libc-alpha wrote:
>> > * H. J. Lu:
>> >
>> >> Linker always sets r_addend to 0 for R_X86_64_GLOB_DAT and
>> >> R_X86_64_JUMP_SLOT.
>> >
>> > Okay, so it's very unlikely that we see non-zero addends.
>> > Maybe you can add this to the commit message?
>> >
>> > Given that, I think the change should be safe.
>> >
>> > Thanks,
>> > Florian
>> >
>>
>> Maybe assert that r_addend is always zero?
>
>I don't think it is necessary since r_addend has never been non-zero.

I agree that an assert is unnecessary. Linkers just guarantee zero
addends.

It seems that by omitting reloc->r_addend, the GCC generated assembler
is slightly shorter. I've tried placing GLOB_DAT/JUMP_SLOT near 64
and don't get simplified assembly.

Reviewed-by: Fangrui Song <maskray@google.com>
H.J. Lu May 26, 2022, 8:59 p.m. UTC | #8 
On Wed, May 25, 2022 at 1:24 PM Fangrui Song <maskray@google.com> wrote:
>
> On 2022-05-25, H.J. Lu via Libc-alpha wrote:
> >On Wed, May 25, 2022 at 11:28 AM Adhemerval Zanella via Libc-alpha
> ><libc-alpha@sourceware.org> wrote:
> >>
> >>
> >>
> >> On 25/05/2022 15:19, Florian Weimer via Libc-alpha wrote:
> >> > * H. J. Lu:
> >> >
> >> >> Linker always sets r_addend to 0 for R_X86_64_GLOB_DAT and
> >> >> R_X86_64_JUMP_SLOT.
> >> >
> >> > Okay, so it's very unlikely that we see non-zero addends.
> >> > Maybe you can add this to the commit message?
> >> >
> >> > Given that, I think the change should be safe.
> >> >
> >> > Thanks,
> >> > Florian
> >> >
> >>
> >> Maybe assert that r_addend is always zero?
> >
> >I don't think it is necessary since r_addend has never been non-zero.
>
> I agree that an assert is unnecessary. Linkers just guarantee zero
> addends.
>
> It seems that by omitting reloc->r_addend, the GCC generated assembler
> is slightly shorter. I've tried placing GLOB_DAT/JUMP_SLOT near 64
> and don't get simplified assembly.
>
> Reviewed-by: Fangrui Song <maskray@google.com>

Here is the patch I am checking in.

Thanks.
Sunil Pandey July 14, 2022, 12:06 a.m. UTC | #9 
On Thu, May 26, 2022 at 2:00 PM H.J. Lu via Libc-alpha
<libc-alpha@sourceware.org> wrote:
>
> On Wed, May 25, 2022 at 1:24 PM Fangrui Song <maskray@google.com> wrote:
> >
> > On 2022-05-25, H.J. Lu via Libc-alpha wrote:
> > >On Wed, May 25, 2022 at 11:28 AM Adhemerval Zanella via Libc-alpha
> > ><libc-alpha@sourceware.org> wrote:
> > >>
> > >>
> > >>
> > >> On 25/05/2022 15:19, Florian Weimer via Libc-alpha wrote:
> > >> > * H. J. Lu:
> > >> >
> > >> >> Linker always sets r_addend to 0 for R_X86_64_GLOB_DAT and
> > >> >> R_X86_64_JUMP_SLOT.
> > >> >
> > >> > Okay, so it's very unlikely that we see non-zero addends.
> > >> > Maybe you can add this to the commit message?
> > >> >
> > >> > Given that, I think the change should be safe.
> > >> >
> > >> > Thanks,
> > >> > Florian
> > >> >
> > >>
> > >> Maybe assert that r_addend is always zero?
> > >
> > >I don't think it is necessary since r_addend has never been non-zero.
> >
> > I agree that an assert is unnecessary. Linkers just guarantee zero
> > addends.
> >
> > It seems that by omitting reloc->r_addend, the GCC generated assembler
> > is slightly shorter. I've tried placing GLOB_DAT/JUMP_SLOT near 64
> > and don't get simplified assembly.
> >
> > Reviewed-by: Fangrui Song <maskray@google.com>
>
> Here is the patch I am checking in.
>
> Thanks.
>
> --
> H.J.

I would like to backport this patch to release branches.
Any comments or objections?

--Sunil
diff mbox series

Patch

diff --git a/sysdeps/x86_64/dl-machine.h b/sysdeps/x86_64/dl-machine.h
index c70af7ab1e..7f607f6dff 100644
--- a/sysdeps/x86_64/dl-machine.h
+++ b/sysdeps/x86_64/dl-machine.h
@@ -325,11 +325,13 @@  and creates an unsatisfiable circular dependency.\n",
 #  endif
 	  /* Set to symbol size plus addend.  */
 	  value = sym->st_size;
+	  *reloc_addr = value + reloc->r_addend;
+	  break;
 # endif
-	  /* Fall through.  */
+
 	case R_X86_64_GLOB_DAT:
 	case R_X86_64_JUMP_SLOT:
-	  *reloc_addr = value + reloc->r_addend;
+	  *reloc_addr = value;
 	  break;
 
 	case R_X86_64_DTPMOD64: