Re: PATCH: Optimize protected call for i386

Re: PATCH: Optimize protected call for i386

[Alan Modra]

On Wed, May 14, 2003 at 07:49:40AM -0700, H. J. Lu wrote:
> On Wed, May 14, 2003 at 06:24:53PM +0930, Alan Modra wrote:
> > On Tue, May 13, 2003 at 02:27:16PM -0700, H. J. Lu wrote:
> > > How about this patch? It added a new relocation type R_386_PC32_DATA
> > > to distinguish PC relative data reference from PC relative branch.
> > 
> > I think it's going to a lot of work just to support non-PIC shared
> > libs.  I don't think it's worth bothering really.
> 
> The current gcc won't use PIC for calls to protected functions. It
> expects linker will resolve it directly.

I see.  I think this is wrong.  gcc needs to emit the normal PIC call,
ie. "call foo@PLT" so the the linker can set up a plt entry if
needed.  A plt entry is needed for protected symbols so that function
pointer comparisons work, since the x86 ABI uses the .plt entry for
the value of a function symbol referenced by a dynamic object.
Consider liba.so exporting a protected symbol foo.  Will a function
pointer, value foo, passed to libb.so compare equal to &foo in
libb.so?  What about a function pointer, value foo, passed from
libb.so back to liba.so?

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre

Re: PATCH: Optimize protected call for i386

[Richard Henderson]

On Thu, May 15, 2003 at 08:32:53AM +0930, Alan Modra wrote:
> I see.  I think this is wrong.  gcc needs to emit the normal PIC call,
> ie. "call foo@PLT" so the the linker can set up a plt entry if
> needed.

I disagree.

> A plt entry is needed for protected symbols so that function
> pointer comparisons work, since the x86 ABI uses the .plt entry for
> the value of a function symbol referenced by a dynamic object.
> Consider liba.so exporting a protected symbol foo.  Will a function
> pointer, value foo, passed to libb.so compare equal to &foo in
> libb.so?  What about a function pointer, value foo, passed from
> libb.so back to liba.so?

None of which has anything to do with call instructions.

If liba.so contains a call to foo, ld *knows* that foo *must*
resolve to the version in liba.so and nowhere else.  Thus we
can direct the call to foo without going through a PLT entry.

If you take the address of foo, in *any* DSO, you will not be
using either R_386_PC32 or R_386_PLT32.  You'll be using either
R_386_32 or R_386_GOT32.


r~

Re: PATCH: Optimize protected call for i386

[H. J. Lu]

On Wed, May 14, 2003 at 04:24:39PM -0700, Richard Henderson wrote:
> On Thu, May 15, 2003 at 08:32:53AM +0930, Alan Modra wrote:
> > I see.  I think this is wrong.  gcc needs to emit the normal PIC call,
> > ie. "call foo@PLT" so the the linker can set up a plt entry if
> > needed.
> 
> I disagree.

Since "call foo" and ".long foo - ." will both generate a R_386_PC32
relocation against foo, linker can't resolve R_386_PC32 against
a protected symbol without knowing how it is used. "call foo@PLT"
solves this problem for linker.

> 
> > A plt entry is needed for protected symbols so that function
> > pointer comparisons work, since the x86 ABI uses the .plt entry for
> > the value of a function symbol referenced by a dynamic object.
> > Consider liba.so exporting a protected symbol foo.  Will a function
> > pointer, value foo, passed to libb.so compare equal to &foo in
> > libb.so?  What about a function pointer, value foo, passed from
> > libb.so back to liba.so?
> 
> None of which has anything to do with call instructions.
> 
> If liba.so contains a call to foo, ld *knows* that foo *must*
> resolve to the version in liba.so and nowhere else.  Thus we
> can direct the call to foo without going through a PLT entry.
> 
> If you take the address of foo, in *any* DSO, you will not be
> using either R_386_PC32 or R_386_PLT32.  You'll be using either
> R_386_32 or R_386_GOT32.

But the currrent gcc uses R_386_GOTOFF for address of a protected
data or function symbol on x86 when PIC is used. I will say it is
a bug. R_386_GOT32 should be used here. I can provide a testcase
if asked.


H.J.

Re: PATCH: Optimize protected call for i386

[Alan Modra]

On Wed, May 14, 2003 at 04:24:39PM -0700, Richard Henderson wrote:
> On Thu, May 15, 2003 at 08:32:53AM +0930, Alan Modra wrote:
> > I see.  I think this is wrong.  gcc needs to emit the normal PIC call,
> > ie. "call foo@PLT" so the the linker can set up a plt entry if
> > needed.
> 
> I disagree.
> 
> > A plt entry is needed for protected symbols so that function
> > pointer comparisons work, since the x86 ABI uses the .plt entry for
> > the value of a function symbol referenced by a dynamic object.
> > Consider liba.so exporting a protected symbol foo.  Will a function
> > pointer, value foo, passed to libb.so compare equal to &foo in
> > libb.so?  What about a function pointer, value foo, passed from
> > libb.so back to liba.so?
> 
> None of which has anything to do with call instructions.

Well, it may be that the linker will set up a plt entry anyway,
thus making the function pointer comparisons work properly.  The
problem is that if gcc emits "call foo" for protected syms, you'll get
an R_386_PC32 reloc on foo.  How is the linker supposed to distinguish
this reloc from an R_386_PC32 on ".long foo - ."?  In the second case,
you really want to use the plt entry address.

> If liba.so contains a call to foo, ld *knows* that foo *must*
> resolve to the version in liba.so and nowhere else.  Thus we
> can direct the call to foo without going through a PLT entry.

Exactly.  ld can do this even when gcc emits "call foo@PLT".

> If you take the address of foo, in *any* DSO, you will not be
> using either R_386_PC32 or R_386_PLT32.  You'll be using either
> R_386_32 or R_386_GOT32.

Does the ABI preclude use of R_386_PC32 from assembly?

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre

Re: PATCH: Optimize protected call for i386

[Richard Henderson]

On Thu, May 15, 2003 at 09:21:07AM +0930, Alan Modra wrote:
> How is the linker supposed to distinguish
> this reloc from an R_386_PC32 on ".long foo - ."?

Who cares?  You can't generate this from C.  As for assembly,
"don't do that if you care about pointer equality".

Indeed, I can think of reasons to use this pcrel form if you
_don't_ want pointer equality.  Such as in .eh_frame data.


r~

Re: PATCH: Optimize protected call for i386

[Richard Henderson]

On Wed, May 14, 2003 at 04:50:07PM -0700, H. J. Lu wrote:
> > If you take the address of foo, in *any* DSO, you will not be
> > using either R_386_PC32 or R_386_PLT32.  You'll be using either
> > R_386_32 or R_386_GOT32.
> 
> But the currrent gcc uses R_386_GOTOFF for address of a protected
> data or function symbol on x86 when PIC is used. I will say it is
> a bug.

Agreed.

> R_386_GOT32 should be used here. I can provide a testcase if asked.

File a PR so it does not get lost.


r~

Re: PATCH: Optimize protected call for i386

[Alan Modra]

On Wed, May 14, 2003 at 05:40:03PM -0700, Richard Henderson wrote:
> On Thu, May 15, 2003 at 09:21:07AM +0930, Alan Modra wrote:
> > How is the linker supposed to distinguish
> > this reloc from an R_386_PC32 on ".long foo - ."?
> 
> Who cares?  You can't generate this from C.  As for assembly,
> "don't do that if you care about pointer equality".
> 
> Indeed, I can think of reasons to use this pcrel form if you
> _don't_ want pointer equality.  Such as in .eh_frame data.

Yeah, OK, fair enough.  HJ, go ahead and commit your
SYMBOL_CALLS_LOCAL patch (and post it to the binutils list).

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre