Re: [PATCH 6/9] Add __vsyslog_internal, with same flags as __v*printf_internal.

[Florian Weimer <fweimer at redhat dot com>]

On 03/13/2018 02:37 PM, Zack Weinberg wrote:
> On Tue, Mar 13, 2018 at 8:43 AM, Florian Weimer <fweimer@redhat.com> wrote:
> > On 03/13/2018 01:39 PM, Zack Weinberg wrote:
> > > On Tue, Mar 13, 2018 at 7:59 AM, Florian Weimer <fweimer@redhat.com>
> > > wrote:
> > > > On 03/07/2018 08:32 PM, Zack Weinberg wrote:
> > > >
> > > > > +/* __vsyslog_internal uses the same mode_flags bits as
> > > > > +   __v*printf_internal; see libio/libioP.h.  */
> > > > > +extern void __vsyslog_internal (int pri, const char *fmt,
> > > > > __gnuc_va_list
> > > > > ap,
> > > > > +                                unsigned int mode_flags)
> > > > > +     __attribute__ ((__format__ (__printf__, 2, 0)));
> > > >
> > > >
> > > >
> > > > I'm surprised that this doesn't need attribute_hidden or
> > > > libc_hidden_proto
> > > > to avoid new PLT calls.
> > >
> > >
> > > That's only needed for functions that will be called _both_ from
> > > inside and outside glibc.  This function is only ever called from
> > > inside glibc, so it doesn't appear in any Versions files and it's
> > > hidden by default.
> >
> >
> > Some architectures will still use indirect calls without attribute_hidden,
> > so please add it.  The existing tests do not catch this reliably
> > unfortunately.
>
> Can you be more specific?  This will affect all of the other new
> __*_internal functions added in this patchset, so I need to know how
> to be sure I got it right.  Also, this seems like something we should
> find a way to automate if at all possible.

Consider this code:

$ cat call.c
int external (void) ATTR;

int
call (void)
{
  return external () + 1;
}

With default visibility, GCC 7 produces:

$ ppc64-linux-gnu-gcc -m32 -fPIC -c -O2 -DATTR= call.c && 
ppc64-linux-gnu-objdump -d --reloc call.o

call.o:     file format elf32-powerpc


Disassembly of section .text:

00000000 <call>:
   0:	94 21 ff e0 	stwu    r1,-32(r1)
   4:	7c 08 02 a6 	mflr    r0
   8:	42 9f 00 05 	bcl     20,4*cr7+so,c <call+0xc>
   c:	93 c1 00 18 	stw     r30,24(r1)
  10:	7f c8 02 a6 	mflr    r30
  14:	90 01 00 24 	stw     r0,36(r1)
  18:	3f de 00 00 	addis   r30,r30,0
			1a: R_PPC_REL16_HA	.got2+0x800e
  1c:	3b de 00 00 	addi    r30,r30,0
			1e: R_PPC_REL16_LO	.got2+0x8012
  20:	48 00 00 01 	bl      20 <call+0x20>
			20: R_PPC_PLTREL24	external+0x8000
  24:	80 01 00 24 	lwz     r0,36(r1)
  28:	83 c1 00 18 	lwz     r30,24(r1)
  2c:	38 21 00 20 	addi    r1,r1,32
  30:	38 63 00 01 	addi    r3,r3,1
  34:	7c 08 03 a6 	mtlr    r0
  38:	4e 80 00 20 	blr

With hidden visibility, we get instead:

$ ppc64-linux-gnu-gcc -m32 -fPIC -c -O2 -DATTR='__attribute ((visibility 
("hidden")))' call.c && ppc64-linux-gnu-objdump -d --reloc call.o

call.o:     file format elf32-powerpc


Disassembly of section .text:

00000000 <call>:
   0:	94 21 ff e0 	stwu    r1,-32(r1)
   4:	7c 08 02 a6 	mflr    r0
   8:	93 c1 00 18 	stw     r30,24(r1)
   c:	90 01 00 24 	stw     r0,36(r1)
  10:	48 00 00 01 	bl      10 <call+0x10>
			10: R_PPC_LOCAL24PC	external
  14:	80 01 00 24 	lwz     r0,36(r1)
  18:	83 c1 00 18 	lwz     r30,24(r1)
  1c:	38 21 00 20 	addi    r1,r1,32
  20:	38 63 00 01 	addi    r3,r3,1
  24:	7c 08 03 a6 	mtlr    r0
  28:	4e 80 00 20 	blr

The linker may have some optimization to eliminate the PLT indirection 
(blinding the localplt test), but it cannot get rid of the other 
unnecessary instructions.

Does this example help?

Thanks,
Florian