Legacy _IO_* symbols and Flaot128 transition

Legacy _IO_* symbols and Flaot128 transition

[Florian Weimer]

libio exports a bunch of symbols for historic reasons:

_IO_fprintf
_IO_printf
_IO_sprintf
_IO_sscanf
_IO_vfprintf
_IO_vfscanf
_IO_vsprintf

These aren't compat symbols yet, but will not be in installed headers 
for glibc 2.28.  Zack's cleanup patches only turns _IO_vfscanf into a 
compat symbol (in “Add __vfscanf_internal and __vfwscanf_internal with 
flags arguments.”).

I think we should turn all of them into compat symbols and drop their 
compatibility wrappers from nldbl_nonshared.a, and avoid adding 
binary128 compatibility wrappers for them.

I can work on a patch if that would help.

Comments?

Thanks,
Florian

Re: Legacy _IO_* symbols and Flaot128 transition

[Gabriel F. T. Gomes]

On Fri, 29 Jun 2018, Florian Weimer wrote:

>libio exports a bunch of symbols for historic reasons:
>
>_IO_fprintf
>_IO_printf
>_IO_sprintf
>_IO_sscanf
>_IO_vfprintf
>_IO_vfscanf
>_IO_vsprintf
>
>These aren't compat symbols yet, but will not be in installed headers 
>for glibc 2.28.  Zack's cleanup patches only turns _IO_vfscanf into a 
>compat symbol (in “Add __vfscanf_internal and __vfwscanf_internal with 
>flags arguments.”).
>
>I think we should turn all of them into compat symbols and drop their 
>compatibility wrappers from nldbl_nonshared.a, and avoid adding 
>binary128 compatibility wrappers for them.

Should they be turned into compat symbols *because* they won't be
installed headers (and we don't want new user code linking against them)?
Or is there something else to it?

As for binary128 compatibility wrappers for them, I agree (I haven't
created these on my branch, so this should already be in sync with your
suggestions).

>I can work on a patch if that would help.

That would help a lot! :)

Thanks,
Gabriel

Re: Legacy _IO_* symbols and Flaot128 transition

[Florian Weimer]

On 06/29/2018 05:26 PM, Gabriel F. T. Gomes wrote:
> On Fri, 29 Jun 2018, Florian Weimer wrote:
> 
>> libio exports a bunch of symbols for historic reasons:
>>
>> _IO_fprintf
>> _IO_printf
>> _IO_sprintf
>> _IO_sscanf
>> _IO_vfprintf
>> _IO_vfscanf
>> _IO_vsprintf
>>
>> These aren't compat symbols yet, but will not be in installed headers
>> for glibc 2.28.  Zack's cleanup patches only turns _IO_vfscanf into a
>> compat symbol (in “Add __vfscanf_internal and __vfwscanf_internal with
>> flags arguments.”).
>>
>> I think we should turn all of them into compat symbols and drop their
>> compatibility wrappers from nldbl_nonshared.a, and avoid adding
>> binary128 compatibility wrappers for them.
> 
> Should they be turned into compat symbols *because* they won't be
> installed headers (and we don't want new user code linking against them)?
> Or is there something else to it?

No, it's just a cleanup, and to clarify what's necessary for future work.

I also want to add some minimal test for libnldbl_nonshared.a, and 
reducing its size helps with that.  (I'm aware that on ppc64, the 
current toolchain doesn't support long double as double anymore, but I 
can just lie to the compiler and use double instead, I guess.)

> As for binary128 compatibility wrappers for them, I agree (I haven't
> created these on my branch, so this should already be in sync with your
> suggestions).

Good to know.

Florian

Re: Legacy _IO_* symbols and Flaot128 transition

[Joseph Myers]

On Fri, 29 Jun 2018, Florian Weimer wrote:

> > Should they be turned into compat symbols *because* they won't be
> > installed headers (and we don't want new user code linking against them)?
> > Or is there something else to it?
> 
> No, it's just a cleanup, and to clarify what's necessary for future work.

Anything not accessible from installed headers is a good candidate for 
making into a compat symbol.

> I also want to add some minimal test for libnldbl_nonshared.a, and reducing
> its size helps with that.  (I'm aware that on ppc64, the current toolchain
> doesn't support long double as double anymore, but I can just lie to the
> compiler and use double instead, I guess.)

It certainly appears to be supported to me (using GCC mainline, compiling 
with -mlong-double-64 produces sizeof (long double) == 8 for both 
powerpc64 and powerpc64le).

However, libnldbl_nonshared.a is for long double = double *with compilers 
not supporting asm redirection*.  Which is a case the (incorrect) patch in 
<https://sourceware.org/ml/libc-alpha/2014-08/msg00497.html> provides 
evidence hasn't worked for a long time.  So I still think there's a case 
for eliminating libnldbl_nonshared.a and explicitly requiring asm 
redirection support for these non-default long double variants, which it's 
in practice required already (if not more generally requiring it for any 
compiler using the glibc headers), as discussed in 
<https://sourceware.org/ml/libc-alpha/2018-03/msg00281.html>.

(The actual __nldbl_* exports from the glibc shared libraries - the 
functions that get used with -mlong-double-64 given asm redirection 
support - *should* be tested, via building some tests with 
-mlong-double-64.  The binary128 work should, as per 
<https://sourceware.org/ml/libc-alpha/2018-06/msg00680.html>, add tests 
that are generic to different long double formats, so that it will then be 
easy to build them for -mlong-double-64 for ldbl-opt configurations.)

-- 
Joseph S. Myers
joseph@codesourcery.com

Re: Legacy _IO_* symbols and Flaot128 transition

[Florian Weimer]

On 06/29/2018 07:15 PM, Joseph Myers wrote:
> On Fri, 29 Jun 2018, Florian Weimer wrote:
> 
>>> Should they be turned into compat symbols *because* they won't be
>>> installed headers (and we don't want new user code linking against them)?
>>> Or is there something else to it?
>>
>> No, it's just a cleanup, and to clarify what's necessary for future work.
> 
> Anything not accessible from installed headers is a good candidate for
> making into a compat symbol.

Agreed.

>> I also want to add some minimal test for libnldbl_nonshared.a, and reducing
>> its size helps with that.  (I'm aware that on ppc64, the current toolchain
>> doesn't support long double as double anymore, but I can just lie to the
>> compiler and use double instead, I guess.)
> 
> It certainly appears to be supported to me (using GCC mainline, compiling
> with -mlong-double-64 produces sizeof (long double) == 8 for both
> powerpc64 and powerpc64le).

Oh, this option isn't documented for POWER, so I didn't try it.

> However, libnldbl_nonshared.a is for long double = double *with compilers
> not supporting asm redirection*.  Which is a case the (incorrect) patch in
> <https://sourceware.org/ml/libc-alpha/2014-08/msg00497.html> provides
> evidence hasn't worked for a long time.  So I still think there's a case
> for eliminating libnldbl_nonshared.a and explicitly requiring asm
> redirection support for these non-default long double variants, which it's
> in practice required already (if not more generally requiring it for any
> compiler using the glibc headers), as discussed in
> <https://sourceware.org/ml/libc-alpha/2018-03/msg00281.html>.

Okay, I review the situation to determine which way is the best way 
forward for the 2.28 release.

> (The actual __nldbl_* exports from the glibc shared libraries - the
> functions that get used with -mlong-double-64 given asm redirection
> support - *should* be tested, via building some tests with
> -mlong-double-64.  The binary128 work should, as per
> <https://sourceware.org/ml/libc-alpha/2018-06/msg00680.html>, add tests
> that are generic to different long double formats, so that it will then be
> easy to build them for -mlong-double-64 for ldbl-opt configurations.)

I suppose we should also add tests for the functions which are not 
redirected to __nldbl_* variants, but to the non-l variants (because 
only for things like printf, separate wrappers are needed if there is 
redirection support).

Thanks,
Florian