Re: [PATCH] Merge glibc ieee128 ldbl redirect support into cdefs.h

[Paul E Murphy]

On 1/1/21 1:49 PM, Bruno Haible wrote:
> Hi,
>
> Paul Murphy wrote in
> <https://lists.gnu.org/archive/html/bug-gnulib/2020-12/msg00106.html>:
> > While trying to build gdb with ldbl == ieee128 on ppc64le, we found out
> > gnulib's copy of cdefs.h needs to be synchronized with upstream glibc.
> >
> > The only notable divergence is to ensure
> > __LDOUBLE_REDIRECTS_TO_FLOAT128_ABI is always defined as this macro is
> > fairly recent.
>
> Why would this be needed? Nothing in Gnulib uses this macro.
> The cdefs.h of Gnulib is meant to be a minimal replacement (you can also
> call it a "mock") of the one in Glibc.
>
> Can you show the error message you got when compiling gdb, and how
> you traced it back to a Gnulib issue?


This was encountered when building IBM's advance toolchain (or 
sufficiently new GCC) with the -mabi=ieeelongdouble option, and glibc 
2.32 or newer.

This can be reproduced by building Tulio's branch of AT with IEEE128 
enabled. https://github.com/tuliom/advance-toolchain/tree/ldbl-ieee128

We ran into the following:

In file included from 
/home/tuliom/opt/at-next-15.0-0-alpha/include/stdio.h:871,
                 from ./stdio.h:43,
                 from 
/home/tuliom/tmp/at-build-tray/at15.0-0-alpha.redhat-8_ppc64le_ppc64le/sources
/gdb/gnulib/import/tempname.h:23,
                 from 
/home/tuliom/tmp/at-build-tray/at15.0-0-alpha.redhat-8_ppc64le_ppc64le/sources
/gdb/gnulib/import/tempname.c:20:
/home/tuliom/opt/at-next-15.0-0-alpha/include/bits/stdio-ldbl.h: In 
function ‘__LDBL_REDIR1_DECL’:
/home/tuliom/opt/at-next-15.0-0-alpha/include/bits/stdio-ldbl.h:36:1: 
error: expected declaration specifiers before ‘__LDBL_REDIR1_DECL’
   36 | __LDBL_REDIR1_DECL (scanf, __isoc99_scanfieee128)
      | ^~~~~~~~~~~~~~~~~~


The missing macro is provided by glibc's cdefs.h.  Thus, the wrong 
cdefs.h was picked up. I found gnulib provides its own copy which does 
not provide these.  Likewise, I verified applying this patch to AT's 
gdb.  This fixed compilation, and the redirects worked as expected.