[PATCH] ia64: fix dl-sysdep miscompilation on gcc-10

[Adhemerval Zanella]

On 16/05/2020 10:53, slyich--- via Libc-alpha wrote:
> From: Sergei Trofimovich <slyfox@gentoo.org>
> 
> gcc-10 enabled -fno-common by default. This caused glibc link failure:
> 
> ```
> $ ia64-unknown-linux-gnu-gcc -pipe -O2 -nostdlib -nostartfiles \
>     -static -o .../elf/sln ...
> libc.a(dl-support.o): in function `setup_vdso':
> glibc-2.31/elf/setup-vdso.h:116:(.text+0x1322):
>     relocation truncated to fit: GPREL22 against `.text'
> collect2: error: ld returned 1 exit status
> ```

I can't reproduce it with master (b6ad64b907ab00) with GCC 10.1.1 20200520.
(built using build-many-glibcs.py). I tried to explicit use -fno-common on
CFLAGS. Are you sing a non default compiler option?

> 
> The difference between gcc-9 and gcc-10 is the way relocation is
> generated for ia64-specific symbol '_dl_sysinfo_break', defined as:
> 
> ```c
>   32 #ifndef __ASSEMBLER__
>   33 /* Don't declare this as a function---we want it's entry-point, not
>   34    it's function descriptor... */
>   35 extern int _dl_sysinfo_break attribute_hidden;
>   36 # define DL_SYSINFO_DEFAULT ((uintptr_t) &_dl_sysinfo_break)
>   37 # define DL_SYSINFO_IMPLEMENTATION              \
>   38   asm (".text\n\t"                              \
>   39        ".hidden _dl_sysinfo_break\n\t"          \
>   40        ".proc _dl_sysinfo_break\n\t"            \
> ```
> 
> Note: it's declared as ".sdata" (defined nearby GOT) but assumed
> to be defined in ".text" as it's really a direct function address.
> 
> As a result generated relocation changed from:
> 
> from `GPREL64I`:
> 
> ```
>     1390:       05 00 84 1e 98 11       [MLX]       st8 [r15]=r33
>                         1391: GPREL64I  .text
>     1396:       00 00 00 00 00 e0                   movl r15=0x0;;
>     139c:       01 00 00 60
>     13a0:       0b 78 04 1e 00 20       [MMI]       add r15=r1,r15;;
> ```
> 
> to `GPREL22`:
> 
> ```
>       if (GLRO(dl_sysinfo) == DL_SYSINFO_DEFAULT)
> 
>     1320:       0b 80 00 1c 18 10       [MMI]       ld8 r16=[r14];;
>                         1322: GPREL22   .text
>     1326:       00 08 3d 30 23 e0                   st8 [r15]=r33
>     132c:       01 08 00 90                         addl r15=0,r1;;
> ```
> 
> This change explicitly declares symbol to be in ".text" section.
> That allows gcc to generate far relocations.
> 
> Regression tested on glibc-master and gcc-10.1.0 on rx3600.

The rationale is sound, but the trick to define a function pointer a 
data variable on text sections seems hacky and fragile. I still think
it could be a fix, but I think we should refactor and simplify
the over-enginnering of NEED_DL_SYSINFO/USE_DL_SYSINFO.

> 
> Bug: https://bugs.gentoo.org/723268
> Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org>
> ---
>  sysdeps/unix/sysv/linux/ia64/dl-sysdep.h | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/sysdeps/unix/sysv/linux/ia64/dl-sysdep.h b/sysdeps/unix/sysv/linux/ia64/dl-sysdep.h
> index 78fa6dd2c9..e526e02ff4 100644
> --- a/sysdeps/unix/sysv/linux/ia64/dl-sysdep.h
> +++ b/sysdeps/unix/sysv/linux/ia64/dl-sysdep.h
> @@ -32,7 +32,9 @@
>  #ifndef __ASSEMBLER__
>  /* Don't declare this as a function---we want it's entry-point, not
>     it's function descriptor... */
> -extern int _dl_sysinfo_break attribute_hidden;
> +/* Use section ".text" to force far GPREL64 relocation instead of
> +   GPREL22 . */
> +extern int _dl_sysinfo_break attribute_hidden __attribute__((section(".text")));
>  # define DL_SYSINFO_DEFAULT ((uintptr_t) &_dl_sysinfo_break)
>  # define DL_SYSINFO_IMPLEMENTATION		\
>    asm (".text\n\t"				\
>