Re: binutils as sparc/setx issue

[Alan Modra]

On Wed, May 18, 2022 at 05:24:45AM -0500, Dullfire wrote:
> 
> 
> It appears that binutils as for sparc(64) is emitting incorrect relocations 
> for some case of 'setx'
> for example
> ```
> 
> 
>  % echo 'setx _data, %g7, %g4' | sparc64-misc-linux-gnu-as -KPIC -o 
> /tmp/test-sparc64.o
> 
>  % sparc64-misc-linux-gnu-objdump -rD /tmp/test-sparc64.o
> 
> 
> 
> /tmp/test-sparc64.o:     file format elf64-sparc
> 
> 
> 
> 
> 
> Disassembly of section .text:
> 
> 
> 
> 0000000000000000 <.text>:
> 
>    0: 0f 00 00 00     sethi  %hi(0), %g7
> 
>                       0: R_SPARC_HH22 _data
> 
>    4: 09 00 00 00     sethi  %hi(0), %g4
> 
>                       4: R_SPARC_LM22 _data
> 
>    8: 8e 11 e0 00     mov  %g7, %g7
> 
>                       8: R_SPARC_HM10 _data
> 
>    c: 88 11 20 00     mov  %g4, %g4
> 
>                       c: R_SPARC_GOT10        _data
> 
>   10: 8f 29 f0 20     sllx  %g7, 0x20, %g7
> 
>   14: 88 11 00 07     or  %g4, %g7, %g4
> 
> 
> ```
> This relocation set will give bits [63:10] of the symbol address, and 
> bits[9:0] of the got entry. Which is probably never the desired result.
> I have observed this on both 2.37 and 
> master(85aaf32e610e01ccde008e5bbfd6df95558dbeae).

Yes, the GOT10 reloc is due to the BFD_RELOC_LO10 emitted by
gas/config/tc-sparc.c:synthetize_setx being converted in
tc_gen_reloc.

That's not the only problem.  The relocs emitted in synthetize_setx
will cause text relocations in a dynamic library or PIE.  So it looks
to me that setx with a non-constant value when PIC is simply
unsupported.  I'd encourage you to post a patch to
binutils@sourceware.org fixing the problem, or at least emitting an
error rather than bad code.

-- 
Alan Modra
Australia Development Lab, IBM