How ld decides offset for relocation type MIPS GPREL16?

[Qinglai Xiao]

>> (ELF GP value in original object's .reginfo)

This is what I'm looking for. Thanks a bunch!

regards,
-Qinglai

On Mon, Jul 9, 2012 at 9:14 PM, Richard Sandiford
<rdsandiford@googlemail.com> wrote:
> Qinglai Xiao <jigsaw@gmail.com> writes:
>> I'm writing a runtime linker for MIPS32 arch, and have problems with
>> GPREL16 relocation type.
>> For instance, the GNU compiler/assembler outputs a relocatable object with
>> an instruction as below :
>>
>>     ff838018        sd      v1,-32744(gp)
>>
>> The symbol involved here belongs to .sbss section, and is set to
>> runtime address 0x12b119e8.
>>
>> After relocated by ld,
>> _gp = .0x12b12220
>> .sdata = 0x12b0a228
>> .sbss =  0x12b11868
>> Size of .sdata is 0x7640 and size of .sbss is 0x770.
>> ( The values are read from target ELF object.)
>>
>> And the instruction becomes:
>>
>>     ff83f7c8        sd      v1,-2104(gp)
>>
>> My question is: How ld calculates the offset 0xf7c8?
>> My guess is that 0xf7c8 is calculated by _gp, .sbss/.sdata addresses and
>> original offset 0x8018.
>
> Yeah.  For local symbols (including .sbss and .sdata section symbols
> like you say) it's:
>
>    (symbol value)
>    + (original offset)
>    + (ELF GP value in original object's .reginfo)
>    - (_gp value in final link)
>
>> And I believe the related code is in routine
>> _bfd_mips_elf_gprel16_reloc of binutils/src/bfd/elf32-mips.c.
>
> That's the code that writes the offset, but it's the R_MIPS_GPREL16
> case in mips_elf_calculate_relocation that does the calculation above.
>
> Richard