Re: mips multi-got error

[Richard Sandiford <richard at codesourcery dot com>]

Thiemo Seufer <ths@networkno.de> writes:
> David Daney wrote:
>> Thiemo Seufer wrote:
>> >David Daney wrote:
>> >[snip]
>> >
>> >>That is not supposed to work.  Multi-got linking only works if there are 
>> >>fewer than 8192 symbols per compilation unit (input .o to the linker). 
>> >>See the multi-got-1 testcase in ld/testsuite/ld-mips-elf.
>> >>
>> >>If you need more than 8192 symbols per compilation unit you must use 
>> >>-xgot for the assembler.  Recent GCC versions have a -mxgot option that 
>> >>does the same thing for GCC.
>> >
>> >
>> >Actually, it is supposed to work. :-)  It is also a known deficiency
>> >in the current assembler/linker implementation.

David's surely right here.  It isn't _supposed_ to work.  You might
be able to invent new schemes that allow it to work, but it won't
work with the current scheme.

>> How?
>> 
>> The gp is set in the function prolog.  If a single function references 
>> more got/plt entries than can be indexed with a 16 bit offset, wouldn't 
>> you need to either load the gp before each got access(very inefficient I 
>> would think) or use a 32 bit offset (somewhat inefficient)?
>
> It doesn't need a gp load before each access, one per 16-bit addressable
> GOT block would be enough. (This implies a Multi-GOT extension for
> relocatable object files, probably with gp0..gpN / _gp_disp0.._gp_dispN
> as per-GOT symbols.)

But how would you know which _gp_dispN applied to a given GOT relocation?
I hope it wouldn't be by comparing relocation field addresses.

Richard