Re: RFC: Patch to fix MIPS -mno-shared with multi-got...

[David Daney <ddaney at avtrex dot com>]

Thiemo Seufer wrote:
> David Daney wrote:
> [snip]
>
> > > WRT possible breakge to things expecting the real "_gp": I agree that
> > > might be a problem.  One possibility would be to define a new symbol,
> > > something like __gnu_local_gp, and make -mno-shared use that instead
> > > of plain _gp.  -mno-shared is a new option, so there's no backwards
> > > compatiblity problem.
> >
> > OK, I kind of like this idea.  But I am not an ABI authority.  I am 
> > assuming the the meaning of "_gp_disp" is defined somewhere in the ABI 
> > specs.  I know _gp is specified.  Does anybody have objections to adding 
> > the new magic __gnu_local_gp symbol?
>
>
> I have the general (and still vague) idea to have multigot-capable object
> files, with _gp _gp.1 .. _gp.n and _gp_disp _gp_disp.1 .. _gp_disp.n
> depending on the GOT in use, and with the single got case staying binary
> compatible. I think this would cover your case as well, but it's of
> course much more work.

I don't think that solves the problem we are talking about.

The main problem is that with -mno-shared you get object files with this:

    lui    $gp,%hi(_gp)
    addiu    $gp,$gp,%lo(_gp)

Where _gp refers to which ever GOT that the linker assignes to the module.

But it is possible to have C code like this:

extern int _gp;


int
foo()
{
  int *bar = &_gp;
/* Do something with the primary GOT */
}

In this case the existing semantics are clear.  It is the primary GOT.


My original patch changed the semantics of the above C program. 
Richard's suggestion to use __gnu_local_gp lets us have the desired 
semantics for .cpload with out changing those of the C program.

David Daney.