Re: [Gold] how to do relax in Gold ?

[Shawn <shaolin dot xie at ia dot ac dot cn>]

On 2013å01æ16æ 14:09, Ian Lance Taylor wrote:
> On Tue, Jan 15, 2013 at 10:02 PM, Shawn <shaolin.xie@ia.ac.cn> wrote:
> > Hi All:
> >      I'am porting Gold to our processor right now ,everything is fine except
> > the label whose value is big.  For Example, the following instruction assign
> > a LABEL value to register R0:
> >     ===========
> >      R0 = LABEL_0;
> >     ===========
> >
> > 1. Our Objects
> >    Our instruction is encoded in 32bit,   if the value of "LABLE_0" is too
> > big, for example 0x40000, there is not enough space to encode the
> > instruction into a 32bit instruction, we have to expand "R0 = LABEL_0" into
> > a 64bit instruction automatically.
> >    Because the value of "LABEL_0" is not known until link-time, we hope
> > "Gold" can handle this .
> >
> > 2. My plan  & Question:
> >    I found the function related to our object is  the "Target::do_relax"
> > function, but only arm.cc & powerpc.cc has overwritten it. The "do_relax" in
> > arm.cc & powerpc.cc is too complicate to analyze. I hope to get a clean
> > implementation of Our Objects.
> >     a)  Scan  relocation info in all input objects, find the relocations that
> > may  need relaxation ( read & compare the symbol value )
> >           How can I read the relocation info ?
> >           How can I reany advice ad the symbol  value ? Is the value read
> > from the Symbol_table represent the value after layout ?
> >     b)  for each relocation that  need relaxation, insert a 32bit stub in the
> > VIEW, and update the value of the symbol which is located after the
> > relocation site.
> >
> >      And is there any other example related to relaxation except arm.cc &
> > powerpc.cc ? or any document ?
> No, sorry.
>
> I think one approach would be to record the possible relaxation points
> in your scan_relocs routine.  You will have all the relocation
> information available there.  That will let you detect out of range
> relocations where you can insert stubs.
>
> Ian
Thanks Ian. Did your mean Target::scan_relocs ? I will try that.

--
Shawn