Encoding of immediates different to label addresses

[Andrew Burgess]

On 02/01/2012 09:44, Julius Baxter wrote:
> I'm working on fixing up the OpenRISC 1000 CGEN port and have come
> across something I find I can't quite do with CGEN.
>
> For the jump and branch instructions, if we use immediate value number
> specified in the assembly, we need that encoded into the instruction
> without any shifting occurring to it.
>
> Eg. "l.bf 1" should become 0x10000001, "l.bf 4" should become 0x10000004 etc.
>
> These target addresses are shifted-left by two during decoding, and so
> "l.bf 1" is actually a branch of 4 bytes.
>
> If we use a label for an address, though, I'm seeing that we're
> getting the offset in bytes during encoding, which then needs to be
> shifted right by two before being put in the instruction.
>
> But it seems we can have the shifting on both numerical immediates and
> labels, or neither. (Pardon my incorrect use of terminology, I have a
> feeling I'm not quite calling everything its right name!)
>
> Upon inspecttion, it looks like the binutils port we have been using
> for the past few years (not CGEN-based) has a right-shift of 2 in the
> md_apply_fix() function in binutils/gas/config/tc-or32.c (
> http://sourceware.org/git/?p=binutils.git;a=blob;f=gas/config/tc-or32.c;h=7234fb837e006f979c255f62206e39a9822046d2;hb=HEAD#l603
> ) when the instruction had the actual value of the label inserted into
> it.
>
> Is there any way to handle this using the CGEN description?

I don't know about that.

> The current instruction field definition being used is:
>
> (df f-disp26     "disp26"              (PCREL-ADDR) 25 26 INT
>      ((value pc) (sra WI (sub WI value pc) (const 2)))
>      ((value pc) (add WI (sll WI value (const 2)) pc)))
>
> This does the shifting on encode and the target label addresses are
> right but the numerical immediates are wrong, but if I remove the sra
> on the encode, the numerical immediates are right but the target label
> addresses are wrong.
>
> I've had a bit of a poke around and seen the function
> gas_cgen_md_apply_fix() is used as a generic md_apply_fix() function,
> but can't quite see where I might be able to hook in things for
> particular encoding cases.

What might work is to drop the shifting from the field definition and 
create your own md_apply_fix function something like this,

void
md_apply_fix (fixS *fixP, valueT *valP, segT seg)
{
   if (fixP->fx_done
       && fixP->fx_r_type == BFD_RELOC_UNUSED + OPERAND_ID_OF_BRANCH)
     {
        gas_assert (fixP->fx_addsy == NULL);
        gas_assert (fixP->fx_subsy == NULL);

        *valP >>= 2;
     }

     gas_cgen_md_apply_fix (fixP, valP, seg);
}


Good luck!

Andrew