[PATCH 5/6] aix: implement R_TOCU and R_TOCL relocations

[Alan Modra]

On Mon, Mar 01, 2021 at 09:20:26AM +0000, CHIGOT, CLEMENT wrote:
> > +  if (rel->r_type == R_TOCU)
> > +    *relocation = (*relocation & 0xffff0000) >> 16 ;
> > 
> > Please double-check this.  The doc might say "the high-order 16 bits
> > of the displacement" but I suspect that isn't 100% correct.  If the
> > code in your testcases is proper usage, and I expect it is, then the
> > high 16 bits needs to be adjusted for the low 16 bits being signed.
> > 
> > You probably want:
> >    *relocation = ((*relocation + 0x8000) >> 16) & 0xffff;
> 
> Yes, you're right. I wasn't aware of that. 
> 
> > +#ifdef OBJ_XCOFF
> > +  /* AIX often generates addis instructions using "addis RT, D(RA)"
> > +     format instead of the classic "addis RT, RA, SI" one.
> > +     Restore it to the default format as it's the one encoded
> > +     in ppc opcodes.  */
> > +  if (!strcmp (opcode->name, "addis") && strchr (str, '(') != NULL)
> > +    ppc_xcoff_fixup_addis (&str);
> > +#endif
> > +
> > 
> > Yikes, can't you tell people to fix their code?  I see so many ways
> > for this to go wrong, for example, if an expression calculating an
> > offset happened to use parentheses.  Also, ppc_xcoff_fixup_addis does
> > no sanity checking and can easily cause buffer overflows.
> 
> Sadly no... AIX assembler wants the 3rd argument of an addis to be absolute. 
> Thus, it's using a different syntax when it's not. 
> To be precise, "addis RT, RA, SI" is allowed only if SI is an absolute and
> "addis RT, D(RA)" is allowed only if D is a label/symbol having a relocation. 
> Thus, I need to find a way to allow both in GNU as. 
> I agree that my solution isn't the best one, but I didn't find any way to 
> add an opcode in ppc-opc.c. 
> If you have any idea how to do it properly, I would be glad. Otherwise, 
> I will have to use this approach.. 

The operand scanning isn't designed to support alternate operand
syntax like this, so you are correct that transforming the input is
likely the easiest way to support an ancient syntax not even mentioned
in current powerpc architecture books.  However, I'd like to see
better syntax checking before attempting the transform.  Something
like the following, perhaps.  Note the check for no comma past the
first one found, use of strrchr, and checks for empty strings.  Do you
need to treat "subis" the same way?

  if (strcmp (opcode->name, "addis") == 0)
    {
      char *rt_e = strchr (str, ',');
      if (rt_e != NULL
	  && strchr (rt_e + 1, ',') == NULL)
	{
	  char *d_e = strchr (rt_e + 1, '(');
	  if (d_e != NULL && d_e != rt_e + 1)
	    {
	      char *ra_e = strrchr (d_e + 1, ')');
	      if (ra_e != NULL && ra_e != d_e + 1)
		ppc_xcoff_fixup_addis (str, rt_e, d_e, ra_e);
	    }
	}
    }


-- 
Alan Modra
Australia Development Lab, IBM