Re: [patch] MIPS: Fix synthesized doubleword transfers (ping)

[Richard Sandiford <rsandifo at redhat dot com>]

Richard Sandiford <rsandifo@redhat.com> writes:
> "Maciej W. Rozycki" <macro@linux-mips.org> writes:
>> 	* config/tc-mips.c (macro): Avoid a reloc overflow for word 
>> 	memory transfers in expansions of the "li.d", "l.d", "s.d", "ld" 
>> 	and "sd" macros.
>
> Don't these instructions require doubleword-aligned addresses?
> If so, I don't see the problem.  The relocs for "sym+offset"
> should always be sorted as:
>
>         HI16 sym        (in-place addend of %hi(offset))
>         LO16 sym        (in-place addend of %lo(offset))
>         LO16 sym        (in-place addend of %lo(offset+4))
>
> and the linker will do the right thing.
>
> I suppose bad things could happen if you composed an aligned
> address from an unaligned symbol and an unaligned offset,
> but is that allowed?

FWIW, I checked what the MIPSpro assembler does.  Given:

        ld      $4,foo

as -32 -non_shared will generate:

   0:   3c010000        lui     at,0x0
                        0: R_MIPS_HI16  foo
   4:   8c240000        lw      a0,0(at)
                        4: R_MIPS_LO16  foo
   8:   8c250004        lw      a1,4(at)
                        8: R_MIPS_LO16  foo

just like gas does.  Not that I'm saying we should do something simply
because MIPSpro does it.  Just thought I'd throw it into the pot.

 [ Although, as is probably clear from earlier messages, my preference is
   to keep things the way they are, and require any symbolic component to
   be 8-byte aligned.  I've never heard of any practical problems with that
   restriction.  OTOH, after the above, I don't have any new data or
   arguments to add, so I'll shut up now ;) ]

Richard