Re: [PATCH] Fix distinction of 32/64bit addresses in MIPS gas

[Richard Sandiford <r.sandiford@redhat.com>]

Thiemo Seufer <ica2_ts@csv.ica.uni-stuttgart.de> writes:

> Richard Sandiford wrote:
> [snip]
> > AFAIK, R_MIPS_64 won't be used unless you have 64-bit addresses (and
> > hence 64-bit registers), or if you have an explicit pseudo-op like
> > .8byte.  o32 binaries wouldn't be using either of those things anyway,
> > would they?
> 
> That's right, but you can't tell if it is actually o32 conformant
> by looking at the ELF header. There is no flag which says
> "Warning! This object does not conform to any established ABI".
> It would be nice to have one.

Well, there's the O32 flag, which is only set if -mabi=32 is given.
That bit does guarantee register and address size, just not the lack of
64-bit relocs generated by .8byte and the like.  Maybe GAS should warn
about that?

> > I thought the consensus some time ago was that o32 implied 32-bit
> > registers, and therefore 32-bit addresses.  The HAVE_??BIT_ macros are
> > already set up like that as long as you specify -mabi=32 on the command
> > line.
> 
> That's all fine for 32bit, but I needed a way to check if 64bit
> addresses can be used generally (e.g. for dli). Something like
> ! HAVE_32BIT_ADDRESSES won't do that, because full 64bit support
> needs a 64bit object format.

But *limited* 64-bit support is available with elf32 as things stand...

> Having a HAVE_64BIT_ADDRESSES macro which is _not_ the inverse of
> HAVE_32BIT_ADDRESSES is ugly.

...so I'm still unsure why that's necessary.

> I also found no reason why there is made use of 64bit instructions
> like daddiu for a 32bit load. It makes the code look different from
> o32 while doing the same.

But if the code was written or compiled with 64-bit addresses in mind,
it seems natural enough to use daddu in pointer arithmetic.  Does it do
any harm?  The only reason I can see for making HAVE_32BIT_ADDRESSES any
stricter is if there's a situation in which code written for 32-bit
addresses can't be assembled that way as things stand, presumably
because GAS's command-line options are too inflexible.

> This was the one part of my patch. The other was to change (d)la to
> chose it's expansion in dependency of the address model instead of
> the insn name. That's the way the SGI assembler behaves, and this
> is the useful behaviour for ABI conformance. For 64bit code in an
> 32bit object file this may have side effects (the usability of
> R_MIPS_64 is not affected).

Right.  For the record, I'm only talking about the other bit (the big
one).

Richard