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

[Thiemo Seufer <ica2_ts@csv.ica.uni-stuttgart.de>]

Richard Sandiford wrote:
> Thiemo Seufer <ica2_ts@csv.ica.uni-stuttgart.de> writes:
> > How could it do that? The o32 ABI does _not_ specify this flag,
> > a file without any ABI flags set is a valid o32 file. I can't
> > see what the use of the O32 header flag could be.
> 
> Ok, that hadn't twigged, sorry!
> 
> So, if I've understood you correctly, your take seems to be:
> 
> - O32 doesn't define an ABI bit for ELF, so every ELF file without
>   an ABI flag set should be compatible with O32.

Not exactly. Most tools will _regard_ it as o32, even if it isn't.

> Whereas mine is:
> 
> - O32 doesn't define an ABI bit for ELF, so users dosn't have that
>   safety net if they use the wrong options.  They need to remember
>   to use the appropriate command-line switches, like -mabi=32.

ABI header flags were invented for remembering it. It would be nice
if the linker had a chance to find out what the file actually is.

>   Fleshing that out a bit, I think:
> 
>   GAS has historically supported all sorts of wierd and wonderful
>   combinations that don't have any representation in the ELF header
>   flags.  And for which no ABI flags will be set.  I don't think that's
>   reason enough not to allow them.  GAS should generate whatever code
>   the command line tells it to, whether or not those options can be
>   determined from the output file.

One NO_ABI_FLAG as a catchall would be enough to let at least the
GNU tools know what the file is _not_.

[snip]
> > AFAICS my patch doesn't break the limited 64bit support.
> 
> Like you said in your previous mail, it means that $gp is now assumed to
> be a 32-bit rather than 64-bit value.

This is always assumed for everything loaded in 32bit address space.

> There might be dynamic loaders
> out there that handle the R_MIPS_64 extension, which could replace them
> with genuine 64-bit (as opposed to sign-extended 32-bit) addresses.

They can't, because R_MIPS_64 holds only 32bit in this case.

[snip]
> > Yes, it inflates the number of variants without need.
> 
> ...I really don't see why maintaining the old variant is a problem.  All
> the main part of your patch did was change the definition of the
> HAVE_32BIT_ADDRESSES macro, it didn't really seem to simplify the code
> as such.

It didn't increase complexity while providing all what I need
for full 64bit support. That's the advantage over a
32BIT -- HALF_64BIT -- FULL_64BIT approach.

[snip]
> I still don't see why you need to do that, sorry!  Sometimes these things
> take a while to sink in.  In what situation would the two be different?
> Like you say, they should be inverses.

Full 64bit support requires a 64bit object format to work, half
64bit support doesn't have one and has it's code in 32bit space.
This means for e.g. "dli" a expansion to

	lui	$a, %highest(sym)
	lui	$b, %hi(sym)
	daddiu	$a, %higher(sym)
	daddiu	$b, %lo(sym)
	dsll32	$a, 0
	daddu	$a, $a, $b

for 64bit addresses, while 32bit addresses should use

	lui	$a, %hi(sym)
	addiu	$a, %lo(sym)

for performance and code size. The half 64bit version currently
uses

	lui	$a, %hi(sym)
	daddiu	$a, %lo(sym)

which does exactly the same but pretends to use 64bit addresses.
I hope it got clearer now.


Thiemo