Re: MIPS patch to avoid lazy binding in la macros

["Maciej W. Rozycki" <macro at ds2 dot pg dot gda dot pl>]

On 24 Feb 2003, Richard Sandiford wrote:

> > > Just to be clear, I'm talking about the gcc mips-3_4-rewrite-branch here.
> > > It has two modes: one which uses assembler macros (as current gcc does)
> > > and one that uses explicit %reloc() operators.  When using assembler
> > > macros, it gets lazy binding by using "jal foo".
> > 
> >  So what's the problem in the first place?
> 
> Well, there are two problems:
> 
> 1) The assembly sequence I posted is an efficient way of testing
>    and calling a weak function.  It works correctly if assembled
>    by an irix assembler.
> 
>    Now you could say, "well it doesn't work with gas; live with it".
>    But IMO this is not really any different from bfd breaking the ABI.
>    There have been examples in the past where bfd implemented the
>    ABI incorrectly, and it was fixed.

 The assembly sequence is OK, but why not simply make gcc use %reloc() 
operators for gas and assembler macros for IRIX as?

> Problem (2) is this...
> 
> > > >  Any "la $25,foo" actually as gcc used not to use the register for
> > > > anything else when emitting SVR4 PIC code.
> > > 
> > > Are you sure?  What stopped gcc from using $25 for anything else?
> > > Was it declared to be a fixed register?
> > 
> >  I haven't seen it elsewhere in generated code.  But I can't find it being
> > declared a fixed register, either (2.95).  It might have been pure luck
> > and I may be wrong, then.
> 
> 2) We have two conflicting precedents.  On the one hand, gcc treats
>    $25 as a general temporary register, on the other, gas treats it
>    as special.  Both are long-standing, but they aren't compatible.

 But with %reloc() operators the problem goes away, doesn't it?

> >  BTW, why the naming of "%got" vs "%call16" is inconsistent?  I fixed it
> > in my code -- you could do that, too.
> 
> Well, I stuck with the NewABI conventions.  There doesn't seem
> any point changing to a different name for o32.

 There is a point, not necessarily for o32 only: how do you express
something like this (using the notation I used previosly):

	la	$25,%call(foo)
	jal	%got(bar)

to get CALL relocations for foo and GOT ones for bar as appropriate for
the selected ABI and GOT model?  OK, a minor one, though -- they may be
named say %call_r and %got_r and the NewABI (inconsistent -- who the hell
does invent that stuff?) conventions be left as is.

 Frankly, given your %reloc stuff in the trunk I am not that much opposed
to removing the heuristics anymore, but how about deferring it until after
2.14 so that there is a release that can be used either way?  That would
ease transition a bit. 

-- 
+  Maciej W. Rozycki, Technical University of Gdansk, Poland   +
+--------------------------------------------------------------+
+        e-mail: macro at ds2 dot pg dot gda dot pl, PGP key available        +