Re: MIPS patch to avoid lazy binding in la macros

[Richard Sandiford <rsandifo at redhat dot com>]

Eric Christopher <echristo at redhat dot com> writes:
> > Hmm.  I see you're responding to my first message.  As I said
> > in a follow-up, I think we should stop gcc splitting the jal
> > macro in final.  Sound reasonable?
> > 
> 
> That's not a bad idea.
> 
> Another one is to have rewrite produce the instruction sequence
> (including relocs) that would be emitted by gas and we could go back and
> leave la alone. I don't much like this, but it is an idea that would
> help the assembly language programmers.

Yeah, the explicit reloc case should work OK.  The problem is the
"legacy" (?) macro support (!TARGET_EXPLICIT_RELOCS).  The new
register constraint for call addresses ('c') only allows $25
for TARGET_ABICALLS.  So instead of:

        la $2,foo
        beqz $2,1f
        move $25,$2
        jal $25
1:

we get:

        la $25,foo
        beqz $25,1f
        jal $25
1:

The point is, most "recent" gcc versions could emit the second
sequence as well.  They only avoid it by chance.  I really think
this is a correctness issue for all gccs, not just the
rewrite branch.

Richard