Re: [PATCH] powerpc: Work around gcc miscompilation of __pa() on 64-bit

[Paul Mackerras <paulus@samba.org>]

On Tue, Aug 27, 2013 at 04:42:35PM +0930, Alan Modra wrote:
> On Tue, Aug 27, 2013 at 04:07:49PM +1000, Paul Mackerras wrote:
> > On 64-bit, __pa(&static_var) gets miscompiled by recent versions of
> > gcc as something like:
> > 
> >         addis 3,2,.LANCHOR1+4611686018427387904@toc@ha
> >         addi 3,3,.LANCHOR1+4611686018427387904@toc@l
> 
> I might argue that this isn't a miscompilation, since -mcmodel=medium
> assumes everything can be accessed within +/-2G of the toc pointer,

And there's the bug right there... the assumption that this number
that I want to compute is the address of something that can be
accessed.  It isn't, and gcc shouldn't treat it as such (not unless I
do something like subsequently casting it back to a pointer and
dereferencing that).

> but it's definitely a problem since gas and/or ld don't give an
> overflow error.  They would except for the fact that our ABI has a
> hole in it.

Well, that and the fact that the code in gcc that generates
instructions to compute addresses blindly uses addis; addi even when
the offset involved is way too large for that to possibly work.

Paul.