New assembler warnings with binutils 2.29

[Catalin Marinas]

On Mon, Aug 14, 2017 at 10:55:48AM +0100, Ramana Radhakrishnan wrote:
> On Mon, Aug 14, 2017 at 09:59:02AM +0100, Catalin Marinas wrote:
> > On Fri, Aug 11, 2017 at 10:26:06AM +0100, Ard Biesheuvel wrote:
> > > On 11 August 2017 at 10:22, Catalin Marinas <catalin.marinas at arm.com> wrote:
> > > > On Thu, Aug 10, 2017 at 01:13:22PM -0700, Laura Abbott wrote:
> > > >> Fedora rawhide recently upgraded to binutils 2.29 and this seems
> > > >> to produce new warnings:
> > > >>
> > > >> ./arch/arm64/include/asm/assembler.h: Assembler messages:
> > > >> ./arch/arm64/include/asm/assembler.h:125: Warning: ignoring attempt to redefine built-in register 'lr'
> > > >>
> > > >> This is
> > > >>
> > > >> /*
> > > >>  * Register aliases.
> > > >>  */
> > > >> lr      .req    x30             // link register
> > > >
> > > > Strange, does gas now think 'lr' is a general purpose register (aliased
> > > > to x30)? It never was and IIRC the toolchain people many years ago
> > > > refused to add it, hence the alias above in the kernel. I wonder if they
> > > > added 'fp' as well...
> 
> I don't remember the reasons why it was refused in the past. Asking around I
> can't seem to find anyone who remembers why either and it would be much
> quicker to fix the problem in binutils than any of this archeology :)

I don't have the old emails around either ;).

> > > > We could remove the alias and replace all 'lr' instances with 'x30'
> > > > throughout the kernel (no too many) or we add some #ifdef around the
> > > > above based on the binutils version.
> > > > This is annoying. Replacing lr with x30 achieves the opposite of the
> > > intent of the binutils change. And using #ifdefs is inaccurate,
> > > because you can't really test the binutils version only the GCC
> > > version, and those are not tightly coupled.
> > > > Can you .unreq it?
> > 
> > Not really, with an older binutils I get:
> > 
> > arch/arm64/include/asm/assembler.h:125: Error: unknown register alias 'lr'
> > 
> > I personally consider this a binutils bug. After 6+ years (probably not
> > all public) of building the kernel just fine, all of a sudden certain
> > strings became reserved in gas. Three options:
> 
> IIRC the ABI states that LR is an alias for X30, FP is an alias for X29,
> IP0 an alias for X16 and IP1 an alias for X17.

It's unfortunate that binutils never define them, so we ended up with
.req in the kernel.

> The warning is unfortunate - looking at the implementation of the
> .req directive in gas it's something controlled in the aarch64 backend
> and thus we could change gas to make LR (and friends) an internal alias thus
> removing the warning.
> 
> Any attempts to realias these to the correct meaning as per the AAPCS
> continues to work from my limited testing of a prototype patch so far.

This would be great. Thanks Ramana!

-- 
Catalin