[PATCH] ARM64: Fix compiling with GCC 6 and Atomics enabled

[Will Deacon]

On Mon, Dec 21, 2015 at 01:58:30PM +0100, Ard Biesheuvel wrote:
> On 21 December 2015 at 13:51, Will Deacon <will.deacon at arm.com> wrote:
> > On Mon, Dec 21, 2015 at 01:46:22PM +0100, Ard Biesheuvel wrote:
> >> On 21 December 2015 at 13:38, Will Deacon <will.deacon at arm.com> wrote:
> >> > On Fri, Dec 18, 2015 at 08:17:35PM -0800, Andrew Pinski wrote:
> >> >> The problem here is that GCC 6 and above emits .arch now
> >> >> for each function so now the global .arch_extension has
> >> >> no effect.  This fixes the problem by putting
> >> >> .arch_extension inside ARM64_LSE_ATOMIC_INSN so
> >> >> it is enabled for each place where LSE is used.
> >> >
> >> > Hmm, this is going to affect arch/arm/ much more heavily than arch/arm64.
> >> > .arch_extension is used for virt, mp and sec over there, and it may be
> >> > tricky to isolate the actual instruction usage (at least, virt looks
> >> > lost in kvm/arm.c).
> >> >
> >> > Why can't gas have an option to accept all instruction encodings that it
> >> > knows about, inspite of any .arch directives?
> >> >
> >>
> >> Modern GAS supports things like -march=armv7-a+mp+sec+virt, so it
> >> probably makes sense to pass that on the command line when building
> >> for v7 (or +sec only for v6) if the assembler is found to support it
> >> at build time.
> >
> > Does that override a more restrictive .arch directive emitted by the
> > compiler?
> >
> 
> It seems to be additive: -march=armv7-a+mp+sec allows a .S file
> containing a virt arch_extension + both hvc and smc instructions to be
> assembled.

The problem I'm seeing is if I have something like:

  .arch_extension lse

before something like:

  .cpu cortex-a57+fp+simd+crc

-or-

  .arch armv8-a+fp+simd+crc

then I can no longer assemble lse instructions. So the .cpu/.arch
directive is undoing the .arch_extension. We can fix this by following
Andrew's suggestion to have .arch_extension before every point of use,
but the whole thing would be much simpler if we could just tell gas to
assemble harder.

Maybe we just need to construct the mother of all -march options based
on build-time checks in the Makefile?

Will