Re: [PATCH 6 of 6] arm: Enable VFP at boot

[Tim Deegan <tim@xen.org>]

At 10:20 +0100 on 01 Jun (1338546010), Ian Campbell wrote:
> > -CFLAGS += -march=armv7-a -mcpu=cortex-a15
> > +CFLAGS += -march=armv7-a -mcpu=cortex-a15 -mfpu=vfpv3 -mfloat-abi=softfp
> 
> This might have been worthy of a comment...

Ah, yes it might.  Would you like me to add one and resubmit?

If I get time next week I'm going to look at VFP context switching; at
that point I might move all the VFP code into its own .c or .S file and
enable the fpu instructions just there, to be sure that gcc doesn't
start using the VFP registers in normal Xen code -- it will do that for
integer arithmetic with -mfpu=neon, for example.  OTOH it might be
OK to do that as long as there are no caller-save VFP registers; I'll
have a look at the calling conventions and see what can be done.

Tim.

> http://wiki.debian.org/ArmHardFloatPort has some discussion, including:
>         * soft: Full software floating point. 
>         * softfp: Use the FPU, but remain compatible with soft-float
>         code. 
>         * hard: Full hardware floating point. 
> 
> and "The combination of -mfpu=vfp and -mfloat-abi=hard is not available
> in FSF GCC 4.4". (there's a reference to a TODO section which seems to
> not exist anymore...). I think Debian's armhf does actually use hard
> (the page seems to imply it)
> 
> Anyway softfp does seem like the right option, at least until
> -mfloat-abi=hard. Since this is all internal to the hypervisor (guests
> can do whatever they want, I'm happily running armhf binaries) I don't
> think this is a big deal to change as and when the compiler will let us.
>