Re: [kbuild-all] mipsel-linux-gnu-gcc: error: unrecognized command line option '-mcompact-branches=optimal'

From: Paul Burton
Date: Fri Apr 22 2016 - 04:54:57 EST


On Thu, Apr 21, 2016 at 10:10:51PM +0100, Maciej W. Rozycki wrote:
> On Thu, 21 Apr 2016, Ralf Baechle wrote:
>
> > > I don't think it makes sense as the compiler won't support MIPSr6 code
> > > anyway, so first it'll bail out on `-march=mips32r6', and if we go even
> > > further and disable that too, then GAS will probably break somewhere on
> > > inline asm and GCC will produce code which does not make sense otherwise.
> >
> > GCC 5.2.0 claims to support mips32r6 and mips64r6. It's just the option
> > -mcompact-branches which seem to have been added later only.
>
> Ah, I see -- I didn't track the timeline of support for this compiler's
> option and I took it from an earlier response that the compiler does not
> support R6 at all.
>
> In that case however it looks to me like these `-mcompact-branches='
> options (all the three we support) need to be wrapped into `$(call
> cc-option,...)'.

An alternative that it could be argued better fits the principle of
least surprise is to add an extra option to the Kconfig choice that
simply leaves -mcompact-branches unspecified. I just submitted a patch
to do so [1].

> They do not affect any functionality and they are an
> optimisation choice only anyway (and therefore I wonder why they've been
> placed in arch/mips/Kconfig.debug rather than arch/mips/Kconfig).

They're in Kconfig.debug because debug is exactly what they've been
useful for - given that compact branches are new to R6 it's been useful
in debugging systems, both hardware & simulators, to sometimes not use
them. It's also been useful to force their use attempting to work around
the compiler bug that [2] works around differently (bug 2179 on DMZ
bugzilla). On the other hand I can't think of a reason we'd want to
specify compact branch policy that isn't for debug - I'd expect for
performance optimisation we're more likely to rely upon the toolchain
using a sensible policy if the kernel is built for a specific CPU (eg.
perhaps -mcpu=p6600 prefers non-compact branches & -mcpu=m6250 prefers
all compact branches, or similar).

Thanks,
Paul

[1] https://patchwork.linux-mips.org/patch/13165/
[2] https://patchwork.linux-mips.org/patch/12556/