Re: subtle side effect of commit a1c48bb160f836

[Michal Marek <mmarek@suse.cz>]

On Thu, Jun 18, 2015 at 09:10:30AM +0200, Geert Uytterhoeven wrote:
> Hi Vineet,
> 
> On Thu, Jun 18, 2015 at 8:47 AM, Vineet Gupta
> <Vineet.Gupta1@synopsys.com> wrote:
> > commit a1c48bb160f8368 "Makefile: Fix unrecognized cross-compiler command line
> > options" moved ARCH specific cc option handling before common -Os/O2 setup.
> >
> > For ARC this had a subtle effect that we can no longer over-ride generic -O2 with
> > -O3, hence a performance regression observed going from 3.13 to 3.18 (the above
> > commit went into 3.16)
> >
> > I want to understand how to properly fix this. Moving the include of arch makefile
> > will bring back the old issue. I can introduce another option to set default optim
> > level, but only arc/m32r care about it anyways.

m32r only sets -O for its assembler.


> Can we include $(srctree)/arch/$(SRCARCH)/Makefile twice?

Please don't, the gcc commandline is long enough already.


> Or perhaps we can not apply the extra -O* if there's already a -O* option?
> 
> Alternatively, as we already have CONFIG_CC_OPTIMIZE_FOR_SIZE,
> a(nother) Kconfig option may make sense.

We can also introduce some ARCH_CFLAGS that is appended near the end of
the list, and have arc/Makefile add its -O3 there. But I'd like to why
the -O3 needs to be there in first place. Obviously, the kernel works
with -O2, otherwise the regression would have been identified earlier.
So why can't users specify -O3 in KCFLAGS like on any other
architecture.

Michal