[RFC PATCH 1/3] ARM: iwmmxt: Fix Makefile rules for building iwmmxt for Thumb-2

[Arnd Bergmann]

On Wednesday 07 September 2011 13:18:21 Nicolas Pitre wrote:
> > > +
> > > +# When enough people have binutils which support -march=...+iwmmxt, this
> > > +# should change to something like if __LINUX_ARM_ARCH__ < 7.
> > > +ifdef CONFIG_THUMB2_KERNEL
> > > +AFLAGS_iwmmxt.o                        := -Wa,-march=armv7-a+iwmmxt
> > > +else
> > >  AFLAGS_iwmmxt.o                        := -Wa,-mcpu=iwmmxt
> > > +endif
> > 
> > It looks more like the switch should depend on the compiler version.
> > Unless there is a clear way to decide if gcc supports this switch, I
> > think it's reasonable to have the change like above.
> 
> Normally the way to go with gcc version dependent alternatives is to use 
> something like:
> 
> AFLAGS_foo.o := $(call cc-option,<the_new_flag>,<the_fallback_flag>)
> 
> This will test if <the_new_flag> is supported by the used gcc, and use 
> the fallback otherwise.

Yes, that's possible here, but it's not actually correct either, because the
CPU core that we are running on is either a v5 XScale with iwmmxt or
a v7 pj4 with iwmmxt. Now, it should not really matter if we build the
code with flags for a different more complex instruction set, but it can
potentially hide bugs.

I think the simple solution that Dave posted is actually more appropriate.
The three possible cases are:

v5+iwmmxt: always use -Wa,-mcpu=iwmmxt as we've always done, and it's correct
v7+iwmmxt+arm: still use -Wa,-mcpu=iwmmxt, not correct but close enough and
               is known to build the file with all existing toolchaings
v7+iwmmxt+thumb2: always use -Wa,-march=armv7-a+iwmmxt, which is correct and
		      the only possible way to build this file anyway. Old toolchains
		      will fail and there is nothing we can do about it.

	Arnd