Re: [PATCH] arm64: drop CROSS_COMPILE for LLVM=1 LLVM_IAS=1

[Nick Desaulniers <ndesaulniers@google.com>]

On Sat, Jul 3, 2021 at 5:47 PM Nathan Chancellor <nathan@kernel.org> wrote:
>
> On Fri, Jul 02, 2021 at 11:29:31AM -0700, Nick Desaulniers wrote:
> > On Fri, Jul 2, 2021 at 4:59 AM Arnd Bergmann <arnd@kernel.org> wrote:
> > >
> > > On Fri, Jul 2, 2021 at 1:55 AM 'Nick Desaulniers' via Clang Built
> > > Linux <clang-built-linux@googlegroups.com> wrote:
> > > >
> > > > +ifneq ($(LLVM),)
> > > > +ifneq ($(LLVM_IAS),)
> > > > +ifeq ($(CROSS_COMPILE),)
> > > > +CLANG_TARGET   :=--target=aarch64-linux
> > > > +CLANG_FLAGS    += $(CLANG_TARGET)
> > > > +KBUILD_CFLAGS  += $(CLANG_TARGET)
> > > > +KBUILD_AFLAGS  += $(CLANG_TARGET)
> > > > +endif
> > > > +endif
> > > > +endif
> > >
> > > I think only the "CLANG_TARGET   :=--target=aarch64-linux" line should
> > > go into the
> > > per-architecture Makefile. It doesn't hurt to just set that
> > > unconditionally here,
> > > and then change the CLANG_FLAGS logic in the top-level Makefile to use this
> > > in place of $(notdir $(CROSS_COMPILE:%-=%)).
> >
> > I don't think we can do that. Based on the order the arch/ specific
> > Makefiles are included, if we don't eagerly add --target to the
> > KBUILD_{C|A}FLAGS, then cc-option, as-option, and as-instr macros
> > (defined in scripts/Makefile.compiler) checks in per arch/ Makefiles
> > may fail erroneously because --target was not set for
> > KBUILD_{C|A}FLAGS yet.
> >
> > Another issue is the order of operations between the top level
> > Makefile and the per arch/ Makefiles.  The `notdir` block you
> > reference occurs earlier than the per-arch includes:
> >
> >  609 TENTATIVE_CLANG_FLAGS += --target=$(notdir $(CROSS_COMPILE:%-=%))
> > ...
> >  648 include $(srctree)/arch/$(SRCARCH)/Makefile
> >
> > We would need the opposite order to do what you describe. Reordering
> > these would effectively be a revert of
> > commit ae6b289a3789 ("kbuild: Set KBUILD_CFLAGS before incl. arch Makefile")
> > which I'm not sure we want to do.  But maybe there's another way I'm
> > not seeing yet?
>
> Is there any reason we cannot just add this sort of logic to the main
> Makefile?
>
> Such as (indentation to emphasis diff):
>
> ifeq ($(CROSS_COMPILE),)
> ifneq ($(LLVM),)
> ifeq ($(LLVM_IAS),1)
>         ifeq ($(ARCH),arm64)
>                 TENTATIVE_CLANG_FLAGS   += --target=aarch64-linux
>         else ifeq ($(ARCH),s390)
>                 TENTATIVE_CLANG_FLAGS   += --target=s390x-linux
>         else ifeq ($(ARCH),x86_64)
>                 TENTATIVE_CLANG_FLAGS   += --target=x86_64-linux
>         else
>                 $(error Specify CROSS_COMPILE or add '--target=' option to Makefile)
>         endif
> endif
> endif
> else
> TENTATIVE_CLANG_FLAGS   += --target=$(notdir $(CROSS_COMPILE:%-=%))
> ifeq ($(LLVM_IAS),1)
> TENTATIVE_CLANG_FLAGS   += -integrated-as
> else
> TENTATIVE_CLANG_FLAGS   += -no-integrated-as
> GCC_TOOLCHAIN_DIR := $(dir $(shell which $(CROSS_COMPILE)elfedit))
> TENTATIVE_CLANG_FLAGS   += --prefix=$(GCC_TOOLCHAIN_DIR)$(notdir $(CROSS_COMPILE))
> endif
> endif
>
> I know this looks a little cumbersome but it does help us avoid
> duplication across architecture Makefiles and ordering dependencies.

Yeah, ok.

I like the use of `include` to compartmentalize the top level Makefile
further.  We can move this whole block of LLVM related flag handling
into something under scripts, then add this block and it doesn't look
too bad IMO.  Masahiro, are you ok with that?  If so, I'd break this
into 2 patches:
1. moving this block of existing code into a new file.
2. adding the CROSS_COMPILE functionality.

See https://groups.google.com/g/clang-built-linux/c/s-voh6WQFxM for
the gist of what I was thinking (though not broken into 2 patches yet,
just testing that it works; it does).

This approach will collide with Miguel's series in -next.  Should I
base the patches on mainline, or linux-kbuild, then have Miguel rebase
his patches on that or what?
-- 
Thanks,
~Nick Desaulniers