Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20210701235505.1792711-1-ndesaulniers@google.com> <20210702112210.GA11084@willie-the-truck>
In-Reply-To: <20210702112210.GA11084@willie-the-truck>
From:   Nick Desaulniers <ndesaulniers@google.com>
Date:   Fri, 2 Jul 2021 10:50:27 -0700
Message-ID: <CAKwvOdk4GLDVrOJsxPBVM+g4sBiYW-64M3rNN6hsBgHP0eYqGg@mail.gmail.com>
Subject: Re: [PATCH] arm64: drop CROSS_COMPILE for LLVM=1 LLVM_IAS=1
To:     Will Deacon <will@kernel.org>
Cc:     Catalin Marinas <catalin.marinas@arm.com>,
        Masahiro Yamada <masahiroy@kernel.org>,
        Arnd Bergmann <arnd@kernel.org>,
        Fangrui Song <maskray@google.com>,
        Nathan Chancellor <nathan@kernel.org>,
        linux-arm-kernel@lists.infradead.org, linux-kernel@vger.kernel.org,
        clang-built-linux@googlegroups.com
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Fri, Jul 2, 2021 at 4:22 AM Will Deacon <will@kernel.org> wrote:
>
> On Thu, Jul 01, 2021 at 04:55:05PM -0700, Nick Desaulniers wrote:
> > diff --git a/arch/arm64/Makefile b/arch/arm64/Makefile
> > index 7bc37d0a1b68..016873fddcc3 100644
> > --- a/arch/arm64/Makefile
> > +++ b/arch/arm64/Makefile
> > @@ -34,6 +34,17 @@ $(warning LSE atomics not supported by binutils)
> >    endif
> >  endif
> >
> > +ifneq ($(LLVM),)
> > +ifneq ($(LLVM_IAS),)
> > +ifeq ($(CROSS_COMPILE),)
> > +CLANG_TARGET :=--target=aarch64-linux
> > +CLANG_FLAGS  += $(CLANG_TARGET)
> > +KBUILD_CFLAGS        += $(CLANG_TARGET)
> > +KBUILD_AFLAGS        += $(CLANG_TARGET)
>
> Do we need to do anything extra for the linker here? I can't see how we
> avoid picking up the host copy.

That's handled by the top level Makefile when LLVM=1 is set.

There is $KBUILD_LDFLAGS, but we don't do anything with it at the
moment in terms of which linker we select; $LD controls which linker
we use.

LLD can figure out the target based on the object files it's given as
input, so it doesn't need any `--target=` flag. When clang is invoked
as the compiler or assembler, it does need --target.

> Have you tested the compat vDSO with this change? I think we'll just end
> up passing two --target options, which is hopefully ok, but thought I'd
> better check.

Good catch.  We don't reuse KBUILD_CFLAGS or KBUILD_AFLAGS for the
compat vdso for this very reason. In arch/arm64/kernel/vdso32/Makefile
you'll see no references to KBUILD_CFLAGS or KBUILD_AFLAGS; instead we
use VDSO_CFLAGS and VDSO_AFLAGS in their stead.

But, we could (and should) make this same change for the compat vdso,
and drop the need for CROSS_COMPILE_COMPAT for LLVM.

Let me play around with the changes Arnd suggested and see if I can
get that working.  I'm a bit nervous about making this depend on
something from the top level Makefile on initial glance; these changes
start to become tree wide rather than isolated per arch/, but let's
see.  Maybe at that point we carry a series in the kbuild tree with
acks for the arch/ specific changes from the respective maintainers?

Either way, I'll send a v2 that nixes CROSS_COMPILE_COMPAT for LLVM.
-- 
Thanks,
~Nick Desaulniers