Re: [PATCH V4 14/14] LoongArch: Adjust build infrastructure for 32BIT/64BIT

[Nathan Chancellor]

Hi Rui,

On Fri, Apr 24, 2026 at 07:06:23PM +0800, WANG Rui wrote:
> > As a result of the above problem, I see
> > 
> >   clang: error: ignoring '-mabi=ilp32s' as it conflicts with that implied by '-msoft-float' (lp64s) [-Werror,-Woption-ignored]
> >   clang: error: ignoring '-mabi=ilp32s' as it conflicts with that implied by '-msoft-float' (lp64s) [-Werror,-Woption-ignored]
> >   clang: error: ignoring '-mabi=ilp32s' as it conflicts with that implied by '-msoft-float' (lp64s) [-Werror,-Woption-ignored]
> > 
> > when building configurations that I expected to be 64-bit but had been
> > turned into 32-bit ones with LLVM. This was also reported by the test
> > robot with allnoconfig.
> > 
> >   https://lore.kernel.org/202604232041.ESJDwVG4-lkp@xxxxxxxxx/
> > 
> > Maybe some change is needed on the LLVM side?
> 
> The root cause is that scripts/Makefile.clang unconditionally sets
> CLANG_TARGET_FLAGS_loongarch to loongarch64-linux-gnusf.
> However, Clang with --target=loongarch64 does not support a 32-bit ABI.
> LLVM clearly separates loongarch32 and loongarch64 targets.
> 
> Since CLANG_TARGET_FLAGS gets initialized before CONFIG is available, the
> current setup effectively assumes that a 64-bit target must also handle
> a 32-bit ABI. That assumption may not hold for LLVM, which models 32-bit
> and 64-bit LoongArch as separate targets.

Hmmm, okay, that definitely explains it. As the comment at the top of
scripts/Makefile.clang states, we rely on targets supporting switching
their word size via '-m32' / '-m64' so that we can use a 64-bit triple
but still get a 32-bit one if necessary. For example:

  --target=x86_64-linux-gnu -m32

effectively becomes

  --target=i386-linux-gnu

It seems like this works for LoongArch in clang as well?

diff --git a/arch/loongarch/Makefile b/arch/loongarch/Makefile
index 47516aeea9d2..2837004d0d75 100644
--- a/arch/loongarch/Makefile
+++ b/arch/loongarch/Makefile
@@ -67,6 +67,9 @@ ifneq ($(SUBARCH),$(ARCH))
 endif
 
 ifdef CONFIG_32BIT
+ifdef CONFIG_CC_IS_CLANG
+cflags-y		+= -m32
+endif
 ifdef CONFIG_32BIT_STANDARD
 ld-emul			= $(32bit-emul)
 cflags-y		+= -march=la32v1.0 -mabi=ilp32s -mcmodel=normal
diff --git a/arch/loongarch/vdso/Makefile b/arch/loongarch/vdso/Makefile
index 42aa96249828..eeda13ecb711 100644
--- a/arch/loongarch/vdso/Makefile
+++ b/arch/loongarch/vdso/Makefile
@@ -14,6 +14,7 @@ ccflags-vdso := \
 	$(filter -E%,$(KBUILD_CFLAGS)) \
 	$(filter -march=%,$(KBUILD_CFLAGS)) \
 	$(filter -m%-float,$(KBUILD_CFLAGS)) \
+	$(filter -m32,$(KBUILD_CFLAGS)) \
 	$(CLANG_FLAGS) \
 	-D__VDSO__
 
--

appears to do the right thing?

  $ make -skj"$(nproc)" ARCH=loongarch LLVM=1 mrproper loongarch32_defconfig all

  $ file vmlinux arch/loongarch/boot/vmlinux.efi
  vmlinux:                         ELF 32-bit LSB executable, LoongArch, version 1 (SYSV), statically linked, BuildID[sha1]=26b4af26bb14c923270429f4ccdc66a519c059b1, not stripped
  arch/loongarch/boot/vmlinux.efi: PE32+ executable for EFI (application), LoongArch 32-bit (stripped to external PDB), 2 section

Cheers,
Nathan