ABI 'o32' is not supported on CPU 'mips64r2'

[tpimh]

Both CONFIG_MIPS32_O32=y and CONFIG_CPU_MIPSR2=y can not be enabled at the same time.

As this comment suggests this should be fixed: https://github.com/ClangBuiltLinux/llvm-project/blob/cccf1ef0c89432339363fc0121e3d97f5a04ac69/clang/test/Driver/mips-abi.c#L12

[tpimh]

From clang source code quoted in #885 (comment):

  // FIXME: It's valid to use O32 on a 64-bit CPU but the backend can't handle
  //        this yet. It's better to fail here than on the backend assertion.
  if (processorSupportsGPR64() && ABI == "o32") {
    Diags.Report(diag::err_target_unsupported_abi) << ABI << CPU;
    return false;
  }

  // FIXME: It's valid to use O32 on a mips64/mips64el triple but the backend
  //        can't handle this yet. It's better to fail here than on the
  //        backend assertion.
  if (getTriple().isMIPS64() && ABI == "o32") {
    Diags.Report(diag::err_target_unsupported_abi_for_triple)
        << ABI << getTriple().str();
    return false;
  }

[nathanchance]

Upstream LLVM issue: https://llvm.org/pr38063

Patch disabling CONFIG_MIPS32_O32 for Clang: https://lore.kernel.org/r/1614838778-3696-1-git-send-email-yangtiezhu@loongson.cn/

[nathanchance]

Upstream thread with some addition info from Arnd: https://lore.kernel.org/CADVatmNtFYyqk0KwYzB1x6hOPkQ=0gV8mPN1BWp6QWrCs5GaxA@mail.gmail.com/

[nickdesaulniers]

@ardbiesheuvel came up with:

diff --git a/arch/mips/Makefile b/arch/mips/Makefile
index bb236de13133..d7b7451d055f 100644
--- a/arch/mips/Makefile
+++ b/arch/mips/Makefile
@@ -409,7 +409,7 @@ ifdef CONFIG_MIPS32_N32
 endif
 ifdef CONFIG_MIPS32_O32
        @$(kecho) '  Checking missing-syscalls for O32'
-       $(Q)$(MAKE) $(build)=. missing-syscalls missing_syscalls_flags="-mabi=32"
+       $(Q)$(MAKE) $(build)=. missing-syscalls missing_syscalls_flags="--target=mips-linux"
 endif
 
 install:
diff --git a/arch/mips/vdso/Makefile b/arch/mips/vdso/Makefile
index f72658b3a53f..2fb69d606547 100644
--- a/arch/mips/vdso/Makefile
+++ b/arch/mips/vdso/Makefile
@@ -141,8 +141,8 @@ ifdef CONFIG_MIPS32_O32
 
 obj-vdso-o32 := $(obj-vdso-y:%.o=$(obj)/%-o32.o)
 
-$(obj-vdso-o32): KBUILD_CFLAGS := $(cflags-vdso) -mabi=32
-$(obj-vdso-o32): KBUILD_AFLAGS := $(aflags-vdso) -mabi=32
+$(obj-vdso-o32): KBUILD_CFLAGS := $(cflags-vdso) --target=mips-linux -march=mips32r6
+$(obj-vdso-o32): KBUILD_AFLAGS := $(aflags-vdso) --target=mips-linux -march=mips32r6
 
 $(obj)/%-o32.o: $(src)/%.S FORCE
        $(call if_changed_dep,vdsoas_o_S)
@@ -151,7 +151,7 @@ $(obj)/%-o32.o: $(src)/%.c FORCE
        $(call cmd,force_checksrc)
        $(call if_changed_rule,cc_o_c)
 
-$(obj)/vdso-o32.lds: KBUILD_CPPFLAGS := $(ccflags-vdso) -mabi=32
+$(obj)/vdso-o32.lds: KBUILD_CPPFLAGS := $(ccflags-vdso) --target=mips-linux -march=mips32r6
 $(obj)/vdso-o32.lds: $(src)/vdso.lds.S FORCE
        $(call if_changed_dep,cpp_lds_S)

Seems to build. With some touch ups to avoid passing that flag to GCC will probably be upstreamable.

[nickdesaulniers]

Possible workaround: https://lore.kernel.org/llvm/20230407102721.14814-5-jiaxun.yang@flygoat.com/

[FlyGoat]

Clang patch: https://reviews.llvm.org/D146269

[nathanchance]

This is fixed in LLVM 17 with llvm/llvm-project@7983f8a. Perhaps now that this is fixed in the toolchain, the kernel will accept a workaround disabling the configuration for older toolchain versions? Otherwise, we could just leave it broken.