Hi Nick, 2017-11-10 1:51 GMT+09:00 Nick Desaulniers <ndesaulniers@xxxxxxxxxx>: > On Wed, Nov 8, 2017 at 8:31 PM, Masahiro Yamada > <yamada.masahiro@xxxxxxxxxxxxx> wrote: >> 2017-11-08 4:46 GMT+09:00 Nick Desaulniers <ndesaulniers@xxxxxxxxxx>: >>> We didn't notice this problem on Android, because we took the original >>> LLVMLinux patch into our 4.4 kernels, which did not have this issue. We >>> ran into this taking the proper upstream patch on newer kernel versions. >>> The original LLVMLinux patch can be seen at: >>> >>> http://git.linuxfoundation.org/?p=llvmlinux/kernel.git;a=blobdiff;f=Makefile;h=389006c4ef494cda3a1ee52bf355618673ab4f31;hp=e41a3356abee83f08288362950bfceebd25ec3c2;hb=ef9126da11b18ff34eb1f01561f53c378860336c;hpb=f800c25b7a762d445ba1439a2428c8362157eba6 >>> >>> It seems that when the patch was re-upstreamed, a V2 was requested that >>> moved the definition of Clang's target triple to be later in the top >>> level Makefile than the inclusion of the arch specific Makefile, >>> breaking macros like ld-option when cross compiling. V2 was requested >>> at: >>> >>> https://lkml.org/lkml/2017/4/21/116 >> >> IMO, this description is not helpful in any way in upstream. >> >> Moreover, >> 785f11aa595bc3d4e74096cbd598ada54ecc0d81 >> did not break anything. It sound unfair to me. > > Ok, I will cut that part description on v3. > >>> diff --git a/Makefile b/Makefile >>> index a7476e6934f1..d349734c7ef7 100644 >>> --- a/Makefile >>> +++ b/Makefile >>> @@ -472,6 +472,38 @@ ifneq ($(KBUILD_SRC),) >>> $(srctree) $(objtree) $(VERSION) $(PATCHLEVEL) >>> endif >>> >>> +ifeq ($(cc-name),clang) >>> +ifneq ($(CROSS_COMPILE),) >>> +CLANG_TARGET := -target $(notdir $(CROSS_COMPILE:%-=%)) >>> +GCC_TOOLCHAIN := $(realpath $(dir $(shell which $(LD)))/..) >>> +endif >>> +ifneq ($(GCC_TOOLCHAIN),) >>> +CLANG_GCC_TC := -gcc-toolchain $(GCC_TOOLCHAIN) >>> +endif >>> +KBUILD_CFLAGS += $(CLANG_TARGET) $(CLANG_GCC_TC) >>> +KBUILD_AFLAGS += $(CLANG_TARGET) $(CLANG_GCC_TC) >>> +KBUILD_CPPFLAGS += $(call cc-option,-Qunused-arguments,) >>> +KBUILD_CFLAGS += $(call cc-disable-warning, unused-variable) >>> +KBUILD_CFLAGS += $(call cc-disable-warning, format-invalid-specifier) >>> +KBUILD_CFLAGS += $(call cc-disable-warning, gnu) >>> +KBUILD_CFLAGS += $(call cc-disable-warning, address-of-packed-member) >>> +# Quiet clang warning: comparison of unsigned expression < 0 is always false >>> +KBUILD_CFLAGS += $(call cc-disable-warning, tautological-compare) >>> +# CLANG uses a _MergedGlobals as optimization, but this breaks modpost, as the >>> +# source of a reference will be _MergedGlobals and not on of the whitelisted names. >>> +# See modpost pattern 2 >>> +KBUILD_CFLAGS += $(call cc-option, -mno-global-merge,) >>> +KBUILD_CFLAGS += $(call cc-option, -fcatch-undefined-behavior) >>> +KBUILD_CFLAGS += $(call cc-option, -no-integrated-as) >>> +KBUILD_AFLAGS += $(call cc-option, -no-integrated-as) >>> +else >>> + >>> +# These warnings generated too much noise in a regular build. >>> +# Use make W=1 to enable them (see scripts/Makefile.extrawarn) >>> +KBUILD_CFLAGS += $(call cc-disable-warning, unused-but-set-variable) >>> +KBUILD_CFLAGS += $(call cc-disable-warning, unused-const-variable) >>> +endif >>> + >> >> >> Could you move this a bit later, please? >> >> >> >> # These warnings generated too much noise in a regular build. >> # Use make W=1 to enable them (see scripts/Makefile.extrawarn) >> KBUILD_CFLAGS += $(call cc-disable-warning, unused-but-set-variable) >> KBUILD_CFLAGS += $(call cc-disable-warning, unused-const-variable) >> endif >> >> << INSERT HERE >> >> >> # The arch Makefile can set ARCH_{CPP,A,C}FLAGS to override the default >> # values of the respective KBUILD_* variables >> ARCH_CPPFLAGS := >> ARCH_AFLAGS := >> ARCH_CFLAGS := >> include arch/$(SRCARCH)/Makefile >> >> >> >> >> >> >> arch/$(SRCARCH)/Makefile is included for config targets >> for KBUILD_DEFCONFIG, but no reason to compute compiler flags. >> I want to cut unnecessary cc-option parsing. > > With the new try-run-cached macros, if the arch/$(SRCARCH)/Makefile > gets included twice, with the compiler flags not set correctly for > clang to cross compile, and the results are cached, wont they be wrong > the second time the arch specific Makefile is included? > -- Good point. The cached data from arch/$(SRCARCH)/Makefile for configuration is not used for the second run. That means some garbage data in the cache file, but less than 10, I think. You do not need to give CC or CROSS_COMPILE for kernel configuration in the first place. So, make ARCH=arm64 defconfig make ARCH=arm64 CC=clang CROSS_COMPILE=aarch64-linux-gnu- should be fine. Even if we place lots of clang's cc-option earlier, users may not give the same CC for configuration and building. I think more optimized way is to skip computing cc-option for "make *config", "make clean" etc. I tried to do that. https://patchwork.kernel.org/patch/9983827/ I decided to take time for cleaner implementation, but that is what I'd like to achieve in the future. -- Best Regards Masahiro Yamada -- To unsubscribe from this list: send the line "unsubscribe linux-kbuild" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
