| Message ID | 20230624184055.3000636-8-kernel@xen0n.name (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | LoongArch: Preliminary ClangBuiltLinux enablement | expand |
On Sun, 2023-06-25 at 02:40 +0800, WANG Xuerui wrote: > From: WANG Xuerui <git@xen0n.name> > > Now the arch code is mostly ready for LLVM/Clang consumption, it is > time > to re-organize the CFLAGS a little to actually enable the LLVM build. > > In particular, -mexplicit-relocs and -mdirect-extern-access are not > necessary nor supported on Clang; feature detection via cc-option > would > not work, because that way the broken combo of "new GNU as + old GCC" > would seem to get "fixed", but actually produce broken kernels. > Explicitly depending on CONFIG_CC_IS_CLANG is thus necessary to not > regress UX for those building their own kernels. > > A build with !RELOCATABLE && !MODULE is confirmed working within a > QEMU > environment; support for the two features are currently blocked on > LLVM/Clang, and will come later. > > Signed-off-by: WANG Xuerui <git@xen0n.name> > --- > arch/loongarch/Makefile | 6 +++++- > 1 file changed, 5 insertions(+), 1 deletion(-) > > diff --git a/arch/loongarch/Makefile b/arch/loongarch/Makefile > index 366771016b99..82c619791a63 100644 > --- a/arch/loongarch/Makefile > +++ b/arch/loongarch/Makefile > @@ -51,7 +51,9 @@ LDFLAGS_vmlinux += -static -n > -nostdlib > > # When the assembler supports explicit relocation hint, we must use it. > # GCC may have -mexplicit-relocs off by default if it was built with an old > -# assembler, so we force it via an option. > +# assembler, so we force it via an option. For LLVM/Clang the desired behavior > +# is the default Actually -mdirect-extern-access is also a part of the desired behavior. Though it's not strictly needed, it is a micro-optimization to avoid unnecessary GOT accesses. Maybe with Clang the LTO pass can optimize them away but I'm not sure. Ok with the flags change (for now) but the comment can be more accurate. > , and the flag is not supported, so don't pass it if Clang is > +# being used. > # > # When the assembler does not supports explicit relocation hint, we can't use > # it. Disable it if the compiler supports it. > @@ -61,8 +63,10 @@ LDFLAGS_vmlinux += -static -n -nostdlib > # combination of a "new" assembler and "old" compiler is not supported. Either > # upgrade the compiler or downgrade the assembler. > ifdef CONFIG_AS_HAS_EXPLICIT_RELOCS > +ifndef CONFIG_CC_IS_CLANG > cflags-y += -mexplicit-relocs > KBUILD_CFLAGS_KERNEL += -mdirect-extern-access > +endif > else > cflags-y += $(call cc-option,-mno-explicit-relocs) > KBUILD_AFLAGS_KERNEL += -Wa,-mla-global-with-pcrel
Hi, Ruoyao, On Sun, Jun 25, 2023 at 2:42 AM WANG Xuerui <kernel@xen0n.name> wrote: > > From: WANG Xuerui <git@xen0n.name> > > Now the arch code is mostly ready for LLVM/Clang consumption, it is time > to re-organize the CFLAGS a little to actually enable the LLVM build. > > In particular, -mexplicit-relocs and -mdirect-extern-access are not > necessary nor supported on Clang; feature detection via cc-option would > not work, because that way the broken combo of "new GNU as + old GCC" > would seem to get "fixed", but actually produce broken kernels. > Explicitly depending on CONFIG_CC_IS_CLANG is thus necessary to not > regress UX for those building their own kernels. > > A build with !RELOCATABLE && !MODULE is confirmed working within a QEMU > environment; support for the two features are currently blocked on > LLVM/Clang, and will come later. > > Signed-off-by: WANG Xuerui <git@xen0n.name> > --- > arch/loongarch/Makefile | 6 +++++- > 1 file changed, 5 insertions(+), 1 deletion(-) > > diff --git a/arch/loongarch/Makefile b/arch/loongarch/Makefile > index 366771016b99..82c619791a63 100644 > --- a/arch/loongarch/Makefile > +++ b/arch/loongarch/Makefile > @@ -51,7 +51,9 @@ LDFLAGS_vmlinux += -static -n -nostdlib > > # When the assembler supports explicit relocation hint, we must use it. > # GCC may have -mexplicit-relocs off by default if it was built with an old > -# assembler, so we force it via an option. > +# assembler, so we force it via an option. For LLVM/Clang the desired behavior > +# is the default, and the flag is not supported, so don't pass it if Clang is > +# being used. > # > # When the assembler does not supports explicit relocation hint, we can't use > # it. Disable it if the compiler supports it. > @@ -61,8 +63,10 @@ LDFLAGS_vmlinux += -static -n -nostdlib > # combination of a "new" assembler and "old" compiler is not supported. Either > # upgrade the compiler or downgrade the assembler. > ifdef CONFIG_AS_HAS_EXPLICIT_RELOCS > +ifndef CONFIG_CC_IS_CLANG > cflags-y += -mexplicit-relocs > KBUILD_CFLAGS_KERNEL += -mdirect-extern-access > +endif I prefer to drop CONFIG_CC_IS_CLANG and use cflags-y += $(call cc-option,-mexplicit-relocs) KBUILD_CFLAGS_KERNEL += $(call cc-option,-mdirect-extern-access) Then Patch-6 can be merged in this. What's your opinion? Huacai > else > cflags-y += $(call cc-option,-mno-explicit-relocs) > KBUILD_AFLAGS_KERNEL += -Wa,-mla-global-with-pcrel > -- > 2.40.0 >
On 2023/6/25 10:13, Huacai Chen wrote: > Hi, Ruoyao, > > On Sun, Jun 25, 2023 at 2:42 AM WANG Xuerui <kernel@xen0n.name> wrote: >> >> From: WANG Xuerui <git@xen0n.name> >> >> Now the arch code is mostly ready for LLVM/Clang consumption, it is time >> to re-organize the CFLAGS a little to actually enable the LLVM build. >> >> In particular, -mexplicit-relocs and -mdirect-extern-access are not >> necessary nor supported on Clang; feature detection via cc-option would >> not work, because that way the broken combo of "new GNU as + old GCC" >> would seem to get "fixed", but actually produce broken kernels. >> Explicitly depending on CONFIG_CC_IS_CLANG is thus necessary to not >> regress UX for those building their own kernels. >> >> A build with !RELOCATABLE && !MODULE is confirmed working within a QEMU >> environment; support for the two features are currently blocked on >> LLVM/Clang, and will come later. >> >> Signed-off-by: WANG Xuerui <git@xen0n.name> >> --- >> arch/loongarch/Makefile | 6 +++++- >> 1 file changed, 5 insertions(+), 1 deletion(-) >> >> diff --git a/arch/loongarch/Makefile b/arch/loongarch/Makefile >> index 366771016b99..82c619791a63 100644 >> --- a/arch/loongarch/Makefile >> +++ b/arch/loongarch/Makefile >> @@ -51,7 +51,9 @@ LDFLAGS_vmlinux += -static -n -nostdlib >> >> # When the assembler supports explicit relocation hint, we must use it. >> # GCC may have -mexplicit-relocs off by default if it was built with an old >> -# assembler, so we force it via an option. >> +# assembler, so we force it via an option. For LLVM/Clang the desired behavior >> +# is the default, and the flag is not supported, so don't pass it if Clang is >> +# being used. >> # >> # When the assembler does not supports explicit relocation hint, we can't use >> # it. Disable it if the compiler supports it. >> @@ -61,8 +63,10 @@ LDFLAGS_vmlinux += -static -n -nostdlib >> # combination of a "new" assembler and "old" compiler is not supported. Either >> # upgrade the compiler or downgrade the assembler. >> ifdef CONFIG_AS_HAS_EXPLICIT_RELOCS >> +ifndef CONFIG_CC_IS_CLANG >> cflags-y += -mexplicit-relocs >> KBUILD_CFLAGS_KERNEL += -mdirect-extern-access >> +endif > I prefer to drop CONFIG_CC_IS_CLANG and use > cflags-y += $(call cc-option,-mexplicit-relocs) > KBUILD_CFLAGS_KERNEL += $(call cc-option,-mdirect-extern-access) > > Then Patch-6 can be merged in this. > > What's your opinion? FYI: with this approach the build no longer instantly dies with binutils 2.40 + gcc 12.3, but there are also tons of warnings that say the model attribute is being ignored. I checked earlier discussions and this means modules are silently broken at runtime, which is not particularly good UX. But after more thought, I find it possible to not regress UX nor explicitly check for Clang: the special point about Clang is that it emits explicit relocs *without* support for the CLI -mexplicit-relocs flag. So assuming: * CC_HAS_EXPLICIT_RELOCS means "-mexplicit-relocs" is supported by $CC, * CC_EMITS_EXPLICIT_RELOCS means explicit relocs are emitted by $CC, with -mexplicit-relocs in CFLAGS if supported, We then have: * gcc 12.x: !CC_HAS_EXPLICIT_RELOCS && !CC_EMITS_EXPLICIT_RELOCS * gcc 13.x: CC_HAS_EXPLICIT_RELOCS && CC_EMITS_EXPLICIT_RELOCS * clang: !CC_HAS_EXPLICIT_RELOCS && CC_EMITS_EXPLICIT_RELOCS So in this case (inside the AS_HAS_EXPLICIT_RELOCS=y block), as long as CC_EMITS_EXPLICIT_RELOCS we are okay. We can $(error) out in the first case and provide helpful diagnostics too. What do you people think about this alternative approach?
On Sun, 2023-06-25 at 15:16 +0800, WANG Xuerui wrote: > On 2023/6/25 10:13, Huacai Chen wrote: > > Hi, Ruoyao, > > > > On Sun, Jun 25, 2023 at 2:42 AM WANG Xuerui <kernel@xen0n.name> wrote: > > > > > > From: WANG Xuerui <git@xen0n.name> > > > > > > Now the arch code is mostly ready for LLVM/Clang consumption, it is time > > > to re-organize the CFLAGS a little to actually enable the LLVM build. > > > > > > In particular, -mexplicit-relocs and -mdirect-extern-access are not > > > necessary nor supported on Clang; feature detection via cc-option would > > > not work, because that way the broken combo of "new GNU as + old GCC" > > > would seem to get "fixed", but actually produce broken kernels. > > > Explicitly depending on CONFIG_CC_IS_CLANG is thus necessary to not > > > regress UX for those building their own kernels. > > > > > > A build with !RELOCATABLE && !MODULE is confirmed working within a QEMU > > > environment; support for the two features are currently blocked on > > > LLVM/Clang, and will come later. > > > > > > Signed-off-by: WANG Xuerui <git@xen0n.name> > > > --- > > > arch/loongarch/Makefile | 6 +++++- > > > 1 file changed, 5 insertions(+), 1 deletion(-) > > > > > > diff --git a/arch/loongarch/Makefile b/arch/loongarch/Makefile > > > index 366771016b99..82c619791a63 100644 > > > --- a/arch/loongarch/Makefile > > > +++ b/arch/loongarch/Makefile > > > @@ -51,7 +51,9 @@ LDFLAGS_vmlinux += -static -n -nostdlib > > > > > > # When the assembler supports explicit relocation hint, we must use it. > > > # GCC may have -mexplicit-relocs off by default if it was built with an old > > > -# assembler, so we force it via an option. > > > +# assembler, so we force it via an option. For LLVM/Clang the desired behavior > > > +# is the default, and the flag is not supported, so don't pass it if Clang is > > > +# being used. > > > # > > > # When the assembler does not supports explicit relocation hint, we can't use > > > # it. Disable it if the compiler supports it. > > > @@ -61,8 +63,10 @@ LDFLAGS_vmlinux += -static -n -nostdlib > > > # combination of a "new" assembler and "old" compiler is not supported. Either > > > # upgrade the compiler or downgrade the assembler. > > > ifdef CONFIG_AS_HAS_EXPLICIT_RELOCS > > > +ifndef CONFIG_CC_IS_CLANG > > > cflags-y += -mexplicit-relocs > > > KBUILD_CFLAGS_KERNEL += -mdirect-extern-access > > > +endif > > I prefer to drop CONFIG_CC_IS_CLANG and use > > cflags-y += $(call cc-option,-mexplicit-relocs) > > KBUILD_CFLAGS_KERNEL += $(call cc-option,-mdirect-extern-access) > > > > Then Patch-6 can be merged in this. > > > > What's your opinion? > > FYI: with this approach the build no longer instantly dies with binutils > 2.40 + gcc 12.3, but there are also tons of warnings that say the model > attribute is being ignored. I checked earlier discussions and this means > modules are silently broken at runtime, which is not particularly good UX. We can add #if defined(MODULE) && !__has_attribute(model) # error some fancy error message #endif into percpu.h to error out in this case. It had been in my earlier drafts of explicit relocs patches, but we dropped it because there was no such configuration (unless a snapshot of development GCC is used, and using such a snapshot is never supported IIUC).
On 2023/6/25 15:36, Xi Ruoyao wrote: > On Sun, 2023-06-25 at 15:16 +0800, WANG Xuerui wrote: >> On 2023/6/25 10:13, Huacai Chen wrote: >>> Hi, Ruoyao, >>> >>> On Sun, Jun 25, 2023 at 2:42 AM WANG Xuerui <kernel@xen0n.name> wrote: >>>> >>>> From: WANG Xuerui <git@xen0n.name> >>>> >>>> Now the arch code is mostly ready for LLVM/Clang consumption, it is time >>>> to re-organize the CFLAGS a little to actually enable the LLVM build. >>>> >>>> In particular, -mexplicit-relocs and -mdirect-extern-access are not >>>> necessary nor supported on Clang; feature detection via cc-option would >>>> not work, because that way the broken combo of "new GNU as + old GCC" >>>> would seem to get "fixed", but actually produce broken kernels. >>>> Explicitly depending on CONFIG_CC_IS_CLANG is thus necessary to not >>>> regress UX for those building their own kernels. >>>> >>>> A build with !RELOCATABLE && !MODULE is confirmed working within a QEMU >>>> environment; support for the two features are currently blocked on >>>> LLVM/Clang, and will come later. >>>> >>>> Signed-off-by: WANG Xuerui <git@xen0n.name> >>>> --- >>>> arch/loongarch/Makefile | 6 +++++- >>>> 1 file changed, 5 insertions(+), 1 deletion(-) >>>> >>>> diff --git a/arch/loongarch/Makefile b/arch/loongarch/Makefile >>>> index 366771016b99..82c619791a63 100644 >>>> --- a/arch/loongarch/Makefile >>>> +++ b/arch/loongarch/Makefile >>>> @@ -51,7 +51,9 @@ LDFLAGS_vmlinux += -static -n -nostdlib >>>> >>>> # When the assembler supports explicit relocation hint, we must use it. >>>> # GCC may have -mexplicit-relocs off by default if it was built with an old >>>> -# assembler, so we force it via an option. >>>> +# assembler, so we force it via an option. For LLVM/Clang the desired behavior >>>> +# is the default, and the flag is not supported, so don't pass it if Clang is >>>> +# being used. >>>> # >>>> # When the assembler does not supports explicit relocation hint, we can't use >>>> # it. Disable it if the compiler supports it. >>>> @@ -61,8 +63,10 @@ LDFLAGS_vmlinux += -static -n -nostdlib >>>> # combination of a "new" assembler and "old" compiler is not supported. Either >>>> # upgrade the compiler or downgrade the assembler. >>>> ifdef CONFIG_AS_HAS_EXPLICIT_RELOCS >>>> +ifndef CONFIG_CC_IS_CLANG >>>> cflags-y += -mexplicit-relocs >>>> KBUILD_CFLAGS_KERNEL += -mdirect-extern-access >>>> +endif >>> I prefer to drop CONFIG_CC_IS_CLANG and use >>> cflags-y += $(call cc-option,-mexplicit-relocs) >>> KBUILD_CFLAGS_KERNEL += $(call cc-option,-mdirect-extern-access) >>> >>> Then Patch-6 can be merged in this. >>> >>> What's your opinion? >> >> FYI: with this approach the build no longer instantly dies with binutils >> 2.40 + gcc 12.3, but there are also tons of warnings that say the model >> attribute is being ignored. I checked earlier discussions and this means >> modules are silently broken at runtime, which is not particularly good UX. > > We can add > > #if defined(MODULE) && !__has_attribute(model) > # error some fancy error message > #endif > > into percpu.h to error out in this case. It had been in my earlier > drafts of explicit relocs patches, but we dropped it because there was > no such configuration (unless a snapshot of development GCC is used, and > using such a snapshot is never supported IIUC). Ah I've seen that. So in this case we simply wrap -mexplicit-relocs with cc-option and error out in case of CONFIG_MODULE but no model attribute, which nicely prevents broken configurations (MODULE && ((old_gcc && new_binutils) || clang)) with feature detection alone. This seems elegant and better to me; Huacai, WDYT?
On Sun, Jun 25, 2023 at 3:48 PM WANG Xuerui <kernel@xen0n.name> wrote: > > On 2023/6/25 15:36, Xi Ruoyao wrote: > > On Sun, 2023-06-25 at 15:16 +0800, WANG Xuerui wrote: > >> On 2023/6/25 10:13, Huacai Chen wrote: > >>> Hi, Ruoyao, > >>> > >>> On Sun, Jun 25, 2023 at 2:42 AM WANG Xuerui <kernel@xen0n.name> wrote: > >>>> > >>>> From: WANG Xuerui <git@xen0n.name> > >>>> > >>>> Now the arch code is mostly ready for LLVM/Clang consumption, it is time > >>>> to re-organize the CFLAGS a little to actually enable the LLVM build. > >>>> > >>>> In particular, -mexplicit-relocs and -mdirect-extern-access are not > >>>> necessary nor supported on Clang; feature detection via cc-option would > >>>> not work, because that way the broken combo of "new GNU as + old GCC" > >>>> would seem to get "fixed", but actually produce broken kernels. > >>>> Explicitly depending on CONFIG_CC_IS_CLANG is thus necessary to not > >>>> regress UX for those building their own kernels. > >>>> > >>>> A build with !RELOCATABLE && !MODULE is confirmed working within a QEMU > >>>> environment; support for the two features are currently blocked on > >>>> LLVM/Clang, and will come later. > >>>> > >>>> Signed-off-by: WANG Xuerui <git@xen0n.name> > >>>> --- > >>>> arch/loongarch/Makefile | 6 +++++- > >>>> 1 file changed, 5 insertions(+), 1 deletion(-) > >>>> > >>>> diff --git a/arch/loongarch/Makefile b/arch/loongarch/Makefile > >>>> index 366771016b99..82c619791a63 100644 > >>>> --- a/arch/loongarch/Makefile > >>>> +++ b/arch/loongarch/Makefile > >>>> @@ -51,7 +51,9 @@ LDFLAGS_vmlinux += -static -n -nostdlib > >>>> > >>>> # When the assembler supports explicit relocation hint, we must use it. > >>>> # GCC may have -mexplicit-relocs off by default if it was built with an old > >>>> -# assembler, so we force it via an option. > >>>> +# assembler, so we force it via an option. For LLVM/Clang the desired behavior > >>>> +# is the default, and the flag is not supported, so don't pass it if Clang is > >>>> +# being used. > >>>> # > >>>> # When the assembler does not supports explicit relocation hint, we can't use > >>>> # it. Disable it if the compiler supports it. > >>>> @@ -61,8 +63,10 @@ LDFLAGS_vmlinux += -static -n -nostdlib > >>>> # combination of a "new" assembler and "old" compiler is not supported. Either > >>>> # upgrade the compiler or downgrade the assembler. > >>>> ifdef CONFIG_AS_HAS_EXPLICIT_RELOCS > >>>> +ifndef CONFIG_CC_IS_CLANG > >>>> cflags-y += -mexplicit-relocs > >>>> KBUILD_CFLAGS_KERNEL += -mdirect-extern-access > >>>> +endif > >>> I prefer to drop CONFIG_CC_IS_CLANG and use > >>> cflags-y += $(call cc-option,-mexplicit-relocs) > >>> KBUILD_CFLAGS_KERNEL += $(call cc-option,-mdirect-extern-access) > >>> > >>> Then Patch-6 can be merged in this. > >>> > >>> What's your opinion? > >> > >> FYI: with this approach the build no longer instantly dies with binutils > >> 2.40 + gcc 12.3, but there are also tons of warnings that say the model > >> attribute is being ignored. I checked earlier discussions and this means > >> modules are silently broken at runtime, which is not particularly good UX. > > > > We can add > > > > #if defined(MODULE) && !__has_attribute(model) > > # error some fancy error message > > #endif > > > > into percpu.h to error out in this case. It had been in my earlier > > drafts of explicit relocs patches, but we dropped it because there was > > no such configuration (unless a snapshot of development GCC is used, and > > using such a snapshot is never supported IIUC). > > Ah I've seen that. So in this case we simply wrap -mexplicit-relocs with > cc-option and error out in case of CONFIG_MODULE but no model attribute, > which nicely prevents broken configurations (MODULE && ((old_gcc && > new_binutils) || clang)) with feature detection alone. > > This seems elegant and better to me; Huacai, WDYT? OK, perfect. Huacai > > -- > WANG "xen0n" Xuerui > > Linux/LoongArch mailing list: https://lore.kernel.org/loongarch/ >
diff --git a/arch/loongarch/Makefile b/arch/loongarch/Makefile index 366771016b99..82c619791a63 100644 --- a/arch/loongarch/Makefile +++ b/arch/loongarch/Makefile @@ -51,7 +51,9 @@ LDFLAGS_vmlinux += -static -n -nostdlib # When the assembler supports explicit relocation hint, we must use it. # GCC may have -mexplicit-relocs off by default if it was built with an old -# assembler, so we force it via an option. +# assembler, so we force it via an option. For LLVM/Clang the desired behavior +# is the default, and the flag is not supported, so don't pass it if Clang is +# being used. # # When the assembler does not supports explicit relocation hint, we can't use # it. Disable it if the compiler supports it. @@ -61,8 +63,10 @@ LDFLAGS_vmlinux += -static -n -nostdlib # combination of a "new" assembler and "old" compiler is not supported. Either # upgrade the compiler or downgrade the assembler. ifdef CONFIG_AS_HAS_EXPLICIT_RELOCS +ifndef CONFIG_CC_IS_CLANG cflags-y += -mexplicit-relocs KBUILD_CFLAGS_KERNEL += -mdirect-extern-access +endif else cflags-y += $(call cc-option,-mno-explicit-relocs) KBUILD_AFLAGS_KERNEL += -Wa,-mla-global-with-pcrel