We get constant feedback that the command line invocation of make is too
long. CROSS_COMPILE is helpful when a toolchain has a prefix of the
target triple, or is an absolute path outside of $PATH, but it's mostly
redundant for a given ARCH.
If CROSS_COMPILE is not set, simply set --target= for CLANG_FLAGS,
KBUILD_CFLAGS, and KBUILD_AFLAGS based on $ARCH.
Previously, we'd cross compile via:
$ ARCH=arm64 CROSS_COMPILE=aarch64-linxu-gnu make LLVM=1 LLVM_IAS=1
Now:
$ ARCH=arm64 make LLVM=1 LLVM_IAS=1
Link: https://github.com/ClangBuiltLinux/linux/issues/1399
Suggested-by: Arnd Bergmann <[email protected]>
Suggested-by: Nathan Chancellor <[email protected]>
Signed-off-by: Nick Desaulniers <[email protected]>
---
Changes RFC -> v1:
* Rebase onto linux-kbuild/for-next
* Keep full target triples since missing the gnueabi suffix messes up
32b ARM. Drop Fangrui's sugguested by tag. Update commit message to
drop references to arm64.
* Flush out TODOS.
* Add note about -EL/-EB, -m32/-m64.
* Add note to Documentation/.
Documentation/kbuild/llvm.rst | 5 +++++
scripts/Makefile.clang | 38 +++++++++++++++++++++++++++++++++--
2 files changed, 41 insertions(+), 2 deletions(-)
diff --git a/Documentation/kbuild/llvm.rst b/Documentation/kbuild/llvm.rst
index b18401d2ba82..80c63dd9a6d1 100644
--- a/Documentation/kbuild/llvm.rst
+++ b/Documentation/kbuild/llvm.rst
@@ -46,6 +46,11 @@ example: ::
clang --target=aarch64-linux-gnu foo.c
+When both ``LLVM=1`` and ``LLVM_IAS=1`` are used, ``CROSS_COMPILE`` becomes
+unnecessary and can be inferred from ``ARCH``. Example: ::
+
+ ARCH=arm64 make LLVM=1 LLVM_IAS=1
+
LLVM Utilities
--------------
diff --git a/scripts/Makefile.clang b/scripts/Makefile.clang
index 297932e973d4..a79088797a50 100644
--- a/scripts/Makefile.clang
+++ b/scripts/Makefile.clang
@@ -1,6 +1,40 @@
-ifneq ($(CROSS_COMPILE),)
+# Individual arch/{arch}/Makfiles should use -EL/-EB to set intended endianness
+# and -m32/-m64 to set word size based on Kconfigs instead of relying on the
+# target triple.
+ifeq ($(CROSS_COMPILE),)
+ifneq ($(LLVM),)
+ifeq ($(LLVM_IAS),1)
+ifeq ($(ARCH),arm)
+CLANG_FLAGS += --target=arm-linux-gnueabi
+else ifeq ($(ARCH),arm64)
+CLANG_FLAGS += --target=aarch64-linux-gnu
+else ifeq ($(ARCH),hexagon)
+CLANG_FLAGS += --target=hexagon-linux-gnu
+else ifeq ($(ARCH),i386)
+CLANG_FLAGS += --target=i686-linux-gnu
+else ifeq ($(ARCH),m68k)
+CLANG_FLAGS += --target=m68k-linux-gnu
+else ifeq ($(ARCH),mips)
+CLANG_FLAGS += --target=mipsel-linux-gnu
+else ifeq ($(ARCH),powerpc)
+CLANG_FLAGS += --target=powerpc64le-linux-gnu
+else ifeq ($(ARCH),riscv)
+CLANG_FLAGS += --target=riscv64-linux-gnu
+else ifeq ($(ARCH),s390)
+CLANG_FLAGS += --target=s390x-linux-gnu
+else ifeq ($(ARCH),x86)
+CLANG_FLAGS += --target=x86_64-linux-gnu
+else ifeq ($(ARCH),x86_64)
+CLANG_FLAGS += --target=x86_64-linux-gnu
+else
+$(error Specify CROSS_COMPILE or add '--target=' option to scripts/Makefile.clang)
+endif # ARCH
+endif # LLVM_IAS
+endif # LLVM
+else
CLANG_FLAGS += --target=$(notdir $(CROSS_COMPILE:%-=%))
-endif
+endif # CROSS_COMPILE
+
ifeq ($(LLVM_IAS),1)
CLANG_FLAGS += -integrated-as
else
--
2.32.0.93.g670b81a890-goog
Hi Nick,
On Thu, Jul 8, 2021 at 1:12 AM Nick Desaulniers <[email protected]> wrote:
> We get constant feedback that the command line invocation of make is too
> long. CROSS_COMPILE is helpful when a toolchain has a prefix of the
> target triple, or is an absolute path outside of $PATH, but it's mostly
> redundant for a given ARCH.
>
> If CROSS_COMPILE is not set, simply set --target= for CLANG_FLAGS,
> KBUILD_CFLAGS, and KBUILD_AFLAGS based on $ARCH.
>
> Previously, we'd cross compile via:
> $ ARCH=arm64 CROSS_COMPILE=aarch64-linxu-gnu make LLVM=1 LLVM_IAS=1
Which didn't really work, I assume? (s/linxu/linux/)
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
On Thu, Jul 8, 2021 at 5:08 PM Geert Uytterhoeven <[email protected]> wrote:
>
> Hi Nick,
>
> On Thu, Jul 8, 2021 at 1:12 AM Nick Desaulniers <[email protected]> wrote:
> > We get constant feedback that the command line invocation of make is too
> > long. CROSS_COMPILE is helpful when a toolchain has a prefix of the
> > target triple, or is an absolute path outside of $PATH, but it's mostly
> > redundant for a given ARCH.
> >
> > If CROSS_COMPILE is not set, simply set --target= for CLANG_FLAGS,
> > KBUILD_CFLAGS, and KBUILD_AFLAGS based on $ARCH.
> >
> > Previously, we'd cross compile via:
> > $ ARCH=arm64 CROSS_COMPILE=aarch64-linxu-gnu make LLVM=1 LLVM_IAS=1
>
> Which didn't really work, I assume? (s/linxu/linux/)
Also
s/gnu/gnu-/
--
Best Regards
Masahiro Yamada
On Thu, Jul 8, 2021 at 7:43 AM 'Nick Desaulniers' via Clang Built
Linux <[email protected]> wrote:
>
> We get constant feedback that the command line invocation of make is too
> long. CROSS_COMPILE is helpful when a toolchain has a prefix of the
> target triple, or is an absolute path outside of $PATH, but it's mostly
> redundant for a given ARCH.
>
> If CROSS_COMPILE is not set, simply set --target= for CLANG_FLAGS,
> KBUILD_CFLAGS, and KBUILD_AFLAGS based on $ARCH.
>
> Previously, we'd cross compile via:
> $ ARCH=arm64 CROSS_COMPILE=aarch64-linxu-gnu make LLVM=1 LLVM_IAS=1
> Now:
> $ ARCH=arm64 make LLVM=1 LLVM_IAS=1
>
> Link: https://github.com/ClangBuiltLinux/linux/issues/1399
> Suggested-by: Arnd Bergmann <[email protected]>
> Suggested-by: Nathan Chancellor <[email protected]>
> Signed-off-by: Nick Desaulniers <[email protected]>
> ---
> Changes RFC -> v1:
> * Rebase onto linux-kbuild/for-next
> * Keep full target triples since missing the gnueabi suffix messes up
> 32b ARM. Drop Fangrui's sugguested by tag. Update commit message to
> drop references to arm64.
> * Flush out TODOS.
> * Add note about -EL/-EB, -m32/-m64.
> * Add note to Documentation/.
>
> Documentation/kbuild/llvm.rst | 5 +++++
> scripts/Makefile.clang | 38 +++++++++++++++++++++++++++++++++--
> 2 files changed, 41 insertions(+), 2 deletions(-)
When I was considering a similar idea, my plan was
to implement this in Kconfig instead of in Makefile
because that will pass the compiler information
in one direction (only from Kconfig to Kbuild), but
that is not so important. We can change it later
if needed.
I did not complete it because I was investigating
some issues (especially [3] below), but maybe
that is something we don't care about.
Can you address [2] below at least?
If we do not have any concern, I can merge it.
It is likely so because people are only discussing
"We want to omit omit CROSS_COMPILE".
[1] explicit target triple for native builds
The default target of my distro clang
is x86_64-pc-linux-gnu.
$ clang --version
Ubuntu clang version 11.0.0-2
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/bin
So, previously, the kernel was built with
implied --target=x86_64-pc-linux-gnu.
With this patch, --target=x86_64-linux-gnu will be
explicitly specified.
The same applies to native-builds of other architectures.
For example, when a user builds the arm64 kernel on
an arm64 server, --target=aarch64-linux-gnu is
explicitly forced.
I guess, this is a good direction because the produced
code will be more deterministic, irrespective of the
Clang's configuration.
[2] 32/64-bit configuration is toggled in Kconfig time.
Initially, you submitted only arm64. Maybe, it was intentional
since arm64 is the simplest case.
In the kernel tree, arch/arm and arch/arm64 are very special
cases where 32-bit and 64-bit are separated by directory.
Some of the other architectures are bi-arch, and
32-bit/64-bit is specified by CONFIG_64BIT in Kconfig time.
When Makefiles are being parsed, we actually do not know
whether the user is planning to configure the kernel
for 32-bit or 64-bit because CONFIG_64BIT is not
specified at this point.
ARCH=x86 + CONFIG_64BIT=y
will build the x86_64 kernel, and
ARCH=x86 + CONFIG_64BIT=n
will build the i386 kernel.
Then, you may wonder
else ifeq ($(ARCH),x86)
CLANG_FLAGS += --target=x86_64-linux-gnu
... works?
Yes, it does fortunately.
-m32/-m64 takes precedence over the
{x86_64,i386} part of the target triple.
As far as I tested,
clang --target=x86_64-linux-gnu -m32
produced i386 code.
Interestingly,
clang --target=i386-linux-gnu -m64
produced x86_64 code.
We must rely on this behavior of Clang because
--target (which is contained in CLANG_FLAGS)
must be specified before the Kconfig time.
Then, a user can toggle CONFIG_64BIT any time
from menuconfig etc.
With this in mind, using $(ARCH) as if-else
switches is pointless.
$(SRCARCH) is the only meaningful input.
else ifeq ($(ARCH),i386)
CLANG_FLAGS += --target=i686-linux-gnu
else ifeq ($(ARCH),x86)
CLANG_FLAGS += --target=x86_64-linux-gnu
else ifeq ($(ARCH),x86_64)
CLANG_FLAGS += --target=x86_64-linux-gnu
should be replaced with:
else ifeq ($(SRCARCH),x86_64)
CLANG_FLAGS += --target=x86_64-linux-gnu
Some architectures are not only bi-arch, but also bi-endian.
You hardcoded 64bit little endian for ppc:
else ifeq ($(ARCH),powerpc)
CLANG_FLAGS += --target=powerpc64le-linux-gnu
But, we must rely on the fact that
clang --target=powerpc64le-linux-gnu -mbig-endian -m32
produces big-endian 32-bit code.
This makes the "64le" part meaningless.
This should be noted. Otherwise it is difficult
to understand why --target=x86_64-linux-gnu works fine
with building the i386 kernel.
[3] User-space compilation
This does not matter to the kernel itself, but
Kbuild compiles some userspace programs for
the target architecture.
See the samples/ directory for example.
Another example is net/bpfilter/Makefile, which
embeds the user mode helper when
CONFIG_BPFILTER_UMH=y.
For this purpose, Kconfig checks if $(CC) is
capable of linking the userspace.
(CONFIG_CC_CAN_LINK).
When cross-building with Clang, I cannot see
CONFIG_CC_CAN_LINK set.
If we care about CONFIG_CC_CAN_LINK, probably,
--sysroot or something should be set according to:
https://clang.llvm.org/docs/CrossCompilation.html
This is an existing issue, but I have no time
for looking into this.
On debian systems, sysroot for cross-compilation
are located in /usr/aarch64-linux-gnu,
/usr/arm-linux-gnueabi, /usr/arm-linux-gnueabihf,
/usr/i686-linux-gnu/ etc. but I do not know if it
is the same across distros.
[4] What is the best target if we hard-code it?
Currently, we require the correct CROSS_COMPILE
is provided by users.
The target might impact the performance
or the ABI.
It was difficult for me to define
which one is better than another.
For example for ARCH=arm, which is better
--target=arm-linux-gnueabi or
--target=arm-lnux-gnueabihf or
something we don't care about?
> diff --git a/Documentation/kbuild/llvm.rst b/Documentation/kbuild/llvm.rst
> index b18401d2ba82..80c63dd9a6d1 100644
> --- a/Documentation/kbuild/llvm.rst
> +++ b/Documentation/kbuild/llvm.rst
> @@ -46,6 +46,11 @@ example: ::
>
> clang --target=aarch64-linux-gnu foo.c
>
> +When both ``LLVM=1`` and ``LLVM_IAS=1`` are used, ``CROSS_COMPILE`` becomes
> +unnecessary and can be inferred from ``ARCH``. Example: ::
> +
> + ARCH=arm64 make LLVM=1 LLVM_IAS=1
> +
> LLVM Utilities
> --------------
>
> diff --git a/scripts/Makefile.clang b/scripts/Makefile.clang
> index 297932e973d4..a79088797a50 100644
> --- a/scripts/Makefile.clang
> +++ b/scripts/Makefile.clang
> @@ -1,6 +1,40 @@
> -ifneq ($(CROSS_COMPILE),)
> +# Individual arch/{arch}/Makfiles should use -EL/-EB to set intended endianness
> +# and -m32/-m64 to set word size based on Kconfigs instead of relying on the
> +# target triple.
> +ifeq ($(CROSS_COMPILE),)
> +ifneq ($(LLVM),)
> +ifeq ($(LLVM_IAS),1)
> +ifeq ($(ARCH),arm)
> +CLANG_FLAGS += --target=arm-linux-gnueabi
> +else ifeq ($(ARCH),arm64)
> +CLANG_FLAGS += --target=aarch64-linux-gnu
> +else ifeq ($(ARCH),hexagon)
> +CLANG_FLAGS += --target=hexagon-linux-gnu
> +else ifeq ($(ARCH),i386)
> +CLANG_FLAGS += --target=i686-linux-gnu
> +else ifeq ($(ARCH),m68k)
> +CLANG_FLAGS += --target=m68k-linux-gnu
> +else ifeq ($(ARCH),mips)
> +CLANG_FLAGS += --target=mipsel-linux-gnu
> +else ifeq ($(ARCH),powerpc)
> +CLANG_FLAGS += --target=powerpc64le-linux-gnu
> +else ifeq ($(ARCH),riscv)
> +CLANG_FLAGS += --target=riscv64-linux-gnu
> +else ifeq ($(ARCH),s390)
> +CLANG_FLAGS += --target=s390x-linux-gnu
> +else ifeq ($(ARCH),x86)
> +CLANG_FLAGS += --target=x86_64-linux-gnu
> +else ifeq ($(ARCH),x86_64)
> +CLANG_FLAGS += --target=x86_64-linux-gnu
> +else
> +$(error Specify CROSS_COMPILE or add '--target=' option to scripts/Makefile.clang)
> +endif # ARCH
> +endif # LLVM_IAS
> +endif # LLVM
> +else
> CLANG_FLAGS += --target=$(notdir $(CROSS_COMPILE:%-=%))
> -endif
> +endif # CROSS_COMPILE
> +
> ifeq ($(LLVM_IAS),1)
> CLANG_FLAGS += -integrated-as
> else
> --
> 2.32.0.93.g670b81a890-goog
>
> --
> You received this message because you are subscribed to the Google Groups "Clang Built Linux" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to [email protected].
> To view this discussion on the web visit https://groups.google.com/d/msgid/clang-built-linux/20210707224310.1403944-3-ndesaulniers%40google.com.
--
Best Regards
Masahiro Yamada
On Thu, Jul 8, 2021 at 12:23 PM Masahiro Yamada <[email protected]> wrote:
> On Thu, Jul 8, 2021 at 7:43 AM 'Nick Desaulniers' via Clang Built
>
> We must rely on this behavior of Clang because
> --target (which is contained in CLANG_FLAGS)
> must be specified before the Kconfig time.
> Then, a user can toggle CONFIG_64BIT any time
> from menuconfig etc.
>
> With this in mind, using $(ARCH) as if-else
> switches is pointless.
> $(SRCARCH) is the only meaningful input.
>
>
> else ifeq ($(ARCH),i386)
> CLANG_FLAGS += --target=i686-linux-gnu
> else ifeq ($(ARCH),x86)
> CLANG_FLAGS += --target=x86_64-linux-gnu
> else ifeq ($(ARCH),x86_64)
> CLANG_FLAGS += --target=x86_64-linux-gnu
>
> should be replaced with:
>
> else ifeq ($(SRCARCH),x86_64)
> CLANG_FLAGS += --target=x86_64-linux-gnu
I think we usually only have to provide the architecture
name, as in "--target=x86_64", though for arm I get a
warning "clang: warning: unknown platform, assuming
-mfloat-abi=soft" unless I provide the full triple.
> For example for ARCH=arm, which is better
> --target=arm-linux-gnueabi or
> --target=arm-lnux-gnueabihf or something we don't care about?
The kernel is always soft-float itself, so it does not matter either way.
Arnd
On Thu, Jul 8, 2021 at 3:23 AM Masahiro Yamada <[email protected]> wrote:
>
> On Thu, Jul 8, 2021 at 7:43 AM 'Nick Desaulniers' via Clang Built
> Linux <[email protected]> wrote:
> >
> > We get constant feedback that the command line invocation of make is too
> > long. CROSS_COMPILE is helpful when a toolchain has a prefix of the
> > target triple, or is an absolute path outside of $PATH, but it's mostly
> > redundant for a given ARCH.
> >
> > If CROSS_COMPILE is not set, simply set --target= for CLANG_FLAGS,
> > KBUILD_CFLAGS, and KBUILD_AFLAGS based on $ARCH.
> >
> > Previously, we'd cross compile via:
> > $ ARCH=arm64 CROSS_COMPILE=aarch64-linxu-gnu make LLVM=1 LLVM_IAS=1
> > Now:
> > $ ARCH=arm64 make LLVM=1 LLVM_IAS=1
> >
> > Link: https://github.com/ClangBuiltLinux/linux/issues/1399
> > Suggested-by: Arnd Bergmann <[email protected]>
> > Suggested-by: Nathan Chancellor <[email protected]>
> > Signed-off-by: Nick Desaulniers <[email protected]>
> > ---
> > Changes RFC -> v1:
> > * Rebase onto linux-kbuild/for-next
> > * Keep full target triples since missing the gnueabi suffix messes up
> > 32b ARM. Drop Fangrui's sugguested by tag. Update commit message to
> > drop references to arm64.
> > * Flush out TODOS.
> > * Add note about -EL/-EB, -m32/-m64.
> > * Add note to Documentation/.
> >
> > Documentation/kbuild/llvm.rst | 5 +++++
> > scripts/Makefile.clang | 38 +++++++++++++++++++++++++++++++++--
> > 2 files changed, 41 insertions(+), 2 deletions(-)
>
>
>
>
>
>
>
>
> When I was considering a similar idea, my plan was
> to implement this in Kconfig instead of in Makefile
> because that will pass the compiler information
> in one direction (only from Kconfig to Kbuild), but
> that is not so important. We can change it later
> if needed.
>
> I did not complete it because I was investigating
> some issues (especially [3] below), but maybe
> that is something we don't care about.
>
> Can you address [2] below at least?
Sure!
> If we do not have any concern, I can merge it.
> It is likely so because people are only discussing
> "We want to omit omit CROSS_COMPILE".
>
>
>
>
>
>
>
> [1] explicit target triple for native builds
>
> The default target of my distro clang
> is x86_64-pc-linux-gnu.
>
> $ clang --version
> Ubuntu clang version 11.0.0-2
> Target: x86_64-pc-linux-gnu
> Thread model: posix
> InstalledDir: /usr/bin
>
> So, previously, the kernel was built with
> implied --target=x86_64-pc-linux-gnu.
>
>
> With this patch, --target=x86_64-linux-gnu will be
> explicitly specified.
Correct. We've been doing this for x86 builds of Android kernels for a
while without issue.
I can add this note to the commit message:
```
For native builds (not involving cross compilation) we now explicitly
specify a target triple
rather than rely on the implicit host triple.
```
The only downside I can think of now is that we've encountered E2BIG
for excessively long command line arguments in the past (mostly for
out of tree drivers in Android). I'm having trouble imagining how the
implicit host triple could differ in a way from these explicit ones
that would break native compilation. Then again, someone did just
submit patches for building Linux on BSD.
If we don't want to do that, perhaps we could check `cross_compiling`.
Why did you make that variable lowercase in
commit f02aa48dde8b ("kconfig: use /boot/config-* etc. as
DEFCONFIG_LIST only for native build")?
Because the "origin" is not the environment?
> The same applies to native-builds of other architectures.
> For example, when a user builds the arm64 kernel on
> an arm64 server, --target=aarch64-linux-gnu is
> explicitly forced.
>
> I guess, this is a good direction because the produced
> code will be more deterministic, irrespective of the
> Clang's configuration.
>
>
>
> [2] 32/64-bit configuration is toggled in Kconfig time.
>
> Initially, you submitted only arm64. Maybe, it was intentional
> since arm64 is the simplest case.
>
> In the kernel tree, arch/arm and arch/arm64 are very special
> cases where 32-bit and 64-bit are separated by directory.
>
> Some of the other architectures are bi-arch, and
> 32-bit/64-bit is specified by CONFIG_64BIT in Kconfig time.
>
> When Makefiles are being parsed, we actually do not know
> whether the user is planning to configure the kernel
> for 32-bit or 64-bit because CONFIG_64BIT is not
> specified at this point.
>
> ARCH=x86 + CONFIG_64BIT=y
> will build the x86_64 kernel, and
> ARCH=x86 + CONFIG_64BIT=n
> will build the i386 kernel.
>
>
> Then, you may wonder
>
> else ifeq ($(ARCH),x86)
> CLANG_FLAGS += --target=x86_64-linux-gnu
>
> ... works?
>
> Yes, it does fortunately.
>
> -m32/-m64 takes precedence over the
> {x86_64,i386} part of the target triple.
>
> As far as I tested,
>
> clang --target=x86_64-linux-gnu -m32
>
> produced i386 code.
>
> Interestingly,
>
> clang --target=i386-linux-gnu -m64
>
> produced x86_64 code.
Correct. -m32/-m64 and -LE/-BE refine the target triple that the
driver builds up.
> We must rely on this behavior of Clang because
> --target (which is contained in CLANG_FLAGS)
> must be specified before the Kconfig time.
> Then, a user can toggle CONFIG_64BIT any time
> from menuconfig etc.
Correct. So we can't quite move all clang flags into one Makefile
under scripts/ if they rely on Kconfig being run first. This new
makefile is a "pre-kconfig" set of flags.
> With this in mind, using $(ARCH) as if-else
> switches is pointless.
> $(SRCARCH) is the only meaningful input.
>
>
> else ifeq ($(ARCH),i386)
> CLANG_FLAGS += --target=i686-linux-gnu
> else ifeq ($(ARCH),x86)
> CLANG_FLAGS += --target=x86_64-linux-gnu
> else ifeq ($(ARCH),x86_64)
> CLANG_FLAGS += --target=x86_64-linux-gnu
>
> should be replaced with:
>
> else ifeq ($(SRCARCH),x86_64)
> CLANG_FLAGS += --target=x86_64-linux-gnu
Sure, it looks like this would simplify the i386 vs x86_64 handling,
and the use of SRCARCH does seem more prevalent throughout the
codebase. I will fix in v2.
> Some architectures are not only bi-arch, but also bi-endian.
>
>
> You hardcoded 64bit little endian for ppc:
>
> else ifeq ($(ARCH),powerpc)
> CLANG_FLAGS += --target=powerpc64le-linux-gnu
>
>
> But, we must rely on the fact that
>
> clang --target=powerpc64le-linux-gnu -mbig-endian -m32
>
> produces big-endian 32-bit code.
>
> This makes the "64le" part meaningless.
>
>
> This should be noted. Otherwise it is difficult
It is noted; in the top part of the newly added make file.
```
1 # Individual arch/{arch}/Makfiles should use -EL/-EB to set
intended endianness
2 # and -m32/-m64 to set word size based on Kconfigs instead of
relying on the
3 # target triple.
```
Is there somewhere/somehow else you'd like me to note that?
> to understand why --target=x86_64-linux-gnu works fine
> with building the i386 kernel.
>
>
>
> [3] User-space compilation
>
> This does not matter to the kernel itself, but
> Kbuild compiles some userspace programs for
> the target architecture.
> See the samples/ directory for example.
>
> Another example is net/bpfilter/Makefile, which
> embeds the user mode helper when
> CONFIG_BPFILTER_UMH=y.
>
> For this purpose, Kconfig checks if $(CC) is
> capable of linking the userspace.
> (CONFIG_CC_CAN_LINK).
>
> When cross-building with Clang, I cannot see
> CONFIG_CC_CAN_LINK set.
Yes, that is a known issue.
https://github.com/ClangBuiltLinux/linux/issues/1290
>
> If we care about CONFIG_CC_CAN_LINK, probably,
> --sysroot or something should be set according to:
>
> https://clang.llvm.org/docs/CrossCompilation.html
>
> This is an existing issue, but I have no time
> for looking into this.
>
> On debian systems, sysroot for cross-compilation
> are located in /usr/aarch64-linux-gnu,
> /usr/arm-linux-gnueabi, /usr/arm-linux-gnueabihf,
> /usr/i686-linux-gnu/ etc. but I do not know if it
> is the same across distros.
Right. If I remember the glibc/binutils/gcc bootstrap dance, I thought
gcc was configured with a known path to a particular version of glibc.
So a cross toolchain knew where to look for its cross libc.
Clang doesn't have such configure time step; it can cross compile
easily with one binary, but cross linking a working executable is
still a PITA due to not knowing which cross libc to link against. I'm
not sure whether we need to improve Clang's logic when cross compiling
to look in "sane default" paths for a cross libc, or if we should just
add some flag when cross compiling with clang in Kbuild (such as
--sysroot) in order for CONFIG_CC_CAN_LINK to work as expected.
Fangrui probably has a good opinion about this.
Zig (the compiler, but also the language name) can do this quite
nicely. I'm envious.
https://twitter.com/andy_kelley/status/1241409388532948992?lang=en
https://www.youtube.com/watch?v=pq1XqP4-qOo
But this is also somewhat orthogonal to the goal of "infer
CROSS_COMPILE (or really, --target=) from ARCH (or really, SRCARCH)" I
think. It's still interesting for us all to discuss on-list though.
>
>
>
>
>
> [4] What is the best target if we hard-code it?
>
> Currently, we require the correct CROSS_COMPILE
> is provided by users.
>
> The target might impact the performance
> or the ABI.
> It was difficult for me to define
> which one is better than another.
>
> For example for ARCH=arm, which is better
> --target=arm-linux-gnueabi or
> --target=arm-lnux-gnueabihf or
> something we don't care about?
Yes, this is a case I was interested in. I've used either
interchangeably without issue for years. That's not to say we get the
same binary image with either.
I get the same .config for the defconfig target with either.
If I zero out KBUILD_BUILD_TIMESTAMP and the build number, I still get
different sha1sums. Though that assumes clang, lld, and kbuild are
all deterministic, which I also haven't spent time to verify.
--
Thanks,
~Nick Desaulniers
On Thu, Jul 8, 2021 at 4:45 AM Arnd Bergmann <[email protected]> wrote:
>
> On Thu, Jul 8, 2021 at 12:23 PM Masahiro Yamada <[email protected]> wrote:
> > On Thu, Jul 8, 2021 at 7:43 AM 'Nick Desaulniers' via Clang Built
> >
> > We must rely on this behavior of Clang because
> > --target (which is contained in CLANG_FLAGS)
> > must be specified before the Kconfig time.
> > Then, a user can toggle CONFIG_64BIT any time
> > from menuconfig etc.
> >
> > With this in mind, using $(ARCH) as if-else
> > switches is pointless.
> > $(SRCARCH) is the only meaningful input.
> >
> >
> > else ifeq ($(ARCH),i386)
> > CLANG_FLAGS += --target=i686-linux-gnu
> > else ifeq ($(ARCH),x86)
> > CLANG_FLAGS += --target=x86_64-linux-gnu
> > else ifeq ($(ARCH),x86_64)
> > CLANG_FLAGS += --target=x86_64-linux-gnu
> >
> > should be replaced with:
> >
> > else ifeq ($(SRCARCH),x86_64)
> > CLANG_FLAGS += --target=x86_64-linux-gnu
>
> I think we usually only have to provide the architecture
> name, as in "--target=x86_64", though for arm I get a
> warning "clang: warning: unknown platform, assuming
> -mfloat-abi=soft" unless I provide the full triple.
Right, Fangrui also made that suggestion, but for that reason (the
error for various architectures when using 2-component triples) I'd
prefer to just always specify a full triple. I picked some to have a
starting point; unless they NEED to change, I'll refrain from
modifying them further.
Technically, I think they can have 4 components, not sure why we still
call them a "target triple." I guess I wouldn't be surprised if they
can contain more than 4 components at this point.
--
Thanks,
~Nick Desaulniers
On Fri, Jul 9, 2021 at 4:02 AM 'Nick Desaulniers' via Clang Built
Linux <[email protected]> wrote:
>
> On Thu, Jul 8, 2021 at 3:23 AM Masahiro Yamada <[email protected]> wrote:
> >
> > On Thu, Jul 8, 2021 at 7:43 AM 'Nick Desaulniers' via Clang Built
> > Linux <[email protected]> wrote:
> > >
> > > We get constant feedback that the command line invocation of make is too
> > > long. CROSS_COMPILE is helpful when a toolchain has a prefix of the
> > > target triple, or is an absolute path outside of $PATH, but it's mostly
> > > redundant for a given ARCH.
> > >
> > > If CROSS_COMPILE is not set, simply set --target= for CLANG_FLAGS,
> > > KBUILD_CFLAGS, and KBUILD_AFLAGS based on $ARCH.
> > >
> > > Previously, we'd cross compile via:
> > > $ ARCH=arm64 CROSS_COMPILE=aarch64-linxu-gnu make LLVM=1 LLVM_IAS=1
> > > Now:
> > > $ ARCH=arm64 make LLVM=1 LLVM_IAS=1
> > >
> > > Link: https://github.com/ClangBuiltLinux/linux/issues/1399
> > > Suggested-by: Arnd Bergmann <[email protected]>
> > > Suggested-by: Nathan Chancellor <[email protected]>
> > > Signed-off-by: Nick Desaulniers <[email protected]>
> > > ---
> > > Changes RFC -> v1:
> > > * Rebase onto linux-kbuild/for-next
> > > * Keep full target triples since missing the gnueabi suffix messes up
> > > 32b ARM. Drop Fangrui's sugguested by tag. Update commit message to
> > > drop references to arm64.
> > > * Flush out TODOS.
> > > * Add note about -EL/-EB, -m32/-m64.
> > > * Add note to Documentation/.
> > >
> > > Documentation/kbuild/llvm.rst | 5 +++++
> > > scripts/Makefile.clang | 38 +++++++++++++++++++++++++++++++++--
> > > 2 files changed, 41 insertions(+), 2 deletions(-)
> >
> >
> >
> >
> >
> >
> >
> >
> > When I was considering a similar idea, my plan was
> > to implement this in Kconfig instead of in Makefile
> > because that will pass the compiler information
> > in one direction (only from Kconfig to Kbuild), but
> > that is not so important. We can change it later
> > if needed.
> >
> > I did not complete it because I was investigating
> > some issues (especially [3] below), but maybe
> > that is something we don't care about.
> >
> > Can you address [2] below at least?
>
> Sure!
>
> > If we do not have any concern, I can merge it.
> > It is likely so because people are only discussing
> > "We want to omit omit CROSS_COMPILE".
> >
> >
> >
> >
> >
> >
> >
> > [1] explicit target triple for native builds
> >
> > The default target of my distro clang
> > is x86_64-pc-linux-gnu.
> >
> > $ clang --version
> > Ubuntu clang version 11.0.0-2
> > Target: x86_64-pc-linux-gnu
> > Thread model: posix
> > InstalledDir: /usr/bin
> >
> > So, previously, the kernel was built with
> > implied --target=x86_64-pc-linux-gnu.
> >
> >
> > With this patch, --target=x86_64-linux-gnu will be
> > explicitly specified.
>
> Correct. We've been doing this for x86 builds of Android kernels for a
> while without issue.
>
> I can add this note to the commit message:
> ```
> For native builds (not involving cross compilation) we now explicitly
> specify a target triple
> rather than rely on the implicit host triple.
Sounds good.
> ```
> The only downside I can think of now is that we've encountered E2BIG
> for excessively long command line arguments in the past (mostly for
> out of tree drivers in Android). I'm having trouble imagining how the
> implicit host triple could differ in a way from these explicit ones
> that would break native compilation. Then again, someone did just
> submit patches for building Linux on BSD.
>
> If we don't want to do that, perhaps we could check `cross_compiling`.
> Why did you make that variable lowercase in
> commit f02aa48dde8b ("kconfig: use /boot/config-* etc. as
> DEFCONFIG_LIST only for native build")?
> Because the "origin" is not the environment?
Yes, something like that.
I'd like to use upper cases for user interface.
I think using lower cases is OK for
internal variables, but the policy
may not be consistent.
> > The same applies to native-builds of other architectures.
> > For example, when a user builds the arm64 kernel on
> > an arm64 server, --target=aarch64-linux-gnu is
> > explicitly forced.
> >
> > I guess, this is a good direction because the produced
> > code will be more deterministic, irrespective of the
> > Clang's configuration.
> >
> >
> >
> > [2] 32/64-bit configuration is toggled in Kconfig time.
> >
> > Initially, you submitted only arm64. Maybe, it was intentional
> > since arm64 is the simplest case.
> >
> > In the kernel tree, arch/arm and arch/arm64 are very special
> > cases where 32-bit and 64-bit are separated by directory.
> >
> > Some of the other architectures are bi-arch, and
> > 32-bit/64-bit is specified by CONFIG_64BIT in Kconfig time.
> >
> > When Makefiles are being parsed, we actually do not know
> > whether the user is planning to configure the kernel
> > for 32-bit or 64-bit because CONFIG_64BIT is not
> > specified at this point.
> >
> > ARCH=x86 + CONFIG_64BIT=y
> > will build the x86_64 kernel, and
> > ARCH=x86 + CONFIG_64BIT=n
> > will build the i386 kernel.
> >
> >
> > Then, you may wonder
> >
> > else ifeq ($(ARCH),x86)
> > CLANG_FLAGS += --target=x86_64-linux-gnu
> >
> > ... works?
> >
> > Yes, it does fortunately.
> >
> > -m32/-m64 takes precedence over the
> > {x86_64,i386} part of the target triple.
> >
> > As far as I tested,
> >
> > clang --target=x86_64-linux-gnu -m32
> >
> > produced i386 code.
> >
> > Interestingly,
> >
> > clang --target=i386-linux-gnu -m64
> >
> > produced x86_64 code.
>
> Correct. -m32/-m64 and -LE/-BE refine the target triple that the
> driver builds up.
>
> > We must rely on this behavior of Clang because
> > --target (which is contained in CLANG_FLAGS)
> > must be specified before the Kconfig time.
> > Then, a user can toggle CONFIG_64BIT any time
> > from menuconfig etc.
>
> Correct. So we can't quite move all clang flags into one Makefile
> under scripts/ if they rely on Kconfig being run first. This new
> makefile is a "pre-kconfig" set of flags.
>
> > With this in mind, using $(ARCH) as if-else
> > switches is pointless.
> > $(SRCARCH) is the only meaningful input.
> >
> >
> > else ifeq ($(ARCH),i386)
> > CLANG_FLAGS += --target=i686-linux-gnu
> > else ifeq ($(ARCH),x86)
> > CLANG_FLAGS += --target=x86_64-linux-gnu
> > else ifeq ($(ARCH),x86_64)
> > CLANG_FLAGS += --target=x86_64-linux-gnu
> >
> > should be replaced with:
> >
> > else ifeq ($(SRCARCH),x86_64)
> > CLANG_FLAGS += --target=x86_64-linux-gnu
>
> Sure, it looks like this would simplify the i386 vs x86_64 handling,
> and the use of SRCARCH does seem more prevalent throughout the
> codebase. I will fix in v2.
>
> > Some architectures are not only bi-arch, but also bi-endian.
> >
> >
> > You hardcoded 64bit little endian for ppc:
> >
> > else ifeq ($(ARCH),powerpc)
> > CLANG_FLAGS += --target=powerpc64le-linux-gnu
> >
> >
> > But, we must rely on the fact that
> >
> > clang --target=powerpc64le-linux-gnu -mbig-endian -m32
> >
> > produces big-endian 32-bit code.
> >
> > This makes the "64le" part meaningless.
> >
> >
> > This should be noted. Otherwise it is difficult
>
> It is noted; in the top part of the newly added make file.
> ```
> 1 # Individual arch/{arch}/Makfiles should use -EL/-EB to set
> intended endianness
> 2 # and -m32/-m64 to set word size based on Kconfigs instead of
> relying on the
> 3 # target triple.
> ```
> Is there somewhere/somehow else you'd like me to note that?
Ah, you already noted it in the new Makefile.
Sorry, I missed it.
Then, this is fine.
>
> > to understand why --target=x86_64-linux-gnu works fine
> > with building the i386 kernel.
> >
> >
> >
> > [3] User-space compilation
> >
> > This does not matter to the kernel itself, but
> > Kbuild compiles some userspace programs for
> > the target architecture.
> > See the samples/ directory for example.
> >
> > Another example is net/bpfilter/Makefile, which
> > embeds the user mode helper when
> > CONFIG_BPFILTER_UMH=y.
> >
> > For this purpose, Kconfig checks if $(CC) is
> > capable of linking the userspace.
> > (CONFIG_CC_CAN_LINK).
> >
> > When cross-building with Clang, I cannot see
> > CONFIG_CC_CAN_LINK set.
>
> Yes, that is a known issue.
> https://github.com/ClangBuiltLinux/linux/issues/1290
>
> >
> > If we care about CONFIG_CC_CAN_LINK, probably,
> > --sysroot or something should be set according to:
> >
> > https://clang.llvm.org/docs/CrossCompilation.html
> >
> > This is an existing issue, but I have no time
> > for looking into this.
> >
> > On debian systems, sysroot for cross-compilation
> > are located in /usr/aarch64-linux-gnu,
> > /usr/arm-linux-gnueabi, /usr/arm-linux-gnueabihf,
> > /usr/i686-linux-gnu/ etc. but I do not know if it
> > is the same across distros.
>
> Right. If I remember the glibc/binutils/gcc bootstrap dance, I thought
> gcc was configured with a known path to a particular version of glibc.
> So a cross toolchain knew where to look for its cross libc.
>
> Clang doesn't have such configure time step; it can cross compile
> easily with one binary, but cross linking a working executable is
> still a PITA due to not knowing which cross libc to link against. I'm
> not sure whether we need to improve Clang's logic when cross compiling
> to look in "sane default" paths for a cross libc, or if we should just
> add some flag when cross compiling with clang in Kbuild (such as
> --sysroot) in order for CONFIG_CC_CAN_LINK to work as expected.
> Fangrui probably has a good opinion about this.
>
> Zig (the compiler, but also the language name) can do this quite
> nicely. I'm envious.
> https://twitter.com/andy_kelley/status/1241409388532948992?lang=en
> https://www.youtube.com/watch?v=pq1XqP4-qOo
>
> But this is also somewhat orthogonal to the goal of "infer
> CROSS_COMPILE (or really, --target=) from ARCH (or really, SRCARCH)" I
> think. It's still interesting for us all to discuss on-list though.
>
> >
> >
> >
> >
> >
> > [4] What is the best target if we hard-code it?
> >
> > Currently, we require the correct CROSS_COMPILE
> > is provided by users.
> >
> > The target might impact the performance
> > or the ABI.
> > It was difficult for me to define
> > which one is better than another.
> >
> > For example for ARCH=arm, which is better
> > --target=arm-linux-gnueabi or
> > --target=arm-lnux-gnueabihf or
> > something we don't care about?
>
> Yes, this is a case I was interested in. I've used either
> interchangeably without issue for years. That's not to say we get the
> same binary image with either.
>
> I get the same .config for the defconfig target with either.
>
> If I zero out KBUILD_BUILD_TIMESTAMP and the build number, I still get
> different sha1sums. Though that assumes clang, lld, and kbuild are
> all deterministic, which I also haven't spent time to verify.
> --
> Thanks,
> ~Nick Desaulniers
>
> --
> You received this message because you are subscribed to the Google Groups "Clang Built Linux" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to [email protected].
> To view this discussion on the web visit https://groups.google.com/d/msgid/clang-built-linux/CAKwvOdmufESjYQVZmaPdTXgZO5Ogz%2BOppVSUGAn6BZaC%2BYZhbw%40mail.gmail.com.
--
Best Regards
Masahiro Yamada