Re: [PATCH] Documentation/llvm: refresh docs for LLVM=1

[Nathan Chancellor <nathan@kernel.org>]

On Tue, Feb 15, 2022 at 12:49:47PM -0800, Nick Desaulniers wrote:
> Refresh the docs for 2022:
> * add link to Chimera Linux.
> * add Quick Start section pointing out that LLVM=1 is the simple
>   recommended method.
> * Re-order the sections on Clang and Cross Compiling to come after the
>   section on LLVM utilities. I think the documentation flows better this
>   way.
> * Add note about LLVM=1 to Omitting CROSS_COMPILE section.
> * Change note in Supported Architectures from discussing CC=clang vs.
>   LLVM=1 to discussion LLVM=1 vs. LLVM_IAS=0.
> * Update make command column of support architecture table; we now
>   support LLVM=1 everywhere; just some more work to get ARCH=s390
>   assembling with clang.
> * Word wrap a few columns over 80 lines.

This patch is doing a lot, it might make sense to split it up. Reviewing
documentation changes is not the easiest as it is.

Comments inline, thanks for the patch!

> Signed-off-by: Nick Desaulniers <ndesaulniers@google.com>
> ---
>  Documentation/kbuild/llvm.rst | 82 +++++++++++++++++++++--------------
>  1 file changed, 50 insertions(+), 32 deletions(-)
> 
> diff --git a/Documentation/kbuild/llvm.rst b/Documentation/kbuild/llvm.rst
> index d32616891dcf..7f52cde70bc7 100644
> --- a/Documentation/kbuild/llvm.rst
> +++ b/Documentation/kbuild/llvm.rst
> @@ -15,36 +15,27 @@ such as GCC and binutils. Ongoing work has allowed for `Clang
>  <https://clang.llvm.org/>`_ and `LLVM <https://llvm.org/>`_ utilities to be
>  used as viable substitutes. Distributions such as `Android
>  <https://www.android.com/>`_, `ChromeOS
> -<https://www.chromium.org/chromium-os>`_, and `OpenMandriva
> -<https://www.openmandriva.org/>`_ use Clang built kernels.  `LLVM is a
> +<https://www.chromium.org/chromium-os>`_, `OpenMandriva
> +<https://www.openmandriva.org/>`_, and `Chimera Linux
> +<https://chimera-linux.org/>`_ use Clang built kernels.  `LLVM is a
>  collection of toolchain components implemented in terms of C++ objects
>  <https://www.aosabook.org/en/llvm.html>`_. Clang is a front-end to LLVM that
>  supports C and the GNU C extensions required by the kernel, and is pronounced
>  "klang," not "see-lang."
>  
> -Clang
> ------
> +Quick Start
> +-----------
> +::
>  
> -The compiler used can be swapped out via ``CC=`` command line argument to ``make``.
> -``CC=`` should be set when selecting a config and during a build. ::
> +	make LLVM=1
>  
> -	make CC=clang defconfig
> +for native builds or
>  
> -	make CC=clang
> +or ::
>  
> -Cross Compiling
> ----------------
> +	ARCH=arm64 make LLVM=1

make ARCH=arm64 LLVM=1

is used throughout the rest of the document and I think it is more
consistent to keep all the variables on the right side of make.

Compare:

$ rg "make .*=" Documentation

vs.

$ rg "=.* make" Documentation

>  
> -A single Clang compiler binary will typically contain all supported backends,
> -which can help simplify cross compiling. ::
> -
> -	make ARCH=arm64 CC=clang CROSS_COMPILE=aarch64-linux-gnu-
> -
> -``CROSS_COMPILE`` is not used to prefix the Clang compiler binary, instead
> -``CROSS_COMPILE`` is used to set a command line flag: ``--target=<triple>``. For
> -example: ::
> -
> -	clang --target=aarch64-linux-gnu foo.c
> +for cross compiling.
>  
>  LLVM Utilities
>  --------------
> @@ -63,6 +54,32 @@ They can be enabled individually. The full list of the parameters: ::
>  The integrated assembler is enabled by default. You can pass ``LLVM_IAS=0`` to
>  disable it.
>  
> +Clang
> +-----
> +
> +The compiler used can be swapped out via ``CC=`` command line argument to

... via the ``CC=``... ?

> +``make``.  ``CC=`` should be set when selecting a config and during a build.

I think

``CC=`` should be set when generating a config and building a kernel.

might read a little bit better?

> +``CC=clang`` is implied by ``LLVM=1``, but if it's desired to simply use clang
> +as the compiler for the target, then ``CC=`` is a quick substitute. ::
> +
> +	make CC=clang defconfig
> +
> +	make CC=clang
> +
> +Cross Compiling
> +---------------
> +
> +A single Clang compiler binary will typically contain all supported backends,
> +which can help simplify cross compiling. ::
> +
> +	make ARCH=arm64 CC=clang CROSS_COMPILE=aarch64-linux-gnu-
> +
> +``CROSS_COMPILE`` is not used to prefix the Clang compiler binary, instead
> +``CROSS_COMPILE`` is used to set a command line flag: ``--target=<triple>``.
> +For example: ::
> +
> +	clang --target=aarch64-linux-gnu foo.c
> +
>  Omitting CROSS_COMPILE
>  ----------------------
>  
> @@ -71,7 +88,8 @@ As explained above, ``CROSS_COMPILE`` is used to set ``--target=<triple>``.
>  If ``CROSS_COMPILE`` is not specified, the ``--target=<triple>`` is inferred
>  from ``ARCH``.
>  
> -That means if you use only LLVM tools, ``CROSS_COMPILE`` becomes unnecessary.
> +That means if you use only LLVM tools via ``LLVM=1``, ``CROSS_COMPILE`` becomes
> +unnecessary.
>  
>  For example, to cross-compile the arm64 kernel::
>  
> @@ -88,13 +106,13 @@ Supported Architectures
>  LLVM does not target all of the architectures that Linux supports and
>  just because a target is supported in LLVM does not mean that the kernel
>  will build or work without any issues. Below is a general summary of
> -architectures that currently work with ``CC=clang`` or ``LLVM=1``. Level
> -of support corresponds to "S" values in the MAINTAINERS files. If an
> -architecture is not present, it either means that LLVM does not target
> -it or there are known issues. Using the latest stable version of LLVM or
> -even the development tree will generally yield the best results.
> -An architecture's ``defconfig`` is generally expected to work well,
> -certain configurations may have problems that have not been uncovered
> +architectures that currently work with ``LLVM=1`` but may need the integrated
> +assembler disabled via ``LLVM_IAS=0``. Level of support corresponds to "S"
> +values in the MAINTAINERS files. If an architecture is not present, it either
> +means that LLVM does not target it or there are known issues. Using the latest
> +stable version of LLVM or even the development tree will generally yield the
> +best results.  An architecture's ``defconfig`` is generally expected to work
> +well, certain configurations may have problems that have not been uncovered
>  yet. Bug reports are always welcome at the issue tracker below!
>  
>  .. list-table::
> @@ -112,16 +130,16 @@ yet. Bug reports are always welcome at the issue tracker below!
>       - ``LLVM=1``
>     * - mips
>       - Maintained
> -     - ``CC=clang``
> +     - ``LLVM=1``
>     * - powerpc
>       - Maintained
> -     - ``CC=clang``
> +     - ``LLVM=1``

Are we ready for this yet?

https://github.com/ClangBuiltLinux/linux/issues?q=is%3Aopen+is%3Aissue+label%3A%22%5BARCH%5D+powerpc%22+label%3A%22%5BTOOL%5D+integrated-as%22

I think we could potentially say LLVM=1 LLVM_IAS=0 but pseries_defconfig
does not build with ld.lld:

https://github.com/ClangBuiltLinux/linux/issues/602

Maybe it is worth breaking up powerpc into powerpc32, powerpc64, and
powerpc64le?

>     * - riscv
>       - Maintained
> -     - ``CC=clang``
> +     - ``LLVM=1``
>     * - s390
>       - Maintained
> -     - ``CC=clang``
> +     - ``LLVM=1 LLVM_IAS=0``

I don't think s390 is ready for this claim. ld.lld and llvm-objcopy do
not support s390, which are both used in a defconfig build, as far as I
am aware.

https://github.com/ClangBuiltLinux/linux/issues/1524
https://github.com/ClangBuiltLinux/linux/issues/1530

>     * - x86
>       - Supported
>       - ``LLVM=1``
> 
> base-commit: c5d9ae265b105d9a67575fb67bd4650a6fc08e25
> -- 
> 2.35.1.265.g69c8d7142f-goog
> 

The rest of the changes look good to me.

Cheers,
Nathan