Re: [gentoo-dev] [PATCH v2] kernel-2.eclass: Respect portage CC & AR variables

[Adrian Ratiu]

Hi James, thank you very much for your feedback!

On Wed, 15 Dec 2021, James Beddek <telans@××××××.de> wrote:
> On Thursday, 16 December 2021 3:58:26 AM NZDT Adrian Ratiu
> wrote:
>> Starting with kernel>=v5.7 the build system can override the
>> tools vars by setting LLVM=1 [1], but older kernels still use
>> the default GNU tools, so to be able to use a full LLVM/Clang
>> build, CC should be set together with AR to the portage set
>> values.   Doing this avoids situations like building the kernel
>> with clang (using the set HOSTCC) but using gcc/gnu-ar for
>> headers.   [1] a0d1c951ef08 kbuild: support LLVM=1 to switch
>> the default tools to Clang/LLVM  Co-authored-by: Manoj Gupta
>> <manojgupta@××××××××.org> Signed-off-by: Adrian Ratiu
>> <adrian.ratiu@×××××××××.com> ---
>>  eclass/kernel-2.eclass | 2 +- 1 file changed, 1 insertion(+),
>>  1 deletion(-)
>>  diff --git a/eclass/kernel-2.eclass b/eclass/kernel-2.eclass
>> index adc1425bc2e..caeec86ff59 100644 ---
>> a/eclass/kernel-2.eclass +++ b/eclass/kernel-2.eclass @@ -692,7
>> +692,7 @@ env_setup_xmakeopts() {
>>  	elif type -p ${CHOST}-ar >/dev/null; then
>>  xmakeopts="${xmakeopts} CROSS_COMPILE=${CHOST}-" fi
>> -	xmakeopts="${xmakeopts} HOSTCC=$(tc-getBUILD_CC)" +
>> xmakeopts="${xmakeopts} HOSTCC=$(tc-getBUILD_CC) CC=$(tc-getCC)
>> AR=$(tc-getAR)" export xmakeopts
>>  }
>
> What's the reason for not using $LLVM here? To preserve building
> with clang  for kernels < 5.7 ?

That is correct - if Gentoo were to support only kernels >= 5.7 
then we could try to use $LLVM directly, but older long-term 
releases are still supported, the upstream kernel will backport 
patches to v4.14 until Jan 2024 and we would really like to build 
these with Clang too.

>
> I might be missing the point, but wouldn't something along the
> lines of "LLVM=$(! [[ $(tc-getCC) =~ clang ]]; echo $?)" work to
> preserve the correct  compiler tools?

In theory Manoj's argument is correct: this is about respecting CC 
and the tools configuration in general so we should try to avoid 
hardcoding for clang if possible, but...

In practice I'm not aware of people building the kernel with 
something other than GNU or LLVM variants. At least these are the 
only two which receive "first class" support by the upstream 
kernel community.

In addition, hardcoding $LLVM does not fix the problem of older 
kernels.
 
I thought about backporting $LLVM support to long-term releases 
but I don't think it would be acceptable for inclusion in older 
kernels which should receive only bug fixes.

This is why I think the best solution is to just set the variables 
in this eclass to the values set by portage.

Thanks,
Adrian

>
> Cheers