Re: linux-parisc compile failure in current git

[Helge Deller <deller@gmx.de>]

Hi Masahiro,

On 6/10/21 4:03 AM, Masahiro Yamada wrote:
> On Thu, Jun 10, 2021 at 7:50 AM Helge Deller <deller@gmx.de> wrote:
>>
>> On 6/1/21 12:21 PM, Meelis Roos wrote:
>>> Upstream Linux git fails to compile on gentoo hppa -  .config below.
>>> I have 2 gcc-s as always:
>>> $ gcc-config -l
>>>    [1] hppa2.0-unknown-linux-gnu-9.3.0
>>>    [2] hppa2.0-unknown-linux-gnu-10.2.0 *
>>>
>>>    [3] hppa64-unknown-linux-gnu-10.2.0 *
>>
>>
>> I see the same issue too, but only when compiling natively on a parisc machine.
>> Cross-compiling on a x86 box works nicely.
>>
>> First I thought it's a problem with setting the "cross_compiling" flag in ./Makefile.
>> But that's not sufficient.
>>
>> On a x86 machine (which builds fine) I get
>> SRCARCH=parisc SUBARCH=x86 UTS_MACHINE=parisc
>> The arch/parisc/kernel/asm-offsets.c file gets preprocessed via:
>> hppa64-linux-gnu-gcc
>>
>> On a native 32bit parisc machine I have:
>> SRCARCH=parisc SUBARCH=parisc UTS_MACHINE=parisc
>> Here the arch/parisc/kernel/asm-offsets.c file gets preprocessed via:
>> gcc
>> Instead here the native hppa64-linux-gnu-gcc (cross compiler) should have been used too, since
>> we build a 64-bit hppa kernel (CONFIG_64BIT is set).
>> Note, on hppa we don't have an "-m64" compiler flag as on x86.
>
> I see.
> hppa is not a bi-arch compiler, in other words,
> http- and hppa64- are separate compilers.

Yes.

>> Mashahiro, do you maybe have an idea what gets wrong here, or which
>> patch has changed the behaviour how the asm-offsets.c file gets preprocessed?
>
> Presumably, commit 23243c1ace9fb4eae2f75e0fe0ece8e3219fb4f3
>
> Prior to that commit, arch/parisc/Makefile was like this:
>
> ifneq ($(SUBARCH),$(UTS_MACHINE))
>          ifeq ($(CROSS_COMPILE),)
>                      ...
>
> Now I understand why arch/parisc/Makefile was written this way.
>
> Reverting the change in arch/parisc/Makefile will restore the original behavior.

Sadly, reverting this change (23243c1ace9fb4eae2f75e0fe0ece8e3219fb4f3) does not
restore the original behavior.

> But, please keep in mind that there is an issue remaining.
>
> Please see this code:
>
> ifdef CONFIG_64BIT
> UTS_MACHINE := parisc64
> CHECKFLAGS += -D__LP64__=1
> CC_ARCHES = hppa64
> LD_BFD := elf64-hppa-linux
> else # 32-bit
> CC_ARCHES = hppa hppa2.0 hppa1.1
> LD_BFD := elf32-hppa-linux
> endif
>
>
> UTS_MACHINE is determined by CONFIG_64BIT.
>
> CONFIG_64BIT is defined only after Kconfig is finished.
> When you are trying to configure the .config,
> CONFIG_64BIT is not defined yet.
> So UTS_MACHINE is always 'parisc'.

Yes.
See above, but it worked when I had SUBARCH=x86 (when running my laptop).


> As you know, Kconfig files now have a bunch of 'cc-option' syntax
> to check the compiler capability in Kconfig time.
> Hence, you need to provide a proper compiler in Kconfig time too.
>
> When you build a 64-bit parisc kernel on a 32-bit parisc machine,

Please note, that we don't have a 64-bit parisc userspace yet (just kernel).
This means, that all builds on parisc machines are 32bit and do a
cross-compilation to a parisc64 kernel if requested in the .config.

> Kconfig is passed with CC=gcc since SUBARCH==UTS_MACHINE==parisc.
> After Kconfig, CROSS_COMPILE=hppa64-* is set,
> and the kernel is built by CC=hppa64-*-gcc.

Right. That is the old behaviour. Based on the CONFIG_64BIT option
the hppa64 compiler is choosen for CROSS_COMPILE.

> So, Kconfig evaluated a compiletely different compiler. This is pointless.

Yes, probably.


> There are some options
>
> [option 1]
>    revert the parisc bit of 23243c1ace9fb4eae2f75e0fe0ece8e3219fb4f3
>    This will restore the functionality you may want, but
>    as I said above, Kconfig is doing pointless things.

as mentioned above: Doesn't solve the issue.

> [option 2]
>     Stop using cc-cross-prefix, and pass CROSS_COMPILE explicitly.
>     This is what many architectures including arm, arm64 do.
>     You need to explicitly pass CROSS_COMPILE=aarch64-linux-gnu- etc.
>     if you are cross-compiling arm64.

Might be an option, but it's not as handy as simply choosing CONFIG_64BIT
and then things are done automatically.

> [option 3]
>     Introduce ARCH=parisc64.
>
>     When you are building 64-bit kernel, you can pass ARCH=parisc64
>      A patch attached.  (but not tested much)

Tried it, but doesn't work.
asm-offsets.c is still preprocessed with 32bit compiler (gcc, not hppa20-gcc).

Thanks for your help so far!
If you like I'm happy to give you an account on a hppa64 machine to reproduce yourself.
I'll now try to bisect where it goes wrong. There must be something else before commit 23243c1ace9fb4eae2f75e0fe0ece8e3219fb4f3.

Helge