[PATCH RESEND] riscv: don't specify -mno-save-restore when building with clang

[Jerome Forissier]

On 7/29/20 10:29 AM, Tobias Klauser wrote:
> On 2020-07-29 at 10:08:11 +0200, Jerome Forissier <jerome at forissier.org> wrote:
>>> On 2020-07-28 at 18:06:36 +0200, Tobias Klauser <tklauser at distanz.ch> wrote:
>>>> On 2020-07-28 at 17:20:45 +0200, Palmer Dabbelt <palmer at dabbelt.com> wrote:
>>>>> On Tue, 28 Jul 2020 06:12:52 PDT (-0700), tklauser at distanz.ch wrote:
>>>>>> Clang doesn't support -msave-restore and -mno-save-restore. This avoids
>>>>>> the following message emitted by clang:
>>>>>>
>>>>>>   '-save-restore' is not a recognized feature for this target (ignoring feature)
>>>>>>
>>>>>> Signed-off-by: Tobias Klauser <tklauser at distanz.ch>
>>>>>> ---
>>>>>> Resent due to infradead.org mailing list issues.
>>>>>>
>>>>>>  arch/riscv/Makefile | 2 ++
>>>>>>  1 file changed, 2 insertions(+)
>>>>>>
>>>>>> diff --git a/arch/riscv/Makefile b/arch/riscv/Makefile
>>>>>> index fb6e37db836d..cd3720bc45e8 100644
>>>>>> --- a/arch/riscv/Makefile
>>>>>> +++ b/arch/riscv/Makefile
>>>>>> @@ -44,7 +44,9 @@ riscv-march-$(CONFIG_RISCV_ISA_C)	:= $(riscv-march-y)c
>>>>>>  KBUILD_CFLAGS += -march=$(subst fd,,$(riscv-march-y))
>>>>>>  KBUILD_AFLAGS += -march=$(riscv-march-y)
>>>>>>
>>>>>> +ifndef CONFIG_CC_IS_CLANG
>>>>>>  KBUILD_CFLAGS += -mno-save-restore
>>>>>> +endif
>>>>>>  KBUILD_CFLAGS += -DCONFIG_PAGE_OFFSET=$(CONFIG_PAGE_OFFSET)
>>>>>>
>>>>>>  ifeq ($(CONFIG_CMODEL_MEDLOW),y)
>>>>>
>>>>> Thanks, this one didn't make it the first time.  Is there a reason we can't use
>>>>> cc-option here?  IIRC that's what we use for the other compiler options that
>>>>> may be unimplemented, and it has the advantage of avoiding encoding specific
>>>>> compilers into the build system.
>>>>
>>>> Thanks for the hint. I don't know how I could've overlooked cc-option.
>>>> Will send a v2 using cc-option.
>>>
>>> Looks like it's a bit more complicated:
>>>
>>> Using just cc-option still leads to the warning being emitted, so I
>>> think the CONFIG_CC_IS_CLANG check is still needed (checked using clang
>>
>> Shouldnt cc-option be fixed instead? The warning effectively means the option
>> is not supported, so why should cc-option pretend it is?
> 
> The problem is that clang somewhat handles the option, i.e. doesn't
> treat it as an unknown option like others and thus doesn't exit with
> non-zero, like it does for other options only available for gcc. This is
> what cc-option relies on.

OK, I was thinking perhaps cc-option should consider a warning as a hint
that the option should not be used. But it is probably not a good idea
after all (could easily break things).

However this makes me think the following could work (untested!):

-KBUILD_CFLAGS += -mno-save-restore
+KBUILD_CFLAGS += $(call cc-option,-mno-save-restore -Werror)

> 
> Like mentioned in the follow-up patch [1] clang 11 will handle
> -m{no,}save-restore correctly, so we need to address that case as well
> (which [1] does). Unfortunately, I currently see no other way than using
> these ifeq's to fix this.
> 
> [1] https://lore.kernel.org/linux-riscv/20200729044428.32460-1-tklauser@distanz.ch/T/#u

The -Werror technique would avoid introducing a test on the version number.

-- 
Jerome