Re: [PATCH v1 2/2] memory: tegra: Enable compile testing for all drivers

From: Krzysztof Kozlowski <krzysztof.kozlowski@canonical.com>
To: Nathan Chancellor <nathan@kernel.org>,
	Dmitry Osipenko <digetx@gmail.com>
Cc: kernel test robot <lkp@intel.com>,
	Thierry Reding <thierry.reding@gmail.com>,
	Jonathan Hunter <jonathanh@nvidia.com>,
	Nick Desaulniers <ndesaulniers@google.com>,
	clang-built-linux@googlegroups.com, kbuild-all@lists.01.org,
	linux-kernel@vger.kernel.org, linux-tegra@vger.kernel.org
Subject: Re: [PATCH v1 2/2] memory: tegra: Enable compile testing for all drivers
Date: Tue, 11 May 2021 13:57:32 -0400	[thread overview]
Message-ID: <6fc3e47b-cef6-0d78-0578-551608856ae6@canonical.com> (raw)
In-Reply-To: <YJrARxhVD7QM/GPv@archlinux-ax161>

On 11/05/2021 13:35, Nathan Chancellor wrote:
> On Tue, May 11, 2021 at 07:00:34PM +0300, Dmitry Osipenko wrote:
>> 11.05.2021 18:31, Krzysztof Kozlowski пишет:
>> ...
>>                                       ~~~~~~~~~~~~~~~~~~~~~^
>>>>>>> drivers/memory/tegra/tegra124-emc.c:802:26: warning: implicit conversion from 'unsigned long' to 'u32' (aka 'unsigned int') changes value from 18446744071562067985 to 2147483665 [-Wconstant-conversion]
>>>>>                    emc_ccfifo_writel(emc, EMC_ZQ_CAL_LONG_CMD_DEV0, EMC_ZQ_CAL);
>>>>>                    ~~~~~~~~~~~~~~~~~      ^~~~~~~~~~~~~~~~~~~~~~~~
>>>>>    drivers/memory/tegra/tegra124-emc.c:154:36: note: expanded from macro 'EMC_ZQ_CAL_LONG_CMD_DEV0'
>>>>>            (DRAM_DEV_SEL_0 | EMC_ZQ_CAL_LONG | EMC_ZQ_CAL_CMD)
>>>>>             ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~
>>>>>    13 warnings generated.
>>>>
>>>> This doesn't look like a useful warning from clang, it should see that
>>>> the constant value itself isn't truncated, hence it should be a problem
>>>> of clang. Do you think it's okay to ignore this nonsense?
>>>
>>> I admit I also do not see the real issue here. The DRAM_DEV_SEL_0 fits
>>> in u32 and there is no other bitwise arithmetic than just OR, so why
>>> clang assumes it can have 32 most signifcant bits toggled on?
>>>
>>> +Cc Nathan and Nick,
>>> Maybe you could shed some light here on this warning?
>>>
>>> Dmitry,
>>> In general you should not ignore it because:
>>> 1. This breaks allyesconfig with clang on powerpc (or it is one of the
>>> stoppers),
>>> 2. We might want in some future to build it with clang.
>>
>> I meant to ignore it from the perspective of the memory drivers, i.e. it
>> likely should be fixed in clang and not worked around in the code. Thank
>> you for pinging the right people.
> 
> I do not think this is a bug in clang, gcc warns the same (just not here
> in this case): https://godbolt.org/z/e9GWobMnd
> 
> DRAM_DEV_SEL_0 and DRAM_DEV_SEL_1 are implicitly signed integers because
> there is no suffix on the literal 1. DRAM_DEV_SEL_0 is 2 << 30, which
> can be turned into 1 << 31. That is equal to INT_MAX + 1, which then
> overflows and becomes INT_MIN (undefined behavior). INT_MIN is then
> promoted to unsigned long because EMC_ZQ_CAL_LONG and EMC_ZQ_CAL_CMD are
> unsigned long due to the BIT macro, resulting in the gigantic number
> that clang reports above.
> 
Thanks, good point.

Best regards,
Krzysztof