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

From: Nathan Chancellor
Date: Tue May 11 2021 - 13:35:12 EST


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.

I assume that this driver only runs on hardware where unsigned int is
the same size as unsigned long, meaning this problem is merely
theoretical?

Regardless, defining DRAM_DEV_SEL_{0,1} with the BIT macro fixes the
warning for me and should make everything work as expected.

diff --git a/drivers/memory/tegra/tegra124-emc.c b/drivers/memory/tegra/tegra124-emc.c
index 5699d909abc2..a21ca8e0841a 100644
--- a/drivers/memory/tegra/tegra124-emc.c
+++ b/drivers/memory/tegra/tegra124-emc.c
@@ -272,8 +272,8 @@
#define EMC_PUTERM_ADJ 0x574

#define DRAM_DEV_SEL_ALL 0
-#define DRAM_DEV_SEL_0 (2 << 30)
-#define DRAM_DEV_SEL_1 (1 << 30)
+#define DRAM_DEV_SEL_0 BIT(31)
+#define DRAM_DEV_SEL_1 BIT(30)

#define EMC_CFG_POWER_FEATURES_MASK \
(EMC_CFG_DYN_SREF | EMC_CFG_DRAM_ACPD | EMC_CFG_DRAM_CLKSTOP_SR | \