On Tue, 25 Oct 2022 02:25:06 -0700, Andi Shyti wrote:
>
> Hi Ashutosh,
Hi Andi :)
> > > If a non-constant variable is used as the first argument of the FIELD_PREP
> > > macro, a build error occurs when using the clang compiler.
A "non-constant variable" does not seem to be the cause of the compile
error with clang, see below.
>
> > > drivers/gpu/drm/i915/i915_hwmon.c:115:16: error: result of comparison of constant 18446744073709551615 with expression of type 'typeof (_Generic((field_msk), char: (unsigned char)0, unsigned char: (unsigned char)0, signed char: (unsigned char)0, unsigned short: (unsigned short)0, short: (unsigned short)0, unsigned int: (unsigned int)0, int: (unsigned int)0, unsigned long: (unsigned long)0, long: (unsigned long)0, unsigned long long: (unsigned long long)0, long long: (unsigned long long)0, default: (field_msk)))' (aka 'unsigned int') is always false [-Werror,-Wtautological-constant-out-of-range-compare]
> >
> > What is 18446744073709551615? You may want to limit the length of this line
> > or checkpatch doesn't complain?
>
> yeah! I am not a clang user, and this must be some ugly error
> output. I don't think it makes sense to break it, though.
18446744073709551615 == ~0ull (see use in __BF_FIELD_CHECK).
>
> > > bits_to_set = FIELD_PREP(field_msk, nval);
> > > ^~~~~~~~~~~~~~~~~~~~~~~~~~~
> > > ./include/linux/bitfield.h:114:3: note: expanded from macro 'FIELD_PREP'
> > > __BF_FIELD_CHECK(_mask, 0ULL, _val, "FIELD_PREP: "); \
> > > ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > > ./include/linux/bitfield.h:71:53: note: expanded from macro '__BF_FIELD_CHECK'
> > > BUILD_BUG_ON_MSG(__bf_cast_unsigned(_mask, _mask) > \
So clang seems to break here at this line in __BF_FIELD_CHECK (note ~0ull
also occurs here):
BUILD_BUG_ON_MSG(__bf_cast_unsigned(_mask, _mask) > \
__bf_cast_unsigned(_reg, ~0ull), \
_pfx "type of reg too small for mask"); \
So it goes through previous checks including the "mask is not constant"
check. As Nick Desaulniers mentions "__builtin_constant_p is evaluated
after most optimizations have run" so by that time both compilers (gcc and
clang) have figured out that even though _mask is coming in as function
argument it is really the constant below:
#define PKG_PWR_LIM_1 REG_GENMASK(14, 0)
But it is not clear why clang chokes on this "type of reg too small for
mask" check (and gcc doesn't) since everything is u32.
It is for this reason I want someone from llvm to chime in.
> > > ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~
> > > ./include/linux/build_bug.h:39:58: note: expanded from macro 'BUILD_BUG_ON_MSG'
> > > ~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~
> > > ./include/linux/compiler_types.h:357:22: note: expanded from macro 'compiletime_assert'
> > > _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__)
> > > ~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > > ./include/linux/compiler_types.h:345:23: note: expanded from macro '_compiletime_assert'
> > > __compiletime_assert(condition, msg, prefix, suffix)
> > > ~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > > ./include/linux/compiler_types.h:337:9: note: expanded from macro '__compiletime_assert'
> > > if (!(condition)) \
> > >
> > > Fixes: 99f55efb7911 ("drm/i915/hwmon: Power PL1 limit and TDP setting")
> > > Cc: Ashutosh Dixit <ashutosh.dixit@xxxxxxxxx>
> > > Cc: Anshuman Gupta <anshuman.gupta@xxxxxxxxx>
> > > Cc: Andi Shyti <andi.shyti@xxxxxxxxxxxxxxx>
> > > Signed-off-by: Gwan-gyeong Mun <gwan-gyeong.mun@xxxxxxxxx>
> > > ---
> > > drivers/gpu/drm/i915/i915_hwmon.c | 12 +++---------
> > > 1 file changed, 3 insertions(+), 9 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/i915/i915_hwmon.c b/drivers/gpu/drm/i915/i915_hwmon.c
> > > index 9e9781493025..782a621b1928 100644
> > > --- a/drivers/gpu/drm/i915/i915_hwmon.c
> > > +++ b/drivers/gpu/drm/i915/i915_hwmon.c
> > > @@ -101,21 +101,16 @@ hwm_field_read_and_scale(struct hwm_drvdata *ddat, i915_reg_t rgadr,
> > >
> > > static void
> > > hwm_field_scale_and_write(struct hwm_drvdata *ddat, i915_reg_t rgadr,
> > > - u32 field_msk, int nshift,
> > > - unsigned int scale_factor, long lval)
> > > + int nshift, unsigned int scale_factor, long lval)
> > > {
> > > u32 nval;
> > > - u32 bits_to_clear;
> > > - u32 bits_to_set;
> > >
> > > /* Computation in 64-bits to avoid overflow. Round to nearest. */
> > > nval = DIV_ROUND_CLOSEST_ULL((u64)lval << nshift, scale_factor);
> > >
> > > - bits_to_clear = field_msk;
> > > - bits_to_set = FIELD_PREP(field_msk, nval);
> > > -
> > > hwm_locked_with_pm_intel_uncore_rmw(ddat, rgadr,
> > > - bits_to_clear, bits_to_set);
> > > + PKG_PWR_LIM_1,
> > > + FIELD_PREP(PKG_PWR_LIM_1, nval));
> >
> > I don't want to give up so easily. We might have future uses for the
> > function where we want field_msk to be passed into the function (rather
> > than set inside the function as in this patch).
> >
> > Do we understand what clang is complaining about? And why this compiles
> > with gcc?
>
> Because we are not compiling the builtin functions with gcc but
> gcc has support for them. The FIELD_PREP checks if the first
> parameter is a constant:
>
> BUILD_BUG_ON_MSG(!__builtin_constant_p(_mask),
>
> where _mask was our field_mask, but we ignore it. Apparently
> clang doesn't.
So we have debunked this above.
> If we want to stick to gcc only, then I still think the patch is
> correct for two reasons:
>
> 1. it's cleaner
> 2. we would get on with the job and if one day we will decide
> to suppport builtin functions in gcc as well, we will sleep
> peacefully :)
I disagree with the patch even if we need to fix this in i915 (rather than
say change the headers or something in clang).
Note that hwm_field_scale_and_write() pairs with hwm_field_read_and_scale()
(they are basically a set/get pair) so it is desirable they have identical
arguments. This patch breaks that symmetry.
If we have to fix this in i915, I prefer the following patch (so just skip
the checks in FIELD_PREP):
@@ -112,7 +112,7 @@ hwm_field_scale_and_write(struct hwm_drvdata *ddat, i915_reg_t rgadr,
nval = DIV_ROUND_CLOSEST_ULL((u64)lval << nshift, scale_factor);
bits_to_clear = field_msk;
- bits_to_set = FIELD_PREP(field_msk, nval);
+ bits_to_set = (nval << __bf_shf(field_msk)) & field_msk;
hwm_locked_with_pm_intel_uncore_rmw(ddat, rgadr,
But I'd wait to hear from clang/llvm folks first.
> > Copying llvm@xxxxxxxxxxxxxxx too.
>
> maybe llvm folks have a better opinion.
>
Thanks.
--
Ashutosh
> > > }
> > >
> > > /*
> > > @@ -406,7 +401,6 @@ hwm_power_write(struct hwm_drvdata *ddat, u32 attr, int chan, long val)
> > > case hwmon_power_max:
> > > hwm_field_scale_and_write(ddat,
> > > hwmon->rg.pkg_rapl_limit,
> > > - PKG_PWR_LIM_1,
> > > hwmon->scl_shift_power,
> > > SF_POWER, val);
> > > return 0;
> > > --
> > > 2.37.1
> > >
