state of clang(1)'s -Wshift-count-negative and -Wshift-overflow warnings

[Alexander Best]

On Thu Nov  3 11, Dimitry Andric wrote:
> On 2011-11-03 20:03, Alexander Best wrote:
> > On Thu Nov  3 11, Dimitry Andric wrote:
> >> On 2011-11-03 11:45, Alexander Best wrote:
> >> ...
> >>> /usr/git-freebsd-head/sys/dev/ath/ath_hal/ar5210/ar5210_power.c:36:3: warning: signed shift result (0x200000000) requires 35 bits to represent, but 'int' only has 32 bits [-Wshift-overflow]
> >>>                 OS_REG_RMW_FIELD(ah, AR_SCR, AR_SCR_SLE, AR_SCR_SLE_ALLOW);
> >>>                 ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >>> /usr/git-freebsd-head/sys/dev/ath/ath_hal/ah_internal.h:471:42: note: expanded from macro 'OS_REG_RMW_FIELD'
> >>>                 (OS_REG_READ(_a, _r) &~ (_f)) | (((_v) << _f##_S) & (_f)))
> >>>                                                        ^
> >>> /usr/git-freebsd-head/sys/dev/ath/ah_osdep.h:127:49: note: expanded from macro 'OS_REG_WRITE'
> >>>             (bus_space_handle_t)(_ah)->ah_sh, (_reg), (_val))
> ...
> >> Those warnings are bogus, and due to this bug:
> 
> Actually, I was too quick with this one, since it isn't bogus.  The
> macro invocation:
> 
>   OS_REG_RMW_FIELD(ah, AR_SCR, AR_SCR_SLE, AR_SCR_SLE_ALLOW);
> 
> ultimately expands to:
> 
>   bus_space_write_4((bus_space_tag_t)(ah)->ah_st, (bus_space_handle_t)(ah)->ah_sh, (0x4004), ((bus_space_read_4((bus_space_tag_t)(ah)->ah_st, (bus_space_handle_t)(ah)->ah_sh, (0x4004)) &~ (0x00030000)) | (((0x00020000) << 16) & (0x00030000))));
> 
> The problem part is ((0x00020000) << 16), which is an overflow.  I'm not
> sure how clang concludes that the result (0x200000000) needs 35 bits to
> represent, as it seems to use 34 bits to me.  But that it doesn't fit
> into a signed integer is crystal-clear.
> 
> E.g. this is a real bug!  Probably something in the macro needs to
> explicitly cast to 64 integers, or another workaround must be found.
> 
> The other warning:
> 
> > In file included from /usr/git-freebsd-head/sys/dev/ath/ath_hal/ah_regdomain.c:99:
> > /usr/git-freebsd-head/sys/dev/ath/ath_hal/ah_regdomain/ah_rd_domains.h:69:15: warning: shift count is negative [-Wshift-count-negative]
> >          .chan11a               = BM4(F1_4950_4980,
> >                                   ^~~~~~~~~~~~~~~~~
> > /usr/git-freebsd-head/sys/dev/ath/ath_hal/ah_regdomain/ah_rd_domains.h:41:4: note: expanded from macro 'BM4'
> >           W1(_fa) | W1(_fb) | W1(_fc) | W1(_fd) }
> >           ^
> > /usr/git-freebsd-head/sys/dev/ath/ath_hal/ah_regdomain/ah_rd_domains.h:34:45: note: expanded from macro 'W1'
> >         (((_a) > 63 && (_a) < 128 ? (((uint64_t) 1)<<((_a)-64)) : (uint64_t) 0))
> >                                                    ^ ~~~~~~~~~
> 
> is indeed bogus, since the macro makes sure the shift count never
> becomes negative.  (N.B.: This only happens for global initializations,
> *not* if you would use the same macro in a function.)

and what about this one?

/usr/git-freebsd-head/sys/dev/ath/ath_hal/ah_regdomain/ah_rd_domains.h:629:15: error: shift count >= width of type [-Werror,-Wshift-count-overflow]
         .chan11a               = BM4(W2_5260_5320,
                                  ^~~~~~~~~~~~~~~~~
/usr/git-freebsd-head/sys/dev/ath/ath_hal/ah_regdomain/ah_rd_domains.h:40:34: note: expanded from macro 'BM4'
        { W0(_fa) | W0(_fb) | W0(_fc) | W0(_fd),                        \
                                        ^
/usr/git-freebsd-head/sys/dev/ath/ath_hal/ah_regdomain/ah_rd_domains.h:32:44: note: expanded from macro 'W0'
        (((_a) >= 0 && (_a) < 64 ? (((uint64_t) 1)<<(_a)) : (uint64_t) 0))
                                                  ^ ~~~~
this is being reported by -Wshift-count-overflow and not
-Wshift-count-negative, like the other ones.

cheers.
alex

> 
> 
> >>   http://llvm.org/bugs/show_bug.cgi?id=10030
> >>
> >> Unfortunately, it is still not fixed for the 3.0 release branch, and I
> >> don't expect it will be fixed for the actual release.
> > 
> > thanks for the info. so how about something like
> > 
> > diff --git a/sys/conf/kern.mk b/sys/conf/kern.mk
> > index a0a595f..3cb13de 100644
> > --- a/sys/conf/kern.mk
> > +++ b/sys/conf/kern.mk
> > @@ -1,12 +1,28 @@
> >  # $FreeBSD$
> >  
> >  #
> > -# Warning flags for compiling the kernel and components of the kernel:
> > +# XXX Disable bogus llvm warnings, complaining about perfectly valid shifts.
> > +# See http://llvm.org/bugs/show_bug.cgi?id=10030 for more details.
> > +# 
> > +.if ${CC:T:Mclang} == "clang"
> > +NOSHIFTWARNS=  -Wno-shift-count-negative -Wno-shift-count-overflow \
> > +               -Wno-shift-overflow
> > +.endif
> > +
> > 
> > ...and then add ${NOSHIFTWARNS} to the end of CWARNFLAGS?
> 
> No, this is a way too big hammer, because it eliminates the useful
> warnings together with the false positives.
> 
> It would be better to only apply this band-aid for the specific source
> files that need it, and even then, I would rather wait for the proper
> fix from upstream.