Re: [PATCH] kvm: x86: mmu: Add cast to negated bitmasks in update_permission_bitmask()

[Matthias Kaehlcke]

On Tue, Jun 19, 2018 at 12:11:27PM -0700, Joe Perches wrote:
> On Tue, 2018-06-19 at 11:36 -0700, Matthias Kaehlcke wrote:
> > On Tue, Jun 19, 2018 at 11:07:47AM -0700, Joe Perches wrote:
> > > On Tue, 2018-06-19 at 19:35 +0200, Paolo Bonzini wrote:
> > > > On 19/06/2018 19:23, Joe Perches wrote:
> > > > > On Tue, 2018-06-19 at 10:08 -0700, Nick Desaulniers wrote:
> > > > > > On Tue, Jun 19, 2018 at 8:19 AM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
> > > > > > > 
> > > > > > > On 15/06/2018 20:45, Nick Desaulniers wrote:
> > > > > > > > > 
> > > > > > > > > > In any case I think it it preferable to fix the code over disabling
> > > > > > > > > > the warning, unless the warning is bogus or there are just too many
> > > > > > > > > > occurrences.
> > > > > > > > > 
> > > > > > > > > Maybe.
> > > > > > > > 
> > > > > > > > Spurious warning today, actual bug tomorrow?  I prefer to not to
> > > > > > > > disable warnings wholesale.  They don't need to find actual bugs to be
> > > > > > > > useful.  Flagging code that can be further specified does not hurt.
> > > > > > > > Part of the effort to compile the kernel with different compilers is
> > > > > > > > to add warning coverage, not remove it.  That said, there may be
> > > > > > > > warnings that are never useful (or at least due to some invariant that
> > > > > > > > only affects the kernel).  I cant think of any off the top of my head,
> > > > > > > > but I'm also not sure this is one.
> > > > > > > 
> > > > > > > This one really makes the code uglier though, so I'm not really inclined
> > > > > > > to applying the patch.
> > > > > > 
> > > > > > Note that of the three variables (w, u, x), only u is used later on.
> > > > > > What about declaring them as negated with the cast, that way there's
> > > > > > no cast in a ternary?
> > > > > 
> > > > > It'd be simpler to cast in the BYTE_MASK macro itself
> > > > 
> > > > I don't think that would work, as the ~ would be done on a zero-extended
> > > > signed int.
> > > 
> > > True, but the whole concept is dubious.
> > > The implicit casts are all over the place,
> > > not just in the file.
> > 
> > Would that have been any different with the solution you proposed (if
> > it worked)?
> > 
> > Apparently both gcc and clang limit the warning to the potentially
> > more problematic case where a value with sign bit is promoted.
> 
> I think the warning is exactly equivalent to -Wsign-conversion
> and we don't normally compile the kernel with that either.

I disagree with "exactly equivalent". With -Wsign-conversion a warning
is generated whenever a signed type is assigned to an unsigned
variable or viceversa.

-Wconstant-conversion is only issued when a *constant value* is assigned to
an incompatible type.

> Trying to allow a "make W=3" to be compiler warning message free
> is also silly.
> 
> I think it's better to make the warning emitted only at a W=3
> level instead.

Another difference with -Wsign-conversion is that enabling it would
probably result in thousands of warnings. Do you have evidence that
there is a significant number of spurious -Wconstant-conversion
warnings?