Re: [Xen-devel] [PATCH v2 01/16] xen/mm: Don't use _{g, m}fn for defining INVALID_{G, M}FN

[Tim Deegan <tim@xxxxxxx>]

At 03:18 -0600 on 23 Jun (1498187924), Jan Beulich wrote:
> >>> On 23.06.17 at 10:55, <julien.grall@xxxxxxx> wrote:
> 
> > 
> > On 23/06/17 09:30, Jan Beulich wrote:
> >>>>> On 22.06.17 at 20:31, <julien.grall@xxxxxxx> wrote:
> >>> Hi,
> >>>
> >>> On 20/06/17 11:32, Jan Beulich wrote:
> >>>>>>> On 20.06.17 at 12:06, <tim@xxxxxxx> wrote:
> >>>>> At 03:36 -0600 on 20 Jun (1497929778), Jan Beulich wrote:
> >>>>>>>>> On 20.06.17 at 11:14, <tim@xxxxxxx> wrote:
> >>>>>>> At 01:32 -0600 on 20 Jun (1497922345), Jan Beulich wrote:
> >>>>>>>>>>> On 19.06.17 at 18:57, <julien.grall@xxxxxxx> wrote:
> >>>>>>>>> --- a/xen/include/xen/mm.h
> >>>>>>>>> +++ b/xen/include/xen/mm.h
> >>>>>>>>> @@ -56,7 +56,7 @@
> >>>>>>>>>
> >>>>>>>>>  TYPE_SAFE(unsigned long, mfn);
> >>>>>>>>>  #define PRI_mfn          "05lx"
> >>>>>>>>> -#define INVALID_MFN      _mfn(~0UL)
> >>>>>>>>> +#define INVALID_MFN      (mfn_t){ ~0UL }
> >>>>>>>>
> >>>>>>>> While I don't expect anyone to wish to use a suffix expression on
> >>>>>>>> this constant, for maximum compatibility this should still be fully
> >>>>>>>> parenthesized, I think. Of course this should be easy enough to
> >>>>>>>> do while committing.
> >>>>>>>>
> >>>>>>>> Are you able to assure us that clang supports this gcc extension
> >>>>>>>> (compound literal for non-compound types)
> >>>>>>>
> >>>>>>> AIUI this is a C99 feature, not a GCCism.
> >>>>>>
> >>>>>> Most parts of it yes (it is a gcc extension in C89 mode only), but the
> >>>>>> specific use here isn't afaict: Compound literals outside of functions
> >>>>>> are static objects, and hence couldn't be used as initializers of other
> >>>>>> objects.
> >>>>>
> >>>>> Ah, I see.  So would it be better to use
> >>>>>
> >>>>>   #define INVALID_MFN ((const mfn_t) { ~0UL })
> >>>>>
> >>>>> ?
> >>>>
> >>>> While I think we should indeed consider adding the const, the above
> >>>> still is a static object, and hence still not suitable as an initializer 
> >>>> as
> >>>> per C99 or C11. But as long as gcc and clang permit it, we're fine.
> >>>
> >>> Actually this solutions breaks on GCC 4.9 provided by Linaro ([1]
> >>> 4.9-2016-02 and 4.9-2017.01).
> >>>
> >>> This small reproducer does not compile with -std=gnu99 (used by Xen) but
> >>> compile with this option. Jan, have you tried 4.9 with this patch?
> >>
> >> That's sort of an odd question - you've sent the patch, so I would
> >> have expected you to have made sure it doesn't break (and
> >> while it was me to add the const, this was discussed, and you don't
> >> make clear whether that's the issue). In any event, I've tried ...
> >>
> >>> typedef struct
> >>> {
> >>>      unsigned long i;
> >>> } mfn_t;
> >>>
> >>> mfn_t v = (const mfn_t){~0UL};
> >>
> >> ... this now with 7.1.0, 6.3.0, 5.4.0, 5.2.0, and 4.9.3, and all
> >> of them compile this without errors or warnings (at -Wall -W).
> > 
> > Actually did you build with -std=gnu99? I just tried 4.9.3 for x86 and
> > also 4.8 for ARM64 on Ubuntu Trusty. Both are broken.
> 
> Ah, indeed - that fails with 4.9.3 but succeeds with 5.2.0. And
> it's not the const getting in the way here. I notice this difference
> in their documentation (4.9.3 first, then 7.1.0):
> 
> Compound literals for scalar types and union types are also allowed,
> but then the compound literal is equivalent to a cast.
> 
> Compound literals for scalar types and union types are also allowed.
> In the following example the variable i is initialized to the value 2,
> the result of incrementing the unnamed object created by the
> compound literal.
> 
>       int i = ++(int) { 1 };
> 
> It is especially this example clarifying that newer compilers don't
> treat this like a cast anymore (albeit a casted expression alone is
> fine as initializer in 4.9.3, so there must be more to the failure).
> 
> While I still view this as a compiler bug (as it accepts the code in
> default mode), as a workaround  I guess we'll need to accept a
> gcc < 5 conditional in the header, which we would really have
> wanted to avoid.

Since we'll have to make some scheme that works for 4.9, I think we
should just use that for all versions.

How about:
 - keep INVALID_MFN as an inline function call for most uses;
 - #define INVALID_MFN_INITIALIZER { ~0UL } for when we need a
   real constant initializer aat file scope.

Tim.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel