Re: Building with recent GCC versions: gdbsupport/gdb_assert.h:35:4: error: 'nonnull' argument 'filename' compared to NULL [-Werror=nonnull-compare]

[Andrew Burgess <andrew.burgess@embecosm.com>]

* Tom de Vries <tdevries@suse.de> [2021-07-27 15:38:19 +0200]:

> On 7/27/21 1:49 PM, Tom de Vries wrote:
> > On 7/27/21 1:35 PM, Andrew Burgess wrote:
> >> * Tom de Vries <tdevries@suse.de> [2021-07-27 12:44:10 +0200]:
> >>
> >>> On 7/27/21 12:03 PM, Andrew Burgess wrote:
> >>>> * Jan-Benedict Glaw <jbglaw@lug-owl.de> [2021-07-26 23:11:01 +0200]:
> >>>>
> >>>>> Hi!
> >>>>>
> >>>>> I'm running some CI builds and noticed that, when building GDB with
> >>>>> quite recent GCC versions, it breaks.
> >>>>>
> >>>>> With ie. this "gcc-snapshot" GCC from Debian:
> >>>>>
> >>>>> /usr/lib/gcc-snapshot/bin/gcc --version
> >>>>> gcc (Debian 20210630-1) 12.0.0 20210630 (experimental) [master revision 6bf383c37e6:93c270320bb:35da8a98026849bd20d16bbf9210ac1d0b44ea6a]
> >>>>>
> >>>>> we see:
> >>>>>
> >>>>> ./configure --target=i686-linux --prefix=/tmp/gdb-i686-linux
> >>>>> [...]
> >>>>> all make V=1 all-gdb
> >>>>> [...]
> >>>>> [all 2021-07-26 20:39:22] /usr/lib/gcc-snapshot/bin/g++ -x c++    -I. -I. -I./config -DLOCALEDIR="\"/tmp/gdb-i686-linux/share/locale\"" -DHAVE_CONFIG_H -I./../include/opcode -I./../readline/readline/.. -I./../zlib -I../bfd -I./../bfd -I./../include -I../libdecnumber -I./../libdecnumber  -I./../gnulib/import -I../gnulib/import -I./.. -I..  -DTUI=1    -I./.. -pthread  -Wall -Wpointer-arith -Wno-unused -Wunused-value -Wunused-variable -Wunused-function -Wno-switch -Wno-char-subscripts -Wempty-body -Wunused-but-set-parameter -Wunused-but-set-variable -Wno-sign-compare -Wno-error=maybe-uninitialized -Wno-mismatched-tags -Wsuggest-override -Wimplicit-fallthrough=3 -Wduplicated-cond -Wshadow=local -Wdeprecated-copy -Wdeprecated-copy-dtor -Wredundant-move -Wmissing-declarations -Wstrict-null-sentinel -Wformat -Wformat-nonliteral -Werror -g -O2   -c -o compile/compile.o -MT compile/compile.o -MMD -MP -MF compile/.deps/compile.Tpo compile/compile.c
> >>>>> [all 2021-07-26 20:39:26] In file included from ./../gdbsupport/common-defs.h:126,
> >>>>> [all 2021-07-26 20:39:26]                  from ./defs.h:28,
> >>>>> [all 2021-07-26 20:39:26]                  from compile/compile.c:20:
> >>>>> [all 2021-07-26 20:39:26] ./../gdbsupport/gdb_unlinker.h: In constructor 'gdb::unlinker::unlinker(const char*)':
> >>>>> [all 2021-07-26 20:39:26] ./../gdbsupport/gdb_assert.h:35:4: error: 'nonnull' argument 'filename' compared to NULL [-Werror=nonnull-compare]
> >>>>> [all 2021-07-26 20:39:26]    35 |   ((void) ((expr) ? 0 :                                                       \
> >>>>> [all 2021-07-26 20:39:26]       |   ~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >>>>> [all 2021-07-26 20:39:26]    36 |            (gdb_assert_fail (#expr, __FILE__, __LINE__, FUNCTION_NAME), 0)))
> >>>>> [all 2021-07-26 20:39:26]       |            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >>>>> [all 2021-07-26 20:39:26] ./../gdbsupport/gdb_unlinker.h:38:5: note: in expansion of macro 'gdb_assert'
> >>>>> [all 2021-07-26 20:39:26]    38 |     gdb_assert (filename != NULL);
> >>>>> [all 2021-07-26 20:39:26]       |     ^~~~~~~~~~
> >>>>> [all 2021-07-26 20:39:27] cc1plus: all warnings being treated as errors
> >>>>> [all 2021-07-26 20:39:27] make[1]: *** [Makefile:1642: compile/compile.o] Error 1
> >>>>> [all 2021-07-26 20:39:27] make[1]: Leaving directory '/var/lib/laminar/run/gdb-i686-linux/4/binutils-gdb/gdb'
> >>>>> [all 2021-07-26 20:39:27] make: *** [Makefile:11410: all-gdb] Error 2
> >>>>>
> >>>>>
> >>>>> I also discussed this on the GCC patches mailing list
> >>>>> (https://gcc.gnu.org/pipermail/gcc-patches/2021-July/575568.html),
> >>>>> where Martin suggested that this should be fixed here in GDB.
> >>>>>
> >>>>> Any thoughts about this?
> >>>>
> >>>> As I understand it the nonnull attribute only provides compile time
> >>>> protection against explicitly passing NULL, there's no compiled in
> >>>> non-null check (well, maybe with -fisolate-erroneous-paths-attribute,
> >>>> but the assert might give a better error message).
> >>>>
> >>>> This means its still possible to pass NULL to a nonnull function, its
> >>>> just the behaviour of the program is undefined in that case.
> >>>>
> >>>> So, it doesn't seem crazy that we might want to both (a) have a
> >>>> function declared nonnull, to prevent explicitly passing NULL, and (b)
> >>>> have a NULL check inside the function to catch logic bugs that result
> >>>> in NULL being passed.
> >>>>
> >>>> We could, of course, push the assert outside of the function, but that
> >>>> would really suck due to code duplication, and the risk of missing an
> >>>> assert, so that seems like a non-starter.
> >>>>
> >>>> We could drop either the assert, or the nonnull attribute, but that
> >>>> would suck as both give a valuable, but different form of protection.
> >>>>
> >>>> After some experimenting, I suspect that the assert is being optimised
> >>>> away anyway, which kind of makes sense, as we're telling the compiler
> >>>> it can assume that the pointer is non-null.
> >>>>
> >>>
> >>> Yes, in fact that's what the nonnull-compare warning specifically warns
> >>> against: there's some code that may be optimized away, due to the
> >>> nonnull attribute.
> >>>
> >>>> So, what we probably want is someway to tell (or trick) GCC into
> >>>> including the null check even in the nonnull function....
> >>>>
> >>>> ... here's what I came up with, add this somewhere:
> >>>>
> >>>>  template<typename T>
> >>>>  bool __attribute__ ((noinline))
> >>>>  nonnull_arg_is_really_not_null (const T *ptr)
> >>>>  {
> >>>>    return ptr != nullptr;
> >>>>  }
> >>>>
> >>>> then change the assert to:
> >>>>
> >>>>  gdb_assert (nonnull_arg_is_really_not_null (filename));
> >>>>
> >>>> Seems to keep the assert, and silence the warning.  Thoughts?
> >>>>
> >>>
> >>> I understand why it works, but it seems fragile to me.  At some point
> >>> some compiler may get smart enough to also optimize this case, and then
> >>> we're back in the same situation.
> >>
> >> Good point.
> >>
> >> The GCC documentation for noinline[1] suggests we can avoid the call
> >> being removed by adding 'asm ("");' into the function:
> >>
> >>   template<typename T>
> >>   bool __attribute__ ((noinline))
> >>   nonnull_arg_is_really_not_null (const T *ptr)
> >>   {
> >>     asm ("");
> >>     return ptr != nullptr;
> >>   }
> >>
> >> I'm not really arguing for this approach over any other, just sharing
> >> what I discovered.
> >>
> > 
> > Ack, understood.  Note that the added asm doesn't stop a compiler from
> > doing:
> > ...
> > gdb_assert (nonnull_arg_is_really_not_null (filename));
> > ...
> > ->
> > ...
> > nonnull_arg_is_really_not_null (filename);
> > gdb_assert (true);
> > ...
> > 
> > Thanks,
> > - Tom
> > 
> >> Thanks,
> >> Andrew
> >>
> >> [1] https://gcc.gnu.org/onlinedocs/gcc-11.1.0/gcc/Common-Function-Attributes.html#Common-Function-Attributes
> >>
> >>>
> >>> I wonder whether using volatile is a better idea (can't try this out
> >>> right now).
> >>>
> 
> I was thinking of something like this:
> ...
> diff --git a/gdbsupport/gdb_unlinker.h b/gdbsupport/gdb_unlinker.h
> index bda6fe7ab54..3d99b41e7ad 100644
> --- a/gdbsupport/gdb_unlinker.h
> +++ b/gdbsupport/gdb_unlinker.h
> @@ -20,6 +20,13 @@
>  #ifndef COMMON_GDB_UNLINKER_H
>  #define COMMON_GDB_UNLINKER_H
> 
> +template<typename T>
> +const T *volatile
> +ignore_nonnull (const T *ptr)
> +{
> +  return ptr;
> +}
> +
>  namespace gdb
>  {
> 
> @@ -35,7 +42,7 @@ class unlinker
>    unlinker (const char *filename) ATTRIBUTE_NONNULL (2)
>      : m_filename (filename)
>    {
> -    gdb_assert (filename != NULL);
> +    gdb_assert (ignore_nonnull (filename) != NULL);
>    }
> 
>    ~unlinker ()
> ...
> 
> This builds for me, though I haven't got a setup yet where the warning
> reproduces, so I can't check whether it actually fixes things.

I've been testing issues like this using:

  https://godbolt.org/z/nfhq6zb7q

Your suggestion gives this error:

  error: 'volatile'-qualified return type is deprecated [-Werror=volatile]
     21 | const T * volatile
        | ^~~~~
  cc1plus: all warnings being treated as errors
  Compiler returned: 1

If we remove the volatile return type then of course GCC inlines and
optimises out the assert.  We could make the 'ignore_nonnull'
noinline, but then we're basically back at my original suggestion.

I suspect the only choice (right now) might be to do:

  template<typename T>
  void __attribute__ ((noinline))
  assert_nonnull (const T *ptr)
  {
    asm ("");
    gdb_assert (ptr != nullptr);
  }

Then replace gdb_assert with 'assert_nonnull (filename)'.

What we'd actually want is for 'assert_nonnull' to be a macro that
passes through the file/function/line just like the existing assert
does so that the failed assert can be reported in the correct place.

The more I look at this the more it feels like this is something GCC
should be able to help us with...

Thanks,
Andrew