Re: [PATCH] c++: Fix wrong modifying const object error for COMPONENT_REF [PR94074]

[Marek Polacek <polacek@redhat.com>]

On Mon, Mar 09, 2020 at 04:25:00PM -0400, Marek Polacek wrote:
> On Mon, Mar 09, 2020 at 03:37:56PM -0400, Jason Merrill wrote:
> > On 3/9/20 9:40 AM, Marek Polacek wrote:
> > > On Mon, Mar 09, 2020 at 09:19:30AM -0400, Jason Merrill wrote:
> > > > On 3/9/20 8:58 AM, Jakub Jelinek wrote:
> > > > > On Fri, Mar 06, 2020 at 07:43:43PM -0500, Jason Merrill wrote:
> > > > > > On 3/6/20 6:54 PM, Marek Polacek wrote:
> > > > > > > I got a report that building Chromium fails with the "modifying a const
> > > > > > > object" error.  After some poking I realized it's a bug in GCC, not in
> > > > > > > their codebase.
> > > > > > > 
> > > > > > > Much like with ARRAY_REFs, which can be const even though the array
> > > > > > > itself isn't, COMPONENT_REFs can be const although neither the object
> > > > > > > nor the field were declared const.  So let's dial down the checking.
> > > > > > > Here the COMPONENT_REF was const because of the "const_cast<const U &>(m)"
> > > > > > > thing -- cxx_eval_component_reference then builds a COMPONENT_REF with
> > > > > > > TREE_TYPE (t).
> > > > > > 
> > > > > > What is folding the const into the COMPONENT_REF?
> > > > > 
> > > > > cxx_eval_component_reference when it is called on
> > > > > ((const struct array *) this)->elems
> > > > > with /*lval=*/true and lval is true because we are evaluating
> > > > > <retval> = (const int &) &((const struct array *) this)->elems[VIEW_CONVERT_EXPR<size_t>(n)];
> > > > 
> > > > Ah, sure.  We're pretty loose with cv-quals in the constexpr code in
> > > > general, so it's probably not worth trying to change that here.  Getting
> > > > back to the patch:
> > > 
> > > Yes, here the additional const was caused by a const_cast adding a const.
> > > 
> > > But this could also happen with wrapper functions like this one from
> > > __array_traits in std::array:
> > > 
> > >        static constexpr _Tp&
> > >        _S_ref(const _Type& __t, std::size_t __n) noexcept
> > >        { return const_cast<_Tp&>(__t[__n]); }
> > > 
> > > where the ref-to-const parameter added the const.
> > > 
> > > > > +      if (TREE_CODE (obj) == COMPONENT_REF)
> > > > > +	{
> > > > > +	  tree op1 = TREE_OPERAND (obj, 1);
> > > > > +	  if (CP_TYPE_CONST_P (TREE_TYPE (op1)))
> > > > > +	    return true;
> > > > > +	  else
> > > > > +	    {
> > > > > +	      tree op0 = TREE_OPERAND (obj, 0);
> > > > > +	      /* The LHS of . or -> might itself be a COMPONENT_REF.  */
> > > > > +	      if (TREE_CODE (op0) == COMPONENT_REF)
> > > > > +		op0 = TREE_OPERAND (op0, 1);
> > > > > +	      return CP_TYPE_CONST_P (TREE_TYPE (op0));
> > > > > +	    }
> > > > > +	}
> > > > 
> > > > Shouldn't this be a loop?
> > > 
> > > I don't think so, though my earlier patch had a call to
> > > 
> > > +static bool
> > > +cref_has_const_field (tree ref)
> > > +{
> > > +  while (TREE_CODE (ref) == COMPONENT_REF)
> > > +    {
> > > +      if (CP_TYPE_CONST_P (TREE_TYPE (TREE_OPERAND (ref, 1))))
> > > +       return true;
> > > +      ref = TREE_OPERAND (ref, 0);
> > > +    }
> > > +  return false;
> > > +}
> > 
> > > here.  A problem arised when I checked even the outermost expression (which is not a
> > > field_decl), then I saw another problematical error.
> > > 
> > > The more outer fields are expected to be checked in subsequent calls to
> > > modifying_const_object_p in next iterations of the
> > > 
> > > 4459   for (tree probe = target; object == NULL_TREE; )
> > > 
> > > loop in cxx_eval_store_expression.
> > 
> > OK, but then why do you want to check two levels here rather than just one?
> 
> It's a hack to keep constexpr-tracking-const7.C working.  There we have
> 
>   b.a.c.d.n
> 
> wherein 'd' is const struct D, but 'n' isn't const.  Without the hack
> const_object_being_modified would be 'b.a.c.d', but due to the problem I
> desribed in the original mail[1] the constructor for D wouldn't have
> TREE_READONLY set.  With the hack const_object_being_modified will be
> 'b.a.c.d.n', which is of non-class type so we error:
> 
> 4710       if (!CLASS_TYPE_P (const_objtype))
> 4711         fail = true;
> 
> I could remove the hack and maybe XFAIL constexpr-tracking-const7.C if you
> want.  Unfortunately I wasn't aware of [1] when I added that feature and
> checking if the whole COMPONENT_REF is const seemed to be enough.
> 
> It's probably not a good idea to make this checking more strict at this
> stage.
> 
> [1] "While looking into this I noticed that we don't detect modifying a const
> object in certain cases like in
> <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94074#c2>.  That's because
> we never evaluate an X::X() CALL_EXPR -- there's none.  So there's no
> CONSTRUCTOR to set TREE_READONLY on.  No idea how to fix this, but it's
> likely something for GCC 11 anyway."

The testcase disappeared from Bugzilla, but it was
<https://paste.centos.org/view/fc9527f6>.

Marek