On Tue, Sep 19, 2023 at 05:25:20PM -0400, Jason Merrill wrote:
> On 9/1/23 08:22, Nathaniel Shead wrote:
> > On Wed, Aug 30, 2023 at 04:28:18PM -0400, Jason Merrill wrote:
> > > On 8/29/23 09:35, Nathaniel Shead wrote:
> > > > This is an attempt to improve the constexpr machinery's handling of
> > > > union lifetime by catching more cases that cause UB. Is this approach
> > > > OK?
> > > >
> > > > I'd also like some feedback on a couple of pain points with this
> > > > implementation; in particular, is there a good way to detect if a type
> > > > has a non-deleted trivial constructor? I've used 'is_trivially_xible' in
> > > > this patch, but that also checks for a trivial destructor which by my
> > > > reading of [class.union.general]p5 is possibly incorrect. Checking for a
> > > > trivial default constructor doesn't seem too hard but I couldn't find a
> > > > good way of checking if that constructor is deleted.
> > >
> > > I guess the simplest would be
> > >
> > > (TYPE_HAS_TRIVIAL_DFLT (t) && locate_ctor (t))
> > >
> > > because locate_ctor returns null for a deleted default ctor. It would be
> > > good to make this a separate predicate.
> > >
> > > > I'm also generally unsatisfied with the additional complexity with the
> > > > third 'refs' argument in 'cxx_eval_store_expression' being pushed and
> > > > popped; would it be better to replace this with a vector of some
> > > > specific structure type for the data that needs to be passed on?
> > >
> > > Perhaps, but what you have here is fine. Another possibility would be to
> > > just have a vec of the refs and extract the index from the ref later as
> > > needed.
> > >
> > > Jason
> > >
> >
> > Thanks for the feedback. I've kept the refs as-is for now. I've also
> > cleaned up a couple of other typos I'd had with comments and diagnostics.
> >
> > Bootstrapped and regtested on x86_64-pc-linux-gnu.
> >
> > @@ -6192,10 +6197,16 @@ cxx_eval_store_expression (const constexpr_ctx
> > *ctx, tree t,
> > type = reftype;
> > - if (code == UNION_TYPE && CONSTRUCTOR_NELTS (*valp)
> > - && CONSTRUCTOR_ELT (*valp, 0)->index != index)
> > + if (code == UNION_TYPE
> > + && TREE_CODE (t) == MODIFY_EXPR
> > + && (CONSTRUCTOR_NELTS (*valp) == 0
> > + || CONSTRUCTOR_ELT (*valp, 0)->index != index))
> > {
> > - if (cxx_dialect < cxx20)
> > + /* We changed the active member of a union. Ensure that this is
> > + valid. */
> > + bool has_active_member = CONSTRUCTOR_NELTS (*valp) != 0;
> > + tree inner = strip_array_types (reftype);
> > + if (has_active_member && cxx_dialect < cxx20)
> > {
> > if (!ctx->quiet)
> > error_at (cp_expr_loc_or_input_loc (t),
>
> While we're looking at this area, this error message should really mention
> that it's allowed in C++20.
>
> > @@ -6205,8 +6216,36 @@ cxx_eval_store_expression (const constexpr_ctx *ctx,
> > tree t,
> > index);
> > *non_constant_p = true;
> > }
> > - else if (TREE_CODE (t) == MODIFY_EXPR
> > - && CONSTRUCTOR_NO_CLEARING (*valp))
> > + else if (!is_access_expr
> > + || (CLASS_TYPE_P (inner)
> > + && !type_has_non_deleted_trivial_default_ctor (inner)))
> > + {
> > + /* Diagnose changing active union member after initialisation
> > + without a valid member access expression, as described in
> > + [class.union.general] p5. */
> > + if (!ctx->quiet)
> > + {
> > + if (has_active_member)
> > + error_at (cp_expr_loc_or_input_loc (t),
> > + "accessing %qD member instead of initialized "
> > + "%qD member in constant expression",
> > + index, CONSTRUCTOR_ELT (*valp, 0)->index);
> > + else
> > + error_at (cp_expr_loc_or_input_loc (t),
> > + "accessing uninitialized member %qD",
> > + index);
> > + if (is_access_expr)
> > + {
> > + inform (DECL_SOURCE_LOCATION (index),
> > + "%qD does not implicitly begin its lifetime "
> > + "because %qT does not have a non-deleted "
> > + "trivial default constructor",
> > + index, inner);
> > + }
>
> The !is_access_expr case could also use an explanatory message.
Thanks for the review, I've updated these messages and will send through
an updated patch once bootstrap/regtest is complete.
> Also, I notice that this testcase crashes with the patch:
>
> union U { int i; float f; };
> constexpr auto g (U u) { return (u.i = 42); }
> static_assert (g({.f = 3.14}) == 42);
This appears to segfault even without the patch since GCC 13.1.
https://godbolt.org/z/45sPh8WaK
I haven't done a bisect yet to work out what commit exactly caused this.
Should I aim to fix this first before coming back with this patch?
Thanks,
Nathaniel
