On 2/25/21 6:47 PM, Jeff Law wrote:
On 2/23/21 2:52 PM, Jason Merrill wrote:
I don't see how the issue here depends on the artificiality of the vptr;
That's what I was trying to get at -- is DECL_ARTIFICIAL really a good
way to detect these kinds of cases. It sounds like it likely isn't, but
it may also be the best we can do. I'm not sure yet.
I'd expect to see the same problem with a data member. The problem is
that a D base subobject is smaller than a complete D object, and in
this case the base subobject is allocated such that if it were a full
D object, it would overlap the end of E. And we're checking the
MEM_REF as though accessing a full D object, so we get a warning.
The general issue about the confusion between complete and as-base
types is PR 22488; I have various thoughts about a proper fix, but
there have always been higher-priority things to work on.
One possible approach to this PR is that we don't need to check an
intermediate _REF; in
MEM[(struct D *)&D.2651 + 24B]._vptr.D
we aren't loading a whole D, we're only looking at the _vptr, which is
within the bounds of E.
That raises a really interesting question. What's the right bounds to
use with these kinds of constructs. Should we be looking at D's type
(struct E) or struct D? I'm not sure there's a one-size-fits-all answer
here and I don't think we have any other context go help guide us.
struct E, definitely; the actual bounds of the struct D object are
unknown here. Unless, I suppose, we were to work harder to find the D
FIELD_DECL in E that corresponds to this offset and get its DECL_SIZE, but:
We know the bounds of the E object, and we can determine that _vptr.D is
within those bounds; that seems sufficient to decide not to warn.
Jason
