Re: [PATCH]: Fix PR tree-optimization/21407

["Giovanni Bajo" <rasky at develer dot com>]

Kenneth Zadeck <zadeck@naturalbridge.com> wrote:

>> Also, it would seem to me that Ada and Java probably have upcast, and
>> that the back end has to work for them too...  I just don't know if
>> they would trip over it.

> At least for java, there is a lot of structure that keep things easy to
> understand.  For one the upcasting is allowed specificly in the context
> of the class heirarchy.  This is different from just taking any field
> that happens to be in a record and recovering a reference to the
> record.    With respect to the gcc implementation, or analysis always
> assumed that if you had a pointer to field, you also had a pointer to
> every kind of object you could get by chaining thru the TYPE_BINFO
> structure.

Notice that exactly the same holds true in C++ for non-POD classes (which
usually are most of the classes in a given translation unit), because
offsetof is only allowed for POD types. Also, non-PODness is viral, which
means that if you embed a non-POD into a structure, you get a non-POD.

In other words, you can reasonably use TBAA in C++ if you limit it to
non-POD types and consider a pointer to non-POD "class A" also a pointer to
any other class in the whole inheritance DAG. And a pointer to a POD "class
B" is also a pointer to any other POD class, but it is *not* a pointer to
non-POD classes.

It seems to me that a language hook should be added to control this
optimization, as it is deeply interwinded with language-specific semantics.
-- 
Giovanni Bajo