In article <4ciauq$2e1@trojan.neta.com> jonnyg@trojan.neta.com (J. Greenblatt) writes:

>  // This is a non RTTI enabled class, no RTTI overhead, no RTTI keywords
>  class foo
>  {
>      int i;
>  };

Actually, that's already more or less the case.  RTTI basically
consists of two parts, dynamic_cast and typeid; in both cases, you
need to distinguish between an expression's static type (its type as
known at compile time) and its dynamic type (the type of an object
pointed to by a pointer or reference).  The two can differ, of course,
since a base class pointer can point to an object of derived class.

If you're simply working with an expression's static type, then there
doesn't need to be any runtime overhead: everything can be done at
compile time.  There's only overhead if you want to find the dynamic
type of an object: that's the only case when you're really doing RTTI
at all.

The standard distinguishes between polymorphic classes (ones that
declare or inherit virtual functions) and nonpolymorphic classes.  You
can still use dynamic_cast and typeid for nonpolymorphic classes, but
only to work with an expression's static type.  The section on
dynamic_cast says "Otherwise [i.e. for downcasting and conversion from
void*], v shall be a pointer to or an lvalue of a polymorphic type",
and the section on typeid says "If the expression is an lvalue of
polymorphic type (_class.virtual_), the type_info for the complete
object (_class.base.init_) referred to is the result."

There is already overhead associated with polymorphic classes (the
virtual table); RTTI adds a little bit more, but it shouldn't add any
overhead to nonpolymorphic classes, which are the ones where
presumably one cares about such things.

Polymorphic classes are defined in section 10.3 of the working paper,
and dynamic_cast and typeid are described in sections 5.2.6 and 5.2.7
respectively.


Matt Austern
SGI: MTI Compilers Group
austern@isolde.mti.sgi.com
--
Matt Austern
SGI: MTI Compilers Group
austern@isolde.mti.sgi.com

 I have been discussing RTTI with a person at work and have a
suggestion I would like to share.

Problem:

 It seems that RTTI requires a compile time flag to be activated. When
activated the code seems to bloat whether you actualy use the feature in the
specific peice of code or not. From what I have learned this overhead is due
to extra work done when the object is constructed (A pointer to this with a
string representation of the class is added to a table). This overhead does
not make RTTI a good addition to C++ but more of a optional addition to those
who want it (this is how it is implemented now). Also are the addition of key
words which is not a biggy in this case given the key word names.

Proposed solution:

 Allow RTTI to be implemented on the whole system as it is now and keep
it as a option. In addition allow RTTI to be turned on on a class by class
basis. On a class by class basis RTTI would be enabled if a base of the
current class is declared with the RTTI keyword. The extra RTTI keywords
(dynamic_cast, static_cast, etc...) are only defined for classes declared
as RTTI classes. The RTTI overhead is now only felt when accessing objects
of RTTI classes.

Example:

 // This is a non RTTI enabled class, no RTTI overhead, no RTTI keywords
 class foo
 {
     int i;
 };

 // This is a RTTI class with the RTTI overhead, RTTI keywords enabled.
 RTTI class bar
 {
     int i;
 };

 // This is also a RTTI class by inheritence
 class baz: public bar
 {
    int k;
 };

Conclusion:
 I may have not considered all the issues here but I think the RTTI
keyword would be a good way of using RTTI without being hit with the extra
overhead when not needed. If the RTTI compiler option is used, all classes
are RTTI classes whether RTTI is specified or not. Some vendors may want
to experiment with it or add this capability as an extension to their
compilers. If the RTTI keyword is used as an "extension", you could #define
RTTI to be nothing on compilers that do not support it and turn the RTTI
compiler flag making all classes RTTI, this will allow your code to be
portable accross compilers that do not support the conditional RTTI keyword.

 A lot of what I seen of RTTI is based on MicroSlof so other
implementations or RTTI may or may not have the problems I described in
the problem section.

     JonG.

In article <4ciauq$2e1@trojan.neta.com> jgreen@amex-trs.com writes:

>  It seems that RTTI requires a compile time flag to be activated. When
> activated the code seems to bloat whether you actualy use the feature in the
> specific peice of code or not. From what I have learned this overhead is due
> to extra work done when the object is constructed (A pointer to this with a
> string representation of the class is added to a table). This overhead does
> not make RTTI a good addition to C++ but more of a optional addition to those
> who want it (this is how it is implemented now). Also are the addition of key
> words which is not a biggy in this case given the key word names.

There's nothing I can think of in the language that requires
it to be implemented that way.  It is intended to be possible
to implement RTTI by adding an extra field to the virtual function
table, which means there is a small amount of space overhead per
class (not per object) and no time overhead except for the places
in the program where RTTI is actually used.  I can see no reason why
any extra work should have to be done when objects are constructed.
--
    --Andrew Koenig
      ark@research.att.com