[boost] [numeric_traits] How to define integer difference_types

From: Sebastian Theophil (stheophil_at_[hidden])
Date: 2007-11-21 09:05:07

Dear all,

Yesterday I had run into issues compiling boost::counting_iterator under
Visual Studio. The reason is  integer_traits<Integer>::difference_type
which is used as a default difference_type for the counting_iterator. 

This has already been discussed here:
http://groups.google.de/group/boost-list/browse_thread/thread/79b5a5bb13
f37576/7773443825609bb1
But I don't think the discussion went far enough:

- First of all, this is not just an issue concerning compiler warnings
about 64 bit conversion issues (/Wp64) as Orhun suggested in the above
thread.
- Trying to convert __int64 to int is a level 1 compiler warning
(http://msdn2.microsoft.com/en-us/library/aa984663(VS.71).aspx) and in
my opinion, the compiler should warn about implicit down-casts.
- David suggested in the thread, that (quote) "These 'integer size
warnings' are nothing more than a nuisance in most generic code.  IMO
they should be turned off on the command line."

To this, I strongly object. The integer size warning is correct and may
be a simple programming error that a developer should at least look at.
It is the supposed greater generality of defining integer_traits<int> to
be __int64 that causes a problem here.

- Defining integer_traits<int>::difference_type as __int64 is first of
all pretty inconsistent
- integer_traits<__int64>::difference_type of course also is __int64 for
lack of a larger type 

- *Most importantly, no one defines std::ptrdiff_t as __int64 although
the same reasoning would apply*

Integer values may overflow. Even __int64 values may overflow. 

There may be cases where you need (int - int) to be a __int64 just like
(int * int) should be __int64. In general however, I would like int +
(int - int) to be an int again. As a developer I know this can overflow
but int = int + __int64(int - int) will overflow too and if I do __int64
+ (__int64 - __int64) then this may overflow too!

(Besides, in the context of iterators it is extremely weird for an
iterator on ints to have an operation advance(__int64).)

There's no general solution preventing ints from overflowing. The
numeric_traits implementation forces me to disable perfectly valid
static type checks for no apparent value at all. Let me take care of
overflowing problems where *I* think they may occur. Let the compiler do
proper type checks.

Regards and I'm looking forward to comments on this,
Sebastian

--
Sebastian Theophil . stheophil_at_[hidden]
Software Engineer
think-cell Software GmbH . Invalidenstr. 34 . 10115 Berlin, Germany 
http://www.think-cell.com . phone [phone] . fax [phone]
Geschaftsfuhrer: Dr. Markus Hannebauer, Arno Schoedl . Amtsgericht Charlottenburg, HRB 85229