From: David Abrahams (dave_at_[hidden])
Date: 2006-05-17 09:54:43
"Robert Ramey" <ramey_at_[hidden]> writes:
> David Abrahams wrote:
>> "Robert Ramey" <ramey_at_[hidden]> writes:
>>> c) "The IEEE standard strongly recommends that implementations allow
>>> trap handlers to be installed."
>>> C++ doesn't permit this.
>> Incorrect. C++ absolutely does permit implementations to allow trap
>> handlers to be installed. C++ simply does not require it.
> My basis for citing this is page 122 of "The C++ Progamming Language"
> by Stroustrup, copyright 2000, reprinted May 2003 with corrections.
> It says, "In particular, underflow, overflow and division by zero do not
> throw standard exceptions". If that's wrong, incomplete or out
> of date, I would be curious to know about it.
What do you want me to tell you about it? B.S. was probably writing
colloquially, as in "there is no guarantee that the implementation
will throw a standard exception."
> It seems to comport with my personal experience with C++ numeric
That means nothing about what implementations are allowed to do, and
I'm pretty sure I can set VC++ up to throw a C++ exception in these
cases. Yep, there it is: http://tinyurl.com/9my88
> Feel free to expand upon this.
No, I'm not gonna expand upon it.
Don't make me do the legwork; get a copy of standard and read what it
says. Describing to you what's plainly written in the standard until
you believe me is a big time and bandwidth waster.
>>> d) "Another ambiguity in most language definitions concerns what
>>> happens on overflow, underflow and other exceptions. The IEEE
>>> standard precisely specifies the behavior of exceptions, and so
>>> languages that use the standard as a model can avoid any ambiguity
>>> on this point. " But C++ doesn't permit exceptions to be thrown in
>>> these instances.
>> Incorrect. Exceptions can be thrown anywhere that undefined behavior
>> is specified. Overflow, underflow, and divide-by-zero all induce
>> undefined behavior.
> Hmmm I suppose that any thing can happen when undefined behavior
> is specified.
Yes, that's intentional latitude for the implementation to be able to
behave gracefully or not, as the situation demands.
> So writing a program that depends upon an undefined
> operation yielding a Nan would be a bad idea - wouldn't it?
No. If your implementation says it is IEEE-754 compliant, as many
are, it's a perfectly good idea. Just like it's a perfectly good idea
to depend on pthreads if you know your implementation is on POSIX, or
to depend on the presence of 64-bit integers if your implementation
tells you it supports long long, or...
>> b) is just a special case of a). I will agree that eliminating
>> undefined floating point behaviors will make C++ more predictable.
>>> Until one of the above (or maybe something else) is done. There can
>>> really be no unambiguous resolution to the problem of passing
>>> results from undefined operations from one machine to another.
>> Of course there can be. All you need to do is write a specification
>> for it that describes what happens in all cases, and it will be
>> unambiguous. If you can do this for ints that have nonportable values
>> greater than 32767, you can do it for floats and doubles, too.
> Besides writing such a specification, wouldn't C++ vendors have to agree
> to implement it?.
No, we're talking about a specification for serialization. That's not
the job of the C++ vendor. The C++ vendor already tells you
everything you need to know, e.g. "I support quiet NaN (or not)" and
"here's how to produce a quiet NaN if I support them," etc.
>>> Obviously, I believe that the adoption of b) above would result in
>>> fewer programs with hidden bugs.
>> That's almost certainly wrong. Floating point divide-by-zero is
>> almost never due to a program bug. And you can get the same effects
>> when dividing by a nonzero number if the result can't be represented.
> The kind of situation I'm thinking of is more like the following. I've got
> a program which among its operations is a matrix inversion. The program
> correctly implements the chosen algorithm. Now I load a near-singular
> matrix and invoke the matrix inversion operation. The sequence of
> operatons results in over/under flows in some intermediate results. No
> is thrown but some NaN's are propagated through the calculations. The
> final result Matrix may or may not have one or man Nan's. So now I
> have a result that is wrong but do not know it and have no way of knowing
If the implementation supports NaNs, of course you do. Check to see
if there are NaNs in the matrix. This is no different from a
calculation on ints that may have produced intermediate values greater
than 32767, except that the condition is easier to detect because NaNs
are sticky. In this regard, the usual implementation of floating
point math is much less error-prone than the usual implementation of
> In FORTRAN this was never a problem as the program aborts at the
> first overflow/underflow or whatever.
I guarantee you the FORTRAN spec doesn't say that the program will
abort upon "overflow/underflow or whatever." FORTRAN was probably the
first language to ever implement IEEE-754. Just google for "fortran
nan" and you'll see what I mean.
> What am I expected to do here? I could recode the matrix inversion
> to check each intermediate result to see if its a NaN?
No, NaNs propagate into every calculation they touch, so if you got
them in the matrix, you'll see them in the output. And if you
multiply that matrix by a vector, the resulting non-NaN elements will
still be meaningful (provided the original matrix was
well-conditioned, which is a whole other matter).
> I can't imagine that's what I'm expected to do. How do people
> handle this now?
I am not a numerics expert, but I know enough to understand that they
do handle it in predicatable ways, and that doing so is important to
them. Why don't you do a little research yourself? I'm sure a few
well-aimed web searches will yield a wealth of information.
-- Dave Abrahams Boost Consulting www.boost-consulting.com
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk