C++ Standard Library IOStreams floating-point problem

[Steven Schweda]

A different floating-point thread reminded me
of this C++ problem.

According to the HP C++ User's Guide:

http://h71000.www7.hp.com/commercial/cplus/docs/ugv_stl.html#_sec

7.1.7 IOStreams Expects Default Floating-Point Format

The C++ standard library IOStreams expects floating-point values in the
default floating-point format for each platform: G_FLOAT on Alpha
systems and IEEE on I64 systems. Using standard library IOStreams for
processing floating-point values in a different format (for example, in
a program compiled /FLOAT=IEEE on Alpha or /FLOAT=G_FLOAT on I64) is not
supported. The C++ class library does not have this restriction.

Here, "is not supported" is a gentle euphemism
for "doesn't work". The current SETI@home
client, for example, uses IOStreams, and also
expects IEEE floating-point. The restriction
described above makes this impossible on Alpha.

It would be nice if an IEEE edition of this
library were available on Alpha. Any hope?

[Travis Craig]

Wow, you're right, Steven. Now I'm wondering how we dodged that bullet with all the code we switched to the standard library a couple of years ago. Nothing I can think off off-hand except luck.

Time to search the source again, I guess. I don't know if it broke anything of ours, but I'd vote for an IEEE version of the library, too.

--Travis Craig

[Steven Schweda]

The secret is to avoid doing the wrong
thing(s). The offending SETI@home code
comprises many statements like:

in >> center;

where "in" is a std::istringstream, and
"center" is a double. The use is like atof()
(or atol[l]() for the integers), apparently,
but so much more cute. They use this stuff
to parse the data in a some ".xml" file(s).
Works fine for integers, too, but the
results are nonsense for floating-point
variables, due to the format difference.

The SETI@home code has hundreds of these,
making manual conversion to harmless, old
code more work than I can justify.

[Martin Vorlaender]

Steven,

>>>
in >> center;
...
The SETI@home code has hundreds of these,
making manual conversion to harmless, old
code more work than I can justify.
<<<

Perhaps one could subclass IOStreams, overwriting operator>> to do the Right Thing(TM). That way not the uses of "in" et al. would have to be reworked, but just the declarations.

Just an idea...

cu,
Martin

[Steven Schweda]

> Perhaps one could subclass IOStreams [...]

As soon as I learn enough C++ to give this
a try, I may give this a try. (Unless
someone who knows something tells me before
then that it's really hopeless.)