Mail Archives: cygwin/2007/09/22/03:36:31
Mike MacFerrin wrote:
> I'm a grad student in a Compiler Tools course, and (as part of the course) we're compiling a subset of the Python language down to C, and from there to x86 Assmebly. Our code has to pass a lengthy series of test-suites for correctness (comparing the output of our compiled C Code with that of native Python code), and I stumbled across what seems to be a quirky bug in Cygwin Python 2.5.1 (the most recent install).
>
> To the point. If I run a simple two-line Python module:
>
> print 0.0
> print -0.0
>
> ...it gives the following output (as you'd expect):
>
> 0.0
> 0.0
Not what I would expect or want given that IEEE 754 supports negative zero.
> However, if I remove the first line:
>
> print -0.0
>
> ...I get...
>
> -0.0
>
> It seems to matter whether I've previously called "print 0.0" in the module.
There is a bug in the Python optimizer
<http://bugs.python.org/issue1678380> and patch at
<http://bugs.python.org/issue1678668>.
> No other "print <float>" commands make a difference.
It's not the printing that does it, it's the compilation. When a unit of
code is compiled in Python 2.5, all the things that compare equal to 0.0
are collapsed. If you had another sign-sensitive function, you could
call it and get the same results.
> Obviously, that seems strange (and doesn't follow the Python language definition).
Where is this "Python language definition" and where does this behavior
contradict it? I believe that this behavior is platform-specific and
undefined for Python.
> Run python and check for yourself if you wanna see it (from the shell it always prints "-0.0", but from a file it's inconsistent.
Not really. This is from the shell:
>>> print 0.0, -0.0
0.0 0.0
> Either way, it should just spit out "0.0").
I disagree. In the future you should expect -0.0 to consistently spit
out -0.0.
> In the meantime, I'm a bit frustrated that my code fails certain tests at home on Cygwin (due to no error of my own),
The error is in the tests that expect 0.0 and -0.0 to produce the same
results. Once this bug is fixed, I don't think they will on Linux
anymore either.
> Any suggestions or workarounds?
You can incorporate the patch I've linked, or you can get consistent
results if you use -(0.0). Also, dealing with IEEE 754 negative zeros
seems like something that would be useful in a compiler course so you
should bring it up with your professor.
--
Unsubscribe info: http://cygwin.com/ml/#unsubscribe-simple
Problem reports: http://cygwin.com/problems.html
Documentation: http://cygwin.com/docs.html
FAQ: http://cygwin.com/faq/
- Raw text -