homepage

This is only on a PowerPC, specifically a p2020, running Debian 7.0 wheezy, python 2.7.3.
Calling round() seems to corrupt something (in the floating point state?) that causes subsequent exponentiation (the ** operator) to be wrong:
Python 2.7.3 (default, Jan 2 2013, 16:38:11)
[GCC 4.6.3] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> 2 ** -2
0.25
>>> round(1)
1.0
>>> 2 ** -2
0.1253019036571362
Cool. Huh?

Here's an example that uses only math library functions and float literals, presumably with a simpler code path:
Python 2.7.3 (default, Jan 2 2013, 16:38:11)
[GCC 4.6.3] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import math
>>> math.exp(-2.0)
0.1353352832366127
>>> math.ceil(1.0)
1.0
>>> math.exp(-2.0)
0.23902241864785234
By the way, the equivalent C program (using exp() and ceil()) on the same platform behaves correctly, so it's not an obvious libc or fpu problem.

I can't reproduce this on 2.7.4. Could you please test 2.7.4?

Hmm: it's not an obvious Python bug, either: both math.exp and math.ceil are simple wrappers around the libm functions, so there's little room for things to go wrong between Python and the OS.
Are you in a position to compile Python from source on your platform?
What was the equivalent C program you tried? What's the compiler on this platform? Note that gcc (I believe) evaluates math function calls for constants at compile time (using MPFR), rather than run time, which would mean if your C program simply does "exp(-2.0);" then it's not even using the OS libm.

Oh, sorry, I missed PowerPC. Please ignore my previous comment.

OK. My bad. I should have been tipped off that the program didn't need -lm to link. Output from C code:
0.135335
1
0.239022
Feel free to close. I'll have to look into this on my own. Thanks and sorry.

No problem; thanks for the update. By the way, if you do file an OS bug report of some form, it would be great if you could add a link to this issue. That might help anyone who encounters this in Python in the future.
Closing.