- Float (and complex) literals in source code were evaluated to full

precision only when running from a .py file; the same code loaded from a

.pyc (or .pyo) file could suffer numeric differences starting at about the

12th significant decimal digit. For example, on a machine with IEEE-754

floating arithmetic,

x = 9007199254740992.0

print long(x)

printed 9007199254740992 if run directly from .py, but 9007199254740000

if from a compiled (.pyc or .pyo) file. This was due to marshal using

str(float) instead of repr(float) when building code objects. marshal

now uses repr(float) instead, which should reproduce floats to full

machine precision (assuming the platform C float<->string I/O conversion

functions are of good quality).

This may cause floating-point results to change in some cases, and

usually for the better, but may also cause numerically unstable

algorithms to break.