Comment 9 for bug 871083

Extra data points here:

The amd64 misbuild produced the same output both times, and was on the same host (allspice).
The armel misbuild only occurred the first time (on a babbage), but was fine the second time (on a panda).

Given that this seems to be reproducible (and potentially hardware-specific), this might be debuggable, with access to the hardware that exhibits the issue.