The test suite is not a good benchmark: it also tests decimal.py. For
numerical performance I'm running:

  cd Modules/_decimal/tests
  ../../../python bench.py

You can hit Ctrl-C after the first cdecimal result, since that's usually
already a pretty good indicator of overall performance. On 64-bit, for
9 digits of precision cdecimal is currently only around 1.5 times slower
than float. I want to keep that.

Running an unpatched _decimal.c three times gives (Linux, 64-bit, Core2 Duo):

0.162576s  0.165146s  0.163242s

With your second patch:

0.204383s  0.204383s  0.206919s

> Regarding the failing test:
> It appears that the hackcheck() method in typeobject.c is responsible for this failure: 

Thanks for the analysis. Perhaps Martin can comment on that.