Speed up _decimal by another 10-15% by caching the thread local context
that was last accessed. In the pi benchmark (64-bit platform, prec=9),
_decimal is now only 1.5x slower than float.
2 files changed