On Fri, Oct 19, 2012 at 8:36 AM, Victor Stinner <victor.stinner@gmail.com> wrote:
2012/10/19 Benjamin Peterson <benjamin@python.org>:
> It would be interesting to see how common it is for strings which have
> their hash computed to be compared.

I implemented a quick hack. When running "./python -m test test_os":
Python calls PyUnicode_RichCompare() 15206 times with Py_EQ or Py_NE
operator. In 41.4% (6295 calls), the hash of the two operands is
known. In 41.2% (6262 times on 15206), the hash of the two operands
are known *and are different*!

The hit rate may depend since when the process was started. For
example, in a fresh interpreter: the hit rate is only 7% (189 hit /
2703 calls).

When running the test suite, the hit rate is around 80% (hashs are
known in 90%) after running 70 tests. At the same time, the average of
string length is 4.1 characters and quite all strings are pure ASCII.

I create the issue http://bugs.python.org/issue16286 to discuss this

If you want to measure the performance impact compared to a clean build then you can use the unladen benchmarks as it contains several Python 3-compatible benchmarks now.