[issue6713] Integer & Long types: Performance improvement of 1.6x to 2x for base 10 conversions

[Mark Dickinson]

Mark Dickinson <dickinsm at gmail.com> added the comment:

Updated patch, with minor changes:
  - remove an incorrect Py_DECREF(str)
  - rename _PyLong_ToDecimal;  no need for the _Py prefix, since this
    function isn't shared across files
  - absorb special case for 0 into the rest of the code
  - whitespace and indentation fixes

Not that it matters much, but it's curious that on my machine (gcc-4.2, OS 
X 10.6.1, x64-64) it's significantly faster (~6% increase in str() speed 
for large integers) to use the line:

    pout[j] = z - (twodigits)hi * _PyLong_DECIMAL_BASE;

in the middle of the inner loop, rather than the line:

    pout[j] = z - hi * _PyLong_DECIMAL_BASE;

I'm wondering whether this is just a quirk of my OS/compiler combination, 
or whether there's a good reason for this difference.  The lines are 
functionally equivalent, since the result is reduced modulo 2**32 either 
way, but the first line involves a 32x32->64 multiplication and a 64-bit 
subtraction, where the second involves a 32x32->32 multiplication and a 
32-bit subtraction;  the generated assembly code for the second line is 
also one instruction shorter (there's a move opcode saved somewhere).

----------
Added file: http://bugs.python.org/file14902/long_decimal_conversion_py3k_2.patch

_______________________________________
Python tracker <report at bugs.python.org>
<http://bugs.python.org/issue6713>
_______________________________________