[NumPy-Tickets] [NumPy] #1604: distutils/command/config.py binary search needs floordiv '//' to work under Python3

NumPy Trac numpy-tickets@scipy....
Sat Sep 4 04:28:04 CDT 2010


#1604: distutils/command/config.py binary search needs floordiv '//' to work under
Python3
--------------------+-------------------------------------------------------
 Reporter:  cb123   |       Owner:  somebody
     Type:  defect  |      Status:  new     
 Priority:  normal  |   Milestone:  1.5.0   
Component:  Other   |     Version:          
 Keywords:          |  
--------------------+-------------------------------------------------------
 There is a bug in the build under Python3.  The symptom is an infinite
 loop of recompiling a _configtest.c file.

 numpy/distutils/command/config.py uses (hi+lo) / 2 in a binary search.  In
 py2 '/' is "integer floor div" but in py3 is "float converting div".  The
 fix is just using '//' which has been available since python2.2 (or '>> 1'
 instead of '/ 2').  As coded, binary search requires floordiv to converge.

 Beyond the non-convergence issue, the tests themselves fail erroneously.
 Using a float divide creates a float.  Upon conversion to a string via
 "%(size)s", that creates constants in the generated code like "<= 16.0".
 Though the value of the expression is integral, as required, these being
 ".0" instead of round numbers cause the left operand of the inequality to
 be promoted to double and '<=' to be evaluated in compile-time floating
 point arithmetic.  Some compilers, e.g. gcc-4.5.1, do not support FP math
 for compile-time constants like an array length even if it evalutes to an
 integer.  Other compilers, e.g. clang-1.1, g++-4.5.1, work fine with it.
 I have not looked into what relevant standards say, but using '//' makes
 forward python behavior imitate the python-2.7 and early behavior and
 seems a reasonable fix.

 Ok.  This is possibly the longest explanation I have ever had for what
 amounts to a one-character patch. :)

-- 
Ticket URL: <http://projects.scipy.org/numpy/ticket/1604>
NumPy <http://projects.scipy.org/numpy>
My example project


More information about the NumPy-Tickets mailing list