[ python-Bugs-1245381 ] log() on a big number fails on powerpc64

[SourceForge.net]

Bugs item #1245381, was opened at 2005-07-26 11:38
Message generated for change (Comment added) made by heffler
You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=105470&aid=1245381&group_id=5470

Please note that this message will contain a full copy of the comment thread,
including the initial issue submission, for this request,
not just the latest update.
Category: Python Interpreter Core
Group: Python 2.3
Status: Open
Resolution: None
Priority: 5
Submitted By: Jiri Dluhos (jdluhos)
Assigned to: Nobody/Anonymous (nobody)
Summary: log() on a big number fails on powerpc64

Initial Comment:
The log() function seems to give incorrect results for
large numbers on a 64-bit powerpc architecture. The
test_math program from the Python test suite fails with
the following message:

math module, testing with eps 1e-05
constants
acos
asin
atan
atan2
ceil
cos
cosh
degrees
exp
fabs
floor
fmod
frexp
hypot
ldexp
log
Traceback (most recent call last):
  File "test_math.py", line 117, in ?
    testit('log(10**40, 10)', math.log(10**40, 10), 40)
  File "test_math.py", line 13, in testit
    raise TestFailed, '%s returned %f, expected %f'%test.test_support.TestFailed: log(10**40, 10) returned
21.938200, expected 40.000000

Tested on a SMP IBM PowerPC machine, with Python 2.3.4
(64-bit build).

----------------------------------------------------------------------

Comment By: Marvin Heffler (heffler)
Date: 2005-10-03 15:06

Message:
Logged In: YES 
user_id=298758

I have tried this with python 2.4, 2.4.1, and now 2.4.2. All of 
them face the same problem where log won't work correctly 
on ppc64. The same code works just fine for me on ia32, 
ia64, ppc32, s390, s390x, and x86_64. Only ppc64 sees the 
problem.

The problem can be reproduced easily on a ppc64 system 
with python having been built as a 64-bit binary. Just start 
python in interactive mode and enter the following commands:

   >>import math
   >>math.log(10**40, 10)

The correct answer is 40, but on ppc64 the result comes 
back as 21.938200260161128. I'm not very familiar with the 
python source code, but with a little playing I came up with a 
way to correct the problem. I don't know why the problem 
gets resolved, but it does. By adding a simple printf near the 
end of the loghelper function in Modules/mathmodule.c, the 
problem goes away. Here's a diff showing the change I made 
to get things working correctly:

--- mathmodule.c.SAV    2005-10-03 12:56:38.468014112 -
0500
+++ mathmodule.c        2005-10-03 12:57:35.192044904 -
0500
@@ -224,6 +224,7 @@
                   log(x) + log(2) * e * SHIFT.
                   CAUTION:  e*SHIFT may overflow using int 
arithmetic,
                   so force use of double. */
+               printf("", (e * (double)SHIFT));
                x = func(x) + (e * (double)SHIFT) * func(2.0);
                return PyFloat_FromDouble(x);
        }

Maybe someone who is more familiar with the python code 
can figure out why a change like this would make a difference.

----------------------------------------------------------------------

Comment By: Neal Norwitz (nnorwitz)
Date: 2005-09-30 01:36

Message:
Logged In: YES 
user_id=33168

Can you test with newer versions of Python, preferrably 
2.4.2 or current CVS?  2.3.4 worked for me on 64-bit Opteron.

----------------------------------------------------------------------

You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=105470&aid=1245381&group_id=5470