Bugs item #678265, was opened at 2003-01-31 12:10 Message generated for change (Comment added) made by mdr0 You can respond by visiting: https://sourceforge.net/tracker/?func=detail&atid=105470&aid=678265&group_id=5470

Category: Python Interpreter Core Group: None Status: Open Resolution: None Priority: 5 Submitted By: Neal Norwitz (nnorwitz) Assigned to: Nobody/Anonymous (nobody) Summary: test_coercion fails on AIX

Initial Comment: Tim, I assigned this to you to see if you had any insight/ideas about this problem.

test_coercion fails on AIX. All the failures are due to the test expecting n+0j, but the output is n-0j.

Here's an example:

- <MethodNumber 1> / (2+0j) = (0.5+0j) ? ^ + <MethodNumber 1> / (2+0j) = (0.5-0j) ? ^

I have another compiler, so I&#039;ll probably try that. Especially since it&#039;s gcc. :-)

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

Comment By: Mark D. Roth (mdr0) Date: 2004-03-10 10:52

Message: Logged In: YES user_id=994239

Looks like the test is still failing, even with this patch.

Any idea what that xlc option might have been? I looked through the list of options, but it wasn't immediately obvious to me which one might have helped.

Any other ideas? Thanks!

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

Comment By: Neal Norwitz (nnorwitz) Date: 2004-03-09 18:16

Message: Logged In: YES user_id=33168

IIRC, there was a command line switch to xlc which would change the math behaviour. I don't recall trying it but from the description it seemed like it would work.

Neil, I'll try to look at your patch. If you think it's not controversial, just check it in. I'll definitely review it in that case. :-)

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

Comment By: Neil Schemenauer (nascheme) Date: 2004-03-09 17:22

Message: Logged In: YES user_id=35752

I really didn't intend for test_coercion to become a platform fp quirks test. The attached patch changes the test to format floating point numbers to one decimal place.

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

Comment By: Mark D. Roth (mdr0) Date: 2004-03-09 15:35

Message: Logged In: YES user_id=994239

I've run into this same problem under AIX 4.3.3 with IBM compiler version 6.0.0.5 and under AIX 5.1 with IBM compiler version 6.0.0.4.

Is there a work-around, other than building with gcc?

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

Comment By: Tim Peters (tim_one) Date: 2003-01-31 15:08

Message: Logged In: YES user_id=31435

You really need a platform wizard for something like this. If there are any compiler options with descriptions like "generate IEEE-754 compatible code" (or "stricter 754 xyz"), or "disable generation of fused multiply-add", they would be relevant.

OTOH, Python promises nothing about how the sign bit of a floating zero behaves, so it would be OK by me if the test were changed not to produce complex results with zero imaginary members.

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

Comment By: Neal Norwitz (nnorwitz) Date: 2003-01-31 13:26

Message: Logged In: YES user_id=33168

The test passes with gcc 3.2. The C compiler is in vac.C version 4.4.0.3 (lslpp -l vac.C).

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