[Orp-bugs] [ orp-Bugs-459040 ] Loss of precision of floating point

[<no...@so...>]

Bugs item #459040, was opened at 2001-09-05 22:36
You can respond by visiting: 
http://sourceforge.net/tracker/?func=detail&atid=384545&aid=459040&group_id=25513

Category: O3 jit
Group: None
>Status: Closed
Resolution: None
Priority: 5
Submitted By: Wu Gansha (wugansha)
Assigned to: Shi Xiaohua (xshi2)
Summary: Loss of precision of floating point

Initial Comment:
When I run SpecJBB2000 with O3, one validation test 
can't pass.
Observation:
The validation data demands printing of $285.00, but 
ORP prints $284.00.
Cause:
I found the internal representation of the data on 
ORP/O3 is 284.99999... (while the correct should be 
285.0, ORP/O1 is correct), and there's a bug in 
JBB2000, which prints 284.9999... as $284.00.
Though it's a bug but after all ORP/O3 is a little 
different with all other JVMs in dealing with floating 
point arithmetic.
Now I have a test case submitted to Xiaohua which will 
demo the difference.
Hopefully it could help.    

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

>Comment By: Shi Xiaohua (xshi2)
Date: 2001-09-12 00:57

Message:
Logged In: YES 
user_id=190089

Since the method and the variables are not constrained 
by "strictfp", such kind of precise lost can not be totally 
predicted and avoided, even though we fixed it in this 
perticular testcase, it could happen in some other places 
in the future. In fact , without "strictfp", it can not be 
considered as a real bug.
Now we can use O1 for this special method to skip it.

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

Comment By: Shi Xiaohua (xshi2)
Date: 2001-09-05 22:54

Message:
Logged In: YES 
user_id=190089

It can be reproduced by Gansha's testcase. I'll trace it 
and find the reason.

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

You can respond by visiting: 
http://sourceforge.net/tracker/?func=detail&atid=384545&aid=459040&group_id=25513