Re: [boost] [ublas] 1.33 regressions in test4 and test5

From: John Maddock (john_at_[hidden])
Date: 2005-05-13 06:59:46

> I should have caught up on my mail before I dove back into looking at test 
> results.  I'd just discovered what you say about an hour ago and have been 
> debugging the test to see if I could see where it should logically be 
> different.  I'm not even sure I know what  lower_, upper_, and size_ are 
> supposed to be for, or where/when things get allocated (if they do) when 
> one accesses something outside the current bounds.

Neither do I.

I've had a reply from Michael Stevens (below) that indicates that the 
problem is basically unfixable, and has always been present: so at present 
no-ones too sure what to do about it :-(

It's not helped by the fact that ublas's original authors are no longer 
around, so maintenance has basically been left to those users who care about 
the library, and probably have better things to do anyway...

John.

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Thanks for putting all the effort looking into the problem. The problem has
actually been present for a very long time. Actually it has been there for
several releases. Just that previously (the last 2 years) the test code
specifically disabled the testing of 'matrix_slice' proxies in test4 and
test5 to avoid the error!!!

After Boost 1.32.0 was out I made sure all such test 'fudge' code was 
disabled
so we could see the true state of affairs. The test4 and test5 failures are
not really regressions. Similarly the failure with VC8 in test2 is I believe
due to a bug in VC7 onwards that was specifically worked around before.

Fixing the bug with 'matrix_slice' proxies is very hard. The problem only
occurs with packed matrices and is due to the design of 'matrix_slice'. 
Sadly
a version that may work would probably be less efficient.

Before 1.33.0 is release we need to decide where to fudge the tests again. 
No
user has yet found it a problem in practice!

Regards,
 Michael