Re: GCC-3.3.4 release status report

[Richard Guenther <rguenth@tat.physik.uni-tuebingen.de>]

Joe Buck wrote:
> On Sat, Mar 27, 2004 at 11:09:52AM +0100, Gabriel Dos Reis wrote:
> 
>>  The number of open PRs targetted for 3.3.4 has grown up to 46
>>(from 41 last report).
>>
>> Some PRs have been fixed, others are new.
>>The set of critical PRs is now 8 and is different from what we got
>>last report:
>>
> ([3.3 regression] -O2 -funroll-loop miscompiles POOMA testcase)
> 
> Now *this* is one that I'd like to see fixed!  It's likely to be tractable
> because 3.2.3 worked, and it's nasty.

I'm the reporter of the bug (and there is another arch-specific for 
powerpc, maybe related, PR13222), and the bug looks indeed serious.  But 
I personally am using already 3.4 for production use and tree-ssa for 
testing - and POOMA based testcases are huge - even if this one was 
reduced a lot already.  Someone with good understanding of the loop 
unroller and rtl needs to look at this.  At least warn about use of 
-funroll-loops as it caused that many regressions in the past.

One problem is that we don't excercise those commonly (in scientific 
apps) used options in any -O level, so this gets too less beating on in 
regular C apps which would result in so much nicer testcases...

So, in conclusion, I'd be not too unhappy if this is _not_ fixed for 3.3.4.

>>  optimization/11841
> 
> 
> Likewise this one; could be related to 13653 (another -funroll-loop bug
> new since 3.2.3)

Let's hope fixing one of 11841, 13222 or 13653 fixes the other cases. 
Or simply strip off handling of "unusual" cases from the unroller and 
rather pessimize the odd cases than producing wrong code.

Richard.