[llvm-commits] Speeding up instruction selection

[Evan Cheng]

On Mar 26, 2008, at 2:29 PM, Roman Levenstein wrote:

> Hi,
>
> 2008/3/27, Evan Cheng <evan.cheng at apple.com>:
>>
>> On Mar 26, 2008, at 12:19 PM, Roman Levenstein wrote:
>>>>
>>>
>>> As long as there is a hope, I can wait ;-)
>>> And to use the opportunity, there is my pending patch for
>>> ScheduleDAGRRList.cpp still around (the one, where I suggested to  
>>> use
>>> std::set instead of priority queues based on heaps, see
>>> http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20080303/059076.html)
>>> .
>>> Evan played with it, but there was a problem with 176.gcc (which is
>>> probably unrelated) and somehow the review is stuck at the moment.  
>>> It
>>> would be nice to proceed with it. In the meantime, I could commit  
>>> some
>>> parts of it, that are not related to the std::set and are just  
>>> obvious
>>> optimizations that cannot break anything, e.g. fix for
>>> SumOfUnscheduledPredsOfSuccs.
>>
>>
>> Unfortunately your patch is triggering a bug somewhere which breaks
>> 176.gcc. Dan is trying to track it down but it's illusive. We don't
>> know what part of your patch is causing this so it's best not to  
>> check
>> anything in. I'm sure we'll eventually figure out the issue. Thanks
>> for your patience.
>
> OK. It is a pity that it break the test case, which is not publicly  
> available...
> Do you know any other testcases, which are also broken by this patch?
> What kind of failure do you have?

No other tests are broken. 176.gcc failure is a miscompilation.

>
>
>>> In the meantime, I could commit some
>>> parts of it, that are not related to the std::set and are just  
>>> obvious
>>> optimizations that cannot break anything, e.g. fix for
>>> SumOfUnscheduledPredsOfSuccs.
>
> But it is OK to commit this small fix for SumOfUnscheduledPredsOfSuccs
> only? This one simply cannot break anything because it does not change
> the logic at all. It just stops iterations earlier than in the current
> version and thus saves some cycles. And it is so obvious, that you can
> immediately see that it is correct.

Yes, that part is obviously safe. Thanks.

Evan

>
>
> -Roman
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits