[llvm] r222142 - [Reassociate] As the expression tree is rewritten make sure the operands are

[Chad Rosier]

Hi Kevin,
The failure looks to only happen at -Ofast with -mcpu=cortex-a57. 
Unfortunately, I tested with cortex-a53 and it doesn't seem to reproduce.
Sorry for the noise! :/

I'll debug the issue now and revert if I can't determine the issue shortly.

 Chad

> Hi Chad,
>
> We found this commit introduced a correctness problem on AArch64. Coremark
> is failed on performance test due to result miscompare, and 403.gcc in
> SPEC2006 got problem in run time.
>
> I think it should be easy to reproduce those problems. Please have a look
> at this.
>
> Thanks,
> Kevin
>
> 2014-11-18 19:12 GMT+00:00 Chad Rosier <mcrosier at codeaurora.org>:
>
>> Thanks for the info, Suyong.  I'm pretty sure the change is trivially
>> correct.  It's very likely this commit perturbed a lot of IR.  Prior to
>> my
>> recent commits the IR that was coming out of the pass was largely
>> unordered.
>>
>>  Chad
>>
>> > Hi Chad, Timur
>> >
>> > Just wanted to make a point here.
>> >
>> > This patch actually eased further transformations like SLP
>> Vectoization.
>> >
>> > Test case :
>> >
>> > sum = a[0] + a[1] + a[2] + a[3]
>> >
>> > Earlier, without this patch, re-associate pass would re-organize this
>> > expression into
>> >
>> > sum = (((a[1] + a[0]) + a[2]) + a[3])
>> >
>> > which made SLP vectorizer difficult to identify this type of
>> horizontal
>> > reduction.
>> > With this patch, the expression would remain
>> >
>> > sum = (((a[0] + a[1]) + a[2]) + a[3])
>> >
>> > which made SLP vectorizer task simpler.
>> >
>> > Patch holds good value for further optimization passes,
>> > though as Timur reports that its breaking some test cases.
>> >
>> > --
>> > With regards,
>> > Suyog Sarda
>> >
>>
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>
>
>
>
> --
> Best Regards,
>
> Kevin Qin
>