[PATCH] Fix a performance problem in gep(gep ...) merging

[Xinliang David Li]

IIUC, GEP merging can be harmful regardless of where the target is in loop
context or not. Wei's heuristic filters out lots of harmful cases. A more
general fix might need introduce more sophisticated cost analysis -- it is
possible (as mentioned by Quentin) for this simple heuristic to filter out
beneficial merging, but test cases are needed for that.

David

On Tue, Apr 14, 2015 at 1:16 AM, Nick Lewycky <nicholas at mxc.ca> wrote:

> Quentin Colombet wrote:
>
>> The thing with the "both constants" approach is that we may miss some
>> cases where an addressing mode may apply. In particular, on x86, when one
>> argument of the add is a constant, we would be able to match reg1 + reg2 +
>> displacement.
>>
>> That being said, your numbers imply this is not a problem in practice,
>> and most targets do not support reg1 + reg2 + reg3/imm addressing mode.
>> Anyway, we could recover from that later on if needed.
>>
>> So LGTM.
>>
>
> I might be misunderstanding why this patch helps, please bear with me. Is
> this not a special case of a general problem, mainly that we're merging
> something into a loop that was previously two operations one inside and one
> outside the loop? Why wouldn't this problem come up in other ways? We
> already have other problems where we introduce loop carried dependencies
> where ones previously didn't exist, shouldn't we approach this by adding an
> API to detect those and then making transforms conditional on that?
>
> Nick
>
>
>
>> Thanks Wei!
>> -Quentin
>>
>>
>> REPOSITORY
>>    rL LLVM
>>
>> http://reviews.llvm.org/D8911
>>
>> EMAIL PREFERENCES
>>    http://reviews.llvm.org/settings/panel/emailpreferences/
>>
>>
>>
>>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150414/cf062979/attachment.html>