[LLVMdev] [RFC] SCEV Enhancements

David A. Greene dag at cray.com
Fri Feb 24 15:41:05 PST 2012

Dan Gohman <gohman at apple.com> writes:

>> We already do that as much as possible but it doesn't cover every case
>> currently.  I will see if we can cover more cases.
> Ok. This is generally worthwhile for other reasons as well.


>> Sure.  I'll try tweaking instcombine first and see where that gets us.
>> I have some pretty good testcases.  I'd be thrilled if we could just
>> dump these SCEV changes.  That's why I did the original GEP instcombine
>> stuff.  I actually ripped out the SCEV changes after that and
>> performance tanked.  :(
> Ok, this seems the best approach. If you hit something you can't fix,
> we can re-evaluate.

Sounds reasonable.

>>> [0] or i1* if you feel like being a purist ;-).
>> Err?  What are the scaling semantics for i1?  I don't immediately see
>> that it should be the same as i8.
> Array elements are guaranteed to start on address-unit boundaries, so i1
> elements get padded to the nearest address-unit size.

Still looks strange to me.  :)


More information about the llvm-dev mailing list