llvm dev - Feb 2020 - InstCombiner bug

I believe there's a bug
in InstCombiner::reassociateShiftAmtsOfTwoSameDirectionShifts(),
illustrated by https://rise4fun.com/Alive/ne8y. I'll post a bug on the
tracker as soon as I can get an account set up.

Short form: when identifying the two shifts to combine, it does so while
ignoring any zero-extensions of those values. In this case, the values are
both i1 zero-extended to i3s, so when taking their sum, SimplifyAddInst
emits an xor, causing the shift to basically disappear.

I could write more, but the Alive link is probably more concise.

Let me know what you think / if there's some followup needed, etc.
Thanks!
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20200204/3a1bce75/attachment.html>

Looks definitely a bug, please report it, thanks!

John


On 2/4/20 10:09 AM, Rob Springer via llvm-dev wrote:
> I believe there's a bug 
> in InstCombiner::reassociateShiftAmtsOfTwoSameDirectionShifts(), 
> illustrated by https://rise4fun.com/Alive/ne8y. I'll post a bug on the 
> tracker as soon as I can get an account set up.
> 
> Short form: when identifying the two shifts to combine, it does so while 
> ignoring any zero-extensions of those values. In this case, the 
> values are both i1 zero-extended to i3s, so when taking their sum, 
> SimplifyAddInst emits an xor, causing the shift to basically disappear.
> 
> I could write more, but the Alive link is probably more concise.
> 
> Let me know what you think / if there's some followup needed, etc.
> Thanks!
> 
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>